text
stringlengths 1
2.56M
| id
stringlengths 40
40
| metadata
dict |
|---|---|---|
\section{Introduction}
Bayesian uncertainty quantification is useful for calibrating physical models to observed data. As well
as inferring the parameters that produce the best fit between a model and observed data, Bayesian methods can
also identify the range of parameters that are consistent with observations and allow for prior beliefs to be
incorporated into inferences. This means that not only can predictions be made but also their uncertainty can be
quantified. In demography, Bayesian analysis is used to forecast the global population
\citep{gerland2014world}. In defence systems, Bayesian analysis is used to track objects from radar signals
\citep{arulampalam2002tutorial}. And in computational neuroscience Bayesian analysis is used to compare
different models of brain connectivity and to estimate physiological parameters in mechanistic models
\citep{kiebel2008dynamic}. Many more examples can be found in the references of
\citep{bishop2006pattern,gelman2013bayesian,lunn2012bugs}.
We focus here on problems which require the use of Markov Chain Monte Carlo (MCMC), a widely applicable
methodology for generating samples approximately drawn from the posterior distribution of model parameters
given observed data. MCMC is useful for problems where a parametric closed form solution for the posterior
distribution cannot be found. MCMC became popular in the statistical community with the re-discovery of Gibbs
sampling \citep{smith1993bayesian}, and the development of the BUGS software \citep{lunn2012bugs}. More
recently it has been found that methods which use derivatives of the posterior distribution with respect to
model parameters, such as the Metropolis Adjusted Langevin Algorithm (MALA) and Hamiltonian Monte Carlo (HMC)
tend to generate samples more efficiently than methods which do not require derivatives \citep{hoffman2014no}.
HMC is used in the popular Stan software \citep{carpenter2017stan}. From the perspective of a C++ programmer,
the limitations of Stan are as follows: it may take a significant investment of effort to get started.
Either the C++ code has to be translated into the Stan modelling language. Or, alternatively, C++ code can be
called from Stan, but it may be challenging to (efficiently) obtain the derivatives that Stan needs in order
to sample efficiently.
The software that we present includes (i) our own implementation of a derivative-based MCMC sampler called
simplified manifold MALA (smMALA). This sampler can be easily be used in conjunction with C++ codes for Bayesian data
analysis, (ii) Stan's MCMC sampler with derivatives computed using \dcocpp{}, an industrial standard tool for
efficient derivative computation.
An alternative approach to the one presented in this paper would be simply to use Stan as a stand-alone tool
without the smMALA sampler and without \dcocpp{}. Determining the most efficient MCMC sampler for a given
problem is still an active area of research, but at least within computational neuroscience, it has been found the
smMALA performs better than HMC methods on certain problems \citep{sengupta2016gradient}. Determining the most
appropriate method for computing derivatives will depend on both the user and the problem at hand. In many
applications Algorithmic Differentiation (AD) is needed for the reasons given in Section \ref{sec:deriv}. The
Stan Math Library includes an open-source AD tool. Commercial AD tools such as \dcocpp{} offer a richer set
of features than open-source tools, and these features may be needed in order to optimize derivative
computations. For example, the computations done using the Eigen linear algebra library \citep{eigenweb} can
be differentiated either using the Stan Math Library or using \dcocpp{} but there are cases where \dcocpp{}
computes derivatives more efficiently than the Stan Math Library \citep{peltzer2019eigen}. The aim of the
software we present is to offer a range of options that both make it easy to get started and to tune
performance.
\section{Methods for spectral time-series analysis}
\pagestyle{main}
\subsection{Whittle likelihood}
The software presented in this paper is targeted at differential equation models with a stable equilibrium
point and stochastic input. We refer to such models as stable SDEs. The methods that are implemented
assume that the system is operating in a regime where we can approximate the dynamics through linearization
around the stable fixed point. If the time-series data is stationary this is a reasonable assumption. Note
that the underlying model may be capable of operating in nonlinear regimes such as limit cycles or chaos in
addition to approximately linear dynamics. However, parameter estimation using data in nonlinear regimes
quickly becomes intractable - see Chapter 2 of \citep{maybank2019bayesian}. The stability and linearity
assumptions are commonly made in the computational neuroscience literature, see for example
\citep{moran2009dynamic}.
In order to simplify the presentation we illustrate the software using a linear state-space model, which is of
the form,
\begin{equation} \label{eq:de}
d\mathbf{X}(t) = A\, \mathbf{X}(t)dt + \mathbf{P}(t),
\end{equation}
\noindent where the term $A\, \mathbf{X}(t)$ represents the deterministic evolution of the system and $\mathbf{P}(t)$
represents the noisy input. The example we analyze in this paper is the noise-driven harmonic oscillator,
which is a linear state-space model with
\begin{equation}
A =
\begin{pmatrix}
0 & 1 \\
-\omega_0^2 & -2 \zeta \omega_0
\end{pmatrix},
\quad
\mathbf{P}(t) =
\begin{pmatrix}
0 \\
dW(t)
\end{pmatrix},
\end{equation}
\noindent and where $dW(t)$ represents a white noise process with variance $\sigma_{in}^2$. The observations are
modelled as $Y_k = X_0(k\cdot \Delta t) + \epsilon_k$ with $\epsilon_k \sim N(0, \sigma_{obs}^2)$.
Our aim is to infer model parameters ($\omega_0, \zeta, \sigma_{in}$) from time-series data. This could be
done in the time-domain using a Kalman filter, but it is often more computationally efficient to do inference
in the frequency domain \citep{bojak2005modeling,maybank2017fast}. In the case where we only have a single
output (indexed by $i$) and a single input (indexed by $j$), we can compute a likelihood of the parameters
$\theta$ in the frequency domain through the following steps.
\begin{enumerate}
\item Compute the (left and right) eigendecomposition of $A$, such that,
\begin{equation}
A \mathcal{R} = \Lambda \mathcal{R}, \quad \mathcal{L} A = \mathcal{L} \Lambda, \quad \mathcal{L} \mathcal{R} = \mathrm{diag}(c) \\
\end{equation}
where $\mathrm{diag}(c)$ is a diagonal matrix, such that $c_i$ is the dot product of the $i$th left
eigenvector with the $i$th right eigenvector.
\item Compute $ij$th element of the transfer matrix for frequencies $\omega_1,\ldots,\omega_K$,
\begin{equation} \label{eq:transfer}
\mathcal{T}(\omega) = \mathcal{R}\ \mathrm{diag}\bigg[ \frac{1}{c_k (i\omega - \lambda_k)} \bigg]\ \mathcal{L}.
\end{equation}
\item Evaluate the spectral density for component $i$ of $\mathbf{X}(t)$, $f_{X_i}(\omega)$, and the spectral
density for the observed time-series, $f_{Y}(\omega)$,
\begin{align}
\label{eq:spec_eq}
f_{X_i}(\omega) &= |\mathcal{T}_{ij}(\omega)|^2\ f_{P_j}(\omega), \\
f_{Y}(\omega) &= f_{X_i}(\omega) + \sigma_{obs}^2 \Delta t,
\end{align}
\noindent where $f_{P_j}(\omega)$ is the spectral density for component $j$ of $\mathbf{P}(t)$.
\item Evaluate the Whittle likelihood,
\begin{equation} \label{eq:whittle}
p(y_0,\ldots,y_{n-1}|\theta) = p(S_0,\ldots,S_{n-1}|\theta) \approx \prod_{k=1}^{n/2-1}
\frac{1}{f_Y(\omega_k)} \exp\left[-\frac{S_k}{f_Y(\omega_k)}\right],
\end{equation}
where $\{S_k\}$ is the Discrete Fourier Transform of $\{y_k\}$. Note that $\theta$ represents a parameter set
(e.g. specific values of $\omega_0, \zeta, \sigma_{in}$) that determines the spectral density.
\end{enumerate}
The matrix $A$ that parameterizes a linear state-space model is typically non-symmetric, which means that
eigenvectors and eigenvalues will be complex-valued. We use Eigen-AD \citep{peltzer2019eigen}, a fork of the \cpp{} linear
algebra library Eigen \citep{eigenweb}. Eigen is templated which facilitates the application of AD by overloading tools and Eigen-AD
provides further optimizations for such tools. The operations above require an AD tool that supports
differentiation of complex variables. AD of complex variables is considered in \citep{pusch1995ad}. It is
available from release 3.2.0 of the Stan Math Library and in \dcocpp{} from release 3.4.3.
\subsection{Markov Chain Monte Carlo}
\label{sec:mcmc}
In the context of Bayesian uncertainty quantification, we are interested in generating samples from the
following probability distribution,
\begin{equation} \label{eq:bayes} p(\theta|y) \propto p(y|\theta) p(\theta),
\end{equation}
\noindent where $p(y|\theta)$ is the likelihood of the parameter set $\theta$ given observed data
$y_0,\ldots,\allowbreak y_{n-1}$, and $p(\theta)$ is the prior distribution of the parameters. In many
applications the likelihood $p(y|\theta)$ is based on some physical model where we cannot derive a closed form expression for the
posterior density $p(\theta|y)$. Markov Chain Monte Carlo (MCMC) has emerged over the last 30 years as one of
the most generally applicable and widely used framework for generating samples from the posterior distribution
\citep{gelman2013bayesian,green2015bayesian}. The software used in this paper makes use of two MCMC
algorithms: the No U-Turn Sampler (NUTS) \citep{hoffman2014no} and the simplified manifold Metropolis Adjusted
Langevin Algorithm (smMALA) \citep{girolami2011riemann}. NUTS is called via the Stan environment
\citep{carpenter2017stan}. It is a variant of Hamiltonian Monte Carlo (HMC), which uses the gradient (first
derivative) of the posterior density, whereas smMALA uses the gradient and Hessian (first and second derivatives) of
the posterior density, smMALA is described in Algorithm \ref{alg:smmala}.
The error in estimates obtained from MCMC is approximately $C / \sqrt{N}$, where $N$ is the number of MCMC
iterations and $C$ is some problem-dependent constant. In general it is not possible to demonstrate that MCMC
has converged, but there are several diagnostics that can indicate non-convergence, see Section 11.4 of
\citep{gelman2013bayesian} for more detail. Briefly, there are two phases of MCMC sampling: burn-in and the
stationary phase. Burn-in is finished when we are in the region of the parameter space containing the true
parameters. In this paper we restrict ourselves to synthetic data examples. In this case it is
straight-forward to assess whether the sampler is burnt in by checking whether the true parameters used to
simulate the data are contained in the credible intervals obtained from the generated samples. In real data
applications, it is good practice to test MCMC sampling on a synthetic data problem that is analogous to the
real data problem. During the stationary phase we assess convergence rate using the effective sample size,
\textit{N Eff}. If we were able to generate independent samples from the posterior then the constant $C$ is
$\mathcal{O}(1)$. MCMC methods generate correlated samples, in which case $C$ may be $\gg 1$. A small
\textit{N Eff} (relative to $N$) is an indicator of slow convergence.
If we are sampling from a multivariate target distribution, \textit{N Eff} for the $i$th component is equal to,
\begin{equation} \frac{S}{1 + 2 \sum_k \hat{\rho}_i(k)},
\end{equation}
\noindent where $S$ is the number of samples obtained during the stationary period, and $\hat{\rho}_i(k)$, is
an estimate of the autocorrelation at lag $k$ for the $i$th component of the samples. This expression can be
derived from writing down the variance of the average of a correlated sequence (Chapter 11 of
\citep{gelman2013bayesian}). The key point to note is that if the autocorrelation decays slowly \textit{N Eff} will
be relatively small.
\begin{algorithm}[ht]
\KwIn{Data, $y$; Initial value for $\theta = (\theta_1,\ldots,\theta_N)$;
\newline
Likelihood $l(y|\theta)$; Prior distribution $p(\theta)$.}
\Parameter{Step size, $h$; Number of iterations, $I$}
\For{$i = 2,\ldots,I$ }{
Evaluate gradient and Hessian of the unnormalized log posterior,\newline
\Indp $g_{\theta} = \nabla \bigg[\log[ \ l(y|\theta)\ p(\theta) \ ] \bigg]$ and $\mathbf{H}_{\theta}$\;
Set $C = h^2 \mathbf{H}_{\theta}^{-1}$, and $m = \theta + \frac{1}{2} C\, g_{\theta}$\;
Propose $\theta^* \sim q(\theta^*|\theta)=N(m, C)$\;
Evaluate acceptance ratio, $\alpha(\theta, \theta^*) =
\mathrm{min}\left[1, \dfrac{l(y|\theta^*) p(\theta^*) q(\theta|\theta^*)
}{l(y|\theta) p(\theta) q(\theta^*|\theta)} \right]$\;
Draw $u \sim \mathrm{Uniform}(0,1)$\;
\lIf{$u < \alpha(\theta, \theta^*)$}{Set $\theta = \theta^*$ }
}
\caption{smMALA}\label{alg:smmala}
\end{algorithm}
\subsection{Derivative computation}
\label{sec:deriv}
A central finite difference approximation to the first and second derivatives of the function $F(\theta): \mathbb{R}^N \rightarrow
\mathbb{R}^M$ can be computed as,
\begin{equation}
\label{eq:d1}
\frac{\partial F}{\partial \theta_i} \approx \frac{F(\theta+h e_i) - F(\theta-h e_i)}{2 h}, \qquad
\frac{\partial^2 F}{\partial \theta_i \partial \theta_j} \approx
\frac{\dfrac{\partial F}{\partial \theta_i}\big(\theta + h e_j\big) -
\dfrac{\partial F}{\partial \theta_i}\big(\theta-h e_i\big)}{2 h}, \nonumber
\end{equation}
\noindent where $e_i$ is the $i$th Cartesian basis vector, and $h$ is a user-defined step-size. For
first-order derivatives the default value we use is $\sqrt{\epsilon}|\theta_i|$, where $\epsilon$ is machine
epsilon for double types, i.e., if $\theta_i$is $\mathcal{O}(1)$, $h\approx10^{-8}$. For second derivatives
the default value is $\epsilon^{1/3}|\theta_i|$, so that $h\approx 5\cdot 10^{-6}$. More details on the
optimal step size in finite difference approximations can be found in \citep{nagblog-fd}. First derivatives
computed using finite differences require $\mathcal{O}(N)$ function evaluations, where $N$ is the number of
input variables, and the optimal relative accuracy is around $10^{-6}-10^{-8}$ for double precision variables. Second
derivatives require $\mathcal{O}(N^2)$ function evaluations, and the optimal accuracy is around $10^{-2}-10^{-4}$.
Derivatives that are accurate to machine precision ($10^{-14}-10^{-16}$ for double precision variables) can be
computed using AD tools. AD tools generally use one of two modes: tangent (forward) or adjoint (reverse).
For a function with $N$ inputs and $M$ outputs, tangent mode requires $\mathcal{O}(N)$ function evaluations
and adjoint mode requires $\mathcal{O}(M)$ function evaluations. In statistical applications our output is
often a scalar probability density, so adjoint AD will scale better with $N$ than either tangent mode or
finite differences in terms of total computation time. Adjoint AD is implemented as follows using the Stan
Math Library, \citep{carpenter2015stan}.
\begin{verbatim}
// Define and init matrix with Stan's adjoint type
Matrix<stan::math::var,Dynamic,1> theta = input_values;
// Compute primal and gradient
stan::math::var lp = computation(theta);
lp.grad();
// Extract gradient
Matrix<double,Dynamic,1> grad_vec(theta.size());
for(int j=0; j<theta.size(); j++) grad_vec(j) = theta(j).adj();
\end{verbatim}
In the code above the function is evaluated using the \texttt{stan::math:var} scalar type rather than
\texttt{double}. During function evaluation, derivative information is recorded. When the \texttt{grad()}
function is called this derivative information is interpreted and derivative information is then accessed by
calling \texttt{adj()} on the input variables. Similarly to the Stan Math Library, derivatives with
machine-precision accuracy can be computed using the NAG \dcocpp{} tool developed in collaborations with RWTH
Aachen University's STCE group, \citep{AIB-2016-08}.
\begin{verbatim}
// Define and init matrix with dco's adjoint type
typedef dco::ga1s<double> DCO_MODE;
typedef DCO_MODE::type Scalar;
Matrix<Scalar,Dynamic,1> theta = input_values;
// Enable dynamic activity analysis
for(int j=0; j<theta.size(); j++)
DCO_MODE::global_tape->register_variable(theta(j));
// Compute primal and gradient
Scalar lp = computation(theta);
dco::derivative(lp) = 1.0;
DCO_MODE::global_tape->interpret_adjoint();
// Extract gradient
Matrix<double,Dynamic,1> grad_vec(theta.size());
for(int j=0; j<theta.size(); j++)
grad_vec(j) = dco::derivative(theta(j));
\end{verbatim}
\dcocpp{} provides its adjoint type using the \texttt{dco::ga1s<T>::type} typedef, where \texttt{T}
is the corresponding primal type (e.g.\ \texttt{double}). Higher order adjoints can be achieved by
recursively nesting the adjoint type. Using this type, the program is first executed in the augmented
forward run, where derivative information is stored in the \texttt{global\_tape} data structure.
The \texttt{register\_variable} function initialises recording of the tape and facilitates dynamic
varied analysis~\citep{hascoet2005tbr}. Derivative information recorded during the augmented primal run is
then interpreted using the \texttt{interpret\_adjoint()} function and \texttt{dco::derivative} is used to
access the propagated adjoints.
\section{Results for noise driven harmonic oscillator}
We now present the results of MCMC sampling using smMALA to estimate parameters of the noise driven
harmonic oscillator. Then we demonstrate that, for this particular model, MCMC sampling can be accelerated by
using NUTS. We also compare run times and sampling efficiency when AD is used to evaluate derivatives. Here
we estimate parameters using synthetic data (i.e. data generated from the model). This is a useful check that
the MCMC sampler is working correctly: we should be able to recover the parameters that were used to simulate
the data. We generate a pair of datasets representing different conditions (which we label $c_1$ and $c_2$).
We assume that the $\omega_0$ and $\sigma_{in}$ parameters vary across conditions, but that $\zeta$ is fixed
across conditions, and that $\sigma_{obs}=0.05$ (i.e. it is known and does not need to be sampled). In Table
\ref{fig:ho_mcmc_spec} we show that the $95\%$ credible intervals for each parameter include the actual
parameter values for all 5 parameters. These results came from one MCMC run of $10,000$ iterations using the
Whittle likelihood and a uniform prior. The step size parameter, $h$, was set to $1.0$. There were no
burn-in iterations as the sampler was initialized at the true parameter values. Figure \ref{fig:ho_mcmc_spec}
shows $95\%$ confidence intervals for the spectral density obtained using the Welch method alongside $95\%$
credible intervals for the spectral density estimated from $10,000$ MCMC iterations. This is another useful
check: the spectral density predictions generated by sampled parameter sets are consistent with non-parametric
estimates of the spectral density.
\begin{figure}
\centering
\hspace{-2.2cm}
\begin{minipage}[b]{0.60\linewidth}
\centering
\includegraphics[width=1.05\linewidth]{harmonic_oscillator_spec_dens_est}
\end{minipage
\begin{minipage}[b]{0.30\linewidth}
\centering
\begin{tabular}[b]{|c||c|c|c|c|}
\hline
\multirow{2}{*} & \multirow{2}{*}{Actual} & \multicolumn{3}{c|}{Quantile} \\
\hhline{|~||~|-|-|-|}
& & 0.025 & 0.50 & 0.975 \\
\hhline{|=||=|=|=|=|}
$\omega_0(c_1)$ & 80 & 78.3 & 80.0 & 82.1 \\
\hline
$\omega_0(c_2)$ & 40 & 37.4 & 38.9 & 40.4 \\
\hline
$\sigma_{in}(c_1)$ & 100 & 94.0 & 101 & 109 \\
\hline
$\sigma_{in}(c_2)$ & 10 & 9.53 & 10.6 & 11.7 \\
\hline
$\zeta$ & 0.2 & 0.17 & 0.19 & 0.22 \\
\hline
\end{tabular}
\vspace{1.5cm}
\end{minipage}
\caption{Spectral density and posterior parameter estimates for noise driven harmonic oscillator. Synthetic data
was generated by simulating from the model: duration = 20.0, time-step = 0.01. Two datasets ($c_1$ and
$c_2$) were generated with different parameter values for $\omega_0$ and $\sigma_{in}$ in each dataset
(values are given in `actual' column). The quantiles are estimated from a sequence of $10,000$ MCMC samples
from the joint (posterior) distribution of all the parameters given all of the data.
For example, 2.5\% of the MCMC samples had $\omega_0(c_1)$ values less than 78.3.}
\label{fig:ho_mcmc_spec}
\end{figure}
Table \ref{tab:bench} uses the same model and the same datasets as above to compare smMALA with NUTS, and
finite differences with AD. The AD implementation used in smMALA was dco's tangent over adjoint mode
(i.e. \texttt{dco::gt1s} combined with \texttt{dco::ga1s}). In NUTS we used the dco's tangent mode for
computing the derivatives of the spectral density and Stan's adjoint mode (\texttt{stan::math::var}) for the
rest of the computation. Given the most recent developments in the Stan Math Library (see release notes for
v3.2.0) it would likely be possible to use Stan's adjoint mode for the whole computation. However this was
not the case when we started writing the code. In general users may find that they need the more advanced
functionality of \dcocpp{} for part of their computation in order to obtain good performance.
The MCMC samplers were each run for 1,000 iterations. The results were analyzed using Stan's
\texttt{stansummary} command-line tool. To account for correlation between MCMC samples we use the \textit{N
Eff} diagnostic defined in Section \ref{sec:mcmc} to measure sampling efficiency, min \textit{N Eff} is the
minimum over the 5 model parameters. Table \ref{tab:bench} shows that, for the noise driven harmonic
oscillator, we can accelerate MCMC sampling by a factor of around 3-4 through the combined effect of more
efficient MCMC sampling (NUTS rather than smMALA) and more accurate derivative evaluations (AD rather than
finite differences). In some cases the \textit{N Eff / s} is higher for finite differences than for AD,
because of the small number of input variables. However, even for this simple model the NUTS min \textit{N
Eff} is higher with AD than with central finite differences suggesting that the extra accuracy in the
derivatives results in more efficient sampling per MCMC iteration.
In the context of Bayesian uncertainty quantification what we are interested in is whether the MCMC samples
are an accurate representation of the true posterior distribution. A necessary condition for posterior
accuracy is that the true parameters are (on average) contained in the credible intervals derived from the
MCMC samples. This is the case for all the different variants of MCMC that we tested. The main differences
we found between the different variants was the sampling efficiency. As discussed in Section \ref{sec:mcmc},
a sampling efficiency that is $3-4$ time greater means a reduction in the value of $C$ in the
Monte Carlo error ($C / \sqrt{N}$) by that factor. Another way of interpreting this is that we could
reduce the number of MCMC iterations by a factor of $3-4$ and expect to obtain the same level of accuracy in
the estimated posterior distribution.
\begin{table}[!htb]
\caption{Noise-driven harmonic oscillator benchmarking results on Intel 2.50GHz CPU, Ubuntu (18.04) Linux.}
\begin{tabular}{|l|l|l|l|l|}
\hline
MCMC sampler & Derivative implementation & CPU time (s) & min \textit{N Eff} & min \textit{N Eff} / s \\
\hline
smMALA & Finite differences & 6.2 & 150 & 24 \\
smMALA & \dcocpp & 8.2 & 150 & 18 \\
NUTS & Finite differences & 8.5 & 384 & 46 \\
NUTS & Stan and \dcocpp & 6.0 & 503 & 84 \\
\hline
\end{tabular}
\label{tab:bench}
\end{table}
\section{Discussion}
The Whittle likelihood function is written to be polymorphic over classes derived from the
a base class called \texttt{Stable\_sde}. The function signature includes a reference to the base class.
\begin{verbatim}
template<typename SCALAR_T>
SCALAR_T log_whittle_like(Stable_sde<SCALAR_T> & m, ...);
\end{verbatim}
Classes that are derived from the \texttt{Stable\_sde} base class must define the following pure virtual
function.
\begin{verbatim}
virtual Matrix<SCALAR_T, Dynamic, Dynamic> sde_jacobian(...) = 0;
\end{verbatim}
The spectral density evaluation (Equation \eqref{eq:spec_eq}) is implemented in the base class, reducing the
effort required to do spectral data analysis for other stable SDEs. The smMALA sampler is written
to be polymorphic over classes derived from a base class called \texttt{Computation}. Classes that are
derived from the \texttt{Computation} base class must define the following pure virtual function.
\begin{verbatim}
virtual SCALAR_T eval(Matrix<SCALAR_T,Dynamic,1>& theta,...) = 0;
\end{verbatim}
This function should evaluate the posterior density in the user's model. The gradient and Hessian of the
posterior density are implemented in the \texttt{Compu\-tation} class. This reduces the effort required to
use NUTS or smMALA for other computational models. Classes derived from \texttt{Stable\_sde} or from
\texttt{Computation} can be instantiated with several different scalar types including \texttt{double},
\texttt{stan\allowbreak::math\allowbreak::var}, and a number of \dcocpp{} types. This enables automatic
evaluation of first and second derivatives. The gradient and Hessian functions need a template specialization
to be defined in the base class for each different scalar type that is instantiated.
We conclude with some comments regarding the choice of MCMC sampler and the method for evaluating derivatives.
Our software only implements derivative-based MCMC as this tends to be more computationally efficient than
other MCMC methods \citep{hoffman2014no,penny2016annealed}. Derivative-based MCMC samplers can be further
subdivided into methods that use higher-order derivatives, such as smMALA (sometimes referred to as
Riemannian) and methods that only require first-order derivatives (such as NUTS). Riemannian methods tend to
be more robust to complexity in the posterior distribution, but first-order methods tend to be more
computationally efficient for problems where the posterior geometry is relatively simple. We recommend
using smMALA in the first instance, then NUTS as a method that may provide acceleration in MCMC sampling, in
terms of effective samples per second, \textit{N Eff /s}. Regarding the derivative method, finite difference
methods often result in adequate performance for problems with moderate input dimension (e.g. 10-20), at least
with smMALA. But for higher-dimensional problems (e.g. Partial Differential Equations or Monte Carlo
simulation) we recommend accelerating derivative computation using adjoint AD
\citep{nagsite-ad,naumann2018adjoint}. The software presented enables users to implement and benchmark all
these alternatives so that the most appropriate methods for a given problem can be chosen.
\bibliographystyle{apalike}
| 9856dcb55082ef484087585c6651ca34b0a53da0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Firm heterogeneity is prevalent in international trade. On the one hand, it
is evident that firms are heterogeneous in terms of firm size, with a few
dominant large exporters and a host of negligible small firms.\footnote
See Parenti (2018).} On the other hand, small firms also display
heterogeneity in the form of productivity differences.\footnote
See Melitz (2003) and Melitz and Ottaviano (2008).} This note attempts to
address these two types of firm heterogeneity in a coherent model and
examine the impact of trade liberalization.
We draw on the model in Parenti (2018) where large and small firms coexist,
but goes beyond that model by introducing cost heterogeneity among small
firms. In our model, large firms are treated as established incumbent
oligopolists, whereas small firms are treated as monopolistic competitors
who endogenously make entry decisions and face initial uncertainty about
their future productivity prior to entry. The heterogeneity among small
firms is in the spirit of Melitz-Ottaviano type's heterogeneity.\footnote
Representative works include Melitz (2003) and Melitz and Ottaviano (2008).}
First, we propose a novel necessary and sufficient condition for the
existence of such a mixed market structure in a closed economy,
accommodating the cost heterogeneity among small firms. Large and small
firms coexist if (i) the marginal cost of large firms is strictly lower than
that of the least efficient small firm in operation, and (ii) consumers'
preferences on the differentiated goods are sufficiently high, the number of
large firms is not too high, and the cost of large firms is not too low. The
former condition guarantees that large firms earn positive profits, and the
latter condition guarantees a positive mass of small firms in operation.
Second, in contrast to Parenti (2018), we show that bilateral trade
liberalization may raise the mass of small firms selling in each country.
Assuming that small firms share the same cost and do not export, Parenti
(2018) finds that trade liberalization leads to the expansion of large firms
and decreases the mass of small firms selling in each country. However, we
introduce cost heterogeneity among small firms and allow the most efficient
ones to export. Therefore, a reduction in trade cost as a result of
bilateral trade liberalization not only reduces the export cost of large
firms, but also lowers the trade cost of small firms. The former effect,
which is consistent with Parenti (2018), generates a negative impact on the
mass of small firms, whereas the latter effect generates a positive impact
on the mass of small firms. The impact of trade liberalization on the mass
of small firms selling in each country depends on these two opposing
effects. When the number and cost of large firms are sufficiently high, the
former effect dominates, and trade liberalization decreases the mass of
small firms. Otherwise, the latter effect dominates, and trade
liberalization increases the mass of small firms. We also find that trade
liberalization may raise the mass of small producers in each country.
This paper contributes to the recent studies on the mixed market structure
with large and small firms. Assuming that small firms have identical cost,
Shimomura and Thisse (2012) and Pan and Hanazono (2018) identify the
condition for the coexistence of large and small firms and examine the
impact of a large firm's entry in a closed economy. Complementary to these
works, this paper derives a new coexistence condition incorporating cost
heterogeneity among small firms and investigates the impact of trade
liberalization. Although our results can be regarded as an extension of
Parenti (2018), they open a new channel through which the most efficient
small firms could export and benefit from trade liberalization.
Our results also relate to the research on trade liberalization. In the
model with monopolistic competition and firm heterogeneity, Melitz (2003)
and Melitz and Ottaviano (2008) show that bilateral trade liberalization
raises the mass of small firms selling in each country. On the contrary, in
our model, bilateral trade liberalization may reduce the mass of small firms
due to the expansion in the export of large firms. In the Stackelberg model
where large firms are leaders and can export whereas small firms are
followers and sell only in the domestic market, Pan and Tabuchi (2019) also
find that free trade reduces the number of small firms. In contrast,
allowing for the export by efficient small firms, we find that trade
liberalization may raise the mass of small firms because the exporting small
firms also benefit from trade liberalization.
\section{Closed Economy}
Consider an economy with $L$ consumers, each supplying one unit of labor
inelastically. Each consumer consumes two types of goods, a homogeneous good
and differentiated goods. The homogeneous good is produced under constant
returns to scale at unit cost and perfect competition, which implies a unit
wage. The differentiated goods are supplied by a discrete number $N$ of
large firms $n=1,...,N$, and a continuum of small firms indexed by $i\in
\lbrack 0,M]$. Occupying a substantial market share, each large firm behaves
strategically as an oligopolist. Being negligible in the market, each small
firm behaves non-strategically as a monopolistic competitor.
\subsection{Preferences}
All consumers share the same utility function given b
\begin{equation}
U=\alpha (\int_{0}^{M}x_{i}di+\sum_{n=1}^{N}X_{n})-\frac{\beta }{2
[\int_{0}^{M}(x_{i})^{2}di+\sum_{n=1}^{N}(X_{n})^{2}]-\frac{\gamma }{2
(\int_{0}^{M}x_{i}di+\sum_{n=1}^{N}X_{n})^{2}+x_{0},
\end{equation
where $x_{i}$ represents the individual consumption of the good produced by
small firm $i$, and $X_{n}$ represents the individual consumption of the
good produced by large firm $n$. The numeraire good is represented by $x_{0}
. The demand parameters $\alpha >0$ and $\gamma >0$ capture the degree of
substitutability between the varieties of the differentiated product and the
numeraire, and $\beta >0$ represents the degree of differentiation between
the differentiated varieties.
Assuming that consumers have positive demands for the numeraire good, the
inverse demand for small firm $i$ and large firm $n$ are respectively
\begin{equation}
p_{i}=\alpha -\beta x_{i}-\gamma \mathbf{X,}\text{ }\forall i\in \lbrack 0,M]
\label{inverse demand small}
\end{equation
\begin{equation}
P_{n}=\alpha -\beta X_{n}-\gamma \mathbf{X,}\text{ }n=1,...,N
\label{inverse demand big}
\end{equation
whenever $x_{i}>0$, $X_{n}>0$, and $\mathbf{X}$\textbf{$=$}
\int_{0}^{M}x_{i}di+\sum_{n=1}^{N}X_{n}$ is the individual total consumption
of the differentiated good$.$
Inverse demands (\ref{inverse demand small}) and (\ref{inverse demand big})
can be inverted to the demands for small firm $i$ and large firm $n$
\begin{equation}
q_{i}=Lx_{i}=\frac{\alpha L}{\beta +\gamma (M+N)}-\frac{L}{\beta }p_{i}
\frac{L}{\beta }\frac{\gamma }{\beta +\gamma (M+N)}\mathbf{P,}\text{
\forall i\in \lbrack 0,M] \label{demand small}
\end{equation
\begin{equation}
Q_{n}=LX_{n}=\frac{\alpha L}{\beta +\gamma (M+N)}-\frac{L}{\beta }P_{n}
\frac{L}{\beta }\frac{\gamma }{\beta +\gamma (M+N)}\mathbf{P,}\text{
n=1,...,N \label{demand big}
\end{equation
where $\mathbf{P}=\int_{0}^{M}p_{i}di+\sum_{n=1}^{N}P_{n}$ is the aggregate
price. To ensure a positive consumption, i.e., $q_{i}\geq 0$ and $Q_{n}\geq
0 $, the prices of large and small firms should satisfy
\begin{equation}
p_{i},P_{n}\leq \frac{1}{\beta +\gamma (M+N)}(\alpha \beta +\gamma \mathbf{P
)=p_{\max }, \label{pmax}
\end{equation
where $p_{\max }$ represents the price at which demand for the good is
driven to zero.
\subsection{Firms}
We consider a 2-stage game. In the first stage, small firms enter the
market. In the second stage, large and small firms who have entered the
market compete in price.
\paragraph{Small Firms}
Entry in the differentiated product sector is costly as each firm incurs
product development and production start-up costs. Subsequent production
exhibits constant returns to scale at a marginal cost $c$. Following Melitz
and Ottaviano (2008), Research and development yield uncertain outcomes for
c,$ and firms learn about this cost level only after making the irreversible
investment $f_{E}$ required for entry. We assume that the marginal cost of
each small firm is a draw from a common and known distribution $G(c)$ with
support $[0,c_{M}].$ Since the entry cost is sunk, firms that can cover
their marginal cost survive and produce. All other firms exit the industry.
Substituting the demand function (\ref{demand small}), the gross profit of
small firm $i$ with marginal cost $c_{i}$ can be expressed a
\begin{equation}
\pi _{i}=(p_{i}-c_{i})(\frac{\alpha L}{\beta +\gamma (M+N)}-\frac{L}{\beta
p_{i}+\frac{L}{\beta }\frac{\gamma }{\beta +\gamma (M+N)}\mathbf{P}),\text{
\forall i\in \lbrack 0,M]. \label{pis}
\end{equation}
\paragraph{Large Firms}
Large firms are established incumbents and produce at marginal cost $C$.
Substituting the demand function (\ref{demand big}), the profit of large
firm $n$ can be expressed a
\begin{equation}
\Pi _{n}=(P_{n}-C)(\frac{\alpha L}{\beta +\gamma (M+N)}-\frac{L}{\beta
P_{n}+\frac{L}{\beta }\frac{\gamma }{\beta +\gamma (M+N)}\mathbf{P}),\text{
n=1,...,N. \label{piL}
\end{equation}
\subsection{The Equilibrium Analysis}
\subsubsection{\textbf{Stage 2}}
\paragraph{Small Firms}
Small firm $i$ maximizes its profit with respect to its price $p_{i}$,
treating the aggregate price $\mathbf{P}$ as given since its market share is
negligible. The profit maximizing price $p_{i}$ and quantity $q_{i}$ then
satisf
\begin{equation}
q_{i}(c_{i})=\frac{L}{\beta }[p_{i}(c_{i})-c_{i}], \label{qs(ps)}
\end{equation}
If the profit maximizing price $p_{i}$ is above the price bound $p_{\max }$
from (\ref{pmax}), then small firm $i$ exits. Let $c_{D}$ denote the cost of
the small firm who is indifferent about remaining in the industry. This firm
earns zero profit as its cost is equal to its marginal cost, that is,
p(c_{D})=c_{D}=p_{\max }$, and its demand level $q(c_{D})$ is driven to $0$.
We assume that $c_{M}$ is high enough to be above $c_{D}$, so that some
firms with cost draws between these two levels exit. All firms with cost
below $c_{D}$ earn positive profits (gross of the entry cost) and remain in
the industry. Substituting (\ref{pmax}), and knowing that $p_{\max }=c_{D}$,
the first order condition of a small firm with $c<c_{D}$ yields its optimal
price $p(c)$ and quantity $q(c)$
\begin{equation}
p(c)=\frac{1}{2}(c_{D}+c), \label{ps(c)}
\end{equation
\begin{equation}
q(c)=\frac{L}{2\beta }(c_{D}-c), \label{qs(c)}
\end{equation
and the profit can be expressed b
\begin{equation}
\pi (c)=\frac{L}{4\beta }(c_{D}-c)^{2}. \label{pis(c)}
\end{equation}
\paragraph{Large Firms}
Different from a small firm, who treats the aggregate price $\mathbf{P}$
parametrically, a large firm behaves strategically and internalizes its
impact on $\mathbf{P}$. Large firm $n$ maximizes its profit given by (\re
{piL}), yielding the optimal price
\begin{equation}
P_{n}=P=\frac{c_{D}+(1-\Theta )C}{2-\Theta },\text{ }n=1,...,N, \label{PL}
\end{equation
where $\Theta =\gamma /[\beta +\gamma (M+N)]\in (0,1)$ represents the
internalization by the large firm$.$ Due to the internalization, a large
firm's price depends on the threshold cost $c_{D}$ of small firms and the
mass $M$ of small firms, which are endogenously determined by the free entry
condition in the first stage$.$ Moreover, incurring the same marginal cost
C $, large firms set the same price $P$.
Similar to a small firm, a large firm stops operation if and only if the
profit maximizing price $P$ is above the price bound $p_{\max }$. Since
p_{\max }=c_{D}$, the necessary and sufficient condition for the existence
of large firms is $C<c_{D}$.\footnote
We assume that large firms incur zero fixed cost. If the fixed cost is
positive, then the condition for the existence of large firms should be
stricter.}
\subsubsection{\textbf{Stage 1}}
Prior to entry, the expected profit of a small firm\ is $\int_{0}^{c_{D}}\pi
(c)dG(c)-f_{E}.$ Substituting (\ref{pis(c)}), the free entry condition can
be rewritten a
\begin{equation}
\frac{L}{4\beta }\int_{0}^{c_{D}}(c_{D}-c)^{2}dG(c)=f_{E},
\label{free entry}
\end{equation
which determines the cost cut-off $c_{D}$.
To obtain tractable results, we assume the Pareto parametrization for the
cost draws of small firms, i.e.,
\begin{equation}
G(c)\,=(\frac{c}{c_{M}})^{k},\text{ }c\in \lbrack 0,c_{M}].
\label{pareto cost}
\end{equation}
Given this parametrization, the cut-off cost level $c_{D}$ determined by
\ref{free entry}) is then
\begin{equation}
c_{D}^{\ast }=(\frac{\beta \phi }{L})^{\frac{1}{k+2}}, \label{cD}
\end{equation
where $\phi =2(k+1)(k+2)(c_{M})^{k}f_{E}$ is the technology index.
Substituting the optimal prices of small and large firms given by (\re
{ps(c)}) and (\ref{PL}), and using that $p_{\max }=c_{D}$, (\ref{pmax}) can
be expressed a
\begin{equation}
\frac{(\alpha -c_{D}^{\ast })\beta }{\gamma }=\frac{Mc_{D}^{\ast }}{2(k+1)
+N(c_{D}^{\ast }-C)\frac{1-\Theta (M)}{2-\Theta (M)}, \label{M*}
\end{equation
In equation (\ref{M*}), the LHS is constant and positive, whereas the RHS
increases with $M$. Let $M^{\ast }$ be the equilibrium mass of small firms.
To ensure that $M^{\ast }>0$, the RHS should be smaller than the LHS when
M=0$. Therefore, equation (\ref{M*}) uniquely determines a positive mass
M^{\ast }$ of small firms in operation if and only if
\[
\frac{(\alpha -c_{D}^{\ast })\beta }{\gamma }>N(c_{D}^{\ast }-C)\frac{\beta
+\gamma (N-1)}{2\beta +\gamma (2N-1)}.
\]
As shown earlier, large firms earn a positive profit if and only if $C<c_{D}
. Substituting the equilibrium cost cut-off of small firms in (\ref{cD}), we
establish the condition for the coexistence of large and small firms in
Proposition 1.
\begin{proposition}
There exists a unique mixed market equilibrium where large and small firms
coexist if and only if (i) $C<(\beta \phi /L)^{\frac{1}{k+2}}$ and (ii)
[\alpha -(\beta \phi /L)^{\frac{1}{k+2}}]\beta >\gamma N[(\beta \phi /L)^
\frac{1}{k+2}}-C][\beta +\gamma (N-1)]/[2\beta +\gamma (2N-1)]$.
\end{proposition}
The first condition, which guarantees that large firms earn positive
profits, requires that the cost of large firms be smaller than the cost
cutoff of small firms, which is satisfied if the technology index $\phi $ is
sufficiently high and the market size $L$ is not too large. The second
condition, which guarantees a positive mass of small firms, is satisfied if
the consumer's preference for the differentiated good $\alpha $ is
sufficiently high, the cost $C$ of large firms is not too low in comparison
with the cost cutoff $(\beta \phi /L)^{\frac{1}{k+2}}$ of small firms, and
the number $N$ of large firms is not too large.
In models with the coexistence of large and small firms, see, e.g.,
Shimomura and Thisse (2012), Parenti (2018), and Pan and Hanazono (2018), it
is commonly assumed that small firms share the same technology. When large
firms are single-product and facing zero fixed cost, their assumption
implies large firms operate if and only if $C<c+2\sqrt{\beta f_{E}}$.
However, when we allow the small firms to be heterogeneous and facing cost
uncertainty prior to entry, Proposition 1 indicates that large firms should
be more efficient than the least productive small firm to survive.
\section{Open Economy}
In the previous section, we considered a closed-economy model. In this
section, we extend the model to an open economy, and examine the impact of
trade liberalization. Following Parenti (2018), we consider two symmetric
countries, $H$ and $F$, each with the same market size and distribution of
firms. Specifically, each country has $L$ consumers who share the same
preferences, leading to the inverse demand functions (\ref{inverse demand
small}) and (\ref{inverse demand big}). In addition, in each country, there
are $N$ large firms with marginal cost $C$, and small firms share the same
cost distribution given by (\ref{pareto cost}). The two markets are
segmented, although firms can produce in one market and sell in the other at
a trade cost $\tau $. We assume that large and small firms exist and export
in both countries, and we will identify the condition in the analysis.
The aggregate price in country $h$ is then defined b
\[
\mathbf{P}^{h}=\int_{0}^{M^{h}}p_{i}di+NP_{D}^{h}+NP_{X}^{f},\text{ }h,f=H,F
\text{ }h\neq f,
\
where $M^{h}$ is the mass of small firms selling in country $h,$ including
both domestic small firms and foreign small exporters, $P_{D}^{h}$
represents the price of domestic large firms, and $P_{X}^{f}$ represents the
delivered price of foreign large firms.
\bigskip The price threshold for positive demand in country $h$ is then
given by
\begin{equation}
p_{\max }^{h}=\frac{1}{\beta +\gamma (M^{h}+2N)}(\alpha \beta +\gamma
\mathbf{P}^{h}).\text{\ } \label{pmax l}
\end{equation}
\subsection{The Equilibrium Analysis}
\subsubsection{Stage 2}
\paragraph{Small Firms}
Let $p_{D}^{h}(c)$ and $q_{D}^{h}(c)$ represent the domestic levels of
profit maximizing price and quantity sold for a small firm producing in
country $h$ with cost $c$. Such a small firm may also export output
q_{X}^{h}(c)$ at a delivered price $p_{X}^{h}(c).$ The delivered cost of a
unit with cost $c$ to the foreign country is $\tau c,$ where $\tau >1$ and
is the same across all exporting firms. Since the markets are segmented,
small firms independently maximize the profits in the domestic and foreign
markets. Let $\pi _{D}^{h}(c)=[p_{D}^{h}(c)-c]q_{D}^{h}(c)$ and $\pi
_{X}^{h}(c)=[p_{X}^{h}(c)-\tau c]q_{X}^{h}(c)$ denote the maximized value of
these profits as a function of the firm's marginal cost $c$.
Analogously to (\ref{qs(ps)}), the profit maximizing prices and quantities
must satisfy: $q_{D}^{h}(c)=(L/\beta )[p_{D}^{h}(c)-c]$ and
q_{X}^{h}(c)=(L/\beta )[p_{X}^{h}(c)-\tau c].$ As was the case in the closed
economy, only the small firms that earn non-negative profits in a market
(domestic or export) will choose to sell in that market. This leads to
similar cost cut-off rules for small firms selling in either market. Let
c_{D}^{h}$ denote the upper bound cost for small firms selling in their
domestic market, and let $c_{X}^{h}$ denote the upper bound cost for small
exporters from $h$ to $f$. These cutoffs must then satisfy
\begin{eqnarray}
c_{D}^{h} &=&\sup \{c:\pi _{D}^{h}(c)>0\}=p_{\max }^{h}, \label{cutoff pmax}
\\
c_{X}^{h} &=&\sup \{c:\pi _{X}^{h}(c)>0\}=\frac{p_{\max }^{f}}{\tau }.
\nonumber
\end{eqnarray}
This implies $c_{X}^{f}=c_{D}^{h}/\tau $: trade barriers make it harder for
small firms to break even in the foreign market than in the domestic market.
As was the case in the closed economy, the optimal prices and quantities are
also functions of the cost cutoffs
\begin{eqnarray}
p_{D}^{h}(c) &=&\frac{1}{2}(c_{D}^{h}+c),\text{ }q_{D}^{h}(c)=\frac{L}
2\beta }(c_{D}^{h}-c), \label{p(c) open} \\
p_{X}^{h}(c) &=&\frac{\tau }{2}(c_{X}^{h}+c),\text{ }q_{X}^{h}(c)=\frac{L}
2\beta }\tau (c_{X}^{h}-c), \nonumber
\end{eqnarray
which yield the maximized profit levels
\begin{eqnarray}
\pi _{D}^{h}(c) &=&\frac{L}{4\beta }(c_{D}^{h}-c)^{2}, \label{pi(c) open} \\
\pi _{X}^{h}(c) &=&\frac{L}{4\beta }\tau ^{2}(c_{X}^{h}-c)^{2}. \nonumber
\end{eqnarray}
\paragraph{Large Firms}
Similar to the closed economy, we assume that all the large firms share the
same marginal cost $C$ in production, and the delivered cost of a unit to
the foreign country is $\tau C$. Let $P_{D}^{h}$ and $Q_{D}^{h}$ represent
the domestic levels of profit maximizing price and quantity sold for a large
firm producing in country $h$. The large firm also exports output $Q_{X}^{h}$
at a delivered price $P_{X}^{h}.$ Analogously to (\ref{PL}), the profit
maximizing domestic and export prices ar
\begin{eqnarray}
P_{D}^{h} &=&\frac{c_{D}^{h}+(1-\Theta ^{h})C}{2-\Theta ^{h}},
\label{P open} \\
P_{X}^{h} &=&\tau \frac{c_{X}^{h}+(1-\Theta ^{f})C}{2-\Theta ^{f}},
\nonumber
\end{eqnarray
where $\Theta ^{h}=\gamma /[\beta +\gamma (M^{h}+2N)]$ and $\Theta
^{f}=\gamma /[\beta +\gamma (M^{f}+2N)]$. A large firm in country $h$ stops
selling in the domestic market if and only if $P_{D}^{h}$ is above the price
bound $p_{\max }^{h}$. Since $p_{\max }^{h}=c_{D}^{h}$, the necessary and
sufficient condition for the domestic operation of large firms is
C<c_{D}^{h}$. Furthermore, it stops exporting if and only if $P_{X}^{h}$ is
above the price bound $p_{\max }^{f}$ (where $f\neq h$). Since $p_{\max
}^{f}=\tau c_{X}^{h}$, the necessary and sufficient condition for the
existence of exporting large firms is $C<c_{X}^{h}.$ These two conditions
indicate that large firms share the same cost cutoffs with small firms in
both domestic and foreign markets.
Therefore, large firms in both countries export if and only if $C<\min
\{c_{X}^{H},c_{X}^{F}\}.$
\subsubsection{Stage 1}
Entry of small firms is unrestricted in both countries. Small firms choose a
production location prior to entry and pay the sunk entry cost. We assume
that both countries share the same technology for small firms -- referenced
by the entry cost $f_{E}$ and cost distribution $G(c)$. Therefore, the free
entry condition\ is expressed a
\[
\frac{L}{4\beta }\int_{0}^{c_{D}^{h}}\pi
_{D}^{h}(c)dG(c)+\int_{0}^{c_{X}^{h}}\pi _{X}^{h}(c)dG(c)=f_{E}.
\]
\bigskip With Pareto parametrization for the cost draws $G(c)$ in both
countries, the free entry condition can be rewritten as
\[
(c_{D}^{h})^{k+2}+\tau ^{2}(c_{X}^{h})^{k+2}=\frac{\beta \phi }{L}.
\]
Since $c_{X}^{f}=c_{D}^{h}/\tau $, the free entry condition (\ref{free entry
open}) can be expressed as
\begin{equation}
(c_{D}^{h})^{k+2}+\rho (c_{D}^{f})^{k+2}=\frac{\beta \phi }{L},
\label{free entry open}
\end{equation
where $\rho =(\tau )^{-k}\in (0,1)$ is an inverse measure of trade costs
(the \textquotedblleft freeness\textquotedblright\ of trade). This system
can be solved for the cutoffs in both countries
\begin{equation}
c_{D}^{h\ast }=c_{D}^{f\ast }=c_{D}^{\ast }=[\frac{\beta \phi }{L(1+\rho )
]^{1/(k+2)}, \label{cD open}
\end{equation
by which the average cost of small firms selling in each country is $\bar{c
_{D}^{\ast }=kc_{D}^{\ast }/(k+1).$
By (\ref{pmax l}) and (\ref{P open}), the aggregate price in country $h$ is
given by
\[
\mathbf{P}^{h}=\frac{M^{h}(2k+1)}{2(k+1)}c_{D}^{\ast }+N\frac{c_{D}^{\ast
}+(1-\Theta ^{h})C}{2-\Theta ^{h}}+N\frac{c_{D}^{\ast }+(1-\Theta ^{f})\tau
}{2-\Theta ^{f}},
\
Substituting $\mathbf{P}^{h}$, (\ref{pmax l}) can be expressed a
\begin{equation}
\frac{(\alpha -c_{D}^{\ast })\beta }{\gamma }=\frac{M^{h}}{2}\frac
c_{D}^{\ast }}{k+1}+N(c_{D}^{\ast }-C)\frac{1-\Theta ^{h}}{2-\Theta ^{h}
+N(c_{D}^{\ast }-\tau C)\frac{1-\Theta ^{h}}{2-\Theta ^{h}}.
\label{Ml(pmax) open pareto}
\end{equation
which uniquely determines the mass of small firms in country $h$ because the
RHS strictly increases with $M^{h}$. (\ref{Ml(pmax) open pareto}) also
implies that $M^{f}=M^{h}=M$ and hence $\Theta ^{f}=\Theta ^{h}=\Theta .$
Therefore, (\ref{Ml(pmax) open pareto}) could be reduced t
\begin{equation}
\frac{(\alpha -c_{D}^{\ast })\beta }{\gamma }=\frac{M}{2}\frac{c_{D}^{\ast
}{k+1}+\frac{1-\Theta (M)}{2-\Theta (M)}N[(c_{D}^{\ast }-C)+(c_{D}^{\ast
}-\tau C)], \label{Ml(pmax) open pareto symmetric}
\end{equation
which uniquely determines the equilibrium mass $M^{\ast }$ of small firms
selling in each country. To ensure that $M^{\ast }>0$, we assume that
(\alpha -c_{D}^{\ast })\beta /\gamma >N[(c_{D}^{\ast }-C)+(c_{D}^{\ast
}-\tau C)][\beta +\gamma (N-1)]/[2\beta +\gamma (2N-1)].$
\subsection{\protect\bigskip Trade Liberalization}
Based on the above equilibrium analysis, now we examine how bilateral trade
liberalization impacts the mass of small firms.
First, (\ref{cD open}) indicates that a reduction in trade cost as a result
of bilateral trade liberalization decreases the cost cut-off of small firms,
that is, $dc_{D}^{\ast }/d\tau >0$.
Now we examine the impact of trade liberalization on the mass of small firms
selling in each country, including both domestic small producers and foreign
small exporters. By (\ref{Ml(pmax) open pareto symmetric}), we hav
\[
\frac{dM}{d\tau }=-\frac{[\frac{\alpha \beta }{\gamma }+(1+\tau )NC]\frac{k}
k+2}\frac{\rho }{\rho +1}\frac{1}{\tau }-\frac{1-\Theta }{2-\Theta }NC}
\frac{c_{D}^{\ast }}{2(k+1)}+(\frac{\Theta }{2-\Theta })^{2}[(c_{D}^{\ast
}-C)+(c_{D}^{\ast }-\tau C)]N},
\
which is negative if and only i
\begin{equation}
\frac{2-\Theta (M^{\ast })}{1-\Theta (M^{\ast })}\frac{\alpha \beta }{\gamma
}>(\frac{2\rho +k+2}{k\rho }\tau -1)NC, \label{M condition}
\end{equation
where $M^{\ast }$ is the equilibrium mass of small firms determined by (\re
{Ml(pmax) open pareto symmetric}).
Proposition 2 establishes the impact of bilateral trade liberalization on
the mass of small firms selling in each country.
\begin{proposition}
A decrease in trade costs as a result of bilateral trade liberalization
increases the mass of small firms selling in each country if $(\alpha \beta
/\gamma )[2-\Theta (M^{\ast })]/[1-\Theta (M^{\ast })]>[(2\rho +k+2)\tau
/(k\rho )-1]NC$, and decreases the mass of small firms selling in each
country otherwise.
\end{proposition}
A reduction in trade cost generates \textit{two opposing effects} on the
mass of small firms selling in each country. First, trade liberalization
reduces the export cost and hence enables more small firms to export, which
potentially increases the mass of small firms. Second, a decrease in trade
cost also generates cost savings on the exported goods of large firms, which
induces large firms to expand export. This generates competitive pressure on
small firms and may reduce the mass of surviving small firms. As marginal
cost $C$ rises, the cost savings become stronger and lead to more aggressive
exporting behavior by large firms. Furthermore, the aggregation of such
aggressive exports increases with the number $N$ of large firms. Therefore,
condition (\ref{M condition}) holds as long as the number $N$ or marginal
cost $C$ of large firms is not too high.
Last, we show that bilateral trade liberalization may also raise the mass of
small producers in each country. By symmetry, in each country, there are
M_{E}G(c_{D})$ domestic small producers and $M_{E}G(c_{X})$ foreign small
exporters. Thus, $M_{E}G(c_{D})+M_{E}G(c_{D}/\tau )=M$, by which the mass of
small entrants in each country can be expressed as
M_{E}=M(c_{M}/c_{D})^{k}/(1+\rho )$. The mass of small producers in each
country is then expressed by
\[
M_{D}=M_{E}G(c_{D})=\frac{M}{1+\rho }.
\
The impact of a reduction in trade cost on the mass of small producers is
then given b
\[
\frac{dM_{D}}{d\tau }=\frac{1}{1+\rho }[\frac{dM}{d\tau }+\frac{k\rho M}
(1+\rho )\tau }].
\]
A reduction in trade cost impacts the mass $M_{D}$ of domestic small
producers through (a) the change in the mass $M$ of small firms selling in
each country and (b) the reallocation between domestic producers and foreign
exporters within the group of small firms. The change in the mass of small
firms selling in each country, according to Proposition 2, is positive as
long as thenumber and marginal cost of large firms are not too high.
Furthermore, trade liberalization invites the most productive small firms to
increase export, which makes competition tougher and may force some of the
least productive small producers to exit. This generates a negative impact
on the mass of domestic small producers. Therefore, if the rise in the total
mass of small firms selling in each country outweighs the negative impact
from the export expansion by the most productive small firms, then trade
liberalization would also raise the mass of small producers in each country.
Otherwise, trade liberalization reduces the mass of small producers.
Proposition 3 establishes this result with the precise condition.
\begin{proposition}
A decrease in trade costs as a result of bilateral trade liberalization
increases the mass of small producers if
\[
\frac{\alpha \beta /\gamma +(1+\tau )NC}{k+2}>\frac{M^{\ast }c_{D}^{\ast }}
2(k+1)}+M^{\ast }N[2c_{D}^{\ast }-(1+\tau )C][\frac{\Theta (M^{\ast })}
2-\Theta (M^{\ast })}]^{2}+NC\frac{(1+\rho )\tau }{\rho k}\frac{1-\Theta
(M^{\ast })}{2-\Theta (M^{\ast })},
\
and decreases the mass of small producers otherwise.
\end{proposition}
Here $c_{D}^{\ast }$ and $M^{\ast }$ are the equilibrium values determined
in (\ref{cD open}) and (\ref{Ml(pmax) open pareto symmetric}). Owing to the
expansion of small exporters as a result of trade liberalization, the range
of parameters where trade liberalization increases the mass of small
producers narrows down.\footnote
We identify both the case when $dM_{D}^{\ast }/d\tau >0$ and that when
dM_{D}^{\ast }/d\tau <0$ by numerical analysis. For instance, when $\alpha
=0.6$, $\beta =\gamma =1,$ $N=1$, $C=0.01$, $L=100$, $f_{E}=1$, and $\tau
=1.5$, $dM_{D}^{\ast }/d\tau <0$; when $\alpha =2.4$, $\beta =\gamma =1,$
N=2$, $C=0.02$, $L=100$, $f_{E}=1$, and $\tau =1.5$, $dM_{D}^{\ast }/d\tau
>0 $.}
We conclude by highlighting the distinction between our results and those in
Parenti (2018). Take Proposition 2 for instance. Assuming that small firms
share the same cost and do not export, Parenti (2018) finds that trade
liberalization decreases the mass of small firms. In Proposition 2, we go a
step further by introducing cost heterogeneity among small firms and
allowing a portion of small firms to export, identifying two opposing
effects of trade liberalization on the mass of small firms. The second
effect, which reduces the mass of small firms, is consistent with Parenti
(2018), whereas the first effect is novel here. The insights of Proposition
2 extend to Proposition 3.
| 9c1e537a17ae375c4ffe48060062254a062cbe1a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\noindent In 1997, Van Hamme \cite{Hamme} observed that $13$ supercongruences on truncated forms of Ramanujan's and Ramanujan-like formulas for $1/\pi$.
In particular, the following supercongruence of Van Hamme \cite[(B.2)]{Hamme},
\begin{equation*}
\sum_{k=0}^{(p-1)/2}\frac{4k+1}{(-64)^k}{2k\choose k}^3\equiv p(-1)^{\frac{p-1}{2}}\pmod {p^3}
\end{equation*}
was first proved by Mortenson \cite{Mortenson} using a $_6F_5$ transformation and a technical evaluation of a quotient of Gamma functions,
where $p$ is an odd prime. Recently, $q$-analogues of congruences and supercongruences have caught the interests of many authors
(see, for example, \cite{Guillera,Guo-m3,Guo-a2,Guo-jmaa,Guo-Zudlin,Guo-result,GS3,GW,Liu-Petrov,Liu-Huang,NP,
Tauraso,WY,Zudilin}).
In \cite[Conjecture 4.3]{Guo-ITSF}, Guo conjectured: for any prime $p>3$ and positive integer $r$,
\begin{align}
\sum_{k=0}^{(p^r-1)/2}\frac{(4k+1)^3}{256^k}{2k\choose k}^4 &\equiv -p^r \pmod {p^{r+3}}, \label{eq1.11}\\
\sum_{k=0}^{p^r-1}\frac{(4k+1)^3}{256^k}{2k\choose k}^4 &\equiv -p^r \pmod {p^{r+3}}. \label{eq1.12}
\end{align}
Later, Guo \cite[Theorem 1.1]{Guo-result} proved \eqref{eq1.11} and \eqref{eq1.12} by establishing the following complete $q$-analogues of them:
for odd integer $n>1$, modulo $[n]_{q^2}\Phi_n(q^2)^3$,
\begin{align*}
\sum_{k=0}^{(n-1)/2}[4k+1]_{q^2}[4k+1]^2\frac{(q^2;q^4)_k^4}{(q^4;q^4)_k^4}q^{-4k}
&\equiv -[n]_{q^2}\frac{2q^{2-n}}{1+q^2}-[n]_{q^2}^3\frac{(n^2-1)(1-q^2)^2q^{2-n}}{12(1+q^2)},
\\
\sum_{k=0}^{n-1}[4k+1]_{q^2}[4k+1]^2\frac{(q^2;q^4)_k^4}{(q^4;q^4)_k^4}q^{-4k}
&\equiv
-[n]_{q^2}\frac{2q^{2-n}}{1+q^2}-[n]_{q^2}^3\frac{(n^2-1)(1-q^2)^2q^{2-n}}{12(1+q^2)}.
\end{align*}
Here and throughout the paper, $(a;q)=(1-a)(1-aq)\cdots(1-aq^{n-1})$ denotes the $q$-\emph{shifted factorial},
$[n]=[n]_q=1+q+\cdots+q^{n-1}$ stands for the \emph{$q$-integer}, and $\Phi_n(q)$ is
the $n$-th \emph{cyclotomic polynomial} in $q$, i.e.,
$$\Phi_n(q)=\prod_{\substack{1\leq k\leq n\\\gcd(n,k)=1}}^n(q-\zeta^k)$$
with $\zeta$ being an $n$-th primitive root of unity.
Guo \cite[Theorem 1.2]{Guo-result} also proved that, for any odd integer $n>1$,
\begin{align}
\sum_{k=0}^{(n+1)/2}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k} &\equiv 0\pmod{[n]_{q^2}\Phi_n(q^2)^3}, \label{eq1.17} \\
\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k} &\equiv 0\pmod{[n]_{q^2}\Phi_n(q^2)^3}, \label{eq1.18}
\end{align}
thus confirming the $m=3$ case \cite[Conjecture 5.2]{GL}.
The aim of this paper is to prove the following refinements of \eqref{eq1.17} and \eqref{eq1.18}, which were originally conjectured by
Guo and Schlosser \cite[Conjecture 3]{GS3} (The modulus $[n]_{q^2}^4$ case was first formulated by Guo \cite[Conjecture 6.3]{Guo-result}).
\begin{theorem}\label{thm3} Let $n>1$ be an odd integer. Then
\begin{align}
\sum_{k=0}^{(n+1)/2}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k} &\equiv (2q+2q^{-1}-1)[n]_{q^2}^4\pmod{[n]_{q^2}^4\Phi_n(q^2)},
\label{eq1.3} \\
\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k} &\equiv (2q+2q^{-1}-1)[n]_{q^2}^4\pmod{[n]_{q^2}^4\Phi_n(q^2)}. \label{eq1.4}
\end{align}
\end{theorem}
Let $n=p^r$ be an odd prime power, and take $q\to 1$ in \eqref{eq1.3} and \eqref{eq1.4}. Noticing that
$$
\lim_{q\to 1}\frac{(q^{-2},q^4)_k}{(q^4;q^4)_k}=\frac{-1}{4^k(2k-1)}{2k\choose k},
$$
we immediately obtain the following conclusion, which was observed by Guo \cite[Conjecture 6.4]{Guo-m3}.
\begin{corollary} Let $p$ be an odd prime and $r$ a positive integer. Then
$$
\sum_{k=0}^{(p^r+1)/2}\frac{(4k-1)^3}{256^k(2k-1)^4}{2k\choose k}^4\equiv 3p^{4r} \pmod{p^{4r+1}},
$$
$$
\sum_{k=0}^{p^r-1}\frac{(4k-1)^3}{256^k(2k-1)^4}{2k\choose k}^4\equiv 3p^{4r} \pmod{p^{4r+1}}.
$$
\end{corollary}
The remainder of this paper is organized as follows. In the next section, we show the proof of Theorem \ref{thm3}. Section 3 contains a parameter-generalization of Theorem \ref{thm2} and a proof of \cite[Theorem 4.3]{Guo-result} modulo $[n]_{q^2}^2(1-aq^{2n})(a-q^{2n})$.
\section{Proof of Theorem \ref{thm3}}
We first prove the following identity, which plays an important role in our proof of Theorem \ref{thm3}.
\begin{theorem}\label{thm2} Let $n$ be a positive integer. Then
\begin{equation}\label{eq:thm2.1}
\begin{split}
&\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k}\\
&\quad=(q^{2n}+1)^4[n]_{q^2}^4\frac{(q^{-2};q^4)^4_{n}}{(q^4;q^4)_{n}^4}
\biggl(2\cdot\frac{q^5+q^{4n+1}(q^{4n-2}-q^2-1)}{(q^2-1)^2}-q^{4n}\biggr).
\end{split}
\end{equation}
\end{theorem}
Note that \eqref{eq:thm2.1} could be regarded as a $q$-analogue of the following identity:
$$
\sum_{k=0}^{n-1}\frac{(4k-1)^3}{256^k(2k-1)^4}{2k\choose k}^4=\frac{16n^4(8n^2-12n+3){2n\choose n}^4}{256^n(2n-1)^4}.
$$
\begin{proof}
For convenience, let
$$
S(n)=\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k},
$$ and
$$
T(n)=(q^{2n}+1)^4[n]_{q^2}^4\frac{(q^{-2};q^4)^4_{n}}{(q^4;q^4)_{n}^4}f_n(q),
$$
where
$$
f_n(q)=2\cdot\frac{q^5+q^{4n+1}(q^{4n-2}-q^2-1)}{(q^2-1)^2}-q^{4n}.
$$
We proceed by induction on $n$. For $n=1,$ it is clear that
$$
S(1)=-\frac1{q^4}=T(1).
$$
Suppose that the statement is true for $n$. We now consider the $n+1$ case.
Using the induction hypothesis, we have
\begin{align*}
S(n+1)&=S(n)+[4n-1]_{q^2}[4n-1]^2\cdot\frac{(q^{-2};q^4)_n^4}{(q^4;q^4)_n^4}q^{4n}\\[5pt]
&=\frac{(q^{-2};q^4)_n^4}{(q^4;q^4)_n^4}\biggl([2n]_{q^2}^4f_n(q)+[4n-1]_{q^2}[4n-1]^2q^{4n} \biggr)\\[5pt]
&=\frac{(q^{-2};q^4)_{n+1}^4(1-q^{4n+4})^4}{(q^4;q^4)_{n+1}^4(1-q^{4n-2})^4}\biggl([2n]_{q^2}^4f_n(q)
+[4n-1]_{q^2}[4n-1]^2q^{4n}\biggr)\\[5pt]
&=(1+q^{2n+2})^4[n+1]_{q^2}^4\frac{(q^{-2};q^4)_{n+1}^4}{(q^4;q^4)_{n+1}^4(1-q^{4n-2})^4}\\[5pt]
&\quad\times\biggl((1-q^{4n})^4f_n(q)+(1-q^{2(4n-1)})(1-q^{4n-1})^2(1-q)(1+q)^3q^{4n}\biggr)\\[5pt]
&=T(n+1),
\end{align*}
where the last equality holds because of the following relation
\begin{align*}
&(1-q^{4n})^4f_n(q)+(1-q^{2(4n-1)})(1-q^{4n-1})^2(1-q)(1+q)^3q^{4n}\\
&\quad=(1-q^{4n-2})^4f_{n+1}(q).
\end{align*}
This completes the proof of Theorem \ref{thm2}.
\end{proof}
We are now able to prove Theorem \ref{thm3}.
\begin{proof}[Proof of \eqref{eq1.3}]
Replacing $n$ by $(n+3)/2$ in \eqref{eq:thm2.1}, we obtain
\begin{align}
\sum_{k=0}^{(n+1)/2}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k}
&=(q^{n+3}+1)^4\frac{(1-q^{n+3})^4(q^{-2};q^4)_{(n+3)/2}^4}{(1-q^2)^4(q^4;q^4)_{(n+3)/2}^4}
f_{\frac{n+3}{2}}(q) \notag\\
&=[n]_{q^2}^4\frac{(q^{-2};q^4)_{(n+1)/2}^4}{(q^4;q^4)_{(n+1)/2}^4}f_{\frac{n+3}2}(q). \label{eq3.4}
\end{align}
In light of \cite[(4.2)]{Guo-result}, we have
$$
\frac{(q^{-2};q^4)_{(n+1)/2}}{(q^4;q^4)_{(n+1)/2}}\equiv(-1)^{(n+1)/2}q^{(n-1)^2/2-2}\pmod{\Phi_n(q^2)},
$$
and so the right-hand side of \eqref{eq3.4} is congruent to
$$
[n]_{q^2}^4q^{2(n-1)^2-8}\cdot q^5(2q^2-q+2)\equiv [n]_{q^2}^4(2q+2q^{-1}-1)
$$
modulo $[n]_{q^2}^4\Phi(q^2)$. This completes the proof.
\end{proof}
\begin{proof}[Proof of \eqref{eq1.4}]
It is easy to see that
$$
\frac{(q;q^2)_n}{(q^2;q^2)_n}=\frac 1{(-q;q)^2_n}{2n\brack n},
$$
where the \emph{$q$-binomial coefficients} ${n\brack k}$ are defined by
$$
{n\brack k}={n\brack k}_q
=\begin{cases}
\dfrac{(q;q)_n}{(q;q)_k(q;q)_{n-k}} & \mbox{if } 0\leq k\leq n, \\[15pt]
0 & \mbox{otherwise}.
\end{cases}$$
So the identity \eqref{eq:thm2.1} may be restated as
\begin{align}
&\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^4}{(q^4;q^4)_k^4}q^{4k} \notag\\
&\quad=\frac{(q^{2n}+1)^4[n]_{q^2}^4(q^2-1)^4}{(q^2-q^{4n})^4(-q^2;q^2)_n^8}{2n\brack n}_{q^2}^4\left(2\cdot\frac{q^5+q^{4n+1}(q^{4n-2}-q^2-1)}{(q^2-1)^2}-q^{4n}\right).
\label{eq3.1}
\end{align}
Since $q^n\equiv 1\pmod{\Phi_n(q)}$, by \cite[Lemma 3.1]{GW} we have
\begin{equation}\label{eq3.2}
{2n\brack n}_{q^2}=(1+q^{2n}){2n-1\brack n-1}_{q^2}\equiv 2(-1)^{n-1}q^{n(n-1)}\equiv 2\pmod{\Phi_n(q^2)}
\end{equation}
for odd $n>1$. In view of \cite[Lemma 3.2]{GW}, the following $q$-congruence holds:
\begin{equation}\label{eq3.3}
(-q^2;q^2)_n=(1+q^{2n})(-q^2;q^2)_{n-1}\equiv 2\pmod {\Phi_n(q^2)}.
\end{equation}
Applying \eqref{eq3.2} and \eqref{eq3.3}, we see that the right-hand side of \eqref{eq3.1} is congruent to
$$
[n]_{q^2}^4\left(2\cdot\frac{q^5+q(q^{-2}-q^2-1)}{(q^2-1)^2}-1\right)=[n]_{q^2}^4(2q+2q^{-1}-1)
$$
modulo $[n]_{q^2}^4\Phi(q^2)$. This completes the proof.
\end{proof}
\section{A further generalization of Theorem \ref{thm2}}
By induction on $n$, we can also prove the following parameter generalization of \eqref{eq:thm2.1}.
\noindent \begin{theorem}\label{thm4} Let $n>1$ be an integer. Then
\begin{align}
&\sum_{k=0}^{n-1}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^2(q^{-2}/a;q^4)_k(aq^{-2};q^4)_k}
{(q^4;q^4)_k^2(q^4/a;q^4)_k(aq^4;q^4)_k}q^{4k} \notag\\
&\quad=q(q^{2n}+1)^2[n]_{q^2}^2{2n\brack n}_{q^2}^2
\frac{(1-q^{-2})^2(q^6/a;q^4)_{n-2}(aq^6;q^4)_{n-2}f_n(a,q)}
{(1-q^{4n-2})^2(-q^2;q^2)_n^4(aq^4;q^4)_{n-1}(q^4/a;q^4)_{n-1}}, \label{eq:thm4}
\end{align}
where
\begin{align*}
f_n(a,q)&=2+2\sum_{i=1}^{4n-6}i(q^i+q^{8n-6-i})
-\frac{(a^2+1)(q^{8n-5}-2q^{4n-1}+2q^{4n-2}-2q^{4n-3}+q)}{a(q-1)^2}
\\&\quad+(8n-10)q^{4n-4}(1+q^2)+(8n-11)q^{4n-5}(1+q^4)+(8n-8)q^{4n-3}+2q^{8n-6}.
\end{align*}
\end{theorem}
We end this paper with the following generalization of \cite[Theorem 4.3]{Guo-result}.
\begin{theorem}\label{thm5} Let $n>1$ be an odd integer and $a$ an indeterminate. Then, modulo $[n]_{q^2}^2(1-aq^{2n})(a-q^{2n})$,
\begin{equation}\label{eq:thm5}
\sum_{k=0}^{M}[4k-1]_{q^2}[4k-1]^2\frac{(q^{-2};q^4)_k^2(q^{-2}/a;q^4)_k(aq^{-2};q^4)_k}
{(q^4;q^4)_k^2(q^4/a;q^4)_k(aq^4;q^4)_k}q^{4k}\equiv 0,
\end{equation}
where $M=(n+1)/2$ or $n-1$.
\end{theorem}
\begin{proof}
Let $n\mapsto (n+3)/2$ in \eqref{eq:thm4}, and notice the following realtion
\begin{align*}
\frac{1}{(1-q^{2n+4})^2}\biggl[\frac{n+3}2\biggr]_{q^2}^2{n+3\brack \frac{n+3}{2}}_{q^2}^2
=[n]_{q^2}^2{n-1\brack\frac{n-1}{2} }_{q^2}^2(1+q^{n+3})^2\frac{(1+q^{n+1})^2}{(1-q^{n+1})^2}.
\end{align*}
Since
\begin{equation*}
\gcd(1-q^n,1+q^m)=1\quad\text{and}\quad\gcd([n],[2n-1])=1
\end{equation*}
for all positive integers $m$ and $n$ with $n$ odd, we conclude that
the $q$-congruence \eqref{eq:thm5} holds for $M=n-1$ and $(n+1)/2$.
\end{proof}
| be0e6c28d9209b55abb6d457ac48af7cd4e73b02 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\vspace{-0.15in}
\begin{figure}
\centering
\includegraphics[height=5.5cm,width=0.9\linewidth]{FIG/new_intro32.pdf}
\caption{Starting from 32M GitHub\xspace repositories, we find 7.5K malware source code repositories using 137 malware keywords (Q137).}
\label{fig:intro}
\end{figure}
Security research could greatly benefit by an extensive database of malware source code, which is currently unavailable.
This is the assertion that motivates this work.
First, security researchers can use malware source code
to: (a) understand malware behavior and techniques,
and (b) evaluate security methods and tools~\cite{darki2019idapro,jerkins2017motivating}.
In the latter, having the source code can provide the groundtruth for
assessing the effectiveness of different techniques, such as reverse engineering methods\cite{healey2019source, schulte2018evolving,chen11novel} and anti-virus methods.
Second, currently, a malware {\bf source code} database is not readily available.
By contrast, there are several databases with malware {\bf binary code}, as collected via honeypots, but even those are often limited in number and not widely available. We discuss existing malware archives in Section~\ref{sec:related}.
{\bf A missed opportunity:}
Surprisingly, software archives, like GitHub\xspace, host many publicly-accessible
malware repositories, but this has not yet been explored to provide
security researchers with malware source code for their studies.
In this work, we focus on GitHub\xspace which is arguably the largest software storing and sharing platform.
As of October 2019, GitHub\xspace
reports more than 34 million users \cite{git:user} and more than 32 million public repositories~\cite{git:pub-repo}.
As we will see later, there are thousands of repositories that have malware source code, which seem to have escaped the radar of the research community so far.
Why do authors create public malware repositories?
This question mystified us: these repositories expose
both the creators and the intelligence behind the malware.
Intrigued, we conducted a small investigation on malware authors, as we discuss below.
{\bf Problem:} How can we find malware source code repositories in a large archive, like GitHub\xspace?
The input to the problem is an online archive and the
desired output is a database of malware repositories.
The challenges include: (a) collecting an appropriate set of repositories from the potentially vast archive, (b) identifying the repositories that contain malware.
Optionally, we also want to further help researchers that will potentially use these repositories, by determining additional properties, such as
the most likely target platform, the malware type or family etc.
Another practical challenge is the need to create the ground truth for validation purposes.
{\bf Related work:}
To the best of our knowledge, there does not seem to be any studies
focusing on the problem above. We group related works in the following categories.
First, several studies analyze software repositories
to find usage and limitations without any focus on
malware~\cite{cosentino2016findings}.
Second, several efforts maintain databases of malware binaries but without source code~\cite{allix2016androzoo,arp2014drebin}.
Third, many efforts attempt to extract higher-level information from binaries, such as lifting to Intermediate Representation(IR)~\cite{vdurfina2011design}, but it is really difficult to re-create the source code~\cite{chen2013refined}. In fact, such studies would benefit
from our malware source-code archive to evaluate and improve their methods. Taking a software engineering angle, an interesting work~\cite{calleja2016look} compares the evolution of 150 malware source code repositories with that of benign software.
We discuss related works in Section~\ref{sec:related}.
{\bf Contributions:}
Our work is arguably the first to systematically identify
malware source code repositories from a massive public archive.
The contribution of this work is three-fold:
(a) we propose SourceFinder\xspace, a systematic approach to
identify malware source-code repositories with high precision,
(b) we create, arguably, the largest non-commercial malware source code archive with \sourceNum repositories,
and (c) we study patterns and trends of the repository ecosystem including temporal and author-centric properties and behaviors.
We apply and evaluate our method on the GitHub\xspace archive, though it could also be used on other archives, as we discuss in Section~\ref{sec:discussion}.
Our key results can be summarized in the following points, and some key numbers are shown in Figure~\ref{fig:intro}.
a. We collect {\bf \totalrepo malware-related repositories} from GitHub\xspace. In the collection, we overcome various practical limitations,
and we also generate an extensive groundtruth with 2013 repositories, as we explain in Section~\ref{sub-sec:data-collection}.
b. {\bf SourceFinder\xspace achieves \precision precision.}
We systematically consider different Machine Learning approaches, and carefully-created representations for the different
fields of the repository, such as title, description etc.
We then systematically evaluate the effect of the different features, as we discuss in Section~\ref{sec:result}.
We show that we classify malware repositories with a \precision precision, \recall recall and \fOne F1-score using five fields from the repository.
c. {\bf We identify \sourceNum malware source-code repositories}, which is arguably the largest malware source-code database in the research community.
We have already downloaded the contents in these repositories, in case GitHub\xspace decides to deactivate them.
We also created a curated database of 250 malware repositories manually verified and spanning a wide range of malware types.
Naturally, we intend to make our datasets available for research purposes.
d. {\bf The number of new malware repositories in our data more than triples every four years}. The increasing trend is interesting and alarming at the same time.
e. {\bf We identify popular and influential repositories.} We identify the malware repositories using three metrics of popularity: the number of watchers, forks and stars.
We find 8 repositories that dominate the top-5 lists of all three metrics.
f. {\bf We identify prolific and influential authors.}
We find that 3\% of the authors have more than 300 followers.
We also find that
0.2\% of the authors have more than 7 malware repositories, with
the most prolific author {\em cyberthreats} having created 336 repositories.
g. {\bf We identify and profile 18\xspace professional hackers.} We find 18\xspace authors of malware repositories, who seem to have created a brand around their activities, as they use the same user names in security forums.
For example, user {\em 3vilp4wn} (pronounced evil-pawn) is the author of a keylogger malware in
GitHub\xspace, which the author is promoting in
the {\em Hack This Site} forum using the same username.
We present our study of malware authors in Section~\ref{sec:authors}.
{\bf Open-sourcing for maximal impact: creating an engaged community.}
We intend to make our datasets and our tools available for research purposes.
Our vision is to create community-driven reference center,
which will provide:
(a) malware source code repositories,
(b) community-vetted labels and feedback,
and
(c) open-source tools for collecting and analyzing malware repositories.
Our goal is to expand our database
with more software archives. Although authors could start hiding their repositories (see Section~\ref{sec:discussion}),
we argue that our already-retrieved database could have significant impact in enabling certain types of research studies.
\section{Background}
\label{backround-data}
\label{sub-sec:background}
\vspace{-0.15in}
We provide background information on GitHub\xspace and the type of information that repositories have.
GitHub\xspace is a massive world-wide software archive, which
enables users to share code through its public repositories
thus creating a global social network of interaction. For instance,
first, users can collaborate on a repository. Second, users often "fork" projects: they copy and evolve projects.
Third, users can follow projects,
and "up-vote" projects using "stars" (think Facebook likes).
Although GitHub\xspace has many private repositories, there are 32 million public software repositories.
We describe the key elements of a GitHub\xspace repository. A repository is equivalent to a project folder, and typically, each repository corresponds to a single software project.
A repository in GitHub\xspace has the following data fields:
a) title, b) description, c) topics, d) README file, e) file and folders, f) date of creation and last modified, g) forks, h) watchers, i) stars, and j) followers and followings, which we explain below.
{\bf 1. Repository title: }
The title is a mandatory field and it usually consists of less than 3 words.
{\bf 2. Repository description: }
This is an optional field that describes the objective of the project
and it is usually 1-2 sentences long.
{\bf 3. Repository topics: }
An author can optionally provide topics for her repository, in the form of tags, for example, \textit{``linux, malware, malware-analysis, anti-virus"}.
Note that 97\% of the repositories in our dataset have less than 8 topics.
{\bf 4. README file: }
As expected, the README file is a documentation and/or light manual for the repository. This field is optional and its size varies from
one or two sentences to many paragraphs.
For example, we found that 17.48\% of the README files
in our repositories are empty.
{\bf 5. File and folders: }
In a well-constructed software, the file and folder names of the source code can provide useful information.
For example, some malware repositories contain files or folders with
indicative names, such as ``malware", ''source code" or even specific malware types or names of specific malware, like {\em mirai}.
{\bf 6. Date of creation and last modification: }
GitHub\xspace maintains the date of creation and last modification of a repository.
We find malware repository created in 2008 are actively being modified by authors till present.
7. {\bf Number of forks:}
Users can fork a public repository: they can create a clone
of the project~\cite{jiang2017and}. An user can fork any public repository to change locally and contribute to the original project if the owner accepts the modification.
The number of forks is an indication of the popularity and impact of a repository.
Note that the number of forks indicates the number of distinct users that have forked a repository.
{\bf 8. Number of watchers:}
Watching a repository is equivalent to ``following" in the social media language. A ``watcher" will get notifications, if there is any new activity in that project. The numbers of watchers
is an indication of the popularity of a repository~\cite{dabbish2012social}.
{\bf 9. Number of stars: }
A user can ``star" a repository, which is equivalent to the ``like" function in social media~\cite{begel2013social}, and places the repository in the users favorite group, but does not provide constant updates as with the ``watching" function.
{\bf 10. Followers: }
Users can also follow other users' work.
If A follows B, A will be added to B's followers and B will be added to A's following list. The number of followers is an indication of the popularity of a user~\cite{lee2013github}.
\section{Data Collection} \label{sub-sec:data-collection}
\vspace{-0.15in}
\begin{table
\centering
\small
\begin{tabular}
{|p{0.1\linewidth}|p{0.65\linewidth}|r|}
\hline
{\bf Set} & {\bf Descriptions} & {\bf Size} \\ \hline
Q1 & Query set = \{"malware"\} & 1 \\ \hline
Q50 & Query with 50 keywords with Q1$\subset$Q50 & 50 \\ \hline
Q137 & Query with 137 keywords with Q50$\subset$Q137 & 137\\ \hline \hline \hline
RD1 & Retrieved repositories from query Q1 & 2775 \\ \hline
RD50 & Retrieved repositories from query Q50 & 14332 \\ \hline
RD137 & Retrieved repositories from query Q137 & 97375 \\ \hline \hline \hline
LD1 & Labeled subset of RD1 dataset & 379 \\ \hline
LD50 & Labeled subset of RD50 dataset & 755 \\ \hline
LD137 & Labeled subset of RD137 dataset & 879 \\ \hline \hline \hline
M1 & Malware source code repositories in RD1 & 680 \\ \hline
M50 & Malware source code repositories in RD50 & 3096 \\ \hline
M137 & Malware source code repositories in RD137 & 7504
\\ \hline \hline \hline
\curDataSet & Manually verified malware source code dataset & 250 \\ \hline
\end{tabular}
\caption{Datasets, their relationships, and their size.}
\label{tab:my_label}
\end{table}
The first step in our work is to collect repositories from GitHub\xspace that have a higher chance of being related to malware.
Extracting repositories at scale from GitHub\xspace hides several subtleties and challenges, which we discuss below.
Using the GitHub\xspace Search API, a user can query with a set of keywords and obtain the most relevant repositories.
We describe briefly how we select appropriate keywords, retrieve related repositories from GitHub\xspace and how we establish our ground truth.
{\bf A. Selecting keywords for querying: }
In this step, we want to retrieve repositories from GitHub\xspace in a way that: (a) provides as many as possible malware repositories,
and (b) provides a wide coverage over different types of malware.
For this reason, we select keywords from three categories: (a)
malware and security related keywords, such as malware and virus,
(b) malware type names, such as ransomware and keylogger, and (c) popular malware names, such as mirai. Due to space limitations, we will provide the full list of keywords in our website at publication time for repeatability purposes.
We define three sets of keywords that we use to query GitHub\xspace. The reason is that we want to assess the sensitivity of the number of keywords on the outcome. Specifically, we use the following query sets: (a) the {\bf Q1 set}, which only contains the keyword ``malware"; (b) the {\bf Q50 set}, which contains 50 keywords, and (c) the {\bf Q137 set} which contains 137 keywords.
The Q137 keyword set is a super-set of Q50, and Q50 is a superset of Q1.
As we will see below, using the query set Q137 provides wider coverage, and we recommend in practice.
We use the other two to assess the sensitivity of the results in the initial set of keywords.
We list our datasets in Table~\ref{tab:my_label}.
{\bf B. Retrieving related repositories:}
Using the Search API, we query GitHub\xspace with our set of keywords.
Specifically, we query GitHub\xspace with every keyword in our set separately.
In an ideal world, this would have been enough to collect
all related repositories:
a query with ``malware" (Q1) should return the many thousands related repositories, but this is not the case.
The search capability hides several subtleties and limitations.
First, there is a limit of 1000 repositories that a single search can return: we get the top 1000 repositories ordered by relevancy to the query.
Second, the GitHub API allows 30 requests per minute for an authenticated user and 10 requests per minute for an unauthenticated user.
{\bf Bypassing the API limitations.}
We were able to find a work around for the first limitation by using ranking option. Namely, a user can specify her preferred ranking order for the results based on: (a) best match, (b) most stars, (c) fewest stars, (d) most forks, (e) fewest forks, (f) most recently updated, and (g) the least recently updated order. By repeating a query with all these seven ranking options, we can maximize the number of distinct repositories that we get.
This way, for each keyword in our set, we search with these seven different ranking preferences to obtain a list of GitHub\xspace repositories.
{\bf C. Collecting the repositories:} We download all the repositories
identified in our queries using PyGithub~\cite{PyGithub}, and we obtain three sets of repositories RD1, RD50 and RD137.
These retrieved datasets have the same "subset" relationship
that they query sets have: RD1 $\subset$ RD50 $\subset$ RD137.
Note that we remove pathological repositories, mainly repositories with no actual content, or repositories "deleted" by GitHub\xspace.
For each repository, we collect and store: (a) repository-specific information, (b) author-specific information, and (c) all the code within the repository.
As we see from Table~\ref{tab:my_label}, using more and specialized malware keywords
returns significantly more repositories. Namely, searching with the keyword ``malware" does return 2775 repositories, but searching with the Q50 and Q137 returns 14332 and 97375 repositories respectively.
\begin{table}
\centering
\footnotesize
\begin{tabular}
{|p{0.28\linewidth}|p{0.28\linewidth}|p{0.28\linewidth}|}
\hline
{\bf Labeled Dataset} & {\bf Malware Repo.} & {\bf Benign Repo.} \\ \hline
LD137 & 313 & 566 \\ \hline
LD50 & 326 & 429 \\ \hline
LD1 & 186 & 193 \\ \hline
\end{tabular}
\caption{Our groundtruth: labeled datasets for each of the three queries, for a total of 2013 repositories.}
\label{tab:dat_def}
\end{table}
{\bf D. Establishing the groundtruth}:
As there was no available groundtruth, we needed to establish our own.
As this is a fairly technical task, we opted for
domain experts instead of Mechanical Turk users, as recommended by recent studies~\cite{gharibshah2020rest}. We use three computer scientists to manually label 1000 repositories, which we selected in a uniformly random fashion, from each of our dataset RD137 and RD50 and 600 repositories from RD1.
The judges were instructed to independently investigate every repository thoroughly.
{\bf Ensuring the quality of the groundtruth.}
To increase the reliability of our groundtruth, we took the following measures.
First, we asked judges to label a repository {\em only}, if they were certain that it is malicious or benign and distinct, and leave it unlabeled otherwise. We only kept the repositories for which the judges agreed unanimously.
Second, duplicate repositories were removed via manual inspection,
and were excluded from the final labeled dataset to avoid overfitting.
It is worth noting that we only found very few duplicates in the order of 3-5 in each dataset with hundreds of repositories.
With this process, we establish three separate labeled datasets named LD137, LD50, and LD1 starting from the respective malware repositories from each of our queries, as shown in Table~\ref{tab:dat_def}.
Although the labeled datasets are not 50-50, they are
representing both classes reasonably well, so that a naive solution
that will label everything as one class, would perform poorly. By contrast, our approach performs sufficiently well, as we will see in Section~\ref{sec:result}.
As there is no available dataset, we argue that we make a sufficient size dataset by manual effort.
\section{Overview of our Identification Approach}
\label{mal-identification}
\vspace{-0.15in}
Here, we describe our supervised learning algorithm to identify the repositories that contain malware.
{\bf Step 1. Data preprocessing:}
As in any Natural Language Processing (NLP) method, we start with some initial processing of the text to improve the effectiveness of the solution. We briefly outline three levels of processing functionality.
\textbf{a. Character level preprocessing:} We handle the character level ``noise" by removing special characters, such as punctuation and currency symbols, and fix Unicode and other encoding issues.
\textbf{b. Word level preprocessing:}
We eliminate or aggregate words following the best practices of Natural Language Processing~\cite{jivani2011comparative}.
First, we remove article words and other words that don't carry significant meaning on their own.
Second, we use a stemming technique
to handle inflected words. Namely, we want to decrease the dimensionality of the data by grouping words with the same "root". For example, we group the words ``organizing'', ``organized'', ``organize'' and ``organizes'' to one word ``organize''.
Third, we filter out common file and folder names that we do not expect to help
in our classification, such as ``LEGAL'', ``LICENSE'', ``gitattributes'' etc.
\textbf{c. Entity level filtering:} We filter entities that are likely not helpful in describing the scope of a repository. Specifically, we remove numbers, URLs, and emails, which are often found in the text.
We found that this filtering improved the classification performance.
In the future, we could consider mining URLs and other information, such as names of people, companies or youtube channels, to identify authors, verify intention, and find more malware activities.
{\bf Step 2. The repository fields:}
We consider fields from the repositories that can be numbers or text.
Text-based fields require processing in order to turn them into
classification features and we explain this below.
We use and evaluate the following text fields: title, description, topics, file and folder names and README file fields.
\textbf{Text field representation: }
We consider two techniques to represent each text field by a feature in the classification.
{\bf i. Bag of Words (BoW):}
The bag-of-words (BoW) model is among the most widely used representations of
a document. The document is represented as the number of occurrences of its words, disregarding grammar and word order~\cite{zhang2010understanding}. This model is commonly used in document classification where the frequency of each word is used as feature value for training a classifier~\cite{mctear2016conversational}. We use the model with the count vectorizer and TF-IDF vectorizer to create the feature vector.
In more detail,
we represent each text field in the repository with a vector $V[K]$,
where $V[i]$ corresponds to the significance of work $i$ for the text.
There are several ways to assign values $V[i]$: (a) zero-one to account for presence, (b) number of occurrences, and (c) the TF-IDF value of the word. We evaluated all the above methods.
{\em Fixing the number of words per field.}
To improve the effectiveness of our approach using BoW, we conduct a feature selection process, ${\chi}^2$ statistic following best practices~\cite{rogati2002high}. The ${\chi}^2$ statistic measures the lack of independence between a word (feature) and a class. A feature with lower chi-square score is less informative for that class, and thus
not useful in the classification.
We discuss this further in Section~\ref{sec:result}.
For each text-based field $f$,
we select the top $K_f$ words for that field, which exhibit the highest discerning power in identifying malware repositories.
Note that we set a value for $K_f$ during the training stage
For each field, we select the value $K_f$,
as we explain in Section~\ref{sec:result}.
{\bf ii. Word embedding:}
The word embedding model is a vector representations of each word in a document: each word is mapped to an M-dimensional vector of real numbers~\cite{mikolov2013distributed}, or equivalently are projected in an M-dimensional space. A good embedding ensures that words that are close in meaning have nearby representations in the embedded space.
In order to create the document vector, word embedding follows two approaches (i) frequency-based vectorizer(unsupervised)~\cite{schnabel2015evaluation} and (ii) content-based vectorizer(supervised)~\cite{kusner2015word}.
Note that in this type of representation, we do not use the {\em word level processing}, which we described in the previous step, since this method can leverage contextual information.
We use frequency-based word embedding with word average and TF-IDF vectorizer. We also use pre-trained model of Google word2vec~\cite{mikolov2013efficient} and Stanford (Glov)~\cite{pennington2014glove} to create the feature vector.
Finally, we create the vector of the repository by concatenating
the vectors of each field of that repository.
{\bf Step 3. Selecting the fields:} Another key question is which fields from the repository to use in our classification. We experiment with all of the fields listed in Section~\ref{sub-sec:background} and we explain our findings in the next Section.
{\bf Step 4. Selecting a ML engine:}
We design classifiers to classify the repositories into two classes: (i) malware repository and (ii) benign repository.
We systematically evaluate many machine learning algorithms \cite{murphy2012machine,bishop2006pattern}: Naive Bayes (NB), Logistic Regression (LR), Decision Tree (CART), Random Forest(RF), K-Nearest Neighbor (KNN), Linear Discriminant Analysis (LDA), and Support Vector Machine (SVM).
{\bf Step 5. Detecting source code repositories:}
We also want to identify the existence of source code in
the repositories, as the final step in providing malware source code to the community.
We propose a heuristic approach, which seems to work fairly well in practice.
First, we want to identify files in the repository that contain source code.
To do this, we start by examining their file extension.
If the file extension is one of the known programming languages: {\em Assembly, C/C++, Batch File, Bash Shell Script, Power Shell script, Java, Python, C\#, Objective-C, Pascal, Visual Basic, Matlab, PHP, Javascript, and Go},
we label it as a source file.
Second, if the number of source files in a repository exceeds
the {\bf \sourceThresh threshold (\sourceThreshS)}, we consider
that the repository contains source code.
{\em How effective is this heuristic?} It turns out that in practice it works pretty well, as we will see in Section \ref{sec:result}.
Given that authors go out of their way to share their malware openly, and even provide appropriate titles and keywords, it seems less likely that
they will attempt to obfuscate the existence of source code in the repository.
\section{Evaluation: Choices and Results}
\label{sec:result}
\vspace{-0.15in}
In this section, we evaluate the effectiveness of the classification based on the proposed methodology defined in Section~\ref{mal-identification}. More specifically, our goal here is to answer the following questions:
\begin{enumerate}
\itemsep-0.3em
\item {\bf Repository Field Selection:} Which repository fields should we consider in our analysis?
\item {\bf Field Representation:} Which feature representation is better between bag of words (BoW) and word embeddings and considering several versions of each?
\item {\bf Feature Selection}:
What are the most informative features in identifying malware repositories?
\item {\bf ML Algorithm Selection:} Which ML algorithm exhibits the best performance?
\item {\bf Classification Effectiveness:}
What is the precision, recall and F1-score of the classification?
\item {\bf Identifying Malware Repositories:} How many malware repositories do we find?
\item {\bf Identifying Malware Source Code Repository: } How many of the malware repositories have source code?
\end{enumerate}
Note that we have a fairly complex task: we want to identify the best fields, representation method and Machine Learning engine, while considering different values for parameters.
What complicates matters is that all these selections are interdependent. We present our analysis in sequence,
but we followed many trial and error and non-linear paths in reality.
{\bf 1. Selecting repository fields:}
We evaluated all the repository fields mentioned earlier.
In fact, we used a significant number of experiments with
different subsets of the features, not shown here due to space limitations.
We find that the title, description, topics, README file, and file and folder names have the most discerning power.
We also considered number of forks, watchers, and stars of the repository and the number of followers and followings of the author of the repository. We found that not only it did not help,
but it usually decreased the classification accuracy by 2-3\%.
One possible explanation is that the numbers of forks, stars and followers reflect the popularity rather than the content of a repository.
\begin{table
\centering
\footnotesize
\begin{tabular}
{|p{0.65\linewidth}|p{0.22\linewidth}|}
\hline
{\bf Representation} & \textbf{Classification Accuracy Range} \\ \hline
Bag of Words with Count Vectorizer & 86\%-51\% \\ \hline
Bag of Words with Count Vectorizer + Feature Selection & 91\%-56\% \\ \hline
Bag of Words with TF-IDF vectorizer & 82\%-63\% \\ \hline \hline
Word Embedding with Word Average & 85\%-72\% \\ \hline
Word Embedding with TF-IDF & 85\%-74\% \\ \hline \hline
Pretrained Google word2vec Model & 76\%-64\% \\ \hline
Pretrained Stanford (Glov) Model & 73\%-62\% \\ \hline
\end{tabular}
\caption{Selecting the feature representation model: We evaluate all the representations across seven machine learning approaches and report the range of the overall classification accuracy. }
\label{tab:feature-representation}
\end{table}
{\bf 2. Selecting a field representation:}
The goal is to find, which representation approach works better.
In Table \ref{tab:feature-representation}, we show the comparison of the range of classification accuracy across the 7 different ML algorithms that we will also consider below.
We find that Bag of Words with the count vectorizer representation
reaches 86\% classification accuracy,
with the word embedding approach nearly matches that with
85\% accuracy.
Note that we finetune the selection of words to represent each field in the next step.
Why does not the embedding approach outperform the bag of words?
One would have expected that the most complex embedding approach would have been the winner and with a significant margin.
We attribute this to the relatively small text size in most text fields, which also do not provide well-structured sentences (think two-three words for the title, and isolated words for the topics). Furthermore, the word co-occurrences does not exist in topics and file names field, which partly what makes embedding approaches work well in large and well structured documents \cite{li2015word, globerson2007euclidean}.
In the rest of this paper, we choose the Bag of Words with count vectorizer to represent our text fields, since it exhibits good performance and is computationally less intensive than the embedding method.
\begin{figure}
\centering
\includegraphics[height=5cm, width=0.99\linewidth]{FIG/model_comparison3.pdf}
\caption{Naive Bayes tops by performance (Accuracy, Precision, Recall and F1-score) comparison among NB, LR, CART, RF, KNN, LDA and SVM on dataset LD137.}
\label{fig:model_comparison}
\end{figure}
\textbf{3. Fixing the number of words per field.}
We want to identify the most discerning words from each text field, which is a standard process in NLP for improving the scalability, efficiency and accuracy of a text classifier \cite{chen2009feature}.
Using the ${\chi}^2$ statistic, we select the top $K_f$ best words from each field.
To select the appropriate number of words per field,
we followed the process below.
We vary $K_f$ = 5,10,20,30,40 and 50 for title, topic and README file, and
we find that the top 30 words in title, 10 words in topic and 10 words in README file exhibit the highest accuracy. Similarly, we try $K_f$ = 80, 90, 100, 110 and 120 for file names and $K_f$ = 300, 325, 350, 375, 400, 425, 450 and 475 for the description field. We find that the top 100 words for file and folder names and top 400 words for description field give the highest accuracy.
Note that we do this during training and refining the algorithm,
and then we continue to use these words as features in testing.
Thus, we select the top: (a) 30 words from the title, (b) 10 words from the topics, (c) 400 words from the description, (d) 100 words from the file names, and 10 words from the README file.
This leads to a total of 550 words across all fields.
For reference, we find 9253 unique words in the repository fields of our training dataset.
Reducing the focus on the top 550 most discerning words per field
increases the classification accuracy by as much as 20\% in some cases.
{\bf 4. Evaluating and selecting ML algorithms:}
We asses the classification performance of Multinomial Naive Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (CART), Random Forest(RF), Linear Discriminant Analysis (LDA), and K-Nearest-Neighbors (KNN),
and show their precision, recall and F1-score in Figure~\ref{fig:model_comparison}.
Multinomial Naive Bayes exhibits the best F1-score with \fOne,
striking a good balance between \precision precision, \recall recall for the malware class.
Detecting the benign class we do even better with 92\% precision, 94\% recall and 93\% F1-score.
By contrast, the F1-score of the other algorithms is below 79\%.
Note that KNN, LR and LDA provide higher precision, but with significantly lower recall. Clearly, one could use these algorithms
to get higher precision at the cost of lower total number of repositories.
We use Multinomial Naive Bayes as our classification engine for the rest of this study.
We attempt to explain the superior F1-Score of the Naive Bayes in our context. The main advantage of Naive Bayes over other algorithms is that it considers the features independently of each other and can handle large number of features better. As a result, it is more robust to noisy or unreliable features.
It also performs well in domains with many equally important features, where other approaches suffer, especially with a small training data, and it is not prone to overfitting~\cite{ting2011naive}. As a result, the Naive Bayes is considered a dependable algorithm for text classification and it is often used as the benchmark to beat~\cite{xu2018bayesian}.
\begin{figure
\includegraphics[width=0.99\linewidth, height=5cm]{FIG/precision-recall-malware2.pdf}
\caption{Assessing the effect of the number of keywords in the query: Precision, Recall and F1-score of our approach on the LD137, LD50 and LD1 labeled datasets.}
\label{fig:precision-recall}
\end{figure}
{\bf 5. Assessing the effect of the query set:}
We have made the following choices in the previous steps:
(a) 5 text-based fields,
(b) bag of words with count vectorization,
(c) 550 total words across all the fields,
and (d) the Multinomial Naive Bayes.
We perform 10-fold cross validation and report the precision, recall and F1-score in Figure~\ref{fig:precision-recall} for our three different labeled data sets.
We see that the
precision stays above 89\% for all three datasets,
with a recall above 77\%.
It is worth noting the relative stability of our approach with respect to the keyword set for the initial query especially between LD50 and LD137 datasets.
The LD1 dataset we observe higher accuracy, but significantly less recall compared to LD137. We attribute this fact to the single keyword used in selecting the repositories in LD1, which may have lead to a more homogeneous group of repositories.
Interestingly, LD50 seem to have the lower recall and F1-score
even though the differences are not that large.
\textbf{6. Identifying \malrepo malware repositories:}
We use LD137 to train our Multinomial Naive Bayes model and apply it on RD137 dataset. We find \malrepo malware repositories.
We also apply the same trained model on RD1 and RD50 and find 809 and 3615 malware repositories respectively, but this repositories are included in the \malrepo. (Recall that RD1 and RD50 are subsets of RD137).
\begin{table
\centering
\footnotesize
\begin{tabular}
{|p{0.14\linewidth}|p{0.13\linewidth}||p{0.18\linewidth}|p{0.32\linewidth}|}
\hline
\textbf{Dataset} & \textbf{Initial} & \textbf{Malware} & \textbf{Mal. + Source} \\ \hline
RD1 & 2775 & 809 & 680 \\ \hline
RD50 & 14332 & 3615 & 3096 \\ \hline
RD137 & 97375 & 8644 & 7504 \\ \hline
\end{tabular}
\caption{The identified repositories per dataset with: (a) malware, and (b) malware and source code.}
\label{tab:mal-identifcation}
\end{table}
\textbf{7. Identifying \sourceNum malware source code repositories:}
We use our heuristic approach to identify source code repositories.
We set our \sourceThresh threshold to 75\%, meaning that:
if more than 75\% of files in a repository are source code files, we label it as a source code repository. Applying this heuristic, we find that \sourceNum repositories
are most likely source code repositories in RD137. We use the name \textbf{M137} to refer to this group of malware source code repositories. We find 680 and 3096 malware source code repositories in RD1 and RD50 as shown in Table \ref{tab:mal-identifcation}. However, these are subset of M137, given that RD1 and RD50 are subsets of RD137.
To evaluate the effectiveness of our heuristic, we manually check 30 randomly-selected repositories from M137. We find that all 30 repositories contain source code\footnote{Apart from a manual verification, these 30 repositories were further stress-tested: (a) 20 where used in a separate static analysis study, and (b) 15 were compiled and run successfully within an emulator.
},
which corresponds to 100\% precision.
We will further evaluate the effectiveness of this heuristic in the future.
\textbf{8. A curated malware source code dataset: \curDataSet}
As a tangible contribution,
we provide, \curDataSet, a dataset of 250 repositories from the M137 dataset, which we manually verify for containing malware source code and
relating to a particular malware type.
Opting for diversity and coverage,
the dataset spans all the identified types: virus, backdoor, botnet, keylogger, worm, ransomware, rootkit, trojan, spyware, spoof, ddos, sniff, spam, and cryptominer.
While constantly updating, we will make this dataset available to researchers.
\section{A large scale study of malware}
\label{sec:longitudinal}
\vspace{-0.15in}
Encouraged by the substantial number of malware repositories,
we study the distributions and longitudinal properties of the
identified malware repositories in M137.
{\bf Caveat:} We provide some key observations in this section,
but they should be viewed as indicative and approximate trends
and only within the context of the collected repositories
and with the general assumption that repository titles and
descriptions are reasonably accurate.
In Section~\ref{sec:discussion}, we discuss issues around the
biases and limitations that our dataset may introduce.
\begin{figure}[h]
\centering
\includegraphics[height=5cm,width=0.40\textwidth]{FIG/loglog_forks_stars_watch.pdf}
\caption{CCDF distributions of forks, stars and watchers per repository.}
\label{fig:ccdf-repos}
\end{figure}
\textbf{A. Identifying influential repositories.}
The prominence of a repository can be measured by the number of \textit{forks},
\textit{stars}, and \textit{watchers}.
In Figure~\ref{fig:ccdf-repos}, we plot the complementary
cumulative distribution function (CCDF) of these three metrics for our
malware repositories.
{\bf Fork distribution:} We find that 2\% of the repositories
seem quite influential with at least 100 forks
as shown in Figure~\ref{fig:ccdf-repos}.
Recall that the fork counter indicates the number of distinct users that have forked a repository.
For reference, 78\% of the repositories have less
than 2 forks.
{\bf Star distribution}: We find that 2\% of the
repositories receive more than 250 stars as shown in Figure~\ref{fig:ccdf-repos}. For reference,
75\% of the repositories have less than 3 stars.
{\bf Watcher distribution}: In
Figure~\ref{fig:ccdf-repos}, we find that 1\% of the
repositories have more than 50 watchers. For reference,
we observe that 84\% of
the repositories have less than 3 watchers.
Note that these distributions are skewed, and follow patterns
that can be approximated by a log-normal distribution.
{\em Which are the most influential repositories?}
We find that 8 repositories dominate the top 5 spots across all three metrics: stars, forks, and watchers. We present a short profile of these dominant repositories in Table \ref{tab:top8repo}. Most of the repositories contain a single malware project, which is an established practice among the authors in GitHub\xspace~\cite{projectperrepo1,projectperrepo2}. We find that the repository ``theZoo''~\cite{thezoo}, created by \textit{ytisf} in 2014
is the most forked, watched, and starred repository with 1393 forks, 730 watchers and 4851 stars as of October, 2019.
However, this repository is quite unique and was created with the intention of being a malware database with 140 binaries and 80 source code repositories.
\begin{table}[t]
\centering
\footnotesize
\begin{tabular}
{|p{0.02\linewidth}|p{0.16\linewidth}|p{0.05\linewidth}|p{0.05\linewidth}|p{0.1\linewidth}|p{0.32\linewidth}|}
\hline
\textbf{R ID} & \textbf{Author} & \textbf{\# Star} & \textbf{\# Fork}& \textbf{\# Watcher}& \textbf{ Content of the Repository}\\ \hline
1 & ytisf & 4851 & 1393 & 730 & 80 malware source code and 140 Binaries\\ \hline
2 & n1nj4sec & 4811 & 1307 & 440 & Pupy RAT\\ \hline
3 & Screetsec & 3010 & 1135 & 380 & TheFatRat Backdoor\\ \hline
4 & malwaredllc & 2515 & 513 & 268 & Byob botnet\\ \hline
5 & RoganDawes & 2515 & 513 & 268 & USB attack platform\\ \hline
6 & Visgean & 626 & 599 & 127 & Zeus trojan horse\\ \hline
7 & Ramadhan & 535 & 283 & 22 & 30 malware samples\\ \hline
8 & dana-at-cp & 1320 & 513 & 125 & backdoor-apk backdoor\\ \hline
\end{tabular}
\caption{The profile of the top 5 most influential malware repositories across all three metrics with 8 unique repositories.}
\label{tab:top8repo}
\end{table}
{\bf Influence metrics are correlated}:
As one would expect, the influence and popularity metrics are correlated.
We use a common correlation metric,
the Pearson Correlation Coefficient ($r$)~\cite{benesty2009pearson}, measured in a scale of $[-1,1]$. We calculate the metric for all pairs of our three popularity metrics. We find that all of them exhibit higher positive correlation: stars vs. forks ($r=0.92$, $p<0.01$), forks vs. watchers ($r=0.91$, $p<0.01$) and watchers vs. stars ($r=0.91$, $p<0.01$).
\textbf{B. Malware Type and Target Platform.}
We wanted to get a feel for what type of malware we have identified.
As a first approximation,
we use the keywords found in the text fields to relate repositories in M137 with the type of malware and the
intended target platform.
Our goal is to create
the two-dimensional distribution per
malware type and the target platform as shown in
Table~\ref{tab:type_os_distribution}.
To create this
table,
we associate a repository
with keywords in its title, topics, descriptions, file names and README file fields of: (a) the 6 target platforms,
and (b) the 13 malware type keywords.
\textit{How well does this heuristic approach work?}
We provide two different indications of its relative effectiveness.
First, the vast majority of the repositories relate to
one platform or type of malware:
(a) less than 8\% relate to more than one platform,
and (b) less than 11\% relate to more than one type of malware.
Second, we manually verify the 250
repositories in our curated data~\curDataSet
and find a 98\% accuracy.
Below, we provide some observations from
Table~\ref{tab:type_os_distribution}.
{\bf a. Keyloggers reign supreme.} We see that one of the largest
categories is the keylogger malware with 679 repositories, which
are mostly affiliated with Windows and Linux platforms.
We discuss the emergence of keyloggers
below in our temporal analysis.
{\bf b. Windows and Linux are the most popular targets.}
Not surprisingly, we find that the majority of the malware repositories
are affiliated with these two platforms: 1592 repositories for
Windows, and 1365 for Linux.
{\bf c. MacOS-focused repositories: fewer, but they exist.}
Although MacOS platform are less common among PC users,
we see that malware repositories targeting such platforms indeed
exist. As shown in Figure~\ref{fig:platform-trend},
MacOS malware repositories are an order of magnitude less compared
to those for Windows and Linux.
\begin{table}[t]
\footnotesize
\begin{tabular}
{|p{0.10\linewidth}|r|r|r|r|r|r|r|}
\hline
\multirow{2}{*}{\textbf{Types}} & \multicolumn{7}{|c|}{\textbf{Target Platform}} \\
\cline{2-8}
& Wind. & Linux & Mac & IoT & Andr. & iOS & {\bf Total} \\
\hline \hline
{\bf Total} & {\bf 1592} & {\bf 1365} & {\bf 380} & {\bf 108} & {\bf 442} & {\bf 131} & {\bf 4018} \\ \hline \hline
key-
logger & 396 & 209 & 42 & 2 & 27 & 3 & {\bf 679} \\ \hline
back-
door & 181 & 227 & 37 & 11 & 51 & 4 & {\bf 511} \\ \hline
virus & 235 & 131 & 34 & 2 & 51 & 16 & {\bf 469} \\ \hline
botnet & 153 & 154 & 43 & 36 & 64 & 17 & {\bf 467} \\ \hline
trojan & 133 & 70 & 24 & 16 & 67 & 19 & {\bf 329} \\ \hline
spoof & 76 & 115 & 88 & 2 & 20 & 9 & {\bf 310} \\ \hline
rootkit & 55 & 163 & 13 & 2 & 19 & 3 & {\bf 255} \\ \hline
ransom-
ware & 117 & 67 & 14 & 1 & 33 & 13 & {\bf 245} \\ \hline
ddos & 71 & 95 & 20 & 10 & 9 & 3 & {\bf 208} \\ \hline
worm & 61 & 45 & 18 & 5 & 25 & 18 & {\bf 172} \\ \hline
spyware & 45 & 22 & 6 & 6 & 38 & 16 & {\bf 133} \\ \hline
spam & 40 & 29 & 18 & 14 & 23 & 5 & {\bf 129} \\ \hline
sniff & 29 & 38 & 23 & 1 & 15 & 5 & {\bf 111} \\ \hline
\end{tabular}
\caption{Distribution of the malware repositories from M137
dataset based on the malware type and malware target platform.
This table demonstrates the repositories that fit with the
criteria defined in Section~\ref{sec:longitudinal}.}
\label{tab:type_os_distribution}
\end{table}
\begin{figure*}[h]
\begin{subfigure}{0.26\textwidth}
\includegraphics[height=5cm,width=5cm]{FIG/overall-trend.pdf}
\caption{New malware repositories created per year.}
\label{fig:overall-trend}
\end{subfigure}
~
\begin{subfigure}{0.42\textwidth}
\includegraphics[height=5cm,width=7.7cm]{FIG/type-trend.pdf}
\caption{New repositories per type of malware per year.}
\label{fig:type-trend}
\end{subfigure}
~
\begin{subfigure}{0.30\textwidth}
\includegraphics[height=5cm,width=5.5cm]{FIG/platform-trend.pdf}
\caption{New malware repositories per target platform per year.}
\label{fig:platform-trend}
\end{subfigure}
\caption{New malware repositories per year: a) all malware, b) per type of malware, and c) per target platform.
}
\label{fig:trend}
\end{figure*}
\textbf{C. Temporal analysis.}
We want to study the evolution
and the trends of malware repositories.
We plot the number of new malware repositories per year: a) total malware, b) per type of malware, and c) per target platform in Figure~\ref{fig:trend}.
We discuss a few interesting temporal behaviors below.
{\bf a. The number of new malware repositories more than triples every four years.}
We see an alarming increase from 117 malware repositories in 2010 to 620 repositories in 2014
and to 2166 repositories in 2018.
We also observe a sharp increase of 70\% between 2015 to 2016 shown in Figure~\ref{fig:overall-trend}.
{\bf b. Keyloggers started a super-linear growth since 2010}
and are by far affiliated with the most new repositories per year since 2013, but their rate of growth reduced in 2018.
{\bf c. Ransomware repositories emerge in 2014 and
gain momentum in 2017}.
Ransomware experienced their highest growth rate in 2017 with 155 new repositories, while that number dropped to 103 in 2018.
{\bf d. Malware activity slowed down in 2018 across the board.}
It seems that 2018 is a slower year for all malware
even when seen by type
( Figure~\ref{fig:type-trend})
and target platform (Figure~\ref{fig:platform-trend}).
We find that the number of new malware repositories has dropped significantly in 2018 for most types of malware except virus, keylogger and trojan.
{\bf e. IoT and iPhone malware repositories become more visible after 2014}.
We find that IoT malware emerges in 2015 and iPhone malware sees an increase after 2014 in Figure~\ref{fig:platform-trend}.
We conjecture that
this is possibly encouraged by the emergence and increasing popularity of
specific malware: (a) WireLurker, Masque, AppBuyer malware~\cite{iosmalware2014} for iPhones, and (b) BASHLITE \cite{iotbashlite}, a Linux based botnet for IoT devices.
We find the names of the aferemntioned malware in many repositories starting in 2014.
Interestingly, the source code of the original BASHLITE botnet is available in a repository created by {\em anthonygtellez} in 2015.
{\bf f. Windows and Linux: dominant but slowing down.} In Figure \ref{fig:platform-trend}, we see that windows and linux malware are flattened between 2017 and 2018. By contrast, IoT and android repositories have increased.
\section{Understanding malware authors}
\label{sec:authors}
\vspace{-0.15in}
Intrigued by the fact that authors create public malware repositories, we attempt to understand and profile their behavior.
As a first step towards understanding the malware authors\xspace, we
want to assess their popularity and influence.
We use the following metrics:
(a) number of malware repositories which they created,
(b) number of followers,
(c) total number of watchers on their repositories,
and (d) total number of stars.
We focus on the first two metrics here.
We use the notation
\textit{top k authors} for any of the metrics above, where \textit{k} can be any positive integer to referring to "heavy-hitters".
\textbf{A. Finding influential malware authors\xspace. }
We study the distribution of the number of malware repositories created and the number followers per author in following.
First, we find that
15 authors are contributing roughly 5\% of all malware repositories
by examining the CCDF of the created repositories in Figure~\ref{fig:cddf-author}.
From the figure, we find an outlier author, {\em cyberthreats}, who doesn't follow power law distribution~\cite{faloutsos1999power},
has created 336 malware repositories.
We also find that 99\% authors have less than 5 repositories.
Second, we study the distribution of the number of followers per author but omit the plot due to space limitations. The distributions is skewed with
3\% (221) of the authors having more than 300 followers each,
while 70\% of the authors have less than 16 followers.
\begin{figure}[]
\centering
\includegraphics[height=4cm, width=0.40\textwidth]{FIG/loglog_repos_per_author_ccdf2.pdf}
\caption{CCDF of malware repositories per author.} \label{fig:cddf-author}
\end{figure}
\textbf{B. Malware authors\xspace strive for an online ``brand":}
In an effort to understand the motive of sharing
malware repositories, we make the following investigation.
{\bf a. Usernames seem persistent across online platforms.}
We find that many malware authors\xspace
use the same username consistently across many
online platforms, such as security forums.
We conjecture that they are developing a reputation
and they use their username as a ``unique" identifier.
We identify 18\xspace malware authors\xspace\footnote
{
Note that this does not mean that the other authors are not doing the same, but they could be active in other security forums or online platforms.
},
who are active in at least one of the three
security forums: Offensive Community, Ethical Hacker and Hack This Site, for which we happen to have access to their data. We conjecture that at least some of these usernames correspond to the same users based on the following two indications.
First, we find direct connections between the usernames across different platforms. For example,
user {\em 3vilp4wn} at the
``Hack This Site'' forum is promoting a keylogger malware
by referring to a
GitHub\xspace repository~\cite{hacktool3vilp4wn}
whose author has the same username.
Second, these usernames are fairly uncommon, which
increases the likelihood of belonging to the same person.
For example,
there is a GitHub\xspace user with the name {\em fahimmagsi},
and someone with the same username is boasting about their hacking successes in the ``Ethical Hacker'' forum.
As we will see below, {\em fahimmagsi} seems to
have a well-established online reputation.
{\bf b. ``Googling" usernames reveals significant hacking activities.}
Given that these GitHub\xspace usernames are fairly unique, it was natural
to look them up on the web at large.
Even a simple Internet
search with the usernames reveals
significant hacking activities, including hacking
websites or social networks, and offering hacking tutorials in YouTube.
We
investigate the \textit{top 40} most prolific malware authors\xspace using a web search with a {\em single} simple query: ``hacked by $<$username$>$''.
We then examine only the first page of search results.
Despite all these self-imposed restrictions,
we identify three users with substantial hacking related activities across Internet.
For example,
we find a number of news articles for hacking a series of websites by GitHub\xspace users {\em fahimmagsi} and {\em CR4SH} \cite{hacknewsfahimmagsi}~\cite{hacktoolcr4sh}. Moreover, we find user {\em n1nj4sec} sharing a multi-functional Remote Access Trojan (RAT) named ``Pupy", developed by her, which received significant news coverage in security articles back in March of 2019~\cite{pupyGithub}\cite{pupyarticle}.
We are confident that well-crafted and targeted searches can connect
more malware authors with hacking activities and usernames in other online forums.
\section{Discussion}
\label{sec:discussion}
\vspace{-0.15in}
We discuss the effectiveness and limitations of SourceFinder\xspace.
\textbf{a. Why is malware publicly available in the first place?}
Our investigation in Section~\ref{sec:authors} provides
strong indications that malware
authors want to actively establish their hacking reputation.
It seems that they want to boost their online credibility, which often translates to money. Recent works~\cite{Portnoff2017, Deb2018_USC3,Sapienza2018_USC2} study the
underground markets of malware services and tools:
it stands to reason that notorious hackers will attract more clients.
At the same time, GitHub\xspace acts as a collaboration platform, which can help hackers improve their tools.
\textbf{b. Do we identify every malware repository in GitHub\xspace?} Our tool can not guarantee that it will identify every malware repository in GitHub\xspace. First,
we can only identify repositories that ``want to be found":
(a) they must be public, and (b) they must be described with the
appropriate text and keywords.
Clearly, if the author wants to hide her repository, we won't be able to find it.
However, we argue that this defeats the purpose of having a public archive: if secrecy was desired, the code would have been shared through private links and services.
Second, our approach is constrained by GitHub\xspace querying limitations,
which we discussed in Section \ref{sub-sec:data-collection}, and the set of 137 keywords that we use.
However, we are encouraged by the number and the reasonable diversity
of the retrieved repositories we see in Table~\ref{tab:type_os_distribution}.
\textbf{c. Are our datasets representative?} This is the typical hard question for any measurement or data collection study.
First of all, we want to clarify that our goal
is to create a large database of malware source code.
So, in that regard, we claim that we accomplished our mission.
At the same time, we seem
to have a fair number of malware samples in each category of interest, as we see in Table~\ref{tab:type_os_distribution}.
Studying the trends of malware is a distant second goal, which we present with the appropriate caveat.
On the one hand, we are limited by GitHub\xspace's API operation, as we discussed earlier.
On the other hand, we attempt to reduce the biases that are
under our control.
To ensure some diversity among our malware,
we added as many words as we could in our 137 malware, which is likely to capture a wide range of malware types.
We argue that the fairly wide breadth of malware types in Table \ref{tab:type_os_distribution} is a good indication.
Note that our curated dataset \curDataSet with 250 malware is reasonably representative in terms of coverage.
{\bf d. What is the overlap among the identified repositories?}
Note that our repository does not include forked repositories,
since {\bf GitHub\xspace does not return forked repositories as answers to a query}.
Similarly, the breadth of the types of the malware as shown in Table~\ref{tab:type_os_distribution} hints at a reasonable diversity.
However, our tool cannot claim that the identified repositories are distinct nor is it attempting do so.
GitHub\xspace does not restrict authors from copying (downloading),
and uploading it as a new repository.
In the future, we intend to study the similarity and evolution among these repositories.
{\bf e. Are the authors of repositories the original creator of the source code?}
This is an interesting and complex question that goes beyond the scope of this work.
Identifying the original creator will require studying the source code of all related repositories, and analyzing the dynamics of the hacker authors, which we intend to do in the future.
{\bf f. Are all the malware authors malicious?}
Not necessarily. This is an interesting question, but it is not
central to the main point of our work.
On the one hand,
we find some white hackers or researchers, such as Yuval Nativ~\cite{yuval}, or Nicolas Verdier~\cite{Nicolas}.
On the other hand, several authors seem to be malicious, as we saw in Section~\ref{sec:authors}.
{\bf g. Are our malware repositories in "working order"?}
It is hard to know for sure, but we attempt to answer indirectly.
First, we pick 30 malware source codes and all of them compiled
and a subset of 15 of them actually run successfully in an emulated environment as we already mentioned.
Second, these public repositories are a showcase for the skills
of the author,
who will be reluctant to have repositories of low quality.
Third, public repositories, especially popular ones,
are inevitably scrutinized by their followers.
\textbf{h. Can we handle evasion efforts?} Our goal
is to create the largest malware source-code database possible
and having collected \sourceNum malware repositories seems like a great start.
In the future,
malware authors\xspace could obfuscate their repositories by using misleading titles, and description, and even filenames.
We argue that authors seem to want their repositories to be found, which is why they are public.
We also have to be clear: it is easy for the authors to hide their repositories, and they could would start by making them private or avoid GitHub\xspace altogether.
However, both these moves will diminish the visibility of the authors.
\textbf{i. Will our approach generalize to other archives? }
We believe that SourceFinder\xspace can generalize to other archives, which
provide public repositories, like GitLab and BitBucket. We find that these
sites allow public repositories and let the users retrieve
repositories. We have also seen equivalent
data fields (title, description, etc). Therefore, we are
confident
that our approach can work with other archives.
\section{Related Work}
\label{sec:related}
\vspace{-0.15in}
There are several works that attempt to determine if a piece of software is malware, usually focusing on a binary, using static or dynamic analysis~\cite{aycock2006computer,kolbitsch2009effective,shankarapani2011malware,damodaran2017comparison}. However,
to the best of our knowledge, no previous study has focused on
identifying malware source code in public software archives, such
as GitHub\xspace, in a systematic manner as we do in this work.
We highlight the related works in the following categories:
{\bf a. Studies that needed source code.}
Several
studies~\cite{lepik2018art,zhong2015stealthy,shen2018javascript}
use malware source code that are manually retrieved from
GitHub\xspace repositories.
Some studies~\cite{calleja2016look}~\cite{calleja2018malsource} compare the evolution and the code reuse of 150 malware source codes (with only some from GitHub\xspace) with that of benign software from a software engineering perspective and study the code reuse.
Overall, various studies
\cite{darki2019idapro,jerkins2017motivating} can benefit from
malware source code to fine-tune their approach.
\textbf{b. Mining and analyzing GitHub\xspace:}
Many studies have analyzed different aspects of GitHub\xspace, but not with
the intention of retrieving malware repositories.
First, there are efforts that
study the user interactions and collaborations on GitHub\xspace and their
relationship to other social media in
\cite{kollanyi2016automation,horawalavithana2019mentions,pletea2014security}.
Second, some efforts discuss the challenges in extracting
and analyzing data from GitHub\xspace with respect to sampling
biases~\cite{cosentino2016findings,gousios2017mining}.
Other works~\cite{kalliamvakou2014promises,kalliamvakou2016depth}
study how users utilize the various features
and functions of GitHub\xspace.
Several studies
\cite{howison2004perils,rainer2005evaluating,treude2018unusual}
discuss the challenges of mining
software archives, like \textit{SourceForge} and GitHub\xspace,
arguing that more information is required to make assertions about
users and software projects.
\textbf{c. Databases of malware source code:}
At the time of writing this paper, there are few malware source
code databases and are rarely updated such as project
\textit{theZoo}~\cite{thezoo}. To the best of our knowledge, there
does not exist an active archive of malware source code, where
malware research community can get an enough number of source code to analyze.
\textbf{d. Database of malware binaries:}
There are well established malware binary collection initiatives,
such as Virustotal~\cite{total2019virustotal} which provides analysis
result for a malware binary. There are also community based projects
such as VirusBay~\cite{virusbay} that serve as malware binary sharing
platform.
\textbf{e. Converting binaries to source code:}
A complementary approach is to try to generate the source code
from the binary, but this is a very hard task.
Some works~\cite{vdurfina2011design,vdurfina2013psybot} focus on
reverse engineering of the malware binary to a high-level language representation, but not source code.
Some other efforts
\cite{healey2019source, schulte2018evolving,chen11novel} introduce binary decompilation into readable source code.
However, malware authors use sophisticated obfuscation techniques \cite{saidi2010experiences}\cite{yakdan2016helping,chen2013refined} to make it difficult to reverse engineer a binary into source code.
{\bf f. Measuring and modeling hacking activity.}
Some other studies analyze the underground black market of hacking activities but their starting
point is security forums~\cite{Portnoff2017, Deb2018_USC3,Sapienza2018_USC2}, and as such they
study the dynamics of that community but without retrieving any malware code.
\section{Conclusion}
\vspace{-0.1in}
Our work capitalizes on a great missed opportunity: there are thousands of malware source code repositories on GitHub\xspace.
At the same time, there is a scarcity of
malware source code, which is necessary for
certain research studies.
Our work is arguably the first to develop a systematic approach to extract malware source-code repositories
at scale from GitHub\xspace.
Our work provides two main tangible outcomes:
(a) we develop SourceFinder\xspace, which identifies malware repositories
with \precision precision, and
(b) we create, possibly, the largest non-commercial malware source code archive with \sourceNum repositories.
Our large scale study provide some interesting trends for both the malware repositories and the dynamics of the malware authors.
We intend to open-source both SourceFinder\xspace and the database of malware source code to maximize the impact of our work.
Our ambitious vision is to become the authoritative
source for malware source code for the research community by providing tools, databases, and benchmarks.
| 723981cc41523e98b4ba8412f8f07e843a802b07 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec_int}
The so-called N\'ed\'elec or also edge element spaces of~\cite{Ned_mix_R_3_80} form,
on meshes consisting of tetrahedra, the most natural piecewise polynomial subspace of
the space $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ composed of square-integrable fields with square-integrable weak curl.
They are instrumental in numerous applications in link with electromagnetism,
see for example~\cite{Ass_Ciar_Lab_foundat_electr_18,Boss_elctr_98,Hiptmair_acta_numer_2002,Monk_FEs_Maxwell_03}.
The goal of this paper is to study two different but connected questions related to these spaces.
\subsection{Stable broken {\em H}$(\boldsymbol{\operatorname{curl}})$ polynomial extensions}
Polynomial extension operators are an essential tool in numerical analysis involving
N\'ed\'elec spaces, in particular in the case of high-order discretizations.
Let $K$ be a tetrahedron. Then, given a boundary datum in the form of a suitable polynomial
on each face of $K$, satisfying some compatibility conditions,
a {\em polynomial extension} operator constructs a curl-free polynomial in the interior of
the tetrahedron $K$ whose tangential trace fits the {\em boundary datum} and which is stable
with respect to the datum in the intrinsic norm. Such an operator was derived
in~\cite{Demk_Gop_Sch_ext_II_09}, as a part of equivalent developments in the
$H^1$ and $\boldsymbol H(\operatorname{div})$ spaces respectively in~\cite{Demk_Gop_Sch_ext_I_09}
and~\cite{Demk_Gop_Sch_ext_III_12}, see also~\cite{MunSol_pol_lift_97} and the references therein.
An important achievement extending in a similar stable way a given polynomial {\em volume datum}
to a polynomial with curl given by this datum in a single simplex, along with a similar result
in the $H^1$ and $\boldsymbol H(\operatorname{div})$ settings, was presented in~\cite{Cost_McInt_Bog_Poinc_10}.
The above results were then combined together and extended
from a single simplex to a patch of simplices sharing the given vertex in several cases:
in $\boldsymbol H(\operatorname{div})$ in two space dimensions in~\cite{Brae_Pill_Sch_p_rob_09} and in $H^1$ and
$\boldsymbol H(\operatorname{div})$ in three space dimensions in~\cite{Ern_Voh_p_rob_3D_20}. These results have
important applications to a posteriori analysis but also
to localization and optimal $hp$ estimates in a priori analysis,
see~\cite{Ern_Gud_Sme_Voh_loc_glob_div_21}. To the best of our knowledge,
a similar patchwise result in the $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ setting is not available yet,
and it is our goal to establish it here. We achieve it in our first main result,
Theorem~\ref{theorem_stability}, see also the equivalent form in
Proposition~\ref{prop_stability_patch} \revision{and the construction in Theorem~\ref{thm_sweep}}.
Let ${\TT^e}$ be a patch of tetrahedra sharing a given edge $e$ from a shape-regular mesh
$\mathcal T_h$ and let ${\omega_\edge}$ be the corresponding patch subdomain. Let $p\ge0$ be a polynomial
degree. \revision{Let $\boldsymbol j_p \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H(\operatorname{div},{\omega_\edge})$ with $\div \boldsymbol j_p = 0$ be
a divergence-free Raviart--Thomas field, and let $\boldsymbol \chi_p$ be in the broken N\'ed\'elec space
$\boldsymbol{\mathcal N}_p({\TT^e})$.} In this work, we establish that
\begin{equation} \label{eq_BPE}
\min_{\substack{\boldsymbol v_p \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},{\omega_\edge}) \\ \grad \times \boldsymbol v_p = \boldsymbol j_p}}
\|\boldsymbol \chi_p - \boldsymbol v_p\|_{\omega_\edge}
\leq
C
\min_{\substack{\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},{\omega_\edge}) \\ \grad \times \boldsymbol v = \boldsymbol j_p}}
\|\boldsymbol \chi_p - \boldsymbol v\|_{\omega_\edge},
\end{equation}
which means that the {\em discrete} constrained {\em best-approximation error} in the patch
is subordinate to the {\em continuous} constrained best-approximation error up to a
constant $C$. Importantly, $C$ only depends on the shape-regularity of the edge patch
and does {\em not depend} on the {\em polynomial degree $p$} under consideration.
Our proofs are constructive, which has a particular application in a posteriori error
analysis, as we discuss now.
\subsection{$p$-robust a posteriori error estimates \revision{by broken patchwise equilibration}
for \revision{the curl--curl problem}}
\label{sec_a_post_intr}
Let $\Omega \subset \mathbb R^3$ be a Lipschitz polyhedral domain with unit outward normal $\boldsymbol n$.
Let ${\Gamma_{\rm D}},{\Gamma_{\rm N}}$ be two disjoint, open, possibly empty subsets
of $\partial \Omega$ such that $\partial \Omega = \overline {\Gamma_{\rm D}} \cup \overline {\Gamma_{\rm N}}$.
Given a divergence-free field $\boldsymbol j: \Omega \to \mathbb R^3$ \revision{with zero normal trace on ${\Gamma_{\rm N}}$},
\revision{the curl--curl problem} amounts to seeking a field $\boldsymbol A: \Omega \to \mathbb R^3$
satisfying
\begin{subequations}
\label{eq_maxwell_strong}
\begin{alignat}{2}
\label{eq_maxwell_strong_volume}
&\grad \times \grad \times \boldsymbol A = \boldsymbol j, \quad \div \boldsymbol A = 0,&\qquad&\text{in $\Omega$}, \\
&\boldsymbol A \times \boldsymbol n = \boldsymbol 0,&\qquad&\text{on ${\Gamma_{\rm D}}$},\\
&(\grad \times \boldsymbol A) \times \boldsymbol n = \boldsymbol 0,\quad \boldsymbol A \cdot \boldsymbol n = 0,&\qquad&\text{on ${\Gamma_{\rm N}}$}.
\end{alignat}
Note that $\boldsymbol A \times \boldsymbol n = 0$ implies that $(\grad \times \boldsymbol A) \cdot \boldsymbol n=0$ on ${\Gamma_{\rm D}}$.
\revision{When $\Omega$ is not simply connected and/or when ${\Gamma_{\rm D}}$ is not connected,
the additional conditions
\begin{equation}
\label{eq_maxwell_cohomology}
(\boldsymbol A,\boldsymbol \theta)_\Omega = 0,
\qquad
(\boldsymbol j,\boldsymbol \theta)_\Omega = 0,
\qquad \forall \boldsymbol \theta \in \boldsymbol {\mathcal H}(\Omega,\GD)
\end{equation}
must be added in order to ensure existence and uniqueness of a solution
to~\eqref{eq_maxwell_strong}, where $\boldsymbol {\mathcal H}(\Omega,\GD)$ is the finite-dimensional ``cohomology''
space associated with $\Omega$ and the partition of its boundary (see Section \ref{sec_notat}).}
\end{subequations}
\revision{The boundary-value problem \eqref{eq_maxwell_strong} appears immediately
in this form in magnetostatics. In this case, $\boldsymbol j$ and $\boldsymbol A$ respectively represent a (known)
current density and the (unknown) associated magnetic vector potential, while the key quantity
of interest is the magnetic field $\boldsymbol h := \grad \times \boldsymbol A$.
We refer the reader to \cite{Ass_Ciar_Lab_foundat_electr_18,Boss_elctr_98,Hiptmair_acta_numer_2002,Monk_FEs_Maxwell_03}
for reviews of models considered in computational electromagnetism.}
In the rest of the introduction, we assume for simplicity that ${\Gamma_{\rm D}}=\partial\Omega$
(so that the boundary conditions reduce to $\boldsymbol A \times \boldsymbol n = \boldsymbol 0$ on $\partial\Omega$)
and that $\boldsymbol j$ is a piecewise polynomial in the Raviart--Thomas space,
$\boldsymbol j \in \boldsymbol{\mathcal{RT}}_p(\mathcal T_h) \cap \boldsymbol H(\operatorname{div},\Omega)$, $p \geq 0$. Let
$\boldsymbol A_h \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H_0(\boldsymbol{\operatorname{curl}},\Omega)$ be a numerical approximation to $\boldsymbol A$
in the N\'ed\'elec space. Then, the Prager--Synge equality~\cite{Prag_Syng_47}, \cf, \eg,
\cite[equation~(3.4)]{Rep_a_post_Maxw_07} or~\cite[Theorem~10]{Braess_Scho_a_post_edge_08},
implies that
\begin{equation}
\label{eq_PS}
\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_\Omega \leq \min_{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)
\\
\grad \times \boldsymbol h_h = \boldsymbol j}}
\|\boldsymbol h_h - \grad \times \boldsymbol A_h\|_\Omega.
\end{equation}
Bounds such as~\eqref{eq_PS} have been used in, \eg,
\cite{%
Creus_Men_Nic_Pir_Tit_guar_har_19,%
Creus_Nic_Tit_guar_Maxw_17,%
Han_a_post_Maxw_08,%
Neit_Rep_a_post_Maxw_10},
see also the references therein.
The estimate~\eqref{eq_PS} leads to a guaranteed and sharp upper bound. Unfortunately,
as written, it involves a global minimization over $\boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)$,
and is consequently too expensive in practical computations.
Of course, a further upper bound follows from~\eqref{eq_PS} for {\em any}
$\boldsymbol h_h \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)$
such that $\grad \times \boldsymbol h_h = \boldsymbol j$. At this stage, though, it is not clear how to find an
{\em inexpensive local} way of constructing a suitable field $\boldsymbol h_h$, called an
{\em equilibrated flux}. A proposition for the lowest degree $p=0$ was
given in~\cite{Braess_Scho_a_post_edge_08}, but suggestions
for higher-order cases were not available until very recently
in~\cite{Ged_Gee_Per_a_post_Maxw_19,Licht_FEEC_a_post_H_curl_19}.
In particular, the authors in~\cite{Ged_Gee_Per_a_post_Maxw_19} also prove efficiency,
\ie, they devise a field $\boldsymbol h_h^* \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)$ such that,
up to a generic constant $C$ independent of the mesh size $h$ but possibly
depending on the polynomial degree $p$,
\begin{equation} \label{eq_eff}
\|\boldsymbol h_h^* - \grad \times \boldsymbol A_h\|_{\Omega} \leq C \|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{\Omega},
\end{equation}
as well as a local version of~\eqref{eq_eff}.
Numerical experiments in~\cite{Ged_Gee_Per_a_post_Maxw_19} reveal very good effectivity indices,
also for high polynomial degrees $p$.
A number of a posteriori error estimates that are {\em polynomial-degree robust}, \ie,
where no generic constant depends on $p$, were obtained recently. For equilibrations
(reconstructions) in the $\boldsymbol H(\operatorname{div})$ setting in two space dimensions, they were first
obtained in~\cite{Brae_Pill_Sch_p_rob_09}. Later, they were extended to the $H^1$ setting
in two space dimensions in~\cite{Ern_Voh_p_rob_15} and to both $H^1$ and $\boldsymbol H(\operatorname{div})$ settings
in three space dimensions in~\cite{Ern_Voh_p_rob_3D_20}. Applications to problems with arbitrarily
jumping diffusion coefficients, second-order eigenvalue problems, the Stokes problem, linear
elasticity, or the heat equation are reviewed in~\cite{Ern_Voh_p_rob_3D_20}.
In the $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ setting, with application to
\revision{the curl--curl problem}~\eqref{eq_maxwell_strong}, however, to the best of
our knowledge, such a result was missing\footnote{We have learned very recently that a
modification of~\cite{Ged_Gee_Per_a_post_Maxw_19} can lead to a polynomial-degree-robust
error estimate, see~\cite{Ged_Gee_Per_Sch_post_Maxw_20}.}. It is our goal to establish it here,
and we do so in our second main result, Theorem~\ref{theorem_aposteriori}.
Our upper bound in Theorem~\ref{theorem_aposteriori} actually does {\em not derive}
from the Prager--Synge equality to take the form~\eqref{eq_PS}, since we do not construct
an equilibrated flux $\boldsymbol h_h^* \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)$. We instead perform a
{\em \revision{broken patchwise} equilibration} producing locally on each edge patch ${\TT^e}$
a piecewise polynomial $\boldsymbol h_h^{e} \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ such that
$\grad \times \boldsymbol h_h^{e} = \boldsymbol j$. Consequently, our error estimate rather takes the form
\begin{equation} \label{eq_up_intr}
\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_\Omega
\leq
\sqrt{6} C_{\rm L} C_{\rm cont} \left (\sum_{e \in \mathcal E_h}
\|\boldsymbol h_h^{e} - \grad \times \boldsymbol A_h\|_{{\omega_\edge}}^2 \right )^{1/2}.
\end{equation}
We obtain each local contribution $\boldsymbol h_h^{e}$ in a single-stage procedure,
in contrast to the three-stage procedure of~\cite{Ged_Gee_Per_a_post_Maxw_19}.
Our \revision{broken patchwise} equilibration is also rather inexpensive,
since the edge patches are smaller than the usual vertex patches employed
in~\cite{Braess_Scho_a_post_edge_08, Ged_Gee_Per_a_post_Maxw_19}.
Moreover, we can either solve the patch
problems, see~\eqref{eq_definition_estimator_2}, or replace them by a {\em sequential sweep}
through tetrahedra sharing the given edge $e$, see~\eqref{eq_definition_estimator_sweep_2}.
This second option yields \revision{a cheaper procedure where merely elementwise, in place of
patchwise, problems are to be solved and even delivers} a {\em fully explicit} a posteriori
error estimate in the {\em lowest-order} setting $p=0$. The price we pay for these
advantages is the emergence of the constant $\sqrt{6} C_{\rm L} C_{\rm cont}$ in our upper
bound~\eqref{eq_up_intr}; here $C_{\rm cont}$ is fully computable, only depends on the mesh
shape-regularity, and takes values around 10 for usual meshes, whereas $C_{\rm L}$ only depends on
the shape of the domain $\Omega$ \revision{and boundaries ${\Gamma_{\rm D}}$ and ${\Gamma_{\rm N}}$}, with in particular
$C_{\rm L} = 1$ whenever $\Omega$ is convex. Crucially, our error estimates are
{\em locally efficient} and polynomial-degree robust in that
\begin{equation} \label{eq_low_intr}
\|\boldsymbol h_h^{e} - \grad \times \boldsymbol A_h\|_{{\omega_\edge}} \leq C \|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{{\omega_\edge}}
\end{equation}
for all edges $e$, where the constant $C$ only depends on the shape-regularity of the
mesh, as an immediate application of our first main result
in Theorem~\ref{theorem_stability}. It is worth noting that the lower bound~\eqref{eq_low_intr}
is completely local to the edge patches ${\omega_\edge}$ and does not comprise any neighborhood.
\subsection{Organization of this contribution}
The rest of this contribution is organised as follows.
In Section~\ref{sec_not}, we recall the functional spaces,
state a weak formulation of problem~\eqref{eq_maxwell_strong},
describe the finite-dimensional Lagrange, N\'ed\'elec, and Raviart--Thomas spaces,
and introduce the numerical discretization of~\eqref{eq_maxwell_strong}.
Our two main results, Theorem~\ref{theorem_stability}
\revision{(together with its sequential form in Theorem~\ref{thm_sweep})}
and Theorem~\ref{theorem_aposteriori}, are formulated and discussed in Section~\ref{sec_main_res}.
Section~\ref{sec_num} presents a numerical illustration of our a posteriori error estimates for
\revision{curl--curl problem}~\eqref{eq_maxwell_strong}. Sections~\ref{sec_proof_a_post}
and~\ref{sec_proof_stability} are then dedicated to the proofs of our two main results.
\revision{Finally, Appendix~\ref{appendix_weber} establishes an auxiliary result of independent
interest: a Poincar\'e-like inequality using the curl of divergence-free fields in an edge patch.}
\section{\revision{Curl--curl problem} and N\'ed\'elec finite element discretization}
\label{sec_not}
\subsection{Basic notation}
\label{sec_notat}
Consider a Lipschitz polyhedral subdomain $\omega \subseteq \Omega$. We denote by
$H^1(\omega)$ the space of scalar-valued $L^2(\omega)$ functions with
$\boldsymbol L^2(\omega)$ weak gradient, $\boldsymbol H(\boldsymbol{\operatorname{curl}},\omega)$ the space of vector-valued $\boldsymbol L^2(\omega)$
fields with $\boldsymbol L^2(\omega)$ weak curl, and $\boldsymbol H(\operatorname{div},\omega)$ the space of vector-valued
$\boldsymbol L^2(\omega)$ fields with $L^2(\omega)$ weak divergence. Below, we use the notation
$({\cdot},{\cdot})_\omega$ for the $L^2(\omega)$ or $\boldsymbol L^2(\omega)$ scalar product and
$\|{\cdot}\|_\omega$ for the associated norm.
$L^\infty(\omega)$ and $\boldsymbol L^\infty(\omega)$ are the spaces of essentially bounded
functions with norm $\|{\cdot}\|_{\infty,\omega}$.
Let $\boldsymbol H^1(\omega) := \{\boldsymbol v \in \boldsymbol L^2(\omega)| \, v_i \in H^1(\omega), \, i=1, 2, 3\}$.
Let ${\gamma_{\rm D}}$, ${\gamma_{\rm N}}$ be two disjoint, open, possibly empty subsets of $\partial \omega$ such that
$\partial \omega = \overline {\gamma_{\rm D}} \cup \overline {\gamma_{\rm N}}$.
Then $H^1_{\gamma_{\rm D}}(\omega) := \{v \in H^1(\omega)| \, v=0$ on ${\gamma_{\rm D}}\}$
is the subspace of $H^1(\omega)$ formed by functions vanishing on ${\gamma_{\rm D}}$ in the sense of traces.
Furthermore, $\boldsymbol H_{\gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\omega)$ is the
subspace of $\boldsymbol H(\boldsymbol{\operatorname{curl}},\omega)$ composed of fields with vanishing tangential trace on
${\gamma_{\rm D}}$, $\boldsymbol H_{\gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\omega) := \{\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},\omega)$ such that
$(\grad \times \boldsymbol v, \boldsymbol \varphi)_\omega - (\boldsymbol v, \grad \times \boldsymbol \varphi)_\omega = 0$ for all functions
$\boldsymbol \varphi \in \boldsymbol H^1(\omega)$ such that $\boldsymbol \varphi \times \boldsymbol n_{\omega} = \boldsymbol 0$ on
$\partial \omega \setminus {\gamma_{\rm D}}\}$, where $\boldsymbol n_{\omega}$
is the unit outward normal to $\omega$.
Similarly, $\boldsymbol H_{\gamma_{\rm N}}(\operatorname{div},\omega)$ is the subspace of $\boldsymbol H(\operatorname{div},\omega)$ composed of fields
with vanishing normal trace on ${\gamma_{\rm N}}$, $\boldsymbol H_{\gamma_{\rm N}}(\operatorname{div},\omega) := \{\boldsymbol v \in \boldsymbol H(\operatorname{div},\omega)$
such that $(\div \boldsymbol v, \varphi)_\omega + (\boldsymbol v, \boldsymbol \nabla \varphi)_\omega = 0$ for all functions
$\varphi \in H^1_{\gamma_{\rm D}}(\omega)\}$. We refer the reader to~\cite{Fer_Gil_Maxw_BC_97}
for further insight on vector-valued Sobolev spaces with mixed boundary conditions.
\revision{The space $\boldsymbol K(\Omega) := \{ \boldsymbol v \in \boldsymbol H_{\Gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\Omega) \; | \;
\grad \times \boldsymbol v = \boldsymbol 0 \}$ will also play an important role. When
$\Omega$ is simply connected and ${\Gamma_{\rm D}}$ is connected, one simply
has $\boldsymbol K(\Omega) = \boldsymbol \nabla \left (H^1_{\Gamma_{\rm D}}(\Omega)\right )$. In the general
case, one has $\boldsymbol K(\Omega) = \boldsymbol \nabla \left (H^1_{\Gamma_{\rm D}}(\Omega)\right ) \oplus \boldsymbol {\mathcal H}(\Omega,\GD)$, where
$\boldsymbol {\mathcal H}(\Omega,\GD)$ is a finite-dimensional space called the ``cohomology space'' associated
with $\Omega$ and the partition of its boundary \cite{Fer_Gil_Maxw_BC_97}.}
\subsection{\revision{The curl--curl problem}}
\label{sec_Maxw}
If $\boldsymbol j \in \revision{\boldsymbol K(\Omega)^\perp}$
\revision{(the orthogonality being understood in $\boldsymbol L^2(\Omega)$)}, then the
classical weak formulation of~\eqref{eq_maxwell_strong} consists in finding a pair
$(\boldsymbol A,\revision{\boldsymbol \varphi}) \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega) \times \revision{\boldsymbol K(\Omega)}$ such that
\begin{equation}
\label{eq_maxwell_weak}
\left \{
\begin{alignedat}{2}
(\boldsymbol A,\revision{\boldsymbol \theta})_\Omega &= 0 &\quad& \forall \revision{\boldsymbol \theta \in \boldsymbol K(\Omega)}
\\
(\grad \times \boldsymbol A,\grad \times \boldsymbol v)_\Omega + (\revision{\boldsymbol \varphi},\boldsymbol v)_\Omega &= (\boldsymbol j,\boldsymbol v)_\Omega &\quad& \forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega).
\end{alignedat}
\right .
\end{equation}
Picking the test function $\boldsymbol v = \revision{\boldsymbol \varphi}$ in the second equation
of~\eqref{eq_maxwell_weak} shows that $\revision{\boldsymbol \varphi = \boldsymbol 0}$, so that we actually have
\begin{equation}
\label{eq_maxwell_weak_II}
(\grad \times \boldsymbol A,\grad \times \boldsymbol v)_\Omega = (\boldsymbol j,\boldsymbol v)_\Omega \quad \forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega).
\end{equation}
\revision{Note that when $\Omega$ is simply connected and ${\Gamma_{\rm D}}$ is connected,
the condition $\boldsymbol j \in \boldsymbol K(\Omega)^\perp$ simply means that $\boldsymbol j$ is
divergence-free with vanishing normal trace on ${\Gamma_{\rm N}}$, $\boldsymbol j \in \boldsymbol H_{{\Gamma_{\rm N}}}(\operatorname{div},\Omega)$
with $\div \boldsymbol j = 0$, and the same constraint follows from the first equation
of~\eqref{eq_maxwell_weak} for $\boldsymbol A$.}
\subsection{Tetrahedral mesh} \label{sec_mesh}
We consider a matching tetrahedral mesh $\mathcal T_h$ of $\Omega$, \ie,
$\bigcup_{K \in \mathcal T_h} \overline K$ $= \overline \Omega$, each $K$ is a tetrahedron,
and the intersection of two distinct tetrahedra is either empty or their common vertex,
edge, or face. We also assume that $\mathcal T_h$ is compatible with the partition
$\partial \Omega = \overline{{\Gamma_{\rm D}}} \cup \overline{{\Gamma_{\rm N}}}$ of the boundary, which means
that each boundary face entirely lies either in $\overline{{\Gamma_{\rm D}}}$ or in $\overline{{\Gamma_{\rm N}}}$.
We denote by $\mathcal E_h$ the set of edges of the mesh $\mathcal T_h$ and by $\mathcal F_h$ the set of
faces. The mesh is oriented which means that every edge
$e\in\mathcal E_h$ is equipped with a fixed unit tangent vector ${\boldsymbol \tau}_e$ and every face
$F\in\mathcal F_h$ is equipped with a fixed unit normal vector $\boldsymbol n_F$
(see~\cite[Chapter~10]{Ern_Guermond_FEs_I_21}). Finally for every mesh cell $K\in\mathcal T_h$,
$\boldsymbol n_K$ denotes its unit outward normal vector.
The choice of the orientation is not relevant in what follows,
but we keep it fixed in the whole work.
If $K \in \mathcal T_h$, $\mathcal E_K \subset \mathcal E_h$ denotes the set of edges of $K$, whereas
for each edge $e \in \mathcal E_h$,
we denote by ${\TT^e}$ the associated ``edge patch'' that consists of those
tetrahedra $K \in \mathcal T_h$ for which $e \in \mathcal E_K$,
see Figure~\ref{fig_patch}. We also employ the notation
${\omega_\edge} \subset \Omega$ for the open subdomain associated with the patch ${\TT^e}$.
We say that $e\in\mathcal E_h$ is a boundary edge if it lies on $\partial\Omega$ and that
it is an interior edge otherwise (in this case, $e$ may touch the boundary at one of its
endpoints). The set of boundary edges is partitioned into the subset of Dirichlet edges $\EE^{\rm D}_h$
with edges $e$ that lie in $\overline{{\Gamma_{\rm D}}}$ and the subset of Neumann edges $\EE^{\rm N}_h$
collecting the remaining boundary edges.
For all edges $e \in \mathcal E_h$, we denote by ${\Gamma_{\rm N}^\edge}$ the
open subset of $\partial {\omega_\edge}$ corresponding to the collection of faces having
$e$ as edge and lying in $\overline {\Gamma_{\rm N}}$. Note that for interior edges,
${\Gamma_{\rm N}^\edge}$ is empty and that for boundary edges, ${\Gamma_{\rm N}^\edge}$ never equals the whole $\partial {\omega_\edge}$.
We also set ${\Gamma_{\rm D}^\edge} := (\partial {\omega_\edge} \setminus {\Gamma_{\rm N}^\edge})^\circ$.
\revision{Note that, in all situations, ${\omega_\edge}$ is simply connected and ${\Gamma_{\rm D}^\edge}$ is connected,
so that we do not need to invoke here the cohomology spaces}.
\begin{figure}[htb]
\centerline{\includegraphics[height=0.35\textwidth]{figures/patch_edge.pdf} \qquad \qquad \includegraphics[height=0.35\textwidth]{figures/patch_edge_bound.pdf}}
\caption{Interior (left) and Dirichlet boundary (right) edge patch ${\TT^e}$}
\label{fig_patch}
\end{figure}
For every tetrahedron $K \in \mathcal T_h$, we denote the diameter and the inscribed ball
diameter respectively by
\begin{equation*}
h_K := \sup_{\boldsymbol x,\boldsymbol y \in K} |\boldsymbol x - \boldsymbol y|,
\quad
\rho_K := \sup \left \{ r > 0 \; | \; \exists \boldsymbol x \in K; B(\boldsymbol x,r) \subset K \right \},
\end{equation*}
where $B(\boldsymbol x,r)$ is the ball of diameter $r$ centered at $\boldsymbol x$.
For every edge $e \in \mathcal E_h$, $|e|$ is its measure (length) and
\begin{equation}
\label{eq_patch_not}
h_{\omega_\edge} := \sup_{\boldsymbol x,\boldsymbol y \in {\omega_\edge}} |\boldsymbol x - \boldsymbol y|, \quad
\rho_e := \min_{K \in {\TT^e}} \rho_K.
\end{equation}
The shape-regularity parameters of the tetrahedron $K$ and of the edge patch ${\TT^e}$
are respectively defined by
\begin{equation}
\label{eq_regularities}
\kappa_K := h_K/\rho_K \quad \text{ and } \quad \kappa_e := h_{{\omega_\edge}}/\rho_e.
\end{equation}
\subsection{Lagrange, N\'ed\'elec, and Raviart--Thomas elements}
If $K$ is a tetrahedron and ${p'} \geq 0$ is an integer,
we employ the notation $\mathcal P_{p'}(K)$ for the space of scalar-valued (Lagrange) polynomials
of degree less than or equal to ${p'}$ on $K$ and $\widetilde{\mathcal P}_{p'}(K)$
for homogeneous polynomials of degree ${p'}$. The notation $\boldsymbol{\mathcal P}_{p'}(K)$ (resp. $\widetilde{\boldsymbol{\mathcal P}}_{p'}(K)$)
then stands for the space of vector-valued polynomials such that all their components belong to
$\mathcal P_{p'}(K)$ (resp. $\widetilde{\mathcal P}_{p'}(K)$).
Following~\cite{Ned_mix_R_3_80} and~\cite{Ra_Tho_MFE_77}, we then define on each tetrahedron
$K \in \mathcal T_h$ the polynomial spaces of N\'ed\'elec and Raviart--Thomas functions as follows:
\begin{equation} \label{eq_RT_N}
\boldsymbol{\mathcal N}_{p'}(K) := \boldsymbol{\mathcal P}_{p'}(K) + \widetilde{\boldsymbol{\mathcal S}}_{{p'}+1}(K) \quad \text{ and }
\quad
\boldsymbol{\mathcal{RT}}_{p'}(K) := \boldsymbol{\mathcal P}_{p'}(K) + \boldsymbol x \widetilde{\mathcal P}_{p'}(K),
\end{equation}
where
$\widetilde{\boldsymbol{\mathcal S}}_{p'}(K) := \big\{
\boldsymbol v \in \widetilde{\boldsymbol{\mathcal P}}_{p'}(K) \; | \; \boldsymbol x \cdot \boldsymbol v(\boldsymbol x) = 0 \quad \forall \boldsymbol x \in \overline{K} \big \}$.
For any collection of tetrahedra $\mathcal T = \bigcup_{K \in \mathcal T}\{K\}$ and the corresponding open
subdomain $\omega = \big(\bigcup_{K \in \mathcal T} \overline{K} \big)^\circ \subset \Omega$,
we also write
\begin{align*}
\mathcal P_{p'}(\mathcal T)
&:=
\left \{
v \in L^2(\omega) \; | \; v|_K \in \mathcal P_{p'}(K) \quad \forall K \in \mathcal T
\right \},
\\
\boldsymbol{\mathcal N}_{p'}(\mathcal T)
&:=
\left \{
\boldsymbol v \in \boldsymbol L^2(\omega) \; | \; \boldsymbol v|_K \in \boldsymbol{\mathcal N}_{p'}(K) \quad \forall K \in \mathcal T
\right \},
\\
\boldsymbol{\mathcal{RT}}_{p'}(\mathcal T)
&:=
\left \{
\boldsymbol v \in \boldsymbol L^2(\omega) \; | \; \boldsymbol v|_K \in \boldsymbol{\mathcal{RT}}_{p'}(K) \quad \forall K \in \mathcal T
\right \}.
\end{align*}
\subsection{N\'ed\'elec finite element discretization}
For the discretization of problem~\eqref{eq_maxwell_weak}, we consider in this work\revision{,
for a fixed polynomial degree $p \geq 0$, the N\'ed\'elec finite element space given by}
\begin{equation*}
\boldsymbol V_{{\!h}} := \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega).
\end{equation*}
\revision{The discrete counterpart of $\boldsymbol K(\Omega)$, namely
\begin{equation*}
\boldsymbol K_h := \left \{\boldsymbol v_h \in \boldsymbol V_{{\!h}} \; | \; \grad \times \boldsymbol v_h = \boldsymbol 0 \right \}
\end{equation*}
can be readily identified as a preprocessing step by introducing cuts in the mesh
\cite[Chapter 6]{gross_kotiuga_2004a}.}
The discrete problem then consists in finding a pair
$(\boldsymbol A_h,\revision{\boldsymbol \varphi}_h) \in \boldsymbol V_{{\!h}} \times \revision{\boldsymbol K_h}$ such that
\begin{equation}
\label{eq_maxwell_discrete}
\left \{
\begin{alignedat}{2}
(\boldsymbol A_h,\revision{\boldsymbol \theta_h})_{\Omega} &= 0 && \quad \forall \revision{\boldsymbol \theta_h \in \boldsymbol K_h}
\\
(\grad \times \boldsymbol A_h,\grad \times \boldsymbol v_h)_{\Omega} + (\revision{\boldsymbol \varphi_h},\boldsymbol v_h)_{\Omega}
&=
(\boldsymbol j,\boldsymbol v_h)_{\Omega} && \quad \forall \boldsymbol v_{h} \in \boldsymbol V_{{\!h}}.
\end{alignedat}
\right.
\end{equation}
Since \revision{$\boldsymbol K_h \subset \boldsymbol K(\Omega)$}, picking $\boldsymbol v_h = \revision{\boldsymbol \varphi_h}$ in the second
equation of~\eqref{eq_maxwell_discrete} shows that \revision{$\boldsymbol \varphi_h = \boldsymbol 0$}, so that
we actually have
\begin{equation}
\label{eq_maxwell_discrete_II}
(\grad \times \boldsymbol A_h,\grad \times \boldsymbol v_h)_{\Omega} = (\boldsymbol j,\boldsymbol v_h)_{\Omega} \quad \forall \boldsymbol v_h \in \boldsymbol V_{{\!h}}.
\end{equation}
\revision{As for the continuous problem, we remark that when $\Omega$ is simply connected
and ${\Gamma_{\rm D}}$ is connected, $\boldsymbol K_h = \boldsymbol \nabla S_h$, where
$S_h := \mathcal P_{p+1}(\mathcal T_h) \cap H^1_{\Gamma_{\rm D}}(\Omega)$ is the usual Lagrange finite element space.}
\section{Main results} \label{sec_main_res}
This section presents our two main results.
\subsection{Stable discrete best-approximation of broken polynomials in {\em H}$(\boldsymbol{\operatorname{curl}})$}
Our first main result is the combination and extension of~\cite[Theorem~7.2]{Demk_Gop_Sch_ext_II_09}
and~\cite[Corollary~3.4]{Cost_McInt_Bog_Poinc_10} to the edge patches ${\TT^e}$, complementing
similar previous achievements in $\boldsymbol H(\operatorname{div})$ in two space dimensions
in~\cite[Theorem~7]{Brae_Pill_Sch_p_rob_09} and in $H^1$ and $\boldsymbol H(\operatorname{div})$ in three space
dimensions~\cite[Corollaries~3.1 and~3.3]{Ern_Voh_p_rob_3D_20}.
\begin{theorem}[$\boldsymbol H(\boldsymbol{\operatorname{curl}})$ best-approximation in an edge patch]
\label{theorem_stability}
Let an edge $e\in\mathcal E_h$ and the associated edge patch ${\TT^e}$ with subdomain ${\omega_\edge}$ be
fixed. Then, for every polynomial degree $p \geq 0$, all
$\boldsymbol j_h^{e} \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$
with $\div \boldsymbol j_h^{e} = 0$, and all $\boldsymbol \chi_h \in \boldsymbol{\mathcal N}_p({\TT^e})$,
the following holds:
\begin{equation}
\label{eq_stab}
\min_{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h_h = \boldsymbol j_h^e}}
\|\boldsymbol h_h - \boldsymbol \chi_h\|_{\omega_\edge}
\leq
C_{{\rm st},e}
\min_{\substack{
\boldsymbol h \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h = \boldsymbol j_h^{e}}}
\|\boldsymbol h - \boldsymbol \chi_h\|_{\omega_\edge}.
\end{equation}
Here, both minimizers are uniquely defined and the constant $C_{{\rm st},e}$ only depends on the
shape-regularity parameter $\kappa_e$ of the patch ${\TT^e}$ defined in~\eqref{eq_regularities}.
\end{theorem}
Note that the converse inequality to~\eqref{eq_stab} holds trivially with constant $1$, \ie,
\[
\min_{\substack{
\boldsymbol h \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h = \boldsymbol j_h^e}}
\|\boldsymbol h - \boldsymbol \chi_h\|_{\omega_\edge}
\leq
\min_{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h_h = \boldsymbol j_h^e}}
\|\boldsymbol h_h - \boldsymbol \chi_h\|_{\omega_\edge}.
\]
This also makes apparent the power of the result~\eqref{eq_stab}, stating that for piecewise
polynomial data $\boldsymbol j_h^e$ and $\boldsymbol \chi_h$, the best-approximation error over a piecewise polynomial
subspace of $\boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ of degree $p$ is, up to a $p$-independent constant,
equivalent to the best-approximation error over the entire space $\boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$.
The proof of this result is presented in Section~\ref{sec_proof_stability}.
We remark that Proposition~\ref{prop_stability_patch} below gives an equivalent
reformulation of Theorem~\ref{theorem_stability} in the form of a stable broken $\boldsymbol H(\boldsymbol{\operatorname{curl}})$
polynomial extension in the edge patch. \revision{Finally, the following form, which follows
from the proof in Section~\ref{sec:proof_stability_patch}, see Remark~\ref{rem_sweep_brok},
has important practical applications:
\begin{theorem}[$\boldsymbol H(\boldsymbol{\operatorname{curl}})$ best-approximation by an explicit sweep through an edge patch]
\label{thm_sweep}
Let the assumptions of Theorem~\ref{theorem_stability} be satisfied.
Consider a sequential sweep over all elements $K$ sharing the edge $e$,
$K \in {\TT^e}$ such that \textup{(i)} the enumeration starts from an arbitrary
tetrahedron if $e$ is an interior edge and from a tetrahedron containing a
face that lies in ${\Gamma_{\rm N}^\edge}$ (if any) or in ${\Gamma_{\rm D}^\edge}$ (if none in ${\Gamma_{\rm N}^\edge}$) if $e$ is a
boundary edge; \textup{(ii)} two consecutive tetrahedra in the enumeration share a face.
On each $K \in {\TT^e}$, consider
\begin{equation}
\label{eq_def_estimator_sweep}
\boldsymbol h_h^{e,\heartsuit}|_K := \min_{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p(K)\\
\grad \times \boldsymbol h_h = \boldsymbol j_h^e\\
\boldsymbol h_h|_{\mathcal F}^{\boldsymbol \tau} = \boldsymbol r_\mathcal F}}
\|\boldsymbol h_h - \boldsymbol \chi_h\|_K.
\end{equation}
Here, $\mathcal F$ is the set of faces that $K$ shares with elements $K'$
previously considered or lying in ${\Gamma_{\rm N}^\edge}$, and $\boldsymbol h_h|_{\mathcal F}^{\boldsymbol \tau}$ denotes
the restriction of the tangential trace of $\boldsymbol h_h$ to the faces of $\mathcal F$
(see Definition \ref{definition_partial_trace} below for details). The boundary datum
$\boldsymbol r_{\mathcal F}$ is either the tangential trace of $\boldsymbol h_h^{e,\heartsuit}|_{K'}$
obtained after minimization over the previous tetrahedron $K'$, or $\mathbf 0$ on ${\Gamma_{\rm N}^\edge}$. Then,
\begin{equation}
\label{eq_stab_sweep}
\|\boldsymbol h_h^{e,\heartsuit} - \boldsymbol \chi_h\|_{\omega_\edge}
\leq
C_{{\rm st},e}
\min_{\substack{
\boldsymbol h \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h = \boldsymbol j_h^{e}}}
\|\boldsymbol h - \boldsymbol \chi_h\|_{\omega_\edge}.
\end{equation}
\end{theorem}
}
\subsection{$p$-robust \revision{broken patchwise} equilibration a posteriori error
estimates for \revision{the curl--curl problem}}
Our second main result is a polynomial-degree-robust a posteriori error analysis of
N\'ed\'elec finite elements~\eqref{eq_maxwell_discrete} applied to
\revision{curl--curl problem}~\eqref{eq_maxwell_strong}. The local efficiency proof is an
important application of Theorem~\ref{theorem_stability}. To present these results in detail,
we need to prepare a few tools.
\subsubsection{Functional inequalities and data oscillation}
\label{sec_Poinc}
For every edge $e \in \mathcal E_h$, we associate with the subdomain ${\omega_\edge}$ a local Sobolev space
$H^1_\star({\omega_\edge})$ with mean/boundary value zero,
\begin{equation}
\label{eq_Hse}
H^1_\star({\omega_\edge}) :=
\begin{cases}
\{ v \in H^1({\omega_\edge})\; | \; v=0 \text{ on faces having $e$ as edge}\\
\hspace{5.3cm}\text{and lying in $\overline {\Gamma_{\rm D}}$}\} \qquad & \text{ if } e \in \EE^{\rm D}_h,\\
\left \{ v \in H^1({\omega_\edge}) \; | \; \int_{\omega_\edge} v = 0 \right \} \qquad & \text{ otherwise}.
\end{cases}
\end{equation}
Poincar\'e's inequality then states that there exists a constant $C_{{\rm P},e}$ only
depending on the shape-regularity parameter $\kappa_e$ such that
\begin{equation}
\label{eq_local_poincare}
\|v\|_{{\omega_\edge}} \leq C_{{\rm P},e} h_{\omega_\edge} \|\boldsymbol \nabla v\|_{{\omega_\edge}} \qquad \forall v \in H^1_\star({\omega_\edge}).
\end{equation}
To define our error estimators, it is convenient to introduce a piecewise polynomial
approximation of the datum $\boldsymbol j\in \boldsymbol H_{{\Gamma_{\rm N}}}(\operatorname{div},\Omega)$ by setting on every
edge patch ${\TT^e}$ associated with the edge $e \in \mathcal E_h$,
\begin{equation}
\label{eq_definition_jj_h}
\boldsymbol j_h^e := \argmin{\boldsymbol j_h \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge}) \\ \div \boldsymbol j_h = 0}
\|\boldsymbol j - \boldsymbol j_h\|_{{\omega_\edge}}.
\end{equation}
This leads to the following data oscillation estimators:
\begin{equation}
\label{eq_definition_osc}
\operatorname{osc}_e := C_{{\rm \revision{PFW}},e} h_{\omega_\edge} \|\boldsymbol j - \boldsymbol j_h^e\|_{{\omega_\edge}},
\end{equation}
\revision{where the constant $C_{{\rm \revision{PFW}},e}$ is such that for every edge $e \in \mathcal E_h$, we have}
\begin{equation}
\label{eq_local_poincare_vectorial}
\|\boldsymbol v\|_{{\omega_\edge}} \leq C_{{\rm \revision{PFW}},e} h_{{\omega_\edge}}\|\grad \times \boldsymbol v\|_{{\omega_\edge}}
\quad
\forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge}) \text{ with } \div \boldsymbol v = 0.
\end{equation}
\revision{We show in Appendix~\ref{appendix_weber} that $C_{{\rm \revision{PFW}},e}$ only depends on
the shape-regularity parameter $\kappa_e$. Notice that \eqref{eq_local_poincare_vectorial}
is a local Poincar\'e-like inequality using the curl of divergence-free fields in the edge patch.
This type of inequality is known under various names in the literature. Seminal
contributions can be found in the work of Friedrichs~\cite[equation~(5)]{Fried:55}
for smooth manifolds (see also Gaffney~\cite[equation~(2)]{Gaffney:55}) and later in
Weber~\cite{Weber:80} for Lipschitz domains. This motivates the present use of the subscript
${}_{\rm PFW}$ in~\eqref{eq_local_poincare_vectorial}.}
\revision{Besides the above local functional inequalities, we shall also use the fact}
that there exists a constant $C_{\rm L}$ such that for all $v \in \boldsymbol H_{\Gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\Omega)$,
there exists $\boldsymbol w \in \boldsymbol H^1(\Omega) \cap \boldsymbol H_{\Gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\Omega)$ such that $\grad \times \boldsymbol w = \grad \times \boldsymbol v$
and
\begin{equation}
\label{eq_estimate_lift}
\|\boldsymbol \nabla \boldsymbol w\|_{\Omega} \leq C_{\rm L} \|\grad \times \boldsymbol v\|_\Omega.
\end{equation}
When either ${\Gamma_{\rm D}}$ or ${\Gamma_{\rm N}}$ has zero measure, the existence of $C_{\rm L}$ follows
from Theorems 3.4 and 3.5 of~\cite{Cost_Dau_Nic_sing_Maxw_99}.
If in addition $\Omega$ is convex, one can take $C_{\rm L} = 1$
(see~\cite{Cost_Dau_Nic_sing_Maxw_99} together with~\cite[Theorem~3.7]{Gir_Rav_NS_86}
for Dirichlet boundary conditions and
\cite[Theorem~3.9]{Gir_Rav_NS_86} for Neumann boundary conditions).
\revision{For mixed boundary conditions, the existence of $C_{\rm L}$ can be obtained
as a consequence of \cite[Section~2]{Hipt_Pechs_discr_reg_dec_19}. Indeed, we first
project $\boldsymbol v \in \boldsymbol H_{\Gamma_{\rm D}}(\boldsymbol{\operatorname{curl}},\Omega)$ onto $\widetilde{\vv} \in \boldsymbol K(\Omega)^\perp$ without
changing its curl. Then, we define $\boldsymbol w \in \boldsymbol H^1(\Omega)$ from $\widetilde{\vv}$ using
\cite{Hipt_Pechs_discr_reg_dec_19}. Finally, we control $\|\widetilde{\vv}\|_\Omega$ by
$\|\grad \times \boldsymbol v\|_\Omega$ with the inequality from \cite[Proposition 7.4]{Fer_Gil_Maxw_BC_97}
which is a global Poincar\'e-like inequality in the spirit of~\eqref{eq_local_poincare_vectorial}.}
\subsubsection{\revision{Broken patchwise} equilibration by edge-patch problems}
Our a posteriori error estimator is constructed via a simple restriction of the right-hand side
of~\eqref{eq_PS} to edge patches, where no hat function is employed, no modification of
the source term appears, and no boundary condition is imposed for interior edges,
in contrast to the usual equilibration
in~\cite{Dest_Met_expl_err_CFE_99, Braess_Scho_a_post_edge_08, Ern_Voh_p_rob_15}.
For each edge $e \in \mathcal E_h$, introduce
\begin{subequations}\label{eq_definition_estimator}
\begin{equation}
\eta_e := \|\boldsymbol h_h^{e,\star} - \grad \times \boldsymbol A_h\|_{\omega_\edge}, \label{eq_definition_estimator_1}
\end{equation}
\revision{where $\boldsymbol h_h^{e,\star}$ is the argument of the left minimizer in~\eqref{eq_stab}
for the datum $\boldsymbol j_h^e$ from~\eqref{eq_definition_jj_h} and
$\boldsymbol \chi_h := (\grad \times \boldsymbol A_h)|_{{\omega_\edge}}$, \ie,}
\begin{equation}
\boldsymbol h_h^{e,\star} := \argmin{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})
\\
\grad \times \boldsymbol h_h = \boldsymbol j_h^e}}
\|\boldsymbol h_h - \grad \times \boldsymbol A_h\|_{\omega_\edge}. \label{eq_definition_estimator_2}
\end{equation}
\end{subequations}
\revision{In practice, $\boldsymbol h_h^{e,\star}$ is computed from the
Euler--Lagrange conditions for the minimization problem~\eqref{eq_definition_estimator_2}.
This leads to the} following patchwise mixed finite element problem:
Find $\boldsymbol h_h^{\revision{e},\star} \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$,
$\boldsymbol \sigma_h^{\revision{e},\star} \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$,
and $\zeta^{\revision{e},\star}_h \in \mathcal P_p({\TT^e})$ such that
\begin{equation} \label{broken_equil}
\left \{ \arraycolsep=2pt
\begin{array}{rclclcl}
(\boldsymbol h_h^{\revision{e},\star},\boldsymbol v_h)_{{\omega_\edge}}
&+&
(\boldsymbol \sigma_h^{\revision{e},\star},\grad \times \boldsymbol v_h)_{{\omega_\edge}}
& &
&=&
(\grad \times \boldsymbol A_h,\boldsymbol v_h)_{{\omega_\edge}},
\\
(\grad \times \boldsymbol h_h^{\revision{e},\star},\boldsymbol w_h)_{{\omega_\edge}}
&&
&+ &
(\zeta^{\revision{e},\star}_h,\div \boldsymbol w_h)_{\omega_\edge}
&=&
(\boldsymbol j,\boldsymbol w_h)_{{\omega_\edge}},
\\
& &
(\div \boldsymbol \sigma^{\revision{e},\star}_h,\varphi_h)_{\omega_\edge}
& &
&=&
0
\end{array}
\right .
\end{equation}
for all
$\boldsymbol v_h \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$,
$\boldsymbol w_h \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$,
and $\varphi_h \in \mathcal P_p({\TT^e})$. We note that from the optimality condition associated with~\eqref{eq_definition_jj_h},
using $\boldsymbol j$ or $\boldsymbol j_h^e$ in~\eqref{broken_equil} is equivalent.
\subsubsection{\revision{Broken patchwise} equilibration by sequential sweeps}
The patch problems~\eqref{eq_definition_estimator_2} lead to the solution of the linear
systems~\eqref{broken_equil}. Although these are local around each edge and are mutually
independent, they \revision{entail} some computational cost. This cost can be significantly
reduced by taking inspiration from~\cite{Dest_Met_expl_err_CFE_99},
\cite{Luce_Wohl_local_a_post_fluxes_04}, \cite[Section~4.3.3]{Voh_guar_rob_VCFV_FE_11},
the proof of~\cite[Theorem~7]{Brae_Pill_Sch_p_rob_09}, or~\cite[Section~6]{Ern_Voh_p_rob_3D_20}
and literally following the proof in Section~\ref{sec:proof_stability_patch} below\revision{,
as summarized in Theorem~\ref{thm_sweep}}.
This leads to \revision{an alternative error estimator} whose price is the sequential sweep through
tetrahedra sharing the given edge, \revision{where for each tetrahedron, one} solve\revision{s
the elementwise problem~\eqref{eq_def_estimator_sweep} for the
datum $\boldsymbol j_h^e$ from~\eqref{eq_definition_jj_h} and $\boldsymbol \chi_h := (\grad \times \boldsymbol A_h)|_{{\omega_\edge}}$, \ie,
\begin{subequations}
\label{eq_definition_estimator_sweep}
\begin{equation}
\label{eq_definition_estimator_sweep_2}
\boldsymbol h_h^{e,\heartsuit}|_K := \min_{\substack{
\boldsymbol h_h \in \boldsymbol{\mathcal N}_p(K)\\
\grad \times \boldsymbol h_h = \boldsymbol j_h^e\\
\boldsymbol h_h|_{\mathcal F}^{\boldsymbol \tau} = \boldsymbol r_\mathcal F}}
\|\boldsymbol h_h - \grad \times \boldsymbol A_h\|_K \qquad \forall K \in {\TT^e},
\end{equation}
and then set
\begin{equation}
\label{eq_definition_estimator_sweep_1}
\eta_e := \|\boldsymbol h_h^{e,\heartsuit} - \grad \times \boldsymbol A_h\|_{{\omega_\edge}}.
\end{equation}
\end{subequations}}
\subsubsection{Guaranteed, locally efficient, and $p$-robust a posteriori error estimates}
For each edge $e \in \mathcal E_h$, let $\boldsymbol \psi_e$ be the (scaled) edge basis functions
of the lowest-order N\'ed\'elec space, in particular satisfying
$\operatorname{supp} \boldsymbol \psi_e = \overline{{\omega_\edge}}$. More precisely, let $\boldsymbol \psi_e$
be the unique function in $\boldsymbol{\mathcal N}_0(\mathcal T_h) \cap \boldsymbol H(\boldsymbol{\operatorname{curl}},\Omega)$ such that
\begin{equation}
\label{eq_BF}
\int_{e'} \boldsymbol \psi_e \cdot {\boldsymbol \tau}_{e'} = \delta_{e,e'} |e|,
\end{equation}
recalling that ${\boldsymbol \tau}_{e'}$ is the unit tangent vector orienting the edge $e'$.
We define
\begin{equation}
\label{eq_definition_Ccont}
C_{{\rm cont},e}
:=
\|\boldsymbol \psi_e\|_{\infty,{\omega_\edge}}
+
C_{{\rm P},e} h_{\omega_\edge} \|\grad \times \boldsymbol \psi_e\|_{\infty,{\omega_\edge}} \quad \forall e \in \mathcal E_h,
\end{equation}
where $C_{{\rm P},e}$ is Poincar\'e's constant from~\eqref{eq_local_poincare} and $h_{\omega_\edge}$ is the diameter
of the patch domain ${\omega_\edge}$. We actually show in Lemma~\ref{lem_c_stab} below that
\begin{equation}
\label{eq_bound_Ccont}
C_{{\rm cont},e} \leq C_{\kappa_e} := \frac{2|e|}{\rho_e} \left ( 1 + C_{{\rm P},e} \kappa_e \right )
\quad \forall e \in \mathcal E_h,
\end{equation}
where $\rho_e$ is defined in~\eqref{eq_patch_not};
$C_{{\rm cont},e}$ is thus uniformly bounded by the patch-regularity parameter $\kappa_e$
defined in~\eqref{eq_regularities}.
\begin{theorem}[$p$-robust a posteriori error estimate]
\label{theorem_aposteriori}
Let $\boldsymbol A$ be the weak solution of \revision{the curl--curl problem}~\eqref{eq_maxwell_weak}
and let $\boldsymbol A_h$ be its N\'ed\'elec finite element approximation
solving~\eqref{eq_maxwell_discrete}. Let the data oscillation estimators $\operatorname{osc}_e$ be defined
in~\eqref{eq_definition_osc} and the \revision{broken patchwise} equilibration estimators
$\eta_e$ be defined in either~\eqref{eq_definition_estimator}
or~\eqref{eq_definition_estimator_sweep}.
Then, with the constants $C_{\rm L}$, $C_{{\rm cont},e}$, and $C_{{\rm st},e}$ from
respectively~\eqref{eq_estimate_lift}, \eqref{eq_definition_Ccont},
and~\eqref{eq_stab}, the following global upper bound holds true:
\begin{equation}
\label{eq_upper_bound}
\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_\Omega
\leq
\sqrt{6} C_{\rm L} \left (\sum_{e \in \mathcal E_h}C_{{\rm cont},e}^2
\left (\eta_e + \operatorname{osc}_e\right )^2 \right )^{1/2},
\end{equation}
as well as the following lower bound local to the edge patches ${\omega_\edge}$:
\begin{equation}
\label{eq_lower_bound}
\eta_e \leq C_{{\rm st},e} \left (\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{\omega_\edge} + \operatorname{osc}_e\right )
\qquad \forall e \in \mathcal E_h.
\end{equation}
\end{theorem}
\subsection{Comments}
A few comments about Theorem~\ref{theorem_aposteriori} are in order.
\begin{itemize}
\item The constant $C_{\rm L}$ from~\eqref{eq_estimate_lift} can be taken as $1$ for convex domains
$\Omega$ and if either ${\Gamma_{\rm D}}$ or ${\Gamma_{\rm N}}$ is empty. In the general case however, we do not know
the value of this constant. The presence of the constant $C_{\rm L}$ is customary in a posteriori
error analysis of \revision{the curl--curl problem}, it appears, e.g., in Lemma 3.10
of~\cite{Nic_Creus_a_post_Maxw_03} and Assumption 2 of~\cite{Beck_Hipt_Hopp_Wohl_a_post_Maxw_00}.
\item The constant $C_{{\rm cont},e}$ defined in~\eqref{eq_definition_Ccont} can be fully computed
in practical implementations. Indeed, computable values of Poincar\'e's constant $C_{{\rm P},e}$
from~\eqref{eq_local_poincare} \revision{are discussed in, \eg,
\cite{Chua_Whee_est_Poin_06, Vees_Verf_Poin_stars_12},
see also the concise discussion in~\cite{Blech_Mal_Voh_loc_res_NL_20}; $C_{{\rm P},e}$ can be taken
as $1/\pi$ for convex interior patches and as $1$ for most Dirichlet boundary patches.}
Recall also that $C_{{\rm cont},e}$ only depends on the shape-regularity parameter $\kappa_e$
of the edge patch ${\TT^e}$.
\item A computable upper bound on the constant $C_{{\rm st},e}$ from~\eqref{eq_stab} can be obtained
by proceeding as in~\cite[Lemma~3.23]{Ern_Voh_p_rob_15}. The crucial property is again that
$C_{{\rm st},e}$ can be uniformly bounded by the shape-regularity parameter $\kappa_e$ of the edge
patch ${\TT^e}$.
\item The key feature of the error estimators of Theorem~\ref{theorem_aposteriori} is their
polynomial-degree-ro\-bust\-ness (or, shortly, $p$-robustness). This suggests to use
them in $hp$-adaptation strategies, \cf, \eg,
\cite{Dan_Ern_Sme_Voh_guar_red_18,Demk_hp_book_07,Schwab_hp_98}
and the references therein.
\item In contrast to~\cite{Braess_Scho_a_post_edge_08, Ged_Gee_Per_a_post_Maxw_19,Ged_Gee_Per_Sch_post_Maxw_20, Licht_FEEC_a_post_H_curl_19},
we do not obtain here an equilibrated flux, \ie, a piecewise polynomial $\boldsymbol h_h^\star$ in the
global space $\boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H_{{\Gamma_{\rm N}}}(\boldsymbol{\operatorname{curl}},\Omega)$ satisfying, for piecewise polynomial
$\boldsymbol j$, $\grad \times \boldsymbol h_h^\star = \boldsymbol j$. We only obtain from~\eqref{eq_definition_estimator_2}
or~\eqref{eq_definition_estimator_sweep_2} that
$\boldsymbol h_h^{e,\star} \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ and
$\grad \times \boldsymbol h_h^{e,\star} = \boldsymbol j$ locally in every edge patch ${\TT^e}$
and similarly for $\boldsymbol h_h^{e,\heartsuit}$, but we do not build
a $\boldsymbol H_{{\Gamma_{\rm N}}}(\boldsymbol{\operatorname{curl}},\Omega)$-conforming discrete field;
we call this process \revision{broken patchwise} equilibration.
\item The upper bound~\eqref{eq_upper_bound} does not come from the Prager--Synge
inequality~\eqref{eq_PS} and is typically larger than those obtained from~\eqref{eq_PS}
with an equilibrated flux $\boldsymbol h_h^\star \in \boldsymbol{\mathcal N}_p(\mathcal T_h) \cap \boldsymbol H_{{\Gamma_{\rm N}}}(\boldsymbol{\operatorname{curl}},\Omega)$,
because of the presence of the multiplicative factors $\sqrt{6} C_{\rm L} C_{{\rm cont},e}$.
On the other hand, it is typically cheaper to compute the upper bound~\eqref{eq_upper_bound}
than those based on an equilibrated flux since
1)
the problems~\eqref{eq_definition_estimator} and~\eqref{eq_definition_estimator_sweep} involve
edge patches, whereas full equilibration would require solving problems also on vertex patches
which are larger than edge patches;
2)
the error estimators are computed in one stage only solving the
problems~\eqref{eq_definition_estimator_2} or~\eqref{eq_definition_estimator_sweep_2};
3)
\revision{
the broken patchwise equilibration procedure enables the construction of a $p$-robust error
estimator using polynomials of degree $p$, in contrast to the usual procedure requiring the
use of polynomials of degree $p+1$,
\cf~\cite{Brae_Pill_Sch_p_rob_09, Ern_Voh_p_rob_15, Ern_Voh_p_rob_3D_20};
the reason is that the usual procedure involves multiplication by the ``hat function''
$\psi_{\boldsymbol a}$ inside the estimators, which increases the polynomial degree by one,
whereas the current procedure only encapsulates an operation featuring $\boldsymbol \psi_e$
into the multiplicative constant $C_{{\rm cont},e}$, see~\eqref{eq_definition_Ccont}.}
\item The sequential sweep through the patch in~\eqref{eq_definition_estimator_sweep_2}
eliminates the patchwise problems~\eqref{eq_definition_estimator_2} and leads instead to
merely elementwise problems. These are much cheaper than~\eqref{eq_definition_estimator_2},
and, in particular, for $p=0$, \ie, for lowest-order N\'ed\'elec elements
in~\eqref{eq_maxwell_discrete} with one unknown per edge, \revision{they can be made} explicit.
\revision{Indeed, there is only one unknown in~\eqref{eq_definition_estimator_sweep_2} for each
tetrahedron $K \in {\TT^e}$ if $K$ is not the first or the last tetrahedron in the sweep. In the
last tetrahedron, there is no unknown left except if it contains a face that lies in ${\Gamma_{\rm D}^\edge}$,
in which case there is also only one unknown in~\eqref{eq_definition_estimator_sweep_2}. If the
first tetrahedron contains a face that lies in ${\Gamma_{\rm N}^\edge}$, there is again only one unknown
in~\eqref{eq_definition_estimator_sweep_2}. Finally, if the first tetrahedron does not contain
a face that lies in ${\Gamma_{\rm N}^\edge}$, it is possible, instead of $\mathcal F = \emptyset$, to consider the
set $\mathcal F$ formed by the face $F$ that either 1) lies in ${\Gamma_{\rm D}^\edge}$ (if any) or 2) is shared with
the last element and to employ for the boundary datum $\boldsymbol r_{\mathcal F}$
in~\eqref{eq_definition_estimator_sweep_2} the 1) value or 2) the mean value of the tangential
trace $\grad \times \boldsymbol A_h$ on $F$. This again leads to only one unknown
in~\eqref{eq_definition_estimator_sweep_2}, with all the theoretical properties maintained.}
\end{itemize}
\section{Numerical experiments}
\label{sec_num}
In this section, we present some numerical experiments to illustrate the a posteriori
error estimates from Theorem~\ref{theorem_aposteriori} and its use within an adaptive
mesh refinement procedure. We consider a test case with a smooth solution and a test
case with a solution featuring an edge singularity.
\revision{Below, we rely on the indicator $\eta_e$ evaluated
using~\eqref{eq_definition_estimator}, \ie, involving the edge-patch
solves~\eqref{broken_equil}. Moreover, we let
\begin{equation} \label{eq_ests}
\left (\eta_{\rm ofree}\right )^2 := 6 \sum_{e \in \mathcal E_h} \left (C_{{\rm cont},e} \eta_e\right )^2, \qquad
\left (\eta_{{\rm cofree}}\right )^2 := \sum_{e \in \mathcal E_h} \left (\eta_e\right )^2.
\end{equation}
Here, $\eta_{\rm ofree}$ corresponds to an ``oscillation-free'' error estimator, obtained
by discarding the oscillation terms $\operatorname{osc}_e$ in~\eqref{eq_upper_bound}, whereas
$\eta_{{\rm cofree}}$ corresponds to a ``constant-and-oscillation-free'' error estimator,
discarding in addition the multiplicative constants $\sqrt{6} C_{\rm L}$ and $C_{{\rm cont},e}$}.
\subsection{Smooth solution in the unit cube}
We first consider an example in the unit cube $\Omega := (0,1)^3$
and Neumann boundary conditions, ${\Gamma_{\rm N}} := \partial\Omega$ in~\eqref{eq_maxwell_strong}
and its weak form~\eqref{eq_maxwell_weak}. The analytical solution reads
\begin{equation*}
\boldsymbol A(\boldsymbol x) := \left (
\begin{array}{c}
\sin(\pi \boldsymbol x_1)\cos(\pi \boldsymbol x_2)\cos(\pi \boldsymbol x_3)
\\
-\cos(\pi \boldsymbol x_1)\sin(\pi \boldsymbol x_2)\cos(\pi \boldsymbol x_3)
\\
0
\end{array}
\right ).
\end{equation*}
One checks that $\div \boldsymbol A = 0$ and that
\begin{equation*}
\boldsymbol j(\boldsymbol x) := (\grad \times \grad \times \boldsymbol A)(\boldsymbol x) =
3\pi^2 \left (
\begin{array}{c}
\sin(\pi \boldsymbol x_1)\cos(\pi \boldsymbol x_2)\cos(\pi \boldsymbol x_3)
\\
-\cos(\pi \boldsymbol x_1)\sin(\pi \boldsymbol x_2)\cos(\pi \boldsymbol x_3)
\\
0
\end{array}
\right ).
\end{equation*}
We notice that $C_{\rm L} = 1$ since $\Omega$ is convex.
We first propose an ``$h$-convergence'' test case in which, for a fixed polynomial degree $p$,
we study the behavior of the N\'ed\'elec approximation $\boldsymbol A_h$
solving~\eqref{eq_maxwell_discrete} and of the error estimator of
Theorem~\ref{theorem_aposteriori}. We consider a sequence of meshes obtained by first
splitting the unit cube into an $N \times N \times N$ Cartesian grid and then splitting
each of the small cubes into six tetrahedra, with the resulting mesh size $h = \sqrt{3}/N$.
More precisely, each resulting edge patch is convex here, so that the constant $C_{{\rm P},e}$
in~\eqref{eq_definition_Ccont} can be taken as $1/\pi$ for all internal patches, see the
discussion in Section~\ref{sec_Poinc}. Figure~\ref{figure_unit_cube_hconv} presents the results.
The top-left panel shows that the expected convergence rates of $\boldsymbol A_h$ are obtained for
$p=0,\dots,3$. The top-right panel presents the local efficiency of the error estimator
based on the \revision{indicator} $\eta_e$ evaluated using~\eqref{eq_definition_estimator}.
We see that it is very good, the ratio of the patch indicator to the patch error being at
most $2$ for $p=0$, and close to $1$ for higher-order
polynomials. This seems to indicate that the constant $C_{{\rm st},e}$ in~\eqref{eq_stab} is rather small.
The bottom panels of Figure~\ref{figure_unit_cube_hconv} report on the global efficiency of the
error indicators \revision{$\eta_{\rm ofree}$ and $\eta_{{\rm cofree}}$ from~\eqref{eq_ests}.}
As shown in the bottom-right panel, the global
efficiency of $\eta_{{\rm cofree}}$ is independent of the mesh size.
The bottom-left panel shows a slight dependency of the global efficiency of
$\eta_{\rm ofree}$ on the mesh size, but this is only due to the fact that
Poincar\'e's constants differ for boundary and internal patches.
These two panels show that the efficiency actually slightly improves as the polynomial
degree is increased, highlighting the $p$-robustness of the proposed error estimator.
\revision{We also notice that the multiplicative factor $\sqrt{6}C_{{\rm cont},e}$ can lead to
some error overestimation.}
\input{figures/numerics/unit_cube/hconv.tex}
We then present a ``$p$-convergence'' test case where for a fixed mesh, we study
the behavior of the solution and of the error estimator when the polynomial degree
$p$ is increased. We provide this analysis for four different meshes. The three first meshes
are structured as previously described with $N=1,2$, and $4$, whereas the last
mesh is unstructured. The unstructured mesh has $358$ elements, $1774$ edges,
and $h = 0.37$. Figure~\ref{figure_unit_cube_pconv} presents the results.
The top-left panel shows an exponential convergence rate as $p$ is increased for all the
meshes, which is in agreement with the theory, since the solution is analytic.
The top-right panel shows that the local patch-by-patch
efficiency is very good, and seems to tend to $1$ as $p$ increases.
The bottom-right panel shows that the global efficiency of $\eta_{{\rm cofree}}$
also slightly improves as $p$ is increased, and it seems to be rather
independent of the mesh. The bottom-left panel shows that the global efficiency of
$\eta_{\rm ofree}$ is significantly worse on the unstructured mesh. This is because
in the absence of convex patches, we employ for $C_{{\rm P},e}$ the estimate
from~\cite{Vees_Verf_Poin_stars_12} instead of the constant $1/\pi$. We believe that this
performance could be improved by providing sharper Poincar\'e constants.
\input{figures/numerics/unit_cube/pconv.tex}
\subsection{Singular solution in an L-shaped domain}
We now turn our attention to an L-shaped domain featuring a singular solution.
Specifically, $\Omega := L \times (0,1)$, where
\begin{equation*}
L := \left \{
\boldsymbol x = (r\cos\theta,r\sin\theta) \ | \
|\boldsymbol x_1|,|\boldsymbol x_2| \leq 1 \quad 0 \leq \theta \leq 3\pi/2
\right \},
\end{equation*}
see Figure~\ref{figure_lshape_errors}, where $L$ is represented.
\revision{We consider the case ${\Gamma_{\rm D}} := \partial \Omega$, and t}he solution reads
$\boldsymbol A(\boldsymbol x) := \big(0,0,\chi(r) r^\alpha \sin(\alpha \theta)\big)^{\mathrm{T}}$,
where $\alpha := 3/2$, $r^2 := |\boldsymbol x_1|^2 + |\boldsymbol x_2|^2$, $(\boldsymbol x_1,\boldsymbol x_2) = r(\cos\theta,\sin\theta)$,
and $\chi:(0,1) \to \mathbb R$ is a smooth cutoff function such that $\chi = 0$ in a neighborhood of
$1$. We emphasize that $\div \boldsymbol A = 0$ and that, since
$\Delta \left (r^\alpha \sin(\alpha \theta)\right ) = 0$ near the origin, the right-hand side
$\boldsymbol j$ associated with $\boldsymbol A$ belongs to $\boldsymbol H(\operatorname{div},\Omega)$.
We use an adaptive mesh-refinement strategy based on D\"orfler's
marking~\cite{Dorf_cvg_FE_96}. The initial mesh we employ for $p=0$ and $p=1$
consists of $294$ elements and $1418$ edges with $h=0.57$, whereas a mesh with $23$ elements,
$86$ edges, and $h=2$ is employed for $p=2$ and $3$. The meshing package {\tt MMG3D} is employed to generate
the sequence of adapted meshes~\cite{Dobr_mmg3d}.
Figure~\ref{figure_lshape_conv} shows the convergence
histories of the adaptive algorithm for different values of $p$. In the top-left panel,
we observe the optimal convergence rate (limited to $N_{\rm dofs}^{2/3}$ for isotropic
elements in the presence of an edge singularity).
We employ the indicator $\eta_{{\rm cofree}}$ defined in~\eqref{eq_ests}.
The top-right and bottom-left panels respectively present the
\revision{local} and \revision{global} efficiency
indices. In both cases, the efficiency is good considering that the mesh is fully
unstructured with localized features. We also emphasize that the efficiency does not
deteriorate when $p$ increases.
Finally, Figure~\ref{figure_lshape_errors} depicts the estimated and the actual errors at the
last iteration of the adaptive algorithm. The face on the top of the domain $\Omega$ is
represented, and the colors are associated with the edges of the mesh. The left panels correspond
to the values of the estimator $\eta_e$ of~\eqref{eq_definition_estimator}, whereas the value
of $\|\grad \times(\boldsymbol A-\boldsymbol A_h)\|_{\omega_\edge}$ is represented in the right panels. Overall, this figure shows
excellent agreement between the estimated and actual error distribution.
\input{figures/numerics/lshape/conv.tex}
\begin{figure}
\includegraphics[width=\linewidth]{figures/numerics/lshape/P0.pdf}
$p=0$
\includegraphics[width=\linewidth]{figures/numerics/lshape/P1.pdf}
$p=1$
\includegraphics[width=\linewidth]{figures/numerics/lshape/P3.pdf}
$p=3$
\caption{Estimated error (left) and actual error (right) for the L-shaped domain experiment}
\label{figure_lshape_errors}
\end{figure}
\section{Proof of Theorem~\ref{theorem_aposteriori} ($p$-robust a posteriori error estimate)}
\label{sec_proof_a_post}
In this section, we prove Theorem~\ref{theorem_aposteriori}.
\subsection{Residuals}
Recall that $\boldsymbol A_h \in \boldsymbol V_{{\!h}} \subset \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega)$ solves
\eqref{eq_maxwell_discrete} and satisfies~\eqref{eq_maxwell_discrete_II}. In
view of the characterization of the weak solution~\eqref{eq_maxwell_weak_II},
we define the residual $\boldsymbol{\mathcal R} \in (\boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega))'$ by setting
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle
:=
(\boldsymbol j,\boldsymbol v)_{\Omega} - (\grad \times \boldsymbol A_h,\grad \times \boldsymbol v)_{\Omega}
=
(\grad \times (\boldsymbol A - \boldsymbol A_h),\grad \times \boldsymbol v)_{\Omega}
\quad
\forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega).
\end{equation*}
Taking $\boldsymbol v = \boldsymbol A - \boldsymbol A_h$ and using a duality characterization,
we have the error--residual link
\begin{equation}
\label{eq_res}
\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{\Omega}
=
\langle \boldsymbol{\mathcal R},\boldsymbol A-\boldsymbol A_h\rangle^{1/2}
=
\sup_{\substack{
\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega)\\
\|\grad \times \boldsymbol v\|_{\Omega} = 1}
} \langle \boldsymbol{\mathcal R},\boldsymbol v \rangle.
\end{equation}
We will also employ local dual norms of the residual $\boldsymbol{\mathcal R}$.
Specifically, for each edge $e \in \mathcal E_h$, we set
\begin{equation}
\label{eq_RR_ome}
\|\boldsymbol{\mathcal R}\|_{\star,e} :=
\sup_{\substack{
\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\\
\|\grad \times \boldsymbol v\|_{{\omega_\edge}} = 1}
}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle.
\end{equation}
For each $e \in \mathcal E_h$, we will also need an oscillation-free
residual $\boldsymbol{\mathcal R}_h^e \in \left (\boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\right )'$
defined using the projected right-hand side introduced in~\eqref{eq_definition_jj_h},
\begin{equation*}
\langle \boldsymbol{\mathcal R}_h^e,\boldsymbol v \rangle
:=
(\boldsymbol j_h^e,\boldsymbol v)_{{\omega_\edge}} - (\grad \times \boldsymbol A_h,\grad \times \boldsymbol v)_{{\omega_\edge}} \quad
\forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge}).
\end{equation*}
We also employ the notation
\begin{equation*}
\|\boldsymbol{\mathcal R}^e_h\|_{\star,e} :=
\sup_{\substack{
\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\\
\|\grad \times \boldsymbol v\|_{{\omega_\edge}} = 1}}
\langle \boldsymbol{\mathcal R}_h^e,\boldsymbol v \rangle
\end{equation*}
for the dual norm of $\boldsymbol{\mathcal R}_h^e$. Note that $\boldsymbol{\mathcal R}^e_h = \boldsymbol{\mathcal R}_{|\boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})}$ whenever the source term $\boldsymbol j$
is a piecewise $\boldsymbol{\mathcal{RT}}_p(\mathcal T_h)$ polynomial.
\subsection{Data oscillation}
Recalling the definition~\eqref{eq_definition_osc} of $\operatorname{osc}_e$, we
have the following comparison:
\begin{lemma}[Data oscillation]
The following holds true:
\begin{subequations}\begin{equation}
\label{eq_data_oscillations_lower_bound}
\|\boldsymbol{\mathcal R}_h^e\|_{\star,e} \leq \|\boldsymbol{\mathcal R}\|_{\star,e} + \operatorname{osc}_e
\end{equation}
and
\begin{equation}
\label{eq_data_oscillations_upper_bound}
\|\boldsymbol{\mathcal R}\|_{\star,e} \leq \|\boldsymbol{\mathcal R}_h^e\|_{\star,e} + \operatorname{osc}_e.
\end{equation}\end{subequations}
\end{lemma}
\begin{proof}
Let $\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ with $\|\grad \times \boldsymbol v\|_{{\omega_\edge}} = 1$ be fixed.
We have
\begin{equation*}
\langle \boldsymbol{\mathcal R}_h^e,\boldsymbol v \rangle
=
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle - (\boldsymbol j-\boldsymbol j_h^e,\boldsymbol v)_{{\omega_\edge}}.
\end{equation*}
We define $q$ as the unique element of $H^1_{{\Gamma_{\rm D}^\edge}}({\omega_\edge})$ such that
\begin{equation*}
(\boldsymbol \nabla q,\boldsymbol \nabla w) = (\boldsymbol v,\boldsymbol \nabla w) \quad \forall w \in H^1_{{\Gamma_{\rm D}^\edge}}({\omega_\edge}),
\end{equation*}
and set $\widetilde \boldsymbol v := \boldsymbol v - \boldsymbol \nabla q$.
Since $\div \boldsymbol j = \div \boldsymbol j_h^e = 0$ and $\boldsymbol j-\boldsymbol j_h^e \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$,
we have $(\boldsymbol j-\boldsymbol j_h^e,\boldsymbol \nabla q)_{{\omega_\edge}} = 0$, and it follows that
\begin{equation*}
\langle \boldsymbol{\mathcal R}_h^e,\boldsymbol v \rangle
=
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle - (\boldsymbol j-\boldsymbol j_h^e,\widetilde \boldsymbol v)_{{\omega_\edge}}
\leq
\|\boldsymbol{\mathcal R}\|_{\star,e} + \|\boldsymbol j - \boldsymbol j_h^e\|_{{\omega_\edge}}\|\widetilde \boldsymbol v\|_{{\omega_\edge}}.
\end{equation*}
Since $\widetilde{\vv} \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge}) \cap
\boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$
with $\div \widetilde{\vv} = 0$ in ${\omega_\edge}$, recalling~\eqref{eq_local_poincare_vectorial}, we have
\begin{equation*}
\|\widetilde \boldsymbol v\|_{{\omega_\edge}}
\leq
C_{{\rm \revision{PFW}},e} h_{\omega_\edge} \|\grad \times \widetilde \boldsymbol v\|_{{\omega_\edge}}
=
C_{{\rm \revision{PFW}},e} h_{\omega_\edge} \|\grad \times \boldsymbol v\|_{{\omega_\edge}}
=
C_{{\rm \revision{PFW}},e} h_{\omega_\edge},
\end{equation*}
and we obtain~\eqref{eq_data_oscillations_lower_bound} by taking the supremum over all $\boldsymbol v$.
The proof of~\eqref{eq_data_oscillations_upper_bound} follows exactly the same path.
\end{proof}
\subsection{Partition of unity and cut-off estimates}
We now analyze a partition of unity for vector-valued functions that we later employ to localize
the error onto edge patches. Recalling the notation ${\boldsymbol \tau}_{e}$ for the
unit tangent vector orienting $e$, we quote the following classical partition
of unity~\cite[Chapter~15]{Ern_Guermond_FEs_I_21}:
\begin{lemma}[Vectorial partition of unity] \label{lem_PU}
Let $\mathbb I$ be the identity matrix in ${\mathbb R}^{3\times 3}$.
The edge basis functions $\boldsymbol \psi_e$ from~\eqref{eq_BF} satisfy
\[
\sum_{e \in \mathcal E_h}{\boldsymbol \tau}_e \otimes \boldsymbol \psi_e =
\sum_{e \in \mathcal E_h}\boldsymbol \psi_e \otimes {\boldsymbol \tau}_e = {\mathbb I},
\]
where $\otimes$ denotes the outer product, so that we have
\begin{equation}
\label{eq_partition_unity}
\boldsymbol w = \sum_{e \in \mathcal E_h} (\boldsymbol w \cdot {\boldsymbol \tau}_e )|_{\omega_\edge} \boldsymbol \psi_e
\qquad \forall \boldsymbol w \in \boldsymbol L^2(\Omega).
\end{equation}
\end{lemma}
\begin{lemma}[Cut-off stability] \label{lem_c_stab}
For every edge $e \in \mathcal E_h$, recalling the space $H^1_\star({\omega_\edge})$ defined in~\eqref{eq_Hse} and the constant $C_{{\rm cont},e}$ defined in~\eqref{eq_definition_Ccont}, we have
\begin{equation}
\label{eq_estimate_cont}
\|\grad \times(v \boldsymbol \psi_e)\|_{{\omega_\edge}}
\leq
C_{{\rm cont},e} \|\boldsymbol \nabla v\|_{\omega_\edge} \qquad \forall v \in H^1_\star({\omega_\edge}).
\end{equation}
Moreover, the upper bound~\eqref{eq_bound_Ccont} on $C_{{\rm cont},e}$ holds true.
\end{lemma}
\begin{proof}
Let an edge $e \in \mathcal E_h$ and $v \in H^1_\star({\omega_\edge})$ be fixed.
Since $\grad \times (v \boldsymbol \psi_e) = v \grad \times \boldsymbol \psi_e + \boldsymbol \nabla v \times \boldsymbol \psi_e$,
we have, using~\eqref{eq_local_poincare},
\begin{align*}
\|\grad \times (v \boldsymbol \psi_e)\|_{{\omega_\edge}}
&\leq
\|\grad \times \boldsymbol \psi_e\|_{\infty,{\omega_\edge}}\|v\|_{\omega_\edge} +
\|\boldsymbol \psi_e\|_{\infty,{\omega_\edge}}\|\boldsymbol \nabla v\|_{\omega_\edge}
\\
&\leq
(\|\grad \times \boldsymbol \psi_e\|_{\infty,{\omega_\edge}}C_{{\rm P},e} h_{\omega_\edge} +
\|\boldsymbol \psi_e\|_{\infty,{\omega_\edge}})\|\boldsymbol \nabla v\|_{\omega_\edge}
\\
&=
C_{{\rm cont},e} \|\boldsymbol \nabla v\|_{\omega_\edge}.
\end{align*}
This proves~\eqref{eq_estimate_cont}. To prove~\eqref{eq_bound_Ccont},
we remark that in every tetrahedron $K \in {\TT^e}$, we have
(see for instance~\cite[Section~5.5.1]{Monk_FEs_Maxwell_03}, \cite[Chapter~15]{Ern_Guermond_FEs_I_21})
\begin{equation*}
\boldsymbol \psi_e|_K = |e| (\lambda_1 \boldsymbol \nabla \lambda_2 - \lambda_2 \boldsymbol \nabla \lambda_1),
\quad
(\grad \times \boldsymbol \psi_e)|_K = 2 |e|\boldsymbol \nabla \lambda_1 \times \boldsymbol \nabla \lambda_2,
\end{equation*}
where $\lambda_1$ and $\lambda_2$ are the barycentric coordinates of $K$
associated with the two endpoints of $e$ such that ${\boldsymbol \tau}_e$
points from the first to the second vertex.
Since $\|\lambda_j\|_{\infty,K} = 1$ and $\|\boldsymbol \nabla \lambda_j\|_{\infty,K} \leq \rho_K^{-1}$,
we have
\begin{equation*}
\|\boldsymbol \psi_e\|_{\infty,K} \leq \frac{2}{\rho_K}|e|, \quad
\|\grad \times \boldsymbol \psi_e\|_{\infty,K} \leq \frac{2}{\rho_K^2}|e|
\end{equation*}
for every $K \in {\TT^e}$.
Recalling the definition~\eqref{eq_patch_not} of $\rho_e$,
which implies that $\rho_e \leq \rho_K$, as well as the definition $\kappa_e$
in~\eqref{eq_regularities}, we conclude that
\begin{align*}
C_{{\rm cont},e} = \|\boldsymbol \psi_e\|_{\infty,{\omega_\edge}} + C_{{\rm P},e} h_{\omega_\edge} \|\grad \times \boldsymbol \psi_e\|_{\infty,{\omega_\edge}}
\leq \frac{2 |e|}{\rho_e}\left(1 + C_{{\rm P},e} \frac{h_{\omega_\edge}}{\rho_e} \right) = C_{\kappa_e}.
\end{align*}
\end{proof}
\subsection{Upper bound using localized residual dual norms}
We now establish an upper bound on the error using the localized residual dual norms
$\|\boldsymbol{\mathcal R}_h^e\|_{\star,e}$, in the spirit of~\cite{Blech_Mal_Voh_loc_res_NL_20},
\cite[Chapter~34]{Ern_Guermond_FEs_II_21}, and the references therein.
\begin{proposition}[Upper bound by localized residual dual norms]
Let $C_{{\rm cont},e}$ and $C_{\rm L}$ be defined in~\eqref{eq_definition_Ccont}
and~\eqref{eq_estimate_lift}, respectively. Then the following holds:
\begin{equation}
\label{eq_upper_bound_residual}
\|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{\Omega}
\leq
\sqrt{6} C_{\rm L} \left (
\sum_{e \in \mathcal E_h} C_{{\rm cont},e}^2 \left (\|\boldsymbol{\mathcal R}_h^e\|_{\star,e} + \operatorname{osc}_e\right )^2
\right )^{1/2}.
\end{equation}
\end{proposition}
\begin{proof}
We start with~\eqref{eq_res}. Let $\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega)$
with $\|\grad \times \boldsymbol v\|_{\Omega} = 1$ be fixed. Following~\eqref{eq_estimate_lift},
we define $\boldsymbol w \in \boldsymbol H^1(\Omega) \cap \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega)$
such that $\grad \times \boldsymbol w = \grad \times \boldsymbol v$ with
\begin{equation}
\label{tmp_estimate_lift}
\|\boldsymbol \nabla \boldsymbol w\|_{\Omega} \leq C_{\rm L} \|\grad \times \boldsymbol v\|_{\Omega}.
\end{equation}
As a consequence of~\eqref{eq_maxwell_weak_II} and~\eqref{eq_maxwell_discrete_II},
the residual $\boldsymbol{\mathcal R}$ is (in particular) orthogonal to
$\boldsymbol{\mathcal N}_0(\mathcal T_h) \cap \boldsymbol H_{{\Gamma_{\rm D}}}(\boldsymbol{\operatorname{curl}},\Omega)$. Thus,
by employing the partition of unity~\eqref{eq_partition_unity}, we have
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle
=
\langle \boldsymbol{\mathcal R},\boldsymbol w \rangle
=
\sum_{e \in \mathcal E_h} \langle \boldsymbol{\mathcal R}, (\boldsymbol w \cdot {\boldsymbol \tau}_e )|_{\omega_\edge} \boldsymbol \psi_e \rangle
=
\sum_{e \in \mathcal E_h}
\langle \boldsymbol{\mathcal R}, (\boldsymbol w \cdot {\boldsymbol \tau}_e - \overline{w_e})|_{\omega_\edge} \boldsymbol \psi_e \rangle,
\end{equation*}
where $\overline{w_e} := 0$ if $e \in \EE^{\rm D}_h$ and
\begin{equation*}
\overline{w_e} := \frac{1}{|{\omega_\edge}|} \int_{\omega_\edge} \boldsymbol w \cdot {\boldsymbol \tau}_e
\end{equation*}
otherwise. Since
$(\boldsymbol w \cdot {\boldsymbol \tau}_e - \overline{w_e}) \boldsymbol \psi_e \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$
for all $e\in \mathcal E_h$, we have from~\eqref{eq_RR_ome}
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle
\leq
\sum_{e \in \mathcal E_h} \|\boldsymbol{\mathcal R}\|_{\star,e} \|\grad \times \left (
\left (\boldsymbol w \cdot {\boldsymbol \tau}_e - \overline{w_e}
\right ) \boldsymbol \psi_e \right )\|_{{\omega_\edge}}.
\end{equation*}
We observe that $\boldsymbol w \cdot {\boldsymbol \tau}_e - \overline{w_e} \in H^1_\star({\omega_\edge})$
for all $e \in \mathcal E_h$ and that
\begin{equation*}
\|\boldsymbol \nabla(\boldsymbol w \cdot {\boldsymbol \tau}_e - \overline{w_e})\|_{{\omega_\edge}}
=
\|\boldsymbol \nabla (\boldsymbol w \cdot {\boldsymbol \tau}_e)\|_{{\omega_\edge}}
\leq
\|\boldsymbol \nabla\boldsymbol w\|_{{\omega_\edge}}.
\end{equation*}
As a result, \eqref{eq_estimate_cont} shows that
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle
\leq
\sum_{e \in \mathcal E_h} C_{{\rm cont},e} \|\boldsymbol{\mathcal R}\|_{\star,e}\|\boldsymbol \nabla \boldsymbol w\|_{{\omega_\edge}}
\leq
\left (\sum_{e \in \mathcal E_h} C_{{\rm cont},e}^2 \|\boldsymbol{\mathcal R}\|_{\star,e}^2\right )^{1/2}
\left (\sum_{e \in \mathcal E_h} \|\boldsymbol \nabla \boldsymbol w\|_{{\omega_\edge}}^2\right )^{1/2}.
\end{equation*}
At this point, as each tetrahedron $K \in \mathcal T_h$ has $6$ edges, we have
\begin{equation*}
\sum_{e \in \mathcal E_h} \|\boldsymbol \nabla \boldsymbol w\|_{{\omega_\edge}}^2
\revision{=}
6 \|\boldsymbol \nabla \boldsymbol w\|_{\Omega}^2,
\end{equation*}
and using~\eqref{tmp_estimate_lift}, we infer that
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle\leq
\sqrt{6} C_{\rm L}
\left (\sum_{e \in \mathcal E_h} C_{{\rm cont},e}^2\|\boldsymbol{\mathcal R}\|_{\star,e}^2\right )^{1/2}
\|\grad \times \boldsymbol v\|_{\Omega}.
\end{equation*}
Then, we conclude with~\eqref{eq_data_oscillations_upper_bound}.
\end{proof}
\subsection{Lower bound using localized residual dual norms}
We now consider the derivation of local lower bounds on the error using
the residual dual norms. We first establish a result for the residual $\boldsymbol{\mathcal R}$.
\begin{lemma}[Local residual]
For every edge $e \in \mathcal E_h$, the following holds:
\begin{equation}
\label{eq_minimization}
\|\boldsymbol{\mathcal R}\|_{\star,e} =
\min_{\substack{
\boldsymbol h \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\\
\grad \times \boldsymbol h = \boldsymbol j}}
\|\boldsymbol h - \grad \times \boldsymbol A_h\|_{{\omega_\edge}},
\end{equation}
as well as
\begin{equation}
\label{eq_lower_bound_residual}
\|\boldsymbol{\mathcal R}\|_{\star,e} \leq \|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{{\omega_\edge}}.
\end{equation}
\end{lemma}
\begin{proof}
Let us define $\boldsymbol h^\star$ as the unique element of $\boldsymbol L^2({\omega_\edge})$ such that
\begin{equation}
\label{tmp_definition_minimizer}
\left \{
\begin{array}{rcll}
\div \boldsymbol h^\star &=& 0 & \text{ in } {\omega_\edge},
\\
\grad \times \boldsymbol h^\star &=& \boldsymbol j & \text{ in } {\omega_\edge},
\\
\boldsymbol h^\star \cdot \boldsymbol n_{\omega_e} &=& \grad \times \boldsymbol A_h \cdot \boldsymbol n_{\omega_e} & \text{ on } {\Gamma_{\rm D}^\edge},
\\
\boldsymbol h^\star \times \boldsymbol n_{\omega_e} &=& \boldsymbol 0 & \text{ on } {\Gamma_{\rm N}^\edge}.
\end{array}
\right .
\end{equation}
The existence and uniqueness of $\boldsymbol h^\star$ follows from
\cite[Proposition 7.4]{Fer_Gil_Maxw_BC_97} after lifting by $\grad \times \boldsymbol A_h$.
The second and fourth equations in~\eqref{tmp_definition_minimizer}
imply that $\boldsymbol h^\star$ belongs to the minimization set of~\eqref{eq_minimization}.
If $\boldsymbol h' \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ with $\grad \times \boldsymbol h' = \boldsymbol j$ is another element of the
minimization set, then $\boldsymbol h^\star - \boldsymbol h' = \boldsymbol \nabla q$ for some
$q \in H^1_{{\Gamma_{\rm N}^\edge}}({\omega_\edge})$, and we see that
\begin{align*}
\|\boldsymbol h' - \grad \times \boldsymbol A_h\|_{\omega_\edge}^2
&=
\|\boldsymbol h^\star - \grad \times \boldsymbol A_h - \boldsymbol \nabla q\|_{\omega_\edge}^2
\\
&=
\|\boldsymbol h^\star - \grad \times \boldsymbol A_h\|_{\omega_\edge}^2 - 2(\boldsymbol h^\star - \grad \times \boldsymbol A_h,\boldsymbol \nabla q)_{\omega_\edge} + \|\boldsymbol \nabla q\|_{\omega_\edge}^2
\\
&=
\|\boldsymbol h^\star - \grad \times \boldsymbol A_h\|_{\omega_\edge}^2 + \|\boldsymbol \nabla q\|_{\omega_\edge}^2
\\
&\geq
\|\boldsymbol h^\star - \grad \times \boldsymbol A_h\|_{\omega_\edge}^2,
\end{align*}
where we used that $\boldsymbol h^\star$ is divergence-free,
$(\boldsymbol h^\star - \grad \times \boldsymbol A_h) \cdot \boldsymbol n_{\omega_e} = \boldsymbol 0$
on ${\Gamma_{\rm D}^\edge}$, and $q=0$ on ${\Gamma_{\rm N}^\edge}$ to infer that $(\boldsymbol h^\star - \grad \times \boldsymbol A_h,\boldsymbol \nabla q)_{\omega_\edge}=0$.
Hence, $\boldsymbol h^\star$ is a minimizer of~\eqref{eq_minimization}.
Let $\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$. Since $(\grad \times \boldsymbol h^\star,\boldsymbol v)_{\omega_\edge}
= (\boldsymbol h^\star,\grad \times \boldsymbol v)_{\omega_\edge}$, we have
\begin{align*}
\langle \boldsymbol{\mathcal R},\boldsymbol v \rangle
&=
(\boldsymbol j,\boldsymbol v)_{\omega_\edge} - (\grad \times \boldsymbol A_h,\grad \times \boldsymbol v)_{\omega_\edge}
\\
&=
(\grad \times \boldsymbol h^\star,\boldsymbol v)_{\omega_\edge} - (\grad \times \boldsymbol A_h,\grad \times \boldsymbol v)_{\omega_\edge}
\\
&=
(\boldsymbol h^\star,\grad \times \boldsymbol v)_{\omega_\edge} - (\grad \times \boldsymbol A_h,\grad \times \boldsymbol v)_{\omega_\edge}
\\
&=
(\boldsymbol \phi,\grad \times \boldsymbol v)_{\omega_\edge},
\end{align*}
where we have set $\boldsymbol \phi := \boldsymbol h^\star - \grad \times \boldsymbol A_h$. As above,
$\div \boldsymbol \phi = 0$ in ${\omega_\edge}$ and $\boldsymbol \phi \cdot \boldsymbol n_{\omega_e} = 0$ on ${\Gamma_{\rm D}^\edge}$.
Therefore, Theorem 8.1 of~\cite{Fer_Gil_Maxw_BC_97}
shows that $\boldsymbol \phi = \grad \times \boldsymbol \omega$ for some $\boldsymbol \omega \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$, and
\begin{equation*}
\langle \boldsymbol{\mathcal R},\boldsymbol v\rangle = (\grad \times \boldsymbol \omega,\grad \times \boldsymbol v)_{\omega_\edge} \quad \forall \boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge}).
\end{equation*}
At this point, it is clear that
\begin{equation*}
\|\boldsymbol{\mathcal R}\|_{\star,e} =
\sup_{\substack{
\boldsymbol v \in \boldsymbol H_{{\Gamma_{\rm D}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\\
\|\grad \times \boldsymbol v\|_{{\omega_\edge}} = 1}}
(\grad \times \boldsymbol \omega,\grad \times \boldsymbol v)_{\omega_\edge}
=
\|\grad \times \boldsymbol \omega\|_{\omega_\edge}
=
\|\boldsymbol h^\star - \grad \times \boldsymbol A_h\|_{\omega_\edge}.
\end{equation*}
Finally, we obtain~\eqref{eq_lower_bound_residual} by observing that
$\widetilde \boldsymbol h := (\grad \times \boldsymbol A )|_{\omega_\edge}$ is in the minimization set of~\eqref{eq_minimization}.
\end{proof}
We are now ready to state our results for the oscillation-free residuals $\boldsymbol{\mathcal R}_h^e$.
\begin{lemma}[Local oscillation-free residual] \label{lem_res}
For every edge $e \in \mathcal E_h$, the following holds:
\begin{equation}
\label{eq_minimization_jh}
\|\boldsymbol{\mathcal R}_h^e\|_{\star,e} =
\min_{\substack{
\boldsymbol h \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})\\
\grad \times \boldsymbol h = \boldsymbol j_h^e}}
\|\boldsymbol h - \grad \times \boldsymbol A_h\|_{{\omega_\edge}},
\end{equation}
as well as
\begin{equation}
\label{eq_lower_bound_residual_osc}
\|\boldsymbol{\mathcal R}_h^e\|_{\star,e} \leq \|\grad \times(\boldsymbol A - \boldsymbol A_h)\|_{{\omega_\edge}} + \operatorname{osc}_e.
\end{equation}
\end{lemma}
\begin{proof}
We establish~\eqref{eq_minimization_jh} by following the same path as for~\eqref{eq_minimization}.
On the other hand, we simply obtain~\eqref{eq_lower_bound_residual_osc} as a consequence
of~\eqref{eq_data_oscillations_lower_bound} and~\eqref{eq_lower_bound_residual}.
\end{proof}
\subsection{Proof of Theorem~\ref{theorem_aposteriori}}
We are now ready to give a proof of Theorem~\ref{theorem_aposteriori}.
On the one hand, the \revision{broken patchwise} equilibration estimator $\eta_e$ defined
in~\eqref{eq_definition_estimator} is evaluated from a field
$\boldsymbol h_h^{e,\star} \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$
such that $\grad \times \boldsymbol h_h^{e,\star} = \boldsymbol j_h^e$, and the sequential
sweep~\eqref{eq_definition_estimator_sweep} produces $\boldsymbol h_h^{e,\heartsuit}$
also satisfying these two properties. Since the minimization set in~\eqref{eq_minimization_jh}
is larger, it is clear that
\begin{equation*}
\|\boldsymbol{\mathcal R}_h^e\|_{\star,e}
\leq
\eta_e
\end{equation*}
for both estimators $\eta_e$.
Then, \eqref{eq_upper_bound} immediately follows from~\eqref{eq_upper_bound_residual}.
On the other hand, Theorem~\ref{theorem_stability} with the choice $\boldsymbol \chi_h := (\grad \times \boldsymbol A_h)|_{\omega_\edge}$
and the polynomial degree $p$ together with~\eqref{eq_minimization_jh} of Lemma~\ref{lem_res}
implies that
\begin{equation*}
\eta_e \leq C_{{\rm st},e} \|\boldsymbol{\mathcal R}_h^e\|_{\star,e}
\end{equation*}
for the estimator~\eqref{eq_definition_estimator}, whereas \revision{the same result for} $\boldsymbol h_h^{e,\heartsuit}$
from~\eqref{eq_definition_estimator_sweep} \revision{follows from Theorem~\ref{thm_sweep} with again $\boldsymbol \chi_h := (\grad \times \boldsymbol A_h)|_{\omega_\edge}$.}
Therefore, \eqref{eq_lower_bound} is a direct consequence of~\eqref{eq_lower_bound_residual_osc}.
\section{Equivalent reformulation and proof of Theorem~\ref{theorem_stability}
($\boldsymbol H(\boldsymbol{\operatorname{curl}})$ best-approximation in an edge patch)}
\label{sec_proof_stability}
In this section, we consider the minimization problem over an edge patch as posed
in the statement of Theorem~\ref{theorem_stability}\revision{, as well as its sweep
variant of Theorem~\ref{thm_sweep}}, which were central tools to establish
the efficiency of the \revision{broken patchwise equilibrated} error estimators in
Theorem~\ref{theorem_aposteriori}. These minimization problems are similar to the ones considered
in~\cite{Brae_Pill_Sch_p_rob_09, Ern_Voh_p_rob_15, Ern_Voh_p_rob_3D_20} in the framework of
$H^1$ and $\boldsymbol H(\operatorname{div})$ spaces. We prove here Theorem~\ref{theorem_stability} via its equivalence
with a stable broken $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ polynomial extension on an edge patch, as formulated in
Proposition~\ref{prop_stability_patch} below. \revision{By virtue of Remark~\ref{rem_sweep_brok},
this also establishes the validity of Theorem~\ref{thm_sweep}}.
\subsection{Stability of discrete minimization in a tetrahedron}
\subsubsection{Preliminaries}
\label{sec:preliminaries_K}
We first recall some necessary notation from~\cite{Chaum_Ern_Voh_curl_elm_20}.
Consider an arbitrary mesh face $F\in \mathcal F_h$ oriented
by the fixed unit vector normal $\boldsymbol n_F$.
For all $\boldsymbol w \in \boldsymbol L^2(F)$, we define the tangential component of $\boldsymbol w$ as
\begin{equation} \label{eq:def_pi_tau_F}
\boldsymbol \pi^{\boldsymbol \tau}_F (\boldsymbol w) := \boldsymbol w - (\boldsymbol w \cdot \boldsymbol n_F) \boldsymbol n_F.
\end{equation}
Note that the orientation of $\boldsymbol n_F$ is not important here.
Let $K\in\mathcal T_h$ and let $\mathcal F_K$ be the collection of the faces of $K$.
For all $\boldsymbol v \in \boldsymbol H^1(K)$ and all $F\in\mathcal F_K$, the tangential trace of $\boldsymbol v$ on $F$
is defined (with a slight abuse of notation) as $\boldsymbol \pi^{\boldsymbol \tau}_F(\boldsymbol v) := \boldsymbol \pi^{\boldsymbol \tau}_F (\boldsymbol v|_F)$.
Consider now a nonempty subset $\mathcal F \subseteq \mathcal F_K$. We denote $\Gamma_\mathcal F \subset \partial K$ the
corresponding part of the boundary of $K$. Let $p\ge0$ be the polynomial degree
and recall that $\boldsymbol{\mathcal N}_p(K)$ is the N\'ed\'elec space on the tetrahedron $K$,
see~\eqref{eq_RT_N}. We define the piecewise polynomial space on $\Gamma_\mathcal F$
\begin{equation}
\label{eq_tr_K}
\boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F) :=
\left \{
\boldsymbol w_\mathcal F \in \boldsymbol L^2(\Gamma_\mathcal F) \; | \;
\exists \boldsymbol v_p \in \boldsymbol{\mathcal N}_p(K);
\boldsymbol w_F := (\boldsymbol w_\mathcal F)|_F = \boldsymbol \pi^{\boldsymbol \tau}_F (\boldsymbol v_p) \quad \forall F \in \mathcal F
\right \}.
\end{equation}
Note that $\boldsymbol w_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$ if and only if
$\boldsymbol w_F\in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\{F\}})$ for all $F\in\mathcal F$ and whenever $|\mathcal F|\ge2$,
for every pair $(F_-,F_+)$ of distinct faces in $\mathcal F$, the following tangential trace
compatibility condition holds true along their common edge $e := F_+\cap F_-$:
\begin{equation}
\label{eq_edge_compatibility_condition}
(\boldsymbol w_{F_+})|_e \cdot {\boldsymbol \tau}_e =
(\boldsymbol w_{F_-})|_e \cdot {\boldsymbol \tau}_e.
\end{equation}
For all $\boldsymbol w_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$, we set
\begin{equation}
\label{eq_scurl_curl_el}
\operatorname{curl}_F (\boldsymbol w_F) := (\grad \times \boldsymbol v_p)|_F \cdot \boldsymbol n_F \qquad \forall F \in \mathcal F,
\end{equation}
which is well-defined independently of the choice of $\boldsymbol v_p$. Note that the orientation of $\boldsymbol n_F$
is relevant here.
The definition~\eqref{eq:def_pi_tau_F} of the tangential trace cannot be applied to fields
with the minimal regularity $\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K)$. In what follows, we use the
following notion to prescribe the tangential trace of a field in $\boldsymbol H(\boldsymbol{\operatorname{curl}},K)$.
\begin{definition}[Tangential trace by integration by parts in a single tetrahedron]
\label{definition_partial_trace}
Let $K$ be a tetrahedron and
$\mathcal F \subseteq \mathcal F_K$ a nonempty (sub)set of its faces.
Given $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$ and $\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K)$, we
employ the notation ``$\boldsymbol v|^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F$'' to say that
\begin{equation*}
(\grad \times \boldsymbol v,\boldsymbol \phi)_K - (\boldsymbol v,\grad \times \boldsymbol \phi)_K = \sum_{F \in \mathcal F} (\boldsymbol r_F,\boldsymbol \phi \times \boldsymbol n_K)_F
\quad
\forall \boldsymbol \phi \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}(K),
\end{equation*}
where
\begin{equation*}
\boldsymbol H_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}^1(K) :=
\left \{
\boldsymbol \phi \in \boldsymbol H^1(K) \; | \; \boldsymbol \phi|_F \times \boldsymbol n_{\revision{K}} = \boldsymbol 0
\quad
\forall F \in \mathcal F^{\mathrm{c}} := \mathcal F_K \setminus \mathcal F
\right \}.
\end{equation*}
Whenever $\boldsymbol v \in \boldsymbol H^1(K)$, $\boldsymbol v|^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F$ if and only if
$\boldsymbol \pi^{\boldsymbol \tau}_F (\boldsymbol v) = \boldsymbol r_F$ for all $F \in \mathcal F$.
\end{definition}
\subsubsection{Statement of the stability result in a tetrahedron}
Recall the Raviart--Thomas space $\boldsymbol{\mathcal{RT}}_p(K)$ on the simplex $K$, see~\eqref{eq_RT_N}.
We are now ready to state a key technical tool
from~\cite[Theorem~2]{Chaum_Ern_Voh_curl_elm_20}, based
on~\cite[Theorem~7.2]{Demk_Gop_Sch_ext_II_09}
and~\cite[Proposition~4.2]{Cost_McInt_Bog_Poinc_10}.
\begin{proposition}[Stable $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ polynomial extension on a tetrahedron]
\label{prop_stability_tetrahedra}
Let $K$ be a tetrahedron and let
$\emptyset\subseteq \mathcal F \subseteq \mathcal F_K$ be a (sub)set of its faces.
Then, for every polynomial degree $p \geq 0$, for all $\boldsymbol r_K \in \boldsymbol{\mathcal{RT}}_p(K)$ such that
$\div \boldsymbol r_K = 0$, and if $\mathcal F \ne \emptyset$, for all $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$
such that $\boldsymbol r_K \cdot \boldsymbol n_{F} = \operatorname{curl}_F (\boldsymbol r_F)$ for all $F \in \mathcal F$, the following holds:
\begin{equation}
\label{eq_minimization_element_K}
\min_{\substack{
\boldsymbol v_p \in \boldsymbol{\mathcal N}_p(K)
\\
\grad \times \boldsymbol v_p = \boldsymbol r_K
\\
\boldsymbol v_{p}|^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F
}}
\|\boldsymbol v_p\|_{K}
\le C_{\mathrm{st},K}
\min_{\substack{
\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K)
\\
\grad \times \boldsymbol v = \boldsymbol r_K
\\
\boldsymbol v|^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F
}}
\|\boldsymbol v\|_{K},
\end{equation}
where the condition on the tangential trace in the minimizing sets is null if $\emptyset=\mathcal F$.
Both minimizers in~\eqref{eq_minimization_element_K} are uniquely defined and the constant
$C_{\mathrm{st},K}$ only depends on the shape-regularity parameter $\kappa_K$ of $K$.
\end{proposition}
\subsection{Piola mappings}
\label{sec_Piola}
This short section reviews some useful properties of Piola mappings used below,
see~\cite[Chapter~9]{Ern_Guermond_FEs_I_21}. Consider two tetrahedra
$\Kin,K_{\rm out} \subset \mathbb R^3$ and an invertible affine mapping
$\boldsymbol T: \mathbb R^3 \to \mathbb R^3$ such that $K_{\rm out} = \boldsymbol T(\Kin)$.
Let $\mathbb J_{\boldsymbol T}$ be the (constant) Jacobian matrix of $\boldsymbol T$.
Note that we do not require that $\det \mathbb J_{\boldsymbol T}$ is positive. The affine mapping $\boldsymbol T$
can be identified by specifying the image of each vertex of $\Kin$. We consider the
covariant and contravariant Piola mappings
\begin{equation*}
\boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T}(\boldsymbol v) = \left (\mathbb J_{\boldsymbol T}\right )^{-T}
\left (\boldsymbol v \circ \boldsymbol T^{-1} \right ),
\qquad
\boldsymbol \psi^{\mathrm{d}}_{\boldsymbol T}(\boldsymbol v) = \frac{1}{\det \left (\mathbb J_{\boldsymbol T}\right )} \mathbb J_{\boldsymbol T}
(\boldsymbol v \circ \boldsymbol T^{-1})
\end{equation*}
for vector-valued fields
$\boldsymbol v: \Kin \to \mathbb R^3$. It is well-known that $\boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T}$
maps $\boldsymbol H(\boldsymbol{\operatorname{curl}},\Kin)$ onto $\boldsymbol H(\boldsymbol{\operatorname{curl}},K_{\rm out})$ and it maps
$\boldsymbol{\mathcal N}_p(\Kin)$ onto $\boldsymbol{\mathcal N}_p(K_{\rm out})$ for any polynomial degree $p\ge0$.
Similarly, $\boldsymbol \psi^{\mathrm{d}}_{\boldsymbol T}$
maps $\boldsymbol H(\operatorname{div},\Kin)$ onto $\boldsymbol H(\operatorname{div},K_{\rm out})$ and it maps
$\boldsymbol{\mathcal{RT}}_p(\Kin)$ onto $\boldsymbol{\mathcal{RT}}_p(K_{\rm out})$. Moreover,
the Piola mappings $\boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T}$
and $\boldsymbol \psi^{\mathrm{d}}_{\boldsymbol T}$ commute with the curl operator in the
sense that
\begin{equation}
\label{eq_piola_commute}
\grad \times \left (\boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T}(\boldsymbol v)\right )
=
\boldsymbol \psi^{\mathrm{d}}_{\boldsymbol T}\left (\grad \times \boldsymbol v\right )
\quad
\forall \boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},\Kin).
\end{equation}
In addition, we have
\begin{equation}
\label{eq_piola_adjoint}
(\boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T}(\boldsymbol v_{\rm in}),\boldsymbol v_{\rm out})_{K_{\rm out}}
=
\operatorname{sign} (\det \mathbb J_{\boldsymbol T}) (\boldsymbol v_{\rm in},(\boldsymbol \psi^{\mathrm{d}}_{\boldsymbol T})^{-1}(\boldsymbol v_{\rm out}))_{\Kin},
\end{equation}
for all $\boldsymbol v_{\rm in} \in \boldsymbol H(\boldsymbol{\operatorname{curl}},\Kin)$ and $\boldsymbol v_{\rm out} \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K_{\rm out})$.
We also have $\|\boldsymbol \psi_{\boldsymbol T}^{\mathrm{c}}(\boldsymbol v)\|_{K_{\rm out}} \leq
\frac{h_{\Kin}}{\rho_{K_{\rm out}}} \|\boldsymbol v\|_{\Kin}$ for all $\boldsymbol v \in \boldsymbol L^2(\Kin)$, so that
whenever $\Kin,K_{\rm out}$ belong to the same edge patch ${\TT^e}$, we have
\begin{equation}
\label{eq_stab_piola_L2}
\|\boldsymbol \psi_{\boldsymbol T}^{\mathrm{c}}(\boldsymbol v)\|_{K_{\rm out}}
\leq
C \|\boldsymbol v\|_{\Kin} \quad \forall \boldsymbol v \in \boldsymbol L^2(\Kin),
\end{equation}
for a constant $C$ only depending on the shape-regularity $\kappa_e$ of the patch ${\TT^e}$
defined in~\eqref{eq_regularities}.
\subsection{Stability of discrete minimization in an edge patch}
\subsubsection{Preliminaries}
In this section, we consider an edge patch ${\TT^e}$ associated with a mesh edge $e\in\mathcal E_h$
consisting of tetrahedral elements $K$ sharing the edge $e$, cf. Figure~\ref{fig_patch}.
We denote by $n := |{\TT^e}|$ the number of tetrahedra in the patch, by ${\FF^e}$ the set of all faces of
the patch, by ${\FF_{\rm int}^e} \subset {\FF^e}$ the set of ``internal'' faces, i.e., those being shared by two
different tetrahedra from the patch, and finally, by
${\FF_{\rm ext}^e} := {\FF^e} \setminus {\FF_{\rm int}^e}$ the set of ``external'' faces.
The patch is either of ``interior'' type, corresponding to an edge in the interior of the
domain $\Omega$, in which case there is a full loop around $e$, see Figure~\ref{fig_patch},
left, or of ``boundary'' type, corresponding to an edge on the boundary of the domain $\Omega$,
in which case there is no full loop around $e$, see Figure~\ref{fig_patch}, right, and
Figure~\ref{figure_numbering_patch}.
We further distinguish three types of patches of boundary type depending on the status of
the two boundary faces sharing the associated boundary edge: the patch is of Dirichlet boundary
type if both faces lie in $\overline {\Gamma_{\rm D}}$, of ``mixed boundary'' type if one face lies
in $\overline {\Gamma_{\rm D}}$ an the other in $\overline {\Gamma_{\rm N}}$, and of Neumann boundary type
if both faces lie in $\overline {\Gamma_{\rm N}}$. Note that for an interior patch, $|{\FF_{\rm int}^e}| = n$,
whereas $|{\FF_{\rm int}^e}| = n-1$ for a boundary patch. The open domain associated with ${\TT^e}$ is denoted
by ${\omega_\edge}$, and $\boldsymbol n_{{\omega_\edge}}$ stands for the unit normal vector to $\partial {\omega_\edge}$ pointing
outward ${\omega_\edge}$.
\begin{figure}[htb]
\centerline{\includegraphics[height=0.35\textwidth]{figures/patch_edge_bound_DN.pdf}
\qquad \qquad
\includegraphics[height=0.35\textwidth]{figures/patch_edge_bound_NN.pdf}}
\caption{Mixed (left) and Neumann (right) boundary patch ${\TT^e}$}
\label{figure_numbering_patch}
\end{figure}
We denote by $\aaa_{\rm d}$ and $\aaa_{\rm u}$ the two vertices of the edge $e$.
The remaining vertices are numbered consecutively in one sense of rotation around the edge $e$
(this sense is only specific for the ``mixed boundary'' patches) and denoted by
$\boldsymbol a_{0},\boldsymbol a_{1},\dots,\boldsymbol a_{n}$, with $\boldsymbol a_{0} = \boldsymbol a_{n}$ if the patch is interior.
Then ${\TT^e} = \bigcup_{j\in\{1:n\}} K_j$ for $K_j := \operatorname{conv}(\boldsymbol a_{j-1},\boldsymbol a_{j},\aaa_{\rm d},\aaa_{\rm u})$;
we also denote $K_0 := K_n$ and $K_{n+1} := K_1$. For all $j\in\{0:n\}$, we define
$F_{j} := \operatorname{conv}(\boldsymbol a_{j},\aaa_{\rm d},\aaa_{\rm u})$, and for all $j\in\{1:n\}$, we let
$F^{\rm d}_j := \operatorname{conv}(\boldsymbol a_{j-1},\boldsymbol a_{j},\aaa_{\rm d})$ and $F^{\rm u}_j := \operatorname{conv}(\boldsymbol a_{j-1},\boldsymbol a_{j},\aaa_{\rm u})$.
Then $\mathcal F_{K_j} = \{F_{j-1},F_{j},F^{\rm d}_j,F^{\rm u}_j\}$,
and $F_0 = F_n$ if the patch is interior. We observe that, respectively for interior
and boundary patches, ${\FF_{\rm int}^e} = \bigcup_{j\in\{0:n-1\}} \{F_{j}\}$,
${\FF_{\rm ext}^e} = \bigcup_{j\in\{1:n\}} \{F^{\rm d}_j,F^{\rm u}_j\}$ and
${\FF_{\rm int}^e} = \bigcup_{j\in\{1:n-1\}} \{F_{j}\}$,
${\FF_{\rm ext}^e} = \bigcup_{j\in\{1:n\}} \{F^{\rm d}_j,F^{\rm u}_j\} \cup \{F_0,F_n\}$.
Finally, if $F_{j} \in {\FF_{\rm int}^e}$ is an internal face, we define its normal vector
by $\boldsymbol n_{F_{j}} := \boldsymbol n_{K_{j+1}} = -\boldsymbol n_{K_j}$, whereas for any external face $F \in {\FF_{\rm ext}^e}$, we
define its normal vector to coincide with the normal vector pointing outward the patch,
$\boldsymbol n_F := \boldsymbol n_{{\omega_\edge}}$.
We now extend the notions of Section~\ref{sec:preliminaries_K} to the edge patch ${\TT^e}$.
Consider the following broken Sobolev spaces:
\begin{align*}
\boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e})
&:=
\left \{
\boldsymbol v \in \boldsymbol L^2({\omega_\edge}) \; | \; \boldsymbol v|_K \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K)
\; \forall K \in {\TT^e}
\right \},
\\
\boldsymbol H^1({\TT^e})
&:=
\left \{
\boldsymbol v \in \boldsymbol L^2({\omega_\edge}) \; | \; \boldsymbol v|_K \in \boldsymbol H^1(K)
\; \forall K \in {\TT^e}
\right \},
\end{align*}
as well as the broken N\'ed\'elec space $\boldsymbol{\mathcal N}_{p}({\TT^e})$.
For all $\boldsymbol v \in \boldsymbol H^1({\TT^e})$, we employ the notation $\jump{\boldsymbol v}_F \in \boldsymbol L^2(F)$
for the ``(strong) jump'' of $\boldsymbol v$ across any face $F\in\mathcal F^e$. Specifically, for an internal face
$F_{j} \in {\FF_{\rm int}^e}$, we set $\jump{\boldsymbol v}_{F_{j}} := (\boldsymbol v|_{K_{j+1}})|_{F_{j}} - (\boldsymbol v|_{K_j})|_{F_{j}}$,
whereas for an external face $F \in {\FF_{\rm ext}^e}$, we set $\jump{\boldsymbol v}_F := \boldsymbol v|_F$. Note in particular that
piecewise polynomial functions from $\boldsymbol{\mathcal N}_{p}({\TT^e})$ belong to $\boldsymbol H^1({\TT^e})$, so that their strong
jumps are well-defined.
To define a notion of a ``weak tangential jump'' for functions of $\boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e})$,
for which a strong (pointwise) definition cannot apply, some preparation is necessary.
Let $\mathcal F$ be a subset of the faces of an edge patch ${\TT^e}$ containing the internal faces,
i.e. ${\FF_{\rm int}^e} \subseteq \mathcal F \subseteq {\FF^e}$, and denote by $\Gamma_\mathcal F$ the corresponding
open set. The set $\mathcal F$ \revision{represents the set of faces appearing in the minimization. It}
depends on the type of edge patch and is reported in Table~\ref{Tab:type_of_calF}. In extension
of~\eqref{eq_tr_K}, we define the piecewise polynomial space on $\Gamma_\mathcal F$
\begin{equation}
\label{eq_tr_Te}
\boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F) :=
\left \{
\boldsymbol w_\mathcal F \in \boldsymbol L^2(\Gamma_\mathcal F) \; | \;
\exists \boldsymbol v_p \in \boldsymbol{\mathcal N}_{p}({\TT^e});
\boldsymbol w_F := (\boldsymbol w_\mathcal F)|_F = \boldsymbol \pi^{\boldsymbol \tau}_F (\jump{\boldsymbol v_p}_F) \quad \forall F \in \mathcal F
\right \}.
\end{equation}
In extension of~\eqref{eq_scurl_curl_el}, for all $\boldsymbol w_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$, we set
\begin{equation}
\label{eq_scurl_curl_patch}
\operatorname{curl}_F (\boldsymbol w_F) := \jump{\grad \times \boldsymbol v_p}_F \cdot \boldsymbol n_F \qquad \forall F \in \mathcal F.
\end{equation}
Then we can extend Definition~\ref{definition_partial_trace} to prescribe weak tangential jumps
of functions in $\boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e})$ as follows:
\begin{table}[tb]
\begin{center}\begin{tabular}{|l|l|l|}
\hline
patch type&\hfil ${\FF_{\rm int}^e}$&\hfil $\mathcal F$\\
\hline
interior&$\{F_1,\ldots,F_n\}$&${\FF_{\rm int}^e} = \{F_1,\ldots,F_n\}$\\
Dirichlet boundary&$\{F_1,\ldots,F_{n-1}\}$&${\FF_{\rm int}^e} = \{F_1,\ldots,F_{n-1}\}$\\
mixed boundary&$\{F_1,\ldots,F_{n-1}\}$&$\{F_0\}\cup {\FF_{\rm int}^e} = \{F_0,F_1,\ldots,F_{n-1}\}$\\
Neumann boundary&$\{F_1,\ldots,F_{n-1}\}$&$\{F_0\}\cup{\FF_{\rm int}^e}\cup\{F_n\}=\{F_0,F_1,\ldots,F_{n-1},F_n\}$\\
\hline
\end{tabular}\end{center}
\caption{The set of internal faces ${\FF_{\rm int}^e}$ and the set $\mathcal F$ used for the minimization problems on the edge patch for the four patch types.}
\label{Tab:type_of_calF}%
\end{table}
\begin{definition}[Tangential jumps by integration by parts in an edge patch]
\label{definition_jumps}
Given $\boldsymbol r_{\mathcal F} \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$ and $\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e})$,
we employ the notation ``$\jump{\boldsymbol v}^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F$'' to say that
\begin{equation}\label{eq_prescription_tang_jump_e}
\sum_{K \in {\TT^e}}
\left \{
(\grad \times \boldsymbol v,\boldsymbol \phi)_K - (\boldsymbol v,\grad \times \boldsymbol \phi)_K
\right \}
=
\sum_{F \in \mathcal F} (\boldsymbol r_F,\boldsymbol \phi \times \boldsymbol n_F)_F
\quad
\forall \boldsymbol \phi \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e}),
\end{equation}
where
\begin{equation} \label{eq_Htpatch}
\boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e})
:=
\left \{
\boldsymbol \phi \in \boldsymbol H^1({\TT^e}) \; | \;
\jump{\boldsymbol \phi}_F \times \boldsymbol n_F = \boldsymbol 0 \quad
\forall F \in {\FF_{\rm int}^e} \cup ({\FF_{\rm ext}^e} \setminus \mathcal F)
\right \}.
\end{equation}
Whenever $\boldsymbol v \in \boldsymbol H^1({\TT^e})$, $\jump{\boldsymbol v}^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F$ if and only if
$\boldsymbol \pi^{\boldsymbol \tau}_F(\jump{\boldsymbol v}_F) = \boldsymbol r_F$ for all $F \in \mathcal F$. Note that
$\boldsymbol \phi\times \boldsymbol n_F$ in~\eqref{eq_prescription_tang_jump_e} is uniquely defined for
all $\boldsymbol \phi \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e})$.
\end{definition}
\subsubsection{Statement of the stability result in an edge patch}
Henceforth, if $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$ is an elementwise Raviart--Thomas function,
we will employ the notation $\boldsymbol r_K := \boldsymbol r_\mathcal T|_K$ for all $K \in {\TT^e}$. In addition,
if $\boldsymbol v_{\mathcal T} \in \boldsymbol{\mathcal N}_{p}({\TT^e})$ is an elementwise N\'ed\'elec function, the notations
$\div \boldsymbol r_\mathcal T$ and $\grad \times \boldsymbol v_{\mathcal T}$ will be understood elementwise.
\begin{definition}[Compatible data]
\label{definition_compatible_data}
Let $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$ and $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$.
We say that the data $\boldsymbol r_\mathcal T$ and $\boldsymbol r_\mathcal F$ are compatible if
\begin{subequations}\begin{align}
\div \boldsymbol r_\mathcal T & = 0, \label{eq_comp_a} \\
\jump{\boldsymbol r_\mathcal T}_{F} \cdot \boldsymbol n_F & = \operatorname{curl}_F (\boldsymbol r_F) \quad \forall F \in \mathcal F, \label{eq_comp_b}
\end{align}
and with the following additional condition whenever the patch is either of interior or Neumann
boundary type:
\begin{alignat}{2}
\sum_{j\in\{1:n\}} \boldsymbol r_{F_j}|_e \cdot {\boldsymbol \tau}_e
&=
0
&\qquad&
\textrm{(interior type)},
\label{eq_comp_c} \\
\sum_{j\in\{0:n-1\}} \boldsymbol r_{F_j}|_e \cdot {\boldsymbol \tau}_e
&=
\boldsymbol r_{F_n}|_e \cdot {\boldsymbol \tau}_e
&\qquad&
\textrm{(Neumann boundary type)}.
\label{eq_comp_d}
\end{alignat}\end{subequations}
\end{definition}
\begin{definition}[Broken patch spaces]\label{def_spaces}
Let $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$ and $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$
be compatible data as per Definition~\ref{definition_compatible_data}. We define
\begin{subequations}\begin{align}
\boldsymbol V({\TT^e}) & :=
\left \{
\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e}) \; \left |
\begin{array}{rl}
\grad \times \boldsymbol v &= \boldsymbol r_\mathcal T
\\
\jump{\boldsymbol v}^{\boldsymbol \tau}_\mathcal F &= \boldsymbol r_\mathcal F
\end{array}
\right .
\right \}, \label{eq_VTe}\\
\boldsymbol V_p({\TT^e}) & := \boldsymbol V({\TT^e}) \cap \boldsymbol{\mathcal N}_p({\TT^e}). \label{eq_VqTe}
\end{align}\end{subequations}
\end{definition}
We will show in Lemma~\ref{lem_EU} below that the space $\boldsymbol V_p({\TT^e})$
(and therefore also $\boldsymbol V({\TT^e})$) is nonempty.
We are now ready to present our central result of independent interest.
To facilitate the reading, the proof is postponed to Section~\ref{sec:proof_stability_patch}.
\begin{proposition}[Stable broken $\boldsymbol H(\boldsymbol{\operatorname{curl}})$ polynomial extension in an edge patch]
\label{prop_stability_patch}
Let an edge $e\in\mathcal E_h$ and the associated edge patch ${\TT^e}$ with subdomain ${\omega_\edge}$
be fixed. Let the set of faces $\mathcal F$ be specified in Table~\ref{Tab:type_of_calF}.
Then, for every polynomial degree $p \geq 0$, all $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$, and
all $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$ compatible as per
Definition~\ref{definition_compatible_data}, the following holds:
\begin{equation} \label{eq_stab_shift}
\min_{\boldsymbol v_p\in {\boldsymbol V}_p({\TT^e})} \|\boldsymbol v_p\|_{{\omega_\edge}}
= \min_{\substack{
\boldsymbol v_p \in \boldsymbol{\mathcal N}_p({\TT^e})
\\
\grad \times \boldsymbol v_p = \boldsymbol r_\mathcal T
\\
\jump{\boldsymbol v_p}^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F
}}
\|\boldsymbol v_p\|_{{\omega_\edge}}
\le C_{\mathrm{st},e}
\min_{\substack{
\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},{\TT^e})
\\
\grad \times \boldsymbol v = \boldsymbol r_\mathcal T
\\
\jump{\boldsymbol v}^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F
}}
\|\boldsymbol v\|_{{\omega_\edge}} =
C_{\mathrm{st},e} \min_{\boldsymbol v\in {\boldsymbol V}({\TT^e})} \|\boldsymbol v\|_{{\omega_\edge}}.
\end{equation}
Here, all the minimizers are uniquely defined and the constant $C_{\mathrm{st},e}$
only depends on the shape-regularity parameter $\kappa_e$ of the patch ${\TT^e}$
defined in~\eqref{eq_regularities}.
\end{proposition}
\begin{remark}[Converse inequality in Proposition~\ref{prop_stability_patch}]
\label{rem_conv}
Note that the converse to the inequality~\eqref{eq_stab_shift} holds trivially with constant one.
\end{remark}
\subsection{Equivalence of Theorem~\ref{theorem_stability} with Proposition~\ref{prop_stability_patch}}
$ $
\revision{We have the following important link, establishing Theorem~\ref{theorem_stability},
including the existence and uniqueness of the minimizers.
\begin{lemma}[Equivalence of Theorem~\ref{theorem_stability} with
Proposition~\ref{prop_stability_patch}]
Theorem~\ref{theorem_stability} holds if and only if Proposition~\ref{prop_stability_patch} holds.
{\color{black} More precisely, let $\boldsymbol h_p^\star \in \boldsymbol{\mathcal N}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$
and $\boldsymbol h^\star \in \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\boldsymbol{\operatorname{curl}},{\omega_\edge})$ by any solutions to the minimization problems of
Theorem~\ref{theorem_stability} for the data
$\boldsymbol j_h^{e} \in \boldsymbol{\mathcal{RT}}_p({\TT^e}) \cap \boldsymbol H_{{\Gamma_{\rm N}^\edge}}(\operatorname{div},{\omega_\edge})$ with $\div \boldsymbol j_h^{e} = 0$ and
$\boldsymbol \chi_h \in \boldsymbol{\mathcal N}_p({\TT^e})$. Let $\boldsymbol v^\star_p \in {\boldsymbol V}_p({\TT^e})$ and
$\boldsymbol v^\star \in \boldsymbol V({\TT^e})$ be any minimizers of
Proposition~\ref{prop_stability_patch} for the data
\begin{equation} \label{eq_data_eq}
\boldsymbol r_\mathcal T := \boldsymbol j_h^{e}-\grad \times \boldsymbol \chi_h, \qquad
\boldsymbol r_F := - \boldsymbol \pi^{\boldsymbol \tau}_F \left (\jump{\boldsymbol \chi_h}_F\right ) \quad \forall F \in \mathcal F,
\end{equation}
where $\mathcal F$ is specified in Table~\ref{Tab:type_of_calF}.} Then
{\color{black}
\begin{equation} \label{eq_eq}
\boldsymbol h_p^\star - \boldsymbol \chi_h = \boldsymbol v^\star_p, \quad
\boldsymbol h^\star - \boldsymbol \chi_h = \boldsymbol v^\star.
\end{equation}
In the converse direction, for given data $\boldsymbol r_\mathcal T$ and $\boldsymbol r_\mathcal F$} in Proposition~\ref{prop_stability_patch},
{\color{black}
compatible as per Definition~\ref{definition_compatible_data}, taking any $\boldsymbol \chi_h \in \boldsymbol{\mathcal N}_p({\TT^e})$
such that $- \boldsymbol \pi^{\boldsymbol \tau}_F \left (\jump{\boldsymbol \chi_h}_F\right ) = \boldsymbol r_F$ for all $F \in \mathcal F$ and
$\boldsymbol j_h^{e} := \boldsymbol r_\mathcal T + \grad \times \boldsymbol \chi_h$} gives minimizers of Theorem~\ref{theorem_stability} such that{\color{black} ~\eqref{eq_eq} holds true}.
\end{lemma}
}
\begin{proof} \revision{The proof follows} via a shift by the datum $\boldsymbol \chi_h$.
In order to show~\eqref{eq_eq} in the forward direction \revision{(the converse
direction is actually easier)}, we merely need to show that
$\boldsymbol r_\mathcal T$ and $\boldsymbol r_\mathcal F$ prescribed by~\eqref{eq_data_eq} are compatible data
as per Definition~\ref{definition_compatible_data}.
Indeed, to start with, since $\boldsymbol j_h^{e}, \grad \times \boldsymbol \chi_h \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$,
we have $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$. In addition, since $\div \boldsymbol j_h^{e} = 0$
from~\eqref{eq_definition_jj_h},
\begin{equation*}
\div \boldsymbol r_\mathcal T = \div \boldsymbol j_h^{e} - \div (\grad \times \boldsymbol \chi_h) = 0,
\end{equation*}
which is~\eqref{eq_comp_a}.
Then, for all $j\in\{1:n\}$ if the patch is of interior type and for all $j\in\{1:n-1\}$
if the patch is of boundary type, we have
\begin{equation*}
\boldsymbol r_{F_{j}}
=
\boldsymbol \pi^{\boldsymbol \tau}_{F_j} (\boldsymbol \chi_h|_{K_j})
-
\boldsymbol \pi^{\boldsymbol \tau}_{F_j} (\boldsymbol \chi_h|_{K_{j+1}}),
\end{equation*}
and therefore, recalling the definition~\eqref{eq_scurl_curl_el} of the surface curl,
we infer that
\begin{align*}
\operatorname{curl}_{F_{j}} (\boldsymbol r_{F_{j}})
&=
\operatorname{curl}_{F_{j}} (\boldsymbol \pi^{\boldsymbol \tau}_{F_j} (\boldsymbol \chi_h|_{K_j}))
-
\operatorname{curl}_{F_{j}} (\boldsymbol \pi^{\boldsymbol \tau}_{F_j} (\boldsymbol \chi_h|_{K_{j+1}}))
\\
&=
(\grad \times \boldsymbol \chi_h)|_{K_{j}}|_{F_j} \cdot \boldsymbol n_{F_{j}}
-
(\grad \times \boldsymbol \chi_h)|_{K_{j+1}}|_{F_j} \cdot \boldsymbol n_{F_{j}}
\\
&=
- \jump{\grad \times \boldsymbol \chi_h}_{F_{j}} \cdot \boldsymbol n_{F_{j}}.
\end{align*}
On the other hand, since $\boldsymbol j_h^{e} \in \boldsymbol H(\operatorname{div},{\omega_\edge})$,
we have $\jump{\boldsymbol j_h^{e}}_{F_{j}} \cdot \boldsymbol n_{F_{j}} = 0$,
and therefore
\begin{align*}
\jump{\boldsymbol r_\mathcal T}_{F_{j}} \cdot \boldsymbol n_{F_{j}}
&=
- \jump{\grad \times \boldsymbol \chi_h}_{F_{j}} \cdot \boldsymbol n_{F_{j}}
=
\operatorname{curl}_{F_{j}} (\boldsymbol r_{F_{j}}).
\end{align*}
Since a similar reasoning applies on the face $F_0$
if the patch is of Neumann or mixed boundary type and on the face $F_n$
if the patch is of Neumann boundary type, \eqref{eq_comp_b} is established.
It remains to show that $\boldsymbol r_\mathcal F$ satisfies the edge compatibility
condition~\eqref{eq_comp_c} or~\eqref{eq_comp_d} if the patch is interior or
Neumann boundary type, respectively. Let us treat the first case
(the other case is treated similarly). Owing to the convention $K_{n+1}=K_1$,
we infer that
\begin{equation*}
\sum_{j\in\{1:n\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e
=
\sum_{j\in\{1:n\}} (\boldsymbol \chi_h|_{K_j} - \boldsymbol \chi_h|_{K_{j+1}})|_e \cdot {\boldsymbol \tau}_e
=
0,
\end{equation*}
which establishes~\eqref{eq_comp_c}. We have thus shown that $\boldsymbol r_\mathcal T$ and $\boldsymbol r_\mathcal F$
are compatible data as per Definition~\ref{definition_compatible_data}.
\end{proof}
\subsection{Proof of Proposition~\ref{prop_stability_patch}}
\label{sec:proof_stability_patch}
The proof of Proposition~\ref{prop_stability_patch} is performed in two steps.
First we prove that ${\boldsymbol V}_p({\TT^e})$ is nonempty by
providing a generic elementwise construction of a field in ${\boldsymbol V}_p({\TT^e})$;
this in particular implies the existence and uniqueness of all minimizers in~\eqref{eq_stab_shift}.
Then we prove the inequality~\eqref{eq_stab_shift} by using
one such field $\boldsymbol \xi_p^\star \in \boldsymbol V_p({\TT^e})$.
Throughout this section, if $A,B \geq 0$ are two real numbers, we employ
the notation $A \lesssim B$ to say that there exists a constant $C$ that only depends on the
shape-regularity parameter $\kappa_e$ of the patch ${\TT^e}$ defined in~\eqref{eq_regularities},
such that $A \leq CB$. We note that in particular we have $n = |{\TT^e}| \lesssim 1$
owing to the shape-regularity of the mesh $\mathcal T_h$.
\subsubsection{Generic elementwise construction of fields in ${\boldsymbol V}_p({\TT^e})$}
The generic construction of fields in ${\boldsymbol V}_p({\TT^e})$ is based
on a loop over the mesh elements composing the edge patch ${\TT^e}$. This loop is
enumerated by means of an index $j\in\{1:n\}$.
\begin{definition}[Element spaces]\label{def_spaces_elm}
For each $j\in\{1:n\}$, let $\emptyset\subseteq \mathcal F_j \subseteq \mathcal F_{K_j}$ be a
(sub)set of the faces of $K_j$. Let $\boldsymbol r_{K_j} \in \boldsymbol{\mathcal{RT}}_p(K_j)$ with $\div \boldsymbol r_{K_j} = 0$,
and if $\emptyset\ne \mathcal F_j$, let $\widetilde \rr_{\mathcal F_j}^j \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\mathcal F_j})$
in the sense of~\eqref{eq_tr_K} be given data. We define
\begin{subequations}\begin{align}
\boldsymbol V(K_j) & := \left \{
\boldsymbol v \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K_j) \; \left | \;
\begin{array}{l}
\grad \times \boldsymbol v = \boldsymbol r_{K_j}
\\
\boldsymbol v|^{\boldsymbol \tau}_{\mathcal F_j} = \widetilde \rr^j_{\mathcal F_j},
\end{array}
\right .
\right \}, \label{eq_VKj} \\
\boldsymbol V_p(K_j) & := \boldsymbol V(K_j) \cap \boldsymbol{\mathcal N}_p(K_j). \label{eq_VqKj}
\end{align}\end{subequations}
\end{definition}
In what follows, we are only concerned with the cases where $\mathcal F_j$ is either empty
or composed of one or two faces of $K_j$. In this situation, the subspace ${\boldsymbol V}_p(K_j)$
is nonempty if and only if
\begin{subequations} \label{eq:cond_comp}
\begin{alignat}{2}
&\operatorname{curl}_{F} (\widetilde \rr_{F}^{j}) = \boldsymbol r_{K_j} \cdot \boldsymbol n_F&\quad&\forall F\in\mathcal F_j,\label{eq:cond_comp1} \\
&\widetilde \rr_{F_+}^{j}|_{e} \cdot {\boldsymbol \tau}_e = \widetilde \rr_{F_-}^{j}|_{e} \cdot {\boldsymbol \tau}_e&\quad&\text{if $\mathcal F_j=\{F_+,F_-\}$ with $e=F_+\cap F_-$}, \label{eq:cond_comp2}
\end{alignat}
\end{subequations}
where $\boldsymbol n_F$ is the unit normal orienting $F$ used in the definition of the surface curl
(see~\eqref{eq_scurl_curl_el}). The second condition~\eqref{eq:cond_comp2} is relevant only if
$|\mathcal F_j|=2$.
\begin{lemma}[Generic elementwise construction]\label{lem_EU}
Let $e\in\mathcal E_h$, let ${\TT^e}$ be the edge patch associated with $e$, and let
the set of faces $\mathcal F$ be specified in Table~\ref{Tab:type_of_calF}.
Let $\boldsymbol r_\mathcal T \in \boldsymbol{\mathcal{RT}}_p({\TT^e})$ and $\boldsymbol r_\mathcal F \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_\mathcal F)$ be compatible data
as per Definition~\ref{definition_compatible_data}. Define
$\boldsymbol r_{K_j}:= \boldsymbol r_{\mathcal T}|_{K_j}$ for all $j\in\{1:n\}$. Then,
the following inductive procedure yields a sequence of nonempty spaces
$({\boldsymbol V}_p(K_j))_{j\in\{1:n\}}$ in the sense of Definition~\ref{def_spaces_elm},
as well as a sequence of fields $(\boldsymbol \xi_p^j)_{j\in\{1:n\}}$
such that $\boldsymbol \xi_p^j\in {\boldsymbol V}_p(K_j)$ for all $j\in\{1:n\}$. Moreover, the field
$\boldsymbol \xi_p$ prescribed by $\boldsymbol \xi_p|_{K_j} := \boldsymbol \xi_p^j$ for all $j\in\{1:n\}$ belongs to the space
$\boldsymbol V_p({\TT^e})$ of Definition~\ref{def_spaces}:
\\
\begin{subequations}
{\bf 1)} First element ($j=1$): Set $\mathcal F_1 := \emptyset$ if the patch is of interior or
Dirichlet boundary type and set $\mathcal F_1 := \{F_0\}$ if the patch is of Neumann or mixed
boundary type together with
\begin{equation}
\widetilde \rr^1_{F_0} := \boldsymbol r_{F_0}.\label{eq_F_0}
\end{equation}
Define the space ${\boldsymbol V}_p(K_1)$ according to~\eqref{eq_VqKj} and pick any
$\boldsymbol \xi_p^1\in {\boldsymbol V}_p(K_1)$. \\
{\bf 2)} Middle elements ($j\in\{2:n-1\}$): Set $\mathcal F_j:= \{F_{j-1}\}$ together with
\begin{equation}
\widetilde \rr^j_{F_{j-1}} := \boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi^{j-1}_p) + \boldsymbol r_{F_{j-1}}, \label{eq_F_j}
\end{equation}
with $\boldsymbol \xi^{j-1}_p$ obtained in the previous step of the procedure.
Define the space ${\boldsymbol V}_p(K_j)$ according to~\eqref{eq_VqKj} and pick any
$\boldsymbol \xi_p^j\in {\boldsymbol V}_p(K_j)$. \\
{\bf 3)} Last element ($j=n$): Set $\mathcal F_{n} := \{F_{n-1}\}$ if the patch is of Dirichlet or mixed boundary type and set $\mathcal F_n := \{ F_{n-1}, F_n \}$ if the patch is of interior or Neumann boundary type and define $\widetilde \rr^n_{\mathcal F_n}$ as follows:
For the four cases of the patch,
\begin{equation}
\widetilde \rr^n_{F_{n-1}} := \boldsymbol \pi^{\boldsymbol \tau}_{F_{n-1}} (\boldsymbol \xi^{n-1}_p) + \boldsymbol r_{F_{n-1}}, \label{eq_F_nn}
\end{equation}
with $\boldsymbol \xi^{n-1}_p$ obtained in the previous step of the procedure, and in the two cases where $\mathcal F_n$ also contains $F_n$:
\begin{alignat}{2}
\widetilde \rr_{F_n}^n & := \boldsymbol \pi^{\boldsymbol \tau}_{F_n} (\boldsymbol \xi_p^1) - \boldsymbol r_{F_n} &\qquad&\text{interior type}, \label{eq_F_int}\\
\widetilde \rr_{F_n}^n & := \boldsymbol r_{F_n} &\qquad&\text{Neumann boundary type}. \label{eq_F_n}
\end{alignat}
Define the space ${\boldsymbol V}_p(K_n)$ according to~\eqref{eq_VqKj} and pick any
$\boldsymbol \xi_p^n\in {\boldsymbol V}_p(K_n)$. \\
\end{subequations}
\end{lemma}
\begin{proof}
We first show that $\boldsymbol \xi^j_p$ is well-defined in ${\boldsymbol V}_p(K_j)$ for all
$j\in\{1:n\}$. We do so by verifying that $\widetilde \rr^{j}_{\mathcal F_{j}} \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\mathcal F_{j}})$
(recall~\eqref{eq_tr_K}) and that the conditions~\eqref{eq:cond_comp} hold true for all
$j\in\{1:n\}$. Then, we show that $\boldsymbol \xi_p \in \boldsymbol V_p({\TT^e})$.
{\bf (1)} First element ($j=1$). If the patch is of interior or Dirichlet boundary type,
there is nothing to verify since $\mathcal F_1$ is empty. If the patch is of Neumann or mixed
boundary type, $\mathcal F_1 = \{F_0\}$ and we need to verify that
$\widetilde \rr^1_{F_0} \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\{F_0\}})$ and that
$\operatorname{curl}_{F_0} (\widetilde \rr^{1}_{F_0})=\boldsymbol r_{K_1} \cdot \boldsymbol n_{F_0}$, see~\eqref{eq:cond_comp1}.
Since $\widetilde \rr^1_{\mathcal F_1}=\boldsymbol r_{F_0}\in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\{F_0\}})$ by assumption, the first
requirement is met. The second one follows from
$\boldsymbol r_{K_1} \cdot \boldsymbol n_{F_0} = \jump{\boldsymbol r_\mathcal T}_{F_0} \cdot \boldsymbol n_{F_0} =
\operatorname{curl}_{F_0}(\boldsymbol r_{F_0}) = \operatorname{curl}_{F_0} (\widetilde \rr^{1}_{F_0})$ owing to~\eqref{eq_comp_b}.
{\bf (2)} Middle elements ($j\in\{2:n-1\}$). Since $\mathcal F_j = \{F_{j-1}\}$, we need to show
that $\widetilde \rr^j_{F_{j-1}} \in \boldsymbol{\mathcal N}_p^{\boldsymbol \tau}(\Gamma_{\{F_{j-1}\}})$ and that
$\operatorname{curl}_{F_{j-1}} (\widetilde \rr^{j}_{F_{j-1}})=\boldsymbol r_{K_j} \cdot \boldsymbol n_{F_{j-1}}$.
The first requirement follows from the definition~\eqref{eq_F_j} of $\widetilde \rr^j_{F_{j-1}}$.
To verify the second requirement, we recall the
definition~\eqref{eq_scurl_curl_el} of the surface curl
and use the curl constraint from~\eqref{eq_VKj} to infer that
\begin{align*}
\operatorname{curl}_{F_{j-1}} (\widetilde \rr_{F_{j-1}}^{j})
&=
\operatorname{curl}_{F_{j-1}} (\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^{j-1})) + \operatorname{curl}_{F_{j-1}} (\boldsymbol r_{F_{j-1}})
\\
&=
\left (\grad \times \boldsymbol \xi_p^{j-1}\right )|_{F_{j-1}} \cdot \boldsymbol n_{F_{j-1}} + \operatorname{curl}_{F_{j-1}} (\boldsymbol r_{F_{j-1}})
\\
&=
\boldsymbol r_{K_{j-1}} \cdot \boldsymbol n_{F_{j-1}} + \operatorname{curl}_{F_{j-1}} (\boldsymbol r_{F_{j-1}}).
\end{align*}
By virtue of assumption~\eqref{eq_comp_b}, it follows that
\begin{align*}
\boldsymbol r_{K_{j}} \cdot \boldsymbol n_{F_{j-1}} - \operatorname{curl}_{F_{j-1}} (\widetilde \rr_{F_{j-1}}^{j})
&=
\boldsymbol r_{K_{j}} \cdot \boldsymbol n_{F_{j-1}} - \boldsymbol r_{K_{j-1}} \cdot \boldsymbol n_{F_{j-1}} - \operatorname{curl}_{F_{j-1}} (\boldsymbol r_{F_{j-1}})
\\
&=
\jump{\boldsymbol r_\mathcal T}_{F_{j-1}} \cdot \boldsymbol n_{F_{j-1}} - \operatorname{curl}_{F_{j-1}} (\boldsymbol r_{F_{j-1}}) = 0.
\end{align*}
{\bf (3)} Last element ($j=n$). We distinguish two cases.
{\bf (3a)} Patch of Dirichlet or mixed boundary type. In this case, $\mathcal F_n = \{F_{n-1}\}$
and the reasoning is identical to the case of a middle element.
{\bf (3b)} Patch of interior or Neumann boundary type. In this case, $\mathcal F_n = \{F_{n-1},F_n\}$.
First, the prescriptions~\eqref{eq_F_nn}--\eqref{eq_F_int}--\eqref{eq_F_n} imply
that $\widetilde \rr^n_{\mathcal F_n} \in \boldsymbol{\mathcal N}^{\boldsymbol \tau}_p(\Gamma_{\mathcal F_n})$ in the sense of~\eqref{eq_tr_K}.
It remains to show~\eqref{eq:cond_comp1}, i.e.
\begin{equation} \label{eq:cond_comp_n1}
\operatorname{curl}_{F_{n-1}}(\widetilde \rr^n_{F_{n-1}})=\boldsymbol r_{K_n}\cdot\boldsymbol n_{F_{n-1}}, \qquad
\operatorname{curl}_{F_{n}}(\widetilde \rr^n_{F_{n}})=\boldsymbol r_{K_n}\cdot\boldsymbol n_{F_{n}},
\end{equation}
and, since $\mathcal F_n$ is composed of two faces, we also need to show the edge compatibility
condition~\eqref{eq:cond_comp2}, i.e.
\begin{equation} \label{eq:cond_comp_n2}
\widetilde \rr^n_{F_{n-1}}|_e \cdot {\boldsymbol \tau}_e = \widetilde \rr^n_{F_{n}}|_e \cdot {\boldsymbol \tau}_e.
\end{equation}
The proof of the first identity in~\eqref{eq:cond_comp_n1} is as above, so we now detail
the proof of the second identity in~\eqref{eq:cond_comp_n1} and the proof
of~\eqref{eq:cond_comp_n2}.
{\bf (3b-I)} Let us consider the case of a patch of interior type. To prove the
second identity in~\eqref{eq:cond_comp_n1}, we use definition~\eqref{eq_scurl_curl_el}
of the surface curl together with the curl constraint in~\eqref{eq_VKj} and infer that
\begin{align*}
\operatorname{curl}_{F_n} (\widetilde \rr_{F_n}^n)
&=
\operatorname{curl}_{F_n} (\boldsymbol \pi^{\boldsymbol \tau}_{F_n}(\boldsymbol \xi_p^1) - \boldsymbol r_{F_n})
\\
&=
\grad \times \boldsymbol \xi_p^1 \cdot \boldsymbol n_{F_n} - \operatorname{curl}_{F_n} (\boldsymbol r_{F_n})
\\
&=
\boldsymbol r_{K_1} \cdot \boldsymbol n_{F_n} - \operatorname{curl}_{F_n} (\boldsymbol r_{F_n}).
\end{align*}
This gives
\begin{align*}
\boldsymbol r_{K_n} \cdot \boldsymbol n_{F_n} - \operatorname{curl}_F (\widetilde \rr_{F_n}^n)
&=
(\boldsymbol r_{K_n} - \boldsymbol r_{K_1}) \cdot \boldsymbol n_{F_n} + \operatorname{curl}_F (\boldsymbol r_{F_n})
\\
&=
-\jump{\boldsymbol r}_{F_n} \cdot \boldsymbol n_{F_n} + \operatorname{curl}_F (\boldsymbol r_{F_n})
=
0,
\end{align*}
where the last equality follows from~\eqref{eq_comp_b}.
This proves the expected identity on the curl.
Let us now prove~\eqref{eq:cond_comp_n2}.
For all $j\in\{1:n-1\}$, since $\boldsymbol \xi_p^j \in \boldsymbol{\mathcal N}_p(K_j)$, its tangential traces satisfy the
edge compatibility condition
\begin{equation} \label{eq:edge_compatibility_xxi}
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^j)\right ) \right |_e \cdot {\boldsymbol \tau}_e
=
\left .\left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{j}} (\boldsymbol \xi_p^j)\right ) \right |_e \cdot {\boldsymbol \tau}_e.
\end{equation}
Moreover, for all $j\in\{1:n-2\}$, we have $F_j\in\mathcal F_{j+1}$, so that by~\eqref{eq_VKj}
and the definition~\eqref{eq_F_j} of $\widetilde \rr^{j+1}_{F_{j}}$, we have
\begin{equation*}
\boldsymbol \pi^{\boldsymbol \tau}_{F_{j}} (\boldsymbol \xi_p^{j+1}) = \widetilde \rr^{j+1}_{F_{j}}
=
\boldsymbol \pi^{\boldsymbol \tau}_{F_{j}} (\boldsymbol \xi_p^j) + \boldsymbol r_{F_{j}},
\end{equation*}
and, therefore, using~\eqref{eq:edge_compatibility_xxi} yields
\begin{equation*}
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^j)\right ) \right |_e \cdot {\boldsymbol \tau}_e
=
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{j}} (\boldsymbol \xi_p^{j+1})\right ) \right |_e \cdot {\boldsymbol \tau}_e -
\boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e.
\end{equation*}
Summing this identity for all $j\in\{1:n-2\}$ leads to
\[
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{0}} (\boldsymbol \xi_p^1)\right ) \right |_e \cdot {\boldsymbol \tau}_e
=
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{n-2}} (\boldsymbol \xi_p^{n-1})\right ) \right |_e \cdot {\boldsymbol \tau}_e
- \sum_{j\in\{1:n-2\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e.
\]
In addition, using again the edge compatibility condition~\eqref{eq:edge_compatibility_xxi}
for $j=n-1$ and the definition~\eqref{eq_F_nn} of $\widetilde \rr_{F_{n-1}}^n$ leads to
\[
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{n-2}} (\boldsymbol \xi_p^{n-1})\right ) \right |_e \cdot {\boldsymbol \tau}_e
= \widetilde \rr_{F_{n-1}}^n|_e \cdot {\boldsymbol \tau}_e - \boldsymbol r_{F_{n-1}}|_e \cdot {\boldsymbol \tau}_e.
\]
Summing the above two identities gives
\begin{equation}
\label{tmp_induction_edge2}
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{0}} (\boldsymbol \xi_p^1)\right ) \right |_e \cdot {\boldsymbol \tau}_e
=
\widetilde \rr_{F_{n-1}}^n|_e \cdot {\boldsymbol \tau}_e -
\sum_{j\in\{1:n-1\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e.
\end{equation}
Since $F_0=F_n$ for a patch of interior type and $\widetilde \rr_{F_n}^n = \boldsymbol \pi^{\boldsymbol \tau}_{F_n} (\boldsymbol \xi_p^1) - \boldsymbol r_{F_n}$ owing to~\eqref{eq_F_int}, the identity~\eqref{tmp_induction_edge2} gives
\begin{align*}
\widetilde \rr_{F_n}^n |_e\cdot {\boldsymbol \tau}_e
&=
\left . \left (\boldsymbol \pi^{\boldsymbol \tau}_{F_{0}} (\boldsymbol \xi_p^1)\right ) \right |_e \cdot {\boldsymbol \tau}_e
-
\left (
\boldsymbol r_{F_n} |_e \cdot {\boldsymbol \tau}_e
\right )
\\
&= \widetilde \rr_{F_{n-1}}^n|_e \cdot {\boldsymbol \tau}_e
-
\sum_{j\in\{1:n-1\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e
-
\left (
\boldsymbol r_{F_n} |_e \cdot {\boldsymbol \tau}_e
\right )
\\
&=
\widetilde \rr^n_{F_{n-1}}|_e \cdot {\boldsymbol \tau}_e
-
\sum_{j\in\{1:n\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e = \widetilde \rr^n_{F_{n-1}}|_e \cdot {\boldsymbol \tau}_e,
\end{align*}
where we used the edge compatibility condition~\eqref{eq_comp_c} satisfied by $\boldsymbol r_\mathcal F$ in the last equality. This proves~\eqref{eq:cond_comp_n2} in the interior case.
{\bf (3b-N)} Let us finally consider a patch of Neumann boundary type. The
second identity in~\eqref{eq:cond_comp_n1} follows directly from~\eqref{eq_comp_b}
and~\eqref{eq_F_n}. Let us now prove~\eqref{eq:cond_comp_n2}.
The identity~\eqref{tmp_induction_edge2} still holds true. Using that
$(\boldsymbol \pi^{\boldsymbol \tau}_{F_{0}} (\boldsymbol \xi_p^1)) |_e \cdot {\boldsymbol \tau}_e = \boldsymbol r_{F_0}|_e \cdot {\boldsymbol \tau}_e$,
this identity is rewritten as
\[
\widetilde \rr_{F_{n-1}}^n|_e \cdot {\boldsymbol \tau}_e = \sum_{j\in\{0:n-1\}} \boldsymbol r_{F_{j}}|_e \cdot {\boldsymbol \tau}_e = \boldsymbol r_{F_n}|_e \cdot {\boldsymbol \tau}_e,
\]
where the last equality follows from the edge compatibility condition~\eqref{eq_comp_d} satisfied by
$\boldsymbol r_\mathcal F$. But since $\widetilde \rr_{F_n}^n = \boldsymbol r_{F_n}$ owing to~\eqref{eq_F_n}, this again
proves~\eqref{eq:cond_comp_n2}.
{\bf (4)} It remains to show that $\boldsymbol \xi_p \in \boldsymbol V_p({\TT^e})$ as per
Definition~\ref{def_spaces}. By construction, we have
$\boldsymbol \pi^{\boldsymbol \tau}_{F_{j}}(\boldsymbol \xi_p|_{K_{j+1}}) - \boldsymbol \pi^{\boldsymbol \tau}_{F_{j}}(\boldsymbol \xi_p|_{K_j}) = \boldsymbol r_{F_{j}}$
for all $j\in\{1:n-1\}$,
$\boldsymbol \pi^{\boldsymbol \tau}_{F_{n}}(\boldsymbol \xi_p|_{K_{0}}) - \boldsymbol \pi^{\boldsymbol \tau}_{F_{n}}(\boldsymbol \xi_p|_{K_n}) = \boldsymbol r_{F_{n}}$
if the patch is of interior type, $\boldsymbol \pi^{\boldsymbol \tau}_{F_{0}}(\boldsymbol \xi_p|_{K_{1}}) = \boldsymbol r_{F_{0}}$
if the patch is of Neumann or mixed boundary type, and
$\boldsymbol \pi^{\boldsymbol \tau}_{F_{n}}(\boldsymbol \xi_p|_{K_{n}}) = \boldsymbol r_{F_{n}}$ if the patch is of Neumann type.
This proves that $\boldsymbol \pi^{\boldsymbol \tau}_F(\jump{\boldsymbol \xi_p}_F) = \boldsymbol r_F$ for all $F \in \mathcal F$, i.e.,
$\jump{\boldsymbol \xi_p}^{\boldsymbol \tau}_\mathcal F = \boldsymbol r_\mathcal F$ in the sense of Definition~\ref{definition_jumps}.
\end{proof}
\subsubsection{The actual proof}
We are now ready to prove Proposition~\ref{prop_stability_patch}.
\begin{proof}[Proof of Proposition~\ref{prop_stability_patch}]
Owing to Lemma~\ref{lem_EU}, the fields
\begin{equation} \label{eq_min_K}
\boldsymbol \xi_p^{\star j} := \argmin{\boldsymbol v_p \in \boldsymbol V_p(K_j)} \|\boldsymbol v_p\|_{K_j},
\qquad j\in\{1:n\},
\end{equation}
are uniquely defined in ${\boldsymbol V}_p(K_j)$, and the field $\boldsymbol \xi_p^{\star}$ such that
$\boldsymbol \xi_p^{\star}|_{K_j} := \boldsymbol \xi_p^{\star j}$ for all $j\in\{1:n\}$ satisfies
$\boldsymbol \xi_p^\star \in \boldsymbol V_p({\TT^e})$. Since the minimizing sets in~\eqref{eq_stab_shift}
are nonempty (they all contain $\boldsymbol \xi_p^{\star}$), both the discrete and the continuous minimizers
are uniquely defined owing to standard convexity arguments. Let us set
\begin{equation*}
\boldsymbol v^\star := \argmin{\boldsymbol v \in \boldsymbol V({\TT^e})} \|\boldsymbol v\|_{{\omega_\edge}},
\qquad \boldsymbol v^\star_j := \boldsymbol v^\star|_{K_j}, \quad j\in\{1:n\}.
\end{equation*}
To prove Proposition~\ref{prop_stability_patch}, it is enough to show that
\begin{equation} \label{eq res}
\|\boldsymbol \xi_p^\star\|_{{\omega_\edge}} \lesssim \|\boldsymbol v^\star\|_{{\omega_\edge}}.
\end{equation}
Owing to Proposition~\ref{prop_stability_tetrahedra} applied with $K:= K_j$ and $\mathcal F:=\mathcal F_j$
for all $j\in\{1:n\}$, we have
\begin{equation} \label{eq_ineq_K}
\|\boldsymbol \xi_p^\star\|_{K_j} \lesssim
\min_{\boldsymbol \zeta \in \boldsymbol V(K_j)} \|\boldsymbol \zeta\|_{K_j},
\end{equation}
where $\boldsymbol V(K_{j})$ is defined in~\eqref{eq_VKj}.
Therefore, recalling that $|{\TT^e}|\lesssim 1$,
\eqref{eq res} will be proved if for all $j\in\{1:n\}$, we can construct a field
$\boldsymbol \zeta_j \in \boldsymbol V(K_j)$ such that
$\|\boldsymbol \zeta_j\|_{K_j} \revision{\lesssim} \|\boldsymbol v^\star\|_{{\omega_\edge}}$.
To do so, we proceed once again by induction.
{\bf (1)} First element ($j=1$).
Since $\boldsymbol v^\star_1 \in \boldsymbol V(K_1)$, the claim is established with
$\boldsymbol \zeta_1:= \boldsymbol v^\star_1$ which trivially satisfies
$\|\boldsymbol \zeta_1\|_{K_1} = \|\boldsymbol v^\star\|_{K_1} \leq \|\boldsymbol v^\star\|_{{\omega_\edge}}$.
{\bf (2)} Middle elements ($j\in\{2:n-1\}$).
We proceed by induction. Given $\boldsymbol \zeta_{j-1}\in {\boldsymbol V}(K_{j-1})$ such that
$\|\boldsymbol \zeta_{j-1}\|_{K_{j-1}} \lesssim \|\boldsymbol v^\star\|_{{\omega_\edge}}$,
let us construct a suitable $\boldsymbol \zeta_{j}\in {\boldsymbol V}(K_{j})$
such that $\|\boldsymbol \zeta_j\|_{K_j} \lesssim \|\boldsymbol v^\star\|_{{\omega_\edge}}$.
We consider the affine geometric mapping $\boldsymbol T_{j-1,j}: K_{j-1} \to K_{j}$
that leaves the three vertices $\aaa_{\rm d}$, $\boldsymbol a_{j-1}$, and $\aaa_{\rm u}$
(and consequently the face $F_{j-1}$) invariant, whereas $\boldsymbol T_{j-1,j}(\boldsymbol a_{j-2}) = \boldsymbol a_{j}$.
We denote by $\boldsymbol \psi^{\mathrm{c}}_{j-1,j} := \boldsymbol \psi^{\mathrm{c}}_{\boldsymbol T_{j-1,j}}$
the associated Piola mapping, see Section~\ref{sec_Piola}.
Let us define the function $\boldsymbol \zeta_{j} \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K_{j})$ by
\begin{equation} \label{eq_zeta_j}
\boldsymbol \zeta_{j} := \boldsymbol v^\star_{j} - \epsilon_{j-1,j} \boldsymbol \psi^{\mathrm{c}}_{j-1,j}(\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1}),
\end{equation}
where $\epsilon_{j-1,j} := \operatorname{sign} \left (\det \mathbb J_{\boldsymbol T_{j-1,j}}\right)$
\revision{(notice that here $\epsilon_{j-1,j}=-1$)}.
Using the triangle inequality, the $L^2$-stability of the Piola mapping
(see~\eqref{eq_stab_piola_L2}), inequality~\eqref{eq_ineq_K}, and the induction hypothesis,
we have
\begin{equation} \label{eq_bound} \begin{split}
\|\boldsymbol \zeta_{j}\|_{K_j}
&\leq
\|\boldsymbol v^\star\|_{K_j} + \|\boldsymbol \psi^{\mathrm{c}}_{j-1,j}(\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1})\|_{K_j}
\\
&\lesssim
\|\boldsymbol v^\star\|_{K_j} + \|\boldsymbol \xi_p^\star - \boldsymbol v^\star\|_{K_{j-1}}
\\
&\leq
\|\boldsymbol v^\star\|_{K_j} + \|\boldsymbol \xi_p^\star\|_{K_{j-1}} + \|\boldsymbol v^\star\|_{K_{j-1}}\\
&\lesssim
\|\boldsymbol v^\star\|_{K_j} + \|\boldsymbol \zeta_{j-1}\|_{K_{j-1}} + \|\boldsymbol v^\star\|_{K_{j-1}}
\lesssim
\|\boldsymbol v^\star\|_{{\omega_\edge}}.
\end{split}\end{equation}
Thus it remains to establish that $\boldsymbol \zeta_{j} \in \boldsymbol V(K_{j})$
in the sense of Definition~\ref{def_spaces_elm}, i.e., we need to show that
$\grad \times \boldsymbol \zeta_{j}=\boldsymbol r_{K_{j}}$ and $\boldsymbol \zeta_{j}|^{\boldsymbol \tau}_{\mathcal F_{j}} = \widetilde \rr^{j}_{\mathcal F_{j}}$.
Recalling the curl constraints in~\eqref{eq_VTe} and~\eqref{eq_VKj} which yield
$\grad \times \boldsymbol \xi_p^\star = \grad \times \boldsymbol v^\star = \boldsymbol r_{\mathcal T}$ and using~\eqref{eq_piola_commute}, we have
\begin{equation}
\label{tmp_curl_vv_middle}
\begin{split}
\grad \times \boldsymbol \zeta_{j}
&=
\grad \times \boldsymbol v^\star_{j} - \epsilon_{j-1,j}\grad \times \boldsymbol \psi^{\mathrm{c}}_{j-1,j}(\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1})
\\
&=
\boldsymbol r_{K_{j}} - \epsilon_{j-1,j} \boldsymbol \psi^{\mathrm{d}}_{j-1,j}\left (\grad \times (\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1})\right )
=
\boldsymbol r_{K_{j}},
\end{split} \end{equation}
which proves the expected condition on the curl of $\boldsymbol \zeta_j$.
It remains to verify the weak tangential trace condition
$\boldsymbol \zeta_{j}|^{\boldsymbol \tau}_{\mathcal F_{j}} = \widetilde \rr^{j}_{\mathcal F_{j}}$
as per Definition~\ref{definition_partial_trace}.
To this purpose, let $\boldsymbol \phi \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F_{j}^{\mathrm{c}}}(K_{j})$
and define $\widetilde{\pphi}$ by
\begin{equation}\label{eq_thpi}
\widetilde{\pphi}|_{K_{j}} := \boldsymbol \phi \quad \widetilde{\pphi}|_{K_{j-1}} := (\boldsymbol \psi_{j-1,j}^{\mathrm{c}})^{-1}(\boldsymbol \phi),
\quad
\widetilde{\pphi}|_{K_l} = \boldsymbol 0 \quad \forall l \in\{1:n\}\setminus\{ j-1,j\}.
\end{equation}
These definitions imply that
$\widetilde{\pphi} \in \boldsymbol H(\boldsymbol{\operatorname{curl}},{\omega_\edge}) \cap \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e})$
(recall~\eqref{eq_Htpatch}) with
\begin{equation*}
\left . \left (\widetilde{\pphi}|_{K_{j-1}} \right ) \right |_{F_{j-1}} \times \boldsymbol n_{F_{j-1}}
=
\left . \left (\widetilde{\pphi}|_{K_{j}} \right ) \right |_{F_{j-1}} \times \boldsymbol n_{F_{j-1}}
=
\boldsymbol \phi|_{F_{j-1}} \times \boldsymbol n_{F_{j-1}},
\end{equation*}
as well as
\begin{equation*}
\widetilde{\pphi}|_F \times \boldsymbol n_F = \boldsymbol 0 \quad \forall F \in {\FF^e} \setminus \{F_{j-1}\}.
\end{equation*}
(Note that $\widetilde{\pphi}|_F\times\boldsymbol n_F$ is uniquely defined by assumption.)
Recalling definition~\eqref{eq_zeta_j} of $\boldsymbol \zeta_j$ and that
$\grad \times \boldsymbol \zeta_j = \boldsymbol r_{K_{j}} = \grad \times\boldsymbol v^\star_{j}$, see~\eqref{tmp_curl_vv_middle}, we have
\begin{align*}
&
(\grad \times \boldsymbol \zeta_j,\boldsymbol \phi)_{K_{j}}
-
(\boldsymbol \zeta_j,\grad \times \boldsymbol \phi)_{K_{j}}
\\
&=
(\grad \times \boldsymbol v^\star,\boldsymbol \phi)_{K_{j}}
-
(\boldsymbol v^\star,\grad \times \boldsymbol \phi)_{K_{j}}
+
\epsilon_{j-1,j}
(\boldsymbol \psi^{\mathrm{c}}_{j-1,j} (\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1}),\grad \times \boldsymbol \phi)_{K_{j}}
\\
&=
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_{j}}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_{j}}
+ (\boldsymbol \xi_p^\star - \boldsymbol v^\star,\grad \times \widetilde{\pphi} )_{K_{j-1}},
\end{align*}
where we used the definition of $\widetilde{\pphi}$, properties~\eqref{eq_piola_adjoint},
\eqref{eq_piola_commute} of the Piola mapping, and the definition of $\epsilon_{j-1,j}$
to infer that
\begin{align*}
\epsilon_{j-1,j}
(\boldsymbol \psi^{\mathrm{c}}_{j-1,j} (\boldsymbol \xi_p^{\star j-1} - \boldsymbol v^\star_{j-1}),\grad \times \boldsymbol \phi)_{K_{j}}
&=
\epsilon_{j-1,j}^2
(\boldsymbol \xi_p^\star - \boldsymbol v^\star,\grad \times \left ((\boldsymbol \psi^{\mathrm{c}}_{j-1,j})^{-1}\boldsymbol \phi|_{K_j} \right ))_{K_{j-1}}.
\\
&=
(\boldsymbol \xi_p^\star - \boldsymbol v^\star,\grad \times \widetilde{\pphi} )_{K_{j-1}}.
\end{align*}
Since $\grad \times \boldsymbol \xi_p^\star = \boldsymbol r_{\mathcal T} = \grad \times \boldsymbol v^\star$ and
$\widetilde{\pphi} = \boldsymbol 0$ outside $K_{j-1} \cup K_{j}$, this gives
\begin{align}
\label{tmp_middle_trace0}
&
(\grad \times \boldsymbol \zeta_j,\boldsymbol \phi)_{K_{j}}
-
(\boldsymbol \zeta_j,\grad \times \boldsymbol \phi)_{K_{j}}
\\
\nonumber
&=
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_{j}}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_{j}}
+
(\boldsymbol \xi_p^\star - \boldsymbol v^\star,\grad \times \widetilde{\pphi} )_{K_{j-1}}
+
(\grad \times (\boldsymbol v^\star-\boldsymbol \xi_p^\star),\widetilde{\pphi})_{K_{j-1}}
\\
\nonumber
&= \sum_{K \in {\TT^e}}
\left \{
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_K - (\boldsymbol v^\star,\grad \times \widetilde{\pphi})_K
\right \}
-
\left (
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_{j-1}} - (\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_{j-1}}
\right ).
\end{align}
Since $\boldsymbol v^\star \in \boldsymbol V({\TT^e})$, $\widetilde{\pphi} \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e})$,
and $\jump{\boldsymbol v^\star}^{\boldsymbol \tau}_\mathcal F=\boldsymbol r_\mathcal F$, we have from Definitions~\ref{definition_jumps}
and~\ref{def_spaces}
\begin{align}
\label{tmp_middle_trace1}
\sum_{K \in {\TT^e}}
\left \{
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_K - (\boldsymbol v^\star,\grad \times \widetilde{\pphi})_K
\right \}
&=
\sum_{F \in \mathcal F} (\boldsymbol r_F,\widetilde{\pphi} \times \boldsymbol n_F)_{F}
\\
\nonumber
&=
(\boldsymbol r_{F_{j-1}},\boldsymbol \phi \times\boldsymbol n_{F_{j-1}})_{F_{j-1}},
\end{align}
where in the last equality, we employed the definition~\eqref{eq_thpi} of $\widetilde{\pphi}$.
On the other hand, since $\boldsymbol \xi^\star_p|_{K_{j-1}},\widetilde{\pphi}|_{K_{j-1}} \in \boldsymbol H^1(K_{j-1})$,
we can employ the pointwise definition of the trace and infer that
\begin{align}
\label{tmp_middle_trace2}
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_{j-1}} - (\boldsymbol \xi_p^{\star},\grad \times \widetilde{\pphi})_{K_{j-1}}
&=
(\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^{\star j-1}),\widetilde{\pphi}|_{K_{j-1}} \times \boldsymbol n_{K_{j-1}})_{F_{j-1}}
\\
\nonumber
&=
-(\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^{\star j-1}),\boldsymbol \phi \times \boldsymbol n_{F_{j-1}})_{F_{j-1}},
\end{align}
where we used that $\boldsymbol n_{K_{j-1}} = -\boldsymbol n_{F_{j-1}}$.
Then, plugging~\eqref{tmp_middle_trace1} and~\eqref{tmp_middle_trace2}
into~\eqref{tmp_middle_trace0} and employing~\eqref{eq_F_j} and $\boldsymbol n_{K_{j}} = \boldsymbol n_{F_{j-1}}$,
we obtain
\begin{align*}
(\grad \times \boldsymbol \zeta_j,\boldsymbol \phi)_{K_{j}}
-
(\boldsymbol \zeta_j,\grad \times \boldsymbol \phi)_{K_{j}}
&=
(\boldsymbol r_{F_{j-1}},\boldsymbol \phi \times\boldsymbol n_{F_{j-1}})_{F_{j-1}}
+(\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^{\star j-1}),\boldsymbol \phi \times \boldsymbol n_{F_{j-1}})_{F_{j-1}}
\\
&=
(\boldsymbol r_{F_{j-1}}+\boldsymbol \pi^{\boldsymbol \tau}_{F_{j-1}} (\boldsymbol \xi_p^{\star j-1}),\boldsymbol \phi \times \boldsymbol n_{F_{j-1}})_{F_{j-1}}
\\
&=
(\widetilde \rr^{j}_{F_{j-1}},\boldsymbol \phi \times \boldsymbol n_{K_{j}})_{F_{j-1}}.
\end{align*}
Since $\mathcal F_j:= \{F_{j-1}\}$, this shows that $\boldsymbol \zeta_j$ satisfies the weak tangential
trace condition in $\boldsymbol V(K_{j})$ by virtue of Definition~\ref{definition_partial_trace}.
{\bf (3)} Last element ($j=n$). We need to distinguish the type of patch.
{\bf (3a)} Patch of Dirichlet or mixed boundary type. In this case, we can employ the same argument as
for the middle elements since $\mathcal F_n=\{F_{n-1}\}$ is composed of only one face.
{\bf (3b)} Patch of interior type. Owing to the induction hypothesis,
we have $\|\boldsymbol \zeta_j\|_{K_j} \lesssim \|\boldsymbol v^\star\|_{{\omega_\edge}}$ for all $j\in\{1:n-1\}$.
Let us first assume that there is an even number of tetrahedra in the patch ${\TT^e}$, as in
Figure~\ref{fig_patch}, left. The case where this number is odd will be discussed below.
We build a geometric mapping $\boldsymbol T_{j,n}:K_j\to K_n$ for all $j\in\{1:n-1\}$ as follows:
$\boldsymbol T_{j,n}$ leaves the edge $e$ pointwise invariant, $\boldsymbol T_{j,n}(\boldsymbol a_{j-1}):= \boldsymbol a_{n}$,
$\boldsymbol T_{j,n}(\boldsymbol a_j):= \boldsymbol a_{n-1}$ if $(n-j)$ is odd, and $\boldsymbol T_{j,n}(\boldsymbol a_{j}):= \boldsymbol a_{n}$,
$\boldsymbol T_{j,n}(\boldsymbol a_{j-1}):= \boldsymbol a_{n-1}$ if $(n-j)$ is even. Since $n$ is by assumption even,
one readily sees that $\boldsymbol T_{j,n}(F_j)=\boldsymbol T_{j+1,n}(F_j)$ with $F_j=K_j\cap K_{j+1}$ for all
$j\in\{1;n-2\}$.
We define $\boldsymbol \zeta_n \in \boldsymbol H(\boldsymbol{\operatorname{curl}},K_n)$ by setting
\begin{equation}
\label{eq_v_def}
\boldsymbol \zeta_n := \boldsymbol v^\star_n -
\sum_{j\in\{1:n-1\}} \epsilon_{j,n} \boldsymbol \psi^{\mathrm{c}}_{j,n}(\boldsymbol \xi_p^{\star j}-\boldsymbol v^\star_j),
\end{equation}
where $\epsilon_{j,n} := \operatorname{sign}(\det \mathbb J_{\boldsymbol T_{j,n}})$ and $\boldsymbol \psi^{\mathrm{c}}_{j,n}$ is the
Piola mapping associated with $\boldsymbol T_{j,n}$.
Reasoning as above in~\eqref{eq_bound} shows that
\begin{equation*}
\|\boldsymbol \zeta_n\|_{K_n} \lesssim \|\boldsymbol v^\star\|_{{\omega_\edge}}.
\end{equation*}
It now remains to establish that $\boldsymbol \zeta_{n} \in \boldsymbol V(K_{n})$
as per Definition~\ref{def_spaces_elm},
i.e. $\grad \times \boldsymbol \zeta_{n}=\boldsymbol r_{K_{n}}$ and $\boldsymbol \zeta_{n}|^{\boldsymbol \tau}_{\mathcal F_{n}} = \widetilde \rr^{n}_{\mathcal F_{n}}$
with $\mathcal F_n:=\{F_{n-1},F_n\}$. Since $\grad \times \boldsymbol \xi_p^\star = \boldsymbol r_{\mathcal T} = \grad \times \boldsymbol v^\star$,
using~\eqref{eq_piola_commute} leads to $\grad \times \boldsymbol \zeta_n = \grad \times \boldsymbol v^\star_n = \boldsymbol r_{K_n}$
as above in~\eqref{tmp_curl_vv_middle}, which proves the expected condition on the curl of
$\boldsymbol \zeta_n$. It remains to verify the weak tangential trace condition as per
Definition~\ref{definition_partial_trace}. To this purpose, let
$\boldsymbol \phi \in \boldsymbol H_{{\boldsymbol \tau},\mathcal F_n^{\mathrm{c}}}^1(K_n)$ and define $\widetilde{\pphi}$ by
\begin{equation}
\label{eq_ttet}
\widetilde{\pphi}|_{K_n} := \boldsymbol \phi, \qquad
\widetilde{\pphi}|_{K_j} := \left (\boldsymbol \psi_{j,n}^{\mathrm{c}}\right )^{-1}(\boldsymbol \phi)
\quad \forall j\in\{1:n-1\}.
\end{equation}
As $\boldsymbol \phi \in \boldsymbol H_{{\boldsymbol \tau},\mathcal F_n^{\mathrm{c}}}^1(K_n)$, its trace is defined in
a strong sense, and the preservation of tangential traces by Piola mappings shows
that $\widetilde{\pphi} \in \boldsymbol H^1_{{\boldsymbol \tau},\mathcal F^{\mathrm{c}}}({\TT^e})$ in the sense of~\eqref{eq_Htpatch}.
Then, using $\grad \times \boldsymbol \zeta_n = \grad \times \boldsymbol v^\star_n$ and~\eqref{eq_v_def}, we have
\begin{align*}
&
(\grad \times \boldsymbol \zeta_n,\boldsymbol \phi)_{K_n}
-
(\boldsymbol \zeta_n,\grad \times \boldsymbol \phi)_{K_n}
\\
&=
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_n}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_n}
+
\sum_{j\in\{1:n-1\}} \epsilon_{j,n}
(\boldsymbol \psi^{\mathrm{c}}_{j,n}(\boldsymbol \xi^{\star j}_p - \boldsymbol v^\star_{j}),\grad \times \boldsymbol \phi)_{K_n},
\end{align*}
where we used the definition of $\widetilde{\pphi}$ for the first two terms on the right-hand side.
Moreover, using~\eqref{eq_piola_adjoint} and~\eqref{eq_piola_commute} for all $j\in\{1:n-1\}$,
we have
\begin{align*}
\epsilon_{j,n}
(\boldsymbol \psi^{\mathrm{c}}_{j,n}(\boldsymbol \xi^{\star j}_p - \boldsymbol v^\star_{j}),\grad \times \boldsymbol \phi)_{K_n}
&=
\epsilon_{j,n}^2
(\boldsymbol \xi^\star_p - \boldsymbol v^\star,\grad \times ((\boldsymbol \psi^{\mathrm{c}}_{j,n})^{-1}(\boldsymbol \phi|_{K_n})))_{K_j}
\\
&=
(\boldsymbol \xi^\star_p - \boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_j}
\\
&=
(\boldsymbol \xi^\star_p - \boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_j} - (\grad \times (\boldsymbol \xi^\star_p - \boldsymbol v^\star),\widetilde{\pphi})_{K_j},
\end{align*}
since $\grad \times \boldsymbol \xi^\star_p = \boldsymbol r_\mathcal T = \grad \times \boldsymbol v^\star$.
It follows that
\begin{align*}
& (\grad \times \boldsymbol \zeta_n,\boldsymbol \phi)_{K_n}
-
(\boldsymbol \zeta_n,\grad \times \boldsymbol \phi)_{K_n}\\
&=
\sum_{j\in\{1:n\}}
\left \{
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}
-
\sum_{j\in\{1:n-1\}}
\left \{
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}
\\
&=
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_n}
-
(\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_n}
+
\sum_{j\in\{1:n\}}
\left \{
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}
\\
& \qquad -
\sum_{j\in\{1:n\}}
\left \{
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}
\\
&=
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_n}
-
(\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_n}
=
(\grad \times \boldsymbol \xi_p^\star,\boldsymbol \phi)_{K_n}
-
(\boldsymbol \xi_p^\star,\grad \times \boldsymbol \phi)_{K_n}\\
&= \sum_{F\in\mathcal F_n} (\widetilde \rr^{n}_{F},\boldsymbol \phi \times \boldsymbol n_{K_n})_{F},
\end{align*}
where we employed the fact that, since both $\boldsymbol \xi_p^\star,\boldsymbol v^\star \in \boldsymbol V({\TT^e})$,
Definition~\ref{definition_jumps} gives
\begin{align*}
\sum_{j\in\{1:n\}}
\left \{
(\grad \times \boldsymbol v^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol v^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}
&=
\sum_{F \in \mathcal F} (\boldsymbol r_F,\widetilde{\pphi} \times \boldsymbol n_F)_{F}
\\
&=
\sum_{j\in\{1:n\}}
\left \{
(\grad \times \boldsymbol \xi_p^\star,\widetilde{\pphi})_{K_j}
-
(\boldsymbol \xi_p^\star,\grad \times \widetilde{\pphi})_{K_j}
\right \}.
\end{align*}
Thus $\boldsymbol \zeta_n|^{\boldsymbol \tau}_{\mathcal F_n} = \widetilde \rr^n_{\mathcal F_n}$ in the sense of
Definition~\ref{definition_partial_trace}. This establishes the weak tangential
trace condition on $\boldsymbol \zeta_n$ when $n$ is even.
If $n$ is odd, one can proceed as
in~\cite[Section~6.3]{Ern_Voh_p_rob_3D_20}. For the purpose of the proof only,
one tetrahedron different from $K_n$ is subdivided into two subtetrahedra as
in~\cite[Lemma~B.2]{Ern_Voh_p_rob_3D_20}. Then, the above construction of $\boldsymbol \zeta_n$
can be applied on the newly created patch which has an even number of elements, and one
verifies as above that $\boldsymbol \zeta_n\in {\boldsymbol V}(K_n)$.
{\bf (3c)} Patch of Neumann boundary type. In this case, a similar argument as for
a patch of interior type applies, and we omit the proof for the sake of brevity.
\end{proof}
\begin{remark}[Quasi-optimality of $\boldsymbol \xi_p^\star$] \label{rem_sweep_brok}
Let $\boldsymbol \xi_p^\star\in{\boldsymbol V}_p({\TT^e})$ be defined in the above proof
(see in particular~\eqref{eq_min_K}).
Since $\|\boldsymbol v^\star\|_{{\omega_\edge}} \le \min_{\boldsymbol v_p\in {\boldsymbol V}_p({\TT^e})}\|\boldsymbol v_p\|_{{\omega_\edge}}$,
inequality~\eqref{eq res} implies that
$\|\boldsymbol \xi_p^\star\|_{{\omega_\edge}} \lesssim
\min_{\boldsymbol v_p\in {\boldsymbol V}_p({\TT^e})}\|\boldsymbol v_p\|_{{\omega_\edge}}$
(note that the converse inequality is trivial with constant one).
This elementwise minimizer is the one used in \revision{Theorem~\ref{thm_sweep} and in}
the simplified a posteriori error estimator~\eqref{eq_definition_estimator_sweep_2}.
\end{remark}
\bibliographystyle{amsplain}
| 018387edce371ab1c01db174473b9492905bb46e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A $^3$He spin filter is a neutron polarization device in which polarized $^3$He gas is enclosed in a special glass cell that does not contain boron. $^3$He spin filters are widely used to polarize neutrons and analyze neutron spins in neutron scattering experiments because of its capabilities to polarize neutrons in broad energy range and cover wide solid angle~\cite{Babcock2016, Dhiman2017, Chen2017, Berna2007, Lee2016, Beecham2011}. Since $^3$He nuclei have a strongly spin-dependent neutron absorption cross section, neutrons passing through polarized $^3$He gas are polarized. $^3$He nuclei are polarized using laser techniques known as Spin Exchange Optical Pumping (SEOP)~\cite{SEOP} or Metastability Optical Pumping (MEOP)~\cite{MEOP}. In SEOP, $^3$He gas, N$_2$ gas, and pure rubidium or rubidium and potassium are encapsulated within a glass cell. The Rb or Rb-K cells are heated to around 170$^\circ$C or 220$^\circ$C, respectively, to evaporate the alkali metals, and circularly polarized laser light irradiates the cell. Unpaired electrons of alkali atoms are polarized by optical pumping, and the spins of the electrons are exchanged with those of $^3$He nuclei by hyperfine interaction. A diode laser array with output power on the order of 100 W and a wavelength of 794.7~nm, which corresponds to the D$_1$ absorption line of Rb, is typically employed. The $^3$He polarization can be improved by using Rb-K compared to the pure Rb case because of the high optical pumping efficiency of Rb-K~\cite{Hybrid}. When the laser irradiation is stopped, the $^3$He polarization exponentially decays due to non-uniformity of the surrounding magnetic field, collisions of $^3$He atom with each other, and collisions of $^3$He atoms with impurities in the glass cell and a glass wall.
{\it Ex-situ} and {\it {\it in-situ}} methods are used for $^3$He spin filters with SEOP. In the {\it ex-situ} method, a $^3$He cell is polarized in a pumping station away from a neutron beamline, transferred in a magnetic cavity to keep the $^3$He polarization, and installed to a neutron beamline. Since the $^3$He polarization decays during a neutron experiment, the $^3$He cell needs to be periodically re-polarized and re-installed for long-running experiments. However, the advantage of using {\it ex-situ} method is that it can be used for a neutron beamline with limited space, and the installation of the instrument is straightforward. This is beneficial at spallation neutron sources, which generally have small experimental spaces. As there are many
experiments with short measuring time at intense neutron beam facilities, those may be done without re-polarizing the $^3$He cell. In order to keep high $^3$He polarization during the experiment, a long relaxation time of $^3$He polarization is required, which makes details of the fabrication process of the $^3$He cells important.
On the other hand, in the {\it in-situ} method, the $^3$He cell is installed along with a laser system onto a neutron beamline. This method is suitable for longer experiments because a stable $^3$He polarization can be maintained for weeks. However, a larger space compared to the space required for the {\it ex-situ} method is needed, and the safety regulations involving the use of high powered lasers on a neutron beamline must be considered. For details of the technique for optically polarized $^3$He and its application, please see a reviewed paper Ref~\cite{Gentile2017}.\\
At MLF of J-PARC, world class intense pulsed neutron beams are provided to 23 neutron beamlines, and many neutron scattering experiments are conducted~\cite{Nakajima2017}. In Japan, the fundamental development of $^3$He polarization technique based on SEOP began in the 1990s~\cite{Sato1994}, and further development is still ongoing by Ino et al. at High Energy Accelerator Research Organization~\cite{Ino2005, Ino2009, Ino2016, Ino2017, InoNMR}. An {\it in-situ} SEOP system dedicated for a polarized neutron spectrometer
POLANO~\cite{Nakajima2017, POLANO} at beamline No. 23 has been developed, and now under commissioning~\cite{Ino2016, Ino2017}. In order to promote user experiments using $^3$He spin filters on the other beamlines at MLF, we are carrying out the development of a $^3$He spin filters with both {\it in-situ} and {\it ex-situ} methods for their versatile uses on site at J-PARC. Experiments demonstrating their use have been performed on several beamlines. User experiments and fabrication of $^3$He cells have recently begun. In this paper, we report development and utilization of $^3$He spin filters at MLF of J-PARC.
\section{gas-filling station}
As mentioned in the previous section, a clean gas-filling station and fabrication process without impurities are important for a $^3$He cell with long relaxation time. The first gas-filling station was constructed in 2018 at J-PARC. The schematic diagram of the gas-filling station is shown in Fig.~\ref{fig:gasstation}. All the gas lines are wrapped with heaters for vacuum bake-out. An ultimate pressure of the gas-filling station was $5 \times 10^{-8}$~Pa after baking at 150$^\circ$C for a few days. \\
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{FillingStation.eps}
\caption{The schematic diagram of the gas-filling station to fabricate $^3$He cells.}
\label{fig:gasstation}
\end{figure}
Boron-free aluminosilicate GE180 glass is widely used for the glass cell of a $^3$He spin filter due to a small wall relaxation effect and low permeability of $^3$He~\cite{Jianga2013, Chen2011,Berna2007, Salhi2014}. Our fabrication process of $^3$He cells is similar to those of NIST, JCNS, and ORNL~\cite{Jianga2013,Salhi2014, Chen2011}. A cylindrical-shaped glass cell made of GE180 is attached to a glassware made of Pyrex glass, which is referred to as "string", (Fig.~\ref{fig:string}). The glass cell and the string are rinsed with neutral detergent, pure water, acetone, and alcohol before connecting to the gas-filling station. Rubidium and potassium ampoules are also rinsed with acetone and alcohol. After connecting the string to the gas-filling station, the ampoules are put in the string without breaking the ampoules, and the string is sealed off with a gas torch. The ampoules are broken in a nitrogen atmosphere by dropping hammers made of an iron rod covered with glass. The string and the glass cell are baked out for a week at 200$^\circ$C and 400$^\circ$C, respectively. In the next step, rubidium and potassium are distilled to retorts at 200$^\circ$C and 220$^\circ$C, respectively, by using the heaters. After that, the glass tube parts containing ampoules and hammers are pulled off with a gas torch. The glass cell and the string are baked out again for one week, and then the alkali metals in the retorts are distilled to the glass cell. Finally, $^3$He and N$_2$ gases purified using GC50 getters (SAES Getters) are filled to the glass cell, and the glass cell is sealed off by placing a gas torch at the Pyrex part. When encapsulating the gases above 1~atm, the glass cell is submerged in liquid nitrogen to keep the pressure inside the glass cell below 1~atm while sealing. So far, nine $^3$He cells have been fabricated with $^3$He pressures of 3.1~atm
using the gas-filling station at J-PARC, and four of them have found to have spin relaxation times over 150~hours. A list of these $^3$He cells and a photograph of Sekichiku cell are presented in Table~\ref{celllist} and Fig.~\ref{fig:cell}, respectively.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{string.eps}
\caption{The glassware used to fabricate a $^3$He cell. The ampoules are put in the parts labeled as K ampoule and Rb ampoule in the figure. The glass cell made of GE180 is attached to the string made of Pyrex at the top of the glass cell. The string is connected to the pumping station using a Swagelok connector at the metal tube. }
\label{fig:string}
\end{figure}
\begin{table}[h]
\caption{List of $^3$He cells using rubidium and potassium with long relaxation times at J-PARC. }
\begin{ruledtabular}
\begin{tabular}{cccc}
Name & Dimensions [mm]& $^3$He pressure [atm] & $T_1$ [h]\\
\hline
Chidori & $\phi$45 $\times$ 75 & 3.1 & 201\\
Karigane & $\phi$45 $\times$ 75 & 3.1 & 180\\
Sekichiku & $\phi$60 $\times$ 60 & 3.1 & 175\\
Hanabishi & $\phi$40 $\times$ 90 & 3.1 & 165\\
\end{tabular}
\end{ruledtabular}
\label{celllist}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=5cm,clip]{cell.eps}
\caption{$^3$He cell (Sekichiku) fabricated at J-PARC. }
\label{fig:cell}
\end{figure}
\section{NMR and Pumping systems}
\subsection{$B_0$ coil and NMR system}
To keep the $^3$He polarization, a coil with double magnetic shields, referred to as $B_0$ coil, is mainly used for the laser pumping~\cite{Kira2013}. Figure~\ref{fig:coils} shows a cross-sectional view of the $B_0$ coil and a photograph of its interior. The coil is consisting of a main solenoid, two compensation coils, and two side coils. The dimensions of the $B_0$ coil are 29.2~cm in diameter and 36.8~cm in length.
The $B_0$ coil was designed using the finite element method to make the field gradient less than 5$\times 10^{-4}$~/cm in a cylindrical region with a diameter of 10~cm and a length of 10~cm at the center of the coil. A magnetic field of 1.5~mT is produced to keep the $^3$He polarization. Frequency-sweep Adiabatic Fast Passage(AFP)-NMR was employed to flip the $^3$He spin as well as to evaluate the $^3$He polarization~\cite{InoNMR}. A cosine winding is used as a drive coil to apply an oscillating magnetic field to the $^3$He cell. The $^3$He precession signal is detected by a pick-up coil located under the $^3$He cell. The coils were installed inside the $B_0$ coil as shown in Fig.~\ref{fig:coils}. The polarization loss due to spin flip was measured as 3.8$\times$10$^{-5}$ per flip. The $^3$He cell is heated by electrical heating using a rubber heater with a power of 100~W wound around the drive coil during optical pumping.
\begin{figure*}[htb]
\centering
\includegraphics[width=15cm,clip]{coils.eps}
\caption{(a) Cross-sectional view of the $B_0$ coil and the NMR coils. The magnetic shields are made from permalloy and its thicknesses are 2~mm. A static magnetic field is generated with the $B_0$ coil to keep $^3$He polarization. The drive coil is used to generate an oscillating magnetic field to flip $^3$He spins with AFP-NMR, and the spin flip signal is detected with the pick-up coil. (b) A photograph of the inside of the coil with the $^3$He cell and the NMR coils.}
\label{fig:coils}
\end{figure*}
\subsection{Pumping station at MLF}
We have set up a laboratory in the MLF beam hall to polarize $^3$He cells. $^3$He cells polarized at the laboratory can be provided to neutron beamlines in MLF. A pumping system using an air-cooling laser module with a laser power of 30\,W~\cite{Oku2015} has been installed in the laboratory (Fig.~\ref{fig:laser}).
The dimensions of the pumping system are 60~cm $\times$ 60~cm $\times$ 40~cm, and it is used for both {\it in-situ} and {\it ex-situ} methods.
Using a volume Bragg grating (VBG) element in the laser module, the center wavelength of the laser is tuned to 794.7~nm with a line width of 0.2~nm in Full Width at Half Maximum (FWHM). The laser light is circularly polarized using a quarter-wavelength plate installed in the module. The polarized laser light from the laser module is reflected by dielectric multi-layer coated quartz mirrors and irradiates the $^3$He cell. The temperature at the cell surface is kept about 170$^\circ$C for a pure Rb cell and about 210$^\circ$C for a cell with rubidium and potassium during optical pumping. A $^3$He polarization of $\sim$70\% is achieved with this laser system. A polarization measurement system using Electron Paramagnetic Resonance (EPR)~\cite{EPR} is also equipped.
\begin{figure*}[hbt]
\centering
\includegraphics[width=14cm,clip]{lasers.eps}
\caption{(a) Compact pumping system which can be installed on a neutron beamline. The optical system has been constructed in the laser shield box. (b) Air cooling laser module. The laser power is 30~W.}
\label{fig:laser}
\end{figure*}
\subsection{Pumping system using a fiber laser}
In order to achieve higher $^3$He polarization, we have developed a separate pumping system using a fiber laser made by DILAS Diode Laser, Inc. (M1F4S22-794.7.[0,6]-100C-IS9), as shown in Fig.~\ref{fig:Fiberlaser}. The pumping system using this fiber laser has been constructed at another laboratory in the J-PARC research building, about 30\,m away from the MLF experimental hall. This fiber laser is an air-cooled diode laser with a maximum output power of 110~W. The center wavelength is tuned to 794.7~nm with a line width of 0.4~nm in FWHM using a VBG element. An optical fiber with an internal diameter of 400~$\rm{\mu}$m is connected to the laser shield box. An unpolarized laser beam from the fiber exit is split into two paths by a polarized beam splitter. The laser beams are circularly polarized with quarter-wavelength plates. The polarized laser beams irradiate both sides of a $^3$He cell. The degree of circular polarization in each path was measured using a polarimeter as more than 98\% at the $^3$He cell position. The dimensions of the pumping system are 70~cm $\times$ 70~cm $\times$ 40~cm, and it can also be used for the {\it in-situ} method. The pumping system is currently used only for the {\it ex-situ} method, and will be installed in the laboratory at MLF in the near future after following a safety review.
\begin{figure*}[hbt]
\centering
\includegraphics[width=16cm,clip]{FilberLaser.eps}
\caption{(a) Schematic diagram of the pumping system using a fiber laser. (b) Photograph of the pumping system. The optical system was constructed in the laser shield box.}
\label{fig:Fiberlaser}
\end{figure*}
\section{Performance evaluation of the $^3$He spin filter}
We conducted an experiment at the NOBORU beamline~\cite{Nakajima2017, NOBORU}, which is a test neutron beamline for general purpose, to evaluate the $^3$He polarization for the Chidori cell, which has the longest $T_1$ as shown in Table~\ref{celllist}. The $^3$He cell was polarized using the pumping system with the fiber laser at J-PARC research building. Time evolution of the $^3$He polarization during optical pumping is shown in Fig.~\ref{fig:pumping}.
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm,clip]{pumping.eps}
\caption{Time evolution of the $^3$He polarization measured by AFP-NMR during the optical pumping. The $^3$He polarization saturates at about 24 hours after the start of the laser irradiation. The temperature at the surface of the cell was 205$^\circ$C during optical pumping.}
\label{fig:pumping}
\end{figure}
After the $^3$He polarization was saturated, the $^3$He cell, the $B_0$ coil, and the NMR system were transferred to the NOBORU beamline in MLF while applying a magnetic field to the $^3$He cell using a battery to keep the $^3$He polarization. Neutron transmission was measured using a Gas Electron Multiplier (GEM) detector~\cite{GEM} for the polarized and unpolarized $^3$He cell as well as for an empty glass cell with the same dimensions. The number of the transmitted neutrons through the $^3$He cell is described as~\cite{nPol}
\begin{eqnarray}
N_0T_{\rm{cell}}\epsilon(\lambda)\exp(-\rho d \sigma(\lambda))\cosh(P_{\rm{He}}\rho d \sigma(\lambda)),
\end{eqnarray}
where, $N_0$ is the number of incident neutrons, $T_{\rm{cell}}$ is the transmission of the cell windows, $\epsilon(\lambda)$ is the detection efficiency of the GEM detector, $P_{\rm{He}}$ is $^3$He polarization, $\rho$ is the number density of $^3$He nuclei, $d$ is the thickness of $^3$He gas, $\sigma(\lambda)$ is the absorption cross section of the $^3$He nucleus for unpolarized neutrons, and $\lambda$ is the wavelength of incident neutrons, which is calculated from the time of flight.
The ratio of transmitted neutrons for the polarized to unpolarized $^3$He cell $N_{\rm{pol}}/N_{\rm{unpol}}$ is described as
\begin{eqnarray}
\frac{N_{\rm{pol}}}{N_{\rm{unpol}}}=\cosh(P_{\rm{He}}\rho d \sigma(\lambda)).
\label{eq:cosh}
\label{ratio}
\end{eqnarray}
The product of the $^3$He gas pressure and thickness $\rho d$ was obtained from the measurement of the ratio of transmitted neutrons for the unpolarized $^3$He cell to the empty glass cell as 20.8$\pm$0.1~atm$\cdot$cm. The measured ratios of $N_{\rm{pol}}/N_{\rm{unpol}}$ are plotted againt the neutron wavelength in Fig.~\ref{fig:transmission} along with the curve of best fit using Eq.~(\ref{eq:cosh}) with $P_{\rm{He}}$ as a free parameter.
The $^3$He polarization was measured to be 85.3$\pm$0.5\%. The uncertainty comes mostly from the measured $\rho d$. \\
The neutron wavelength dependences on the neutron polarization is obtained using the following equation:
\begin{eqnarray}
P_n=\sqrt{1-\frac{T^2_0}{T^2}},
\label{eq:nPol}
\end{eqnarray}
where $P_n$ is the polarization of neutrons passed through the polarized $^3$He cell, $T$ is the transmission of polarized $^3$He cell, and $T_0$ is the transmission of the unpolarized $^3$He cell. The transmissions $T_0$ and $T$ are described as
\begin{eqnarray}
T_0=\frac{N_{\rm{unpol}}}{N_{\rm{cell}}}, \ \ T=\frac{N_{\rm{pol}}}{N_{\rm{cell}}},
\end{eqnarray}
where $N_{\rm{cell}}$ is the transmitted neutrons for the empty glass cell. The neutron wavelength dependence of the neutron polarization and the neutron transmission are shown in Fig.~\ref{fig:PolTrans}. The proton beam power to the neutron source was 500~kW, and each measurement took 15~min. The relaxation time $T_1$ on the beamline was 170~h, which was measured with AFP-NMR.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{3HePolarization.eps}
\caption{Ratio of the numbers of transmitted neutrons for the polarized to the unpolarized $^3$He cell. The red line shows the best fit. The error bars show the statistical uncertainty.}
\label{fig:transmission}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{nPol.eps}
\caption{Neutron wavelength dependences of neutron polarization (open circles) and neutron transmission (filled circles). }
\label{fig:PolTrans}
\end{figure}
\section{Trial and User Experiments using $^3$He spin filter}
We have carried out several trial experiments with {\it in-situ} and {\it ex-situ} SEOP methods at neutron beamlines of MLF~\cite{Hayashida16, Hayashida19, Kira15, Sakai14}. In 2017, the $^3$He spin filter was first supplied to a user experiment. Although no user experiment was conducted in 2018, five user experiments were performed in 2019--2020. Neutron beamlines, where we conducted experiments using $^3$He spin filters, are listed in Table~\ref{beamlinelist}. In this section, we report on experiments performed on typical beamlines at MLF using $^3$He spin filters, including the results already reported.
\begin{table}[h]
\caption{List of neutron beamlines, where experiments were conducted with $^3$He spin filters.}
\begin{ruledtabular}
\begin{tabular}{ccccc}
No. &Name & Purpose of beamline& Type of $^3$He spin filter\\
\hline
04 & ANNRI & Nuclear reaction & {\it ex-situ} \\
05 & NOP & Fundamental physics & {\it ex-situ} \\
06 & VIN ROSE& Spin echo & {\it ex-situ} \\
10 & NOBORU & General purpose & {\it ex-}, {\it in-situ} \\
15 & TAIKAN & SANS & {\it ex-}, {\it in-situ} \\
17 & SHARAKU & Reflectometer & {\it in-situ} \\
18 & SENJU & Single crystal diffraction &{\it ex-situ} \\
22 & RADEN & Pulsed neutron imaging &{\it in-situ} \\
\end{tabular}
\end{ruledtabular}
\label{beamlinelist}
\end{table}
\subsection{BL10 NOBORU}
A trial experiment for magnetic imaging was carried out~\cite{Hayashida16} with {\it in-situ} and {\it ex-situ} systems using $^3$He cell with pure Rb and the 30~W laser as shown in Fig.~\ref{fig:laser}. The {\it in-situ} system with a $^3$He cell of 17~atm$\cdot$cm and the {\it ex-situ} system with $^3$He cell of 11~atm$\cdot$cm were used as a neutron spin polarizer and analyzer, respectively. The dimensions of the $^3$He cells were 35~mm in diameter and 55~mm in length for both. A coil with a 0.35 mm-thick ring-shaped magnetic steel core was placed between the polarizer and the analyzer as a sample. A two-dimensional RPMT neutron detector~\cite{Hirota2005} was set after the analyzer. The magnetic field applied to the sample was controlled with the current sent to the coil. Two-dimensional images of the neutron polarization were obtained, and obvious differences in the spin rotations of neutrons were observed as the wavelength dependence of the neutron polarization in two current conditions, $I$ = 1.42~A and $I$ = 0~A. Our result concluded that the magnetization patterns in these two conditions are different. \\\\
Furthermore, a study for magnetic atomic resolution holography with polarized neutrons using a $^3$He spin filter is currently underway. Two user experiments for this study were conducted at NOBORU in 2019--2020.
\subsection{BL04 ANNRI}
BL04 ANNRI~\cite{Nakajima2017, ANNRI} is a neutron beamline used for studies of nuclear science such as nuclear data for nuclear technology, astrophysics, and quantitative analyses. A germanium detector assembly using 22 high-quality germanium crystals is installed on this beamline to measure the ($n$,$\gamma$) reaction. Measurements of the ($n$,$\gamma$) reaction for the study of fundamental symmetry violation in nuclei have been conducted as a user experiment using a $^3$He spin filter~\cite{Yamamoto2018,Yamamoto2019,Yamamoto20}. A $^3$He spin filter with the {\it ex-situ} system was used to polarize epithermal neutrons to measure the ($n$,$\gamma$) reaction with polarized neutrons at the 0.74~eV neutron resonance of $^{139}$La. A lanthanum sample was placed at the center of the germanium detector assembly, and the emitted $\gamma$-rays from the sample were detected by the germanium detectors. The $^3$He cell, solenoid, and compensation coils with a magnetic shield made of permalloy were installed between the detector and a neutron collimator (Fig.~\ref{fig:BL04}). The magnetic field used to keep the $^3$He polarization was perpendicular to the neutron beam in order to obtain a vertically polarized neutron beam. The magnetic shield had two openings in the path of the neutron beam to avoid a decay of the neutron polarization. Neutron polarization was held by the magnetic filed produced with the guide magnets to a lanthanum target. A $^3$He cell using pure Rb with a pressure thickness of 19.3~atm$\cdot$cm and dimensions of 50~mm diameter and 70~mm length was used. The $^3$He cell was polarized with the 30\,W laser system at MLF, and the $^3$He polarization at the start of the measurement was typically 60\%. The relaxation time of the $^3$He polarization was 130~h. In 2019, as a result of three separate measurements using the $^3$He spin filter, a significant spin-dependent angular distribution of $\gamma$-rays from the 0.74~eV neutron resonance of $^{139}$La was observed~\cite{Yamamoto20}. \\
Furthermore, a ($n$,$\gamma$) reaction measurement using polarized epithermal neutrons to obtain nuclear data for reactor science is being planned, and the development of a $^3$He cell for epithermal neutrons is an ongoing process.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{BL04.eps}
\caption{Photograph of the {\it ex-situ} system installed at ANNRI beamline. The coil wrapped with permalloy sheets, NMR system, and $^3$He cell were placed between the iron shields.}
\label{fig:BL04}
\end{figure}
\subsection{BL17 SHARAKU}
A trial experiment using the {\it in-situ} system on a beamline for neutron reflectometry: BL17 SHARAKU~\cite{Nakajima2017,SHARAKU} was performed~\cite{Hayashida16}. The SHARAKU beamline has a neutron polarizer and an analyzer consisting of supermirrors. Some of the supermirrors are stacked vertically in order to cover a wide area, which analyze the scattered neutron spins with wavelengths longer than 0.2 nm. However, some neutrons, which transmitted through the supermirrors are scattered by the supermirrors, and spacial distribution of neutrons reflected from a sample is disturbed. A $^3$He spin filter was used as the analyzer instead of the supermirrors to reduce uncertainty produced with the supermirrors (Fig.~\ref{fig:BL17}). The $^3$He cell using pure Rb with a pressure thickness of 11.1~atm$\cdot$cm, diameter of 35~mm, and length of 55~mm was used. A Fe-Cr multi-layered thin film was used as a sample, and off-specular measurements were performed using the {\it in-situ} system. The $^3$He polarization was 60\% during the experiment. The {\it in-situ} system successfully demonstrated and worked stably over 4 days of the experimental time. The sample had been measured previously at ISIS, and we obtained the identical results on SHARAKU. This satisfactory implies that our {\it in-situ} system worked as designed.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{BL17.eps}
\caption{Photograph of the {\it in-situ} system installed at SHARAKU beamline.}
\label{fig:BL17}
\end{figure}
\subsection{ BL06 VIN ROSE}
BL06 VIN ROSE~\cite{Nakajima2017, Hino2017} is equipped with a modulated intensity by zero effort (MIEZE) spectrometer to analyze the slow dynamics of condensed matter by measuring the intermediate scattering function. Experiments using a combination of a MIEZE spectrometer and the time of flight (TOF-MIEZE~\cite{VINROSE}) method have been carried out, and additionally, polarization analysis of scattered neutrons by a sample is planned in order to study spin dynamics by installing a spin analyzer. We installed a $^3$He cell after the sample position as an analyzer to confirm that the $^3$He spin filter can be used for the MIEZE spectrometer~\cite{Hayashida19}. The MIEZE signal was measured with and without the $^3$He cell, and no difference was observed in the signal contrast. This result implies that diffusion of $^3$He gas does not affect the MIEZE signal and a $^3$He spin filter can be used as a second analyzer in the MIEZE spectrometer. Furthermore, in 2020, an experiment to measure the spin dynamics of ferrofluid was conducted by using a $^3$He spin filter as shown in Fig.~\ref{fig:BL06}. The $^3$He polarization of 72\% and relaxation time of 120~h was achieved on this beamline in the experiment, and the analysis is currently underway.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{BL06.eps}
\caption{Photograph of the {\it ex-situ} system installed on the VIN ROSE beamline. The polarized neutron beam was scattered by the sample, and the spin of the scattered neutron was analyzed with the $^3$He spin filter. The vacuum duct was wrapped with boron sheets to reduce background neutrons.}
\label{fig:BL06}
\end{figure}
\subsection{BL15 TAIKAN}
BL15 TAIKAN~\cite{Nakajima2017, TAIKAN} is a small and wide angle neutron scattering instrument. We conducted a trial experiment to measure coherent and incoherent scattering cross sections from hydrogen contained material by analyzing scattered neutrons using the {\it in-situ} system~\cite{Kira15}. Supermirrors were installed upstream of the beamline as a neutron polarizer, and the $^3$He spin filter was used as a spin analyzer. Silver behenate was used as the sample, and the $^3$He cell with pure Rb was placed 120~mm downstream of the sample. The diameter of $^3$He cell was 35~mm, the length was 55~mm, and the gas pressure was 3~atm. A stable $^3$He polarization of 68\% was kept during the experiment. Scattered neutrons reaching a part of the small-angle detector bank were spin analyzed. The coherent and incoherent scattering components were successfully separated by analyzing the scattered neutrons.
Although the trial experiment was successful, the solid angle of the $^3$He cell was inadequate and the pressure-thickness was not suitable for the neutron wavelength used in TAIKAN. In order to cover a larger solid angle, an {\it ex-situ} SEOP system using a larger $^3$He cell, which is to be placed 10~mm downstream of the sample, is being prepared (Fig.~\ref{fig:BL15}). A new $^3$He cell with rubidium and potassium has been fabricated with a diameter of 60~mm, a thickness of 40~mm, and a gas pressure of 1.5~atm for cold neutrons. More trial and user experiments will be carried out in 2020.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{BL15Coil.eps}
\caption{Photograph of an {\it ex-situ} SEOP system for a SANS experiment at TAIKAN beamline. The {\it ex-situ} system is equipped with a sample changer which can be used to place up to six samples. The sample position is moved with the X-stage. The $^3$He polarization is kept by the Helmholtz coil. The $^3$He cell with the dimensions of $\phi$60$\times$40~mm is placed at 10~mm downstream of the sample to cover a large solid angle. }
\label{fig:BL15}
\end{figure}
\subsection{BL21 NOVA}
Experiments using $^3$He spin filters at BL21 NOVA~\cite{Nakajima2017, NOVA}, which is a total diffractometer, are also planned. NOVA is equipped with large solid angle neutron detectors. The detectors and a sample are installed in a large vacuum chamber. We are planning to install two {\it ex-situ} systems upstream of the beamline and around the sample to perform polarized neutron scattering experiments. The schematic view of the {\it ex-situ} systems at NOVA is shown in Fig.~\ref{fig:BL21Config}. The {\it ex-situ} system for the polarizer is installed at the upstream of the beamline, and the neutron polarization is held by the magnetic field produced with a guide coil to the sample. The scattered neutrons by the sample are analyzed with two $^3$He cells. This {\it ex-situ} system for the analyzer which can be installed to the vacuum chamber is developed as shown in Fig.~\ref{fig:BL21}. A sample is suspended from the top of the vacuum flange. The neutrons scattered by the sample are analyzed with two $^3$He cells placed at a distance of 20~mm from the sample. The stage for the $^3$He cells can be rotated so that the scattered neutrons of any angle can be analyzed. The $^3$He polarization is kept using the Helmholtz coil that produces 40~W of heat. The heat from the coil flows to the flange through the 30~mm thick aluminum plates connected to the coil, and the flange is air-cooled. A thermal simulation of the {\it ex-situ} system in a vacuum was performed. As a result of the simulation, the maximum temperature was determined to be 76$^\circ$C at the bottom of the coil, which does not cause a problem. A test experiment using the {\it ex-situ} system will be carried out in 2020.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,clip]{BL21Config2.eps}
\caption{Schematic view from the top of the {\it ex-situ} systems at NOVA. A neutron beam is polarized by the $^3$He cell installed upstream of the beamline, and the neutron polarization is held by the magnetic field produced with the guide coil to the sample. The scattered neutrons by the sample are analyzed with two $^3$He cells.}
\label{fig:BL21Config}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=7cm,clip]{BL21Coil.eps}
\caption{Photograph of the {\it ex-situ} system for the analyzer. The sample, the $^3$He cells, and the Helmholtz coil are suspended from the vacuum flange. The $^3$He polarization is kept with the Helmholtz coil.}
\label{fig:BL21}
\end{figure}
\section{Conclusion}
Development of $^3$He spin filters for user experiments are being carried out at MLF at J-PARC. The {\it in-situ} system using a 30~W laser has been developed, and several trial experiments using $^3$He cells with pure Rb have been conducted. Recently, a gas-filling station has been constructed at J-PARC and several high quality $^3$He cells using rubidium and potassium were fabricated. Additionally, the pumping system using a 110~W fiber laser was developed. High $^3$He polarization of 85\% was achieved on a neutron beamline using the $^3$He cell with rubidium and potassium and the pumping system using the fiber laser. The first user experiment utilizing a $^3$He spin filter took place in 2017, and five user experiments were conducted in 2019--2020. These experiments are beginning to yield scientific results. Preparations are underway for further user experiments using $^3$He spin filters.
\section*{Acknowledgement}
The authors would like to thank the staff of each beamline, and MLF and J-PARC for operating the accelerators and the neutron production target. This work was supported by MEXT KAKENHI Grant Nos.~19K21047, JP19GS0210, JSPS KAKENHI Grant Nos.~18K11929, JP18H05518, and JP17H02889. The neutron scattering experiments at BL06 and BL21 were approved by the Neutron Scattering Program Advisory Committee of IMSS, KEK (Proposal Nos. 2019S06 and 2019S07). The neutron scattering experiments at BL04 were performed under the user program of MLF (Proposals Nos.~2018B0148 and 2019A0185). The neutron scattering experiments at BL10, BL15, and BL17 were performed under the program of project use (Proposal Nos.~2012P0802, 2014P0802, and 2019P0201).
| 6a79a0412a467edd8fc1fefaa461a03c37eb4717 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{X-ray photoevaporative wind model} \label{sec:met:model_pe}
A photoevaporative wind is launched when radiation heats up the gas in the disc and deposits enough energy to leave the gas gravitationally unbound. In the planet-forming region of a disc the heating process is generally dominated by irradiation from the central star. We use the velocity and density structure of the gas in a photoevaporating wind from the radiation-hydrodynamical calculations of \citet{Picogna2019}. We rerun the hydrodynamical calculation in order to extend the simulation domain to 1000 au above the disc midplane. This was necessary to account for lamppost illumination from a jet source. The inner and outer radius remained at 0.35 and 200 au as in \citet{Picogna2019}. In this model a primordial disc of mass 0.07 M$_{\odot}$ is irradiated by its 0.7 M$_{\odot}$ host star with the X-ray + EUV spectrum presented by \citet{Ercolano2008a, Ercolano2009} with a luminosity of $L_X = 2\cdot10^{30}$ erg/s. Solar abundances from \citet{Asplund2005}, depleted according to \citet{Savage1996} were assumed. The gas and dust radiative transfer code MOCASSIN \citep{Ercolano2003, Ercolano2005a, Ercolano2008} was used to obtain a parametrisation for the gas temperature as a function of ionisation parameter and gas column density. This parametrisation was then used in the hydro-code PLUTO to calculate the hydrodynamic structure of the wind. More details are given by \citet{Picogna2019}.
We post-processed the density and velocity grids of the model by mapping it to a 2-dimensional Cartesian grid with 120 logarithmically spaced grid points in disc radius and 2000 linearly spaced points in height. The remapped grid extends from the disc radius R = 0 to 66 au and the height z = 0 to 200 au, however the inner radius of the hydrodynamical model is 0.35 au, below which all cells are set to 0. We have verified that the choice of 0.35 au for the inner radius is sufficient to accurately calculate profiles by running an additional hydrodynamical model with an inner radius of 0.05 au and comparing the resulting profiles. The emission from inside 0.35 au does not contribute significantly to the total flux, because the volume of this region is small. We choose the extent and resolution in z that high in order to be able to properly resolve the high regions of the wind when investigating the effect of lamppost illumination from a jet. The final grid results from averaging the last 25 orbits of the simulation in order to increase the smoothness.
\subsection{Magnetocentrifugal MHD wind model} \label{sec:met:model_mhd}
A number of magneto-hydrodynamical calculations of discs have shown that under given assumptions for magnetic flux and ionisation level in the disc a wind is launched, which can be as vigorous as a photoevaporative wind and which could, in theory, extend to radii well within the gravitational radius for photoevaporation. Indeed \citet{Banzatti2018} (B19) suggest that an MHD wind might be responsible for the observed broad component of [OI] forbidden line emission. Unfortunately a post-processing of currently available MHD calculations is not suitable for this purpose, since the grids of the available calculations do not extend to small enough radii where the emission would have to originate to give the observed FWHM of the BLVC of the [OI] 6300 line.
We have indeed experimented by post-processing the recent calculation of \citet{Rodenkirch2019a} which include a MHD wind and a photoevaporative wind for a range of the plasma parameter $\beta$, i.e. the ratio of thermal pressure over magnetic pressure. Unfortunately similar to other calculations \citep[e.g.][]{Wang2019} their grid only extends to 1 au in inner radius, while the observed FWHM suggest emission regions $<$ 0.5 au. Furthermore a limitation of these calculations is the need to impose a floor density for the inner regions, which produces numerical artefacts that significantly affect the line profiles. The floor density is needed because with decreasing density the Alfv\'{e}n velocity, which has to be resolved in every dynamical time-step, increases, imposing a severe limit to the time-step which leads to an unfeasible increase of the computational time.
These are serious limitations for the calculations of line intensities and profiles originating from the inner regions of the computational domain, which preclude the use of these numerical calculations at present.
In this work we thus rely on an analytical description of a magnetocentrifugal MHD wind. For a description of the density and velocity structure of the MHD wind we use the model by \citet{Milliner2018} that is based on the \citet{Blandford1982} (BP82) axisymmetric self-similar solutions for a magnetocentrifugally driven wind for thin discs. It was originally applied to FU Ori type objects, but is not specific to such. The model uses cylindrical coordinates [$R$, $\Phi$, $z$] = [$R_f\xi$, $\Phi$, $R_f\chi$] with non-dimensional functions [$\xi'$, $\Phi$, $\chi$] and the footpoint $R_f$ to describe self-similar streamlines of the form
\begin{align}
\chi = a\xi^2 + b\xi - (a + b).
\end{align}
\citet{Milliner2018} follow \citetalias{Blandford1982} by using the velocity components
\begin{align}
[v_R, ~v_{\Phi}, ~v_z] = [\xi' f(\chi), ~g(\chi), ~f(\chi)] \sqrt{GM_*/R_f},
\end{align}
to solve the structure of the flow along the streamlines and to finally derive the mass density
\begin{align}
\rho v_z = (\rho v_z)(z = 0) \frac{(R_{in}/R_f)^2}{\xi(\xi - \chi \xi')}
\end{align}
with
\begin{align}
(\rho v_z)(z = 0) = \frac{\dot{M}_w}{4\pi R_{in}^2 \mathrm{ln}(R_{out}/R_{in})}.
\end{align}
where $\dot{M}_w$ is the wind mass-loss rate and $R_{in}$ and $R_{out}$ are the inner and outer disc radii, respectively. More details can be found in \citet{Milliner2018} and \citetalias{Blandford1982}.
We use this model to construct 2-dimensional cartesian grids of the velocity and density structure with 100 grid points in disc radius and 600 points in height, which we found to be sufficient to properly resolve the emission. The grid points are spaced evenly in square root space. The grid extends from R = 0 to 10 au and z = 0 to 60 au. The parameters for the wind solution are $a = 0.0936$, $b = 0.612$, $\kappa = 0.03$, $\lambda = 30$, $R_{in} = 20$~$R_{\odot}$, $R_{out} = 10$ au and M$_*$ = 0.7 M$_{\odot}$. Since the parameters for MHD wind models are poorly constrained, we have chosen values that yield a poloidally-collimated flow similar to that seen in many numerical simulations. Based on the assumption that MHD winds drive accretion, we increase the mass-loss rates for models with a higher accretion luminosity to a rate comparable to the accretion rate \citep{Wang2019, Fang2018}. We choose $\dot{M}_w = 10^{-8.6}~\mathrm{M_{\odot}/yr}$ for models with an accretion luminosity of $2.6\cdot10^{-2}$~L$_{\odot}$, $\dot{M}_w = 10^{-7.5}~\mathrm{M_{\odot}/yr}$ for models with 0.31~L$_{\odot}$ and , $\dot{M}_w = 10^{-6.9}~\mathrm{M_{\odot}/yr}$ for 1~L$_{\odot}$. A summary of the input parameters for the different models is shown in table \ref{tab:input_parameters}. To calculate the hydrogen number density from the mass density we use the mean molecular weight $\mu = 1.3$. We assume the same abundances as in the X-ray photoevaporative model.
\begin{table*}
\centering
\include{tables/input_parameters}
\caption{Relevant parameters that were used to set up our models. PEJ models are photoevaporation models with an included "lamppost" illumination jet source.}
\label{tab:input_parameters}
\end{table*}
\subsection{Photoionisation calculations} \label{sec:met:ionisation}
We follow the approach of \citet{Ercolano2010} (EO10) and \citet{Ercolano2016} (EO16) to perform photoionisation calculations of the wind structures with the goal of predicting line luminosities and spectral profiles of the [OI]~6300, [OI]~5577, [SII]~4068 and [SII]~6730 lines. We use the previously mentioned MOCASSIN code with the same settings and input spectrum as in \citetalias{Ercolano2016} and \citet{Picogna2019}: A soft X-ray spectrum with a luminosity of $2\cdot10^{30}$ erg/s and an additional blackbody spectrum at a temperature of 12000~K that is loosely representative of an accretion component. We express the luminosity of the accretion component L$_{\mathrm{acc}}$ as the bolometric luminosity of the blackbody and perform our calculations for three different values: $2.6\cdot10^{-2}$~L$_{\odot}$, 0.31~L$_{\odot}$ and 1~L$_{\odot}$, corresponding to ionizing EUV luminosities (h$\nu$ > 13.6~eV) of $8.56\cdot10^{28}$~erg/s, $1.39\cdot10^{30}$~erg/s and $3.29\cdot10^{30}$~erg/s, respectively. Since the wind in our photoevaporative model is driven by the X-ray, the accretion component has no effect on the dynamics of this wind. For models where we investigate the effect of a "lamppost" illumination from a radiation source in the jet we add a third component, an EUV source modelled as a blackbody of temperature 25000~K. The bolometric luminosity of the component is chosen as a fraction of the accretion luminosity, where we calculate models with 10~\% and 32~\%. These parameters are observationally unconstrained and we use them because they work well to show the effect that a jet could have on the wind. The jet source is then placed on the z-axis at a height of 50 au. At that height, the wind velocity and density are approximately constant and a source higher in the jet would behave similarly.
\subsection{Line profile calculations and fits} \label{sec:met:profiles}
We have used the two-dimensional grids of emissivities from our photoionisation calculations and our grids of the wind structure to construct a three-dimensional axisymmetric cylindric volume and calculated the spectral profiles of our lines when viewed along different lines of sight. To construct the cylinder we used 256 evenly spaced azimuthal grid points. Based on the assumption that the disc midplane is optically thick to our lines we do not model the other half of the disc. This does have implications for our profiles at very high inclinations: Contribution from high regions in the receding side of the disc wind where the line of sight does not cross the midplane within the grid is lost in our calculation. However, we assume that this contribution is absorbed in the bound disc, because the line of sight does have a very long path through it, at the inclinations where this is relevant. We do account for line attenuation by dust absorption by calculating the column number density $N_H$ from each grid cell in the 3D-grid to the point where the line of sight leaves the grid. To calculate the absorption cross-section $C_{\lambda}^{abs}$ (5.49$\cdot10^{-23}$, 6.83$\cdot10^{-23}$, 1.52$\cdot10^{-22}$ and 4.92$\cdot10^{-23}$~cm$^2$ for [OI]~6300, [OI]~5577, [SII]~4068 and [SII]~6730, respectively), we assume a fixed dust-to-gas mass ratio of 10$^{-2}$ for the entire grid and a standard MRN type distribution \citep{Mathis1977}. Scattering is neglected. Recent work \citep{Franz2020} has shown that only grains smaller than approximately 10~$\mu$m can be entrained in a photoevaporative flow \citep[see also][]{Owen2011, Hutchison2016}. However we use the MRN distribution here as we are interested in an upper limit on the effect of dust absorption of these lines. As expected the effect is small.
Since our MHD model does not include a thick disc, we use the grids of the column number densities that we calculated for the photoevaporative model and remap them to be compatible with the MHD model.
To calculate the spectral profile we follow \citetalias{Ercolano2010} and \citetalias{Ercolano2016}. For each cell in our 3D-grid we calculate the line-of-sight velocity of the gas $v_{los}$, the thermal rms velocity of the emitting atom $v_{th}$ and the optical depth $\tau = C_{\lambda}^{abs} N_H$. We then create bins of velocity u with a resolution of 0.25 km/s and numerically integrate the luminosity in each bin by summing the contribution of the volume-averaged power $l(\vec{r})$ emitted in each cell of the 3D-grid:
\begin{align}
L(u) = \int{d\vec{r} \frac{l(\vec{r})}{\sqrt{2\pi v_{th}(\vec{r})}}} exp\bigg( - \frac{[u - v_{los}]^2}{2 v_{th}(\vec{r})^2} -\tau (\vec{r}) \bigg),
\end{align}
where $\vec{r}$ is the vector pointing to the grid cell. We convolve our profiles with a Gaussian of width $\sigma = c / R$, where we choose the spectral resolution $R = 50000$, a resolution similar to the observations that we compare our profiles to. The observational sample is comprised of the high resolution ($\sim$7 km/s) observations analyzed by \citet{Fang2018} and \citetalias{Banzatti2018}. To facilitate the comparison we include observed profile components only if they are classified as either NLVC, BLVC or HVC.
We use a similar approach as \citetalias{Banzatti2018} to create fits to our synthetic profiles: We perform multi-Gaussian fits, beginning with a single Gaussian component and adding another component only if it improves the $\chi^2$ value by at least 20~\%, up to a maximum of 6 components. Since our profiles are much smoother and thus compare better to Gaussian components than real observed spectra, we introduce an additional exit condition: If $\chi^2 < \frac{2}{3}$~y$_{\mathrm{max}}^2$, where y$_{\mathrm{max}}$ is the flux at the peak of the profile, the fit is considered to be sufficiently accurate and no more components will be added. This is to prevent the algorithm from fitting to a too high level of detail that cannot be achieved in observations. Figures of all fits are available as online material. We adopt the categorization of \citetalias{Banzatti2018}, where all components with a blue- or redshift of more than 30~km/s are classified as HVC. Components slower than that are either classified as NLVC if their FWHM does not exceed 40~km/s or as BLVC, otherwise. We do not adopt the further distinction between BLVC+NLVC and SC or SCJ components, because correlations with the n$_{13-31}$ infrared-index suggest that single components could trace winds in more evolved discs, such as transition discs \citepalias{Banzatti2018}, whereas in this work we model primordial discs. We do note, however, that our photoevaporation model can only produce SC-type components, and that only the MHD model can produce more complex profiles with multiple components.
\subsection{Emission regions} \label{sec:res:emission_regions}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{T_emission_all}
\caption{Temperature of the photoevaporative wind model (top panels) and the MHD wind model (bottom panels). Overlain are the contours of the 80\% emission regions of [OI]~6300 (green), [OI]~5577 (magenta), [SII]~4068 (yellow) and [SII]~6730 (cyan) for an accretion luminosity of $2.6\cdot 10^{-2} ~\mathrm{L_{\odot}}$ (left panels), $0.31~\mathrm{L_{\odot}}$ (middle panels) and 1~L$_{\odot}$ (right panels).}
\label{fig:T_emission_all}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{ne_emission_all}
\caption{Electron number density of the photoevaporative wind model (top panels) and the MHD wind model (bottom panels). Overlain are the contours of the 80\% emission regions of [OI]~6300 (green), [OI]~5577 (magenta), [SII]~4068 (yellow) and [SII]~6730 (cyan) for an accretion luminosity of $2.6\cdot 10^{-2} ~\mathrm{L_{\odot}}$ (left panels), $0.31~\mathrm{L_{\odot}}$ (middle panels) and 1~L$_{\odot}$ (right panels).}
\label{fig:ne_emission_all}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{nh_v_emission_all}
\caption{Density structure of the photoevaporative wind model (top panels) and the MHD wind model (bottom panels). The arrows show the velocity vectors in the RZ-plane. Overlain are the contours of the 80\% emission regions of [OI]~6300 (green), [OI]~5577 (magenta), [SII]~4068 (yellow) and [SII]~6730 (cyan) for an accretion luminosity of $2.6\cdot 10^{-2} ~\mathrm{L_{\odot}}$.}
\label{fig:nh_v_emission_all}
\end{figure*}
\citet{Ercolano2016} have shown that their X-ray photoevaporative wind model can successfully reproduce the correlation between accretion luminosity and [OI]~6300 line luminosity by comparing their simulations with observations of \citet{Rigliaco2013} and \citet{Natta2014}. They suggest that this correlation is a consequence of the size of the wind region that the EUV flux, which is dominated by and proportional to the accretion luminosity, can reach and heat up. Since the luminosity of collisionally excited lines depends exponentially on the temperature through the Boltzmann term $\mathrm{exp(-\Delta E/k_B T)}$, a larger heated region leads to increased line luminosities. Figure \ref{fig:T_emission_all} shows temperature maps of the different models with the 80\% emission regions of all four lines overlain. Since the MHD model does not have a thick disc, we masked away the emission that would originate within the bound disc. The mask was constructed by remapping the bound disc of the photoevaporative model to the grid of the MHD model. It is clear from the figure that a higher accretion luminosity increases the size of the emission region. Another important factor for the luminosity of collisionally excited lines is the abundance of the emitting species (e.g. neutral oxygen for [OI] lines) and of the colliding particles that excite the species, which in our case are electrons and neutral hydrogen. It is important to note that our photoionisation calculations include neutral hydrogen as colliding particles only for the [OI]~6300 line, due to the lack of collisional rates with neutral hydrogen for the relevant excitation levels of the other lines. The calculated luminosities of the [OI]~5577, [SII]~4068 and [SII]~6730 should thus be regarded as a lower limit. The rate coefficients for [OI]~6300 are taken from \citet{Launay1977}, who report only the first four levels, $^3$P$_2$, $^3$P$_1$, $^3$P$_0$ and $^1$D$_2$, but not the $^1$S$_0$ level that is relevant for [OI]~5577. We verified that this only affects the total line luminosity but has no significant effect on the shape of the line profiles by recalculating the [OI]~6300 profiles with neutral hydrogen collisions turned off and comparing them to the original profiles. Figure \ref{fig:ne_emission_all} shows maps of the electron number density in our models and figure \ref{fig:nh_v_emission_all} shows maps of the hydrogen number density and the outflow velocities in the rz-plane. From these three figures we can read off the physical properties of the emission regions. The emission only becomes significant above a critical density, where the rate of collisional excitation to the upper level of the relevant transition matches the rate of radiative de-excitation. If the density is too low, not enough atoms/molecules are in the excited state. If it is too high, collisional de-excitation will dominate and suppress the transition. The temperature and the electron and hydrogen number densities in the emission regions are listed in table \ref{tab:emission_properties}. The [OI] 5577 line is only emitted in a small region close to the star where the density is comparatively high. As a result the line is very weak, which makes observations of this line challenging. The [SII]~6730 line is emitted even at very low densities and temperatures and therefore arises in a much larger region that reaches deep into the wind. Of our four lines it would be the best candidate to trace an extended disc wind.
\begin{table}
\centering
\input{tables/emission_properties.tex}
\caption{Temperature, electron number density and hydrogen number density of the line emission regions.}
\label{tab:emission_properties}
\end{table}
\subsection{Line profiles} \label{sec:res:profiles}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{profiles_pe}
\caption{Simulated line profiles for the photoevaporative wind models. The profiles were artificially degraded to a spectral resolution of R = 50000 and normalized to the peak flux of the 0\degree{} profile.}
\label{fig:profiles_pe}
\end{figure*}
The line profiles resulting from the X-ray photoevaporative wind models are shown in figure \ref{fig:profiles_pe}, normalized to the peak flux of the 0\degree{} inclination profiles. It is clear that the [OI]~6300 and [SII]~4068 profiles are very similar although the [SII]~4068 emission is reaching farther into the wind and along the inner edge of the wind. As can be seen in figure \ref{fig:nh_v_emission_all} (top panels) the regions have a very similar velocity structure, which explains the similarity of the line profiles. They both have blueshifted peaks of a few~km/s with the highest blueshift at 60\degree{} inclination. The [SII]~6730 profile is emitted up to higher regions and therefore less affected by dust attenuation at higher inclinations. The broadest profiles, when observed at an inclination, are those of the [OI]~5577 line. This is due to the very small emission region that lies very close to the star and is thus subject to strong Keplerian broadening. Although the other lines are emitted in that region, too, the contribution to their line profiles manifests only weakly in their wings. Unsurprisingly, the [OI]~5577 line is also the weakest and shows very little blueshift, because a photoevaporative wind cannot effectively be launched that close to the star. We fitted the profiles as described in section \ref{sec:met:profiles} and show the resulting component properties in table \ref{tab:fits_pe} in appendix \ref{sec:appendix:fits}. Figures of all fits are available as online material.
\begin{table*}
\centering
\input{tables/component_stats.tex}
\caption{Mean values of the centroid velocities and FWHM and their standard deviation for all lines and components.}
\label{tab:component_stats}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{profiles_mhd}
\caption{Simulated line profiles for the MHD wind models. The profiles were artificially degraded to a spectral resolution of R = 50000 and normalized to the peak flux of the 0\degree{} profile.}
\label{fig:profiles_mhd}
\end{figure*}
Figure \ref{fig:profiles_mhd} shows the profiles for the MHD wind model. By far the most notable difference when comparing these profiles to the photoevaporative wind model are the very high blueshifts. At higher accretion luminosities and mass-loss rates all profiles show significant blueshifts up to velocities of $\sim$300~km/s. Despite the simple model, some of the profiles are remarkably similar to observed profiles, although the model generally underestimates the line luminosities. Figure \ref{fig:DGTau_comp_mhd} compares the 30\degree{} inclined [OI]~6300 profile to the profile of DG Tau, which has a similar inclination. Note that using the [OI]~6300 luminosity of DGTau from \citet{Simon2016a}, we find that the luminosity of our synthetic line is a factor of $\sim$25 too low. The profiles have thus been normalized to account for this factor. We highlight that our work shows that the observed HVC can be reproduced by our analytical MHD wind without the need of explicitly including a 'classical' collimated jet. Indeed, with blueshifts up to $\sim$300~km/s, the high-velocity wings extend to even higher blueshifts than the wing of the DG Tau profile. In the understanding that the fastest, innermost edge of the wind model is a jet, this demonstrates that the jet does not need to be collimated or exhibit shocks in order to produce the observed HVCs, as proposed in previous works. At lower blueshifts the emission of our synthetic profile is too broad and with hints of Keplerian double peaks compared to that of DG Tau. In fact, we find pronounced double peaks in all our inclined profiles, but they are rarely observed. This could indicate that the Keplerian broadening is too strong in our model, but this is unlikely considering the inner radius of 20~R$_{\odot}$, which is relatively large for an MHD wind. An alternative explanation is that the Keplerian throught is filled out by another narrow low velocity component originating in a different outflow, such as a photoevaporative wind. We explore this possibility in section \ref{sec:discussion:profiles_combi}. The reduced high-velocity flux in the profiles of the model with the lowest accretion rate can be easily understood with the help of the emission maps in figures \ref{fig:T_emission_all} -- \ref{fig:nh_v_emission_all}: The density in the fast flowing inner region of the inner wind is low. As a result the temperature is high, as is the degree of ionization. With the low density and high degree of ionization, not enough neutral oxygen is present for the [OI] lines to be emitted. The singly ionized sulfur is still abundant enough to produce a significant flux of the [SII] lines in the high-velocity region. As the accretion luminosity and with it the density of the wind increases, the fast moving inner edge is less ionized and the high-velocity flux increases. The [OI]~5577 line that is only emitted close to the star has no significant contribution in the high-velocity region. The MHD profiles are typically best fitted with 3 to 5 components. All fitted components are listed in tables \ref{tab:fits_mhd_pt1} and \ref{tab:fits_mhd_pt2}.
Table \ref{tab:component_stats} shows the mean values of the centroid velocity and FWHM of the NLVCs and where applicable the BLVCs and HVCs for the photoevaporative and MHD models. Table \ref{tab:luminosities} shows the line luminosities.
\begin{figure}
\centering
\includegraphics[width=.48\textwidth]{DGTau_comp_mhd}
\caption{Simulated [OI]~6300 profiles (orange lines) of the MHD wind model with an accretion luminosity of $0.31~\mathrm{L_{\odot}}$ at 30\degree{} inclination compared to the observed spectrum of DG Tau (black lines). The dashed lines show the individual components from the Gaussian decomposition. The synthetic profile has been artificially degraded to a spectral resolution of R = 50000. The profiles have been normalized in such a way that they are comparable to each other. The DG Tau spectrum is taken from \citetalias{Banzatti2018}.}
\label{fig:DGTau_comp_mhd}
\end{figure}
\begin{table*}
\centering
\input{tables/luminosities.tex}
\caption{Logarithm of the line luminosities in units of~L$_{\odot}$.}
\label{tab:luminosities}
\end{table*}
\subsubsection{Line profile decomposition} \label{sec:res:profile_decomposition}
It is true for all lines that the line broadening is entirely dominated by the velocity gradient of the wind, when the disc is viewed at low inclinations. In that case the profiles do not allow for a clear distinction between low-velocity flux originating close to the star in the slow base of a fast wind or jet and emission originating in a slow extended disc wind far away from the star. The formation of Keplerian double peaks at inclinations $\gtrsim$ 40\degree{} indicates that Keplerian rotation begins to affect the broadening of the low velocity-flux around that inclination. However, on the blueshifted side, an increase of the inclination will result in contamination of the low-velocity component with emission from high-velocity regions that is projected to the low-velocity regime. This contamination is strengthened by the fact that the high-velocity wind regions have the smallest outflow-angles with respect to the z-axis in the rz-plane, resulting in stronger projection effects for HVCs and stronger broadening for LVCs. To illustrate this, figure \ref{fig:OI-6300_v_emis} shows the 80~\% emission regions of the [OI]~6300 line in the MHD models for different velocities in increments of 20 km/s. At 0\degree{} inclination the regions are well separated and could be traced using the centroid velocity of narrow components, if they are no broader than $\sim$20~km/s. At higher inclinations, starting already at 20\degree{}, the emission regions from flux observed in the range between +40~km/s and -40~km/s overlap significantly. At high inclinations projection effects are strong and a distinction of emission regions based on the blue- or redshift is difficult. As a result, a fitted Gaussian component will contain flux from many different parts of the wind. We therefore advise against interpreting different fit components as tracing physically distinct outflow regions. For the wind models with higher accretion luminosities, it is true that the majority of the emission traces those parts of the wind that are launched inside 1 au or even 0.5 au at higher inclinations. In our fits these models do reproduce NLVCs that are consistent with observations and typically attributed to being launched at larger radii (c.f. figure \ref{fig:corr_inc}). This demonstrates that although the velocity slices do reflect the onion-like velocity structure of the wind, the line width of the components is not a good indicator for the emission region, neither when the disc is viewed at low inclinations, nor at high inclinations. Line ratios between components should only be considered with caution, because the components are likely to contain flux from multiple regions with different physical properties. We expect the centroid velocity, as a proxy for the velocity slice, to hold most of the physical meaning of the Gaussian components. It could be worth considering alternative measurements (e.g. a spectral slope) for future studies of correlations between line profiles and physical properties of the disc.
\begin{figure*}
\centering
\includegraphics[width=.9\textwidth]{OI-6300_v_emis}
\caption{80~\% emission regions of [OI] 6300 flux that is observed at different velocities for the three MHD models with a spectral resolution of R = 50000.}
\label{fig:OI-6300_v_emis}
\end{figure*}
Nevertheless the Gaussian fits to observed profiles do reveal some correlations. We can compare our fits with the observations, in order to see if we can reproduce the general trends. Figure \ref{fig:OI-6300_overview} shows an overview of the centroid velocities and FWHM of the [OI]~6300 line. Except for the narrowest observations and the two observations with a blueshift around $\sim$12~km/s, the NLVCs are well reproduced by the photoevaporation models. The SCs, while having similar blueshifts, are typically much broader than the NLVCs and cannot be reproduced by our photoevaporation model. Correlations with infrared-index reported by \citetalias{Banzatti2018} suggest that these components could trace winds in more evolved discs, such as transition discs. As has been shown by \citetalias{Ercolano2016}, transition disc profiles are generally broader, although still not broad enough to match the observed SC line widths.
BLVCs and HVCs are only produced by the MHD model. The HVCs tend to be too broad around a blueshift of $\sim$150~km/s when compared to observations. This discrepancy can be explained by the fitting procedure which prefers one single broad Gaussian over two narrower Gaussians that would better match the observations. The one redshifted HVC comes from a profile at 80\degree{} and is a fit to the redshifted part of the Keplerian double peak. The MHD model consistently overestimates the blue or redshift of the NLVCs. With only 4 BLVCs at the narrower end of the spectrum, the model clearly tends to favour NLVCs and HVCs over BLVCs. These discrepancies are a consequence of the Keplerian double peaks that are found in the majority of our profiles. They are better fitted by two narrow components than a broad component.
\begin{figure*}
\centering
\includegraphics[width=.9\textwidth]{OI-6300_overview}
\caption{Overview of the parameter space of centroid velocities and FWHM of the [OI]~6300 profile components. Components from the MHD models are marked with a triangle, those from the photoevaporation model with a circle. The plusses mark observations by \citetalias{Banzatti2018}. The colors indicate the type of the components: Green: HVC, red: BLVC, blue: NLVC, grey: SC, pink: SCJ.}
\label{fig:OI-6300_overview}
\end{figure*}
\subsection{Correlations with inclination} \label{sec:res:correlations_inc}
\begin{figure*}
\centering
\includegraphics[width=.8\textwidth]{corr_inc}
\caption{Correlations between the disc inclination and the [OI]~6300 line component centroid (left panels) and FWHM normalized with the square root of the stellar mass (right panel). The grey lines show the values expected from purely Keplerian broadening at different radii. The plusses mark observations by \citetalias{Banzatti2018}. The colors indicate the type of the components: Green: HVC, red: BLVC, blue: NLVC, grey: SC.}
\label{fig:corr_inc}
\end{figure*}
\citetalias{Banzatti2018} investigate the correlation between [OI]~6300 centroid velocities and viewing angle in their sample and find a weak trend of the LVCs to have their maximum centroid velocity at an inclination around 35\degree{}. As can be seen in figure \ref{fig:corr_inc} none of our models is able to clearly reproduce this trend. The top left panel shows the inclination against centroid velocities of our photoevaporation model profile components compared to the observed NLVCs. While the model with the lowest accretion luminosity has its peak at 60\degree{} inclination, the models with higher accretion luminosities show only a weak trend with a maximum around 20\degree{}. As is clear from the figure and as has already been pointed out by \citetalias{Banzatti2018} the centroid velocities are dominantly influenced by the accretion luminosity and not by the viewing angle. We will investigate the correlations between the component properties and accretion luminosity in the next section. We have seen in the previous section that the photoevaporation model fails to reproduce the two observed components with the highest blueshifts. As we will show in section \ref{sec:res:jet}, a lamppost-type illumination of the upper layers of the X-ray photoevaporative wind by a UV source in a jet could help close the gap to these observations. The right panel shows the correlation of the component width with inclination. The widths are consistent with Keplerian broadening between 0.5 and 5~au, which is also consistent with the emission maps, but as seen at inclinations < 20\degree{}, Kepler broadening is not the dominating mechanism.
The MHD model seems to produce NLVCs with decreasing blueshift that transitions into a redshift with increasing inclination. A visual inspection of the fits in appendix \ref{sec:appendix:fits} reveals that most profiles are fitted with a NLVC at their right edge, which explains this behaviour. The model is able to produce NLVCs that reach up to 30~km/s but similar to the photoevaporation model it cannot reproduce the narrowest of the observed NLVCS. At inclinations < 25\degree{} the observations show a reduced number of NLVCs and BLVCs with intermediate blueshifts of $\sim$10 - 30~km/s. One possible cause could be the 30~km/s threshold above which components get classified as HVCs. Outflows in a jet along the z-axis with velocities slightly above that threshold would have a projected velocity just below it, resulting in a classification as BLVC. If that was the case we would expect a reduced number of HVCs at higher viewing angles but this is difficult to assess with the given distribution of observational samples. In fact, it is much more likely that the low number of components at low inclinations is a direct consequence of the non-uniform distribution of samples, as is shown in figure \ref{fig:inclinations}. This could also explain the observed peak at $\sim$35\degree{}. The general lack of BLVCs makes it impossible to obtain useful information about the effect of the inclination on the BLVCs. The centroid velocities of the HVCs generally decrease with increasing inclination, matching the observations well. The figure helps to support two arguments that were made in the previous section: The redshifted HVC belongs to the 80\degree{} profile of the 1~L$_{\mathrm{acc}}$ model and the very broad HVCs are mostly present at low inclinations up to 40\degree{}. At higher inclination the extended blueshifted wing is projected to lower velocities and the HVCs become somewhat narrower.
\subsection{Correlations with accretion luminosity} \label{sec:res:correlations_acc}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{corr_acc}
\caption{Correlations between the accretion luminosity and [OI]~6300 line component luminosities (left panels), centroids (middle panels) and FWHM (right panels). The plusses mark observations by \citetalias{Banzatti2018}, the crosses mark observations by Fang et al. (2018). The colours have the same meaning as in figure \ref{fig:corr_inc}. }
\label{fig:corr_acc}
\end{figure*}
The more interesting correlations, which we suggest to be the underlying cause of all three NLVC-BLVC connections reported by \citetalias{Banzatti2018} are those between the component properties and the accretion luminosity. We have shown in appendix \ref{sec:appendix:correlations} that in order to reproduce these observed connections it is sufficient to reproduce the correlations with L$_{\mathrm{acc}}$. Figure \ref{fig:corr_acc} shows these correlations for our models and the comparison to the observations. The first reported connection is a positive correlation between the equivalent width of the NLVCs and that of the BLVCs. The NLVCs produced by our photoevaporative wind model are in good agreement with the observations, which shows that in a scenario where the NLVCs are produced by a photoevaporative wind and the BLVCs by a different wind type, one would still find the observed connection between the NLVC and BLVC luminosities. It is worth noticing at this point that the luminosities of our MHD model components are at or below the lower limit of the observed luminosities but the slope matches the observations. Compared to the photoevaporative wind model the densities in our MHD model are much lower for a wind with the same mass loss-rate. As a result the collisionally excited lines are weaker, too.
The second connection reported by \citetalias{Banzatti2018} is a positive correlation between the centroid velocity of the NLVCs and that of the BLVCs. As is shown in the middle panel of figure \ref{fig:corr_acc}, the NLVC centroids of our photoevaporative wind model increase with increasing accretion luminosities. This can be explained by the larger emission regions for the models with higher L$_{\mathrm{acc}}$, as we will discuss in section \ref{sec:discussion:correlations}. The correlation is again consistent with the observed NLVCs, which shows that the observed connection between NLVC and BLVC centroids could also be achieved when the two components trace different wind types. We have seen before that the MHD model does not reproduce the NLVC and BLVC centroids very well, but the HVC centroids tend to be more blueshifted with increasing accretion luminosity, although there are a number of HVCs with lower blueshifts than would be expected from the observations. It is likely that these components are another consequence of our fits to Keplerian double peaks.
Finally, \citetalias{Banzatti2018} reported a positive correlation between the FWHM of the NLVCs and that of the BLVCs. The observations show a decreasing FWHM with L$_{\mathrm{acc}}$ for both, the NLVC and the BLVC. The right panel of figure \ref{fig:corr_acc} clearly demonstrates that the FWHM of the NLVCs in our photoevaporative model matches this correlation. There is no clear trend in the observed HVCs but the MHD model does not have any broad HVCs in the model with low L$_{\mathrm{acc}}$ as a direct consequence of the lack of highly blushifted flux.
By showing that the photoevaporation model can reproduce the correlations of the NLVCs with L$_{\mathrm{acc}}$, we have demonstrated that all three reported NLVC-BLVC connections are compatible with a scenario in which the NLVC traces a photoevaporative wind and the BLVCs are produced in a different wind. The correlation between [OI]~6300 luminosity and L$_{\mathrm{acc}}$ does not distinguish between different wind scenarios, because both, photoevaporative and MHD winds produce such a positive correlation. This argues against conclusions \citep[e.g. by][]{Nisini2017} that this correlation alone suggests an MHD origin for the wind traced by [OI]~6300. What distinguishes different winds is truly the velocity structure and not necessarily the luminosity.
\subsection{Lamppost illumination from a jet}\label{sec:res:jet}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{T_emission_jet_pe}
\caption{Electron number density (top) and temperature (bottom) of the photoevaporative wind model with accretion luminosity $2.6\cdot 10^{-2}~\mathrm{L_{\odot}}$ and an additional EUV jet source at Z = 50 AU on the z-axis that is modelled as a blackbody with T = 25000 K and luminosity $2.6\cdot 10^{-3}~\mathrm{L_{\odot}}$. Overlain are the contours of the 80\% emission regions of [OI]~6300 (green), [OI]~5577 (magenta), [SII]~4068 (yellow) and [SII]~6730 (blue) for an accretion luminosity of $2.6\cdot 10^{-2}~\mathrm{L_{\odot}}$ (left panels), $0.31~\mathrm{L_{\odot}}$ (middle panels) and $1~\mathrm{L_{\odot}}$ (right panels). The [OI]~5577 emission region lies entirely inside 1 au and is not visible at the scale of the figure.}
\label{fig:T_emission_jet_pe}
\end{figure*}
The observations show that a HVC is often present when both a broad and a narrow component is detected in the LVC of [OI] 6300. If the HVC is indicative of a jet, as often assumed in the literature, then it is useful to investigate how the profiles change when the wind is heated at higher heights above the midplane. To this aim, we modelled a scenario in which the wind is illuminated in a lamppost-type fashion from a source placed higher up on the z-axis. This source could potentially be an EUV or X-ray source created by shocks in a jet at a height of a few tens of au. We repeated our calculations for the photoevaporative and MHD models with an accretion luminosity of $2.6\cdot 10^{-2}~\mathrm{L_{\odot}}$ and 0.31~L$_\mathrm{{\odot}}$, both with an additional EUV input spectrum at a height of 50 au on the z-axis, which we modelled as a blackbody of temperature 25000 K and varying luminosity. As is shown in figure \ref{fig:T_emission_jet_pe}, upper wind levels can contribute significantly to the emission when the photoevaporative wind model is heated by an illumination source in the jet. Table \ref{tab:luminosities_pe_jet} shows the total line luminosities for the different lamppost-models and their ratio to their luminosity in the models without a jet. The [OI]~5577 line is not significantly affected by the jet source, as it requires a higher density to be emitted efficiently. The resulting [OI]~6300 profiles are shown in figure \ref{fig:profiles_jet_pe}. Compared to the profiles without a lamppost-type illumination, the centroid velocity of the fitted NLVC can increase the blueshift from $\sim$5~km/s to up to 10~km/s. The individual fits for the [OI]~6300 profiles are listed in table \ref{tab:fits_pe_jet}. With increasing luminosity of the jet source or decreasing accretion luminosity, the emission from the higher wind layers can dominate over the emission from lower layers, which will lead to more blueshifted but narrower profiles. This scenario is able to close the gap to the observed NLVCs with the highest blueshifts of up to $\sim$12~km/s. While the centroid velocities increase, the FWHM tend to decrese with increasing luminosity of the jet source. Apart from the [SII]~6730 line, the lamppost illumination has no effect on the emission in the MHD wind model. In that model, the density of the upper wind is too low for the other lines to be emitted efficiently. Since the MHD wind is very fast at high heights, the additional [SII]~6730 flux manifests in an increase in high-velocity flux. Because the velocity structure is relatively constant that high in the upper wind, the height of the jet source has little effect, as long as it is not placed low enough for the launching region of the wind to be heated by the jet source. The temperature from the jet source, however, does play an important role. A higher temperature implies a higher EUV flux, which is able to heat a larger volume of the wind, but also increases the degree of ionization, suppressing emission of neutral or lowly ionized species. X-ray sources, such as those observed in multiple jets \citep[e.g.][]{Gudel2007a} have little effect on the line profiles, as X-rays are mainly absorbed by heavier elements and not by hydrogen, making them inefficient in heating the wind.
\begin{figure}
\centering
\includegraphics[width=.48\textwidth]{OI-6300_profiles_jet_pe}
\caption{Simulated [OI]~6300 profiles at 30\degree{} inclination from the photoevaporative wind models with and without jet components. The dotted lines show the profiles without a jet component, the solid lines with a jet source that is 10~\% of the accretion luminosity. The luminosities in the legend are given in units of L$_{\odot}$. All jet components are modelled as a blackbody with T = 25000 K at a height of 50 AU on the z-axis. All profiles have been artificially degraded to a spectral resolution of R = 50000.}
\label{fig:profiles_jet_pe}
\end{figure}
\begin{table}
\centering
\input{tables/luminosities_pe_jet.tex}
\caption{Total line luminosities from the photoevaporative wind models with an additional EUV jet source at a height of 50 AU on the z-axis that is modelled as a blackbody with T = 25000 K with a luminosity that is given in the table as L$_{\mathrm{jet}}$. The last column shows the ratios of the luminosities between a model with and without the jet source.}
\label{tab:luminosities_pe_jet}
\end{table}
\begin{table}
\centering
\input{tables/fits_pe_jet.tex}
\caption{Centroid velocities, FWHM, luminosity and type of the profile components from the photoevaporative wind models with an additional EUV jet source at a height of 50 AU on the z-axis that is modelled as a blackbody with T = 25000 K with a luminosity that is given in the table as L$_{\mathrm{jet}}$.}
\label{tab:fits_pe_jet}
\end{table}
\subsection{Origin of the correlations with L\texorpdfstring{$_{\mathrm{acc}}$}{}} \label{sec:discussion:correlations}
We have shown in section \ref{sec:res:correlations_acc} that the photoevaporation model has the same correlations with the accretion luminosities as observed NLVCs. However BLVCs are not reproduced in a photoevaporative wind, therefore we have shown that two different wind types can produce the same correlations. To explain why these correlations exist we can again use the emission maps in Figure \ref{fig:nh_v_emission_all}: In the photoevaporation model a higher accretion luminosity increases the heated wind region and thus the emission region. This has not only the consequence of higher luminosities, but also more emission from the upper, faster regions of the wind, increasing the blueshift of the line. At the same time the line width is reduced, because the azimuthal velocity is very low in the upper wind regions and the Kepler broadening is reduced. The top right panel of figure \ref{fig:corr_acc} shows that this effect is strongest for the highest inclinations. The 0\degree{} profile shows exactly the opposite behaviour, because Kepler rotation is not at play and the broadness is entirely dominated by the velocity gradient, which is greater in a larger emission region.
In the case of an MHD wind the increase in size of the emission region with higher L$_{\mathrm{acc}}$ is less pronounced, because it is countered by the increase in gas density. As a result, there is a strong correlation between accretion luminosities and line luminosity, but the line profiles remain similar. An exception is the model with the lowest accretion luminosity where the inner edge of the wind is highly ionized such that no flux with a blueshift > 100~km/s can be observed, as discussed in detail in section \ref{sec:res:emission_regions}. With similar profiles, the correlation between the profile components and L$_{\mathrm{acc}}$ are not as obvious as in the observations. This could be an indication that our model overestimates the link between wind mass-loss and accretion rate. With a weaker correlation we would expect a bigger increase in the size of the emission regions and consequently correlations that are similar to those found in the photoevaporation models and the observations. Alternatively, as proposed by \citetalias{Banzatti2018}, an MHD wind where the increase in mass-loss manifests not only in a higher gas density but also in higher outflow velocities could restore the correlations.
\subsection{Correlations with viewing angle}
\label{sec:discussion:corr_inc}
\citetalias{Banzatti2018} interpret the observed correlation between NLVC line widths and viewing angle and the lack of such a correlation in the BLVC as possible evidence that the NLVCs are emitted close to the disc surface at larger radii, rotating close to the Keplerian speed, while the BLVCs are emitted at higher parts of the wind, where the poloidal velocity is increased and the toroidal velocity is lower than the Keplerian speed at the footpoint where the flow was launched. As discussed in section \ref{sec:res:profile_decomposition} the Gaussian decomposition and with it the correlations of their properties should only be treated with caution, but we can nevertheless investigate the plausibility of this scenario: In our MHD wind models with higher accretion luminosities the emission region of the flux with zero blueshift is indeed located at a height of $\gtrsim$ 2 au, but this includes the narrow component as well. The scenario would thus only be plausible if the MHD profile was supplemented by another narrow low velocity component produced by a different wind at larger radii and closer to the disc surface. Evidence for such a combination of profiles is the overabundance of Keplerian double peaks in our models. Double peaks are rarely observed, which means that either our model overestimates azimuthal wind velocities or in reality the Keplerian throughts are filled in by another component.
\subsection{Combination of MHD and photoevaporation profiles} \label{sec:discussion:profiles_combi}
One possible solution for both problems described in the previous section is the combination of a higher-density inner MHD wind and a photoevaporative wind that dominates at larger radii and fills the Kepler throught of the MHD profile. If the MHD wind arises from a region that is MRI turbulent we would expect the wind itself to become turbulent and introduce asymmetries that allow the X-ray to reach the disc at high-enough radii. Furthermore, knots/ Herbig-Haro objects in jets are evidence for a temporal variability of the outflow velocity. It might be plausible to assume a (possibly time dependent) filling factor to modulate the amount of radiation finally reaching the disc at radii where it can drive a photoevaporative wind.
\begin{figure}
\centering
\includegraphics[width=.48\textwidth]{profiles_combi_1}
\caption{Combination of MHD and photoevaporative wind profile for the [OI]~6300 line of our models with with L$_{\mathrm{acc}}$ = 0.31~L$_{\odot}$ at 40\degree{} inclination.}
\label{fig:profile_combi}
\end{figure}
We performed this experiment by combining our MHD with the photoevaporation profiles to see whether this could indeed bring the total profiles closer to the observations. For this purpose, we added a fraction of the flux of the photoevaporative profiles to that of the MHD profile. We chose fractions of 24~\%, 12~\% and 6~\% for the models with the lowest, intermediate and highest accretion luminosity, respectively. It is important to note that this choice is chosen only because it provided acceptable results and has no further justification, besides our expectation that a denser inner MHD wind, which is the case in the models with higher L$_{\mathrm{acc}}$, will shield more of the X-ray radiation that drives the photoevaporative wind. Figure \ref{fig:profile_combi} shows the combined [OI]~6300 profile at 40\degree{} inclination for the model with L$_{\mathrm{acc}}$ = 0.31~L$_{\odot}$. The Keplerian double peaks are indeed filled by the photoevaporation profile. Note that the fits do not properly separate the photoevaporative from the MHD component. Figure \ref{fig:OI-6300_overview_combi} shows that combining the profiles does improve the centroid and FWHM distribution of the modelled profiles. Especially the number of highly blue- or redshifted NLVCs is greatly reduced and the number of BLVCs increased from 4 to 7. The BLVCs with high FHWM $\gtrsim$ 100~km/s are still not reproduced, however, it is reassuring that our simple experiment can already significantly improve the distribution of component properties.
\begin{figure*}
\centering
\includegraphics[width=.9\textwidth]{OI-6300_overview_combi}
\caption{Overview of the parameter space of centroid velocities and FWHM of the components of the combined [OI]~6300 MHD + photoevaporation profiles. Components from the MHD+photoevaporation combination models are marked with a triangle, those from the photoevaporation model with a circle. The plusses mark observations by \citetalias{Banzatti2018}. The colors indicate the type of the components: Green: HVC, red: BLVC, blue: NLVC, grey: SC, pink: SCJ}
\label{fig:OI-6300_overview_combi}
\end{figure*}
\subsection{Model limitations} \label{sec:discussion:limitations}
Our MHD model, while to a good degree consistent with observations, uses many simplifying assumptions. Most obviously, it lacks a thick disc which will certainly affect the velocity and density structure of the wind at larger radii. This could significantly affect the line profiles, especially in the case of high accretion luminosities, when the emission regions reach farther into that area. Moreover we have seen that the model tends to underestimate the line luminosities and to overestimate the outflow velocities despite the relatively large inner radius. Detailed numerical models with much lower inner radii and lower floor densities than existing models would be valuable to accurately calculate synthetic line profiles from MHD winds. Models that combine MHD with photoevaporation with an adjustable ratio between them, such as the model by \citet{Rodenkirch2019a} could be particularly useful to determine the relative importance of MHD and photoevaporative winds.
Our synthetic line profiles have all been calculated from wind models with fixed parameters, with the luminosity of the illuminating UV component being the only variable. Although \citet{Ercolano2010} have shown that the X-ray luminosity has no effect on the luminosities of the collisionally excited lines, it does have an effect on the photoevaporative wind mass-loss rates in the sense that a star with a higher X-ray luminosity is able to drive more rigorous winds \citep{Picogna2019}. Similarly, the stellar mass can affect the mass-loss rates of a photoevaporative wind (Picogna et al., in preparation). In a more extensive study that includes multiple wind models calculated with different initial conditions, we would expect the synthetic profile components to cover an even larger parameter space of observations.
\section{Correlation analysis of NLVC and BLVC component properties with L\texorpdfstring{$_{\mathrm{acc}}$}{}} \label{sec:appendix:correlations}
\begin{figure*}
\centering
\includegraphics[width=.9\textwidth]{expected_corr_comps}
\caption{Black dashed lines: correlations between the component properties. Orange solid lines: correlations as we would expect them if they were solely a result of the correlations with the accretion luminosity. The correlations and observations in the left and middle panels are those reported by \citetalias{Banzatti2018}, the observations in the right panel by \citet{Fang2018}. The black dashed line in the right panel was obtained in this work as described in appendix \ref{sec:appendix:correlations}.}
\label{fig:expected_corr_comps}
\end{figure*}
As described in section \ref{sec:results}, \citetalias{Banzatti2018} have found that the centroid velocities, FWHM and equivalent widths of the NLVCs correlate with those of the BLVCs. They also found that the centroid velocities and FWHM of both components have similar correlations with the accretion luminosity and another correlation between component luminosity and accretion luminosity is well known (section \ref{sec:introduction}). We will show here that the correlations between NLVC properties and L$_{\mathrm{acc}}$ and between BLVC properties and L$_{\mathrm{acc}}$ are indeed sufficient to explain the correlation between BLVC and NLVC properties. If the component property $y$ is correlated with the accretion luminosity L$_{\mathrm{acc}}$ via
\begin{equation}
y_{\mathrm{BLVC}} = a + b ~log\big(\mathrm{L_{acc}}\big)
\end{equation}
and
\begin{equation}
y_{\mathrm{NLVC}} = c + d ~log\big(\mathrm{L_{acc}}\big)
\end{equation}
and if we assume that $y$ is not correlated with another variable, we would expect that
\begin{equation}
y_{\mathrm{BLVC}} = a - \frac{b}{d} \big(c + y_{\mathrm{NLVC}}\big).
\end{equation}
\begin{table}
\centering
\input{tables/fits_corr_lacc.tex}
\caption{Fit parameters for the correlations of the centroids, FWHMs and component luminosities with the accretion luminosity and for the correlations between the LVC properties.}
\label{tab:fits_corr_lacc}
\end{table}
We tested this expectation against the observed correlations using the fit parameters listed in table \ref{tab:fits_corr_lacc}. Since \citetalias{Banzatti2018} do not report fit parameters for the correlation between component luminosity and accretion luminosity we use the parameters reported by \citet{Fang2018} and perform a fit to the BLVC - NLVC luminosity correlation in their sample to test the connection between the component luminosities. The relation we found is
\begin{equation}
log\big(\mathrm{L_{BLVC}}\big) = -6.48 + 4.31 log\big(\mathrm{L_{NLVC}}\big).
\end{equation}
Errors are not stated, because we lack the uncertainties of the underlying data, but for the purpose of this test we can safely ignore them. The results are shown in figure \ref{fig:expected_corr_comps}, which shows that the connections between NLVC and BLVC properties can indeed be explained by their correlations with the accretion luminosity.
\section{Complementary figures and tables}\label{sec:appendix:fits}
\begin{table*}
\centering
\input{tables/fits_pe.tex}
\caption{Centroid velocities, FWHM, luminosity and type of the individual components obtained by fitting the profiles from the photoevaporative wind models.}
\label{tab:fits_pe}
\end{table*}
\begin{table*}
\centering
\input{tables/fits_mhd_1.tex}
\caption{Centroid velocities, FWHM, luminosity and type of the individual components obtained by fitting the [OI] profiles from the MHD wind models. The fits to the [SII] profiles are listed in table \ref{tab:fits_mhd_pt2}.}
\label{tab:fits_mhd_pt1}
\end{table*}
\begin{table*}
\centering
\input{tables/fits_mhd_2.tex}
\caption{Centroid velocities, FWHM, luminosity and type of the individual components obtained by fitting the [SII] profiles from the MHD wind models. The fits to the [OI] profiles are listed in table \ref{tab:fits_mhd_pt1}.}
\label{tab:fits_mhd_pt2}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=.3\textwidth]{inclinations}
\caption{Number of objects in the sample of observations with a given inclination.}
\label{fig:inclinations}
\end{figure*}
\section{Introduction}\label{sec:introduction}
\input{1-introduction}
\section{Methods}\label{sec:methods}
\input{2-methods}
\section{Results}\label{sec:results}
\input{3-results}
\section{Discussion}\label{sec:discussion}
\input{4-discussion}
\section{Conclusions}\label{sec:conclusions}
\input{5-conclusions}
\section*{Acknowledgements}\label{sec:acknowledgements}
\input{6-acknowledgements}
\bibliographystyle{mnras}
| 5f199187012d678c2fc4f883fdcbd191089abd7d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
\input{02_introduction.tex}
\section{Privacy-Protection Drone Patrol System}
\label{sec:main}
\input{03_system_main.tex}
\section{Approach, Modifications, and Algorithm}
\label{sec:approach_modifications}
\input{03_1_Approach_and_Modification.tex}
\section{Evaluation}
\label{sec:Experiments}
\input{04_experiments.tex}
\section{Conclusion}
\label{sec:Conclusion}
\input{05_conclusion.tex}
\bibliographystyle{ieeetr}
\subsection{Related Work}
\subsubsection{Removal of Privacy Sensitive Information}
In robot cameras, the privacy infringement has attracted attention to develop a method removing the privacy-sensitive information in images \cite{PlaceAvoider, Jason, MU_Low_Anonymizer}.
In \cite{PlaceAvoider}, the authors introduced scene recognition from a image.
The scheme determines if a person is in a privacy-sensitive location.
If a image is taken in a privacy-sensitive place, the proposal allows a camera device to be automatically turned off.
However, faces are still exposed in privacy-insensitive places, and thus this scheme is not suitable for patrol drone visions.
Jason et al. \cite{Jason} developed the privacy preserving action detection via a face modifier by using Generative Adversarial Networks (GANs).
They proposed pixel-level modifications to change each person's face with minimal effect on action recognition performance.
The proposed generator modifies each pixel in a original face to remove features of the face.
To train the generator, the authors design their discriminator based on a face identification network that recognizes who s/he is.
The generator learns the way to make the discriminator believe that a generated face is different from the original image.
However, the generator tends to replace a input face to another face in the training dataset.
In other words, the problem is that the generator doesn't learn how to change a face not in the training dataset.
Moreover, the generator observes pixels of a face for modification, which means that a modified face is generated based on the original face.
This approach could still leave some information of a original face in a modified face.
The work \cite{MU_Low_Anonymizer} proposed a dynamic resolution face detection architecture to blur faces.
The framework detects faces from extreme low resolution images via the proposed deep learning-based algorithm.
Except for the detected faces, other privacy-insensitive pixels are enhanced to high resolution.
Hence, in result images, only faces are blurred, which protects privacy-sensitive parts while preserving the performance of robot perception.
However, in the case that a face are big in a frame, an intimate person can recognize who the person is in the frame even if the face is blurred.
In addition, it would be possible not to detect a face in a low-resolution image, and then this scheme couldn't protect a person's privacy.
\subsubsection{Generative Adversarial Networks}
Generative Adversarial Networks (GANs) have had impressive success in generating realistic images \cite{GAN_Goodfellow}.
The goal of this learning framework is to train a neural network to model a image distribution in an unsupervised manner.
The trained network can generate a fake image indistinguishable from a real image.
This training approach have been adopted to image-to-image translation \cite{Huang_18, Isola_17, Karacan_16, Karacan_18, Liu_17, Zhao_19, Zhu_17, Zhu_17_2, Park_19}.
Those works learn a mapping from input to output images, meaning that a input image is translated to a image in a different image distribution.
To construct our training architecture for obtaining deep-learning networks for our purpose, we have adopted the latest two works \cite{Park_19, Zhu_17}.
By using the training framework in \cite{Zhu_17}, we make a generator to translate a photorealistic image to a segmentation mask, and the work \cite{Park_19} is used to train another generator that converts the resultant segmentation mask to a photorelistic image.
\subsubsection{SLAM} To verify that our anonymization method has no effect on vision-based robot perception, we utilize simultaneous localization and mapping (SLAM) techniques.
By using SLAM, in an unknown environment, a robot constructs a map around itself and localizes itself in the resultant map.
Hence, in a video frame, manipulation of some pixels could affect the performance of SLAM since a map is drawn by extracting feature points of lines, edges, and corners of objects in images.
In this work, ORB-SLAM2 \cite{ORB-SLAM2} is implemented on our system, which is one of the most popular algorithms of the vision-based SLAM.
\subsection{Face Anonymizing Approach}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Face_Anonymization_Procedure.eps}
\caption{Our Approach to Anonymize Faces in a Video Frame}
\label{Fig:Anonymization_Approach}
\end{figure*}
Fig.~\ref{Fig:Anonymization_Approach} illustrates our anonymization approach for our system.
To anonymize faces, a companion computer has three different networks: Face detection network, Segmentation network, and Synthesis network.
The face detection network operates to detect faces whenever a video frame is fetched to the companion computer.
Detected faces are cropped and resized to meet the required input size of the segmentation network.
Via the segmentation network, face's images are translated to semantic images.
Note that the semantic images have no privacy information but the outline of detected faces is still maintained in the resultant images.
The synthesis network generates photorelistic images based on the semantic images.
Finally, the phtorelistic images replace the original faces.
Since semantic images still have the outline of each facial component, phtorelistic images could retain facial expressions.
\subsection{Training Architecture for Segmentation and Synthesis Networks}
\if\mycmd1
\begin{figure*}[t]
\centering
\includegraphics[width=0.8\paperwidth]{Training_Architecture_3.eps}
\caption{Training Architecture for Segmentation and Synthesis Networks}
\label{Fig:Training_Architecture}
\end{figure*}
\else
\begin{figure}[t]
\centering
\includegraphics[width=0.99\columnwidth]{Training_Architecture_4.eps}
\caption{Training Architecture for Segmentation and Synthesis Networks}
\label{Fig:Training_Architecture}
\end{figure}
\fi
Fig.~\ref{Fig:Training_Architecture} presents our training architecture for segmentation and synthesis networks.
To train those networks, we combine CycleGAN and GauGAN each of which is one of the spotlight image-to-image translation frameworks using GAN.
The segmentation network is trained by CycleGAN's framework, where Generator $G$ called segmentation generator is to generate a semantic images from a photorealistic image.
The synthesis network is obtained from GauGAN's framework.
Generator $G^\text{s}$ will make a photorealistic image from the output of segmentation generator $G$.
To obtain well-trained networks suitable for anonymization, we make modifications to generators' and discriminators' losses of both translation frameworks.
\subsubsection{Training Architecture Model}
To explain our modifications, we formulate our training architecture as follows.
For training samples, $X$ and $Y$ denote a photorelistic domain and a semantic domain, respectively.
For each domain, training samples are represented by $\{x_i\}_{i=1}^{N}$ and $\{y_j\}_{j=1}^{N}$.\footnote{In this work, $x_i$ is paired with a corresponding $y_j$, and thus both domains have the same number of samples.}
Samples of each domain are followed by a data distribution, which represents $x \sim p_\text{data}(x)$ and $y \sim p_\text{data}(y)$, respectively.
In this architecture, we have three generators, $G$, $F$, and $G^\text{s}$.
Each generator is a mapping function: $G: X \rightarrow Y$, $F: Y \rightarrow X$, and $G^\text{s}: Y \rightarrow X$.
For adversarial networks of those generators, there are discriminators $D_X$, $D_Y$, and $D_X^\text{s}$ each of which distinguishes if a input is from a data distribution or is generated by a generator.
$D_Y$, $D_X$, and $D_X^\text{s}$ examine output of $G$, $F$, and $G^\text{s}$, respectively.
\subsubsection{Well-Known Basic Definitions}
In this subsection, we introduce well-known definitions of losses \cite{Zhu_17, Park_19}.
Based on the losses, we will describe our modifications.
\textbf{Adversarial Loss}:
for each generator and discriminator pair, the \textit{adversarial loss} is defined as follows.
\if\mycmd1
\begin{equation}
\label{eq:adversarial_loss}
\mathcal{L}_\text{adv}(G, D_Y, X, Y) = \mathbb{E}_{x \sim p_\text{data}(x)}[\text{log}(1-D_Y(G(x)))] + \mathbb{E}_{y \sim p_\text{data}(y)}[\text{log}(D_Y(y))],
\end{equation}
\else
\begin{align}
\label{eq:adversarial_loss}
\mathcal{L}_\text{adv}(G, D_Y, X, Y) &= \mathbb{E}_{x \sim p_\text{data}(x)}[\text{log}(1-D_Y(G(x)))] \nonumber \\
&+ \mathbb{E}_{y \sim p_\text{data}(y)}[\text{log}(D_Y(y))],
\end{align}
\fi
where $G(x)$ is a generated image by a generator $G$.
$G$ tries to make $D_Y$ as difficult as possible to distinguish generated samples $G(x)$ from real samples $y$ whereas $D_Y$ should not be deceived by $G$.
The relationship can be formulated as $\min_G \max_{D_Y} \mathcal{L}_{\text{adv}}(G, D_Y, X, Y)$.
Hence, the generator actually should minimize the following loss.
\if\mycmd1
\begin{align}
\label{eq:loss_adv_G}
\mathcal{L}_{\text{adv}, G}(G, D_Y, X, Y) &= \mathbb{E}_{x \sim p_\text{data}(x)}[\text{log}(1-D_Y(G(x)))].
\end{align}
\else
\begin{align}
\label{eq:loss_adv_G}
\!\!\!\!\mathcal{L}_{\text{adv}, G}(G, D_Y, X, Y) = \mathbb{E}_{x \sim p_\text{data}(x)}[\text{log}(1\!-\!D_Y(G(x)))].
\end{align}
\fi
For other pairs, ($F$, $D_X$) and ($G^\text{s}$, $D_X^\text{s}$), the adversarial loss can be obtained by replacing ($G$, $D_Y$) in (\ref{eq:adversarial_loss}) with ($F$, $D_X$) and ($G^\text{s}$, $D_X^\text{s}$).
In addition, in (\ref{eq:adversarial_loss}), $X$ and $Y$ are replaced with $Y$ and $X$.
\textbf{Cycle-Consistency Loss}: the \textit{cycle-consistency loss} \cite{Zhu_17} is defined as
\begin{equation}
\mathcal{L}_\text{cyc}(G, F) = \mathcal{L}_{ \text{cyc}, G}(F) + \mathcal{L}_{\text{cyc}, F}(G),
\end{equation}
where $\mathcal{L}_{G, \text{cyc}}(F)$ and $\mathcal{L}_{F, \text{cyc}}(G)$ is a cycle-consistency loss for each generator, which is defined as
\begin{align}
\mathcal{L}_{\text{cyc}, G}(F) &= \mathbb{E}_{x \sim p_\text{data}(x)}[\lVert F(G(x)) - x \rVert_1], \\
\mathcal{L}_{\text{cyc}, F}(G) &= \mathbb{E}_{y \sim p_\text{data}(y)}[\lVert G(F(y)) - y \rVert_1].
\end{align}
This cycle-consistency loss is used to induce a sample $x_i$ to be mapped to a desired sample $y_j$.
Note that the adversarial loss guarantees that via a learned mapping function samples in a domain $X$ are mapped to samples in a domain $Y$, but the learned mapping function cannot translate a sample $x_i$ to a intended $y_j$.
In other words, the adversarial loss can guarantee translation between data distributions.
Hence, to obtain a mapping function between individual samples, the cycle-consistency loss should be used in the training procedure for generators.
\textbf{Multi-Scale Discriminators' Feature Loss}:
multiple discriminators are utilized to distinguish between a sample $y_j$ and a synthesized output $G(x_i)$ \cite{Park_19}.
$M$ Discriminators are trained to distinguish $y_j$ and $G(x_i)$ at $M$ different scales, which allows each discriminator to examine $y_j$ and $G(x_i)$ at a different view.
As the size of $y_j$ and $G(x_i)$ becomes smaller, a discriminator has a wider view of $y_j$ and $G(x_i)$ since receptive field sizes of all the discriminators are the same.
By using multiple discriminators, for a generator $G^\text{s}$, a GAN feature matching loss is defined as follows.
\if\mycmd1
\begin{equation}
\label{eq:loss_FM_G}
\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}) = \sum_{k=1}^{M} \frac{1}{M} \mathbb{E}_{(y,x)} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i} \lVert D_{X,k}^{\text{s}, i}(x) - D_{X,k}^{\text{s}, i}(G^\text{s}(y)) \rVert_{1} \Bigr ],
\end{equation}
\else
\begin{align}
\label{eq:loss_FM_G}
\!\!\!\!&\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}) \nonumber \\
\!\!\!\!&= \!\! \sum_{k=1}^{M} \!\frac{1}{M} \mathbb{E}_{(y,x)} \!\Bigl [ \sum_{i=1}^{T} \frac{1}{N_i} \lVert D_{X,k}^{\text{s}, i}(x) \!-\! D_{X,k}^{\text{s}, i}(G^\text{s}(y)) \rVert_{1} \! \Bigr ],
\end{align}
\fi
where $\mathbb{E}_{(y, x) \sim p_\text{data}(y, x)} \triangleq \mathbb{E}_{(y,x)}$ for simplicity.
$M$ is the number of discriminators, $D_{X, k}^\text{s}(\cdot)$ is the $M$-th discriminator, and $D_{X, k}^{\text{s}, i}$ is denoted as the $i$-th layer feature extractor of $D_{X, k}^\text{s}(\cdot)$.
$N_i$ means the number of elements in each layer, and $T$ is the number of feature layers.
Note that $G^\text{s}$ can learn how to translate a semantic image to a photorealistic image at both coarse and fine views since discriminators distinguish $x$ and $G^\text{s}(y)$ at $M$ different views.
\textbf{VGG Perceptual Loss}: the work \cite{Park_19} utilizes the perceptual loss in \cite{Johnson_16}.
The VGG perceptual loss is obtained by the 19-layer VGG network \cite{vgg}.
The VGG perceptual loss is defined as
\if\mycmd1
\begin{equation}
\label{eq:loss_VGG_G}
\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y))=\sum_{i \in S_\text{I}} \frac{1}{C_i H_i W_i}\lVert \psi_i(G^\text{s}(y)) - \psi_i(x) \rVert_{1},
\end{equation}
\else
\begin{equation}
\label{eq:loss_VGG_G}
\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y))=\sum_{i \in S_\text{I}} \frac{\lVert \psi_i(G^\text{s}(y)) - \psi_i(x) \rVert_{1}}{C_i H_i W_i},
\end{equation}
\fi
where $S_\text{I}$ is the set including VGG's layer indexes, $\psi$ is the 19-layer VGG network and $\psi_i$ is denoted as the $i$-th layer of $\psi$.
For $\psi_i$, $C_i$, $H_i$, and $W_i$ are the number of channels, the height, and the width, respectively.
By minimizing $\mathcal{L}_\text{VGG}(\cdot)$, $G^\text{s}$ can generate a photorealistic image $G^\text{s}(y)$, visually indistinguishable from $x$ in the feature-level perspective.
\subsection{Our Modifications for Anonymization System}
\subsubsection{Modification on Segmentation Generator's Loss}
The goal of a segmentation generator $G$ is to generate semantic images with which a synthesis generator $G^\text{s}$ makes well-synthesized images.
\textbf{Modification}: to consider the performance of $G^\text{s}$ in the loss of $G$, we define the loss of $G$ as follows.
\begin{align}
\label{eq:loss_G_segmentation}
&\mathcal{L}_{G}(G, F, D_Y, G^\text{s}) \nonumber\\
&= \underbrace{\mathcal{L}_\text{\text{adv}, G}(G, D_Y, X, Y) + \lambda_\text{cyc} \mathcal{L}_{\text{cyc}, G}(F)}_{\text{the original loss of $G$}} \nonumber \\
& +
\underbrace{\lambda_\text{s} \mathcal{L}_{G^\text{s}}(G^\text{s}, D_{X, 1}^\text{s}, \ldots ,D_{X, M}^\text{s}, G) + \lambda_\text{dist} \mathcal{L}_\text{dist}(G)}_{\text{the newly added term}},
\end{align}
where $\mathcal{L}_\text{dist}(G) = \mathbb{E}_{x \sim p_\text{data}(x)}[\lVert G(x)-y \rVert_{1}]$, and $\mathcal{L}_\text{dist}(G)$ could further reduce the space of possible mapping functions with $\mathcal{L}_{\text{cyc}, G}(F)$.
$\mathcal{L}_{G^\text{s}}$ is the loss of a synthesis generator $G^\text{s}$, and will be explained in Section~\ref{sec:Modifications_on_Synthesis} in detail.
In addition, $\lambda_\text{cyc}, \lambda_\text{s}$, and $\lambda_\text{dist}$ control the relative importance of each loss.
By adding $\mathcal{L}_{G^\text{s}}(G^\text{s}, D_{X, 1}^\text{s}, \ldots ,D_{X, M}^\text{s}, G)$, the generator $G$ will be trained to generate a semantic image minimizing the loss of $G^\text{s}$.
\textbf{Objective of Segmentation Learning Part}:
Hence, the full objective of segmentation-learning part is defined as:
\if\mycmd1
\begin{align}
\label{eq:objective_segmentation}
\mathcal{L}_\text{seg}(G, F, D_X, D_Y, G^\text{s}) &= \mathcal{L}_{G}(G, F, D_Y, G^\text{s}) + \mathbb{E}_{y\sim p_\text{data}(y)}[D_Y(y)]
\nonumber \\
&+ \mathcal{L}_{F}(G, F, D_X) +\mathbb{E}_{x\sim p_\text{data}(x)}[D_X(x)],
\end{align}
\else
\begin{align}
\label{eq:objective_segmentation}
&\mathcal{L}_\text{seg}(G, F, D_X, D_Y, G^\text{s}) \nonumber \\
&= \mathcal{L}_{G}(G, F, D_Y, G^\text{s}) + \mathbb{E}_{y\sim p_\text{data}(y)}[D_Y(y)]
\nonumber \\
&+ \mathcal{L}_{F}(G, F, D_X) +\mathbb{E}_{x\sim p_\text{data}(x)}[D_X(x)],
\end{align}
\fi
where $\mathcal{L}_{F}(G,\! F,\! D_X) \!\!=\!\! \mathcal{L}_{\text{adv}, F}(F, \!D_X, \!Y, \!X) + \lambda_\text{cyc} \mathcal{L}_{\text{cyc}, F}(G)$.
The segmentation learning part will solve (\ref{eq:objective_segmentation}) as follows:
\begin{equation}
\label{eq:solve_segmentation_objective}
G^{*}, F^{*} = \operatornamewithlimits{argmin}\limits_{G, F} \max_{D_X, D_Y} \mathcal{L}_\text{seg}(G, F, D_X, D_Y, G^\text{s}).
\end{equation}
\subsubsection{Modifications on Synthesis Generator's Loss}
\label{sec:Modifications_on_Synthesis}
There are two main challenges to hinder the learning of synthesis generator to anonymize faces.
We introduce loss of a synthesis generator $G^\text{s}$, and then explain each challenge.
To obtain a synthesis generator to achieve our system purpose, we have a modification on the loss of $G^\text{s}$ and the loss of $D_{X,k}^\text{s}, \forall k$.
In addition, we modify the way to train the discriminator $D_{X,k}^\text{s}$.
In \cite{Park_19}, by using (\ref{eq:loss_adv_G}), (\ref{eq:loss_FM_G}), and (\ref{eq:loss_VGG_G}), the loss of a synthesis generator can be written as
\if\mycmd1
\begin{align}
\label{eq:loss_G_synthesis}
\mathcal{L}_{G^\text{s}}(G^\text{s}, D_{X,1}^\text{s}, \ldots D_{X,M}^\text{s}, G) &= \sum_{k=1}^{M} \frac{1}{M} \Bigl \{ \mathcal{L}_{\text{adv}, G^\text{s}}(G^\text{s}, D_{X, k}^\text{s}, Y, X, G(x)) + \mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,k}^\text{s}) \Bigr \} \nonumber \\
&+ \mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(G(x))),
\end{align}
\else
\begin{align}
\label{eq:loss_G_synthesis}
&\mathcal{L}_{G^\text{s}}(G^\text{s}, D_{X,1}^\text{s}, \ldots D_{X,M}^\text{s}, G) \nonumber \\
&=\!\! \sum_{k=1}^{M} \! \frac{1}{M} \Bigl \{ \! \mathcal{L}_{\text{adv}, G^\text{s}}(G^\text{s}, D_{X, k}^\text{s}, Y, X, G(x)) + \mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,k}^\text{s}) \! \Bigr \} \nonumber \\
&+ \mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(G(x))),
\end{align}
\fi
where we introduce for simplicity $\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,k}^\text{s})=\mathbb{E}_{(y,x)} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i} \lVert D_{X,k}^{\text{s}, i}(x) - D_{X,k}^{\text{s}, i}(G^\text{s}(y)) \rVert_{1} \Bigr ]$ in (\ref{eq:loss_FM_G}).
In addition, $\mathcal{L}_{\text{adv}, G^\text{s}}(G^\text{s}, D_{X, k}^\text{s}, Y, X, G(x))$ is redefined as
\if\mycmd1
\begin{equation}
\label{eq:loss_synthesis_G_redefined}
\mathcal{L}_{\text{adv}, G^\text{s}}(G^\text{s}, D_{X, k}^\text{s}, Y, X, G(x)) = \mathbb{E}_{y\sim p_\text{data}(y)} \Bigl [\text{log}(1-D_{X,k}^\text{s}(G^\text{s}(G(x)))) \Bigr ].
\end{equation}
\else
\begin{align}
\label{eq:loss_synthesis_G_redefined}
&\mathcal{L}_{\text{adv}, G^\text{s}}(G^\text{s}, D_{X, k}^\text{s}, Y, X, G(x)) \nonumber \\
&= \mathbb{E}_{\hat{y}\sim p_\text{data}(y)} \Bigl [\text{log}(1-D_{X,k}^\text{s}(G^\text{s}(\hat{y}))) \Bigr ],
\end{align}
\fi
where $\hat{y}=G(x)$.
\textbf{Challenge in VGG Perceptual Loss}:
in the synthesis-learning part, the loss (\ref{eq:loss_G_synthesis}) should be minimized to train the generator $G^\text{s}$.
The minimization leads to reduce $\mathcal{L}_{\text{VGG}}(\psi, x, G^\text{s}(G(x)))$, and thus the distance between features of $x$ and $G^\text{s}(G(x))$ is also reduced during the training of $G^\text{s}$.
As a result, the generator $G^\text{s}$ is trained to generate a photorealistic image $G^\text{s}(G(x))$ that can be almost the same as the original photorealistic image $x$.
This trained generator cannot be utilized for our face-anonymizing system.
\textbf{Modification on VGG Perceptual Loss}:
to prevent the distance between $G^\text{s}(G(x))$ and $x$ from being reduced to a very small value, we introduce margins to the VGG perceptual loss (\ref{eq:loss_VGG_G}) as follows.
\if\mycmd1
\begin{equation}
\label{eq:loss_VGG_G_with_margin}
\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y), S_\text{I})=\sum_{i \in S_\text{I}} \max \Bigl (0, \frac{1}{C_i H_i W_i} \lVert \psi_i(G^\text{s}(y)) - \psi_i(x) \rVert_{1} - \epsilon_{m(i)} \Bigr),
\end{equation}
\else
\begin{align}
\label{eq:loss_VGG_G_with_margin}
&\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y), S_\text{I})\nonumber \\
&=\sum_{i \in S_\text{I}} \max \Bigl (0, \frac{\lVert \psi_i(G^\text{s}(y)) - \psi_i(x) \rVert_{1}}{C_i H_i W_i} - \epsilon_{m(i)} \Bigr),
\end{align}
\fi
where $\Upsilon = \{ \epsilon_1, \ldots, \epsilon_{|S_\text{I}|} \}$ and $|S_\text{I}|$ is the number of elements in $S_\text{I}$.
$m(i)$ is a mapping function to find for $S_\text{I}$ a VGG's layer index corresponding to $i$.
Note that a margin value allows the distance between the $i$-th VGG layers for $x$ and $G^\text{s}(G(x))$ to be at least $\epsilon_i$ value.
Hence, a photorealistic image $G^\text{s}(G(x))$ can have different features from features of the original image $x$, which could make $G^\text{s}(G(x))$ look different from $x$.
\textbf{Challenge in Adversarial Loss and Multi-Scale Discriminators' Feature Loss}:
to explain our additional modifications, we need to comprehend about how a discriminator $D_{X, k}^\text{s}$ works.
Based on the understanding, we describe a hindrance to the learning of our synthesis generator $G^\text{s}$.
In addition, we modify the adversarial losses of $G^\text{s}$ and $D_{X, k}^\text{s}$, and the multi-scale discriminators' feature loss of $G^\text{s}$.
To minimize (\ref{eq:loss_synthesis_G_redefined}), $G^\text{s}$ should make a synthesized face $G^\text{s}(G(x))$ look like a face in the training dataset, and thus tends to translate $G(x)$ to $x$ that is what our system should anonymize.
Specifically, in (\ref{eq:loss_synthesis_G_redefined}), a discriminator $D_{X,k}^\text{s}$ examines $G^\text{s}(G(x))$ to determine if $G^\text{s}(G(x))$ is from the training dataset $X$ or is arbitrarily generated.
The generated image $G^\text{s}(G(x))$ contains an entire face, and thus $D_{X,k}^\text{s}$ is trained to determine whether the entire face in $G^\text{s}(G(x))$ is from the training dataset.
As a result, to deceive $D_{X,k}^\text{s}$, $G^\text{s}$ is trained to generate $x$ from $G(x)$.
\textbf{Modifications on Adversarial Loss and Multi-Scale Discriminators' Feature Loss}:
to prevent that $G^\text{s}$ regenerates the almost same face as $x$, we have modification on the adversarial losses of $G^\text{s}$ and $D_{X,k}^\text{s}$.
The reason that $G^\text{s}$ reproduces $x$ is because a discriminator $D_{X,k}^\text{s}$ examines if the entire face in $G^\text{s}(G(x))$ is from a training dataset including $x$.
In other words, to deceive $D_{X,k}^\text{s}$ checking an entire face, $G^\text{s}$ necessarily makes a face from the domain $X$, which greatly reduces the space of possible mapping.
In addition, $G^\text{s}$ should produce a photorealistic face by maintaining the shape and location of each facial part in a semantic-face image, which also further reduces the space of possible mapping.
For expanding the space of possible mapping, we limit a discriminator $D_{X,k}^\text{s}$ to investigate each facial component, not entire face.
By applying the idea, the adversarial loss is rewritten as follows.
\if\mycmd1
\begin{align}
\label{eq:loss_adv_synthesis_modified}
\mathcal{L}_\text{adv}^{s}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, Y, X, S_\xi) &=
\sum_{k=1}^{M} \frac{1}{M} \Bigl \{ \mathbb{E}_{y\sim p_\text{data}(y)} \Bigl [ \sum_{i\in S_{\xi}} \frac{1}{|S_{\xi}|} \text{log}(1-D_{X,k}^\text{s}(\xi_{i}(G^\text{s}(\hat{y})))) \Bigr ] \Bigr \} \nonumber \\
&+
\sum_{k=1}^{M} \frac{1}{M} \Bigl \{ \mathbb{E}_{x\sim p_\text{data}(x)} \Bigl [ \sum_{i\in S_{\xi}} \frac{1}{|S_{\xi}|} \text{log}(D_{X,k}^\text{s}(\xi_i(x))) \Bigr ] \Bigr \},
\end{align}
\else
\begin{align}
\label{eq:loss_adv_synthesis_modified}
&\mathcal{L}_\text{adv}^{s}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, Y, X, S_\xi) \nonumber \\
&=
\sum_{k=1}^{M}\! \frac{1}{M} \! \Bigl \{ \mathbb{E}_{y\sim p_\text{data}(y)} \Bigl [ \sum_{i\in S_{\xi}} \!\!\frac{1}{|S_{\xi}|} \text{log}(1\!-\!D_{X,k}^\text{s}(\xi_{i}(G^\text{s}(\hat{y})))) \Bigr ] \! \Bigr \} \nonumber \\
&+
\sum_{k=1}^{M} \frac{1}{M} \Bigl \{ \mathbb{E}_{x\sim p_\text{data}(x)} \Bigl [ \sum_{i\in S_{\xi}} \frac{1}{|S_{\xi}|} \text{log}(D_{X,k}^\text{s}(\xi_i(x))) \Bigr ] \Bigr \},
\end{align}
\fi
where we denote the output of our segmentation generator as $\hat{y}=G(x)$, $\xi_i$ is the extractor to extract pixels corresponding to the label index $i$, $S_{\xi}$ is the set including extracted labels' index, and $|S_{\xi}|$ is the number of elements in $S_\xi$.
For example, if $i=2$ and the label index 2 indicates a nose in a face, all pixels in $\xi_2(G^\text{s}(\hat{y}))$ become zero except for the pixels corresponding to the nose.
According to (\ref{eq:loss_adv_synthesis_modified}), our modification allows a discriminator $D_{X,k}^\text{s}$ to examine a part of a face instead of observing all facial parts at a time.
This approach allows discriminators to learn the distribution of each facial component instead of learning the distribution of an entire face.
Hence, a discriminator tries to distinguish each part of a face in $G^\text{s}(\hat{y})$ from that of a face in $x$, which could widen the space of possible mapping in the entire face's point of view.
In addition, by setting $|S_\xi|<N_\text{f}$ where $N_\text{f}$ denotes the number of labels in a face, we make discriminators observe some parts of an entire face, and thus the generator $G^\text{s}$ could have wider space of possible mapping for the other parts not examined by discriminators.
\if\mycmd1
In the same vein, the multi-scale discriminators' feature loss can be also redefined as
\begin{equation}
\label{eq:loss_multi_scale_disc_feature_modified}
\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) =
\sum_{k=1}^{M} \frac{1}{M}
\underbrace{\mathbb{E}_{(y,x)} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i}
\sum_{j\in S_\xi} \frac{1}{|S_\xi|} \lVert D_{X,k}^{\text{s}, i}(\xi_j(x)) - D_{X,k}^{\text{s}, i}(\xi_j(G^\text{s}(\hat{y}))) \rVert_{1} \Bigr ]}_{\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)}.
\end{equation}
\else
In the same vein, the multi-scale discriminators' feature loss can be also redefined as (\ref{eq:loss_multi_scale_disc_feature_modified}).
\begin{figure*}
\begin{align}
\label{eq:loss_multi_scale_disc_feature_modified}
&\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) =
\sum_{k=1}^{M} \frac{1}{M}
\underbrace{\mathbb{E}_{(y,x)} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i}
\sum_{j\in S_\xi} \frac{1}{|S_\xi|} \lVert D_{X,k}^{\text{s}, i}(\xi_j(x)) D_{X,k}^{\text{s}, i}(\xi_j(G^\text{s}(\hat{y}))) \rVert_{1} \Bigr ]}_{\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)}.
\end{align}
\end{figure*}
\fi
In spite of our modifications on $\mathcal{L}_\text{adv}^\text{s}(\cdot)$ and $\mathcal{L}_{\text{FM}, G^\text{s}}(\cdot)$, there is still room for $G^\text{s}$ to learn to regenerate $x$ since $\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)$ still compares features of $G^\text{s}(G(x))$ and $x$, which are extracted by $D_{X,k}^\text{s}$.
\if\mycmd1
Hence, we slightly modify (\ref{eq:loss_multi_scale_disc_feature_modified}) as follows.
\begin{align}
\label{eq:loss_multi_scale_disc_feature_modified_2}
\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) =
\sum_{k=1}^{M} \frac{1}{M}
\underbrace{\mathbb{E}_{(y,x,\tilde{x})} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i}
\sum_{j\in S_\xi} \frac{1}{|S_\xi|} \lVert D_{X,k}^{\text{s}, i}(\xi_j(\tilde{x})) - D_{X,k}^{\text{s}, i}(\xi_j(G^\text{s}(\hat{y}))) \rVert_{1} \Bigr ]}_{\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)},
\end{align}
where $\tilde{x} \neq x$ but $\tilde{x}$ is in the same training dataset of $x$.
By the modification, $\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)$ compares features of $G^\text{s}(G(x))$ to features of $\tilde{x}$, which can help $G\text{s}$ learning to generate a different face from $x$.
\else
Hence, we slightly modify (\ref{eq:loss_multi_scale_disc_feature_modified}) to (\ref{eq:loss_multi_scale_disc_feature_modified_2}).
\begin{figure*}
\begin{align}
\label{eq:loss_multi_scale_disc_feature_modified_2}
\mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) =
\sum_{k=1}^{M} \frac{1}{M}
\underbrace{\mathbb{E}_{(y,x,\tilde{x})} \Bigl [ \sum_{i=1}^{T} \frac{1}{N_i}
\sum_{j\in S_\xi} \frac{1}{|S_\xi|} \lVert D_{X,k}^{\text{s}, i}(\xi_j(\textcolor{blue}{\tilde{x}})) - D_{X,k}^{\text{s}, i}(\xi_j(G^\text{s}(\hat{y}))) \rVert_{1} \Bigr ]}_{\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)},
\end{align}
\noindent\makebox[\linewidth]{\rule{17cm}{0.4pt}}
\end{figure*}
In (\ref{eq:loss_multi_scale_disc_feature_modified_2}), $\mathbb{E}_{(x,y,\tilde{x})\sim p_\text{data}(x, y, x)} \triangleq \mathbb{E}_{x,y,\tilde{x}}$, and $\tilde{x} \neq x$ but $\tilde{x}$ is from the same training dataset of $x$.
By the modification, $\mathcal{L}_{\text{FM},G^\text{s}}(D_{X,k}^\text{s}, S_\xi)$ compares features of $G^\text{s}(G(x))$ to features of $\tilde{x}$, which can help $G^\text{s}$ learning to generate a different face from $x$.
\fi
\textbf{Objective of Synthesis Learning Part}: Based on (\ref{eq:loss_VGG_G_with_margin}), (\ref{eq:loss_adv_synthesis_modified}), and (\ref{eq:loss_multi_scale_disc_feature_modified_2}), our full objective of synthesis-learning part is defined as:
\if\mycmd1
\begin{align}
\label{eq:objective_synthesis_learning}
\mathcal{L}_\text{syn}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi, G) &= \mathcal{L}_\text{adv}^\text{s}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, Y, X, S_\xi)
+ \mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) \nonumber \\
&+
\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y), S_\text{I}) +
\mathcal{L}_{\text{cyc}, G^\text{s}}(G),
\end{align}
\else
\begin{align}
\label{eq:objective_synthesis_learning}
&\mathcal{L}_\text{syn}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi, G) \nonumber \\
&= \mathcal{L}_\text{adv}^\text{s}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, Y, X, S_\xi) \nonumber \\
&+ \mathcal{L}_{\text{FM}, G^\text{s}}(D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi) \nonumber \\
&+
\mathcal{L}_\text{VGG}(\psi, x, G^\text{s}(y), S_\text{I}) +
\mathcal{L}_{\text{cyc}, G^\text{s}}(G),
\end{align}
\fi
where $\mathcal{L}_{\text{cyc}, G^\text{s}}(G) = \mathbf{E}_{x \sim p_\text{data}(x)}[\lVert G(G^\text{s}(G(x))) - G(x) \rVert_{1}]$ that allows a synthesized image $G^\text{s}(G(x))$ to maintain the shape and location of each facial part in $x$.
Since (\ref{eq:loss_multi_scale_disc_feature_modified_2}) compares features of $G^\text{s}(G(x))$ to those of $\tilde{x}$, the generator $G^\text{s}$ can make a synthesized image to remain the shape and location of each facial part in $\tilde{x}$ not in $x$.
Hence, $\mathcal{L}_{\text{cyc}, G^\text{s}}(G)$ helps $G^\text{s}$ to generate synthesized images to remain the shape and location of facial parts in $x$.
Hence, by $\mathcal{L}_{\text{cyc}, G^\text{s}}(G)$ and (\ref{eq:loss_multi_scale_disc_feature_modified_2}), $G^\text{s}$ can generate a synthesized face $G^\text{s}(G(x))$ including facial features of $\tilde{x}$ while maintaining the shape and location of facial components in $x$.
Finally, the synthesis learning part will solve (\ref{eq:objective_synthesis_learning}) as follows:
\begin{equation}
(G^\text{s})^* = \operatornamewithlimits{argmin}\limits_{G^\text{s}} \max\limits_{D_{X,k}^\text{s},\forall k} \mathcal{L}_\text{syn}(G^\text{s}, D_{X,1}^\text{s}, \ldots, D_{X,M}^\text{s}, S_\xi, G).
\end{equation}
\subsection{Training Procedure and Details}
\input{Algorithm_Training_Procedure}
\textbf{Procedure}: The overall training procedure is summarized in Algorithm~\ref{alg:training_procedure}, where $N_\text{data}$ is the number of data in the dataset $X$ and $Y$.
By repeating the training procedure, we obtain the optimized $G^*$ and $(G^\text{s})^*$.
\textbf{Details}: In segmentation learning part, for the segmentation generator $G$, we adopt the network architecture in \cite{Johnson_16} that is known for powerful neural-type transfer and use 9 resnet blocks for $256 \times 256$ images, which is applied equally to the generator $F$.
For the discriminators $D_X$ and $D_Y$, we use $70 \times 70$ PatchGANs \cite{Isola_17, Li_16, Ledig_17, Zhu_17}.
In synthesis learning part, we construct our network architecture by applying the Spectrum Norm \cite{Spectrum_Norm} to all the layers in both generator and discriminator.
For our synthesis generator, we use the SPADE generator in \cite{Park_19}.
Finally, we set $M=3$ for our discriminators.
For our training, we set $\lambda_\text{cyc}, \lambda_\text{s}, \lambda_\text{dist}$ to $10$ in (\ref{eq:loss_G_segmentation}).
A solver is set to the ADAM solver \cite{ADAM} with a batch size of 1.
For the solver, $\beta_1=0.5$ and $\beta_2=0.999$ in segmentation learning part and $\beta_1=0$ and $\beta_2=0.999$ in synthesis learning part.
\textbf{Training Dataset}: We conduct our training procedure with CelebA-HQ dataset \cite{CelebAMask-HQ}.
This dataset contains 30,000 high-resolution face images with 19 semantic classes.
In this work, we modify the semantic dataset by extracting 10 main facial components.
In our modified semantic dataset, semantic classes include skin, nose, eyes, eyebrows, ears, mouth, lip, hair, neck, and eyeglass.
In addition, we utilize a face detector to crop a face in high-resolution face images.
The reason that cropped face images are needed is that in our system faces in a video frame will be detected by using a face detector.
By conducting this preprocessing, we can train segmentation and synthesis generators, optimized for our system.
Moreover, we create and use various resolution images for a image.
Our segmentation generator requires a specific sized image as input, and thus detected face images should be resized to the specific size.
If the size of a detected face is smaller than the required size, a resized face image is low resolution.
In reality, our system can detect a face in various sizes, various resolutions.
To make our segmentation and synthesis generators work well with various resolutions, we utilize various resolution images for a face image during our training procedure.
\subsection{Face-Anonymizing Algorithm}
\input{Algorithm_Face_Anonymizing}
With the optimized segmentation and synthesis generators, our face-anonymizing procedure is conducted as in Algorithm~\ref{alg:face_anonymizing_algorithm}.
A companion computer conducts Algorithm~\ref{alg:face_anonymizing_algorithm} whenever it receives a video frame.
$f_\text{D}$ denotes detected faces, $N_\text{face}$ is the number of detected faces, $f_\text{A}$ includes all anonymized faces, and $R_\text{thr}$ is a threshold for the ratio of the face size to the image size.
Note that $f_\text{D}$ and $f_\text{A}$ also include the background as well as faces.
\textbf{Maintenance of Background}: In Algorithm~\ref{alg:face_anonymizing_algorithm}, $B(\cdot)$ is a function that makes all nonzero elements in a input semantic image 1.
In a semantic image, zero indicates the background.
$B^{-1}(\cdot)$ is the opposite function of $B(\cdot)$.
Hence, in Algorithm~\ref{alg:face_anonymizing_algorithm}, \textcircled{1} creates a image that mixes the anonymized face in $f_\text{A}$ and the background in $f_\text{D}$, which could preserve the background in the original image.
\textbf{Repetitive Face Detection}: In Algorithm \ref{alg:face_anonymizing_algorithm}, we repeat to conduct the face detector with detected faces until the size of a face in $f_\text{D}$ occupies $R_\text{thr}\times100$ of the size of $f_\text{D}$.
The reason for this repetitive detection is that our generators well anonymize a input image that is full of one face.
\begin{figure}[!t]
\centering
\includegraphics[width=1\columnwidth]{Repetitive_Detection.eps}
\caption{Bounding box regression and NMS}
\label{Fig:Repetitive_Detection}
\end{figure}
The face detector finds a face with the bounding box regression and non-maximum suppression (NMS).
The bounding box regression provides several bounding boxes (red boxes) on a face in Fig.~\ref{Fig:Repetitive_Detection}(a).
Then, NMS trims the bounding boxes to obtain a bounding box (the yellow box) in Fig.~\ref{Fig:Repetitive_Detection}(b).
In Fig.~\ref{Fig:Repetitive_Detection}(c), however, the resultant image is not suitable as input for our generators.
Hence, to increase the ratio of the face size to the image size, we repeatedly conduct the face detector on a face until we obtain a image in Fig.~\ref{Fig:Repetitive_Detection}(d).
\subsection{Drone Patrol System}
The drone system consists of the following components:
(1) Drone control computer operating motors so that the drone can move physically,
(2) Companion computer conducting face-anonymizing neural networks and performing SLAM,
(3) Wireless chipset receiving commands from the ground station and transmitting anonymized video frames to it,
(4) High-resolution camera recording a video;
In our drone, the companion computer is connected with the drone control computer, the wireless chipset, and the high-resolution camera.
We utilize ROS to allow all the components to communicate with each other.
Via wireless communication, the companion computer communicates with the ground station.
The ground station can transmit a command message to the companion computer.
The companion computer sends the received message to drone control computer controls.
The moving drone continuously records images via the camera, which is passed to our face-anonymzing networks implemented in the companion computer. The anonymized images are sent back to the ground station via the wireless chipset, and thus we can immediately check the results on the screen of the ground station.
At the same time, the anonymized images are processed by the ORB-SLAM2 algorithm.
\subsection{Face-Anonymizing Generators Evaluation}
\textbf{Face Detection}: Our system utilizes a lightweight but accurate face detector, called FaceBoxes \cite{zhang2017faceboxes}.
The computing speed is invariant no matter how many faces are in a image, and the accuracy was verified with various face datasets.
\textbf{Test Dataset}: We test our face-anonymizing generators on several datasets.
\begin{itemize}
\item \textit{CelebA-HQ}: This dataset contains 30,000 high-resolution face images.
We randomly select 1,500 images for our test. Note that the remaining 28,500 images are used for the training.
\item \textit{Helen}: This dataset has 2,330 face images \cite{Helen}.
In addition, this provides 2,330 annotation images to locate 8 facial components, which are skin, eyebrow, eye, nose, lip, inner mouth, and hair.
\item \textit{Facescrub}: This dataset is large face datset, which contains 106,863 face images of male and female 530 celebrities \cite{Facescrub}.
\item \textit{FaceForensic}: This dataset provides 1000 video sequences, which has been sourced from 977 youtube videos \cite{roessler2019faceforensicspp}.
All videos contain a mostly frontal face without occlusions.
\end{itemize}
\textbf{Qualitative Evaluation of Face Anonymization}:
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{Test_CelebA_0.eps}
\vspace{-0.15in}
\caption{Test Results on CelebA-HQ Dataset}
\label{Fig:Result_CelebA-HQ_part}
\vspace{-0.1in}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{Test_Helen_0.eps}
\vspace{-0.15in}
\caption{Test Results on Helen Dataset}
\label{Fig:Result_Helen_part}
\vspace{-0.1in}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{Test_FaceScrub_0.eps}
\vspace{-0.15in}
\caption{Test Results on FaceScrub Dataset}
\label{Fig:Result_FaceScrub_part}
\end{figure}
Figs.~\ref{Fig:Result_CelebA-HQ_part}, \ref{Fig:Result_Helen_part}, and \ref{Fig:Result_FaceScrub_part} provide the quality of our face-anonymizing generators.
We can confirm that our method produces well anonymized faces for diverse faces in CelebA-HQ, Helen, and FaceScrub dataset.
For each dataset, the additional results are shown in Figs.~\ref{Fig:Result_CelebA-HQ}, \ref{Fig:Result_Helen}, and \ref{Fig:Result_FaceScrub}.
\textbf{Quantitative Evaluation of Face Anonymization}:
To evaluate our anonymization system quantitatively, we utilize siamese network \cite{siamese_net}.
The siamese network is widely utilized to measure the dissimilarity between two images.
Via the siamese network, we measure how dissimilar an anonymized face and an original face are.
For our evaluation, we train a siamese network to calculate Euclidean distance between two faces.
The larger the euclidean distance is, the more dissimilar two faces are.
The inception resnet \cite{InceptionResnet} is adopted for the backbone network of the siamese network.
We perform 50 epochs of training on the CASIA-WebFace dataset \cite{CASIA}, which includes about 500,000 images, and the image sizes are $256 \times 256$.
\begin{table}[t]
\centering
\caption{Average Euclidean distance for each test dataset}
\label{table:siamese_result}
\begin{tabular}{|c||c|c||c|l|c|c|}
\hline
& \multicolumn{2}{c||}{Criterion} & \multicolumn{4}{c|}{Test Dataset} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Average\\ Euclidean\\ Distance\end{tabular}} &
Same &
Different &
\multicolumn{2}{c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}CelebA\\ HQ\end{tabular}}} &
\multirow{2}{*}{Helen} &
\multirow{2}{*}{FaceScrub} \\
& person & people & \multicolumn{2}{c|}{} & & \\ \cline{2-7}
& 1.15 & 2.13 & \multicolumn{2}{c|}{1.57} & 2.59 & 1.94 \\ \hline
\end{tabular}
\end{table}
Table~\ref{table:siamese_result} presents the average Euclidean distance for each test dataset.
In the criterion, the value for 'Same person' is average Euclidean distance between faces of the same person, and another value for 'Different people' is obtained between faces of different people.
Those values are measured during training the siamese network.
For each test dataset, the value is the average Euclidean distance between an original face and an anonymized face by our system.
Note that the larger the value is, the more dissimilar two faces are.
According to Table~\ref{table:siamese_result}, our anonymization system indeed can make a face that looks different from an original face.
For all test dataset, the average Euclidean distance is larger than that of criterion.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{Test_FaceForensic_3.eps}
\caption{Test Results on FaceForensic Dataset via Algorithm~\ref{alg:face_anonymizing_algorithm}}
\label{Fig:Result_FaceForensic_3}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=1.8\columnwidth]{Demo_Desk_1.eps}
\caption{SLAM Results on a original video and an anonymized video}
\label{Fig:SLAM_Results_Desk}
\vspace{-0.1in}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=2\columnwidth]{Demo_Drone.eps}
\caption{Our face-anonymizing system results on drone}
\label{Fig:Evaluation_on_drone}
\vspace{-0.1in}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=2\columnwidth]{Demo_Drone_SLAM.eps}
\caption{SLAM results on drone}
\label{Fig:Evaluation_SLAM_on_drone}
\vspace{-0.1in}
\end{figure*}
\subsection{Evaluation of Algorithm~\ref{alg:face_anonymizing_algorithm}}
To test Algorithm~\ref{alg:face_anonymizing_algorithm}, we conduct our algorithm on FaceForensic video dataset.
Fig.~\ref{Fig:Result_FaceForensic_3} presents results for various videos.
In each result for 'original' and 'Anonymized', a enlarged face image is attached for readability.
Those results confirm that our algorithm finds a face in a video frame, and then anonymizes the detected face well.
The additional results are shown in Figs.~\ref{Fig:Result_FaceForensic_1} and \ref{Fig:Result_FaceForensic_2}.
As a result, for various dataset, our privacy-protection drone vision system can remove privacy information of a person by anonymizing his/her face.
\subsection{SLAM Results}
In this section, we investigate the impact of our proposal on vision-based robot perception, ORB-SLAM2.
Fig.~\ref{Fig:SLAM_Results_Desk} shows two types of results:
\begin{itemize}
\item \textit{Frames with SLAM's features} are obtained via ORB-SLAM2, in the upper figures of Fig.~\ref{Fig:SLAM_Results_Desk}. The resultant images include green boxes that are feature points extracted by ORB-SLAM2.
The feature points indicate edges, lines, and corners of objects in a image.
\item \textit{PointCloud2's results} are drawn via RVIZ \cite{pointcloud2, rviz} based on the extracted feature points.
In the lower figures of Fig.~\ref{Fig:SLAM_Results_Desk}, the white points are stamped by the calculated distance between a ZED camera and extracted feature points.
\end{itemize}
In Fig.~\ref{Fig:SLAM_Results_Desk}, the upper figures present the comparison of feature extraction results on a original video frame and an anonymized video frame.
In both frames, the feature points are well extracted.
The extracted feature points of our anonymization scheme are almost the same as that of the original video.
In addition, we can notice that the face is also anonymized well by our scheme.
The lower figures present the distance between a camera and a face, measured by ORB-SLAM2.
In the red circle, the white points are extracted from the face.
Hence, the distance between those points and a ZED camera means the distance between the face and the camera.
Under our scheme, the distance is almost the same as that under the original video.
In summary, our anonymization scheme has no impact on the performance of ORB-SLAM2, which confirms that our system can preserve vision-based drone perception well.
In the drone perception, inaccuracy can indeed incur an accident where a drone hits a person's face.
Since our approach effectively protects the person's privacy with good perception, our anonymization scheme is certainly proper to privacy-protection drone patrol system.
\subsection{Drone Hardware Setup}
\textbf{Drone Hardware Specification}:
We build a customized drone each component of which is as follows.
We selected DJI F550 for our drone frame.
1137 T-motor V2 carbon fiber propeller and T-motor MN3110 KV 780 are selected as propellers and motors.
T-motor Air 40A is chosen as electronic speed controllers.
We utilize Holybro Pixhawk4 for our drone control computer.
Finally, we selected ZED stereo camera for our high-resolution camera.
\textbf{Experiments of Face-Anonymizing Networks with Nvidia Xavier}:
NVIDIA Jetson Xavier is an machine-learning computer for autonomous machines with the high-computing power.
Especially, this companion computer has a 512-core GPU, and thus it is indeed suitable for large-scale matrix operations which are needed for efficient neural networks computation.
A 8800mah 4S1P LiPo battery was installed for power supply, which provides a voltage of 14.8V.
\subsection{Our Face-Anonymizing System Result on Drone}
Fig.~\ref{Fig:Evaluation_on_drone} shows our face-anonymizing system results on a video recorded in a drone.
The drone is hovering in our laboratory, and a person is walking in front of the drone.
From this result, we can confirm that our system can well anonymize a face with varying its size.
Fig.~\ref{Fig:Evaluation_SLAM_on_drone} presents the comparison of feature extraction results on an original video and an anonymized video.
Both videos confirm that the feature points are well extracted.
In our anonymization scheme, the extracted feature points are almost the same as that of the original video.
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Test_FaceForensic_1.eps}
\caption{Additional Test Results on FaceForensic Dataset via Algorithm~\ref{alg:face_anonymizing_algorithm}}
\label{Fig:Result_FaceForensic_1}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Test_FaceForensic_2.eps}
\caption{Additional Test Results on FaceForensic Dataset via Algorithm~\ref{alg:face_anonymizing_algorithm}}
\label{Fig:Result_FaceForensic_2}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Test_CelebA.eps}
\caption{Additional Test Results on CelebA-HQ Dataset}
\label{Fig:Result_CelebA-HQ}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Test_Helen.eps}
\caption{Additional Test Results on Helen Dataset}
\label{Fig:Result_Helen}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\paperwidth]{Test_FaceScrub.eps}
\caption{Additional Test Results on FaceScrub Dataset}
\label{Fig:Result_FaceScrub}
\end{figure*}
| 2c7bca17f73b7f091b7ab9b5470e90d920197aa1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Deformation quantization, also referred to as phase space quantum mechanics by many authors, consists in a general quantization procedure based on the idea that a quantum system is obtained by deforming both the geometrical and the algebraic structures of classical phase space \cite{Bayen}, \cite{Flato}. One of the main features of this approach lies in the primordial role acquired by the algebra of quantum observables which, from this perspective, is not given by a family of self-adjoint operators acting on a Hilbert space as in ordinary quantum mechanics, but instead these observables correspond to real or complex valued functions defined on the classical phase space, where the usual commutative point-wise product is replaced by an associative, non-commutative product, denoted as the star-product. Consequently, this star-product induces a deformation of the Poisson bracket in such a way that all the information included on the commutators between self-adjoint operators is contained in the deformed algebraic classical structure. Considering that one may circumvent the use of operators as quantum observables, within the deformation quantization formalism an essential ingredient resides on the definition of the Wigner distribution \cite{Wigner}. This distribution corresponds to a phase space representation of the density matrix, and thus it contains all the information of the auto-correlation properties and the transition amplitudes associated to a given quantum mechanical system \cite{Zachos}. A prominent property of the Wigner distribution for a quantum system lies on the possibility to acquire negative values on particular regions of phase space and, therefore, it can not be merely interpreted as a probability density in the standard sense and, in consequence, it is usually referred to as a quasi-probability distribution in the literature. However, this seemingly odd feature impress the Wigner distribution with a relevant property since it allows to visualize the fact that quantum trajectories in phase space that correspond to regions with negative probability values may be useful to determine joint-correlation functions and entanglement properties within the quantum system \cite{Weinbub}.
The formalism of deformation quantization not only has provided important contributions in pure mathematics \cite{Kontsevich}, \cite{Reichert}, \cite{Bahns}, \cite{Waldmann}, but it also has proved to be a reliable tool in the understanding of many physical quantum systems \cite{Zachos}, \cite{Fredenhagen}, \cite{Compean}, \cite{Compean2}, recently including the treatment of constrained systems \cite{DQconstraints}, and certain aspects on loop quantum cosmology and quantum gravity \cite{DQpoly}, \cite{PolyWigner}.
In this paper we propose to analyze the formalism of quantum field theory by means of coherent states as seen from the deformation quantization point of view. Motivated by the representation of quantum mechanics in phase space in terms of quasiprobability distributions, widely used in quantum optics and quantum information theory \cite{Cahill}, \cite{Manko}, \cite{Scully}, we address the construction of the quantum field analogues of the Weyl-Stratonovich quantizer, the Wigner functional, the star-product and the $s$-ordered symbols associated to quantum field operators by adopting the holomorphic representation, as first introduced by Cahill and Glauber in the context of standard quantum mechanics \cite{Cahill}. In particular, via the quantum field analogue of the Husimi distribution we obtain the Moyal and the normal star-products for a scalar field in terms of holomorphic variables \cite{Dito}. Then, following some ideas formulated in \cite{Wunsche}, \cite{Moya}, we provide a series representation of the Wigner functional with respect to the Fock basis and coherent states.
This paper is organized as follows, in section 2, we briefly introduce the basic ideas of the deformation quantization scheme applied to fields and we obtain its holomorphic representation. In section 3, the coherent representation, the star products and the quantum field counterparts of the Wigner and Husimi distributions are obtained. Then, by employing the coherent states, a series representation of the Wigner functional is reviewed. Finally, we introduce some concluding remarks in section 4.
\section{Deformation quantization of fields }
\label{sec:DQF}
In this section we briefly review the Wigner-Weyl quantization scheme for fields. We will closely follow the description of the formalism as described in \cite{Compean} and \cite{PolyWigner}. For simplicity, we will focus on the real scalar field, nevertheless a generalization to other fields follows straightforwardly. In order to get a full perspective on the deformation quantization program we refer the reader to the more detailed reviews \cite{Bordemann}, \cite{Blaszak}, \cite{Gutt}.
\subsection{The Wigner-Weyl-Stratonovich quantization}
Consider a real scalar field $\varphi$ defined on a four dimensional background Minkowski spacetime $\mathcal{M}$. Let us perform a $3+1$ decomposition of the spacetime $\mathcal{M}$ in the form $\mathcal{M}=\Sigma\times\mathbb{R}$, for any Cauchy surface $\Sigma$, which at the present case of study may be thought of as topologically equivalent to $\mathbb{R}^{3}$. The spacetime manifold $\mathcal{M}$ is endowed with local coordinates $(x,t)\in\mathbb{R}^{3}\times\mathbb{R}$ and a metric $\eta=diag(+1,+1,+1,-1)$. In what follows, we deal with fields evaluated at the instant $t=0$, and we write $\varphi(x,0):=\varphi(x)$, and $\varpi(x,0):=\varpi(x)$, where $\varpi(x)$ stands for the canonical conjugate momentum associated to the field $\varphi(x)$. Consequently, the phase space of the theory is locally written through coordinates $\Gamma=(\varphi,\varpi)$, which in turn can be related to initial data on a Cauchy surface $\Sigma$.
By analogy to quantum mechanics, in order to define the Weyl quantization rule \cite{Weyl}, i.e., a one to one mapping from the set of classical observables to the set of quantum observables, usually given by self-adjoint operators defined on a Hilbert space, we need to provide a quantization map such that it takes the Poisson brackets
\begin{eqnarray}\label{Poisson}
\left\lbrace \varphi(x),\varpi(y)\right\rbrace
& = & \delta(x-y) \,,\nonumber\\
\left\lbrace \varphi(x),\varphi(y)\right\rbrace
& = &
\left\lbrace \varpi(x),\varpi(y)\right\rbrace = 0 \,,
\end{eqnarray}
to the commutator of operators (in natural units where $\hbar=1$)
\begin{eqnarray}
\left[\hat{\varphi}(x),\hat{\varpi}(y) \right]\Psi
& = &
i\delta(x-y)\Psi \,, \nonumber\\
\left[\hat{\varphi}(x),\hat{\varphi}(y) \right]\Psi
& = & \left[\hat{\varpi}(x),\hat{\varpi}(y) \right]\Psi=0 \,,
\end{eqnarray}
where the vector state $\Psi[\varphi]$ is given by a functional of the field $\varphi\in\mathcal{S}'(\mathbb{R}^{3})$, where $\mathcal{S}'(\mathbb{R}^{3})$ stands for the Schwartz space of tempered distributions, that is, the space of continuous linear functionals on the space of rapidly decreasing smooth test functions $\mathcal{S}(\mathbb{R}^{3})$. To be more precise, the state $\Psi$ belongs to the Hilbert space $\mathcal{H}=L^{2}\left( \mathcal{S}'(\mathbb{R}^{3}),d\mu\right) $. Here d$\mu$ is given by the formal measure $\mathcal{D}\varphi:=\prod_{x\in\mathbb{R}^{3}}d\varphi(x)$, which denotes a uniform Lebesgue measure on the configuration space \cite{Glimm}. Although, it is known that in an infinite dimensional vector space a translational invariant measure cannot be properly defined, other measures such as Gaussian measures of mean zero, can be used in order to construct different representations of the Hilbert space $\mathcal{H}$ by employing proper probability measures, as shown in \cite{PolyWigner}, \cite{Corichi}, \cite{Velinho},.
The Weyl-Stratonovich quantization, $W:L^{2}(\Gamma)\to \mathcal{L}(\mathcal{H})$, is defined as the linear map from the space of functionals on the phase space $\Gamma$, to the linear operators acting on the Hilbert space $\mathcal{H}$ \cite{Compean}, \cite{Stratonovich}, by
\begin{equation}\label{Weylmap}
\hat{F}\Psi:=W(F[\varphi,\varpi])\Psi=\int\mathcal{D}\varphi\mathcal{D}\left(\frac{\varpi}{2\pi}\right)F[\varphi,\varpi]\hat{\Omega}[\varphi,\varpi]\Psi \,,
\end{equation}
where the operator $\hat{\Omega}$ is given by
\begin{equation}\label{Stratonovich}
\hat{\Omega}[\varphi,\varpi]=\int\mathcal{D}\left(\frac{\lambda}{2\pi} \right)\mathcal{D}\mu\;e^{-i\varphi(\lambda)+i\hat{\varphi}(\lambda)-i\varpi(\mu)+i\hat{\varpi}(\mu)} \,,
\end{equation}
and we have introduced the notation
\begin{equation}
\varphi(\lambda):=\int_{\mathbb{R}^{3}}dx\;\varphi(x)\lambda(x)\,,\;\;\;\mathrm{and}\;\;\;\;\hat{\varphi}(\lambda):=\int_{\mathbb{R}^{3}}dx\;\hat{\varphi}(x)\lambda(x) \,,
\end{equation}
for all $\lambda, \mu\in \mathcal{S}(\mathbb{R}^{3})$, and we have used similar expressions to denote the other terms appearing in the definition of $\hat{\Omega}[\varphi,\varpi]$. The operator $\hat{\Omega}$ represented in (\ref{Stratonovich}) corresponds to the quantum field analogue of the Weyl-Stratonovich quantizer, and as we shall see, its inverse allows us to define an associative and noncommutative product, the so called star-product, which proves to be one of the fundamental algebraic structures within the deformation quantization scheme. Employing the relations $\int\mathcal{D}\varphi\ket{\varphi}\bra{\varphi}=\hat{1}$ and $\int\mathcal{D}\left( \frac{\varpi}{2\pi}\right)\ket{\varpi}\bra{\varpi}=\hat{1}$, where $\ket{\varphi}$ and $\ket{\varpi}$ stand for vector eigenstates of the operators $\hat{\varphi}$ and $\hat{\varpi}$, respectively, the Weyl-Stratonovich operator satisfy the following properties
\begin{eqnarray}
\hat{\Omega}^{\dagger}[\varphi,\varpi]
& = &
\hat{\Omega}[\varphi,\varpi] \,,\\
\mathrm{tr}{\left\lbrace\hat{\Omega}^{\dagger}[\varphi,\varpi]\right\rbrace}
& = & 1 \,, \\
\mathrm{tr}{\left\lbrace \hat{\Omega}[\varphi,\varpi]\hat{\Omega}[\varphi',\varpi']\right\rbrace}
& = &
\delta(\varphi-\varphi')\delta\left( \frac{\varpi-\varpi'}{2\pi}\right) \label{trace} \,.
\end{eqnarray}
By multiplying the expression (\ref{Weylmap}) by the operator $\hat{\Omega}$ and taking the trace on both sides, using property (\ref{trace}), one obtains that the phase space function associated to the operator $\hat{F}$ reads
\begin{equation}
F[\varphi,\varpi]:=W^{-1}(\hat{F})=\mathrm{tr}\left\lbrace \hat{\Omega}[\varphi,\varpi]\hat{F}\right\rbrace.
\end{equation}
This map, also known as the Wigner map, corresponds to the inverse relation of the Weyl-Stratonovich quantizer and it associates quantum operators to real functions or symbols, following the standard terminology in harmonic analysis \cite{Reed}. The symbols determine an associative algebra endowed with a noncommutative product denominated as the Moyal star-product in field theory, given by
\begin{equation}\label{star}
(F_{1}\star F_{2})[\varphi,\varpi]:=W^{-1}(\hat{F}_{1}\hat{F}_{2})=\mathrm{tr}\left\lbrace \hat{\Omega}[\varphi,\varpi]\hat{F}_{1}\hat{F}_{2}\right\rbrace \,.
\end{equation}
By substituting the expression (\ref{Weylmap}) into (\ref{star}) and using the expansion of the functionals $F_{1}$ and $F_{2}$ in Taylor series we obtain (we encourage the reader to see \cite{Compean}, \cite{Hirshfeld} for detailed calculations)
\begin{eqnarray}
\label{Moyal}
\hspace{10ex}\mkern-95mu(F_{1}\star F_{2})[\varphi,\varpi]
= & &
\nonumber\\
F_{1}[\varphi,\varpi]\exp\left\lbrace\frac{i}{2}\int_{\mathbb{R}^{3}} dx \left(\frac{\overleftarrow{\delta}}{\delta\varphi(x)}\frac{\overrightarrow{\delta}}{\delta\varpi(x)}-\frac{\overleftarrow{\delta}}{\delta\varpi(x)}\frac{\overrightarrow{\delta}}{\delta\varphi(x)} \right) \right\rbrace F_{2}[\varphi,\varpi] & . &
\end{eqnarray}
Finally, let $\hat{\rho}=\ket{\Psi}\bra{\Psi}$ be the density operator of a quantum state $\Psi\in\mathcal{H}$, the symbol $\rho[\varphi,\varpi]$ corresponding to the operator $\hat{\rho}$ reads
\begin{equation}\label{Wignerfunctional}
\rho[\varphi,\varpi]=\mathrm{tr}\left\lbrace \hat{\Omega}[\varphi,\varpi]\hat{\rho}\right\rbrace=\int\mathcal{D}\left(\frac{\mu}{2\pi}\right)e^{-i\varpi(\mu)}\Psi^{*}\left[\varphi-\frac{\mu}{2}\right]\Psi\left[\varphi+\frac{\mu}{2} \right] \,.
\end{equation}
This functional is the quantum field analogue of the Wigner distribution associated to the Hilbert space $\mathcal{H}$ in quantum mechanics.
As mentioned before, a distinguished property for the Wigner distribution lies on the possibility to acquire negative values on some regions of phase space, therefore, it can not be interpreted merely as a probability distribution, and thus it is usually referred to as a quasi-probability distribution. This particularity endows the Wigner distribution with the quality of being an excellent candidate to
characterize the quantum properties of a system, not only as it contains the information of the density matrix, but also as negative contributions to the Wigner function may be interpreted as non classical interferences, despite the fact that not all quantum states introduce negative contributions to the Wigner distribution \cite{Zachos}.
\subsection{The Bargamnn-Fock representation}
In order to obtain the Bargmann-Fock representation (also known as the holomorphic representation), let us consider the expansion of the fields $\varphi$ and $\varpi$ in terms of an infinite set of harmonic oscillators
\begin{eqnarray}
\varphi(x,t)=\frac{1}{(2\pi)^{3/2}}\int_{\mathbb{R}^{3}} dk\left(\frac{1}{2\omega(k)}\right)^{1/2}\left(a(k,t)e^{ikx}+a^{*}(k,t)e^{-ikx} \right) \,,\label{varphi} \\
\varpi(x,t)=\frac{1}{(2\pi)^{3/2}}\int_{\mathbb{R}^{3}} dk\left(\frac{\omega(k)}{2}\right)^{1/2}i\left(a^{*}(k,t)e^{-ikx}-a(k,t)e^{ikx} \right)\label{varpi} \,,
\end{eqnarray}
where $\omega(k)=\sqrt{k^{2}+m^{2}}$, $a(k,t)=a(k)e^{-i\omega(k)t}$ and
$kx:=\sum_{i=1}^3 k_{i}x_{i}$, with $i=1,2,3$. From expressions (\ref{varphi}) and (\ref{varpi}) we may obtain
\begin{equation}\label{a}
a(k,t)=\frac{1}{(2\pi)^{3/2}(2\omega(k))^{1/2}}\int dx\, e^{-ikx}\left(\omega(k)\varphi(x,t)+i\varpi(x,t) \right) \,,
\end{equation}
and its complex conjugate, $a^{*}(k,t)$. By using the Poisson bracket relations (\ref{Poisson}), we can observe that $a(k,t)$ and $a^{*}(k,t)$ satisfy the Poisson algebra
\begin{eqnarray}\label{Poissona}
\left\lbrace a(k,t),a^{*}(k',t)\right\rbrace
& = &
-i\delta(k-k') \,,\\
\left\lbrace a(k,t),a(k',t)\right\rbrace
& = &
\left\lbrace a^{*}(k,t),a^{*}(k',t)\right\rbrace=0\label{Poissonas} \,.
\end{eqnarray}
In order to compute the Weyl-Stratonovich quantizer in terms of the holomorphic variables $a(k,t)$ and $a^{*}(k,t)$, let us introduce the following canonical variables \cite{Lifshitz}
\begin{eqnarray}
Q(k,t)=\left( \frac{1}{2\omega(k)}\right)^{1/2}\left(a(k,t)+a^{*}(k,t) \right)\,,\label{Q}\\
P(k,t)=i\left(\frac{\omega(k)}{2} \right)^{1/2}\left(a^{*}(k,t)-a(k,t) \right)\,. \label{P}
\end{eqnarray}
From the Poisson algebra given in (\ref{Poissona}) and (\ref{Poissonas}), and the definition of the variables (\ref{Q}) and (\ref{P}), one can derive the Poisson relations
\begin{eqnarray}
\left\lbrace Q(k,t),P(k',t)\right\rbrace
& = & \delta(k-k') \,,\label{PoissonQP}\\
\left\lbrace Q(k,t),Q(k',t)\right\rbrace
& = &
\left\lbrace P(k,t),P(k',t)\right\rbrace=0\label{PoissonQQ} \,.
\end{eqnarray}
Further, whenever we insert the holomorphic variables (\ref{a}) into equations (\ref{Q}) and (\ref{P}) we get
\begin{eqnarray}
\nonumber
Q(k,t)
& = &
\frac{1}{(2\pi)^{3/2}\omega(k)}\int_{\mathbb{R}^{3}} dx\left(\varpi(x,t)\sin(kx)+\omega(k)\varphi(x,t)\cos(kx) \right) \,,\\
P(k,t)
& = & \frac{1}{(2\pi)^{3/2}}\int_{\mathbb{R}^{3}} dx\left( \varpi(x,t)\cos(kx)-\omega(k)\varphi(x,t)\sin(kx)\right)\,.
\label{Q&P}
\end{eqnarray}
Then, it is evident from the algebraic relations (\ref{PoissonQP}), (\ref{PoissonQQ}) and (\ref{Q&P}) that $Q(k,t)$ and $P(k,t)$ defines a linear canonical transformation. This means, that instead of using the field variables $\varphi(x)$ and $\varpi(x)$, it should be possible to write the Weyl-Stratonovich quantizer in terms of $Q(k,t)$ and $P(k,t)$.
With the computation of the canonical variables $Q(k,t)$ and $P(k,t)$ in hand, we can rewrite the Weyl-Stratonovich map in the following form (see \cite{Compean} for further details)
\begin{equation}\label{WSQP}
\hat{\Omega}[Q,P]=\hat{\Omega}[\varphi[Q,P],\varpi[Q,P]]=\int\mathcal{D}\left(\frac{\lambda}{2\pi}\right)\mathcal{D}\mu\, e^{-iQ(\lambda)-iP(\mu)+i\hat{Q}(\lambda)+\hat{P}(\mu)} \,,
\end{equation}
where we have used the notation
\begin{equation}
Q(\lambda)=\int_{\mathbb{R}^{3}}dk\;Q(k)\lambda(k)\,, \;\;\; \mathrm{and} \;\;\;\; \hat{Q}(\lambda)=\int_{\mathbb{R}^{3}}dk\;\hat{Q}(k)\lambda(k)\,,
\end{equation}
for all $\lambda, \mu\in \mathcal{S}(\mathbb{R}^{3})$, and similar expressions to denote the other terms appearing in the definition of $\hat{\Omega}[Q,P]$. The field operators $\hat{Q}$ and $\hat{P}$ satisfy
the relations $\hat{Q}(k)\ket{Q}=Q(k)\ket{Q}$, $\hat{P}(k)\ket{P}=P(k)\ket{P}$, where $\ket{Q}$ and $\ket{P}$ stand for their corresponding eigenvectors fulfilling the properties $\int DQ \ket{Q}\bra{Q}=\hat{1}$ and
$\int \mathcal{D}\left( \frac{P}{2\pi}\right)\ket{P}\bra{P}=\hat{1}$, respectively.
Consequently, it can be shown that the Moyal star-product can be written as (see \cite{Compean} for more details)
\begin{eqnarray}
\mkern-110mu(F_{1}\star F_{2})[Q,P]=\mathrm{tr}\left\lbrace \hat{\Omega}[Q,P]\hat{F}_{1}\hat{F}_{2} \right\rbrace = \nonumber\\
F_{1}[Q,P]\exp\left\lbrace\frac{i}{2}\int_{\mathbb{R}^{3}} dk \left(\frac{\overleftarrow{\delta}}{\delta Q}\frac{\overrightarrow{\delta}}{\delta P(k)}-\frac{\overleftarrow{\delta}}{\delta P(k)}\frac{\overrightarrow{\delta}}{\delta Q(k)} \right) \right\rbrace F_{2}[Q,P] \,.
\end{eqnarray}
Finally, it is possible, using the formalism in terms of the $Q(k)$ and $P(k)$ variables to express the symbol of the density operator $\hat{\rho}=\ket{\Psi}\bra{\Psi}$ as
\begin{equation}
\rho[Q,P]=\mathrm{tr}\left\lbrace \hat{\Omega}[Q,P]\hat{\rho}\right\rbrace=\int\mathcal{D}\left(\frac{\mu}{2\pi}\right)e^{-i P(\mu)}\Psi^{*}\left[Q-\frac{\mu}{2}\right]\Psi\left[Q+\frac{\mu}{2} \right] \,.
\end{equation}
In the next section we will find that, by making use of coherent states, the Weyl-Stratonovich quantizer and the Wigner functional written in terms of holomorphic variables will allow us to compute the quantum field analogue of the $s$-ordered symbols introduced within the context of quantum mechanics by Cahill and Glauber in~\cite{Cahill} with the aim to study different orderings in comparison to the Weyl-Wigner symmetric ordering of operators and their corresponding star-products.
\section{Coherent representation and deformation quantization}
\subsection{Coherent representation and star products}
In order to construct coherent states in quantum field theory, i.e., the analogues of well-localized states of the harmonic oscillator with fluctuations similar as those for the vacuum state \cite{Itzykson}, let us assume a normalizable function $\alpha(k)\in\mathcal{S}(\mathbb{R}^{3})$, and consider the state given by
\begin{equation}\label{displacement}
\ket{\alpha}=e^{\hat{a}^{\dagger}(\alpha)-\hat{a}(\alpha^{*})}\ket{0}=:\hat{D}(\alpha)\ket{0} \,,
\end{equation}
where, as before, we have used the notation
\begin{equation}\label{coherent}
\hat{a}^{\dagger}(\alpha)=\int_{\mathbb{R}^{3}}dk\,\alpha(k)\hat{a}^{\dagger}(k)
\,, \;\;\; \mathrm{and}\;\;\;\; \hat{a}(\alpha^{*})=\int_{\mathbb{R}^{3}}dk\;\alpha^{*}(k)\hat{a}(k) \,,
\end{equation}
and $\ket{0}$ stands for the vacuum ground state defined through
\begin{eqnarray}\label{vacuum}
\hat{a}(k)\ket{0} = 0 \,, \;\; \mathrm{and} \;\;\;\; \braket{0|0}=1 \,,
\end{eqnarray}
for all $k\in\mathbb{R}^{3}$, and $\hat{a}^{\dagger}(k)$, $\hat{a}(k)$ denote the creation and annihilation operators for a given mode $k$, respectively. In the definition (\ref{displacement}), the operator $\hat{D}(\alpha)$ corresponds to the field counterpart of the Glauber displacement operator \cite{Cahill}, and it satisfies
the relation\footnote{Throughout the paper, given any complex number $a$, we denote by $\Re (a)$ and $\Im (a)$ its real and imaginary parts, respectively.}
\begin{equation}
\label{DisplProp}
\hat{D}(\alpha)\hat{D}(\beta)\ket{0}=\hat{D}(\alpha)\ket{\beta}=e^{i\Im(\braket{\beta,\alpha})}\ket{\alpha+\beta} \,,
\end{equation}
where $\braket{\beta,\alpha}$ represents the inner product in $L^{2}(\mathbb{R}^{3})$
\begin{equation}
\braket{\beta,\alpha}=\int_{\mathbb{R}^{3}}dk\,\beta^{*}(k)\alpha(k) \,.
\end{equation}
Moreover, the coherent states satisfy the non-orthogonality relations
\begin{equation}
\braket{\beta|\alpha}=e^{\frac{1}{2}(\braket{\beta,\alpha}-\braket{\alpha,\beta})}e^{-\frac{1}{2}||\beta-\alpha||^{2}} \,,
\end{equation}
where $||\beta-\alpha||^{2}$ denotes the norm induced by the inner product $\braket{\beta-\alpha,\beta-\alpha}$. Even though the coherent states are
non-orthogonal, one may show that they fulfill the completeness relation \cite{Glimm}, \cite{Itzykson}
\begin{equation}\label{completeness}
\int\mathcal{D}^{2}\left(\frac{\alpha}{\pi}\right)\ket{\alpha}\bra{\alpha}=\hat{1} \,,
\end{equation}
where the measure is explicitly written as $\mathcal{D}^{2}\left(\frac{\alpha}{\pi}\right)=\mathcal{D}\left(\frac{\Re({\alpha})}{\pi}\right)\mathcal{D}\left(\Im({\alpha})\right)$, that is, it is a formal integral over the complex $\alpha(k)$-plane, for all $k\in\mathbb{R}^{3}$.
In order to write the Weyl-Stratonovich quantizer in terms of the coherent state formulation of fields, let us substitute equations (\ref{Q}) and (\ref{P}) into the expression (\ref{WSQP}), and by making a change of variables we find
\begin{eqnarray}
\mkern-50mu \hat{\Omega}[a,a^{*}]
& =& \int \mathcal{D}^{2}\left(\frac{\xi}{\pi^{2}}\right)e^{a(\xi^{*})-a^{*}(\xi)+\hat{a}^{\dagger}(\xi)-\hat{a}(\xi^{*})} \nonumber\\
& = &
\int \mathcal{D}^{2}\left(\frac{\xi}{\pi^{2}}\right)e^{a(\xi^{*})-a^{*}(\xi)}\hat{D}(\xi) \,,
\label{WSa}
\end{eqnarray}
where the complex function $\xi(k)$ is such that $\Re(\xi),\Im(\xi)\in\mathcal{S}(\mathbb{R}^{3})$, and the formal measure is explicitly given by $\mathcal{D}^{2}\left(\frac{\xi}{\pi^{2}}\right)=\mathcal{D}\left(\frac{\Re(\xi)}{\pi}\right)\mathcal{D}\left(\frac{\Im(\xi)}{\pi}\right)$. It is evident now, that the operator $\hat{\Omega}[a,a^{*}]$ defined in (\ref{WSa}) corresponds to the normalized quantum field analogue of the symmetric ordered Weyl-Stratonovich operator employed in quantum optics and quantum information \cite{Manko}, \cite{Scully},
\cite{Marmo}. By using some standard techniques on Gaussian functional integration \cite{Glimm} and the completeness relation (\ref{completeness}), the operator $\hat{\Omega}[a,a^{*}]$ satisfies $\mathrm{tr}\left\lbrace\hat{\Omega}[a,a^{*}]\right\rbrace=1$.
Now, with the Weyl-Stratonovich operator at hand, the symbol associated to an operator $\hat{F}$ in terms of holomorphic variables is given by
\begin{equation}
F[a,a^{*}]:=W^{-1}[\hat{F}]=\mathrm{tr}\left\lbrace \hat{\Omega}[a,a^{*}]\hat{F} \right\rbrace \,.
\end{equation}
By using the Baker-Campbell-Hausdorff formula \cite{Hall} and the explicit expression for $\hat{\Omega}[a,a^*]$ in (\ref{WSa}), we can write the symbol of an operator as
\begin{equation}\label{FG}
F[a,a^{*}]=\int \mathcal{D}\left(\frac{\xi}{\pi^{2}}\right) e^{\braket{\xi,a}-\braket{a,\xi}+\frac{1}{2}||\xi||^{2}}G(\xi) \,,
\end{equation}
where
\begin{equation}
G(\xi):=\mathrm{tr}\left\lbrace \hat{D}(\xi)\hat{F}\right\rbrace e^{-\frac{1}{2}||\xi||^{2}} \,.
\end{equation}
From equation (\ref{FG}), we note that the symbol associated to the operator $\hat{F}$, can be written as
\begin{equation}
F[a,a^*]=\sum_{n=0}^{\infty}\frac{1}{2^{n}n!}\int \mathcal{D}\left(\frac{\xi}{\pi^{2}}\right)||\xi||^{2n} e^{\braket{\xi,a}-\braket{a,\xi}}G(\xi) \,.
\end{equation}
To simplify this expression, we note that
\begin{equation}
\int_{\mathbb{R}^{3}} dk \frac{\delta}{\delta a(k)}\frac{\delta}{\delta a^{*}(k)}e^{\braket{\xi,a}-\braket{a,\xi}}=-||\xi||^{2}e^{\braket{\xi,a}-\braket{a,\xi}} \,,
\end{equation}
then, the symbol $F[a,a^{*}]$ reads
\begin{equation}\label{WH}
F[a,a^{*}]=\exp\left\lbrace {-\frac{1}{2}\int_{\mathbb{R}^{3}}} dk \frac{\delta}{\delta a(k)}\frac{\delta}{\delta a^{*}(k)}\right\rbrace \int \mathcal{D}\left(\frac{\xi}{\pi^{2}}\right)e^{\braket{\xi,a}-\braket{a,\xi}}G(\xi) \,.
\end{equation}
From this last identity we realize, by comparing with the $s$-parametrized quasiprobability distributions in quantum mechanics \cite{Cahill}, \cite{Moya}, that $F[a,a^{*}]$ is related to the symbol associated to the quantum field analogue of the Husimi $Q$-representation of a normal ordered operator given by
\begin{equation}
F^{(N)}[a,a^{*}]:= Q^{-1}[\hat{F}]=\int \mathcal{D}\left(\frac{\xi}{\pi^{2}}\right)e^{\braket{\xi,a}-\braket{a,\xi}}G(\xi) \,.
\end{equation}
Further, by means of the completeness relation (\ref{completeness}) and the functional integration of Gaussians \cite{Glimm}, it is possible to express the symbol for $F^{(N)}[a,a^{*}]$ as
\begin{equation}\label{Husimi}
F^{(N)}[a,a^{*}]=\mathrm{tr}\left\lbrace\hat{F}\hat{\rho}\right\rbrace= \braket{a|\hat{F}|a} \,.
\end{equation}
Finally, we can observe that, within the context of quantum field theory, the symbol of an operator in the symmetric ordering stated by the Weyl-Stratonovich operator is related to the normal ordering given by the Husimi $Q$-representation, so that
\begin{equation}
F[a,a^{*}]=W^{-1}[\hat{F}]=\exp\left\lbrace {-\frac{1}{2}\int_{\mathbb{R}^{3}}} dk \frac{\delta}{\delta a(k)}\frac{\delta}{\delta a^{*}(k)}\right\rbrace F^{(N)}[a,a^{*}] \,.
\end{equation}
We now look for the composition rule for symbols, which determines the star product. By definition, the Moyal star-product is given as follows
\begin{equation}
(F_{1}\star F_{2})[a,a^{*}]=\mathrm{tr}\left\lbrace\hat{\Omega}[a,a^{*}]\hat{F}_{1}\hat{F}_{2} \right\rbrace \,,
\end{equation}
by applying formulas (\ref{WH}) and (\ref{Husimi}) we obtain
\begin{equation}\label{Moyala}
(F_{1}\star F_{2})[a,a^{*}]=\exp\left\lbrace {-\frac{1}{2}\int_{\mathbb{R}^{3}}} dk \frac{\delta}{\delta a(k)}\frac{\delta}{\delta a^{*}(k)}\right\rbrace \braket{a|\hat{F}_{1}\hat{F}_{2}|a} \,.
\end{equation}
Since the Husimi $Q$-representation deals with normal ordered operators, in such scheme we can write them as the normal ordered expansion
\begin{equation}
\hat{F}=\sum_{m,n=0}\int_{\mathbb{R}^{3}}dk\;c_{mn}(k)(\hat{a}^{\dagger}(k))^{m}(\hat{a}(k))^{n} \,,
\end{equation}
so that, in terms of the expectation value over coherent states and employing the completeness relation (\ref{completeness}), we observe that
\begin{eqnarray}
(F_{1}\star_{N}F_{2})[a,a^{*}]
:= \braket{a|\hat{F}_{1}\hat{F}_{2}|a} = \nonumber\\
\sum_{m,n,p,q}\int_{\mathbb{R}^3\times\mathbb{R}^{3}}dkdk'c_{mn}(k)b_{pq}(k')\braket{a|(\hat{a}^{\dagger}(k))^{m}(\hat{a}(k))^{n}(\hat{a}^{\dagger}(k'))^{p}(\hat{a}(k'))^{q}|a} =\nonumber\\
\int\mathcal{D}^{2}\left(\frac{\xi}{\pi}\right)F_{1}[\xi,a^{*}]F_{2}[a,\xi^{*}]e^{-||a-\xi||^{2}} \,.
\end{eqnarray}
As in the case of the Moyal star-product (\ref{Moyal}), we can perform an expansion of the functionals $F_{1}$ and $F_{2}$ in Taylor series \cite{Alexanian}, \cite{Lizzi}, and then by performing the functional Gaussian integral, yields
\begin{equation}
(F_{1}\star_{N}F_{2})[a,a^{*}]=F_{1}[a,a^{*}]\exp\left\lbrace\int_{\mathbb{R}^{3}}dk\left(\frac{\overleftarrow{\delta}}{\delta a(k)}\frac{\overrightarrow{\delta}}{\delta a^{*}(k)} \right) \right\rbrace F_{2}[a,a^{*}] \,.
\end{equation}
This $\star_{N}$ product corresponds to the normal star-product obtained by means of Berezin calculus in \cite{Dito}. Then, it turns out that from the equation (\ref{Moyala}), the Moyal star-product is given by
\begin{equation}\label{cequivalence}
(F_{1}\star F_{2})[a,a^{*}]=\exp\left\lbrace {-\frac{1}{2}\int_{\mathbb{R}^{3}}} dk \frac{\delta}{\delta a(k)}\frac{\delta}{\delta a^{*}(k)}\right\rbrace (F_{1}\star_{N}F_{2})[a,a^{*}]\,.
\end{equation}
The relation between the two star-products as depticted in (\ref{cequivalence}) is known as a $c$-equivalence within the formalism of deformation quantization \cite{Dito}, \cite{HirshfeldP}, and by expanding the last exponential we obtain
\begin{eqnarray}
\mkern-90mu (F_{1}\star F_{2})[a,a^{*}]=
\nonumber\\
F_{1}[a,a^{*}]\exp\left\lbrace\frac{1}{2}\int_{\mathbb{R}^{3}}dk\left(\frac{\overleftarrow{\delta}}{\delta a(k)}\frac{\overrightarrow{\delta}}{\delta a^{*}(k)}-\frac{\overleftarrow{\delta}}{\delta a^{*}(k)}\frac{\overrightarrow{\delta}}{\delta a(k)} \right) \right\rbrace F_{2}[a,a^{*}] \,,
\end{eqnarray}
which corresponds to the star-product obtained in \cite{Dito}.
To conclude this section, and within the deformation quantization scheme for quantum field theories developed in this work, we note that it is possible to generalize the Weyl-Stratonovich quantization map for any $s$-ordered symbol by a straightforwardly comparison with the standard quantum mechanics counterpart, as
explored in~\cite{Cahill}. Indeed, by considering the s-ordered quantization operator
\begin{equation}
\hat{\Omega}_{s}[a,a^{*}]=\int \mathcal{D}^{2}\left(\frac{\xi}{\pi^{2}}\right)e^{a(\xi^{*})-a^{*}(\xi)}e^{\frac{1}{2}s||\xi||^{2}}\hat{D}(\xi) \,,
\end{equation}
where $s$ is a continuous parameter which defines the ordering of the creation and annihilation operators. The case $s=0$ corresponds to the standard Weyl-Wigner symmetric ordering, and for $s=-1$ we have the Husimi normal ordering, both cases analyzed above. An additional relevant value corresponds to $s=1$, which yields the Glauber-Sudarshan anti-normal ordering in quantization \cite{Cahill}, \cite{Scully}. The importance of the $s$-parametrization as a useful tool for understanding the convergence properties of the expansion of operator functions in terms of integral representations and its corresponding integral kernels has been characterized in \cite{Soloviev}.
\subsection{Series representation of the Wigner functional}
It is possible to obtain a series representation of the Wigner functional $\rho[\varphi,\varpi]$ in complete analogy to the Wigner function appearing in quantum mechanics \cite{Wunsche}, \cite{Moya}. By making the substitution of $\gamma=-\mu/2$ in the definition (\ref{Wignerfunctional}), we obtain
\begin{equation}
\rho[\varphi,\varpi]=\int\mathcal{D}\left( \frac{\gamma}{\pi}\right)e^{2i\varpi(\gamma)}\braket{\varphi-\gamma|\hat{\rho}|\varphi+\gamma} \,,
\end{equation}
where $\hat{\rho}=\ket{\Psi}\bra{\Psi}$ is the density operator associated to a quantum state $\Psi\in\mathcal{H}$. Then
\begin{eqnarray}
\rho[\varphi,\varpi]&=&\int\mathcal{D}\left( \frac{\gamma}{\pi}\right)e^{2i\varpi(\gamma)}\braket{-\gamma|e^{-i\hat{\varpi}(\varphi)}\hat{\rho}e^{i\hat{\varpi}(\varphi)}|\gamma} \nonumber \\
&=&\int\mathcal{D}\left( \frac{\gamma}{\pi}\right)\braket{-\gamma|e^{-i\hat{\varphi}(\varpi)}e^{-i\hat{\varpi}(\varphi)}\hat{\rho}e^{i\hat{\varpi}(\varphi)}e^{i\hat{\varphi}(\varpi)}|\gamma} \,,
\end{eqnarray}
by using the Baker-Campbell-Hausdorff formula and introducing the parity operator $\hat\Pi\ket{\gamma}=\ket{-\gamma}$, it follows that
\begin{eqnarray}
\rho[\varphi,\varpi]&=&\int\mathcal{D}\left( \frac{\gamma}{\pi}\right)\braket{\gamma|\hat{\Pi}e^{-i\hat{\varphi}(\varpi)}e^{-i\hat{\varpi}(\varphi)}\hat{\rho}e^{i\hat{\varpi}(\varphi)}e^{i\hat{\varphi}(\varpi)}|\gamma} \,\nonumber\\
&=&\mathrm{tr}{\left\lbrace \hat{\Pi}\hat{D}^\dagger(\xi)\hat{\rho}\hat{D}(\xi)\right\rbrace } \,,
\end{eqnarray}
where $\hat{D}(\xi)$ is the displacement operator defined in (\ref{displacement}) and $\xi\in\mathcal{S}(\mathbb{R}^{3})$ depends on the field variables $\varphi$ and $\varpi$. From now on, we will consider the Fock basis \cite{Zeidler} spanned
by vector states with occupation numbers $n_{s}$ for all modes $k$
\begin{equation}\label{Fock}
\ket{n_{1},n_{2},\ldots}=\left[ \prod_{s}\frac{\left( \hat{a}^{\dagger}(k_{s})\right)^{n_{s}}}{\sqrt{n_{s}!}}\right]\ket{0} \,,
\end{equation}
that is, the vector (\ref{Fock}) corresponds to the quantum state in which $n_{1}$ particles have momentum $k_{1}$, $n_{2}$ particles have momentum $k_{2}$, etc., and $\ket{0}$ denotes the vacuum ground state defined in (\ref{vacuum}) above.
The set of state vectors $\ket{n_{1},n_{2},\ldots}$, with all possible choices of $n_{s}$, form a complete orthonormal basis in the Hilbert space. Thus, we finally obtain the following representation of the Wigner functional
\begin{equation}\label{Wseries}
\rho[\varphi,\varpi]=\sum_{n_{1},n_{2},\ldots}\braket{n_{1},n_{2},\ldots|\hat{\Pi}\hat{D}^\dagger(\xi)\hat{\rho}\hat{D}(\xi)|n_{1},n_{2},\ldots} \,.
\end{equation}
By writing the density operator in terms of coherent states, $\hat{\rho}=\ket{\alpha}\bra{\alpha}$, and recalling the definition of the displacement operators (\ref{displacement}) and using the property~(\ref{DisplProp}), we can express the formula (\ref{Wseries}) as
\begin{equation}
\rho[\varphi,\varpi]=\sum_{n_{1},n_{2},\ldots}\braket{n_{1},n_{2},\ldots|\hat{\Pi}\hat{D}^\dagger(\xi)\hat{D}(\alpha)\ket{0}\bra{0}\hat{D}^{\dagger}(\alpha)\hat{D}(\xi)|n_{1},n_{2},\ldots} \,.
\end{equation}
Further, using the fact that \cite{Scully},
\begin{equation}
\hat{D}^{\dagger}(\xi)\hat{D}(\alpha)\ket{0}=e^{\frac{1}{2}(\braket{\xi,\alpha}-\braket{\alpha,\xi})}\ket{\alpha-\xi} \,,
\end{equation}
and expressing the parity operator as $\hat{\Pi}=(-1)^{\hat{N}}$, where $\hat{N}$ corresponds to the number operator in field theory \cite{Zeidler}, $\hat{N}=\frac{1}{(2\pi)^{3}}\int_{\mathbb{R}^{3}}dk\,\hat{a}^{\dagger}(k)\hat{a}(k)$,
the Wigner functional reads
\begin{equation}\label{Wseriescoherent}
\rho[\varphi,\varpi]=\sum_{n_{1},n_{2},\ldots}(-1)^N\braket{n_{1},n_{2},\ldots|\alpha-\xi}\braket{\alpha-\xi|n_{1},n_{2},\ldots} \,.
\end{equation}
The series representation of the Wigner functional contained in (\ref{Wseries}) and (\ref{Wseriescoherent}) allow us to calculate the probability distribution of quantum field observables by avoiding formal phase space integrals. This approach could be useful in circumstances where integration methods are very challenging as it may be the case, for instance, while studying either quantum effects in curved backgrounds or in quantum gravity \cite{Birrell}.
\section{Conclusions}
\label{sec:conclu}
In this paper we have analyzed the formalism of quantum field theory by means of coherent states appropriately introduced within the context of deformation quantization.
In particular, by selecting the holomorphic representation for a scalar field, we have explicitly encountered the quantum field analogues of the Weyl-Stratonovich quantizer, the Wigner distribution and the Moyal star-product by means of the Husimi distribution that corresponds to normal ordered operators. By construction, we also have set
a $c$-equivalence between the Moyal product and the normal star-product, $\star_{N}$,
which is commonly obtained by applying Berezin calculus.
Further, within our context, we have also noted that the correspondence between the symmetric and normal ordered symbols may be enlarged by considering the extension of the Weyl-Stratonovich quantization map for any $s$-ordered symbol in terms of the Glauber
displacement operator and a real parameter $s$. This may be a very relevant issue as the $s$-parametrization
has been proposed in the literature as a useful tool in order to analyze convergence of the expansion of operator functions in terms of integral representations.
Moreover, we have discussed a series representation of the Wigner functional which is obtained by considering the action of the Glauber displacement operators on the density operator written in terms of coherent states associated to quantum fields. This series representation may demonstrate relevant in order to determine the probability distribution of quantum field observables. We expect this will be the case in quantum field theory in a curved background where cumbersome phase space integrals are involved even in the simpler cases.
From a different perspective, it will also be relevant to implement the techniques described here within the context of either gauge field theories or the non-regular quantum representations that naturally emerge in Loop Quantum Gravity (LQG). Indeed, we expect that our results here will be significant, in addition to the developments in~\cite{DQconstraints}, \cite{DQpoly} and \cite{PolyWigner}, in order to explore the LQG coherent state formulation in the cosmological scenario. Work along these lines is in progress, and will be reported elsewhere.
\section*{Acknowledgments}
The authors would like to acknowledge financial support from CONACYT-Mexico
under project CB-2017-283838.
\bibliographystyle{unsrt}
| 0f8cb144c5611637e86ccdf929ed58aaf85ebdc2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Simultaneous machine translation is the task of generating partial translations before observing the entire source sentence. The task fits scenarios such as
live captioning and
speech-to-speech translation, where the user expects a translation before the speaker finishes the sentence. Simultaneous
MT has to balance between latency and translation quality. If more input is consumed before translation,
quality is likely to improve due to increased context, but latency also increases. On the other hand, consuming limited input decreases latency, but degrades quality.
There have been several approaches to solve simultaneous machine translation. In \citep{dalvi2018incremental,ma2019stacl}, a fixed policy is
introduced to delay translation by a fixed number of words.
Alternatively,
\citet{satija2016simultaneous,gu2017learning,alinejad2018prediction} use reinforcement learning to learn a dynamic policy to determine
whether to read or output words. \citet{cho2016sinmt} adapt the decoding algorithm without relying on additional components. However,
these
methods do not modify the training of the underlying NMT model.
Instead, it is trained on full sentences. \citet{arivazhagan2019monotonic} introduce a holistic framework that relaxes the hard notion
of read/write decisions at training time, allowing it to be trained jointly with the rest of the NMT model.
In this paper, we integrate a source chunk boundary detection component into a bidirectional recurrent NMT model. This component corresponds to segmentation or
read/write decisions in the literature. It is however trained jointly with the rest of the NMT model. We propose an algorithm to chunk the training data based on automatically learned word alignment.
The chunk boundaries are used as a training signal along with the parallel
corpus.
The main contributions of this work are as follows:
\begin{itemize}
\item We introduce a source chunk boundary detection component and train it jointly with the NMT model. Unlike in \citep{arivazhagan2019monotonic},
our component is trained using hard decisions, which is consistent with inference.
\item We propose a method based on word alignment to generate the source and target chunk boundaries, which are needed for training.
\item We study the use of bidirectional vs unidirectional encoder layers for simultaneous machine translation. Previous work focuses mostly on the use of unidirectional encoders.
\item We provide results
using text and speech input. This is in contrast to previous work that only
simulates simultaneous NMT on text input.
\end{itemize}
\section{Related Work}
\citet{oda2014optimizing} formulate segmentation as an optimization problem solved using dynamic programming to optimize translation quality.
The approach is applied to phrase-based machine translation. Our chunking approach is conceptually simpler, and we explore its use with neural machine translation.
\citet{cho2016sinmt} devise a greedy decoding algorithm for simultaneous neural machine translation. They use a model that is trained on full sentences. In
contrast, we train our models on chunked sentences to be consistent with the decoding condition.
\citet{satija2016simultaneous}, \citet{alinejad2018prediction}, and \citet{gu2017learning} follow a reinforcement learning approach to make decisions
as to when to read source words or to write target words.
\citet{Zheng2019SimplerAF} propose the simpler approach to use the position of the reference target word in the beam of an existing MT system to generate training examples of read/write decisions.
We extract such decisions from statistical word alignment instead.
In \citet{ma2019stacl,dalvi2018incremental}, a wait-$k$ policy is proposed to delay the first target word until $k$ source words are read. The model
alternates between generating $s$ target words and reading $s$ source words, until the source words are exhausted. Afterwards, the rest
of the target words are generated.
In addition, \citet{dalvi2018incremental} convert the training data into
chunks of predetermined fixed size.
In contrast, we train models that learn to produce dynamic context-dependent chunk lengths.
The idea of exploiting word alignments to decide for necessary translation context can be found in several recent papers. \citet{arthur2020learning} train an agent to imitate read/write decisions derived from word alignments. In our architecture such a separate agent model is replaced by a simple additional output of the encoder.
\citet{Xiong2019DuTongChuanCT} use word alignments to tune a pretrained language representation model to perform word sequence chunking. In contrast, our approach integrates alignment-based chunking into the translation model itself, avoiding the overhead of having a separate component and the need for a pretrained model. Moreover, in this work we
improve on pure alignment-based chunks using language models (Section \ref{subsec:improved_chunking})
to avoid leaving relevant future source words out of the chunk.
\citet{press2018noatt} insert $\epsilon$-tokens into the target using word alignments to develop an NMT model without an attention mechanism. Those tokens fulfill a similar purpose to \textit{wait} decisions in simultaneous MT policies.
\newcite{arivazhagan2019monotonic} propose an attention-based model that integrates an additional monotonic attention component. While the motivation is
to use hard attention to select the encoder state at the end of the source chunk, they avoid using discrete attention to keep the model differentiable, and use
soft probabilities instead. The hard mode is only used during decoding. We do not have to work around discrete decisions in this work, since the chunk
boundaries are computed offline before training, resulting in a simpler model architecture.
\section{Simultaneous Machine Translation}
The problem of offline machine translation is focused on finding the target sequence $e_1^I=e_1...e_I$ of length $I$ given the source sequence $f_1^J$ of
length $J$. In contrast, simultaneous MT does not necessarily require the full source input to generate the target output. In this work,
we formulate the problem by assuming latent monotonic chunking underlying the source and target sequences.
Formally, let $s_1^K=s_1...s_k...s_K$ denote
the chunking sequence of $K$ chunks, such that $s_k=(i_k,j_k)$, where $i_k$ denotes the target position of last target word in the $k$-th chunk, and $j_k$
denotes the source position of the last source word in the chunk. Since the source and target chunks are monotonic, the beginnings of the source and
target chunks do not have to be defined explicitly.
The chunk positions are subject to the following constraints:
\begin{align}
i_0 = j_0 = 0, & \ \ \ i_K = I, \ \ \ j_K = J, \nonumber \\
i_{k-1} < i_{k}, & \ \ \ j_{k-1} < j_{k}.
\label{eq:constraints}
\end{align}
We use $\tilde{e}_k=e_{i_{k-1}+1}...e_{i_k}$ to denote the $k$-th target chunk, and
$\tilde{f}_k=f_{j_{k-1}+1}...f_{j_k}$ to denote its corresponding source chunk. The target sequence $e_1^I$ can be rewritten as $\tilde{e}_1^K$, similarly,
the source sequence can be rewritten as $f_1^J=\tilde{f}_1^K$.
We introduce the chunk sequence $s_1^K$ as a latent variable as follows:
\begingroup
\allowdisplaybreaks
\begin{align}
&\;p(e_1^I|f_1^J) = \sum_{K,s_1^K} p(e_1^I,s_1^K|f_1^J) \label{eq:latent}\\
&= \sum_{K,s_1^K} p(\tilde{e}_1^K,s_1^K|\tilde{f}_1^K) \label{eq:rewrite}\\
&= \sum_{K,s_1^K} \prod_{k=1}^K p(\tilde{e}_k,s_k|\tilde{e}_1^{k-1}, s_1^{k-1}, \tilde{f}_1^K)
\label{eq:chain} \\
&= \sum_{K,s_1^K} \prod_{k=1}^K
p(i_k|\tilde{e}_1^{k}, s_1^{k-1}, j_k, \tilde{f}_1^K) \nonumber \\
& \hspace{16.5mm}\cdot p(\tilde{e}_k|\tilde{e}_1^{k-1}, s_1^{k-1}, j_k, \tilde{f}_1^K) \nonumber \\
& \hspace{16.5mm}\cdot p(j_k|\tilde{e}_1^{k-1}, s_1^{k-1}, \tilde{f}_1^K), \label{eq:word_chunk_decomp}
\end{align}
\endgroup
where Equation~\ref{eq:latent} introduces the latent sequence $s_1^K$ with a marginalization sum over all possible chunk sequences and all possible number of
chunks $K$. In Equation~\ref{eq:rewrite} we rewrite the source and target sequences using the chunk notation, and we apply the chain rule of probability
in Equation~\ref{eq:chain}. We use the chain rule again in Equation~\ref{eq:word_chunk_decomp} to decompose the probability further into a
\textit{target chunk boundary probability} $p(i_k|\tilde{e}_1^{k}, s_1^{k-1}, j_k, \tilde{f}_1^K)$, a \textit{target chunk translation probability}
$p(\tilde{e}_k|\tilde{e}_1^{k-1}, s_1^{k-1}, j_k, \tilde{f}_1^K)$,
and a \textit{source chunk boundary probability} $p(j_k|\tilde{e}_1^{k-1}, s_1^{k-1}, \tilde{f}_1^K)$. This creates a generative story, where
the source chunk boundary is determined first, followed by the translation of the chunk, and finally by the target chunk boundary.
The translation probability can be further decomposed to reach the word level:
\begin{align}
&p(\tilde{e}_k,|\tilde{e}_1^{k-1}, s_1^{k-1}, j_k, \tilde{f}_1^K) \nonumber \\
&= \prod_{i=i_{k-1}+1}^{i_k} p(e_i|\underbrace{e_{i_{k-1}+1}^{i-1}, \tilde{e}_1^{k-1}}_{=e_1^{i-1}}, s_1^{k-1}, j_k, \tilde{f}_1^K) \nonumber \\
&\approx \prod_{i=i_{k-1}+1}^{i_k} p(e_i|e_1^{i-1}, f_1^{j_k}, k).
\end{align}
In this work, we drop the marginalization sum over chunk sequences and use fixed chunks during training. The chunk sequences are generated as described in
Section \ref{sec:chunking}.
\section{Model}
\subsection{Source Chunk Boundary Detection}
\label{sec:boundary_detection}
We simplify the chunk boundary probability, dropping the dependence on the target sequence and previous target boundary decisions
\begin{align}
p(j_k|\tilde{e}_1^{k-1}, s_1^{k-1}, \tilde{f}_1^K) \approx p(j_k|f_1^{j_k}, j_1^{k-1}),
\end{align}
where the distribution is conditioned on the source sequence up to the last word of the $k$-th chunk. It is also conditioned on the previous source
boundary decisions $j_1...j_{k-1}$. Instead of computing a distribution over
the source positions, we introduce a binary random variable $b_j$ such that for each source position we estimate the probability of a chunk boundary:
\begin{equation}
b_{j,k} =
\begin{cases}
1 & \text{if\,} j \in \lbrace j_1, j_2 ... j_k \rbrace \\
0 & \mathrm{otherwise.}
\end{cases}
\end{equation}
For this, we use a forward stacked RNN encoder. The $l$-th forward encoder layer is given by
\begin{equation}
\hspace{-0.05cm}
\label{eq:forward}
\overrightarrow{h}_{j,k}^{(l)} =
\begin{cases}
\scalebox{.8}[1.0]{$LSTM$}\big( [\hat{f}_j; \hat{b}_{{j-1},k}], \overrightarrow{h}_{j-1,k}^{(l)}\big) & \hspace*{4mm} l = 1 \vspace{2mm} \\
\scalebox{.8}[1.0]{$LSTM$}\big( \overrightarrow{h}_{j,k}^{(l-1)}, \overrightarrow{h}_{j-1,k}^{(l)}\big) & \hspace*{-7mm} 1 < l < L_{enc},
\end{cases}
\end{equation}
where $\hat{f}_j$ is the word embedding of the word $f_j$, which is concatenated to the embedding of the boundary decision at the previous
source position $\hat{b}_{j-1,k}$. $L_{enc}$ is the number of encoder layers.
On top of the last layer a softmax estimates $p(b_{j,k})$:
\begin{equation}
\label{eq:source_chunk_detection}
\hspace{-0cm}p(b_{j,k}) = \text{softmax}\big(g([\overrightarrow{h}_{j,k}^{(L_{enc})}; \hat{f}_j; \hat{b}_{j-1,k}]) \big),\hspace{-0.2cm}
\end{equation}
where $g(\cdot)$ denotes a non-linear function.
\subsection{Translation Model}
We use an RNN attention model based on~\cite{Bahdanau14:softalign} for $p(e_i|e_1^{i-1}, f_1^{j_k})$. The model shares the forward encoder with the chunk boundary detection model. In addition, we extend the encoder with a stacked backward RNN encoder. The $l$-th backward layer is given by
\begin{equation}
\hspace{-0.07cm}
\label{eq:backward}
\overleftarrow{h}_{j,k}^{(l)} =
\begin{cases}
\textbf{0} & j>j_k, \forall l \vspace{1.5mm} \\
\scalebox{.8}[1.0]{$LSTM$}\big( [\hat{f}_j; b_{j,k}], \overleftarrow{h}_{j+1,k}^{(l)}\big) & l = 1 \vspace{2mm} \\
\scalebox{.8}[1.0]{$LSTM$}\big( \overleftarrow{h}_{j,k}^{(l-1)}, \overleftarrow{h}_{j+1,k}^{(l)}\big) & \hspace*{-3mm} 1 < l < L_{enc},
\end{cases}
\end{equation}
where the backward layer is computed within a chunk starting at the last position of the chunk $j=j_k$. $\textbf{0}$ indicates a vector of zeros for positions beyond the current chunk. The source representation is given by the concatenation of the last forward and backward layer
\begin{equation}
\label{eq:concat}
h_{j,k} = [\overrightarrow{h}_{j,k}^{(L_{enc})}; \overleftarrow{h}_{j,k}^{(L_{enc})}].
\end{equation}
We also stack $L_{dec}$ LSTM layers in the decoder
\begin{equation}
u_{i,k}^{(l)} =
\begin{cases}
\scalebox{.8}[1.0]{$LSTM$}\big( u_{i,k}^{(l-1)}, u_{i-1,\hat{k}}^{(l)}\big) & \hspace*{-2mm} 1 < l < L_{dec} \\
\scalebox{.8}[1.0]{$LSTM$}\big( [\hat{e}_{i}; d_{i,k}], u_{i-1,\hat{k}}^{(l)}\big) & \hspace*{5mm} l = 1,
\end{cases}
\end{equation}
where $\hat{e}_{i}$ is the target word embedding of the word $e_{i}$, $\hat{k}=k$ unless the previous decoder state belongs to the previous chunk, then
$\hat{k}=k-1$. The vector $d_{i,k}$ is the context vector computed over source positions
up to the last source position $j_k$ in the $k$-th chunk
\begin{align}
d_{i,k} =& \sum_{j=1}^{j_k} \alpha_{i,j,k} h_{j,k} \\
\alpha_{i,j,k} =& \text{softmax}(r_{i,1,k}...r_{i,j_k ,k})|_j \\
r_{i,j,k} =& f(h_{j,k}, u_{i-1,\hat{k}}^{(L_{dec})}),
\end{align}
where $\alpha_{i,j,k}$ is the attention weight normalized over the source positions $1\leq j \leq j_k$, and $r_{i,j,k}$ is the energy computed via the
function $f$ which uses $\tanh$ of the previous top-most decoder layer and the source representation at position $j$. Note the difference to the
attention component used in offline MT, where the attention weights are computed considering the complete source sentence $f_1^J$.
The output distribution is computed using a softmax function of energies from the top-most decoder layer $u_{i-1,k}^{(L_{dec})}$, the target embedding
of the previous word $\hat{e}_{i-1}$, and the context vector $d_{i-1,k}$
\begin{align}
p(e_i |& e_1^{i-1}, f_1^{j_k},k) = \nonumber \\
&\text{softmax}\big(g([u_{i-1,k}^{(L_{dec})}; \hat{e}_{i-1}; d_{i-1,k}])\big).
\end{align}
\subsection{Target Chunk Boundary Factor}
Traditionally, the translation model is trained to produce a sentence end token to know when to stop the decoding process. In our approach, this decision has to be made for each chunk (see next section). Hence, we have to train the model to predict the end positions of the chunks on the target side. For this, we use a target factor \citep{garcia2016factored, wilken2019novel}, i.e. a second output of the decoder in each step:
\begin{align}
\label{eq:target_chunk_detection}
p(b_i |& e_1^{i}, f_1^{j_k},k) =
\nonumber \\
&\text{softmax}(g( u_{i-1,k}^{(L_{dec})}, \hat{e}_{i}, \hat{e}_{i-1}, d_{i-1,k}))
\end{align}
where $b_i$ is a binary random variable representing target chunk boundaries analogous to $b_j$ on the source side. This probability corresponds to the first term in Equation \ref{eq:word_chunk_decomp}, making the same model assumptions as for the translation probability.
Note however, that we make the boundary decision dependent on the embedding $\hat{e}_{i}$ of the target word produced in the current decoding step.
\section{Search}
Decoding in simultaneous MT can be seen as an asynchronous process that takes a stream of source words as input and produces a stream of target words as output. In our approach, we segment the incoming source stream into chunks and output a translation for each chunk individually, however always keeping the full source and target context.
\begin{algorithm}
\DontPrintSemicolon
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\Input{ source word stream $f_1^J$}
\Output{ target word stream $e_1^I$}
\BlankLine
$\hat{\bm{f}_k}$ = [],
$\overrightarrow{\bm{h}}$ = [],
$\overleftarrow{\bm{h}}$ = [] \;
\For{$f_j$ \textbf{\upshape in} $f_1^J$}{
$\hat{f}_j$ = Embedding($f_j$) \;
$\overrightarrow{h}_j$, $p(b_j)$ = runForwardEncoder($\hat{f}_j$) \;
$\hat{\bm{f}_k}$ += $\hat{f_j}$ \;
$\overrightarrow{\bm{h}}$ += $\overrightarrow{h}_j$ \;
\If{$p(b_i) > t_b$ \textbf{\upshape or} j = J}{
$\overleftarrow{\bm{h}}_k$ = runBackwardEncoder($\hat{\bm{f}_k}$) \;
$\overleftarrow{\bm{h}}$ += $\overleftarrow{\bm{h}}_k$ \;
$\tilde{\bm{e}}_k$ = runDecoder($\overrightarrow{\bm{h}}$, $\overleftarrow{\bm{h}}$) \;
$e_1^I$ += $\tilde{\bm{e}}_k$ \;
$\hat{\bm{f}_k}$ = [] \;
}
}
\caption{Simultaneous Decoding \\ \footnotesize{lists in bold, [] is the empty list, += appends to a list}} \label{algo:decoding}
\end{algorithm}
Algorithm \ref{algo:decoding} explains
the simultaneous decoding process. One source word $f_j$ (i.e. its embedding $\hat{f}_j$) is read at a time.
We
calculate the next step of the shared forward encoder (Equation \ref{eq:forward}), including source boundary detection (Equation \ref{eq:source_chunk_detection}). If the boundary probability $p(b_j)$ is below a certain threshold $t_b$, we continue reading the next source word $f_{j+1}$. If, however, a chunk boundary is detected, we first feed all word embeddings $\hat{\bm{f}}_k$ of the current chunk into the backward encoder (Equation \ref{eq:backward}), resulting in representations $\overleftarrow{\bm{h}}_k$ for each of the words in the current chunk. After that, the decoder is run according to Equations \ref{eq:concat}--\ref{eq:target_chunk_detection}. Note, that it attends to representations $\overrightarrow{\bm{h}}$ and $\overleftarrow{\bm{h}}$ of all source words read so far, not only the current chunk. Here, we perform beam search such that in each decoding step those combinations of target words and target chunk boundary decisions are kept that have the highest joint probability. A hypothesis is considered final as soon as it reaches a position $i$ where a chunk boundary $b_i = 1$ is predicted. Note that
the length of a chunk translation is not restricted
and hypotheses of different lengths compete. When all hypotheses in the beam are final, the first-best hypothesis is declared as the translation $\tilde{e}_k$ of the current chunk and all its words are flushed into the output stream at once.
During search, the internal states of the forward encoder and the decoder are saved between consecutive different calls while the backward decoder is initialized with a zero state for each chunk.
\section{Alignment-Based Chunking}
\label{sec:chunking}
\subsection{Baseline Approach}
\label{subsec:baseline_chunking}
We aimed at a
meaningful segmentation of sentence pairs into bilingual chunks which could then be translated in monotonic sequence
and each chunk is -- in terms of aligned words -- translatable without consuming source words
from succeeding chunks.
We extract such a segmentation from unsupervised word alignments in source-to-target and target-to-source directions that we trained using the Eflomal toolkit~\citep{Ostling2016efmaral}
and combined using the \textit{grow-diag-final-and} heuristic~\citep{koehn03:spb}. Then, for each training sentence pair, we extract a sequence of ``minimal-length'' monotonic phrase pairs, i.e. a sequence of the smallest possible bilingual chunks which do not violate the alignment constraints\footnote{This means that all source words within a bilingual chunk are aligned only to the target words within this chunk and vice versa.} and at the same time are
conform with the segmentation constraints in Equation~\ref{eq:constraints}. By this we allow word reordering between the two languages to happen only within the chunk boundaries. The method roughly follows the approach of~\cite{marino05:tuples}, who extracted similar chunks as units for n-gram based statistical MT.
For fully monotonic word alignments, only chunks of length 1 either on the source or the target side are extracted (corresponding to 1-to-1, 1-to-M, M-to-1 alignments).
For non-monotonic alignments larger chunks are obtained, in the extreme case the whole sentence pair is one chunk.
Any unaligned source or target words are attached to the chunk directly preceding them,
also any non-aligned words that may start the source/target sentence are attached to the first chunk.
We perform the word alignment and chunk boundary extraction on the word level, and then convert words to subword units for the subsequent use in NMT.
\begin{figure*}[t]
\centering
\footnotesize{
\begin{verbatim}
EN: And | along came | a | brilliant | inventor, | a | scientist, |
who | came up with a partial cure for that disease
DE: Dann | kam | ein | brillanter | Erfinder des Wegs, | ein | Wissenschaftler, |
der | eine teilweise Heilung für diese Krankheit fand
EN: And | along came | a brilliant inventor, | a scientist, |
who | came up with a partial cure for that disease
DE: Dann | kam | ein brillanter Erfinder des Wegs, | ein Wissenschaftler, |
der | eine teilweise Heilung für diese Krankheit fand
\end{verbatim}
}
\caption{Examples of the baseline and the improved approach of extracting chunk boundaries. Note how in the improved approach noun phrases were merged into single bigger chunks. Also note the long last chunk that corresponds to the non-monotonic alignment of the English and German subordinate clause.}\label{fig:chunk-example}
\end{figure*}
\subsection{Delayed Source Chunk Boundaries}\label{subsec:delay}
We observed that the accuracy of source boundary detection can be improved significantly by including the words immediately following the source chunk boundary into the context. Take e.\,g. the source word sequence \texttt{I have seen it}. It
can be translated into German as soon as the word \texttt{it} was read: \texttt{Ich habe es gesehen}. Therefore the model is likely to predict a chunk boundary after \texttt{it}. However, if the next read source word is \textcolor{red}{\texttt{coming}}, it becomes clear that we should have waited
because the correct German translation is now \texttt{Ich habe es \textcolor{red}{kommen} gesehen}. There is a reordering which invalidates the previous partial translation.
To be able to resolve such cases, we shift the source chunks
by a constant delay $D$ such that
$j_1, ..., j_k, ..., j_K$ becomes $j_1 + D, ..., j_k + D, ..., j_K + D$.\footnote{If $j_K + D > J$, we shift the boundary to $J$, allowing empty source chunks at sentence end.} Note that the target chunks remain unchanged, thus the extra source words also provide an expanded context for translation. In preliminary experiments we saw large improvements in translation quality when using a delay of 2 or more words, therefore we use it in all further experiments.
\subsection{Improved Chunks for More Context}
\label{subsec:improved_chunking}
The baseline chunking method (Section \ref{subsec:baseline_chunking}) considers word reordering to determine necessary context for translation. However,
future context is often necessary for correct translation. Consider the translation \texttt{The beautiful woman} $\rightarrow$ \texttt{Die schöne Frau}. Here, despite of the monotonic alignment, we need the context of the third English word \texttt{woman} to translate the first two words as we have to decide on the gender and number of the German article \texttt{Die} and adjective \texttt{schöne}.
In part, this problem is already addressed by adding future source words into the context as described in Section~\ref{subsec:delay}.
However, this method causes a general increase in latency by $D$ source positions and yet covers only short-range dependencies. A better approach is to remove any chunk boundary for which the words following it are important for a correct translation of the words preceding it. To this end, we introduce a heuristic that uses two bigram target language models (LMs). The first language model yields the probability
$p(e_{i_k} | e_{i_k - 1})$ for the last word $e_{i_k}$ of chunk $s_k$, whereas the second one computes the probability $p(e_{i_k}|e_{i_k+1})$ for the last word in the chunk given the first word $e_{i_k+1}$ of the next chunk $s_{k+1}$ that follows the word $e_{i_k}$. The chunk boundary after $e_{i_k}$ is removed if the probability of the latter reverse bigram LM is higher than the probability of the first one by a factor $l = \sqrt{i_k - i_{k-1}}$, i.e. dependent on the length of the current chunk. The motivation for this factor is that shorter chunks should be merged with the context to the right more often than chunks which are already long, provided that the right context word has been frequently observed in training to follow the last word of such a chunk candidate. The two bigram LMs are estimated on the target side of the bilingual data, with the second one trained on sentences printed in reverse order.
Examples of the chunks extracted with the baseline and the improved approach for a given training sentence pair are shown in Figure~\ref{fig:chunk-example}.
\section{Streaming Speech Recognition}
\label{sec:asr}
To translate directly from speech signal, we use a cascaded approach. The proposed simultaneous NMT system consumes words from a streaming automatic speech recognition (ASR) system. This system is based on a hybrid LSTM/HMM acoustic model~\citep{bourlard89,Hochreiter:1997:LSTM}, trained on a total of approx.~2300 hours of transcribed English speech from the corpora allowed by IWSLT 2020 evaluation, including MUST-C, TED-LIUM, and LibriSpeech.
The acoustic model takes 80-dim.~MFCC features as input and estimates state posterior probabilities for 5000 tied triphone states. It consists of 4 bidirectional layers with 512 LSTM units for each direction. Frame-level alignment and state tying were
bootstrapped with a Gaussian mixtures acoustic model.
The LM of the streaming recognizer is a 4-gram count model trained with Kneser-Ney smoothing on English text data (approx.~2.8B running words) allowed by the IWSLT 2020 evaluation. The vocabulary consists of 152K words
and the out-of-vocabulary rate is below~1\%.
Acoustic training and the HMM decoding were performed with the RWTH ASR toolkit~\cite{wiesler2014:rasr}.
The streaming recognizer implements a version of chunked processing~\citep{chen2016training,zeyer16} which allows for the same BLSTM-based acoustic model to be used in both offline and online
applications.
By default, the recognizer updates the current first-best hypothesis by Viterbi decoding starting from the most recent frame and returns the resulting word sequence to the client. This makes
the first-best hypothesis
``unstable'', i.e.~past words can change depending on the newly received evidence due to the
global optimization of the Viterbi decoding. To make the output more stable, we made the decoder delay the recognition results until all active word sequences share a common prefix.
This prefix is then guaranteed to remain unchanged independent of the rest of the utterance and thus can be sent out to the MT model.
\section{Experiments}
We conduct experiments on the IWSLT simultaneous translation task for speech translation of TED talks from English to German.
\subsection{Setup}
For training the baseline NMT system, we utilize the parallel data allowed for the IWSLT 2020 evaluation. We divide it into 3 parts: in-domain, clean, and out-of-domain. We consider data from the TED and MUST-C corpora~\citep{mustc19} as in-domain and use it for subsequent fine-tuning experiments, as well as the ``ground truth'' for filtering the out-of-domain data based on sentence embedding similarity with the in-domain data; details are given in \citep{pbahar2020iwsltsystem}. As ``clean'' we consider the News-Commentary, Europarl, and WikiTitles corpora and use their full versions in training. As out-of-domain data, we consider OpenSubtitles, ParaCrawl, CommonCrawl, and \textit{rapid} corpora, which we reduce to 40\% of their total size, or to 23.2M parallel lines, with similarity-based filtering. Thus, in total, we use almost 26M lines of parallel data to train our systems, which amounts to ca.~327M running words on the English side. Furthermore, we added 7.9M sentence pairs or ca.~145M running words of similarity-filtered back-translated\footnote{ The German-to-English system that we used to translate these data into English is an off-line system trained using the Transformer Base architecture~\citep{transformer} on the in-domain and clean parallel data.} German monolingual data allowed by the IWSLT 2020 evaluation.
In training, the in-domain and clean parallel data had a weight of 5. All models were implemented and trained with the RETURNN toolkit \cite{returnn_acl2018}. We used
an embedding size of 620
and LSTM state sizes of 1000.
\begin{table*}[!h]
\centering
\begin{tabular}{llllllll}
\hline
\multirow{2}{*}{\textbf{System}} & \textbf{\hspace*{-2mm} Delay} &\multicolumn{2}{c}{\textbf{tst2015}} & \multicolumn{2}{c}{\textbf{must-c-HE}} & \multicolumn{2}{c}{ \textbf{must-c-COMMON}} \\
&& \textbf{\BLEU} & \textbf{\TER} & \textbf{\BLEU} & \textbf{\TER} & \textbf{\BLEU} & \textbf{\TER} \\
\hline
\textbf{Offline baseline, Transformer} \\
\hspace{2mm} using reference transcript & n/a & 32.7 & 50.9 & 30.1 & 54.3 & 32.6 & 48.9 \\
\hspace{2mm} using streaming ASR & n/a & 28.6 & 56.3 & 26.0 & 59.2 & 26.4 & 57.3 \\
\hline
\textbf{Proposed simultaneous NMT} & 2 & 24.8 & 60.2 & 21.7 & 63.0 & 21.9 & 60.2 \\
\hspace{2mm} unidirectional, & 3 & 24.6 & 60.2 & 22.6 & 62.7 & 21.8 & 60.8 \\
\hspace{2mm} (6 enc. 2 dec.) & 4 & 24.6 & 61.1 & 22.8 & 62.8 & 21.7 & 61.5 \\
\hline
\textbf{Proposed simultaneous NMT} & 2 & 24.6 & 60.0 & 21.4 & 62.8 & 21.9 & 60.6 \\
\hspace{2mm} bidirectional, & 3 & 24.4 & 60.5 & 22.0 & 62.7 & 21.7 & 61.1 \\
\hspace{2mm} (2x4 enc.\,1 dec.) & 4 & 24.6 & 61.0 & 21.8 & 63.1 & 21.9 & 61.4 \\
\hline
\end{tabular}
\caption{\label{tbl:speech_mt} Experimental results (in \%) for simultaneous NMT of speech, IWSLT 2020 English$\to$German.}
\end{table*}
As heldout tuning set, we use a combination of IWSLT dev2010, tst2014, and MUST-C-dev corpora. To obtain bilingual chunks as described in Section~\ref{sec:chunking}, we word-align all of the filtered parallel/back-translated and tuning data
in portions of up to 1M sentence pairs, each of them combined with all of the in-domain and clean parallel data. As heldout evaluation sets, we use IWSLT tst2015, as well as MUST-C HE and COMMON test data.
For the text input condition, we applied almost no preprocessing,
tokenization was handled as part of the
subword segmentation with the sentencepiece toolkit~\citep{kudo2018sentencepiece}. The vocabularies for both the source and the target subword models had a size of 30K. For the speech input condition, the additional preprocessing applied to the English side of the parallel data had the goal to make it look like speech transcripts. We lowercased the text, removed all punctuation marks, expanded common abbreviations, especially for measurement units, and converted numbers, dates, and other entities expressed with digits into their spoken form. For the cases of multiple readings of a given number (e.g. \texttt{one oh one}, \texttt{one hundred and one}), we selected one randomly, so that the system could learn to convert alternative readings in English to the same number expressed with digits in German. Because of this preprocessing, our system for the speech condition learned to insert punctuation marks, restore word case, and convert spoken number and entity forms to digits as part of the translation process.
The proposed chunking method (Section \ref{sec:chunking}) is applied to the training corpus as a data preparation step. We measured average chunk lengths of 2.9 source words and 2.7 target words. 40\% of both the source and target chunks consist of a single word, about 20\% are longer than 3 words.
We compute case-sensitive BLEU~\citep{papineni02:bleu} and TER~\citep{snover06:ter} scores as well as the average lagging (AL) metric \citep{ma2019stacl}.
\subsection{Results}
Table \ref{tbl:speech_mt} shows results for the proposed simultaneous MT system. For reference, we first provide the translation quality of an offline system
that is trained on full sentences. It is a transformer ``base'' model \cite{transformer} that we trained on the same data as the online systems. Row 1 shows BLEU and TER scores for translation of the human reference transcription of the speech input (converted to lower-case, punctuation removed), whereas row 2 uses the automatic transcription generated by our streaming ASR system (Section \ref{sec:asr}). The ASR system has a word error rate (WER) of 8.7 to 11.2\% on the three test sets, causing a drop of 4-6\% BLEU absolute.
All following systems are cascaded streaming ASR + MT online systems that produce translations from audio input in real-time. Those systems have an overall AL of 4.1 to 4.5 seconds, depending on $D$. We compare between two categories of models: unidirectional and bidirectional. For the unidirectional models the backwards decoder (Equation \ref{eq:backward}) was removed from the architecture. We show results for different values of source boundary delay $D$ (see Section \ref{subsec:delay}). For the number of layers we choose $L_\text{enc}=6$ and $L_{\text{dec}}=2$ for the unidirectional models, and $L_{\text{enc}}=4$ (both directions) and $L_\text{dec}=1$ for the bidirectional models, such that the number of parameters is comparable.
Contradicting our initial assumption,
bidirectional models do not outperform unidirectional models. This might indicate be due to the fact that the majority of chunks are too short to benefit from a backwards encoding. Also, the model is not sensitive to the delay $D$. This \textit{confirms} our assumption that the additional context of future source words is primarily useful for making the source boundary decision, and for this a context of 2 following (sub-)words is sufficient. For translation, the model does not depend on this ``extra'' context but instead is able to make sufficiently good chunking decisions.
Table \ref{tbl:text_mt} shows results for the case of streamed text input (cased and with punctuation marks). We compare our results to a 4-layer unidirectional system that was trained using the wait-$k$ policy \citep{ma2019stacl}. For this, we chunk the training data into single words, except for a first chunk of size $k=9$ on the source side, and set the delay to $D=0$.
All of our systems
outperform this wait-$k$ system by large margins. We conclude that the alignment-based chunking proposed in Section \ref{sec:chunking} is able to provide better source context than a fixed policy and that the source boundary detection component described in Section \ref{sec:boundary_detection} successfully learns to reproduce this chunking at inference time. Also for the text condition, we do not observe large differences between uni- and bidirectional models and between different delays.
\begin{table*}[!h]
\centering
\begin{tabular}{lclllllll}
\hline
\multirow{2}{*}{\textbf{System}} & \multirow{2}{*}{\hspace*{-4mm}\textbf{Delay}}& \multirow{1}{*}{\textbf{Avg.}} & \multicolumn{2}{c}{\textbf{tst2015}} & \multicolumn{2}{c}{\textbf{must-c-HE}} & \multicolumn{2}{c}{\textbf{must-c-COMMON}} \\
& & \textbf{AL} & \textbf{\BLEU} & \textbf{\TER} & \textbf{\BLEU} & \textbf{\TER} & \textbf{\BLEU} & \textbf{\TER} \\
\hline
\textbf{Baseline} wait-k (k=9) & - & & 27.4 & 55.8 & 25.1 & 59.5 & 27.4 & 54.2 \\
\hline
\textbf{Proposed simultaneous MT} & 2 & 4.72 & 30.2 & 52.6 & 28.8 & 55.0 & 30.0 & 50.8 \\
\hspace{2mm} unidirectional & 3 & 5.26 & 30.3 & 53.2 & 28.8 & 55.2 & 29.7 & 50.8 \\
\hspace{2mm} (6 enc.\,2 dec.) & 4 & 6.17 & 29.9 & 53.1 & 28.6 & 55.1 & 29.6 & 50.8 \\
\hline
\textbf{Proposed simultaneous MT} & 2 & 4.65 & 29.3 & 53.4 & 28.1 & 55.4 & 29.7 & 50.9 \\
\hspace{2mm} bidirectional & 3 & 5.46 & 29.6 & 53.7 & 29.0 & 54.8 & 29.7 & 51.6 \\
\hspace{2mm} (2x4 enc.\,1 dec.) & 4 & 6.15 & 29.2 & 54.0 & 28.3 & 55.3 & 29.7 & 51.5\\
\hline
\end{tabular}
\caption{\label{tbl:text_mt} Experimental results (in \%) for simultaneous NMT of text input, IWSLT 2020 English$\to$German.}
\end{table*}
\begin{figure}
\vspace*{-1.5mm}
\centering
\include{BLEU_vs_AL}
\vspace*{-6mm}
\caption{BLEU vs. AL for bidirectional systems from Table \ref{tbl:text_mt}, generated using a delay $D$ of $2$, $3$, and $4$. }
\label{fig:bleu_al}
\end{figure}
For all systems, we report AL scores averaged over all test sets.
Figure \ref{fig:bleu_al} breaks down the scores to the individual test sets for the bidirectional models.
For a source boundary delay $D=2$ we observe an AL of $4.6$ to $4.7$ words. When increasing $D$, we increase the average lagging score by roughly the same amount, which is expected, since the additional source context for the boundary decision is not translated in the same step where it is added. As discussed before, translation quality does not consistently improve from increasing $D$.
\begin{figure}
\centering
\include{len_norm_tuning}
\vspace*{-6mm}
\caption{BLEU vs. the length normalization factor ($\alpha$) on the tuning set (dev2010 + tst2014 + MUST-C-dev).}
\label{fig:len_norm_tuning}
\vspace*{-5mm}
\end{figure}
We found tuning of length normalization to be important, as the average decoding length for chunks is much shorter than in offline translation. For optimal results, we divided the model scores by $I^\alpha$, $I$ being the target length, and tuned the parameter $\alpha$. Figure \ref{fig:len_norm_tuning} shows that $\alpha=0.9$ works best in our experiments, independent of the source boundary delay $D$. This value is used in all experiments.
Furthermore, we found the model to be very sensitive to a source boundary probability threshold $t_b$ different than $0.5$ regarding translation quality. This means the ``translating'' part of the network strongly adapts to the chunking component.
\section{Conclusion}
We proposed a novel neural model architecture for simultaneous MT that incorporates a component for splitting the incoming source stream into translatable chunks. We presented how we generate training examples for such chunks from statistical word alignment and how those can be improved via language models. Experiments on the IWSLT 2020 English-to-German task proved that the proposed learned source chunking outperforms a fixed wait-$k$ strategy by a large margin. We also investigated the value of backwards source encoding in the context of simultaneous MT by comparing uni- and bidirectional versions of our architecture.
\section*{Acknowledgements}
We would like to thank our colleague Albert Zeyer for fruitful discussions and RETURNN implementation support.
| c1e05b3eac53d85e18b4c65580cf74ab2ca270ad | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Methods}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/networks.pdf}
\caption{An overview of CoDiNet. \textbf{Similar Image Generation}: Each sample is augmented randomly and get several similar augmentations. \textbf{Dynamic Routing Network}: Through a dynamic routing network, we get routing paths and prediction for each sample. \textbf{Path Distribution Optimization}: We optimize the path distribution with the consistency loss $\mathcal{L}_{con}$ and the diversity loss $\mathcal{L}_{div}$. Note that the augmentation methods are commonly used in classification, which are cropping and horizontally flipping. Best viewed in color.
}
\label{fig.min_max}
\end{figure*}
In this section, we illustrate our method CoDiNet in detail.
First, we formulate dynamic routing as a space mapping and introduce the basics of dynamic routing accordingly.
Second, we show how to model the relationship between the two spaces with the regularization of consistency and diversity. Third, we design a training strategy to make our dynamic routing network adaptive to different computational budgets. Finally, we illustrate our process of training and inference.
For convenience, \cref{tab:notation} summarizes the notations.
\begin{table}[tb]
\centering
\caption{Notations}
\label{tab:notation}
\vspace{-1em}
\begin{tabu} to 0.49\textwidth {X[c]X[4]}
\toprule
$\mathcal{S}$ & a sample space\\
$s_i$ & a sample \\
$\mathcal{R}$ & a routing space\\
$r_i$ & the routing path for $s_i$\\
$\mathcal{\phi}$& an $n$-block network\\
$F_k$ & the $k$-th block of $\mathcal{\phi}$\\
$a_k$ & the input of the $(k+1)$-th block of $\mathcal{\phi}$\\
$U_k$ & the router for the $k$-th block \\
$u_k$ & the execution decision of the $k$-th block \\
$v_k$ & the relaxation of $u_k$ \\
\midrule
$m_c$ & the margin for consistency \\
$m_d$ & the margin for diversity \\
$\mathcal{L}_{con}$ & the loss function for consistency \\
$\mathcal{L}_{div}$ & the loss function for diversity \\
$\mathcal{L}_{cost}$ & the loss function for customizable dynamic routing\\
$\alpha$ & the hyper-parameter for $\mathcal{L}_{con}$ \\
$\beta$ & the hyper-parameter for $\mathcal{L}_{div}$ \\
$\gamma$ & the hyper-parameter for $\mathcal{L}_{cost}$ \\
\bottomrule
\end{tabu}
\end{table}
\subsection{Dynamic Routing as a Space Mapping}
\subsubsection{Routing Space}
We see dynamic routing as a mapping from a sample space to a routing space. A sample space $\mathcal{S}$ is a set of samples, and a routing space $\mathcal{R}$ is the set of all the possible routing paths of a dynamic routing network $\mathcal{\phi}$. In this way, dynamic routing can be considered as a mapping $\mathcal{\phi}:\mathcal{S} \rightarrow \mathcal{R}$. That is, for each sample $s_i\in\mathcal{S}$, its routing path $r_i \in \mathcal{R}$ in a dynamic routing network $\mathcal{\phi}$ is
\begin{equation}
r_i = \mathcal{\phi}(s_i).
\end{equation}
A routing path consists of a sequence of to-be-skipped and to-be-run blocks. In this paper, we use the same block as in ResNet~\cite{he2016deep}. Let $\mathcal{\phi}$ be an $n$-block network and $u_k \in \{0,1\}$ be the execution decision for the $k$-th block, where $0$ stands for to-be-skipped and $1$ stands for to-be-run. Then, a routing path for the sample $s_i$ is a concatenation of decisions for all the blocks, i.e., $r_i = (u_1, u_2, \cdots, u_n)$.
Hence, the routing space of $\mathcal{\phi}$ is $\mathcal{R}={\{0, 1\}^{n}}$, which contains $2^n$ routing paths in total
\subsubsection{Routers}
For each block in the network $\mathcal{\phi}$, there is a router used to decide whether the block should be executed for a specific sample. Let $F_{k}$ be the $k$-th block of $\mathcal{\phi}$, $a_{k-1}$ be the input of $F_{k}$, and $U_k$ be the router for the $k$-th block.
Then, the dynamic routing result $a_k$ of the $k$-th block is defined as
\begin{equation}
a_{k} = u_{k} \cdot F_{k}(a_{k-1}) + (1-u_{k}) \cdot a_{k-1},
\end{equation}
where the value of $u_k$ stands for the decision of $U_k$.
Routers are supposed to find the correct path while incurring a low computational cost.
To minimize the cost, we use a lightweight structure for each router, which only contains two fully connected layers, as shown in \cref{sfig:routing_module}.
First, the router $U_k$ gathers information from the block input $a_{k-1} \in \mathbb{R}^{W_k \times H_k \times C_k}$ across channels by global average pooling, which is written as
\begin{equation}
z_k = \mathrm{gap}(a_{k-1}).
\end{equation}
Then, the fused features $z_k$ are processed by two fully connected layers sequentially. Let $\mathrm{W}_1 \in \mathbb{R}^{d \times C_k}$ and $\mathrm{W}_2 \in \mathbb{R}^{2 \times d}$ be the weights of $U_k$'s first layer and second layer respectively, where $d$ denotes the output dimension of the first layer. The router $U_k$ is defined as
\begin{equation}
U_k = \mathrm{W}_2 \circ \sigma(\mathrm{W}_1 \circ z_k),
\end{equation}
where $\sigma(\cdot)$ is an activation function, and $\circ$ denotes matrix multiplication. It is worth noting that $U_k$ is a two-element vector as a result. $U_k[0]$ is the first element of $U_k$, and $U_k[1]$ is the second element. In this way, the execution decision $u_k$ of the $k$-th block is calculated by
\begin{equation}
u_k = \mathop{\arg\max}_j(U_k[j]),
\end{equation}
where $u_k=0$ means to-be-skipped, and $u_k=1$ means to-be-run as a result.
The simple yet effective structure incurs less than one percent of the computational cost of a convolution block.
\begin{figure*}[t]
\centering
\subfigure[Routing Space]{
\centering
\includegraphics[height=1.6in, width=0.23\linewidth]{figures/insight1.png}
\label{sfg.insight_a}
}
\subfigure[Augmentation]{
\centering
\includegraphics[height=1.6in, width=0.23\linewidth]{figures/insight2.png}
\label{sfg.insight_b}
}
\subfigure[Consistency]{
\centering
\includegraphics[height=1.6in, width=0.23\linewidth]{figures/insight3.png}
\label{sfg.insight_c}
}
\subfigure[Diversity]{
\centering
\includegraphics[height=1.6in, width=0.23\linewidth]{figures/insight4.png}
\label{sfg.insight_d}
}
\caption{Illustration of the optimization of path distribution. A solid point means a routing path of an image. A hollow circle means a routing path center of similar images. (a) Suppose we have four different input images whose paths distribute in the routing space. (b) With augmentations, we can get four similar instances of each image. (c) The consistency regularization makes routing paths of similar instances cluster around their center. (d) The diversity regularization disperses the center of dissimilar instances. Best viewed in color.
}
\label{fig.insight}
\end{figure*}
\subsubsection{Route Relaxation}
\label{sec.relaxation}
To make the binary routing decisions $u_1, u_2, \cdots, u_n$ optimizable in an end-to-end fashion, we utilize a continuous, differentiable relaxation function, Gumbel-Softmax \cite{gumbel1948statistical}, at the end of each router as shown in \cref{sfig:routing_module}. Gumbel-Softmax turns discrete values into continuous ones, enabling backpropagation.
Let $g$ be the noise samples from a Gumbel distribution, and let $\mathcal{T}$ be the temperature which is fixed to $1$ in our experiments. The relaxation for $u_k$, called $v_k$, is calculated by
\begin{equation}
v_k = \mathrm{softmax}(\frac{\log(U_k)+g}{\mathcal{T}})[1],
\end{equation}
where $\mathrm{softmax}(\cdot)[1]$ is the second element.
In this way, the dynamic routing result of the $i$-th block is relaxed by
\begin{equation}
a_{k} = v_k \cdot F_{k}(a_{k-1}) + (1-v_k) \cdot a_{k-1}.
\label{eq1}
\end{equation}
The routing path $r_i$ for an $n$-block network is also relaxed by $r_i = (v_1, v_2, \cdots, v_n)$, where each element is in $[0,1]$.
It is worth noting that a block in $\mathcal{\phi}$ is either run or skipped at inference time. We will describe it in \cref{sssec:inference}.
Next, we propose to regularize routing paths in the routing space.
\subsection{Consistency and Diversity Regularization}
\label{loss}
We expect that the routing paths of similar samples should be consistent in the routing space. Otherwise, the routing paths should be diverse. In real scenarios, the similarity of different images is difficult to measure. Therefore, we propose to regularize the routing paths of samples according to their self-supervised similarity, i.e., considering the augmentations of an image as its similar samples. To this end, we first generate similar images in \cref{sssec:generation}. To realize consistency for similar samples and diversity for dissimilar samples, we directly regularize path distributions in \cref{sssec:consistency} and \cref{sssec:diversity}. An overview of the proposed optimization method is shown in \cref{fig.min_max}.
\subsubsection{Similar Images Generation}
\label{sssec:generation}
To obtain similar samples, we randomly augment each original sample several times by random cropping and horizontal flipping. We treat the set of augmentations as similar inputs as shown in \cref{sfg.insight_a} and \cref{sfg.insight_b}. Suppose a training batch containing $L$ samples, and $M$ augmentations for each sample. Then, we have an augmented set $\{s_{i,1}, s_{i, 2}, \cdots, s_{i,M}\}$ for a sample $s_i$, where $s_{i,j}$ stands for the $j$-th augmentation for $s_i$. Therefore, there are $L \times M$ inputs in an iteration, i.e., $\{s_{1,1}, s_{1, 2}, \cdots, s_{1,M}\}, \cdots, \{s_{L,1}, s_{L,2}, \cdots, s_{L,M}\}$.
\subsubsection{Consistency Regularization}
\label{sssec:consistency}
As shown in \cref{sfg.insight_c}, consistency regularization is an attractive force in each group of similar samples to make them closer. Let $r_{i,j}$ be the routing path of sample $s_{i,j}$, $\overline{r_{i,:}}$ be the mean of $\{r_{i,1}, r_{i,2}, \cdots, r_{i,M}\}$, which represents the routing path center of sample $s_{i}$'s augmentations.
In this way, the optimization of consistency is written as
\begin{equation}
\begin{aligned}
\label{eq.consistent}
\mathcal{L}_{con}= & \dfrac{1}{L}\dfrac{1}{M}\sum_{i=1}^{L} \sum_{j=1}^{M}[\left \|\, r_{i,j}-\overline{r_{i,:}}\, \right \| - m_c]_+^2,
\end{aligned}
\end{equation}
where $L$ is the batch size, $M$ is the number of augmentations,
$\| \cdot \|$ is the $L^2$ norm, $m_c$ is the margin for consistency, and $[x]_+ = max(0, x)$ denotes the hinge.
The minimization of $\mathcal{L}_{con}$ is to narrow down the differences between all the routing paths and the mean routing path. To further illustrate the consistency, we visualize the path distribution of five classes from the CIFAR-10 test set in \cref{fig.consistency}. In the first row of \cref{fig.consistency}, the paths of the vanilla dynamic routing method scatter in the routing space randomly. With consistency regularization, paths cluster around the center in the routing space, becoming consistent as shown in the second row of \cref{fig.consistency}.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figures/consistency_new.pdf}
\caption{Visualization of the routing path distribution under different constraints through t-SNE on the CIFAR-10 test set. Figures in the first row show the path distribution of the vanilla dynamic routing method without $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$. Figures in the second row show the path distribution of the experiments with $\mathcal{L}_{con}$ only. Figures in the third row show the path distribution of our method. All figures are in the same coordinate scales.
}
\label{fig.consistency}
\end{figure*}
\subsubsection{Diversity Regularization}
\label{sssec:diversity}
Conversely, as shown in \cref{sfg.insight_d}, diversity regularization is a repulsive force between each group of similar samples to push them away. The optimization of diversity is defined as
\begin{equation}
\label{eq.diverse}
\begin{aligned}
\mathcal{L}_{div}= & \dfrac{1}{L}\dfrac{1}{L-1} \sum_{i=1}^{L} \sum_{j \neq i} [m_d-\left \|\overline{r_{i,:}} -\overline{r_{j,:}}\,\right \|]_+^2,
\end{aligned}
\end{equation}
where $m_d$ is the margin for diversity, and $\sum_{j\neq i}$ means all samples in the batch except $s_i$. Within a batch, the mean routing path is optimized to maximize the differences for different groups. In this way, the routing paths of different groups are dispersed and the diversity of routing paths can be guaranteed.
Another advantage of diversity regularization is that it helps the network explore more paths. Path distribution in the real scenario under diversity regularization is also shown in \cref{fig.consistency}. Compared with the first two rows, samples in the third row cluster around several centers and different clusters keep distant from each other, which is in line with our expectations.
In our method, consistency and diversity are two facets of the problem. Although $\mathcal{L}_{con}$ makes the routing space compact, it also runs a risk of making the whole routing space collapse into a small space, impairing the diversity of routing paths. Thus, introducing $\mathcal{L}_{div}$ can compensate for the disadvantage of stand-alone $\mathcal{L}_{con}$ and make the routing paths diverse at the same time. Similarly, $\mathcal{L}_{div}$ can enhance the routing space exploration; however, making routing paths scattered without constraint can be harmful to the parameter sharing among routing paths. As a result, we propose to make use of $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$ together.
\subsection{Customizable Dynamic Routing}
\label{sec.cost}
Dynamic routing is aimed at saving the cost of a network at inference time. How much cost dynamic routing should save for a network depends on the application scenario. Since the computational budget is different from device to device,
it is better to make a dynamic routing network adaptive to devices.
We design a learning strategy to make a dynamic routing network adaptive to different computational budgets.
Let $c_k$ be the computational cost of the $k$-th block. The total cost for an $n$-block network is defined as
\begin{equation}
{cost}_{all} = \sum_{k=1}^{n} c_k \cdot u_k.
\end{equation}
To make it learnable, we use the relaxed continuous routing variable, $v_k$, similar to \cref{eq1}. After relaxation, the loss function for cost optimization $\mathcal{L}_{cost}$ becomes
\begin{equation}
\mathcal{L}_{cost} = \sum_{k=1}^{n} c_k \cdot v_k.
\end{equation}
It is worth noting that we build a computational cost lookup table which records the floating-point operations (FLOPs) of each block. During optimization, each block $F_k$ will be assigned a cost $c_k$ given by the lookup table.
By putting all the losses together, the overall objective of our method is
\begin{equation}
\mathcal{L}_{total} = \mathcal{L}_{cls} + \alpha \cdot \mathcal{L}_{con} + \beta \cdot \mathcal{L}_{div}+ \gamma \cdot \mathcal{L}_{cost},
\label{eq.total}
\end{equation}
in which $\mathcal{L}_{cls}$ is the cross entropy loss used for classification, $\mathcal{L}_{con}$ is the loss defined in \cref{eq.consistent}, and $\mathcal{L}_{div}$ is the loss defined in \cref{eq.diverse}. $\alpha$, $\beta$, and $\gamma$ are the hyper-parameters for respective losses.
To make a dynamic routing network adaptive to different computational budgets, we tune the hyper-parameter $\gamma$ of $\mathcal{L}_{cost}$. Therefore, we can customize the learned network with different computational costs and performances. That is, a smaller $\gamma$ encourages an expensive model, and a larger $\gamma$ encourages an inexpensive model.
\begin{table*}[t]
\caption{The Efficiency and Accuracy Trade-off based on ResNet-110 on CIFAR-10}
\label{tab.Detailed_Data}
\vspace{-1em}
\renewcommand\arraystretch{1.35}
\setlength{\tabcolsep}{4.85mm}{
\begin{tabular}{lccccccc}
\toprule
\multicolumn{1}{l}{Settings} & \multicolumn{1}{c}{GMACCs} & \begin{tabular}[c]{@{}c@{}}Speedup\\ (in GMACCs)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Inference Time\\ GTX-1080 (ms)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Inference Time\\ GTX-1080ti (ms)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Inference Time\\ RTX-2080ti (ms)\end{tabular} & Acc. (\%)\\ \midrule
ResNet-110 & 0.51 & 1.0$\times$ & 17.2 & 16.4 & 15.3 & 93.60 \\ \hline
$\gamma=0.01$ & 0.29 & 1.8$\times$ & 9.43 & 8.69 & 8.56 & 94.47 \\
$\gamma=0.02$ & 0.27 & 1.9$\times$ & 8.53 & 8.41 & 8.16 & 94.30 \\
$\gamma=0.04$ & 0.22 & 2.3$\times$ & 7.69 & 7.62 & 7.37 & 93.94 \\
$\gamma=0.08$ & 0.20 & 2.6$\times$ & 7.66 & 7.18 & 6.98 & 93.71 \\
$\gamma=0.10$ & 0.10 & 5.1$\times$ & 3.85 & 3.74 & 3.29 & 92.45 \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize GMACCs refers to billions of multiply-accumulates. $\gamma$ is the weight for $\mathcal{L}_{cost}$ defined in Eq.(12). The unit of inference time is millisecond.}
\vspace{-1.0em}
\end{table*}
\begin{table}[t]
\caption{Ablation Study of Routers, $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$ on CIFAR-10}
\renewcommand\arraystretch{1.3}
\label{tab:ablation_study}
\vspace{-1em}
\setlength{\tabcolsep}{2.4mm}{
\begin{tabular}{l|ccccc}
\toprule
\multicolumn{1}{l|}{Methods} & Routers & $\mathcal{L}_{con}$ & \multicolumn{1}{c|}{$\mathcal{L}_{div}$} &\#Path & Acc. (\%)\\ \midrule
ResNet-110 & --- &--- & \multicolumn{1}{c|}{---} & 1 & 93.60 \\
Vanilla & $\checkmark$ &--- & \multicolumn{1}{c|}{---} & 113 & 93.66 \\ \midrule
Vanilla + $\mathcal{L}_{con}$& $\checkmark$ &$\checkmark$ & \multicolumn{1}{c|}{---} &20 & 92.88 \\
Vanilla + $\mathcal{L}_{div}$& $\checkmark$ & --- & \multicolumn{1}{c|}{$\checkmark$} & 1079 & 91.34 \\
CoDiNet &$\checkmark$ & $\checkmark$ & \multicolumn{1}{c|}{$\checkmark$} &276 & \textbf{94.47} \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize Vanilla is the vanilla dynamic routing network. The number of activated routing paths and the accuracy under different constraints based on ResNet-110.}
\end{table}
\subsection{Training and Inference}
\label{sssec:inference}
At training time, we optimize our network in two stages. First, we train the parameters of all the blocks and all the routers together by \cref{eq.total}.
Second, we finetune our network to narrow down the decisions' difference between training and inference.
We use continuous decisions $(v_1, v_2, \cdots, v_n)$ as a relaxation of binary decisions $(u_1, u_2, \cdots, u_n)$ to make the first training stage end-to-end, but training with the continuous relaxation $v_k$ of the binary decision $u_k$ will inevitably cause a gap between training and inference.
Therefore, we finetune the network with the parameters in each router fixed.
In terms of inference, a block in the network is either run or skipped exclusively. That is,
\begin{equation}
a_{k} = \begin{dcases}
F_{k}(a_{k-1}), & \text{if } v_k\geq 0.5;\\
a_{k-1}, & \text{otherwise.}
\end{dcases}
\end{equation}
\section{Conclusion}
In this paper, we see routing mechanisms from a novel perspective that regards a dynamic routing network as a mapping from a sample space to a routing space. From this view, path distribution in routing space is a fundamental problem in a dynamic routing network. We propose a novel framework CoDiNet to regularize path distribution with diversity and consistency. Moreover, we design a customizable dynamic routing module enabling the network to adapt to different computational budgets.
We compare CoDiNet with state-of-the-art methods on four benchmark datasets, demonstrating that it can effectively reduce the computational cost without compromising performance.
\section*{Acknowledgement}
This work is supported in part by National Key Research and Development Program of China under Grant 2020AAA0107400, National Natural Science Foundation of China under Grant U20A20222, Zhejiang Provincial Natural Science Foundation of China under Grant LR19F020004, and key scientific technological innovation research project by Ministry of Education.
\section{Introduction}\label{sec:introduction}}
\else
\section{Introduction}
\label{sec:introduction}
\fi
\IEEEPARstart{D}{ynamic} routing is a sample-adaptive inference mechanism for neural networks. At inference time, only part of a dynamic routing network would be activated for each sample, which is aimed at reducing computational cost with little performance compromised~\cite{veit2018convolutional, wang2018skipnet, wu2018blockdrop, almahairi2016dynamic}.
In essence, dynamic routing can be considered as a mapping from a sample space to a routing space. As shown in \cref{fig.mapping}, when samples are presented to a dynamic routing model, they are mapped into a routing space. Each sample's routing path in the routing space can be represented as a binary vector, consisting of a sequence of to-be-run and to-be-skipped blocks. The model walks through the to-be-run blocks.
From the perspective of space mapping, how inference
paths should be distributed in the routing space remains relatively unexplored.
Since dynamic routing is related to two spaces, i.e., a sample space and a routing space, we focus on this question: \textit{how do we model the relationship between the two spaces?}
We expect that the distance between similar samples should be close in the routing space. Otherwise, the distance should be far.
Moreover, \textit{what kind of samples are similar?} We utilize the self-supervised similarity and encourage the self-supervised augmentations walking through close routing paths. Since the images obtained by self-supervised augmentations remain semantically and visually unchanged, their routing paths should be close in the routing space. In comparison, images belong to the same semantic class might be significantly different due to different background, object layout, and color. The feature extraction patterns of these samples are different so that it is infeasible to utilize close routing paths for such visually different images.
As shown in \cref{fig_idea}, images $a$ and $a'$, images $b$ and $b'$ are similar to each other, while $c$ and $c'$, $d$ and $d'$ are dissimilar. Therefore, the routing paths of images $a$ and $a'$, images $b$ and $b'$ should be the same or similar, while routing paths of images $c$ and $c'$, $d$ and $d'$ should be different.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/mapping.pdf}
\caption{Illustration of dynamic routing. A dynamic routing network is a mapping from a sample space to a routing space. Each routing path consists of a sequence of to-be-run and to-be-skipped blocks. The dynamic routing model walks through the to-be-run blocks. Best viewed in color.
}
\label{fig.mapping}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figures/introduction.pdf}
\caption{Illustration of our motivation. \textbf{Left:} Schematic diagram of a dynamic routing network, in which each layer can be either executed or skipped. \textbf{Middle:} A demonstration of the routing space. All potential routing paths compose a binary routing space. \textbf{Right:} Routing paths for similar images should be the same or similar, e.g., $a$ and $a'$, $b$ and $b'$, while routing paths for dissimilar images should be different, e.g., $c$ and $c'$, $d$ and $d'$. Best viewed in color.}
\label{fig_idea}
\end{figure*}
To this end, we explicitly model the relationship between the sample space and the routing space, therein establishing the connection between samples and routing paths. We propose a novel dynamic routing method, termed CoDiNet, to regularize the path distribution in a routing space with the properties of consistency and diversity, based on the aforementioned space mapping.
Firstly, the consistency regularization makes augmentations of the same sample have similar feature activations, thus forming a specific routing paths for specific samples. Parameters on the specific routing path are consistently stimulated by similar samples, which is favorable to parameter sharing among similar samples and robustness of the network.
At the same time, the diversity regularization makes the routing paths generated by dynamic routing more diverse, which strengthens the exploration of the network. It is evident that the more routing paths are used, the more capacity of the network is utilized.
Computational cost is another important issue because the computational ability of different platforms varies considerably. For instance, the inference speed of ResNet-50~\cite{he2016deep} on GTX 1080ti \textit{(30fps)} is much faster than that on Maxwell TitanX \textit{(18fps)} at a resolution of $224\times224$. The average cost of dynamic routing networks should be customizable, when the networks are applied to platforms with different computational budgets.
To this end, we propose a differentiable computational cost loss to optimize the average computational cost of the model and make the computational budgets customizable.
The main contributions of CoDiNet can be summarized into three parts:
\begin{itemize} [leftmargin=*]
\item We explicitly model the relationship between the sample space and the routing space, therein establishing the connection between samples and routing paths.
\item We propose a novel consistency-and-diversity regularized optimizing method modeling the relationships between the two spaces, which makes routing paths optimizable.
\item We design a customizable dynamic routing module and achieve state-of-the-art results in terms of computational cost reduction and performance.
\end{itemize}
\begin{figure*}[h]
\subfigure[Dynamic routing]{
\label{sfig:dynamic_routing}
\centering
\includegraphics[width=.45\linewidth, height=1.5in]{figures/router_a.pdf}
}
\subfigure[Our router]{
\label{sfig:routing_module}
\centering
\includegraphics[width=.53\linewidth, height=1.5in]{figures/router.pdf}
}
\caption{Illustration of dynamic routing and our router. (a) Comparison between static inference and dynamic routing. In the dynamic routing network, whether the current convolution block would be executed depends on the output of the previous block.
(b) Illustration of the structure of our router. The cost of our router is negligible compared with a convolution block.}
\label{fig.route_structure}
\end{figure*}
\section{Experiments}
In this section, we conduct a series of experiments to evaluate the performance of CoDiNet. First, we introduce the experimental setup. Second, we perform the ablation studies for the proposed regularization. Third, we compare our results with the state-of-the-art works and other related methods. Next, we show qualitative analysis on the proposed modules in our method. In the end, we compare different routing strategies and structures.
\subsection{Experimental Setup}
\label{detail}
\subsubsection{Datasets and Metrics}
We evaluate our method on four widely used datasets, which are CIFAR-10~\cite{krizhevsky2009learning}, CIFAR-100~\cite{krizhevsky2009learning}, SVHN~\cite{netzer2011reading}, and ImageNet~\cite{deng2009imagenet} (ILSVRC2012). CIFAR-10/100 consists of 60,000 colored images, which are resized to 32$\times$32. Out of the 60,000 images, 50,000 images are used for training, and the other 10,000 images are used for testing. SVHN includes 73,257 training images and 26,032 testing images. ImageNet contains 1,281,167 training images and 50,000 validation images that are annotated with 1,000 classes and resized to 224$\times$224. We use classification accuracy (top-1) as an evaluation metric.
\begin{table}[t]
\caption{Results on SVHN}
\label{tab:SVHN_result}
\vspace{-1em}
\setlength{\tabcolsep}{2.6mm}{
\renewcommand\arraystretch{1.3}
\begin{tabular}{lccc}
\toprule
Networks & ResNet-110 & Vanilla dynamic routing & CoDiNet \\ \midrule
Acc. (\%) & 94.19 & 93.15 & 94.28 \\
GMACCs & 0.51 & 0.38 & 0.39 \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize Results are based on ResNet-110.
Vanilla dynamic routing is the method without $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$.}
\end{table}
\subsubsection{Implementation Details}
For data augmentation, we follow the settings as in~\cite{veit2018convolutional, wang2018skipnet}.
Images from CIFAR-10/100 are padded with 4 pixels on each side. Images from all datasets except SVHN are randomly cropped and horizontally flipped with a probability of 0.5. Those from SVHN are randomly cropped only.
On CIFAR-10/100 and SVHN, the lightweight ResNets~\cite{he2016deep} are adopted as the backbones, including ResNet-32, ResNet-74 and ResNet-110. On ImageNet, ResNet-50 and ResNet-101 are adopted as the backbones. Finally, the computational is measured in GMACCs, i.e., billions of multiply-accumulate operations as used in~\cite{veit2018convolutional, wang2018skipnet, wu2018blockdrop, almahairi2016dynamic}.
During training, we use SGD as an optimizer with a momentum of 0.9 and a weight decay of 1e-4. The initial learning rate is set to 0.1 and a multi-step scheduler is adopted. On CIFAR10/100, the step-wise learning rate decays by 0.1 at 150 and 200 epochs. As for ImageNet, it decays by 0.1 every 30 epochs. As for the loss hyper-parameters, $\alpha$ and $\beta$ are set to 0.2 respectively. To control the computational cost precisely, the hyper-parameter $\gamma$ for $\mathcal{L}_{cost}$ is tuned adaptive. Besides, the margin for consistency $m_c$ and the margin for diversity $m_d$ are set to 0.2 and 0.5 respectively.
\begin{table*}[t]
\caption{Results on CIFAR-10/100}
\label{tab:result_10_100}
\vspace{-1em}
\setlength{\tabcolsep}{1.7mm}{
\renewcommand\arraystretch{1.4}
\begin{tabular}{clccccccccc}
\toprule
& & \multicolumn{3}{c}{ResNet-32} & \multicolumn{3}{c}{ResNet-74} & \multicolumn{3}{c}{ResNet-110} \\ \cmidrule{3-11}
& & Acc. (\%) & \#Param (M) & \multicolumn{1}{c|}{GMACCs} & Acc. (\%) & \#Param (M) & \multicolumn{1}{c|}{GMACCs} & Acc. (\%) & \#Param (M) & GMACCs \\ \midrule
\multirow{3}{*}{\rotatebox{90}{CIFAR-10}} & \multicolumn{1}{l|}{ResNets} & 92.40 & 0.46 & \multicolumn{1}{c|}{0.14} & 93.30 & 1.13 & \multicolumn{1}{c|}{0.34} & 93.60 & 1.71 & 0.51 \\
&\multicolumn{1}{l|}{Vanilla dynamic routing} & 91.55 & 0.49 & \multicolumn{1}{c|}{0.09} & 93.17 & 1.21 & \multicolumn{1}{c|}{0.18} & 93.66 & 1.83 & 0.30 \\
&\multicolumn{1}{l|}{CoDiNet} & \bf{92.48} & 0.49 & \multicolumn{1}{c|}{0.09} & \bf{93.61} & 1.21 & \multicolumn{1}{c|}{0.19} & \bf{94.47} & 1.83 & 0.29 \\ \midrule
\multirow{3}{*}{\rotatebox{90}{CIFAR-100}} & \multicolumn{1}{l|}{ResNets} & 68.7 & 0.46 & \multicolumn{1}{c|}{0.14} & 70.5 & 1.13 & \multicolumn{1}{c|}{0.34} & 71.2 & 1.71 & 0.51 \\
&\multicolumn{1}{l|}{Vanilla dynamic routing} & 66.4 & 0.49 & \multicolumn{1}{c|}{0.09} & 69.7 & 1.21 & \multicolumn{1}{c|}{0.20} & 70.3 & 1.83 & 0.24 \\
&\multicolumn{1}{l|}{CoDiNet} & \bf{69.2} &0.49 & \multicolumn{1}{c|}{0.11} & \bf{70.9} & 1.21 & \multicolumn{1}{c|}{0.21} & \bf{72.9} & 1.83 & 0.24 \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize Vanilla dynamic routing only uses our routers without $\mathcal{L}_{con}$ or $\mathcal{L}_{div}$. GMACCs refers to billions of multiply-accumulates. \#Param is the number of parameters.}
\end{table*}
\begin{table}[t]
\caption{Comparison with State-of-the-Arts on CIFAR-10}
\label{tab.sota_result}
\vspace{-1em}
\setlength{\tabcolsep}{4.0mm}{
\renewcommand\arraystretch{1.3}
\begin{tabular}{llcccc}
\toprule
Methods & Backbones & GMACCs & Acc. (\%) \\
\midrule
\emph{baseline} \\
ResNet-32 & --- & 0.14 & 92.40 \\
ResNet-110& --- & 0.50 & 93.60 \\ \midrule
\emph{dynamic routing} \\
SkipNet~\cite{wang2018skipnet} & ResNet-74&0.09 & 92.38 \\
BlockDrop~\cite{wu2018blockdrop} & ResNet-110&0.17 & 93.60 \\
Conv-AIG~\cite{veit2018convolutional} & ResNet-110&0.41 & 94.24 \\
IamNN~\cite{leroux2018iamnn} & ResNet-101 & 1.10 & 94.60 \\
CGap~\cite{du2019efficient} & ResNet-110 & 0.19 & 93.43\\
\midrule
\emph{early prediction}\\
ACT~\cite{graves2016adaptive} & ResNet-110& 0.38 & 93.50 \\
SACT~\cite{figurnov2017spatially} & ResNet-110&0.31 & 93.40 \\
DDI~\cite{wang2020dual} & ResNet-74 &0.14 &93.88 \\
DG-Net~\cite{shafiee2019dynamic} & ResNet-101& 3.20 & 93.99 \\
DG-Net (light) & ResNet-101& 2.22 & 91.99 \\
\midrule
\emph{ours}\\
CoDiNet-32 &ResNet-32& 0.09 & 92.48 \\
CoDiNet-110 & ResNet-110& 0.29 & \textbf{94.47} \\
\bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize The results of the others are the best results reported in their papers. GMACCs refer to billions of multiply-accumulates.}
\end{table}
\subsection{Ablation Study} \label{sec_ablation}
In this part, we discuss the effectiveness of each module in CoDiNet.
First, we perform the ablation studies on consistency and diversity. Then, we discuss the customizable dynamic routing module, which is proposed to strike the balance between computational cost reduction and accuracy.
\begin{table}[t]
\caption{Comparison with State-of-the-Arts on ImageNet}
\label{tab.sota_imagenet}
\vspace{-1em}
\renewcommand\arraystretch{1.3}
\setlength{\tabcolsep}{3.4mm}{
\begin{tabular}{llcc}
\toprule
Methods & Backbones & GMACCs & Acc. (\%) \\ \midrule
\emph{baseline} & & &\\
ResNet-50 & --- & 3.86 & 75.36 \\
ResNet-101 & --- & 7.63 & 76.45 \\
\midrule
\emph{dynamic routing} & & &\\
Conv-AIG 50~\cite{veit2018convolutional} & ResNet-50 & 3.06 & 76.18 \\
Conv-AIG 101 & ResNet-101 & 5.11 & 77.37\\
SkipNet~\cite{wang2018skipnet} & ResNet-101 & 6.70 & 77.40 \\
SkipNet (light) & ResNet-101 & 3.60 & 75.22 \\
LC-Net~\cite{xia2020fully} & ResNet-50 & 2.89 & 74.10 \\
BlockDrop~\cite{wu2018blockdrop} & ResNet-101 & 7.32 & 76.80 \\
DG-Net~\cite{shafiee2019dynamic} & ResNet-101 & 7.05 & 76.80\\
DDI~\cite{wang2020dual} & DenseNet-201 & 3.50 & 76.50 \\ \midrule
\emph{early prediction} & & &\\
MSDN~\cite{Huang2017MultiScaleDN} & DenseNets & 2.30 & 74.24 \\
RA-Net~\cite{yang2020resolution} & DenseNets & 2.40 & 75.10 \\
IamNN~\cite{leroux2018iamnn} & ResNet-101 & 4.00 & 69.50 \\
ACT~\cite{graves2016adaptive} & ResNets & 6.70 & 75.30 \\
SACT~\cite{figurnov2017spatially} & ResNets & 7.20 & 75.80 \\
\midrule
\emph{ours} & & &\\
CoDiNet-50 & ResNet-50& 3.10 & 76.63 \\
CoDiNet-101 & ResNet-101& 5.02 & \textbf{77.85}\\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize The results of the others are the best results reported in their papers. GMACCs refer to billions of multiply-accumulates.}
\end{table}
\subsubsection{Effectiveness of Regularization}
First, we conduct experiments on CIFAR-10 to show the ablation studies on each component based on ResNet-110. As shown in \cref{tab:ablation_study}, the accuracy of the vanilla dynamic routing method is 93.66\% with 113 paths. With consistency regularization, limited paths are utilized, and the performance decreases to 92.88\%. When utilizing the consistency and diversity regularization at the same time, it achieves an accuracy of 94.47\% with 276 paths.
We also provide the numerical improvement on other datasets to verify the proposed consistency and diversity-based optimization as a whole. As shown in \cref{tab.sota_imagenet}, our method achieves a comparable result with 18.4\% computational cost reduction on ImageNet with 1.2\% accuracy improvements.
The results on SVHN are shown in \cref{tab:SVHN_result}. Compared with the result of the vanilla dynamic routing method on SVHN, our method gains 1.13\% improvement with about 5\% extra computational cost, which is about 22\% computation reduction against original ResNet-110.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figures/diversity_path.pdf}
\caption{Visualization of the routing paths distribution under different $m_d$ through t-SNE. $m_d$ is the margin of diversity, which is defined in \cref{eq.diverse}. Figures in row 1, 2 and 3 show routing path distributions for $m_d=0.25, 0.5$, and $0.75$ respectively. Images are of 5 classes (airplane, automobile, bird, cat, and deer) from the CIFAR-10 test set.}
\label{fig.diversity}
\end{figure*}
\begin{figure}[t]
\centering
\subfigure[CIFAR-10]{
\centering
\includegraphics[width=.485\linewidth]{minor/cifar10_sota.pdf}
\label{sfg.flop_compare_a}
}
\hspace{-1.4em}
\subfigure[ImageNet]{
\centering
\includegraphics[width=.485\linewidth]{minor/imagenet_sota.pdf}
\label{sfg.flop_compare_b}
}
\caption{The accuracy against computational cost (GMACCs) of CoDiNet comparing to related methods on CIFAR-10 and ImageNet.}
\label{fig.results_comparison}
\end{figure}
\subsubsection{Effect of Customizable Dynamic Routing}
Another benefit of our method is that we can optimize the computational cost explicitly. As we discussed in \cref{sec.cost}, the proposed $\mathcal{L}_{cost}$ can balance the trade-off between accuracy and cost. And we only need to tune the weight of $\mathcal{L}_{cost}$, i.e., $\gamma$ in \cref{eq.total}, to obtain a desired model. \cref{tab.Detailed_Data} shows the trade-off between classification accuracy, GMACCs, and the average inference time on the CIFAR-10 dataset. With the increasing of the computational cost, the accuracy tends to be upward. In the extreme case, our method achieves even only 0.10 GMACCs with a comparable accuracy, which can meet the requirements on low power platforms.
\subsection{Performance Comparison}
In this part, we compare the results of CoDiNet with related methods. First, we compare CoDiNet with the results of ResNets on CIFAR-10/100. Next, we compare CoDiNet with the state-of-the-art dynamic routing networks, early prediction models\footnote{The results of ACT and SACT are quoted from~\cite{wu2018blockdrop} because ACT and SACT did not report corresponding results.}, and related compression methods on CIFAR-10 and ImageNet.
\subsubsection{Comparison with ResNets}
We make a comparison between CoDiNet and ResNets, w.r.t., accuracy and GMACCs. As shown in \mbox{\cref{tab:result_10_100}}, our method achieves higher accuracy with less cost in all experimental settings. In particular, compared with ResNet-110 on the CIFAR-10 dataset, our method needs 60\% cost (0.29 GMACCs) compared to the original network (0.51 GMACCs), and achieves 0.87\% improvement on accuracy. Similarly, on CIFAR-100, our method achieves significant improvement compared with ResNets and vanilla dynamic routing networks (without $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$). It achieves an accuracy of 72.9\% with 0.24 GMACCs. Besides, the cost reduction on deep networks is much larger than cost reduction on shallow networks. It shows that deep networks are more redundant than shallow networks.
\subsubsection{Comparison on CIFAR-10}
\label{sssec:sotas}
We compare CoDiNet with other state-of-the-art dynamic routing methods, early prediction networks, and related compression methods.
As shown in \cref{tab.sota_result}, we compare with the following methods: BlockDrop~\cite{wu2018blockdrop}, SkipNet~\cite{wang2018skipnet}, Conv-AIG~\cite{veit2018convolutional}, ACT~\cite{graves2016adaptive}, SACT~\cite{figurnov2017spatially}, CGAP~\cite{du2019efficient}, DDI~\cite{wang2020dual}, Iamm~\cite{leroux2018iamnn}, DG-Net~\cite{shafiee2019dynamic}. Following \cite{wu2018blockdrop}, PFEC \cite{li2016pruning} and LCCL \cite{dong2017more} are used for comparison. BlockDrop and SkipNet are prevalent methods, applying reinforcement learning and LSTM respectively to implement the dynamic routing. BlockDrop achieves an accuracy of $93.6\%$ with 0.17 GMACCs on CIFAR-10 with ResNet-110. Conversely, SkipNet focuses more on computational cost reduction, obtaining an accuracy of $92.38\%$ with 0.09 GMACCs. Conv-AIG achieves an accuracy of 94.24\% with about 0.41 GMACCs.
As shown in \cref{sfg.flop_compare_a}, the CoDiNet outperforms other methods in most cases with a comparable computational cost. Our method achieves an accuracy of 94.47\% with only 0.29 GMACCs. More importantly, our method does not conflict with compression methods, and it can be used along with compression methods for better performance.
\subsubsection{Comparison on ImageNet}
We compare CoDiNet with state-of-the-art methods on ImageNet as shown in \cref{tab.sota_imagenet}, and the efficiency-accuracy trade-off in \cref{sfg.flop_compare_b}. Among these methods, Conv-AIG~\cite{veit2018convolutional} reports results based on ResNet-50 and ResNet-101, which are 76.18\% and 77.37\% with 3.06 and 5.11 GMACCs respectively. SkipNet~\cite{wang2018skipnet} achieves an accuracy of 75.22\% with 3.6 GMACCs. Besides, RA-Net~\cite{yang2020resolution} is an early prediction method that processes different samples in different resolutions, which achieves an accuracy of 75.10\% with 2.40 GMACCs. In comparison, CoDiNet outperforms these methods, which achieves an accuracy of 76.63\% with 3.10 GMACCs based on ResNet-50 and an accuracy of 77.85\% with 5.02 GMACCs based on ResNet-101.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{minor/path_KL1.pdf}
\caption{KL divergence of the predictions between original and its augmentation on the CIFAR-10 test set. Green bars stands for the vanilla dynamic routing ResNet-110. Blue bars stands for the dynamic ResNet-110 with $\mathcal{L}_{con}$. Best viewed in color.}
\label{fig.KD}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{minor/Focus1_2.png}
\caption{The visualization of the routing paths of multi-augmentations for samples. We randomly visualize 150 groups of augmentations from the CIFAR-10 test set. Dots in the same color stand for a group of augmentations from the same sample. Best viewed in color.}
\label{fig.std_vis}
\end{figure}
\subsection{Qualitative Analysis}
In this part, we conduct experiments to qualitatively analyze our proposals. First, we analyze the effect of the consistency regularization and the diversity regularization. Next, we visualize the distribution of relaxed routing paths. Finally, we show the images sharing the same routing paths.
\begin{table}[t]
\caption{Analysis of the Margin of Diversity}
\renewcommand\arraystretch{1.3}
\label{tab:path_number}
\vspace{-1.0em}
\setlength{\tabcolsep}{6.8mm}{
\begin{tabular}{c|ccc}
\toprule
$m_d$ &\#Path & GMACCs & Acc. (\%) \\ \midrule
0.25 &196 & 0.22 & 93.12 \\
0.50 &276 & 0.29 & \textbf{94.47} \\
0.75 &396 & 0.24 & 92.75 \\
1.00 &800 & 0.32 & 92.46 \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize The number of activated routing paths and accuracy under different margins of diversity with ResNet-110 on CIFAR-10. $m_d$ is the margin of diversity defined in \cref{eq.diverse}, \#Path is the number of utilized routing paths and GMACCs refer to billions of multiply-accumulates.}
\end{table}
\subsubsection{Analysis of Consistency}
\label{sssec:ana_consistency}
To illustrate the effectiveness of the consistency regularization, we adopt KL divergence as an indicator to measure the difference of paths between original test set and augmented test set. That is, a smaller KL divergence indicates a better consistency. Compared with the vanilla dynamic routing on the CIFAR-10 test set, we found that the consistency regularization can significantly enhance the consistency between routing paths of original and augmented images. It also improves the performance of the augmented test set. As shown in \cref{fig.KD}, the KL divergence between original and augmented images decreases considerably with $\mathcal{L}_{con}$.
Additionally, we design a qualitative experiment to show the effect of the consistency regularization. We adopt various augmentation methods including random cropping, horizontal flipping, vertical flipping, and rotation on the 10,000 images of the CIFAR-10 test set. For each original image, we compare its routing path with that of its augmentation
under two models: CoDiNet and vanilla dynamic routing. As a result, 5,342 out of 10,000 image pairs have consistent routing paths with CoDiNet, i.e., the original image and its augmented image have the same routing path. With vanilla dynamic routing, only 2,265 images have consistent paths with their augmentation.
\subsubsection{Analysis of Diversity}
\label{sssec:ana_diversity}
The number of distinct routing paths at inference time under different settings of $m_d$ is shown in \cref{tab:path_number}. With a larger margin, more routing paths will be obtained. When $m_d$ is 0.5, our method achieves the best performance on CIFAR-10 based on ResNet-110. When $m_d$ is larger than 0.5, the performance drops. The reason for that might lie in too dispersed routing paths resulting in under-fitting.
Next, we visualize the routing path distribution on the CIFAR-10 test set to demonstrate the tendency of different margins of diversity $m_d$. In \cref{fig.diversity}, paths are dispersed significantly with an increase of $m_d$. When $m_d$ is 0.25, paths cluster into two groups. When $m_d$ is 0.75, the scale of coordinate is similar as $m_d = 0.5$, but the distribution is more disperse.
\begin{figure*}[t]
\centering
\subfigure[Path distribution without $\mathcal{L}_{con}$ or $\mathcal{L}_{div}$]{
\centering
\includegraphics[width=0.48\linewidth, height=2.3in]{minor/disconsist.pdf}
\label{sfg.disconsist}
}
\subfigure[Path distribution with $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$]{
\centering
\includegraphics[width=0.48\linewidth, height=2.3in]{minor/consist.pdf}
\label{sfg.consist}
}
\caption{Visualization of continuous routing paths by t-SNE on the CIFAR-10 test set.
The color of each dot is the mathematical expectation numbers of to-be-run blocks. Red refers to more to-be-run blocks, while blue to fewer. Images in the green rectangle are augmentations from the same image. Images in the red rectangle are from the same category. Images in the blue rectangle are from different categories. Best viewed in color.}
\label{fig.similar}
\end{figure*}
\subsubsection{Illustration of Routing Paths for Similar Samples}
In this section, we obtain groups of similar images by applying self-supervised augmentation methods (random cropping, horizontal flipping, vertical flipping, and rotation) on the CIFAR-10 test set ten times and then visualize the distribution of routing paths for self-supervised similar images to show the routing paths of ``similar samples'' clustering together by directly visualizing the distribution of routing paths for self-supervised similar images. We mark the augmented images from the same raw images with the same color. As shown in \cref{fig.std_vis}, points with the same color, i.e., augmentations of the same sample, tend to cluster together. As a result, the routing paths of self-supervised similar samples tend to cluster together in our method.
\subsubsection{Correlation between Samples Similarity and Routing Paths Similarity}
In this section, we calculate the PCC (Pearson correlation coefficient) between sample feature similarity and sample routing path similarity. A higher correlation coefficient value indicates the sample feature similarity is more positively correlated to the path similarity, therefore, the routing paths of similar images are closer in the routing space.
Specifically, we use the Cosine Similarity on every pair of routing paths and sample features as the routing paths similarity and sample similarity. To better represent samples, the sample features are extracted by a third-party unsupervised model (an ImageNet pre-trained MoCo~\cite{moco}). As shown in \cref{tab.similarity_method}, the Pearson correlation coefficient for our CoDiNet is 0.581, while the one for the vanilla dynamic routing is 0.024. Thus, our method is about 24 times larger than the vanilla one. Besides, we plot the correlation diagrams for different sample pairs in \cref{fig.similar_core}. Clearly, our method is more likely to encourage the consistency between routing path similarity and sample similarity.
\subsubsection{Visualization of Relaxed Routing Paths}
As shown in \cref{fig.similar}, we visualize the relaxed routing paths of the vanilla dynamic routing network and CoDiNet by t-SNE. Different colors correspond to different mathematical expectations of the numbers of to-be-run blocks. Red refers to more to-be-run blocks, while blue refers to less to-be-run block. The path distribution of the method without $\mathcal{L}_{con}$ or $\mathcal{L}_{div}$, is shown in \cref{sfg.disconsist}, where the paths gather in a small space around the center. In comparison, the path distribution of CoDiNet are regularly distributed and scattered throughout a much larger space as shown in \cref{sfg.consist}.
\begin{table}[t]
\caption{Illustration of the PCC (Pearon correlation coefficient) between sample similarity and the routing path similarity.}
\vspace{-1.0em}
\label{tab.similarity_method}
\setlength{\tabcolsep}{4.4mm}{
\begin{tabular}{lcc}
\toprule
Methods & Vanilla Dynamic Routing & Our CoDiNet \\ \midrule
PCC & 0.024 & 0.581 \\ \bottomrule
\end{tabular}}
\medskip
\emph{\footnotesize PCC refers to the Pearon correlation coefficient. The PCC value ranges from -1 to 1. A PCC value of 0 implies that there is no linear correlation between the similarities. Experiments is on CIFAR-10.}
\end{table}
\begin{figure}[t]
\centering
\subfigure[Vanilla dynamic routing]{
\includegraphics[width=0.485\linewidth]{minor/van_alpha.pdf}
\label{sfg.NMI}
}
\hspace{-1.4em}
\subfigure[Our CoDiNet]{
\includegraphics[width=0.485\linewidth]{minor/our_alpha.pdf}
\label{sfg.EUR}
}
\caption{The correlation between the similarity of sample features and the similarity of sample routing paths with our CoDiNet and the vanilla dynamic routing. We use Cosine Similarity to measure the similarity. It is worth noting that the closer the point to the dotted line, the more positively correlated the similarity of samples and the similarity of paths.}
\label{fig.similar_core}
\end{figure}
Moreover, we provide three groups of samples to present the routing paths of the self-supervised similar images cluster together no matter they belong to the same category or not. As shown in~\cref{fig.similar}, we provide three groups of images and mark out their routing paths. Firstly, images in the green rectangle are augmentations from the same image. Then, images in the red rectangle are from the same category. Next, images in the blue rectangle are from different categories. As shown in~\cref{sfg.disconsist}, the routing paths of images in each rectangle are scattered among the whole distribution without $\mathcal{L}_{con}$ or $\mathcal{L}_{div}$. In comparison, with $\mathcal{L}_{con}$ and $\mathcal{L}_{div}$, the routing paths of images in each group are respectively cluster together as shown in~\cref{sfg.consist}. Therefore, the proposed consistency regularization term can effectively make the routing paths of similar images clustering together.
\section{Related Works}\label{sec:related}
In this section, we revisit relevant methods and divide them into three categories: dynamic routing networks, early prediction networks, and model compression methods. Dynamic routing and early prediction are two typical approaches to dynamic inference.
The former focus on skipping unnecessary units at inference time, while the latter is characterized by multiple exits. Model compression methods are also popular for cost reduction with static inference.
\subsection{Dynamic Routing Networks}
Layer dropping has long been used as a regularization technique in neural networks, e.g., DropConnection~\cite{wan2013regularization} and Dropout~\cite{JMLR:v15:srivastava14a}. Veit {et al}.~\cite{EnsemblesShallow} found that only short paths of deep residual networks are needed. Motivated by this, dynamic routing networks have emerged as a promising technique to skip blocks or layers at inference time for acceleration~\cite{veit2018convolutional, wang2018skipnet, wu2018blockdrop, almahairi2016dynamic, su2020dynamic}. Specifically, ConvNet-AIG~\cite{veit2018convolutional} proposed a convolutional network that adaptively defines its inference graph conditioned on the input images. it proposes a router to make the execution decision for each convolutional block. SkipNet~\cite{wang2018skipnet} introduced a method with LSTM gate-ways to determine whether the current block would be skipped or not. Besides, BlockDrop~\cite{wu2018blockdrop} adopted an extra policy network to sample routing paths from the whole routing space to speed up ResNets' inference. Slimmable Nets~\cite{yu2018slimmable} intended to train a model to support multiple widths to fit different computational constraints. Recursive network~\cite{guo2019dynamic} proposes to execute a convolutional layer multiple times. RNR~\cite{rao2018runtime} models the dynamic process as a Markov decision process and uses reinforcement learning for training.
Spatial dynamic convolutions for fast inference were proposed in~\cite{verelst2020dynamic, yu2019universally, sun2020computation, xie2020spatially}. Multi-scale networks were introduced in~\cite{huang2018multi, yang2020resolution}. They learn easy samples at low resolutions, while hard samples at high resolutions. Channel-based dynamic routing methods~\cite{su2020dynamic, jordao2020discriminative} were introduced as well.
Recently, various dynamic methods with different kinds of selection have been proposed.
Multi-kernel methods~\cite{Chen_2020_CVPR, chen2020dynamic} select different CNN kernels for better performance. Recursive network~\cite{guo2019dynamic} are introduced to reuse the networks.
At training time, dynamic routing models are prone to early convergence to suboptimal states. To deal with the issue, SkipNet~\mbox{\cite{wang2018skipnet}} uses multiple training stages, Blockdrop~\mbox{\cite{wu2018blockdrop}} uses curriculum learning, dynamic conv~\mbox{\cite{verelst2020dynamic}} uses annealing, sparsity network~\mbox{\cite{sun2020computation}} uses a non-conditional pre-training.
In comparison, we focus on how to learn proper paths. To this end, we model the relation between samples and their routing paths explicitly and optimize the routing paths directly, achieving a more stable dynamic routing model.
\subsection{Early Prediction Networks}
While a dynamic routing network has only one exit, an early prediction network is characterized by multiple exits.
In an early prediction network, the network exits once the criterion for a certain sample is satisfied.
Traditional methods~\cite{reyzin2011boosting, hu2014efficient} applied heuristic and greedy algorithms to reduce the executed layers.
BranchyNet~\cite{teerapittayanon2016branchynet} proposed a multiple-branch framework by attaching fully connected layers to intermediate layers of the backbone.
ACT~\cite{graves2016adaptive} proposed a halting unit for a recurrent neural network (RNN) to realize early prediction. Following ACT, SACT~\cite{figurnov2017spatially} proposed a CNN-based early prediction network, adopting a stopping unit for each point on feature maps.
Since then, early prediction frameworks have been widely used in classification for efficient inference.
Considering multi-scale inputs, MSDN~\cite{Huang2017MultiScaleDN} introduced early-exit branches based on DenseNet~\cite{huang2017densely}. According to the allowed time budget, McIntosh et al.~\cite{mcintosh2018recurrent} proposed an RNN architecture to dynamically determine the exit. Li {et al}.~\cite{li2019improved} proposed a self-distillation mechanism to supervise inter-layer outputs with deeper layers. Instead of bypassing residual units, DCP~\cite{gao2018dynamic} generated decisions to save the computational cost for channels. Hydranets~\cite{Hydranets} proposed to replace the last residual block with a Mixture-of-Experts layer. Recently, methods have been adopted to other applications, such as action recognition~\cite{hussein2020timegate, meng2020ar} and object detection~\cite{zhang2019slimyolov3}.
Our method belongs to dynamic routing networks. Thus, our method does not have multiple exits as early prediction networks do. We also compare our method with early prediction networks in \cref{sssec:sotas}.
\subsection{Model Compression Methods}
Compression methods are proposed for high-performance models on platforms with limited computational resources. Knowledge distillation~\cite{hinton2015distilling, chen2017learning, chen2018distilling, yu2018nisp}, low-rank factorization~\cite{ioannou2015training, tai2015convolutional, jaderberg2014speeding}, and quantization~\cite{han2015deep, wu2016quantized, polino2018model} have been widely used to compress the structures and to prune the parameters of neural networks. Besides, recent researches tend to prune unimportant filters or features~\cite{li2016pruning, he2017channel, luo2017thinet, wen2016learning, huang2018condensenet} to compress or speed-up the model. They identify ineffective channels or layers by examining the magnitude of the weight or activation. The relatively ineffective channels and layers are pruned from the model. Then the pruned model is finetuned to mitigate the accuracy loss. With the iteration of pruning unnecessary parts and then finetuning the model, computational cost and model size can reduce effectively. Perforated CNN~\cite{Perforated} speeds up the inference by skipping the computations at fixed spatial locations. In addition, Neural Architecture Search provides other technique plans achieving low-cost models including MnasNet \cite{tan2019mnasnet}, ProxylessNAS \cite{cai2018proxylessnas}, EfficientNet \cite{tan2019efficientnet}, and FbNet \cite{wu2019fbnet}. In contrast to this line of work where the same amount of computation is applied to all samples, we focus on efficient inference by dynamically choosing a series of blocks to be executed conditioned on the input.
\section{Discussion on Routers}
In this section, we compare CoDiNet with two implementations of Gumbel-Softmax, and different router structures to analyze the design of our router. The discussion focuses on the following questions. First, which variant of Gumbel-Softmax is suitable to utilize in our method? Next, what are the advantages of the router used in our method compared to other kinds of routers?
\begin{figure}[h]
\centering
\subfigure[CIFAR-10]{
\centering
\includegraphics[width=.487\linewidth]{minor/gumbel_cifar10.pdf}
\label{sfg.gumbel_a}
}
\hspace{-1.6em}
\subfigure[CIFAR-100]{
\centering
\includegraphics[width=.487\linewidth]{minor/gumbel_cifar100.pdf}
\label{sfg.gumbel_b}
}
\caption{The accuracy against computational cost of the re-parameterized and the straight-through Gumbel-Softmax variants on CIFAR-10/100. Results are based on ResNet-110. Best viewed in color.}
\label{fig.Gumbel_Softmax}
\end{figure}
\subsection{Effectiveness of Gumbel-Softmax}\label{gumbel_softmax}
To train the dynamic routing network end-to-end, relaxation methods are employed because the binary routing paths are not differentiable. As discussed in \cref{sec.relaxation}, we adopt Gumbel-Softmax for in our method. In this section, we compare the two variants of Gumbel-Softmax, i.e., the re-parameterized variant and the straight-through variant. As shown in \cref{fig.Gumbel_Softmax}, the re-parameterized variant performs better than the straight-through variant in most cases on CIFAR-10 and CIFAR-100. Weighing the pros and cons, we take the re-parameterized variant Gumbel-Softmax in our method.
\begin{figure}[h]
\centering
\subfigure[CIFAR-10]{
\centering
\includegraphics[width=.487\linewidth]{minor/CIFAR10_router.pdf}
\label{sfg.router_a}
}
\hspace{-1.6em}
\subfigure[CIFAR-100]{
\centering
\includegraphics[width=.487\linewidth]{minor/CIFAR100_router.pdf}
\label{sfg.router_b}
}
\caption{The accuracy against computational cost of the CNN router, the RNN router, and the FC router on CIFAR-10/100. Results are based on ResNet-110. Best viewed in color.}
\label{fig.routers_compare}
\end{figure}
\subsection{Advantages of Our Router}\label{model_compare}
Routers are key components in a dynamic routing network, which make execution decision for blocks. How to design a lightweight yet effective router has always been the focus in dynamic routing.
In this section, we discuss different types of routers: the CNN router, the RNN router, and the FC router. As shown in~\cref{fig.routers_compare}, we show the accuracy against computational cost, when a network equipped with different routers.
Specifically, a CNN router is composed of a $3\times 3$ convolutional layer followed by a global average pooling layer and a linear layer to output $1\times2$ vector. A RNN router is composed of a global average pooling, a shared linear layer, and a shared LSTM layer with a hidden unit size of 10. For the FC router, it uses two linear layers, after a global average pooling. Please refer to~\cite{veit2018convolutional, wang2018skipnet} for more details.
As a result, the FC router achieves the highest accuracy under multiple computational settings, comparing with the RNN router and the CNN router.
| 30ae25b40dce579e554349b0b3ab4584b6031a1b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{GEMM vs SpDM}\label{subsection:gemm_vs_spdm}
A naive method to solve SpDM is directly using high-performance GEMM libraries, such as cuBLAS, if the sparsity of matrix A is not very high. High-performance GEMM libraries focus on improving the computational throughput of GPUs. For example, cuBLAS can achieve over 90\% of theoretical peak computational throughput of Nvidia Titan X Pascal. Here we briefly analyze the performance of cuBLAS using the roofline model. Note that we only discuss the square matrix multiplication for simplicity. Generally, for calculation of $\ma{C}=\ma{A} \times \ma{B}$, we need $O(n^3)$ multiplications and additions. Meanwhile, accessing the elements of $\ma{A}$ and $\ma{B}$ is also time consuming in modern GPU architectures. Ideally we need $2 \times n^2$ memory access for $\ma{A}$ and $\ma{B}$ at least. However, due to the limited cache size and registers, practical GEMM implementations have much more main memory access. The previous high-performance GEMM works struggle to reduce the overhead brought by slow memory access and promote the compute core utilization. The optimization principles include [citation].
However, consider SpDM, in which $\ma{A}$ is of high sparsity. Simply using those high-performance GEMM libraries may bring many useless calculations. Only $1-s$ of $n^3$ calculations have contributions to $\ma{C}$. That is, even high-performance GEMM libraries can achieve nearly peak computational performance of the devices, the effective percentage of those flops is as low as $1-s$.
\begin{equation}
T = (\alpha n^3+\beta)\times (1-s)
\label{eq:time_sparse}
\end{equation}
\section{Introduction}
Sparse-dense matrix-matrix multiplication (SpDM) has many application areas. It is not only exploited in traditional research fields (e.g., graph analytics \cite{tiskin2001all}, biology \cite{vazquez2010matrix}), but becoming a potential faster implementation for sparse deep learning \cite{liu2015sparse}\cite{shi2017speeding}\cite{wen2017learning}\cite{sun2017meprop}\cite{shi2019distributed}. However, it requires very high sparsity of the model to achieve accelerated speed compared to the original dense implementations \cite{narang2017exploring}.
Dense matrix multiplication, i.e., $C=A\times B$ or general purpose matrix multiplication (GEMM) has been well studied on GPUs to achieve high efficiency \cite{volkov2008benchmarking}\cite{chu2009practical}\cite{nath2010improved}\cite{matsumoto2011multi}\cite{kurzak2012autotuning}\cite{lai2013performance}\cite{abdelfattah2016performance}\cite{zhang2017understanding}\cite{yan2020demystifying}\cite{liu2018g}. However, multiplication of a sparse matrix to a dense matrix (SpDM), in which the sparse matrix is stored with memory-saving formats like compressed row storage (CRS) \cite{dongarra2000compressed}, is understudied, and it easily loses efficiency on modern GPUs. For example, the time cost of calculating the multiplication of a $8000\times 8000$ sparse matrix with sparsity of $0.9$ (i.e., $90\%$ of elements are zeros) to a dense matrix with single precision requires $780ms$ by using cuSPARSE on an Nvidia Tesla P100 GPU, while the corresponding dense algorithm by cuBLAS only requires $121ms$.\footnote{Both cuSPARSE and cuBLAS are from the library of CUDA-8.0.} In other words, though the sparse matrix can reduce the number of multiplication and accumulation operations (MACs) by $90\%$ (since a zero element times any numbers produces zeros that has no contribution to the final results, so such operations can be avoided.), the highly optimized cuBLAS is about $7\times$ faster than cuSPARSE in the above example. For a much higher sparsity of $0.995$, cuSPARSE can be about $50\%$ faster than cuBLAS at the dimension of $8000\times 8000$ matrices on the P100 GPU. High sparsity requirement on SpDM makes it difficult to be deployed as the efficient implementation of matrix multiplication because of the inefficient algorithm design of the SpDM algorithm in cuSPARSE. In practical problems, on one hand, if the sparsity is not high enough, doing SpDM could result in very low efficiency, while using the dense form could get results faster if there is enough memory; on the other hand, if the sparsity is very high, using the dense form not only leads to low efficiency, but it also wastes memory. From our empirical studies of cuSPARSE and cuBLAS, the sparse algorithm of cuSPARSE requires the matrix sparsity to be larger than $0.99$ to outperform the dense counterpart of cuBLAS. One of our key observations of cuSPARSE is that it has many slow memory access that easily leaves the computational resources (i.e., cores) stale in its SpDM APIs. To this end, we would like to design an efficient SpDM algorithm to better utilize the GPU computational resources.
Only a small number of research works focus on high-performance SpDM algorithms for modern GPUs. The most relevant work is \cite{ortega2013fastspmm}, \cite{yang2018design} and \cite{parger2020speck}\cite{jiang2020novel}. On one hand, Ortega et al. \cite{ortega2013fastspmm} try to better optimize the GPU memory access pattern (i.e., coalesced memory access) to achieve higher efficiency. On the other hand, besides the optimization of coalesced memory access, Yang et al. \cite{yang2018design} use the principles of row split \cite{bell2009implementing} and merge path \cite{merrill2016merge} in sparse matrix-dense vector multiplication (SpMV) to design more efficient algorithms for SpDM on GPUs. Jiang et al. \cite{jiang2020novel} mainly re-order the row data and Parger et al. \cite{parger2020speck} propose the parameter tuning technique to optimize the performance of SpDM. However, in \cite{yang2018design}, the authors design their algorithms mainly for the cases that the dense matrices are tall-skinny, and it requires a heuristic to choose whether to use merge-based or row split for better performance. In this paper, we not only exploit the GPU algorithm optimization principles (e.g., coalesced memory access), but also revisit the popular roofline performance model \cite{williams2009roofline} on GPUs to analyze how to increase operational intensity, and then we propose an efficient SpDM algorithm. Our contributions are summarized as follows:
\begin{itemize}
\item We design an efficient SpDM algorithm called GCOOSpDM on GPUs with several optimization techniques including coalescing memory access, bank conflict avoidance of the shared memory and high computation-to-memory ratios.
\item We evaluate the proposed algorithm on a large number of sparse matrices including the public dataset and randomly generated matrices, and the experimental results show that GCOOSpDM outperforms cuSPARSE 1.5-8$\times$ faster in a large proportion of matrices on Nvidia GPUs.
\item We conduct instruction-level analysis for the kernels of GCOOSpDM and cuSPARSE, and the profiled results confirm that our proposed algorithm uses much less slow memory access (DRAM and L2 cache) than cuSPARSE.
\item As compared to cuSPARSE, GCOOSpDM decreases the sparsity requirement from $0.99$ to $0.98$ in order to outperform dense implementation of cuBLAS.
\end{itemize}
The rest of the paper is organized as follows. Section \ref{section:preliminaries} gives introductions to the preliminaries related to SpDM and GEMM. We present our proposed algorithm for efficient SpDM in Section \ref{section:algorithm}. The experimental evaluation and analysis are illustrated in Section \ref{section:evaluation}. Section \ref{section:relatedwork} introduces the related work, and finally we conclude this paper in Section \ref{section:conclusion}.
\section{Preliminaries}\label{section:preliminaries}
A multiplication of two matrices $\ma{A}\in \mathbb{R}^{m\times k}$ and $\ma{B}\in \mathbb{R}^{k\times n}$ produces an result matrix $\ma{C}\in \mathbb{R}^{m\times n}$, i.e.,
\begin{equation}
\ma{C}(i,j)=\sum_{l=0}^{l=k-1}\ma{A}(i,l)\times \ma{B}(l, j).
\end{equation}
To simplify the analysis of the algorithms, we assume that the dimensions of $\ma{A}$ and $\ma{B}$ are both $n\times n$. The sparsity $s$ of matrix $\ma{A}$ is defined as the ratio of the number of zero elements over the total number of elements.
\subsection{The roofline model}
The roofline model \cite{williams2009roofline} is commonly used in performance modeling of multi-core/many-core architectures like GPUs \cite{kim2011performance}\cite{zhang2017understanding}\cite{konstantinidis2017quantitative}. The term operational intensity $r$ (operations per byte of DRAM traffic) is defined to predict the performance of kernels. In the model, there is an upper bound of the GPU throughput when $r$ reaches some threshold, which indicates the program is computation-bound. If $r$ is smaller than the threshold, the GPU throughput is a linear function with respect to $r$, which indicates the program is memory-bound. Using cuBLAS GEMM as an example, in Fig. \ref{fig:gpucublas}, we compare the experimental throughput of dense matrix multiplication with the theoretical throughput from roofline model on two different Nvidia GPUs, GTX980 and Titan X.
Though GEMM in cuBLAS has achieved nearly optimal throughput on matrix multiplication, directly applying GEMM for sparse matrices could result in many useless calculations due to the large amount of zeros. The irregular non-zero elements in sparse matrices make the data access from global memory to registers become the bottleneck of matrix multiplication. In other words, each time of data reading from the sparse matrix, only a limited number of computational operations. Therefore, algorithms for SpDM are generally memory-bound, and for such problems, one should design the algorithm to increase $r$ to achieve higher efficiency.
\begin{figure}[!h]
\centering
\includegraphics[width=0.8\linewidth]{cublas_model.pdf}
\caption{The roofline models for theoretical peak throughput and cuBLAS throughput with single-precision on GPUs.}
\label{fig:gpucublas}
\end{figure}
\subsection{GPU memory hierarchy}
From the roofline model, one should improve the memory access efficiency to fully utilize the computational power of GPUs. There are several types of memories in the GPU memory hierarchy. From fast to slow of access speed, it contains registers, the shared memory (or L1 cache), L2 cache and the global memory \cite{volkov2008benchmarking}\cite{mei2017dissecting}\cite{mei2014benchmarking}. The shared memory and global memory are two kinds of memories that can be flexibly manipulated by programming. In general, data that is repeatedly used could be put into the shared memory or registers for better utilization of GPU cores.
\subsection{COO: The coordinate storage format}
Assume that the matrix is a row-major matrix. The coordinate storage format (COO) \cite{bell2009implementing} is a simple storage scheme for sparse matrices. COO uses an array $values$ to store the values of all non zero elements. The coordinate information of each non zero element is sequentially stored in array $rows$ and array $cols$ respectively. Take a real-valued example of a $4\times 4$ sparse matrix as follows:
\[
\ma{A}=
\begin{bmatrix}
7 & 0 & 0 & 8 \\
0 & 10 &0 & 0 \\
9 & 0 & 0 & 0 \\
0 & 0 & 6 & 3
\end{bmatrix},
\]
the COO format of $\ma{A}$ is represented by
\begin{align*}
values&=[7,8,10,9,6,3],\\
rows &=[0,0,1,2,3,3],\\
cols &=[0,3,1,0,2,3].
\end{align*}
\section{Efficient Algorithm Design}\label{section:algorithm}
In this section, we describe the design of our proposed efficient SpDM algorithm on GPUs including the customized storage format for sparse matrices and its conversion from the dense ones. According to the above analysis in operations of SpDM on GPUs, we first design a new sparse format called grouped COO (GCOO), which is convenient for coalesced memory access and is useful to increase the operational intensity $r$. Then we propose an efficient SpDM algorithm by using GCOO.
\subsection{GCOO: Grouped COO storage format}
A similar format of GCOO is the sliced COO (SCOO) format proposed in \cite{dang2012sliced}, with which the authors achieved higher throughput on sparse matrix-vector multiplication (SpMV) on both CPUs and GPUs. In this paper, we bring the idea of SCOO to propose GCOO for matrix multiplication. The sparse matrix is partitioned to $g$ groups according to the number of columns $n$, and each group is stored in the COO format, so we call it GCOO. For an $n\times n$ matrix stored in the GCOO format, there are $g=\lfloor \frac{n+p-1}{p} \rfloor$ groups, and each group contains $p$ columns except the last one who has $n-(g-1)\times p$ columns. If $n$ is divisible by $p$, then the last group also has $p$ columns. In GCOO, each group is stored in the COO format, and COOs from all groups are concatenated into one array. Let group $i$ be stored in the COO format with $rows_i$, $cols_i$ and $values_i$, where $i=0,1,...,g-1$. We have the stored values of GCOO with $rows$, $cols$ and $values$ that are generated from the concatenation of $rows_i$, $cols_i$ and $values_i$ respectively.
\begin{figure}[!h]
\centering
\includegraphics[width=0.4\linewidth]{gcooformat.pdf}
\caption{An example of GCOO. It has 2 groups, and each group contains 2 columns (i.e., $p=2$).}
\label{fig:gcoo}
\end{figure}
An example of grouping in matrix $\ma{A}$ is shown in Fig. \ref{fig:gcoo}. Matrix $\ma{A}$ is divided into to 2 groups. Group $0$ is represented by $rows_0=[0,1,2]$, $cols_0=[0,1,0]$ and $values_0=[7,10,9]$; and group $1$ is represented by $rows_1=[0,3,3]$, $cols_1=[3,2,3]$ and $values_1=[8,6,3]$. Finally, two groups are concatenated into one array with an extra index array $gIdxes$ to indicate which positions are corresponding to related groups. Therefore, the final stored format of GCOO is as follows:
\begin{align*}
values&=[7,10,9,8,6,3],\\
rows &=[0,1,2,0,3,3],\\
cols &=[0,1,0,3,2,3],\\
gIdxes &=[0,3],
\end{align*}
where $gIdxes$ is an auxiliary array to store the group indexes. It is noted that $rows$, $cols$ and $values$ in GCOO are not the same as those of COO since a single group in GCOO is in a COO format. In order to easily access each group's elements, we use an extra auxiliary array, $nnzPerGroup$, to store the number of non-zero elements in each group. In the above example, the values of $nnzPerGroup$ should be:
\begin{align*}
nnzPerGroup &=[3,3].
\end{align*}
In practice, GCOO spends slightly more memory space than COO and CSR, but it provides more convenient access of data with a higher probability. The comparison of memory consumption to store an $n\times n$ matrix with a sparsity of $s$ (note that $nnz=s\times n^2$) is shown in Table \ref{table:memcon}.
\begin{table}[!ht]
\centering
\caption{Memory consumption of different formats.}
\label{table:memcon}
\begin{tabular}{|l|l|}
\hline
Format& Memory complexity \\\hline
\hline
CSR & $2\times nnz+n$ \\\hline
COO & $3\times nnz$ \\\hline
GCOO & $3\times nnz + 2\times \lfloor \frac{n+p-1}{p} \rfloor$\\\hline
\end{tabular}
\end{table}
The main advantage of GCOO is to help reuse the data from slow memories (e.g., global memory and L2 cache). Specifically, if there exist two or more continuous non-zero elements in one group that are in the same row, then the fetched element from the dense matrix $\ma{B}$ can be reused in the register instead of being read from the slow memory again.
\subsection{Matrix conversion to GCOO}
For the cases that the input matrices $\ma{A}$ and $\ma{B}$ are stored in the dense form, there would be an extra overhead in the format conversion to apply the SpDM algorithm. For example, cuSPARSE provides an API ``cusparseSdense2csr'' to convert the dense matrix to the CSR format so that one can apply the SpDM APIs. For our proposed GCOO, we also need to provide an efficient conversion scheme to convert the dense matrix to GCOO. We use two steps to convert the dense matrix to the GCOO storage.
Step 1: Count the number of non-zero elements. To convert a dense form of a matrix to the sparse form, one should first count the number of non-zero elements ($nnz$) of that matrix in order to allocate the memory according to the value of $nnz$. As for GCOO, we have pre-grouped the matrix by pre-defined $p$, so it is straightforward to calculate the non-zero elements in parallel for different groups such that the array $nnzPerGroup$ can also be calculated. Therefore, in this step, $nnz$, $gIdxes$ and $nnzPerGroup$ can be calculated by scanning the original dense matrix.
Step 2: Store the non-zero elements to $rows$, $cols$ and $values$. First, the memories of $rows$, $cols$ and $values$ are allocated according to $nnz$, and then we can read the non-zero elements with their coordinate information and write them to $rows$, $cols$, and $values$ according to the indexes by $nnzPerGroup$ in parallel.
The pseudocode of the matrix conversion on the GPU from the dense form to GCOO is shown in Algorithm \ref{algo:gcooconv}.
\begin{algorithm}[!ht]
\caption{convertToGCOOFormat}\label{algo:gcooconv}
\textbf{Input: } $A, wA, hA, p$\\
\textbf{Output: $values, cols, rows, gIdxes, nnzPerGroup$}
\begin{algorithmic}[1]
\small
\State $nGroup = (hA + p - 1) / p$;
\State Allocate memory for $gIdxes$ and $nnzPerGroup$ according to $nGroup$;
\State Calculate $gIdxes$ and $nnzPerGroup$ and $nnz$ by scanning $\ma{A}$;
\State Allocate memory for $values$, $cols$, and $rows$ according to $nnz$;
\State Set values of $values$, $cols$ and $rows$ by scanning $\ma{A}$;
\end{algorithmic}
\end{algorithm}
\subsection{GCOOSpDM: an efficient SpDM algorithm}
In the proposed algorithm GCOOSpDM, we focus on three factors that have major impact on the performance. 1) Data partition for the CUDA execution context \cite{nvidia2010programming}. 2) The coalesced memory access of global memory on the sparse matrix $\ma{A}$ and the two dense matrices $\ma{B}$ and $\ma{C}$. 3) When exploiting the faster memory on Nvidia GPUs with the shared memory, we guarantee that the access of the shared memory has no bank conflict. 4) After accessing a single element of the sparse matrix $\ma{B}$, we strive to calculate more results for $\ma{C}$, i.e., achieving higher operational intensity, so that we can achieve higher GFLOPS.
\textbf{Data partition of matrices}. In the context of CUDA, a \textit{thread} is the smallest execution unit of instructions. A group of threads forms a \textit{thread block}, which is executed in a stream multiprocessor (SM) of GPU. Multiple \textit{thread blocks} form a \textit{grid}, and some \textit{thread blocks} are executed in parallel on different SMs at one time. Let $b$ denote the size of a thread block. In our algorithm, each thread block calculates $b\times p$ elements of $\ma{C}$ separately, so a resulting $n\times n$ matrix requires $\lceil \frac{n}{b} \rceil\times \lceil \frac{n}{p} \rceil$ thread blocks. All threads in a thread block share a group of sparse data of $\ma{A}$, but each thread reads continuous columns $\ma{B}$ to do the operations of multiplication and addition to the continuous columns of $\ma{C}$. An example of data partition for $b=4, p=2$ and $n=6$ is shown in Fig. \ref{fig:multiplication}. In the grid, it has 6 thread blocks. Each thread block contains $b=4$ threads, and it calculates 8 elements of $\ma{C}$. Each thread calculates $p=2$ elements of $\ma{C}$.
\textbf{Coalesced memory access}. Three matrices including one sparse matrix $\ma{A}$ with the GCOO format and two dense arrays ($\ma{B}$ and $\ma{C}$) are needed to interactive with the global memory. Irregular global memory access would result in performance degradation on modern GPUs, so we should read the input matrices ($\ma{A}$ and $\ma{B}$) and write the output matrix $\ma{C}$ in a coalesced way.
First, we consider the sparse matrix $\ma{A}$ stored with the GCOO format. Since each group in GCOO of $\ma{A}$ is assigned to one thread block, we just need to consider the block level access of one group of GCOO, i.e., a COO format that has $p$ columns. The number of floating point operations is determined by the number of nonzero elements of $\ma{A}$, so we scan COO to find the corresponding columns of $\ma{B}$. Due to the sparse property, COO could not have many elements, which means we can load COO to the shared memory such that all the threads can read the data fast. Therefore, the $b$ threads in one thread block read $b$ elements of COO from the global memory to the shared memory in a coalesced way. After $\ma{A}$ has been put into the shared memory, it is no need to re-read the elements of $\ma{A}$ from the global memory.
Second, the dense matrix of $\ma{B}$ should be read-aware. The matrix $\ma{B}$ only needs to be accessed when a $(col, row, a)$ of COO has been read from the shared memory, so every thread reads the same $(col, row, a)$, the corresponding column of $\ma{B}$ should be same while the rows should be different to keep all the threads busy and work balance. So threads $t_0, t_1, ..., t_{b-1}$ need to read $\ma{B}(row_0, col), \ma{B}(row_1, col), ..., \ma{B}(row_{b-1}, col)$ in the current block respectively. In order to support the coalesced memory read of $\ma{B}$, the row elements should be in the continuous memory. It is easy to do this because we can just transpose $\ma{B}$ or store $\ma{B}$ in a column-major matrix such that the above elements are in the continuous memory.
Finally, for the result matrix $\ma{C}$, we should only write the matrix once with the final result for each thread to achieve higher throughput. As discussed above, thread $t_i$ reads $(col, row, a)$ of $\ma{A}$, and multiplies with the elements indexed by $(row_i, col)$ in $\ma{B}$, so the write position of $\ma{C}$ should be $(row, row_i)$. As a result, $\ma{C}$ should also be column-major or transposed for the coalesced memory writing.
\textbf{None bank conflict access of the shared memory}. The shared memory used in our algorithm is only proper to the sparse matrix of $\ma{A}$ with the COO format (in one thread block). The kernel allocates a fixed size $b$ of shared memory, and the threads in one thread block read $b$ non-zero elements from $\ma{A}$ each time. Since all the threads in one thread block need to read all elements of $\ma{A}$ to calculate the corresponding columns of $\ma{C}$, all threads read the same element of $\ma{A}$. Therefore, the data in the shared memory can be accessed by all threads in a broadcast way \cite{nvidia2010programming}, which would not result in any bank conflict, and the broadcast access of the shared memory requires only a very small number of clock cycles to fetch the data.
\textbf{High computation-to-memory ratio}. Achieving a high operational intensity $r$ is very important to a high throughput. Regarding the multiplication and accumulation of each thread, each thread reads the shared memory of $\ma{A}$ to get $(col, row, a)$ (donated by $a_r$), and then multiplies $\ma{B}(row_i, col)$ (donated by $b_r$) of $\ma{B}$. In such scenario, we have two opportunities to have more calculations with $a_r$ and $b_r$ since they have been loaded into the registers. The first chance is to find other element of $\ma{B}$ to be multiplied with $a_r$, but the other element that can be multiplied with $a_r$ has been assigned to the other block, so this chance cannot be fulfilled. The second one is to find a next element of $\ma{A}$ who has the same column with the previous one while its row is different, i.e., $(col, row_1, a)$. Therefore, we can search the next $a_{r1}$ (since $\ma{A}$ has been loaded in the shared memory, the time cost of searching is low.) to reuse $b_r$. If such an $a_{r1}$ exists, then we can have $b$ times of multiplication and accumulation without an extra global memory (or L2 cache) access, which results in a higher $r$. For a uniform distributed sparse matrix with sparsity of $s$, there could be $(1-s)\times n$ non-zero elements in the same column.
According to the above four criteria, we conclude the GCOOSpDM algorithm with the following three steps.
Step 1. Each thread block iteratively reads the COO values into the shared memory such that all threads in this thread block can read the COO values for their rows. We exactly know the columns that we need to calculate in the current thread block.
Step 2. The $t^{th}$ thread scans the COO items from the shared memory, and the item contains $row$, $col$ and $value$. According to $col$, the thread reads the element $B(t,col)$ of $\ma{B}$, and then performs the multiplication of $value \times B(t, col)$, whose result is added to the local variable $c_{t,col}$. I.e., $c_{t,col}+=value \times B(t, col)$.
Step 3. Since the current group of data is stored as the COO format, for the current element $(row, col, value)$, its next element should have the same $col$ index if that column has more than one element. So we continue scanning the shared memory to check if there are elements that have the same $col$ such that we can reuse the element of $B(t, col)$.
\begin{figure}[!h]
\centering
\includegraphics[width=\linewidth]{multiplication.pdf}
\caption{Partition of matrices. $\ma{A}$ is the sparse matrix, $\ma{B}$ is the dense matrix, and $\ma{C}$ is the result matrix.}
\label{fig:multiplication}
\end{figure}
The visualization of the algorithm executed with the CUDA programming model is shown in Fig. \ref{fig:multiplication}. On the grid level, there are 6 thread blocks, and each thread block calculates the results of sub-matrix with size of $b\times p$ from $b$ rows of $\ma{B}$, and $p$ columns (i.e., one group in GCOO) of $\ma{A}$. On the thread block level, the GCOO data of sparse matrix are loaded into faster memory once (the shared memory) which is shared among all the threads in the thread block. On the thread level, each thread independently takes charge of computing $p$ elements of $\ma{C}$, say the thread scans the shared memory to read $row$, $col$ and $value$, and then reads the values in column $row$ of $\ma{B}$, which are multiplied by $value$ separately, and each result is accumulated to column $col$ of $\ma{C}$. The algorithm of GCOOSpDM is shown in Algorithm \ref{algo:gcoospdm}.
In Algorithm \ref{algo:gcoospdm}, we first (line 1-10) initialize some local variables including the thread level indexes of output and COO for the current thread block. Then we iteratively scan a block of COO in the for-loop of line 11, and at each iteration, a thread block of COO values are loaded into the shared memory (line 12-15). After that each value of COO in the shared memory is read by all the threads in one thread block, and the corresponding value $b$ in $\ma{B}$ is also read to calculate the result (line 21-26). Instead of continuing the above step, we keep the value of $b$ in the register, and scan the shared COO to check whether we can reuse $b$ so that less memory operations are required (line 28-36). By this way, we can achieve higher operational intensity, i.e., $b$ is reused to do more floating point calculations. At the end, the local results of each thread are written back to $\ma{C}$ that is stored in the global memory with corresponding indexes (line 38-39). Note that both reading of matrix $\ma{A}$ and matrix $\ma{B}$ from the global memory is in a coalescent way, the result writing to matrix $\ma{C}$ is also coalescent. In term of access of the shared memory, it broadcast the data to all the threads in a warp with a small number of cycles.
\begin{algorithm}[!ht]
\caption{GCOOSpDM}\label{algo:gcoospdm}
\textbf{Input: } $values, cols, rows, gIdxes, nnzPerGroup, wA, hA, \\B, wB, hB, C$\\
\textbf{Output: $C$}
\begin{algorithmic}[1]
\small
\State $Cj = blockIdx.y*b+threadIdx.x$;
\State $Ci0 = blockIdx.x*p$;
\State Initial local temporary results $c[0...p]$;
\State Set number of non-zero elements of current group: $nnz$;
\State // Set the current group of COO
\State $vals=values+gIdxes[blockIdx.x]$;
\State $cols=cols+gIdxes[blockIdx.x]$;
\State $rows=rows+gIdxes[blockIdx.x]$;
\State $iter=(b+nnz-1)/b$;
\State $extra = nnz \& (b - 1)$;
\For{$i=0 \rightarrow iter$}
\State $cooOffset=i*b$;
\State $sVals[threadIdx.x]=vals[cooOffset]$;
\State $sCols[threadIdx.x]=cols[cooOffset]$;
\State $sRows[threadIdx.x]=rows[cooOffset]$;
\State $cnnz=\text{max}(extra, b)$;
\State $\_\_$syncthreads();
\If{$Cj<wB$} // Not exceed the boundary
\State $k=1$;
\For{$j=0\rightarrow cnnz, step=k$}
\State $col = sCols[j]$;
\State $row = sRows[j]$;
\State $av=sVals[j]$;
\State $bv=B[col*wB+Cj]$; // Registered.
\State $outIdx=row\&(p-1)$;
\State $c[outIdx]+=av*bv$;
\State $k=1$;
\While{$j+k<cnnz$} // Search $A$ to reuse $bv$
\State $newCol=sCols[j+k]$;
\If{$newCol \neq col$}
\State break;
\EndIf
\State $av=sVals[k+j]$;
\State $row = sRows[k+j]$;
\State $outIdx=row\&(CPG-1)$;
\State $c[outIdx]+=av*bv$;
\State $k+=1$;
\EndWhile
\EndFor
\EndIf
\State $\_\_$syncthreads();
\EndFor
\For{$i=0 \rightarrow p$} // Write results to the global memory
\State $C[Cj+(Ci0+i)*wB]=c[i]$;
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Evaluation and Analysis}\label{section:evaluation}
To show the effectiveness of our proposed algorithm, we do varies of experiments across three Nvidia GPU cards (i.e., GTX 980, GTX Titan X Pascal and Tesla P100) using two kinds of data. The first one is the public sparse matrix dataset \cite{davis2011university} which has different patterns of matrices, and the second one is randomly generated matrices whose zero-valued elements have a uniform distribution.\footnote{Codes of GCOOSpDM and scripts of performance evaluation can be found in \url{https://github.com/hclhkbu/gcoospdm}. And the raw data of our experimental results can be found in: \url{https://github.com/hclhkbu/gcoospdm/tree/master/results}.} The characteristics of tested GPUs are shown in Table \ref{table:gpus}. And the software installed is CUDA-8.0.
\begin{table}[!ht]
\centering
\caption{Characteristics of tested GPUs.}
\label{table:gpus}
\begin{tabular}{|l|l|l|l|}
\hline
Model& GTX980 &TitanX & P100 \\\hline
\hline
SMs $\times$ cores per SM & 16$\times$128 & 28$\times$128 & 56$\times$64 \\\hline
Peak TFLOPS &4.981 & 10.97 & 9.5 \\\hline
Memory Bandwidth (GB/s) &224 & 433 & 732 \\\hline
\end{tabular}
\end{table}
\subsection{Results on public sparse matrices}
We use the public sparse matrices in \cite{davis2011university}. Since we only consider the schemes of square matrices, we pick up all the square matrices in the dataset to evaluate the performances of GCOOSpDM and cuSPARSE. The chosen dataset contains 2694 matrices, whose sparsity is in the range of $[0.98, 0.999999]$, and their dimensions are in the range of $[64, 36720]$. The performance comparison between GCOOSpDM and cuSPARSE is shown in Fig. \ref{fig:realdata}, where $T_{algorithm}$ is used to denote the execution time of $algorithm$. We first compare the overall performance of our algorithm with cuSPARSE on the 2694 matrices, and we then choose 14 types of matrices from varies of applications to compare the performance of the algorithms.
\textbf{Overall performance}. In the 2694 tested matrices, there are about $78\%$ matrices that GCOOSpDM outperforms cuSPARSE on the P100 GPU, and there are more than $90\%$ matrices that GCOOSpDM achieves better performance than cuSPARSE on both GTX980 and TitanX. The average speedups are $1.66\times$, $1.7\times$ and $1.68\times$ on GTX980, TitanX and P100 respectively. Moreover, the maximum speedups of GCOOSpDM are $4.5\times$, $6.3\times$ and $4.2\times$ on GTX980, TitanX and P100 GPUs respectively. By contrast, on the $22\%$ matrices that cuSPARSE is better than GCOOSpDM on the P100 GPU, cuSPARSE only outperforms GCOOSpDM about $1.2\times$ on average. On GTX 980 and Titan X GPUs, there are about $10\%$ cuSPARSE outperforming GCOOSpDM about $1.14\times$. cuSPARSE performs better on the P100 GPU than GTX 980 and TitanX GPUs mainly because the P100 GPU has a much higher memory bandwidth than the other two GPUs as shown in Table \ref{table:gpus}.
\begin{figure}[!h]
\centering
\subfigure[GTX 980]
{
\includegraphics[width=0.32\linewidth]{real980_frequency.pdf}
}\hspace{-3.5mm}
\subfigure[Titan X Pascal]
{
\includegraphics[width=0.32\linewidth]{realtitanx_frequency.pdf}
}\hspace{-3.5mm}
\subfigure[Tesla P100]
{
\includegraphics[width=0.32\linewidth]{realp100_frequency.pdf}
}
\caption{The performance comparison with the frequency of the time ratio between cuSPARSE and GCOOSpDM with the public dataset on three GPUs. The last value (i.e., 2.0+) of x-axis means that $T_{cuSPARSE}/T_{GCOOSpDM} \ge 2.0$.}
\label{fig:realdata}
\end{figure}
\begin{table}[!ht]
\centering
\caption{Details of selected sparse matrices.}
\label{table:selectedmatrx}
\begin{tabular}{|l|l|l|l|}
\hline
Matrix & $n$ & Sparsity & Related Problem\\\hline
\hline
nemeth11&9506&2.31e-03&Quantum Chemistry \\\hline
human\_gene1&22283&2.49e-02&Undirected Weighted Graph\\\hline
Lederberg&8843&5.32e-04&Directed Multigraph\\\hline
m3plates&11107&5.38e-05&Acoustics \\\hline
aug3dcqp&35543&6.16e-05&2D/3D \\\hline
Trefethen\_20000b&19999&7.18e-04&Combinatorial \\\hline
ex37&3565&5.32e-03&Computational Fluid\\\hline
g7jac020sc&5850&1.33e-03&Economic \\\hline
LF10000&19998&1.50e-04&Model Reduction \\\hline
epb2&25228&2.75e-04&Thermal \\\hline
plbuckle&1282&9.71e-03&Structural \\\hline
wang3&26064&2.61e-04&Semiconductor Device \\\hline
fpga\_dcop\_01&1220&3.96e-03&Circuit Simulation \\\hline
viscoplastic2\_C\_1&32769&3.55e-04&Materials \\\hline
\end{tabular}
\end{table}
\textbf{14 types of matrices}. It can be seen that GCOOSpDM does not always outperform cuSPARSE. To further understand the main reasons, we select 14 types of matrices that have different structures and non-zero patterns from a range of areas to analyze their performance differences. The details of the selected matrices are shown in Table \ref{table:selectedmatrx}. To normalize the algorithm performances, we use effective GFLOPS to measure the algorithms as the following Equation
\begin{equation}\label{equ:perf}
P_{algorithm}=\frac{2\times n^3\times (1-s)}{T_{algorithm}}.
\end{equation}
The performance comparison is shown in Fig. \ref{fig:realbar}. On three matrices (``nemeth11'', ``plbuckle'' and ``fpga\_dcop\_01''), GCOOSpDM is worse than cuSPARSE due to the non-zero distribution of the matrices. On these three matrices, the non-zero elements are mainly located on the diagonal of the matrices, such that there is little opportunity to reuse the pre-fetched value of $bv$ (i.e., line 30 will intermediately hold and no further calculations for current $bv$), but it still spends extra overheads to search $\ma{A}$.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\linewidth]{realbarp100.pdf}
\caption{The performance comparison of selected matrices on a Tesla P100 GPU. (The higher the better.)}
\label{fig:realbar}
\end{figure}
\subsection{Random sparse matrices}
We randomly generate square matrices whose dimension are in the range of $[400, 14500]$ with a step size of $100$. For each size of a matrix, we generate the elements with the sparsity in two ranges (i.e, $[0.8,0.995]$ at a $0.005$ step and $[0.995, 0.9995]$ at a $0.0005$ step). In total, there are $6968$ matrices with uniformly distributed non-zero elements for evaluation.
\textbf{Overall performance}. The performance comparison between GCOOSpDM and cuSPARSE using the randomly generated matrices is shown in Fig. \ref{fig:sythperf}. Our GCOOSpDM algorithm outperforms cuSPARSE in $99.51\%$, $99.23\%$ and $97.37\%$ matrices on GTX980, TitanX and P100 GPUs respectively, and the average speedups are $2.13\times$, $2\times$ and $1.57\times$ respectively. Particularly, the maximum speedups on the three GPUs are $4.7\times$, $6.5\times$ and $8.1\times$ respectively. On the cases that cuSPARSE is better GCOOSpDM, they only occupy a very small proportion (less than $3\%$), and the average performance ratio is only around $1.17$, which indicates very close performance on less than $3\%$ cases.
\begin{figure}[!h]
\centering
\subfigure[GTX 980]
{
\includegraphics[width=0.32\linewidth]{syth980_frequency.pdf}
}\hspace{-3.5mm}
\subfigure[Titan X Pascal]
{
\includegraphics[width=0.32\linewidth]{sythtitanx_frequency.pdf}
}\hspace{-3.5mm}
\subfigure[Tesla P100]
{
\includegraphics[width=0.32\linewidth]{sythp100_frequency.pdf}
}
\caption{The performance comparison with the frequency of the time ratio between cuSPARSE and GCOOSpDM with the random generated sparse matrices on three GPUs. The last value (i.e., 2.0+) of x-axis means that $T_{cuSPARSE}/T_{GCOOSpDM} \ge 2.0$.}
\label{fig:sythperf}
\end{figure}
\begin{figure}[!h]
\centering
\subfigure[$n=4000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsity980_4000.pdf}
}\hspace{-3mm}
\subfigure[$n=14000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsity980_14000.pdf}
}
\caption{Performance vs. sparsity on the GTX980 GPU. The lower the better.}
\label{fig:perfvssparsity980}
\end{figure}
\begin{figure}[!h]
\centering
\subfigure[$n=4000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsitytitanx_4000.pdf}
}\hspace{-3mm}
\subfigure[$n=14000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsitytitanx_14000.pdf}
}
\caption{Performance vs. sparsity on the TitanX GPU. The lower the better.}
\label{fig:perfvssparsitytitanx}
\end{figure}
\begin{figure}[!h]
\centering
\subfigure[$n=4000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsityp100_4000.pdf}
}\hspace{-3mm}
\subfigure[$n=14000$]
{
\includegraphics[width=0.48\linewidth]{perfvssparsityp100_14000.pdf}
}
\caption{Performance vs. sparsity on the P100 GPU. The lower the better.}
\label{fig:perfvssparsityp100}
\end{figure}
\textbf{Time vs. sparsity}. As we have shown the efficiency of GCOOSpDM in large range of matrices and sparsity, we want to study further about the performance related to the sparsity $s$. We take two matrices with medium ($n=4000$) and large ($n=14000$) dimensions to show the relationship between performance and sparsity. The range of sparsity is kept at $[0.95, 0.9995]$. Here we also put the time cost of the dense algorithm from cuBLAS into comparison so that we can understand under what sparsity GCOOSpDM can outperform cuBLAS. The results for these two sizes of matrices on GTX980, TitanX and P100 GPUs are shown in Fig. \ref{fig:perfwithspar980}, \ref{fig:perfwithspartitanx} and Fig. \ref{fig:perfwithsparp100}, respectively. On one hand, it can be seen that cuBLAS has a constant time cost when the sparsity of matrix increases since the dense algorithm does not consider zero values. On the other hand, the sparse algorithms of cuSPARSE and GCOOSpDM tend to have a linear speedup when the sparsity increases. Given the two specific dimensions of matrices, GCOOSpDM outperforms cuSPARSE with all sparsity. When the sparsity becomes larger than some thresholds, the sparse algorithm would have advantages than the dense one. However, cuSPARSE needs the sparsity be up to $0.995$ to outperform cuBLAS, while our proposed algorithm GCOOSpDM can outperform cuBLAS with sparsity larger than $0.98$. In summary, the GCOOSpDM algorithm is more applicable for matrix multiplication on GPUs than cuSPARSE and cuBLAS under sparsity larger than $0.98$ to achieve higher performance on current GPUs.
\begin{figure}[!h]
\centering
\subfigure[$s=0.98$]
{
\includegraphics[width=0.48\linewidth]{effective_flops980098.pdf}
}\hspace{-3mm}
\subfigure[$s=0.995$]
{
\includegraphics[width=0.48\linewidth]{effective_flops9800995.pdf}
}
\caption{Performance vs. dimension on GTX980. The higher the better.}
\label{fig:perfwithspar980}
\end{figure}
\begin{figure}[!h]
\centering
\subfigure[$s=0.98$]
{
\includegraphics[width=0.48\linewidth]{effective_flopstitanx098.pdf}
}\hspace{-3mm}
\subfigure[$s=0.995$]
{
\includegraphics[width=0.48\linewidth]{effective_flopstitanx0995.pdf}
}
\caption{Performance vs. dimension on TitanX. The higher the better.}
\label{fig:perfwithspartitanx}
\end{figure}
\begin{figure}[!h]
\centering
\subfigure[$s=0.98$]
{
\includegraphics[width=0.48\linewidth]{effective_flopsp100098.pdf}
}\hspace{-3mm}
\subfigure[$s=0.995$]
{
\includegraphics[width=0.48\linewidth]{effective_flopsp1000995.pdf}
}
\caption{Performance vs. dimension on P100. The higher the better.}
\label{fig:perfwithsparp100}
\end{figure}
\textbf{Performance vs. matrix size}. To further show the sensitivity of the algorithm to the matrix size, we demonstrate the throughput (GFLOPS) in a range of matrix dimensions (i.e., $n\in [400, 14000]$) at two sparsity $0.98$ and $0.995$. The experimental results with sparsity of $0.98$ and $0.995$ are in Fig. \ref{fig:perfwithspar980}, \ref{fig:perfwithspartitanx} and \ref{fig:perfwithsparp100} on three different GPUs. On the three tested GPUs, GCOOSpDM outperforms cuSPARSE with different values of $n$ and two sparsity. For small matrices (e.g., $n<1500$), cuBLAS still outperforms GCOOSpDM since it takes only a small number of cycles in calculating small matrices while GCOOSpDM needs extra overheads on memory allocation and matrix conversion. Given the sparsity of $0.98$ and $n>2000$, GCOOSpDM achieves similar performance as (or slightly better than) cuBLAS. With the sparsity of $0.995$, cuSPARSE achieves close performance with cuBLAS, while GCOOSpDM outperforms cuBLAS up to $2$ times.
\subsection{Breakdown of time costs}
In this subsection, assume that given $\ma{A}$ and $\ma{B}$ are both in the dense form, while $\ma{A}$ is of high sparsity, we would like to present the time costs of matrix conversion and the kernel calculation to finish the matrix multiplication using the sparse algorithm. The different overheads are summarized into three categories: memory allocation for sparse matrix storage, matrix conversion from the dense form to the sparse form, and SpDM kernel calculation. We summarize the first two categories as an extra overhead (EO), and the third one as the real time cost of kernel calculation (KC). The metrics of EO and KC are used to compare GCOOSpDM and cuSPARSE. Instead of using three GPUs, we only choose a TitanX GPU as our analysis platform, since three GPUs should have similar time distribution. Similar to the previous subsection, we use two sizes of matrices (i.e., $n=4000$ and $n=14000$) with sparsity of $[0.95, 0.96, 0.97, 0.98, 0.99]$ for comparison. The results are shown in Fig. \ref{fig:breakdown}. It can be seen that EO has only a small proportion of the total time, and both GCOOSpDM and cuSPARSE have a very close overhead of EO. The dominated part is the execution time of the kernel that calculates the matrix multiplication.
\begin{figure}[!h]
\centering
\subfigure[$n=4000$]
{
\includegraphics[width=0.48\linewidth]{breakdown_4000.pdf}
}\hspace{-3mm}
\subfigure[$n=14000$]
{
\includegraphics[width=0.48\linewidth]{breakdown_14000.pdf}
}\hspace{-3mm}
\caption{Time breakdown for two sizes of matrices. ``GCOO.'' represents the GCOOSpDM algorithm, and ``cuSPA.'' represents the algorithm in cuSPARSE.}
\label{fig:breakdown}
\end{figure}
\subsection{Instruction analysis}
\input{performance_analysis}
\section{Related work}\label{section:relatedwork}
Multiplication of sparse matrices to dense vectors (SpMV) on GPUs have been well studied (e.g., \cite{greathouse2014efficient}\cite{hou2017auto}\cite{bell2009implementing}\cite{merrill2016merge}). Even SpDM can be implemented by multiple SpMVs, the performance could be bad due to a large number of kernel invokes if the matrix is with a large dimension. However, some optimization principles can be applied for SpDM. For example, Yang et al. \cite{yang2018design} use split row \cite{bell2009implementing} and merged path \cite{merrill2016merge} to design SpDM algorithms particularly for tall-skinny matrices.
Regarding the SpDM algorithm analysis, Greiner et al. \cite{greiner2010complexity} propose an I/O model to interpret the lower bound of efficient serial algorithms. Cache oblivious dense and sparse matrix algorithms are presented by Bader et al. for multi-core CPUs \cite{bader2008cache}. Performance benchmarks \cite{ezouaoui2013performance} are conducted to evaluate the efficiency of different sparse matrix formats for SpDM. Koanantakool et al., \cite{koanantakool2016communication} introduce the communication-avoiding SpDM algorithms that are applied in distributed memory systems. Recent work in designing the row reordering technique to achieve better data temporal locality \cite{jiang2020novel} and the dynamic parameter tuning \cite{parger2020speck} to improve the SpDM performance on GPUs.
\section{Conclusion and Future Work}\label{section:conclusion}
Sparse-dense matrix-matrix multiplication is commonly used in many scientific computing areas, while designing such algorithms on modern GPUs is non-trivial due to the irregular structure of the sparse matrix. In this paper, we propose an efficient sparse matrix-dense matrix multiplication algorithm on GPUs, called GCOOSpDM. The main optimization techniques used in our algorithm are the coalesced global memory access, proper usage of the shared memory, and reuse the data from the slow global memory. The experimental results show that our proposed algorithm outperforms the vendor-based library: cuSPARSE several times on both the public sparse dataset and randomly generated matrices on three recent Nvidia GPUs (i.e., GTX 980, Titan X Pascal, and Tesla P100). We also analyze in depth the performance improvement on instruction-level to understand why GCOOSpDM performs better than cuSPARSE. The key observation of the instruction-level analysis is that the reduced number of global memory access contributes a lot to the performance gain.
It is difficult for a single algorithm to fit all structures of matrices, sparsity and different types of GPUs. Auto-tune algorithms play an important role for algorithms to find efficient configuration or implementations in different cases. We would like to consider the auto-tune scheme to set proper $p$ and $b$ for our GCOOSpDM algorithm in the future work, and try to extend the GCOO storage format to the multiplication of two sparse matrices.
\bibliographystyle{IEEEtran}
\Urlmuskip=0mu plus 1mu
| 77d20b1ca6ab3877600fab04c84a1c6766291b2c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $G$ be a finite group and $S\subseteq G$. The \emph{Cayley digraph} $\cay(G,S)$ over $G$ with the connection set $S$ is defined to be the digraph with the vertex set $G$ and the arc set $\{(g,sg): g\in G, s\in S\}$. Two Cayley digraphs over $G$ are called \emph{Cayley isomorphic} if there exists an isomorphism between them which is also an automorphism of $G$. Clearly, two Cayley isomorphic Cayley digraphs are isomorphic. The converse statement is not true in general (see~\cite{AlPar, ET}). A subset $S\subseteq G$ is called a \emph{$\CI$-subset} if for each $T\subseteq G$ the Cayley digraphs $\cay(G,S)$ and $\cay(G,T)$ are isomorphic if and only if they are Cayley isomorphic. A finite group $G$ is called a \emph{$\DCI$-group} (\emph{$\CI$-group}, respectively) if each subset of $G$ (each inverse-closed subset of $G$, respectively) is a $\CI$-subset.
The investigation of $\DCI$-groups was initiated by \'Ad\'am~\cite{Adam} who conjectured, in our terms, that every cyclic group is a $\DCI$-group. This conjecture was disproved by Elspas and Turner in~\cite{ET}. The problem of determining of all finite $\DCI$- and $\CI$-groups was suggested by Babai and Frankl in~\cite{BF}. For more information on $\DCI$- and $\CI$-groups we refer the readers to the survey paper~\cite{Li}.
In this paper we are interested in abelian $\DCI$-groups. The cyclic group of order~$n$ is denoted by $C_n$. Elspas and Turner~\cite{ET} and independently Djokovi\'c~\cite{Dj} proved that every cyclic group of prime order is a $\DCI$-group. The fact that $C_{pq}$ is a $\DCI$-group for distinct primes $p$ and $q$ was proved by Alspach and Parsons in~\cite{AlPar} and independently by Klin and P\"oschel in~\cite{KP}. The complete classification of all cyclic $\DCI$-groups was obtained by Muzychuk in~\cite{M1,M2}. He proved that a cyclic group of order~$n$ is a $\DCI$-group if and only if $n=k$ or $n=2k$, where $k$ is square-free.
Denote the class of all finite abelian groups whose all Sylow subgroups are elementary abelian by $\mathcal{E}$. From~\cite[Theorem~1.1]{KM} it follows that every $\DCI$-group is the coprime product (i.e., the direct product of groups of coprime orders) of groups from the following list:
$$C_{p}^k,~C_4,~Q_8,~A_4,~H\rtimes \langle z \rangle,$$
where $p$ is a prime, $H$ is a group of odd order from $\mathcal{E}$, $|z|\in \{2,4\}$, and $h^z=h^{-1}$ for every $h\in H$. One can check that the class of $\DCI$-groups is closed under taking subgroups. So one of the crucial steps towards the classification of all $\DCI$-groups is to determine which groups from $\mathcal{E}$ are $\DCI$.
The following non-cyclic groups from $\mathcal{E}$ are $\DCI$-groups ($p$ and $q$ are assumed to be distinct primes): $C_p^2$~\cite{AlN,God}; $C_p^3$~\cite{AlN,Dob0}; $C_2^4$, $C_2^5$~\cite{CLi}; $C_p^4$, where $p$ is odd~\cite{HM} (a proof for $C_p^4$ with no condition on $p$ was given in~\cite{Mor}); $C_p^5$, where $p$ is odd~\cite{FK}; $C_p^2\times C_q$~\cite{KM}; $C_p^3\times C_q$~\cite{MS}; $C_p^4\times C_q$~\cite{KR2}. The smallest example of a non-$\DCI$-group from $\mathcal{E}$ was found by Nowitz~\cite{Now}. He proved that $C_2^6$ is non-$\DCI$. This implies that $C_2^n$ is non-$\DCI$ for every $n\geq 6$. Also $C_3^n$ is non-$\DCI$ for every $n\geq 8$~\cite{Sp} and $C_p^n$ is non-$\DCI$ for every prime $p$ and $n\geq 2p+3$~\cite{Som}.
In this paper we find a new infinite family of $\DCI$-groups from $\mathcal{E}$ which are close to the smallest non-$\DCI$-group from $\mathcal{E}$. The main result of the paper can be formulated as follows.
\begin{theo}\label{main}
Let $p$ be a prime. Then the group $C_2^5\times C_p$ is a $\DCI$-group if and only if $p\neq 2$.
\end{theo}
Theorem~\ref{main} extends the results obtained in~\cite{KM,KR2,MS} which imply that the group $C_p^k\times C_q$ is a $\DCI$-group whenever $p$ and $q$ are distinct primes and $k\leq 4$. Note that the ``only if'' part of Theorem~\ref{main}, in fact, was proved by Nowitz in~\cite{Now}. The next corollary immediately follows from~\cite[Theorem~1.1]{KM} and Theorem~\ref{main}.
\begin{corl1}
Let $p$ be a prime. Then a group $G$ of order $32p$ is a $\DCI$-group if and only if $p\neq 2$ and $G\cong C_2^5\times C_p$.
\end{corl1}
To prove Theorem~\ref{main}, we use the $S$-ring approach. An \emph{$S$-ring} over a group $G$ is a subring of the group ring $\mathbb{Z}G$ which is a free $\mathbb{Z}$-module spanned by a special partition of $G$. If every $S$-ring from a certain family of $S$-rings over $G$ is a $\CI$-$S$-ring then $G$ is a $\DCI$-group (see Section~4). The definition of an $S$-ring goes back to Schur~\cite{Schur} and Wielandt~\cite{Wi}. The usage of $S$-rings in the investigation of $\DCI$-groups was proposed by Klin and P\"oschel~\cite{KP}. Many of recent results on $\DCI$-groups were obtained with using $S$-rings (see~\cite{HM,KM,KR1,KR2,MS}).
The text of the paper is organized in the following way. In Section~2 we provide definitions and basic facts concerned with $S$-rings. Section~3 contains a necessary information on isomorphisms of $S$-rings. In Section~4 we discuss $\CI$-$S$-rings and their relation with $\DCI$-groups. Also in this section we prove a sufficient condition of $\CI$-property for $S$-rings (Lemma~\ref{minnorm}). Section~5 is devoted to the generalized wreath and star products of $S$-rings. Here we deduce from previously obtained results two sufficient conditions for the generalized wreath product of $S$-rings to be a $\CI$-$S$-ring (Lemma~\ref{kerci} and Lemma~\ref{citens}). Section~6 and~7 are concerned with $p$-$S$-rings and $S$-rings over groups of non-powerful order respectively. In Section~8 we provide properties of $S$-rings over the groups $C_2^n$, $n\leq 5$, and prove that all $S$-rings over these groups are $\CI$. The material of this section is based on computational results obtained with the help of the GAP package COCO2P~\cite{GAP}. Finally, in Section~9 we prove Theorem~\ref{main}.
\\
\\
\\
{\bf Notation.}
Let $G$ be a finite group and $X\subseteq G$. The element $\sum \limits_{x\in X} {x}$ of the group ring $\mathbb{Z}G$ is denoted by $\underline{X}$.
The set $\{x^{-1}:x\in X\}$ is denoted by $X^{-1}$.
The subgroup of $G$ generated by $X$ is denoted by $\langle X\rangle$; we also set $\rad(X)=\{g\in G:\ gX=Xg=X\}$.
Given a set $X\subseteq G$ the set $\{(g,xg): x\in X, g\in G\}$ of arcs of the Cayley digraph $\cay(G,X)$ is denoted by $R(X)$.
The group of all permutations of $G$ is denoted by $\sym(G)$.
The subgroup of $\sym(G)$ consisting of all right translations of $G$ is denoted by $G_{\r}$.
The set $\{K \leq \sym(G):~ K \geq G_{\r}\}$ is denoted by $\Sup(G_{\r})$.
For a set $\Delta\subseteq \sym(G)$ and a section $S=U/L$ of $G$ we set $\Delta^S=\{f^S:~f\in \Delta,~S^f=S\}$, where $S^f=S$ means that $f$ permutes the $L$-cosets in $U$ and $f^S$ denotes the bijection of $S$ induced by $f$.
If $K\leq \sym(\Omega)$ and $\alpha\in \Omega$ then the stabilizer of $\alpha$ in $K$ and the set of all orbits of $K$ on $\Omega$ are denoted by $K_{\alpha}$ and $\orb(K,\Omega)$ respectively.
If $H\leq G$ then the normalizer of $H$ in $G$ is denoted by $N_G(H)$.
The cyclic group of order $n$ is denoted by $C_n$.
The class of all finite abelian groups whose every Sylow subgroup is elementary abelian is denoted by $\mathcal{E}$.
\section{$S$-rings}
In this section we give a background of $S$-rings. In general, we follow~\cite{KR2}, where the most part of the material is contained. For more information on $S$-rings we refer the readers to~\cite{CP, MP1}.
Let $G$ be a finite group and $\mathbb{Z}G$ the integer group ring. Denote the identity element of $G$ by $e$. A subring $\mathcal{A}\subseteq \mathbb{Z} G$ is called an \emph{$S$-ring (a Schur ring)} over $G$ if there exists a partition $\mathcal{S}(\mathcal{A})$ of~$G$ such that:
$(1)$ $\{e\}\in\mathcal{S}(\mathcal{A})$,
$(2)$ if $X\in\mathcal{S}(\mathcal{A})$ then $X^{-1}\in\mathcal{S}(\mathcal{A})$,
$(3)$ $\mathcal{A}=\Span_{\mathbb{Z}}\{\underline{X}:\ X\in\mathcal{S}(\mathcal{A})\}$.
\noindent The elements of $\mathcal{S}(\mathcal{A})$ are called the \emph{basic sets} of $\mathcal{A}$ and the number $\rk(\mathcal{A})=|\mathcal{S}(\mathcal{A})|$ is called the \emph{rank} of $\mathcal{A}$. If $X,Y \in \mathcal{S}(\mathcal{A})$ then $XY \in \mathcal{S}(\mathcal{A})$ whenever $|X|=1$ or $|Y|=1$.
Let $\mathcal{A}$ be an $S$-ring over a group $G$. A set $X \subseteq G$ is called an \emph{$\mathcal{A}$-set} if $\underline{X}\in \mathcal{A}$. A subgroup $H \leq G$ is called an \emph{$\mathcal{A}$-subgroup} if $H$ is an $\mathcal{A}$-set. From the definition it follows that the intersection of $\mathcal{A}$-subgroups is also an $\mathcal{A}$-subgroup. One can check that for each $\mathcal{A}$-set $X$ the groups $\langle X \rangle$ and $\rad(X)$ are $\mathcal{A}$-subgroups. By the \emph{thin radical} of $\mathcal{A}$ we mean the set defined as
$$\O_\theta(\mathcal{A})=\{ x \in G:~\{x\} \in \mathcal{S}(\mathcal{A}) \}.$$
It is easy to see that $\O_\theta(\mathcal{A})$ is an $\mathcal{A}$-subgroup.
\begin{lemm}~\cite[Lemma~2.1]{EKP}\label{intersection}
Let $\mathcal{A}$ be an $S$-ring over a group $G$, $H$ an $\mathcal{A}$-subgroup of $G$, and $X \in \mathcal{S}(\mathcal{A})$. Then the number $|X\cap Hx|$ does not depend on $x\in X$.
\end{lemm}
Let $L \unlhd U\leq G$. A section $U/L$ is called an \emph{$\mathcal{A}$-section} if $U$ and $L$ are $\mathcal{A}$-subgroups. If $S=U/L$ is an $\mathcal{A}$-section then the module
$$\mathcal{A}_S=Span_{\mathbb{Z}}\left\{\underline{X}^{\pi}:~X\in\mathcal{S}(\mathcal{A}),~X\subseteq U\right\},$$
where $\pi:U\rightarrow U/L$ is the canonical epimorphism, is an $S$-ring over $S$.
\section{Isomorphisms and schurity}
Let $\mathcal{A}$ and $\mathcal{A}^{\prime}$ be $S$-rings over groups $G$ and $G^{\prime}$ respectively. A bijection $f:G\rightarrow G^{\prime}$ is called an \emph{isomorphism} from $\mathcal{A}$ to $\mathcal{A}^{\prime}$ if
$$\{R(X)^f: X\in \mathcal{S}(\mathcal{A})\}=\{R(X^{\prime}): X^{\prime}\in \mathcal{S}(\mathcal{A}^{\prime})\},$$
where $R(X)^f=\{(g^f,~h^f):~(g,~h)\in R(X)\}$. If there exists an isomorphism from $\mathcal{A}$ to $\mathcal{A}^{\prime}$ then we say that $\mathcal{A}$ and $\mathcal{A}^{\prime}$ are \emph{isomorphic} and write $\mathcal{A}\cong \mathcal{A}^{\prime}$.
The group of all isomorphisms from $\mathcal{A}$ onto itself contains a normal subgroup
$$\{f\in \sym(G): R(X)^f=R(X)~\text{for every}~X\in \mathcal{S}(\mathcal{A})\}$$
called the \emph{automorphism group} of $\mathcal{A}$ and denoted by $\aut(\mathcal{A})$. The definition implies that $G_{\r}\leq \aut(\mathcal{A})$. The $S$-ring $\mathcal{A}$ is called \emph{normal} if $G_{\r}$ is normal in $\aut(\mathcal{A})$. One can verify that if $S$ is an $\mathcal{A}$-section then $\aut(\mathcal{A})^S\leq \aut(\mathcal{A}_S)$. Denote the group $\aut(\mathcal{A})\cap \aut(G)$ by $\aut_G(\mathcal{A})$. It easy to check that if $S$ is an $\mathcal{A}$-section then $\aut_G(\mathcal{A})^S\leq \aut_S(\mathcal{A}_S)$. One can verify that
$$\aut_G(\mathcal{A})=N_{\aut(\mathcal{A})}(G_{\r})_e.$$
Let $K\in \Sup(G_{\r})$. Schur proved in~\cite{Schur} that the $\mathbb{Z}$-submodule
$$V(K,G)=\Span_{\mathbb{Z}}\{\underline{X}:~X\in \orb(K_e,~G)\},$$
is an $S$-ring over $G$. An $S$-ring $\mathcal{A}$ over $G$ is called \emph{schurian} if $\mathcal{A}=V(K,G)$ for some $K\in \Sup(G_{\r})$. One can verify that given $K_1,K_2\in \Sup(G_{\r})$,
$$\text{if}~K_1\leq K_2~\text{then}~V(K_1,G)\geq V(K_2,G).~\eqno(1)$$
If $\mathcal{A}=V(K,G)$ for some $K\in \Sup(G_{\r})$ and $S$ is an $\mathcal{A}$-section then $\mathcal{A}_S=V(K^S,G)$. So if $\mathcal{A}$ is schurian then $\mathcal{A}_S$ is also schurian for every $\mathcal{A}$-section $S$. It can be checked that
$$V(\aut(\mathcal{A}),G)\geq \mathcal{A}~\eqno(2)$$
and the equality is attained if and only if $\mathcal{A}$ is schurian.
An $S$-ring $\mathcal{A}$ over a group $G$ is defined to be \emph{cyclotomic} if there exists $K\leq\aut(G)$ such that $\mathcal{S}(\mathcal{A})=\orb(K,G)$. In this case we write $\mathcal{A}=\cyc(K,G)$. Obviously, $\mathcal{A}=V(G_{\r}K,G)$. So every cyclotomic $S$-ring is schurian. If $\mathcal{A}=\cyc(K,G)$ for some $K\leq \aut(G)$ and $S$ is an $\mathcal{A}$-section then $\mathcal{A}_S=\cyc(K^S,G)$. Therefore if $\mathcal{A}$ is cyclotomic then $\mathcal{A}_S$ is also cyclotomic for every $\mathcal{A}$-section $S$.
Two permutation groups $K_1$ and $K_2$ on a set $\Omega$ are called \emph{$2$-equivalent} if $\orb(K_1,\Omega^2)=\orb(K_2,\Omega^2)$ (here we assume that $K_1$ and $K_2$ act on $\Omega^2$ componentwise). In this case we write $K_1\approx_2 K_2$. The relation $\approx_2$ is an equivalence relation on the set of all subgroups of $\sym(\Omega)$. Every equivalence class has a unique maximal element. Given $K\leq \sym(\Omega)$, this element is called the \emph{$2$-closure} of $K$ and denoted by $K^{(2)}$. If $\mathcal{A}=V(K,G)$ for some $K\in \Sup(G_{\r})$ then $K^{(2)}=\aut(\mathcal{A})$. An $S$-ring $\mathcal{A}$ over $G$ is called \emph{$2$-minimal} if
$$\{K\in \Sup(G_{\r}):~K\approx_2 \aut(\mathcal{A})\}=\{\aut(\mathcal{A})\}.$$
Two groups $K_1,K_2\leq \aut(G)$ are said to be \emph{Cayley equivalent} if $\orb(K_1,G)=\orb(K_2,G)$. In this case we write $K_1\approx_{\cay} K_2$. If $\mathcal{A}=\cyc(K,G)$ for some $K\leq \aut(G)$ then $\aut_G(\mathcal{A})$ is the largest group which is Cayley equivalent to $K$. A cyclotomic $S$-ring $\mathcal{A}$ over $G$ is called \emph{Cayley minimal} if
$$\{K\leq \aut(G):~K\approx_{\cay} \aut_G(\mathcal{A})\}=\{\aut_G(\mathcal{A})\}.$$
It is easy to see that $\mathbb{Z}G$ is Cayley minimal.
\section{$\CI$-$S$-rings}
Let $\mathcal{A}$ be an $S$-ring over a group $G$. Put
$$\iso(\mathcal{A})=\{f\in \sym(G):~\text{f is an isomorphism from}~\mathcal{A}~\text{onto}~\text{$S$-ring over}~G\}.$$
One can see that $\aut(\mathcal{A})\aut(G)\subseteq \iso(\mathcal{A})$. However, the converse inclusion does not hold in general. The $S$-ring $\mathcal{A}$ is defined to be a \emph{$\CI$-$S$-ring} if $\aut(\mathcal{A})\aut(G)=\iso(\mathcal{A})$. It is easy to check that $\mathbb{Z}G$ and the $S$-ring of rank~$2$ over $G$ are $\CI$-$S$-rings.
Put
$$\Sup_2(G_\r)=\{K \in \Sup(G_\r):~ K^{(2)}=K\}.$$
The group $M\leq \sym(G)$ is said to be \emph{$G$-regular} if $M$ is regular and isomorphic to $G$. Following~\cite{HM}, we say that a group $K \in \Sup(G_\r)$ is \emph{$G$-transjugate} if every $G$-regular subgroup of $K$ is $K$-conjugate to $G_{\r}$. Babai proved in~\cite{Babai} the statement which can be formulated in our terms as follows: a set $S\subseteq G$ is a $\CI$-subset if and only if the group $\aut(\cay(G,S))$ is $G$-transjugate. The next lemma provides a similar criterion for a schurian $S$-ring to be $\CI$.
\begin{lemm}\label{trans}
Let $K \in \Sup_2(G_\r)$ and $\mathcal{A}=V(K,G)$. Then $\mathcal{A}$ is a $\CI$-$S$-ring if and only if $K$ is $G$-transjugate.
\end{lemm}
\begin{proof}
The statement of the lemma follows from~\cite[Theorem~2.6]{HM}.
\end{proof}
Let $K_1, K_2\in \Sup(G_\r)$ such that $K_1 \le K_2$. Then $K_1$ is called a \emph{$G$-complete
subgroup} of $K_2$ if every $G$-regular subgroup of $K_2$ is $K_2$-conjugate to some $G$-regular subgroup of $K_1$ (see \cite[Definition~2]{HM}). In this case we write $K_1 \preceq_G K_2$. The relation $\preceq_G$ is a partial order on $\Sup(G_\r)$. The set of the minimal elements of $\Sup_2(G_\r)$ with respect to $\preceq_G$ is denoted by $\Sup_2^{\min}(G_\r)$.
\begin{lemm}\cite[Lemma~3.3]{KR2}\label{cimin}
Let $G$ be a finite group. If $V(K, G)$ is a $\CI$-$S$-ring for every $K \in \Sup_2^{\rm min}(G_\r)$ then $G$ is a $\DCI$-group.
\end{lemm}
\begin{rem1}
The condition that $V(K, G)$ is a $\CI$-$S$-ring for every $K \in \Sup_2^{\rm min}(G_\r)$ is equivalent to, say, that every schurian $S$-ring over $G$ is a $\CI$-$S$-ring. However, it is not known whether the statement converse to Lemma~\ref{cimin} is true.
\end{rem1}
We finish the subsection with the lemma that gives a sufficient condition for an $S$-ring to be a $\CI$-$S$-ring. In order to formulate this condition, we need to introduce some further notations. Let $\mathcal{A}$ be a schurian $S$-ring over an abelian group $G$ and $L$ a normal $\mathcal{A}$-subgroup of $G$. Then the partition of $G$ into the $L$-cosets is $\aut(\mathcal{A})$-invariant. The kernel of the action of $\aut(\mathcal{A})$ on the latter cosets is denoted by $\aut(\mathcal{A})_{G/L}$. Since $\aut(\mathcal{A})_{G/L}$ is a normal subgroup of $\aut(\mathcal{A})$, we can form the group $K=\aut(\mathcal{A})_{G/L}G_{\r}$. Clearly, $K\leq\aut(\mathcal{A})$. From~\cite[Proposition~2.1]{HM} it follows that $K=K^{(2)}$.
\begin{lemm}\label{minnorm}
Let $\mathcal{A}$ be a schurian $S$-ring over an abelian group $G$, $L$ an $\mathcal{A}$-subgroup of $G$, and $K=\aut(\mathcal{A})_{G/L}G_{\r}$. Suppose that both $\mathcal{A}_{G/L}$ and $V(K, G)$ are $\CI$-$S$-rings and $\mathcal{A}_{G/L}$ is normal. Then $\mathcal{A}$ is a $\CI$-S-ring.
\end{lemm}
\begin{proof}
Firstly we prove that the group $\aut(\mathcal{A})^{G/L}$ is $G/L$-transjugate. Suppose that $F$ is a $G/L$-regular subgroup of $\aut(\mathcal{A})^{G/L}$. The $S$-ring $\mathcal{A}_{G/L}$ is a $\CI$-$S$-ring by the assumption of the lemma. So Lemma~\ref{trans} implies that the group $\aut(\mathcal{A}_{G/L})$ is $G/L$-transjugate. Since $F\leq \aut(\mathcal{A})^{G/L}\leq \aut(\mathcal{A}_{G/L})$, we conclude that $F$ and $(G/L)_{\r}$ are $\aut(\mathcal{A}_{G/L})$-conjugate. However, $\mathcal{A}_{G/L}$ is normal and hence $F=(G/L)_{\r}$. Therefore $\aut(\mathcal{A})^{G/L}$ is $G/L$-transjugate.
Now let us show that $K\preceq_G \aut(\mathcal{A})$. Let $H$ be a $G$-regular subgroup of $\aut(\mathcal{A})$. Then $H^{G/L}$ is abelian transitive subgroup of $\aut(\mathcal{A})^{G/L}$ and hence $H^{G/L}$ is regular on $G/L$. Therefore $H^{G/L}\cong (G/L)_{\r}=(G_{\r})^{G/L}$. There exists $\gamma\in \aut(\mathcal{A})$ such that $(H^{G/L})^{\gamma^{G/L}}=(G/L)_{\r}=(G_{right})^{G/L}$ because $\aut(\mathcal{A})^{G/L}$ is $G/L$-transjugate. This yields that $H^{\gamma}\leq K$. Thus, $K\preceq_G \aut(\mathcal{A})$.
Finally, prove that $\aut(\mathcal{A})$ is $G$-transjugate. Again, let $H$ be a $G$-regular subgroup of $\aut(\mathcal{A})$. Since $K\preceq_G \aut(\mathcal{A})$, there exists $\gamma\in\aut(\mathcal{A})$ such that $H^{\gamma}\leq K$. The $S$-ring $V(K, G)$ is a $\CI$-$S$-ring by the assumption of the lemma. So $K$ is $G$-transjugate by Lemma~\ref{trans}. Therefore $H^{\gamma}$ and $G_{\r}$ are $K$-conjugate and hence $H$ and $G_{\r}$ are $\aut(\mathcal{A})$-conjugate. Thus, $\aut(\mathcal{A})$ is $G$-transjugate and $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{trans}. The lemma is proved.
\end{proof}
It should be mentioned that the proof of Lemma~\ref{minnorm} is quite similar to the proof of~\cite[Lemma~3.6]{KR2}.
\section{Generalized wreath and star products}
Let $\mathcal{A}$ be an $S$-ring over a group $G$ and $S=U/L$ an $\mathcal{A}$-section of $G$. An $S$-ring~$\mathcal{A}$ is called the \emph{$S$-wreath product} or the \emph{generalized wreath product} of $\mathcal{A}_U$ and $\mathcal{A}_{G/L}$ if $L\trianglelefteq G$ and $L\leq\rad(X)$ for each basic set $X$ outside~$U$. In this case we write $\mathcal{A}=\mathcal{A}_U\wr_{S}\mathcal{A}_{G/L}$ and omit $S$ when $U=L$. The construction of the generalized wreath product of $S$-rings was introduced in~\cite{EP}.
The $S$-wreath product is called \emph{nontrivial} or \emph{proper} if $L\neq \{e\}$ and $U\neq G$. An $S$-ring $\mathcal{A}$ is said to be \emph{decomposable} if $\mathcal{A}$ is the nontrivial $S$-wreath product for some $\mathcal{A}$-section $S$ of $G$; otherwise $\mathcal{A}$ is said to be indecomposable. We say that an $\mathcal{A}$-subgroup $U<G$ has a \emph{gwr-complement} with respect to $\mathcal{A}$ if there exists a nontrivial normal $\mathcal{A}$-subgroup $L$ of $G$ such that $L\leq U$ and $\mathcal{A}$ is the $S$-wreath product, where $S=U/L$.
\begin{lemm}\cite[Theorem~1.1]{KR1}\label{cigwr}
Let $G\in \mathcal{E}$, $\mathcal{A}$ an $S$-ring over $G$, and $S=U/L$ an $\mathcal{A}$-section of $G$. Suppose that $\mathcal{A}$ is the nontrivial $S$-wreath product and the $S$-rings $\mathcal{A}_U$ and $\mathcal{A}_{G/L}$ are $\CI$-$S$-rings. Then $\mathcal{A}$ is a $CI$-$S$-ring whenever
$$\aut_{S}(\mathcal{A}_{S})=\aut_U(\mathcal{A}_U)^{S}\aut_{G/L}(\mathcal{A}_{G/L})^{S}.$$
In particular, $\mathcal{A}$ is a $\CI$-$S$-ring if $\aut_{S}(\mathcal{A}_{S})=\aut_U(\mathcal{A}_U)^{S}$ or $\aut_{S}(\mathcal{A}_{S})=\aut_{G/L}(\mathcal{A}_{G/L})^{S}$.
\end{lemm}
\begin{lemm}\cite[Proposition~4.1]{KR1}\label{trivial}
In the conditions of Lemma~\ref{cigwr}, suppose that $\mathcal{A}_S=\mathbb{Z}S$. Then $\mathcal{A}$ is a $\CI$-$S$-ring. In particular, if $U=L$ then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{lemm}\cite[Lemma~4.2]{KR2}\label{cicayleymin}
In the conditions of Lemma~\ref{cigwr}, suppose that at least one of the $S$-rings $\mathcal{A}_U$ and $\mathcal{A}_{G/L}$ is cyclotomic and $\mathcal{A}_S$ is Cayley minimal. Then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{lemm}\label{kerwreath}
Let $\mathcal{A}$ be an $S$-ring over an abelian group $G$. Suppose that $\mathcal{A}$ is the nontrivial $S=U/L$-wreath product for some $\mathcal{A}$-section $S=U/L$ and $L_1$ is an $\mathcal{A}$-subgroup containing $L$. Then $\mathcal{B}=V(K, G)$, where $K=\aut(\mathcal{A})_{G/L_1}G_{\r}$, is also the $S$-wreath product.
\end{lemm}
\begin{proof}
Since $K\leq \aut(\mathcal{A})$, from Eqs.~(1) and~(2) it follows that
$$\mathcal{B}=V(K,G)\geq V(\aut(\mathcal{A}),G)\geq \mathcal{A}.$$
So $U$ and $L$ are also $\mathcal{B}$-subgroups.
Let $\mathcal{C}=\mathbb{Z}U\wr_{S}\mathbb{Z}(G/L)$. The $S$-rings $\mathcal{C}_U$ and $\mathcal{C}_{G/L}$ are schurian and $\mathcal{C}_S$ is $2$-minimal. So $\mathcal{C}$ is schurian by~\cite[Corollary~10.3]{MP2}. This implies that
$$\mathcal{C}=V(\aut(\mathcal{C}),G).~\eqno(3)$$
Every element from $\aut(\mathcal{C})_e$ fixes every basic set of $\mathcal{C}$ and hence it fixes every $L$-coset. Since $L_1\geq L$, every element from $\aut(\mathcal{C})_e$ fixes every $L_1$-coset. We conclude that $\aut(\mathcal{C})_e\leq \aut(\mathcal{A})_{G/L_1}$ and hence $\aut(\mathcal{C})\leq K$. Now from Eqs.~(1) and~(3) it follows that
$$\mathcal{C}=V(\aut(\mathcal{C}),G)\geq V(K,G)=\mathcal{B}.~\eqno(4)$$
The group $U$ is both $\mathcal{B}$-,$\mathcal{C}$-subgroup. Due to Eq.~(4), every basic set of $\mathcal{B}$ which lies outside $U$ is a union of some basic sets of $\mathcal{C}$ which lie outside $U$. So $L\leq \rad(X)$ for every $X\in \mathcal{S}(\mathcal{B})$ outside $U$. Thus, $\mathcal{B}$ is the $S$-wreath product. The lemma is proved.
\end{proof}
\begin{lemm}\label{kerci}
In the conditions of Lemma~\ref{cigwr}, suppose that: $(1)$ every $S$-ring over $U$ is a $\CI$-$S$-ring; $(2)$ $\mathcal{A}_{G/L}$ is $2$-minimal or normal. Then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
Let $\mathcal{B}=V(K,G)$, where $K=\aut(\mathcal{A})_{G/L}G_{\r}$. From Lemma~\ref{kerwreath} it follows that $\mathcal{B}$ is the $S$-wreath product. Since $L_1=L$, the definition of $\mathcal{B}$ implies that $\mathcal{B}_{G/L}=\mathbb{Z}(G/L)$ and hence $\mathcal{B}_S=\mathbb{Z}S$. Clearly, $\mathcal{B}_{G/L}$ is a $\CI$-$S$-ring. The $S$-ring $\mathcal{B}_U$ is a $\CI$-$S$-ring by the assumption of the lemma. Therefore $\mathcal{B}$ is a $\CI$-$S$-ring by Lemma~\ref{trivial}. The $S$-ring $\mathcal{A}_{G/L}$ is a $\CI$-$S$-ring by the assumption of the lemma. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by~\cite[Lemma~3.6]{KR2} whenever $\mathcal{A}_{G/L}$ is $2$-minimal and by Lemma~\ref{minnorm} whenever $\mathcal{A}_{G/L}$ is normal. The lemma is proved.
\end{proof}
Let $V$ and $W$ be $\mathcal{A}$-subgroups. The $S$-ring $\mathcal{A}$ is called the \emph{star product} of $\mathcal{A}_V$ and $\mathcal{A}_W$ if the following conditions hold:
$(1)$ $V\cap W\trianglelefteq W$;
$(2)$ each $T\in \mathcal{S}(\mathcal{A})$ with $T\subseteq (W\setminus V) $ is a union of some $V\cap W$-cosets;
$(3)$ for each $T\in \mathcal{S}(\mathcal{A})$ with $T\subseteq G\setminus (V\cup W)$ there exist $R\in \mathcal{S}(\mathcal{A}_V)$ and $S\in \mathcal{S}(\mathcal{A}_W)$ such that $T=RS$.
\noindent In this case we write $\mathcal{A}=\mathcal{A}_V \star \mathcal{A}_W$. The construction of the star product of $S$-rings was introduced in~\cite{HM}. The star product is called \emph{nontrivial} if $V\neq \{e\}$ and $V\neq G$. If $V\cap W=\{e\}$ then the star product is the usual \emph{tensor product} of $\mathcal{A}_V$ and $\mathcal{A}_W$ (see~\cite[p.5]{EKP}). In this case we write $\mathcal{A}=\mathcal{A}_V\otimes \mathcal{A}_W$. One can check that if $\mathcal{A}=\mathcal{A}_V\otimes \mathcal{A}_W$ then $\aut(\mathcal{A})=\aut(\mathcal{A}_V)\times \aut(\mathcal{A}_W)$. If $V\cap W\neq \{e\}$ then $\mathcal{A}$ is the nontrivial $V/(V\cap W)$-wreath product.
\begin{lemm}\label{cistar}
Let $G\in\mathcal{E}$ and $\mathcal{A}$ a schurian $S$-ring over $G$. Suppose that $\mathcal{A}=\mathcal{A}_V \star \mathcal{A}_W$ for some $\mathcal{A}$-subgroups $V$ and $W$ of $G$ and the $S$-rings $\mathcal{A}_V$ and $\mathcal{A}_{W/(V\cap W)}$ are $\CI$-$S$-rings. Then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
The statement of the lemma follows from~\cite[Proposition~3.2, Theorem~4.1]{KM}.
\end{proof}
\begin{lemm}\cite[Lemma 2.8]{FK}\label{tenspr}
Let $\mathcal{A}$ be an $S$-ring over an abelian group $G=G_1\times G_2$. Assume that $G_1$ and $G_2$ are $\mathcal{A}$-groups. Then $\mathcal{A}=\mathcal{A}_{G_1}\otimes \mathcal{A}_{G_2}$ whenever $\mathcal{A}_{G_1}$ or $\mathcal{A}_{G_2}$ is the group ring.
\end{lemm}
\begin{lemm}\label{citens}
In the conditions of Lemma~\ref{cigwr}, suppose that $|G:U|$ is a prime and there exists $X\in \mathcal{S}(\mathcal{A}_{G/L})$ outside $S$ with $|X|=1$. Then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
Let $X=\{x\}$ for some $x\in G/L$. Due to $G\in \mathcal{E}$, we conclude that $|\langle x \rangle|$ is prime. So $|\langle x \rangle \cap S|=1$ because $x$ lies outside $S$. Since $|G:U|$ is a prime, $G/L=\langle x \rangle \times S$. Note that $\mathcal{A}_{\langle x \rangle}=\mathbb{Z}\langle x \rangle$. Therefore
$$\mathcal{A}_{G/L}=\mathbb{Z}\langle x \rangle \otimes \mathcal{A}_S$$
by Lemma~\ref{tenspr}.
Let $\varphi \in \aut_{S}(\mathcal{A}_S)$. Define $\psi \in \aut(G/L)$ in the following way:
$$\psi^{S}=\varphi,~x^\psi=x.$$
Then $\psi \in \aut_{G/L}(\mathcal{A}_{G/L})$ because $\mathcal{A}_{G/L}=\mathbb{Z}\langle x \rangle \otimes \mathcal{A}_S$. We obtain that $\aut_{G/L}(\mathcal{A}_{G/L})^S \geq \aut_S(\mathcal{A}_S)$, and hence $\aut_{G/L}(\mathcal{A}_{G/L})^S=\aut_S(\mathcal{A}_S)$. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{cigwr}. The lemma is proved.
\end{proof}
\section{$p$-$S$-rings}
Let $p$ be a prime. An $S$-ring $\mathcal{A}$ over a $p$-group $G$ is called a \emph{$p$-$S$-ring} if every basic set of $\mathcal{A}$ has a $p$-power size. Clearly, if $|G|=p$ then $\mathcal{A}=\mathbb{Z}G$. In the next three lemmas $G$ is a $p$-group and $\mathcal{A}$ is a $p$-$S$-ring over $G$.
\begin{lemm}\label{pextension}
If $\mathcal{B}\geq \mathcal{A}$ then $\mathcal{B}$ is a $p$-$S$-ring.
\end{lemm}
\begin{proof}
The statement of the lemma follows from~\cite[Theorem~1.1]{Pon}.
\end{proof}
\begin{lemm}\label{psection}
Let $S=U/L$ be an $\mathcal{A}$-section of $G$. Then $\mathcal{A}_S$ is a $p$-$S$-ring.
\end{lemm}
\begin{proof}
From Lemma~\ref{intersection} it follows that for every $X\in \mathcal{S}(\mathcal{A})$ the number $\lambda=|X \cap Lx|$ does not depend on $x\in X$. So $\lambda$ divides $|X|$ and hence $\lambda$ is a $p$-power. Let $\pi:G\rightarrow G/L$ be the canonical epimorphism. Note that $|\pi(X)|=|X|/\lambda$ and hence $|\pi(X)|$ is a $p$-power. Therefore every basic set of $\mathcal{A}_S$ has a $p$-power size. Thus, $\mathcal{A}_S$ is a $p$-$S$-ring. The lemma is proved.
\end{proof}
\begin{lemm}\cite[Proposition~2.13]{FK}\label{psring}
The following statements hold:
$(1)$ $|\O_{\theta}(\mathcal{A})|>1$;
$(2)$ there exists a chain of $\mathcal{A}$-subgroups $\{e\}=G_0<G_1<\ldots G_s=G$ such that $|G_{i+1}:G_i|=p$ for every $i\in \{0,\ldots,s-1\}$.
\end{lemm}
\begin{lemm}\label{minpring}
Let $G$ be an abelian group, $K \in \Sup_2^{\min}(G_\r)$, and $\mathcal{A}=V(K, G)$. Suppose that $H$ is an $\mathcal{A}$-subgroup of $G$ such that $G/H$ is a $p$-group for some prime $p$. Then $\mathcal{A}_{G/H}$ is a $p$-S-ring.
\end{lemm}
\begin{proof}
The statement of the lemma follows from~\cite[Lemma~5.2]{KM}.
\end{proof}
\section{$S$-rings over over an abelian group of non-powerful order}
A number $n$ is called \emph{powerful} if $p^2$ divides $n$ for every prime divisor $p$ of $n$. From now throughout this subsection $G=H\times P$, where $H$ is an abelian group and $P\cong C_p$, where $p$ is a prime coprime to $|H|$. Clearly, $|G|$ is non-powerful. Let $\mathcal{A}$ be an $S$-ring over $G$, $H_1$ a maximal $\mathcal{A}$-subgroup contained in $H$, and $P_1$ the least $\mathcal{A}$-subgroup containing $P$. Note that $H_1P_1$ is an $\mathcal{A}$-subgroup.
\begin{lemm}\cite[Lemma~6.3]{KR2}\label{nonpower2}
In the above notations, if $H_1 \neq (H_1P_1)_{p'}$, the Hall $p'$-subgroup of $H_1P_1$, then $\mathcal{A}_{H_1P_1}=\mathcal{A}_{H_1} \star \mathcal{A}_{P_1}$.
\end{lemm}
\begin{lemm}\cite[Proposition 15]{MS}\label{nonpower3}
In the above notations, if $\mathcal{A}_{H_1P_1/H_1}\cong \mathbb{Z}C_p$ then $\mathcal{A}_{H_1P_1}=\mathcal{A}_{H_1} \star \mathcal{A}_{P_1}$.
\end{lemm}
\begin{lemm}\cite[Lemma 6.2]{EKP}\label{nonpower4}
In the above notations, suppose that $H_1<H$. Then one of the following statements holds:
$(1)$ $\mathcal{A}=\mathcal{A}_{H_1}\wr \mathcal{A}_{G/H_1}$ with $\rk(\mathcal{A}_{G/H_1})=2$;
$(2)$ $\mathcal{A}=\mathcal{A}_{H_1P_1}\wr_S \mathcal{A}_{G/P_1}$, where $S=H_1P_1/P_1$ and $P_1<G$.
\end{lemm}
\section{$S$-rings over $C_2^n$, $n\leq 5$}
All $S$-rings over the groups $C_2^n$, where $n\leq 5$, were enumerated with the help of the GAP package COCO2P~\cite{GAP}. The lists of all $S$-rings over these groups are available on the web-page~\cite{Reich} (see also~\cite{Ziv}). The next lemma is an immediate consequence of the above computational results (see also~\cite[Theorem~1.2]{EKP}).
\begin{lemm}\label{schurian}
Every $S$-ring over $C_2^n$, where $n\leq 5$, is schurian.
\end{lemm}
The main goal of this section is to describe $2$-$S$-rings over $C_2^n$, where $n\leq 5$, using computational results and to check that all $S$-rings over the above groups are $\CI$-$S$-rings. From now until the end of the section $G$ is an elementary abelian $2$-group of rank~$n$ and $\mathcal{A}$ is a $2$-$S$-ring over $G$.
\begin{lemm}\label{p3}
Let $n\leq 3$. Then $\mathcal{A}$ is cyclotomic. Moreover, $\mathcal{A}$ is Cayley minimal except for the case when $n=3$ and $\mathcal{A} \cong \mathbb{Z}C_2 \wr \mathbb{Z}C_2\wr \mathbb{Z}C_2$.
\end{lemm}
\begin{proof}
The first part of the lemma follows from~\cite[Lemma~5.2]{KR2}; the second part follows from~\cite[Lemma~5.3]{KR2}.
\end{proof}
Analyzing the lists of all $S$-rings over $C_2^4$ and $C_2^5$ available on the web-page~\cite{Reich}, we conclude that up to isomorphism there are exactly nineteen $2$-$S$-rings over $G$ if $n=4$ and there are exactly one hundred $2$-$S$-rings over $G$ if $n=5$. It can can be established by inspecting the above $2$-$S$-rings one after the other that there are exactly fifteen decomposable and four indecomposable $2$-$S$-rings over $G$ if $n=4$ and there are exactly ninety six decomposable and four indecomposable $2$-$S$-rings over $G$ if $n=5$.
\begin{lemm}\label{p45}
Let $n\in\{4,5\}$ and $\mathcal{A}$ indecomposable. Then $\mathcal{A}$ is normal. If in addition $n=5$ then $\mathcal{A}\cong \mathbb{Z}C_2\otimes \mathcal{A}^{\prime}$, where $\mathcal{A}^{\prime}$ is indecomposable $2$-$S$-ring over $C_2^4$.
\end{lemm}
\begin{proof}
Let $n=4$. One can compute $|\aut(\mathcal{A})|$ and $|N_{\aut(\mathcal{A})}(G_{\r})|$ with using the GAP package COCO2P~\cite{GAP}. It turns out that for each of four indecomposable $2$-$S$-rings over $G$ the equality
$$|\aut(\mathcal{A})|=|N_{\aut(\mathcal{A})}(G_{\r})|$$
is attained. So every indecomposable $2$-$S$-ring over $G$ is normal whenever $n=4$.
Let $n=5$. The straightforward check for each of four indecomposable $2$-$S$-rings over $G$ yields that $\mathcal{A}=\mathcal{A}_H\otimes \mathbb{Z}L$, where $H\cong C_2^4$, $L\cong C_2$, and $\mathcal{A}_H$ is indecomposable $2$-$S$-ring. Clearly, $\mathbb{Z}L$ is normal. By the above paragraph, $\mathcal{A}_H$ is normal. Since $\aut(\mathcal{A})=\aut(\mathcal{A}_H)\times \aut(\mathcal{A}_L)$, we obtain that $\mathcal{A}$ is normal. The lemma is proved.
\end{proof}
Note that if $p>2$ then Lemma~\ref{p45} does not hold. In fact, if $p>2$ then there exists an indecomposable $p$-$S$-ring over $C_p^5$ which is not normal (see~\cite[Lemma~6.4]{FK}).
\begin{lemm}\label{pnorm}
Let $n\leq 5$. Then $\mathcal{A}$ is normal whenever one of the following statements holds:
$(1)$ $\mathcal{A}$ is indecomposable;
$(2)$ $|G:\O_{\theta}(\mathcal{A})|=2$;
$(3)$ $n=4$ and $\mathcal{A}\cong (\mathbb{Z}C_2 \wr \mathbb{Z}C_2)\otimes (\mathbb{Z}C_2 \wr \mathbb{Z}C_2)$.
\end{lemm}
\begin{proof}
If $n\leq 3$ and $\mathcal{A}$ is indecomposable then $\mathcal{A}=\mathbb{Z}G$ by~\cite[Lemma~5.2]{KR2}. Clearly, in this case $\mathcal{A}$ is normal. If $n\in\{4,5\}$ and $\mathcal{A}$ is indecomposable then $\mathcal{A}$ is normal by Lemma~\ref{p45}. There are exactly~$n-1$ $2$-$S$-rings over $G$ for which Statement~2 of the lemma holds. For every $\mathcal{A}$ isomorphic to one of these $2$-$S$-rings and for $\mathcal{A}\cong (\mathbb{Z}C_2 \wr \mathbb{Z}C_2)\otimes (\mathbb{Z}C_2 \wr \mathbb{Z}C_2)$ one can compute $|\aut(\mathcal{A})|$ and $|N_{\aut(\mathcal{A})}(G_{\r})|$ with using the GAP package COCO2P~\cite{GAP}. It turns out that in each case the equality $|\aut(\mathcal{A})|=|N_{\aut(\mathcal{A})}(G_{\r})|$ holds and hence $\mathcal{A}$ is normal. The lemma is proved.
\end{proof}
\begin{lemm}\label{p4cyclotom}
Let $n=4$. Then $\mathcal{A}$ is cyclotomic.
\end{lemm}
\begin{proof}
If $\mathcal{A}$ is decomposable then $\mathcal{A}$ is cyclotomic by~\cite[Lemma~5.6]{KR2}. If $\mathcal{A}$ is indecomposable then $\mathcal{A}$ is normal by Lemma~\ref{p45}. This implies that
$$\aut(\mathcal{A})_e=N_{\aut(\mathcal{A})}(G_{\r})_e\leq \aut(G).$$
The $S$-ring $\mathcal{A}$ is schurian by Lemma~\ref{schurian}. So from Eq.~(2) it follows that $\mathcal{A}=V(\aut(\mathcal{A}),G)$ and hence $\mathcal{A}=\cyc(\aut(\mathcal{A})_e,G)$. The lemma is proved.
\end{proof}
\begin{lemm}\label{p5cyclotom}
Let $n=5$. Suppose that $\mathcal{A}$ is decomposable and $|\O_{\theta}(\mathcal{A})|=8$. Then $\mathcal{A}$ is cyclotomic.
\end{lemm}
\begin{proof}
Let $\mathcal{A}$ be the nontrivial $S$-wreath product for some $\mathcal{A}$-section $S=U/L$. Note that $|U|\leq 16$, $|G/L|\leq 16$, and $|S|\leq 8$. The $S$-rings $\mathcal{A}_{U}$, $\mathcal{A}_{G/L}$, and $\mathcal{A}_{S}$ are $2$-$S$-rings by Lemma~\ref{psection}. So each of these $S$-rings is cyclotomic by Lemma~\ref{p3} whenever the order of the corresponding group is at most~$8$ and by Lemma~\ref{p4cyclotom} otherwise. Since $|\O_\theta(\mathcal{A})|=8$, we conclude that $|S|\leq 4$ or $|S|=8$ and $|\O_\theta(\mathcal{A}_{S})|\geq 4$. In both cases $\mathcal{A}_{S}$ is Cayley minimal by Lemma~\ref{p3}. This implies that
$$\aut_{U}(\mathcal{A}_{U})^{S}=\aut_{G/L}(\mathcal{A}_{G/L})^{S}=\aut_{S}(\mathcal{A}_{S}).$$
Now from~\cite[Lemma~4.3]{KR2} it follows that $\mathcal{A}$ is cyclotomic. The lemma is proved.
\end{proof}
\begin{lemm}\label{p5decomp1}
Let $n=5$. Suppose that $\mathcal{A}$ is decomposable and $|\O_\theta(\mathcal{A})|=4$. Then one of the following statements holds:
$(1)$ there exists an $\mathcal{A}$-subgroup $L\leq \O_\theta(\mathcal{A})$ of order~$2$ such that $\mathcal{A}=\mathbb{Z}\O_\theta(\mathcal{A})\wr_{S} \mathcal{A}_{G/L}$, where $S=\O_\theta(\mathcal{A})/L$;
$(2)$ $|\aut_G(\mathcal{A})|\geq |\aut_U(\mathcal{A}_U)|$ for every $\mathcal{A}$-subgroup $U$ with $|U|=16$ and $U\geq \O_\theta(\mathcal{A})$;
$(3)$ $\mathcal{A}$ is normal;
$(4)$ there exist an $\mathcal{A}$-subgroup $L\leq \O_\theta(\mathcal{A})$ and $X\in \mathcal{S}(\mathcal{A})$ such that $|L|=|X|=2$, $L\neq \rad(X)$, and $\mathcal{A}_{G/L}$ is normal.
\end{lemm}
\begin{proof}
There are exactly forty five decomposable $2$-$S$-rings over $G$ whose thin radical has order~$4$. This can be checked by inspecting all $2$-$S$-rings over $G$ one after the other. Let $\mathcal{A}$ be one of them and $R=\O_\theta(\mathcal{A})$. The straightforward check of basic sets of each of the above forty five $2$-$S$-rings shows that Statement~1 of the lemma holds for twenty six of them. The analysis of basic sets of the remaining nineteen $2$-$S$-rings implies that ten of them have an $\mathcal{A}$-subgroup $L\leq R$ and $X\in \mathcal{S}(\mathcal{A})$ satisfying the following: (1) $|L|=|X|=2$; (2) $L\neq \rad(X)$; (3) one of the conditions from Lemma~\ref{pnorm} holds for $\mathcal{A}_{G/L}$. We conclude that $\mathcal{A}_{G/L}$ is normal and hence Statement~4 of the lemma holds for these ten $2$-$S$-rings.
It remains to consider nine $2$-$S$-rings for which neither Statement~1 nor Statement~4 of the lemma holds. Let $U$ be an $\mathcal{A}$-subgroup with $|U|=16$ and $U\geq R$. Statement~2 of Lemma~\ref{psring} yields that there exists an $\mathcal{A}_U$-subgroup $U_1$ such that
$$|U_1|=8~\text{and}~R<U_1<U.$$
Let $X_1$ be a basic set of $\mathcal{A}$ inside $U_1\setminus R$ of the least possible size and $X_2$ a basic set of $\mathcal{A}$ inside $U\setminus U_1$ of the least possible size. Clearly, $|X_1|\leq 4$ and $|X_2|\leq 8$. Choose $x_1\in X_1$ and $x_2\in X_2$. From the choice of $x_1$ and $x_2$ it follows that $\langle R, x_1, x_2\rangle=U$. So if $\varphi\in \aut_U(\mathcal{A}_U)$, $x_1^{\varphi}=x_1$, and $x_2^{\varphi}=x_2$ then $\varphi$ is trivial. This implies that
$$|\aut_U(\mathcal{A}_U)|\leq |X_1||X_2|\leq 32.~\eqno(5)$$
One can compute $|\aut_G(\mathcal{A})|=|N_{\aut(\mathcal{A})}(G_{\r})_e|$ with using the GAP package COCO2P~\cite{GAP}. The inequality $|\aut_G(\mathcal{A})|\geq 32$ holds for four of the remaining $2$-$S$-rings. Due to Eq.~(5), Statement~2 of the lemma holds for them. For three of the remaining $2$-$S$-rings, we have $|\aut_G(\mathcal{A})|=16$. However, in this situation there are no basic sets of size~$4$ and hence $|X_1|=2$ or there are no basic sets of size $8$ and hence $|X_2|\leq 4$. In both case Eq.~(5) implies that $|\aut_U(\mathcal{A}_U)|\leq 16$ and Statement~2 of the lemma holds.
Now it remains to consider two $2$-$S$-rings for which $|\aut_G(\mathcal{A})|=8$. One of these $2$-$S$-rings does not have a basic set of size~$8$ and every of its $\mathcal{A}$-subgroups of order~$16$ contains a basic set of size~$2$. So $|X_1||X_2|\leq 8$. In view of Eq.~(5), Statement~2 of the lemma holds for this $2$-$S$-ring. For the other of these $2$-$S$-rings computer calculations made with the help the GAP package COCO2P~\cite{GAP} imply that $|\aut(\mathcal{A})_e|=|\aut_G(\mathcal{A})|=8$ and hence it is normal, i.e. Statement~3 of the lemma holds for it. The lemma is proved.
\end{proof}
\begin{lemm}\label{p5decomp2}
Let $n=5$. Suppose that $\mathcal{A}$ is decomposable, $|\O_\theta(\mathcal{A})|=2$, and there exists $X\in \mathcal{S}(\mathcal{A})$ with $|X|>1$ and $|\rad(X)|=1$. Then $|X|=4$ and one of the following statements holds:
$(1)$ $\mathcal{A}\cong \mathcal{B} \wr \mathbb{Z}C_2$, where $\mathcal{B}$ is a $2$-$S$-ring over $C_2^4$;
$(2)$ $|\aut_G(\mathcal{A})|\geq |\aut_U(\mathcal{A}_U)|$ for every $\mathcal{A}$-subgroup $U$ with $|U|=16$;
$(3)$ there exists an $\mathcal{A}$-subgroup $L$ such that $|L|\in\{2,4\}$ and $\mathcal{A}_{G/L}$ is normal.
\end{lemm}
\begin{proof}
There are exactly twenty nine decomposable $2$-$S$-rings over $G$ whose thin radical has order~$2$ (in fact, every $2$-$S$-ring with the thin radical of order~$2$ is decomposable). This can be verified by inspecting all $2$-$S$-rings over $G$ one after the other. Only ten of these twenty nine $2$-$S$-rings have a basic set with the trivial radical and each of such basic sets with the trivial radical has size~$4$. Let $\mathcal{A}$ be one of the ten above $2$-$S$-rings. From the direct check it follows that Statement~1 of the lemma holds for two of these ten $2$-$S$-rings. The analysis of basic sets of the remaining eight $2$-$S$-rings yields that six of them have an $\mathcal{A}$-subgroup $L$ satisfying the following: (1) $|L|\in\{2,4\}$; (2) one of the conditions from Lemma~\ref{pnorm} holds for $\mathcal{A}_{G/L}$. We conclude $\mathcal{A}_{G/L}$ is normal and hence Statement~3 of the lemma holds for these six $2$-$S$-rings.
It remains to consider two $2$-$S$-rings for which neither Statement~1 nor Statement~3 of the lemma holds. Each of these two $2$-$S$-rings has exactly three distinct $\mathcal{A}$-subgroups of order~$16$, say $U_1$, $U_2$, and $U_3$. Computations made by using the GAP package COCO2P yield that for one of them the following holds:
$$|\aut_G(\mathcal{A})|=64,~|\aut_{U_i}(\mathcal{A}_{U_i})|\in \{8,64\}~\text{for}~i\in\{1,2,3\};$$
and for the other of them the following holds:
$$|\aut_G(\mathcal{A})|=32,~|\aut_{U_i}(\mathcal{A}_{U_i})|\in \{4,32\}~\text{for}~i\in\{1,2,3\}.$$
In both cases Statement~2 of the lemma holds. The lemma is proved.
\end{proof}
\begin{lemm}\label{2cigwr}
Let $D\in \mathcal{E}$ such that every $S$-ring over a proper section of $D$ is $\CI$, $\mathcal{D}$ an $S$-ring over $D$, and $S=U/L$ a $\mathcal{D}$-section. Suppose that $\mathcal{D}$ is the nontrivial $S$-wreath product. Then $\mathcal{D}$ is a $\CI$-$S$-ring whenever $D/L\cong C_2^k$ for some $k\leq 4$ and $\mathcal{D}_{D/L}$ is a $2$-$S$-ring.
\end{lemm}
\begin{proof}
The $S$-ring $\mathcal{D}_{D/L}$ is cyclotomic by Lemma~\ref{p3} whenever $|D/L|\leq 8$ and by Lemma~\ref{p4cyclotom} whenever $|D/L|=16$. The $S$-ring $\mathcal{D}_S$ is a $2$-$S$-ring by Lemma~\ref{psection}. If $\mathcal{D}_S\ncong \mathbb{Z}C_2\wr \mathbb{Z}C_2\wr \mathbb{Z}C_2$ then $\mathcal{D}_S$ is Cayley minimal by Lemma~\ref{p3}. The $S$-rings $\mathcal{D}_U$ and $\mathcal{D}_{D/L}$ are $\CI$-$S$-rings by the assumption of the lemma. So $\mathcal{D}$ is a $\CI$-$S$-ring by Lemma~\ref{cicayleymin}.
Assume that
$$\mathcal{D}_S\cong \mathbb{Z}C_2\wr \mathbb{Z}C_2\wr \mathbb{Z}C_2.$$
In this case $|D/L|=16$, $|S|=8$, and there exists the least $\mathcal{D}_S$-subgroup $A$ of $S$ of order~$2$. Every basic set of $\mathcal{D}_{D/L}$ outside $S$ is contained in an $S$-coset because $\mathcal{D}_{(D/L)/S}\cong \mathbb{Z}C_2$. So $\rad(X)$ is an $\mathcal{D}_S$-subgroup for every $X\in \mathcal{S}(\mathcal{D}_{D/L})$ outside $S$. If $|\rad(X)|>1$ for every $X\in \mathcal{S}(\mathcal{D}_{D/L})$ outside $S$ then $\mathcal{D}_{D/L}$ is the $S/A$-wreath product because $A$ is the least $\mathcal{D}_S$-subgroup. This implies that $\mathcal{D}$ is the $U/\pi^{-1}(A)$-wreath product, where $\pi:D\rightarrow D/L$ is the canonical epimorphism. One can see that $|D/\pi^{-1}(A)|\leq 8$ and $|U/\pi^{-1}(A)|\leq 4$. The $S$-rings $\mathcal{D}_{D/\pi^{-1}(A)}$ and $\mathcal{D}_{U/\pi^{-1}(A)}$ are $2$-$S$-rings by Lemma~\ref{psection}. The $S$-ring $\mathcal{D}_{D/\pi^{-1}(A)}$ is cyclotomic by Lemma~\ref{p3} and the $S$-ring $\mathcal{D}_{U/\pi^{-1}(A)}$ is Cayley minimal by Lemma~\ref{p3}. The $S$-rings $\mathcal{D}_U$ and $\mathcal{D}_{D/\pi^{-1}(A)}$ are $\CI$-$S$-rings by the assumption of the lemma. Thus, $\mathcal{D}$ is a $\CI$-$S$-ring by Lemma~\ref{cicayleymin}.
Suppose that there exists a basic set $X$ of $\mathcal{D}_{D/L}$ outside $S$ with $|\rad(X)|=1$. If $\mathcal{D}_{D/L}$ is decomposable then
$$\aut_{D/L}(\mathcal{D}_{D/L})^S=\aut_S(\mathcal{D}_S)$$
by~\cite[Lemma~5.8]{KR2}. Therefore $\mathcal{D}$ is a $\CI$-$S$-ring by Lemma~\ref{cigwr}.
If $\mathcal{D}_{D/L}$ is indecomposable then $\mathcal{D}_{D/L}$ is normal by Lemma~\ref{p45}. So all conditions of Lemma~\ref{kerci} hold for $\mathcal{D}$. Thus, $\mathcal{D}$ is a $\CI$-$S$-ring. The lemma is proved.
\end{proof}
From the results obtained in~\cite{AlN, CLi} it follows that the group $C_2^n$ is a $\DCI$-group for $n\leq 5$. However, this does not imply that every $S$-ring over $C_2^n$, where $n\leq 5$, is a $\CI$-$S$-ring (see Remark~1). Below we check that all $S$-rings over the above groups are $\CI$-$S$-rings.
\begin{lemm}\label{everyci}
Let $n\leq 5$. Then every $S$-ring over $G$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
Every $S$-ring over $G$ is schurian by Lemma~\ref{schurian}. So to prove the lemma, it is sufficient to prove that $\mathcal{B}=V(K,G)$ is a $\CI$-$S$-ring for every $K\in\Sup_2^{\rm min}(G_\r)$ (see Remark~1). The $S$-ring $\mathcal{B}$ is a $2$-$S$-ring by Lemma~\ref{minpring}. If $n\leq 4$ then $\mathcal{B}$ is $\CI$ by~\cite[Lemma~5.7]{KR2}. Thus, if $n=4$ then the statement of the lemma holds.
Let $n=5$. Suppose that $\mathcal{B}$ is indecomposable. Then the second part of Lemma~\ref{p45} implies $\mathcal{B}\cong \mathbb{Z}C_2\otimes \mathcal{B}^{\prime}$, where $\mathcal{B}^{\prime}$ is indecomposable $2$-$S$-ring over $C_2^4$. Since $\mathcal{B}$ is schurian by Lemma~\ref{schurian} and every $S$-ring over an elementary abelian group of rank at most~$4$ is $\CI$ by the above paragraph, we conclude that $\mathcal{B}$ is a $\CI$-$S$-ring by Lemma~\ref{cistar}.
Now suppose that $\mathcal{B}$ is decomposable, i.e. $\mathcal{B}$ is the nontrivial $S=U/L$-wreath product for some $\mathcal{B}$-section $S=U/L$. Clearly, $|G/L|\leq 16$. The $S$-ring $\mathcal{B}_{G/L}$ is a $2$-$S$-ring by Lemma~\ref{psection}. Since every $S$-ring over an elementary abelian group of rank at most~$4$ is $\CI$, $\mathcal{B}$ is a $\CI$-$S$-ring by Lemma~\ref{2cigwr}. The lemma is proved.
\end{proof}
\section{Proof of Theorem~\ref{main}}
Let $G=H \times P$, where $H \cong C_2^5$ and $P \cong C_p$, where $p$ is a prime. These notations are valid until the end of the paper. If $p=2$ then $G$ is not a $\DCI$-group by~\cite{Now}. So in view of Lemma~\ref{cimin}, to prove Theorem~\ref{main}, it is sufficient to prove the following theorem.
\begin{theo}\label{main2}
Let $p$ be an odd prime and $K \in \Sup_2^{\rm min}(G_\r)$. Then $\mathcal{A}=V(K, G)$ is a $\CI$-$S$-ring.
\end{theo}
The proof of Proposition~\ref{main2} will be given in the end of the section. We start with the next lemma concerned with proper sections of $G$.
\begin{lemm}\label{subgroup}
Let $S$ be a section of $G$ such that $S\neq G$. Then every schurian $S$-ring over $S$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
If $S\cong C_2^n$ for some $n\leq 5$ then we are done by Lemma~\ref{everyci}. Suppose that $S\cong C_2^n \times C_p$ for some $n\leq 4$. Then the statement of the lemma follows from~\cite[Remark~3.4]{KR2} whenever $n\leq 3$ and from~\cite[Remark~3.4,~Theorem~7.1]{KR2} whenever $n=4$. The lemma is proved.
\end{proof}
A key step towards the proof of Theorem~\ref{main2} is the following lemma.
\begin{lemm}\label{main3}
Let $\mathcal{A}$ be an $S$-ring over $G$ and $U$ an $\mathcal{A}$-subgroup with $U\geq P$. Suppose that $P$ is an $\mathcal{A}$-subgroup, $\mathcal{A}$ is the nontrivial $S$-wreath product, where $S=U/P$, $|S|=16$, and $\mathcal{A}_{G/P}$ is a $2$-$S$-ring. Then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{lemm}\label{proof1}
In the conditions of Lemma~\ref{main3}, suppose that $S$ has a gwr-complement with respect to $\mathcal{A}_{G/P}$. Then $\mathcal{A}$ is a a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
The condition of the lemma implies that there exists an $\mathcal{A}_{G/P}$-subgroup $A$ such that $\mathcal{A}_{G/P}$ is the nontrivial $S/A$-wreath product. This means that $\mathcal{A}$ is the nontrivial $U/\pi^{-1}(A)$-wreath product, where $\pi: G\rightarrow G/P$ is the canonical epimorphism. Note that $|G/\pi^{-1}(A)|\leq 16$ and $\mathcal{A}_{G/\pi^{-1}(A)}\cong \mathcal{A}_{(G/P)/A}$ is a $2$-$S$-ring by Lemma~\ref{psection}. Therefore $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{2cigwr}. The lemma is proved.
\end{proof}
\begin{lemm}\label{proof2}
In the conditions of Lemma~\ref{main3}, suppose that $S$ does not have a gwr-complement with respect to $\mathcal{A}_{G/P}$. Then
$$|\aut_{G/P}(\mathcal{A}_{G/P})^S|=|\aut_{G/P}(\mathcal{A}_{G/P})|.$$
\end{lemm}
\begin{proof}
To prove the lemma it is sufficient to prove that the group
$$(\aut_{G/P}(\mathcal{A}_{G/P}))_S=\{\varphi\in \aut_{G/P}(\mathcal{A}_{G/P}):~\varphi^S=\id_S\}$$ is trivial. Let $\varphi\in (\aut_{G/P}(\mathcal{A}_{G/P}))_S$. Put $\mathcal{C}=\cyc(\langle \varphi \rangle, G/P)$. Clearly, $\langle \varphi \rangle\leq \aut(\mathcal{A}_{G/P})$. So from Eqs.~(1) and~(2) it follows that $\mathcal{C}\geq \mathcal{A}_{G/P}$. Lemma~\ref{pextension} yields that $\mathcal{C}$ is a $2$-$S$-ring. Since $\varphi^S=\id_S$, we conclude that $\O_\theta(\mathcal{C})\geq S$.
If $\mathcal{C}\neq \mathbb{Z}(G/P)$ then $\O_\theta(\mathcal{C})=S$. Therefore $\mathcal{C}=\mathbb{Z}S\wr_{S/A} \mathbb{Z}((G/P)/A)$ for some $\mathcal{C}$-subgroup $A$ by Statement~1 of~\cite[Proposition~4.3]{KR1}. This implies that $\mathcal{A}_{G/P}=\mathcal{A}_S\wr_{S/A} \mathcal{A}_{((G/P)/A)}$ because $\mathcal{C}\geq \mathcal{A}_{G/P}$ and $S$ is both $\mathcal{A}_{G/P}$, $\mathcal{C}$-subgroup. We obtain a contradiction with the assumption of the lemma. Thus, $\mathcal{C}=\mathbb{Z}(G/P)$ and hence $\varphi$ is trivial. So the group $(\aut_{G/P}(\mathcal{A}_{G/P}))_S$ is trivial. The lemma is proved.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{main3}.]
If $\mathcal{A}_{G/P}$ is indecomposable then $\mathcal{A}_{G/P}$ is normal by Lemma~\ref{p45}. So $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{kerci}. Further we assume that $\mathcal{A}_{G/P}$ is decomposable. Due to Lemma~\ref{proof1}, we may assume also that
$$S~\text{does not have a gwr-complement with respect to}~\mathcal{A}_{G/P}.~\eqno(6)$$
If there exists $X\in \mathcal{S}(\mathcal{A}_{G/P})$ outside $S$ with $|X|=1$ then $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{citens}. So we may assume that
$$\O_\theta(\mathcal{A}_{G/P})\leq S.~\eqno(7)$$
Note that $|\O_\theta(\mathcal{A}_{G/P})|>1$ by Statement~1 of Lemma~\ref{psring} and $|\O_\theta(\mathcal{A}_{G/P})|\leq 16$ by Eq.~(7). So $|\O_\theta(\mathcal{A}_{G/P})|\in \{2,4,8,16\}$. We divide the rest of the proof into four cases depending on $|\O_\theta(\mathcal{A}_{G/P})|$.
\\
\\
\noindent~\textbf{Case~1: $|\O_\theta(\mathcal{A}_{G/P})|=16$.}
\\
\\
Due to Eq.~(7), we conclude that $\mathcal{A}_S=\mathbb{Z}S$. So $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{trivial}.
\\
\\
\noindent~\textbf{Case~2: $|\O_\theta(\mathcal{A}_{G/P})|=8$.}
\\
\\
Since $\mathcal{A}_{G/P}$ is decomposable, Lemma~\ref{p5cyclotom} implies that $\mathcal{A}_{G/P}$ is cyclotomic. The $S$-ring $\mathcal{A}_S$ is a $2$-$S$-ring by Lemma~\ref{psection}. In view of Eq.~(7), we obtain that $|\O_\theta(\mathcal{A}_{S})|=8$. So Statement~2 of~\cite[Proposition~4.3]{KR1} yields that the $S$-ring $\mathcal{A}_S$ is Cayley minimal. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{cicayleymin}.
\\
\\
\noindent~\textbf{Case~3: $|\O_\theta(\mathcal{A}_{G/P})|=4$.}
\\
\\
In this case one of the statements of Lemma~\ref{p5decomp1} holds for $\mathcal{A}_{G/P}$. If Statement~1 of Lemma~\ref{p5decomp1} holds for $\mathcal{A}_{G/P}$ then we obtain a contradiction with Eq.~(6).
If Statement~2 of Lemma~\ref{p5decomp1} holds for $\mathcal{A}_{G/P}$ then $|\aut_{G/P}(\mathcal{A}_{G/P})|\geq |\aut_S(\mathcal{A}_S)|$. From Lemma~\ref{proof2} it follows that $|\aut_{G/P}(\mathcal{A}_{G/P})^S|=|\aut_{G/P}(\mathcal{A}_{G/P})|$ and hence
$$|\aut_{G/P}(\mathcal{A}_{G/P})^S|\geq |\aut_S(\mathcal{A}_S)|.$$
Since $\aut_{G/P}(\mathcal{A}_{G/P})^S\leq \aut_S(\mathcal{A}_S)$, we conclude that $\aut_{G/P}(\mathcal{A}_{G/P})^S=\aut_S(\mathcal{A}_S)$. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{cigwr}.
If Statement~3 of Lemma~\ref{p5decomp1} holds for $\mathcal{A}_{G/P}$ then $\mathcal{A}_{G/P}$ is normal. In this case $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{kerci}.
Suppose that Statement~4 of Lemma~\ref{p5decomp1} holds for $\mathcal{A}_{G/P}$, i.e. there exists an $\mathcal{A}_{G/P}$-subgroup $A\leq \O_\theta(\mathcal{A}_{G/P})$ of order~$2$ and $X=\{x_1,x_2\}\in \mathcal{S}(\mathcal{A}_{G/P})$ such that $\mathcal{A}_{(G/P)/A}$ is normal and $A\neq \rad(X)$. Let $L=\pi^{-1}(A)$, where $\pi:G\rightarrow G/P$ is the canonical epimorphism, and $\mathcal{B}=V(N,G)$, where $N=\aut(\mathcal{A})_{G/L}G_{\r}$.
Prove that $\mathcal{B}$ is a $\CI$-$S$-ring. Lemma~\ref{kerwreath} implies that $\mathcal{B}$ is the $S$-wreath product. From Eqs.~(1) and~(2) it follows that $\mathcal{B}\geq \mathcal{A}$. So $\mathcal{B}_{G/P}\geq \mathcal{A}_{G/P}$ and hence $\mathcal{B}_{G/P}$ is a $2$-$S$-ring by Lemma~\ref{pextension}. We obtain that $\mathcal{B}$ and $U$ satisfy the conditions of Lemma~\ref{main3}.
One can see that $X$ is a $\mathcal{B}_{G/P}$-set and
$$\O_{\theta}(\mathcal{B}_{G/P})\geq \O_{\theta}(\mathcal{A}_{G/P})~\eqno(8)$$
because $\mathcal{B}_{G/P}\geq \mathcal{A}_{G/P}$. The definition of $\mathcal{B}$ yields that every basic set of $\mathcal{B}$ is contained in an $L$-coset and hence every basic set of $\mathcal{B}_{G/P}$ is contained in an $A$-coset. Therefore
$$\{x_1\},\{x_2\}\in \mathcal{S}(\mathcal{B}_{G/P})~\eqno(9)$$
because $X$ is a $\mathcal{B}_{G/P}$-set and $A\neq \rad(X)$. Now from Eqs.~(8) and~(9) it follows that
$$|\O_{\theta}(\mathcal{B}_{G/P})|\geq 8.~\eqno(10)$$
If $\mathcal{B}_{G/P}$ is indecomposable then $\mathcal{B}_{G/P}$ is normal by Lemma~\ref{p45} and hence $\mathcal{B}$ is $\CI$ by Lemma~\ref{subgroup} and Lemma~\ref{kerci}; if $S$ has a gwr-complement with respect to $\mathcal{B}_{G/P}$ then $\mathcal{B}$ is $\CI$ by Lemma~\ref{proof1}; if $\O_\theta(\mathcal{B}_{G/P})\nleq S$ then $\mathcal{B}$ is $\CI$ by Lemma~\ref{subgroup} and Lemma~\ref{citens}; otherwise $\mathcal{B}$ is $\CI$ by Eq.~(10) and one of the Cases~1 or~2.
Clearly, $\mathcal{A}_{G/L}\cong\mathcal{A}_{(G/P)/A}$ and hence $\mathcal{A}_{G/L}$ is normal. Also $\mathcal{A}_{G/L}$ is $\CI$ by Lemma~\ref{subgroup}. The $S$-ring $\mathcal{B}$ is $\CI$ by the above paragraph. Thus, $\mathcal{A}$ is $\CI$ by Lemma~\ref{minnorm}.
\\
\\
\noindent~\textbf{Case~4: $|\O_\theta(\mathcal{A}_{G/P})|=2$.}
\\
\\
Let $A=\O_\theta(\mathcal{A}_{G/P})$. Clearly, $A$ is the least $\mathcal{A}_{G/P}$-subgroup. If $|\rad(X)|>1$ for every $X\in \mathcal{S}(\mathcal{A}_{G/P})$ outside $S$ then $A\leq \rad(X)$ for every $X\in \mathcal{S}(\mathcal{A}_{G/P})$ outside $S$ and we obtain a contradiction with Eq.~(6). So there exists $X\in \mathcal{S}(\mathcal{A}_{G/P})$ outside $S$ with $|\rad(X)|=1$. From Eq.~(7) it follows that $|X|>1$. Lemma~\ref{p5decomp2} implies that $|X|=4$. The number $\lambda=|X \cap Ax|$ does not depend on $x\in X$ by Lemma~\ref{intersection}. If $\lambda=2$ then $A\leq \rad(X)$, a contradiction. Therefore
$$\lambda=1.~\eqno(11)$$
One of the statements of Lemma~\ref{p5decomp2} holds for $\mathcal{A}_{G/P}$. If Statement~1 of Lemma~\ref{p5decomp2} holds for $\mathcal{A}_{G/P}$ then there exists $Y\in \mathcal{S}(\mathcal{A}_{G/P})$ with $|Y|=16$ and $|\rad(Y)|=16$. Since $|S|=16$, we conclude that $Y$ lies outside $S$ and hence $Y=(G/P)\setminus S$. This means that $S$ is a gwr-complement to $S$ with respect to $\mathcal{A}_{G/P}$. However, this contradicts to Eq.~(6).
If Statement~2 of Lemma~\ref{p5decomp2} holds for $\mathcal{A}_{G/P}$ then $|\aut_{G/P}(\mathcal{A}_{G/P})|\geq |\aut_S(\mathcal{A}_S)|$. So Lemma~\ref{proof2} implies that $\aut_{G/P}(\mathcal{A}_{G/P})^S=\aut_S(\mathcal{A}_S)$. Therefore, $\mathcal{A}$ is $\CI$ by Lemma~\ref{subgroup} and Lemma~\ref{cigwr}
Suppose that Statement~3 of Lemma~\ref{p5decomp2} holds for $\mathcal{A}_{G/P}$, i.e. there exists an $\mathcal{A}_{G/P}$-subgroup $B$ such that $|B|\in\{2,4\}$ and $\mathcal{A}_{(G/P)/B}$ is normal. Let $L=\pi^{-1}(B)$, where $\pi:G\rightarrow G/P$ is the canonical epimorphism, and $\mathcal{B}=V(N,G)$, where $N=\aut(\mathcal{A})_{G/L}G_{\r}$.
Prove that $\mathcal{B}$ is a $\CI$-$S$-ring. As in Case~3, $\mathcal{B}$ is the $S$-wreath product by Lemma~\ref{kerwreath} and $\mathcal{B}\geq \mathcal{A}$ by Eqs.~(1) and~(2). So $\mathcal{B}_{G/P}\geq \mathcal{A}_{G/P}$ and hence $\mathcal{B}_{G/P}$ is a $2$-$S$-ring by Lemma~\ref{pextension}. Therefore $\mathcal{B}$ and $U$ satisfy the conditions of Lemma~\ref{main3}.
Note that $X$ is a $\mathcal{B}_{G/P}$-set and Eq.~(8) holds because $\mathcal{B}_{G/P}\geq \mathcal{A}_{G/P}$. By the definition of $\mathcal{B}$, every basic set of $\mathcal{B}$ is contained in an $L$-coset and hence every basic set of $\mathcal{B}_{G/P}$ is contained in a $B$-coset. The set $X$ is a $\mathcal{B}_{G/P}$-set with $|X|=4$ and $|\rad(X)|=1$. So there exists $X_1\in \mathcal{S}(\mathcal{B}_{G/P})$ such that
$$X_1\subset X~\text{and}~|X_1|\in\{1,2\}.$$
If $|X_1|=1$ then $X_1\subseteq \O_{\theta}(\mathcal{B}_{G/P})$. If $|X_1|=2$ then $X_1$ is a coset by a $\mathcal{B}_{G/P}$-subgroup $A_1$ of order~$2$. Clearly, $A_1\subseteq \O_{\theta}(\mathcal{B}_{G/P})$. In view of Eq.~(11), we have $A_1\neq A$. Thus, in both cases $\O_{\theta}(\mathcal{B}_{G/P})\nleq A$. Together with Eq.~(8) this implies that
$$|\O_{\theta}(\mathcal{B}_{G/P})|\geq 4.~\eqno(12)$$
If $\mathcal{B}_{G/P}$ is indecomposable then $\mathcal{B}_{G/P}$ is normal by Lemma~\ref{p45} and hence $\mathcal{B}$ is $\CI$ by Lemma~\ref{subgroup} and Lemma~\ref{kerci}; if $S$ has a gwr-complement with respect to $\mathcal{B}_{G/P}$ then $\mathcal{B}$ is $\CI$ by Lemma~\ref{proof1}; if $\O_\theta(\mathcal{B}_{G/P})\nleq S$ then $\mathcal{B}$ is $\CI$ by Lemma~\ref{subgroup} and Lemma~\ref{citens}; otherwise $\mathcal{B}$ is $\CI$ by Eq.~(12) and one of the Cases~1, 2, or~3.
The $S$-ring $\mathcal{A}_{G/L}$ is normal because it is isomorphic to $\mathcal{A}_{(G/P)/B}$. The $S$-rings $\mathcal{A}_{G/L}$ and $\mathcal{B}$ are $\CI$ by Lemma~\ref{subgroup} and the above paragraph respectively. Thus, $\mathcal{A}$ is $\CI$ by Lemma~\ref{minnorm}.
All cases were considered. The lemma is proved.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{main2}.]
Let $H_1$ be a maximal $\mathcal{A}$-subgroup contained in $H$ and $P_1$ the least $\mathcal{A}$-subgroup containing $P$.
\begin{lemm}\label{proof3}
If $H_1=H$ then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
The $S$-ring $\mathcal{A}_{G/H}$ is a $p$-$S$-ring over $G/H\cong C_p$ by Lemma~\ref{minpring}. So $\mathcal{A}_{G/H}\cong \mathbb{Z}C_p$. Clearly, $G=HP_1$. Therefore $\mathcal{A}=\mathcal{A}_{H} \star \mathcal{A}_{P_1}$ by Lemma~\ref{nonpower3}. Since $H$ and $P_1/(H\cap P_1)$ are proper sections of $G$, the $S$-rings $\mathcal{A}_H$ and $\mathcal{A}_{P_1/(H\cap P_1)}$ are $\CI$-$S$-ring by Lemma~\ref{subgroup}. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{cistar}. The lemma is proved.
\end{proof}
\begin{lemm}\label{proof4}
If $H_1<H$ and $H_1P_1=G$ then $\mathcal{A}$ is a $\CI$-$S$-ring.
\end{lemm}
\begin{proof}
Since $H_1 \neq (H_1P_1)_{p'}=H$, Lemma~\ref{nonpower2} implies that $\mathcal{A}=\mathcal{A}_{H_1} \star \mathcal{A}_{P_1}$. The $S$-rings $\mathcal{A}_{H_1}$ and $\mathcal{A}_{P_1/(H\cap P_1)}$ are $\CI$-$S$-ring by Lemma~\ref{subgroup} because $H_1$ and $P_1/(H_1\cap P_1)$ are proper sections of $G$. Therefore $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{cistar}. The lemma is proved.
\end{proof}
In view of Lemma~\ref{proof3}, we may assume that $H_1<H$. Then one of the statements of Lemma~\ref{nonpower4} holds for $\mathcal{A}$. If Statement~1 of Lemma~\ref{nonpower4} holds for $\mathcal{A}$ then
$$\mathcal{A}=\mathcal{A}_{H_1}\wr \mathcal{A}_{G/H_1},$$
where $\rk(\mathcal{A}_{G/H_1})=2$. If $H_1$ is trivial then $\rk(\mathcal{A})=2$ and obviously $\mathcal{A}$ is a $\CI$-$S$-ring. If $H_1$ is nontrivial then $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{trivial}.
Assume that Statement~2 of Lemma~\ref{nonpower4} holds for $\mathcal{A}$, i.e.
$$\mathcal{A}=\mathcal{A}_U\wr_S \mathcal{A}_{G/P_1},$$
where $U=H_1P_1$, $S=U/P_1$, and $P_1<G$. In view of Lemma~\ref{proof4}, we may assume that $H_1P_1<G$, i.e. $\mathcal{A}$ is the nontrivial $S$-wreath product. The group $G/P_1$ is a $2$-group of order at most~$32$ because $P_1\geq P$. So Lemma~\ref{minpring} implies that $\mathcal{A}_{G/P_1}$ is a $2$-$S$-ring. If $|G/P_1|\leq 16$ then $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{subgroup} and Lemma~\ref{2cigwr}. So we may assume that $|G/P_1|=32$. Clearly, in this case
$$P_1=P.$$
In view of Statement~2 of Lemma~\ref{psring}, we may assume also that
$$|S|=16.$$
Indeed, if $|S|<16$ then $S$ is contained in an $\mathcal{A}_{G/P}$-subgroup $S^{\prime}$ of order $16$ by Statement~2 of Lemma~\ref{psring}. Clearly, $\mathcal{A}=\mathcal{A}_{U^{\prime}}\wr_{S^{\prime}}\mathcal{A}_{G/P}$, where $U^{\prime}=\pi^{-1}(S^{\prime})$ and $\pi:G\rightarrow G/P$ is the canonical epimorphism. Replacing $S$ by $S^{\prime}$, we obtain the required.
Now all conditions of Lemma~\ref{main3} hold for $\mathcal{A}$ and $U$. Thus, $\mathcal{A}$ is a $\CI$-$S$-ring by Lemma~\ref{main3}. The theorem is proved.
\end{proof}
| d055299bedb583cd671224af49678e2583374e42 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
We shall study the sets whose intrinsic density is in the open unit interval. By a result of Astor \cite{intrinsicdensity}, intrinsic density $r$ is equivalent to injection stochasticity $r$, which is the uniform variant of Kolmogorov-Loveland stochasticity. Our goal is to separate intrinsic density from randomness: for almost all $r$ in the open unit interval, we shall construct sets which have intrinsic density $r$, or injection stochasticity $r$, which cannot compute any set random with respect to the $r$-Bernoulli measure.
\\
\\
We briefly recall the notion of (asymptotic) density in the natural numbers:
\begin{definition}
Let $A\subseteq\omega$.
\begin{itemize}
\item The density of $A$ at $n$ is $\rho_n(A)=\frac{|A\upharpoonright n|}{n}$, where $A\upharpoonright n=A\cap\{0,1,2,\dots,n-1\}$.
\item The upper density of $A$ is $\overline{\rho}(A)=\limsup_{n\to\infty} \rho_n(A)$.
\item The lower density of $A$ is $\underline{\rho}(A)=\liminf_{n\to\infty} \rho_n(A)$.
\item If $\overline{\rho}(A)=\underline{\rho}(A)=\alpha$, we call $\alpha$ the density of $A$ and denote it by $\rho(A)$.
\end{itemize}
\end{definition}
Astor \cite{intrinsicdensity} introduced intrinsic density, which requires that the asymptotic density remain fixed under any computable permutation.
\begin{definition}
\begin{itemize}
\item The absolute upper density of $A$ is
\[\overline{P}(A)=\sup\{\overline{\rho}(\pi(A)):\pi\text{ a computable permutation}\}\]
\item The absolute lower density of $A$ is \[\underline{P}(A)=\inf\{\underline{\rho}(\pi(A)):\pi\text{ a computable permutation}\}\]
\item If $\overline{P}(A)=\underline{P}(A)=\alpha$, we call $\alpha$ the intrinsic density of $A$ and denote it by $P(A)$. In particular, $P(A)=\alpha$ if and only if $\rho(\pi(A))=\alpha$ for every computable permutation $\pi$.
\end{itemize}
\end{definition}
A natural question to ask is what reals in the unit interval can be achieved as the intrinsic density of some set. Sets of intrinsic density $0$ and $1$ are well-known. In the open unit interval, Bienvenu (personal communication) gave a straightforward probabilistic argument that shows the sets of intrinsic density $r$ have measure $1$ under the $r$-Bernoulli measure. In fact, something stronger is true.
\begin{definition}
\label{randomnessmeasure}
Let $0<r<1$ be a real number. The Bernoulli measure with parameter $r$, $\mu_r$, is the measure on Cantor space such that for any $\sigma\in2^{<\omega}$,
\[\mu_r(\sigma)=r^{|\{n<|\sigma|:\sigma(n)=1\}|}(1-r)^{|\{n<|\sigma|:\sigma(n)=0\}|}\]
If $X$ is ML-random with respect to $\mu_r$, we say it is $r$-random.
\end{definition}
If we say 1-random, we are specifically referring to $\frac{1}{2}$-ML-randomness. For a review of randomness with respect to noncomputable measures, see Reimann and Slaman \cite{ReimannSlaman}.
\begin{proposition}
\label{randomdensity}
Let $r\in(0,1)$. If $X$ is $r$-random, then $X$ has intrinsic density $r$.
\end{proposition}
Standard arguments for $\frac{1}{2}$ can be generalized to $r$ to prove this. It is true for $r$-Schnorr randomness. See, for example, Nies \cite[Theorem 3.5.21]{Nies} for the proof in the case $r=\frac{1}{2}$.
\\
\\
Thus every real in the unit interval is achieved as the intrinsic density of some set. Furthermore, we will not be able to find an ML-random set which does not have intrinsic density as was done for computable randomness and MWC stochasticity by Ambos-Spies \cite{Ambosspies} and was done for Schnorr randomness and Church stochasticity by Wang \cite{wang}. However, it is still possible to find sets which have intrinsic density $r$ but are not $r$-random. We shall not only do this, but in fact provide examples which cannot even compute an $r$-random set. To achieve this, we shall next develop some tools which work well with both asymptotic and intrinsic density and allow us to change densities. Computable operations like the join can only preserve intrinsic density, and set operations such as the intersection and union behave unpredictably with respect to asymptotic density.
\\
\\
We shall follow the convention, unless otherwise stated, that capital English letters represent sets of natural numbers and the lowercase variant, indexed by a subscript of natural numbers, represents the elements of the set. As an example, if $E$ is the set of even numbers, then $e_n=2n$. Recall that the principal function for a set $A$, $p_A$, is defined via $p_A(n)=a_n$.
\\
\\
Using this representation, it is not hard to see the following characterization of upper and lower density:
\begin{lemma}
\label{technicallimit}
Let $A\subseteq\omega$ be $\{a_0<a_1<a_2<\dots\}$. Then
\begin{itemize}
\item $\overline{\rho}(A)=\limsup_{n\to\infty}\frac{n}{a_n}$
\item $\underline{\rho}(A)=\liminf_{n\to\infty}\frac{n}{a_n}$
\end{itemize}
\end{lemma}
\begin{proof}
Note that if $A\upharpoonright n+1$ has a $0$ in the final bit, then
\[\rho_n(A)=\frac{|A\upharpoonright n|}{n}>\frac{|A\upharpoonright n|}{n+1}=\rho_{n+1}(A)\]
Therefore, to compute the upper density it suffices to check only those numbers $n$ for which $A\upharpoonright n$ has a $1$ as its last bit. Those numbers are exactly $a_n+1$ by the definition of $a_n$, and $|A\upharpoonright a_n+1|=n+1$. Therefore $\{\frac{n+1}{a_n+1}\}_{n\in\omega}$ is a subsequence of $\{\rho_n(A)\}_{n\in\omega}$ which dominates the original sequence, so $\overline{\rho}(A)=\limsup_{n\to\infty}\rho_n(A)=\limsup_{n\to\infty}\frac{n+1}{a_n+1}$. Finally, $\limsup_{n\to\infty}\frac{n+1}{a_n+1}=\limsup_{n\to\infty}\frac{n}{a_n}$.
\\
\\
Similarly, to compute the lower density it suffices to check only the numbers $n$ such that the final digit of $A\upharpoonright n$ is a $0$, but the final digit of $A\upharpoonright n+1$ is a $1$. (That is, if there is a consecutive block of zeroes in the characteristic function of $A$, we only need to check the density at the end of the block when computing lower density, as each intermediate point of the zero block has a higher density than the end.) These numbers are exactly $a_n$ by definition, and $|A\upharpoonright a_n|=n$. Therefore $\{\frac{n}{a_n}\}_{n\in\omega}$ is a subsequence of $\{\rho_n(A)\}_{n\in\omega}$ which is dominated by the original sequence, so $\underline{\rho}(A)=\liminf_{n\to\infty}\rho_n(A)=\liminf_{n\to\infty}\frac{n}{a_n}$.
\end{proof}
A critical proof technique will involve proving that two sets $A$ and $B$ cannot have different intrinsic densities by creating a computable permutation which sends $A$ to $B$ modulo a set of density zero. The following lemma shows that if we can do this, then the density of the image of $A$ is the same as the density of $B$, and therefore that they cannot have different intrinsic densities.
\begin{lemma}
\label{densityzero}
If $\overline{\rho}(H)=0$, then $\overline{\rho}(X\setminus H)=\overline{\rho}(X\cup H)=\overline{\rho}(X)$ and $\underline{\rho}(X\setminus H)=\underline{\rho}(X\cup H)=\underline{\rho}(X)$.
\end{lemma}
\begin{proof}
Notice that
\[\rho_n(X)=\rho_n(X\setminus H)+\rho_n(X\cap H)\]
By definition. Therefore
\[\overline{\rho}(X)=\limsup_{n\to\infty}\rho_n(X)=\limsup_{n\to\infty} \rho_n(X\setminus H)+\rho_n(X\cap H)\]
By subadditivity of the limit superior,
\[\overline{\rho}(X)\leq\limsup_{n\to\infty} \rho_n(X\setminus H)+\limsup_{n\to\infty} \rho_n(X\cap H)\]
As $\overline{\rho}(H)=0$ and $X\cap H\subseteq H$,
\[\overline{\rho}(X)\leq\limsup_{n\to\infty} \rho_n(X\setminus H)=\overline{\rho}(X\setminus H)\]
However, $\overline{\rho}(X\setminus H)\leq\overline{\rho}(X)$ because $X\setminus H\subseteq X$, so $\overline{\rho}(X)=\overline{\rho}(X\setminus H)$ as desired.
\\
\\
The argument for the union and the argument for lower density are functionally identical. (For the union we use $X\cup H$, $X$, and $H\setminus X$ in place of $X$, $X\setminus H$, and $X\cap H$ respectively.)
\end{proof}
Our strategy is to find some process which takes a set $A$ of intrinsic density $\alpha$ and a set $B$ of intrinsic density $\beta$ and codes $B$ and $A$ in such a way that we are left with a set which has new intrinsic density obtained as some function of $\alpha$ and $\beta$. The following lemma will be useful for computing intrinsic densities.
\begin{lemma}
\label{permutationlemma}
Let $f_0,f_1,\dots,f_k$ be a finite collection of injective computable functions and let $C$ be an infinite computable set. Then there is an infinite computable set $H\subseteq C$ such that $\overline{\rho}(f_i(H))=0$ for all $i$.
\end{lemma}
\begin{proof}
Let $h_0=c_0$. Then given $h_n$, define $h_{n+1}$ to be the least element $c$ of $C$ with $f_i(c)\geq h_n!$ for all $i$. Set $H=\{h_0<h_1<h_2<\dots\}$. Then $\overline{\rho}(f_i(H))=0$ for all $i$ because $|f_i(H)\upharpoonright n|\leq |\{n!:n\in\omega\}\upharpoonright n|$.
\end{proof}
We shall see in Section 2 that we cannot hope for such a process to be computable in a way that allows us to recover the original sets. Instead we shall use the following noncomputable coding methods. They are natural and have been used informally by others such as Jockusch and Astor.
\begin{definition}
\label{intowithin}
Let $A$ and $B$ be sets of natural numbers.
\begin{itemize}
\item $B\triangleright A$, or $B$ into $A$, is
\[\{a_{b_0}<a_{b_1}<a_{b_2}<\dots\}\]
That is, $B\triangleright A$ is the subset of $A$ obtained by taking the ``$B$-th elements of $A$.''
\item $B\triangleleft A$, or $B$ within $A$, is
\[\{n:a_n\in B\}\]
That is, $B\triangleleft A$ is the set $X$ such that $X\triangleright A=A\cap B$.
\end{itemize}
\end{definition}
With $B\triangleright A$, we are thinking of $A$ as a copy of $\omega$ as a well-order and $B\triangleright A$ is the subset corresponding to $B$ under the order preserving isomorphism between $A$ and $\omega$. We make a few elementary observations to illustrate how these operations behave:
\begin{itemize}
\item For all $A$, $A=A\triangleright \omega=\omega\triangleright A=A\triangleleft\omega$.
\item For all $A$ and $B$ and any $i$, $a_i$ is either in $B$ or $\overline{B}$. Therefore $i$ is either in $B\triangleleft A$ or $\overline{B}\triangleleft A$ respectively, so $(B\triangleleft A)\sqcup(\overline{B}\triangleleft A)=\omega$.
\item If $B\cap C=\emptyset$, then $(B\triangleright A)\cap(C\triangleright A)=\emptyset$. Furthermore, $A=(X\triangleright A)\sqcup(\overline{X}\triangleright A)$.
\item $\triangleright$ is associative, i.e.\ $B\triangleright(A\triangleright C)=(B\triangleright A)\triangleright C$: By definition, $(A\triangleright C)=\{c_{a_0}<c_{a_1}<c_{a_2}<\dots\}$ and thus
\[B\triangleright (A\triangleright C)=\{c_{a_{b_0}}<c_{a_{b_1}}<c_{a_{b_2}}<\dots\}\]
Similarly, $(B\triangleright A)=\{a_{b_0}<a_{b_1}<a_{b_2}<\dots\}$, and therefore by definition
\[(B\triangleright A)\triangleright C=\{c_{a_{b_0}}<c_{a_{b_1}}<c_{a_{b_2}}<\dots\}\]
\item $\triangleleft$ is not associative: Consider the set of evens $E$, the set of odds $O$, and the set $N$ of evens which are not multiples of $4$. Then
\[(O\triangleleft N)\triangleleft E=\emptyset\triangleleft N=\emptyset\]
However,
\[O\triangleleft(N\triangleleft E)=O\triangleleft O=\omega\]
\item $\triangleright$ and $\triangleleft$ do not associate with each other in general:
\[B\triangleright(A\triangleleft(B\triangleright A))=B\triangleright\omega=B\]
but
\[(B\triangleright A)\triangleleft(B\triangleright A)=\omega\]
Similarly, $B\triangleleft(A\triangleright B)=\omega$, but $(B\triangleleft A)\triangleright B$ is a subset of $B$.
\end{itemize}
In Section 2, we shall construct examples of sets which have intrinsic density $r$ but are not $r$-random. However, all of these examples will trivially compute $r$-random sets, so we shall then turn our attention to constructing examples which cannot. In Section 3 we shall prove our main technical theorems which allow us to complete the construction in Section 4.
\section{Nonrandom sets of intrinsic density $r$}
The following theorem allows us to easily generate sets of intrinsic density $r$ which are not $r$-random from Proposition \ref{randomdensity}.
\begin{theorem}
\label{join}
$P(A\oplus B)=r$ if and only if $P(A)=P(B)=r$.
\end{theorem}
\begin{proof}
We shall first argue the forward direction via contrapositive. We proceed by showing that, for any given computable permutation $\pi$, there are computable permutations which send $A\oplus B$ to $\pi(A)$ modulo a set of density $0$, and similarly for $\pi(B)$. Then the upper (and lower) density of $A\oplus B$ under these permutations will match that of $\pi(A)$ and $\pi(B)$ respectively. Therefore if $P(A)\neq P(B)$, the density of $A\oplus B$ is not invariant under computable permutation.
\\
\\
Let $F=\{n!:n\in\omega\}$ and $G=\overline{F}$. For any fixed computable permutation $\pi$, there is another computable permutation $\hat{\pi}$ defined via enumerating the odds onto the factorials in order and enumerating the evens onto the nonfactorials according to the ordering induced by $\pi$. That is, $\hat{\pi}(2n+1)=f_n$ and $\hat{\pi}(2n)=g_{\pi(n)}$.
\\
\\
Then as $F$ has density $0$, Lemma \ref{densityzero} shows
\[\overline{\rho}(\hat{\pi}(A\oplus B))=\overline{\rho}(\hat{\pi}(A\oplus B)\setminus F)\]
As the image of the odds under $\hat{\pi}$ is a subset of $F$,
\[\hat{\pi}(A\oplus B)\setminus F=\hat{\pi}(A\oplus\emptyset)\]
and
\[\overline{\rho}(\hat{\pi}(A\oplus B))=\overline{\rho}(\hat{\pi}(A\oplus \emptyset))\]
Notice that $\hat{\pi}(A\oplus\emptyset)$ is just $\pi(A)$ with each element $n$ increased by $|F\upharpoonright n|$. Thus
\[\rho_n(\pi(A))\geq\rho_n(\hat{\pi}(A\oplus\emptyset))\geq\frac{|\pi(A)\upharpoonright n|-|F\upharpoonright n|}{n}\]
As $F$ is the factorials, the final expression tends to $\rho_n(\pi(A))$ in the limit, so we see that
\[\overline{\rho}(\hat{\pi}(A\oplus\emptyset))=\overline{\rho}(\pi(A))\]
and
\[\overline{\rho}(\hat{\pi}(A\oplus B))=\overline{\rho}(\hat{\pi}(A\oplus\emptyset))=\overline{\rho}(\pi(A))\]
$\underline{\rho}(\hat{\pi}(A\oplus B))=\underline{\rho}(\pi(A))$ by a nearly identical argument.
\\
\\
In particular, $\overline{P}(A\oplus B)\geq \overline{P}(A)$ and $\underline{P}(A\oplus B)\leq \underline{P}(A)$ because we are taking the limit superior and inferior over all computable permutations, of which $\hat{\pi}$ is but one. Reversing the use of the evens and the odds in the definition of $\hat{\pi}$, we get that the same is true for $B$ in place of $A$, so $\underline{P}(A\oplus B)\leq min(\underline{P}(A),\underline{P}(B))$ and $\overline{P}(A\oplus B)\geq max(\overline{P}(A),\overline{P}(B))$. Therefore if $P(A)\neq P(B)$, $\underline{P}(A\oplus B)\neq\overline{P}(A\oplus B)$ as desired.
\\
\\
We now prove the reverse direction using a technical lemma.
\begin{sublemma}
\label{joinlemma}
Let $\pi$ be a computable permutation and let $P(A)=P(B)=r$. Then
\[\rho(\pi(A\oplus B)\triangleleft\pi(E))=\rho(\pi(A\oplus B)\triangleleft\pi(O))=r\]
\end{sublemma}
\begin{proof}
Let $h:\pi(E)\to\omega$ send the $n$-th element of $\pi(E)$ to $n$ (the inverse of the principal function), and let $d:\omega\to E$ be defined via $d(n)=2n$. Then notice that $d(A)=A\oplus\emptyset$. Furthermore, observe that for any $X\subseteq\pi(E)$, $h(X)=X\triangleleft \pi(E)$ by the definition of $h$ and the {\tt within} operation. Therefore
\[h(\pi(d(A)))=h(\pi(A\oplus\emptyset))=\pi(A\oplus\emptyset)\triangleleft\pi(E)\]
As $\pi(A\oplus B)\cap \pi(E)\subseteq \pi(A\oplus \emptyset)$,
\[\pi(A\oplus\emptyset)\triangleleft\pi(E)=\pi(A\oplus B)\triangleleft\pi(E)\]
Thus $h(\pi(d(A)))=\pi(A\oplus B)\triangleleft\pi(E)$. We shall now massage $h$ and $d$ into permutations which preserve the relevant densities.
\\
\\
By Lemma \ref{permutationlemma}, there is a computable set $H\subseteq \pi(E)$ with $\overline{\rho}(h(H))=0$. Now define the computable permutation $\pi_h$ via $\pi_h(n)=h(n)$ for $n\in\pi(E)\setminus H$, and have $\pi_h$ enumerate $\pi(O)\sqcup H$ onto $h(H)$ in order. Similarly, define the computable permutation $\pi_d$ via $\pi_d(n)=d(n)$ for $n\in\omega\setminus d^{-1}(\pi^{-1}(H))$, and have $\pi_d$ enumerate $d^{-1}(\pi^{-1}(H))$ onto $O\sqcup \pi^{-1}(H)$.
\\
\\
As $\pi_d$ agrees with $d$ on $\overline{d^{-1}(\pi^{-1}(H))}$, we now see that
\[\pi_d(A\setminus \pi_d^{-1}(\pi^{-1}(H)))=(A\oplus\emptyset)\setminus \pi^{-1}(H)\]
Furthermore, applying $\pi$ shows that
\[\pi(\pi_d(A\setminus \pi_d^{-1}(\pi^{-1}(H))))=\pi((A\oplus\emptyset)\setminus \pi^{-1}(H))=\pi(A\oplus\emptyset)\setminus H\]
As $\pi_h$ agrees with $h$ on $\pi(E)\setminus H$ and $h(\pi(A\oplus\emptyset))=\pi(A\oplus B)\triangleleft\pi(E)$, we have
\[\pi_h(\pi(A\oplus\emptyset)\setminus H)=(\pi(A\oplus B)\triangleleft\pi(E))\setminus h(H)\]
Therefore $(\pi(A\oplus B)\triangleleft\pi(E))\setminus h(H)\subseteq\pi_h(\pi(\pi_d(A)))$ and $\pi_h(\pi(\pi_d(A)))\subseteq (\pi(A\oplus B)\triangleleft\pi(E))\cup h(H)$.
\\
\\
By choice of $H$, $\overline{\rho}(h(H))=0$, so Lemma \ref{densityzero} shows that
\[\overline{\rho}(\pi_h(\pi(\pi_d(A))))=\overline{\rho}((\pi(A\oplus B)\triangleleft\pi(E))\setminus h(H))=\overline{\rho}(\pi(A\oplus B)\triangleleft\pi(E))\]
and
\[\underline{\rho}(\pi_h(\pi(\pi_d(A))))=\underline{\rho}((\pi(A\oplus B)\triangleleft\pi(E))\setminus h(H))=\underline{\rho}(\pi(A\oplus B)\triangleleft\pi(E))\]
Therefore, as $P(A)=r$ and $\pi_h\circ\pi\circ\pi_d$ is a computable permutation,
\[\rho(\pi(A\oplus B)\triangleleft\pi(E))=r\]
A nearly identical argument with $O$ in place of $E$ and $B$ in place of $A$ shows similarly that
\[\rho(\pi(A\oplus B)\triangleleft\pi(O))=r\]
\end{proof}
We shall now show that this implies that $\rho(\pi(A\oplus B))=r$.
Consider $\rho_n(\pi(A\oplus B))$. By definition,
\[\rho_n(\pi(A\oplus B))=\frac{|\pi(A\oplus B)\upharpoonright n|}{n}\]
As $\omega=\pi(E)\sqcup\pi(O)$,
\[\frac{|\pi(A\oplus B)\upharpoonright n|}{n}=\frac{|\pi(A\oplus B)\cap\pi(E)\upharpoonright n|+|\pi(A\oplus B)\cap\pi(O)\upharpoonright n|}{n}\]
The latter expression can be rewritten as
\[\frac{|\pi(E)\upharpoonright n|}{|\pi(E)\upharpoonright n|}\cdot\frac{|\pi(A\oplus B)\cap\pi(E)\upharpoonright n|}{n}+\frac{|\pi(O)\upharpoonright n|}{|\pi(O)\upharpoonright n|}\cdot\frac{|\pi(A\oplus B)\cap\pi(O)\upharpoonright n|}{n}\]
Let $m$ be the largest number such that the $m$-th element of $\pi(E)$ is less than $n$, and let $k$ be the analogous number for $\pi(O)$. Now notice that
\[\frac{|\pi(A\oplus B)\cap\pi(E)\upharpoonright n|}{|\pi(E)\upharpoonright n|}=\rho_m(\pi(A\oplus B)\triangleleft\pi(E))\]
and
\[\frac{|\pi(A\oplus B)\cap\pi(O)\upharpoonright n|}{|\pi(O)\upharpoonright n|}=\rho_k(\pi(A\oplus B)\triangleleft\pi(O))\]
by the definition of the {\tt within} operation. Therefore, we can rewrite $\rho_n(\pi(A\oplus B))$ as \[\rho_m(\pi(A\oplus B)\triangleleft\pi(E))\cdot\rho_n(\pi(E))+\rho_k(\pi(A\oplus B)\triangleleft\pi(O))\cdot\rho_n(\pi(O))\]
Using the fact that $\rho_n(\pi(E))+\rho_n(\pi(O))=1$,
\[\rho_n(\pi(A\oplus B))=\rho_m(\pi(A\oplus B)\triangleleft\pi(E))\cdot\rho_n(\pi(E))+\rho_k(\pi(A\oplus B)\triangleleft\pi(O))\cdot(1-\rho_n(\pi(E)))\]
Rearranging, this is equal to
\[\rho_k(\pi(A\oplus B)\triangleleft\pi(O))+\rho_n(\pi(E))\cdot(\rho_m(\pi(A\oplus B)\triangleleft\pi(E))-\rho_k(\pi(A\oplus B)\triangleleft\pi(O)))\]
Taking the limit as $n$ goes to infinity, $m$ and $k$ both go to infinity. Thus
\[\rho_m(\pi(A\oplus B)\triangleleft\pi(E))-\rho_k(\pi(A\oplus B)\triangleleft\pi(O))\]
goes to $0$ by Lemma \ref{joinlemma}. As $\rho_n(\pi(E))$ is bounded between $0$ and $1$ by definition, the second term vanishes. Therefore
\[\lim_{n\to\infty}\rho_n(\pi(A\oplus B))=\lim_{n\to\infty} \rho_k(\pi(A\oplus B)\triangleleft\pi(O))=\rho(\pi(A\oplus B)\triangleleft\pi(O))=r\]
as desired.
\end{proof}
Using Theorem \ref{join}, we may take any set $X$ which is $r$-random and then $X\oplus X$ has intrinsic density $r$ but is not random as desired. However, this is merely a structural difference between the notion of randomness and the notion of intrinsic density as opposed to a computational difference.
\section{Core Theorems}
We would now like to build examples of sets with intrinsic density $r$ which cannot compute an $r$-random set. To do so, we shall prove two main theorems which allow us to apply the {\tt into} and {\tt within} operations to study asymptotic and intrinsic density.
\begin{theorem}
\label{theoremwithin}
Let $C$ be computable and $P(A)=r$. Then $P(A\triangleleft C)=r$.
\end{theorem}
\begin{proof}
Under the map which takes $c_n$ to $n$, $A\cap C$ is mapped to $A\triangleleft C$. However unless $C$ is $\omega$, this is not a permutation. Using Lemma \ref{permutationlemma}, we are able to massage this map into a permutation which takes $c_n$ to $n$ modulo a set of density $0$. Then under this permutation, $A\cap C$ (and $A$) goes to $A\triangleleft C$ modulo a set of density $0$. Therefore if $A\triangleleft C$ did not have intrinsic density $r$, $A$ could not either by Lemma \ref{densityzero}.
\\
\\
Formally, assume $P(A\triangleleft C)\neq r$. Suppose $\pi$ is a computable permutation with $\overline{\rho}(\pi(A\triangleleft C))>r$. Let $f:C\to\omega$ be defined via $f(c_n)=n$. Then $f(A\cap C)=A\triangleleft C$:
\begin{center}
\begin{tikzcd}
A\cap C \arrow [r, "f"] & A\triangleleft C \arrow [r,"\pi"] & \pi(A\triangleleft C)
\end{tikzcd}
\end{center}
By Lemma \ref{permutationlemma}, there is $H\subseteq C$ computable with $\overline{\rho}(\pi(f(H)))=0$. Define $\pi_f:\omega\to\omega$ via $\pi_f(n)=f(n)$ for $n\in C\setminus H$, and for $n\in \overline{C}\sqcup H$ define $\pi_f(n)$ to be the least element of $f(H)$ not equal to $\pi_f(j)$ for some $j<n$. As $f$ agrees with $\pi_f$ on $C\setminus H$,
\[\pi_f((A\cap C)\setminus H)=f(A\cap C)\setminus f(H)=(A\triangleleft C)\setminus f(H)\]
Therefore by applying $\pi$,
\[\pi(\pi_f((A\cap C)\setminus H))=\pi((A\triangleleft C)\setminus f(H))=\pi(A\triangleleft C)\setminus \pi(f(H))\]
Using the above equality,
\[\overline{\rho}(\pi(\pi_f((A\cap C)\setminus H)))=\overline{\rho}(\pi(A\triangleleft C)\setminus\pi(f(H)))\]
As $\overline{\rho}(\pi(f(H)))=0$, we can apply Lemma \ref{densityzero} and see
\[\overline{\rho}(\pi(A\triangleleft C)\setminus\pi(f(H)))=\overline{\rho}(\pi(A\triangleleft C))\]
As $(A\cap C)\setminus H\subseteq A$,
\[\overline{\rho}(\pi(\pi_f(A)))\geq\overline{\rho}(\pi(\pi_f((A\cap C)\setminus H)))=\overline{\rho}(\pi(A\triangleleft C))\]
However, we assumed that $\overline{\rho}(\pi(A\triangleleft C))>r$, so $\overline{\rho}(\pi(\pi_f(A)))>r$. As $\pi\circ\pi_f$ is a computable permutation, this implies $P(A)\neq r$.
\\
\\
This proves that if $\pi$ is a computable permutation with $\overline{\rho}(\pi(A\triangleleft C))>r$, then $P(A)\neq r$. If there is no such permutation, there must be a computable permutation $\pi$ with $\underline{\rho}(\pi(A\triangleleft C))<r$ because we assumed that $P(A\triangleleft C)\neq r$. Then because
\[(\pi(A\triangleleft C))\sqcup(\pi(\overline{A}\triangleleft C))=\pi((A\triangleleft C)\sqcup(\overline{A}\triangleleft C))=\pi(\omega)=\omega\]
we have $\rho_n(\pi(\overline{A}\triangleleft C))=1-\rho_n(\pi(A\triangleleft C))$ for all $n$. Therefore by the subtraction properties of the limit superior,
\[\overline{\rho}(\pi(\overline{A}\triangleleft C))\geq1-\underline{\rho}(\pi(A\triangleleft C))\]
As we assumed $\underline{\rho}(\pi(A\triangleleft C))<r$,
\[1-\underline{\rho}(\pi(A\triangleleft C))>1-r\]
Thus $\overline{\rho}(\pi(\overline{A}\triangleleft C))>1-r$. We now apply the previous case to get that $P(\overline{A})\neq1-r$, which automatically implies $P(A)\neq r$.
\end{proof}
If we are successful in building a set of intrinsic density $r$ which cannot compute an $r$-random set, then we will be able to use this theorem to automatically get many more such examples.
\\
\\
We now turn our attention {\tt into}, which is more useful. We first make a crucial observation about the asymptotic density of $B\triangleright A$, which will be critical for investigating the intrinsic density of sets obtained via use of the {\tt into} operation.
\begin{lemma}
\label{densitylimit}
\mbox{}
\begin{itemize}
\item $\overline{\rho}(B\triangleright A)\leq \overline{\rho}(B)\overline{\rho}(A)$.
\item $\underline{\rho}(B\triangleright A)\geq \underline{\rho}(B)\underline{\rho}(A)$.
\end{itemize}
\end{lemma}
\begin{proof}
By Lemma \ref{technicallimit},
\[\overline{\rho}(B\triangleright A)=\limsup_{n\to\infty}\frac{n}{a_{b_n}}=\limsup_{n\to\infty}\frac{n}{a_{b_n}}\cdot 1=\limsup_{n\to\infty}\frac{n}{a_{b_n}}\cdot\frac{b_n}{b_n}\]
By the submultiplicativity of the limit superior,
\[\overline{\rho}(B\triangleright A)\leq (\limsup_{n\to\infty}\frac{b_n}{a_{b_n}})(\limsup_{n\to\infty}\frac{n}{b_n})=(\limsup_{n\to\infty}\frac{b_n}{a_{b_n}})\overline{\rho}(B)\]
Now $\{\frac{b_n}{a_{b_n}}\}_{n\in\omega}$ is a subsequence of $\{\frac{n}{a_n}\}_{n\in\omega}$, so
\[\limsup_{n\to\infty}\frac{b_n}{a_{b_n}}\leq\limsup_{n\to\infty}\frac{n}{a_n}=\overline{\rho}(A)\]
Therefore $\overline{\rho}(B\triangleright A)\leq \overline{\rho}(B)\overline{\rho}(A)$ as desired.
\\
\\
The case for the limit inferior is nearly identical, reversing $\leq$ to $\geq$ and using supermultiplicativity along with the corresponding identity from Lemma \ref{technicallimit}.
\end{proof}
\begin{corollary}
If $\rho(A)=\alpha$ and $\rho(B)=\beta$, then $\rho(B\triangleright A)=\alpha\beta$.
\end{corollary}
Therefore, if $B\triangleright A$ has intrinsic density, its intrinsic density must be the product of the densities of $A$ and $B$. This will occur in certain instances, which is our second main theorem. Recall that a set $X$ has $Y$-intrinsic density, or intrinsic density relative to $Y$, if its density is invariant under all $Y$-computable permutations as opposed to just the computable ones. We use $P_Y(X)$ to denote the $Y$-intrinsic density of $X$ if it exists.
\begin{theorem}
\label{core}
If $P(A)=\alpha$ and $P_A(B)=\beta$, then $P(B\triangleright A)=\alpha\beta$.
\end{theorem}
\begin{proof}
The proof is similar to the proof of Theorem \ref{theoremwithin}, however we shall present it fully here without referring to techniques from that proof, as it is quite technical. The idea is that for any fixed computable permutation $\pi$, there is an $A$-computable permutation which sends $B$ to $\pi(B\triangleright A)\triangleleft\pi(A)$ modulo a set of density $0$. Therefore if $\pi$ witnesses that $B\triangleright A$ does not have intrinsic density $\alpha\beta$, i.e.\ $\pi(B\triangleright A)$ does not have density $\alpha\beta$, and then as $A$ has intrinsic density $\alpha$ by assumption, Lemma \ref{densitylimit} will show that $\pi(B\triangleright A)\triangleleft\pi(A)$ does not have density $\beta$. Thus $B$ does not have $A$-intrinsic density $\beta$.
\\
\\
Formally, assume $P(A)=\alpha$. Assume that $P(B\triangleright A)\neq \alpha\beta$. We shall show that $P_A(B)\neq \beta$. First suppose that there is some computable permutation $\pi$ such that $\overline{\rho}(\pi(B\triangleright A))>\alpha\beta$. We shall let $\pi(A)=\{p_0<p_1<p_2<\dots\}$. Let $f:A\to\omega$ be defined via $f(a_n)=n$ and $g:\pi(A)\to\omega$ via $g(p_n)=n$, $f$ maps $A$ to its indices and $g$ maps $\pi(A)$ to its indices. Then $f(B\triangleright A)=B$ and $g(\pi(B\triangleright A))=\pi(B\triangleright A)\triangleleft\pi(A)$:
\begin{center}
\begin{tikzcd}
B\triangleright A \arrow[r, "\pi"] \arrow [d, "f"] & \pi(B\triangleright A) \arrow[d, "g"] \\
B & \pi(B\triangleright A)\triangleleft\pi(A)
\end{tikzcd}
\end{center}
Note by Lemma \ref{densitylimit} that $\overline{\rho}(\pi(B\triangleright A)\triangleleft\pi(A))>\beta$: From the definition,
\[(\pi(B\triangleright A)\triangleleft\pi(A))\triangleright \pi(A))=\pi(B\triangleright A)\]
and $\overline{\rho}(B\triangleright A)>\alpha\beta$ by assumption. $\overline{\rho}(\pi(A))=\alpha$ because $P(A)=\alpha$, so $\overline{\rho}(\pi(B\triangleright A)\triangleleft\pi(A))\leq\beta$ would contradict Lemma \ref{densitylimit}.
\\
\\
From this point forward we shall let
\[X=\pi(B\triangleright A)\triangleleft\pi(A)\]
for the sake of readability.
\\
\\
By Lemma \ref{permutationlemma} relativized to $A$ and applied to $g\circ\pi$, there is an $A$-computable set $H\subseteq A$ such that:
\[\overline{\rho}(g(\pi(H)))=0\]
We shall now define permutations which preserve the properties of $f$ and $g$ outside of $H$. Define $\pi_f:\omega\to\omega$ via $\pi_f(k)=f(k)$ for $k\in A\setminus H$, and for $k\in \overline{A}\sqcup H$, let $\pi_f(k)$ be the least element of $f(H)$ not equal to $\pi_f(m)$ for some $m<k$. Define $\pi_g:\omega\to\omega$ similarly using $\pi(A)$, $\pi(H)$, and $g(\pi(H))$ in place of $A$, $H$, and $f(H)$ respectively. Then $\pi_f$ and $\pi_g$ are $A$-computable because $H$, $f$, and $g$ are, and it is a permutation because $f$ and $g$ are bijections (from $A$ and $\pi(A)$ to $\omega$ respectively) which have been modified to be total without violating injectivity or surjectivity.
\\
\\
Now we shall compute $\pi_g(\pi(\pi_f^{-1}(B\setminus f(H))))$. As $f(B\triangleright A)=B$ and $f$ agrees with $\pi_f$ on $\overline{H}$,
\[\pi_f^{-1}(B\setminus f(H))=(B\triangleright A)\setminus H\]
Furthermore
\[\pi((B\triangleright A)\setminus H)=\pi(B\triangleright A)\setminus\pi(H)\]
As $g(\pi(B\triangleright A))=X$ and $\pi_g$ agrees with $g$ on $\overline{\pi(H)}$,
\[\pi_g(\pi(B\triangleright A)\setminus\pi(H))=g(\pi(B\triangleright A))\setminus g(\pi(H))=X\setminus g(\pi(H))\]
Thus $\pi_g(\pi(\pi_f^{-1}(B\setminus f(H))))=X\setminus g(\pi(H))$. As $\overline{\rho}(g(\pi(H))=0$, Lemma \ref{densityzero} shows
\[\overline{\rho}(X\setminus g(\pi(H)))=\overline{\rho}(X)\]
By the definition of $X$,
\[\overline{\rho}(X)=\overline{\rho}(\pi(B\triangleright A)\triangleleft\pi(A))\]
which is greater than $\beta$ by the above. As $B\setminus f(H)\subseteq B$,
\[\pi_g(\pi(\pi_f^{-1}(B\setminus f(H))))\subseteq\pi_g(\pi(\pi_f^{-1}(B)))\]
and thus
\[\overline{\rho}(\pi_g(\pi(\pi_f^{-1}(B))))\geq \overline{\rho}(\pi_g(\pi(\pi_f^{-1}(B\setminus f(H)))))\]
Therefore
\[\overline{\rho}(\pi_g(\pi(\pi_f^{-1}(B))))\geq \overline{\rho}(\pi(B\triangleright A)\triangleleft\pi(A))>\beta\]
As $\pi_g\circ\pi\circ\pi_f^{-1}$ is an $A$-computable permutation, $P_A(B)\neq\beta$.
\\
\\
Therefore we have proved that if there is some computable permutation $\pi$ such that $\overline{\rho}(\pi(B\triangleright A))>\alpha\beta$, then $P_A(B)\neq \beta$. If there is no such permutation, then there must be a computable permutation $\pi$ such that $\underline{\rho}(\pi(B\triangleright A))<\alpha\beta$ because we assumed $P(B\triangleright A)\neq\alpha\beta$. As $A=(B\triangleright A)\sqcup(\overline{B}\triangleright A)$, $\pi(A)=\pi(B\triangleright A)\sqcup\pi(\overline{B}\triangleright A)$. Therefore
\[\overline{\rho}(\pi(\overline{B}\triangleright A))=\overline{\rho}(\pi(A)\setminus\pi(B\triangleright A))\]
The fact that $\rho_n(\pi(A))=\rho_n(\pi(B\triangleright A))+\rho_n(\pi(A)\setminus\pi(B\triangleright A))$ combined with the properties of the limit superior with regards to subtraction implies
\[\overline{\rho}(\pi(A)\setminus\pi(B\triangleright A))\geq\overline{\rho}(\pi(A))-\underline{\rho}(\pi(B\triangleright A))\]
We know that $\overline{\rho}(\pi(A))=\alpha$ because $P(A)=\alpha$. As we assumed that $\underline{\rho}(\pi(B\triangleright A))<\alpha\beta$,
\[\overline{\rho}(\pi(\alpha))-\underline{\rho}(\pi(B\triangleright A))>\alpha-\alpha\beta=\alpha(1-\beta)\]
Bringing this together,
\[\overline{\rho}(\pi(\overline{B}\triangleright A))>\alpha(1-\beta)\]
Thus we can apply the first case of the proof to show that $P_A(\overline{B})\neq 1-\beta$, which automatically implies $P_A(B)\neq\beta$, so we are done.
\end{proof}
\begin{corollary}
If $P(A)=\alpha$ and $P_A(B)=\beta$, then $P(A\cap B)=\alpha\beta$.\footnote{Astor \cite{intrinsicdensity} proved this for the special case when $A$ has intrinsic density $\alpha$ and $B$ is 1-random relative to $A$.}
\end{corollary}
\begin{proof}
By definition,
\[A\cap B=(B\triangleleft A)\triangleright A\]
As $P_A(B)=\beta$, Theorem \ref{theoremwithin} relativized to $A$ shows that $P_A(B\triangleleft A)=\beta$. Therefore we can apply Theorem \ref{core} to $A$ and $B\triangleleft A$ to get that
\[P((B\triangleleft A)\triangleright A)=P(A\cap B)=\alpha\beta\]
\end{proof}
It is natural to ask if any of the intrinsic density requirements in Theorem \ref{core} can be dropped. It is immediate that we cannot drop the requirement that $A$ has intrinsic density: $P_A(\omega)=1$ for any $A$, so $\omega$ always satisfies the requirements on $B$, but $\omega\triangleright A=A$, so $A$ must have intrinsic density. Similarly, $B\triangleright\omega=B$ for any $B$, so $B$ must have intrinsic density. Therefore the only possible weakening of Theorem \ref{core} would be to require $P(B)=\beta$ as opposed to $P_A(B)=\beta$. However, this fails.
\begin{proposition}
\label{weakening}
Let $P(A)=\frac{1}{2}$. Then $P(A\oplus\overline{A})=\frac{1}{2}$ but $A\triangleright(A\oplus\overline{A})$ does not have intrinsic density.
\end{proposition}
\begin{proof}
Note that $A\oplus\overline{A}$ has intrinsic density $\frac{1}{2}$ by Theorem \ref{join} as $P(A)=\frac{1}{2}$ implies $P(\overline{A})=\frac{1}{2}$.
\\
\\
Let $E$ represent the set of even numbers. Notice that $A\oplus \overline{A}$ contains exactly one of $2k$ or $2k+1$ for all $k\in\omega$. Therefore the $n$-th element of $A\oplus \overline{A}$ is $2n$ if $n\in A$ and $2n+1$ if $n\not\in A$. Thus
\[E\triangleleft (A\oplus\overline{A})=A\]
by definition. By the properties of the {\tt within} operation,
\[A\triangleright(A\oplus \overline{A})=(E\triangleleft(A\oplus\overline{A}))\triangleright (A\oplus\overline{A})=E\cap(A\oplus\overline{A})=A\oplus\emptyset\]
By Proposition \ref{join}, however, $A\oplus\emptyset$ does not have intrinsic density.
\end{proof}
\section{Intrinsic Density $r$ sets which cannot compute an $r$-random}
We are now ready to construct sets of intrinsic density $r$ which cannot compute an $r$-random set. To do this, we would like to have a countable collection of sets which all have intrinsic density relative to each other so that we may apply Theorem \ref{core} repeatedly.
\begin{lemma}
\label{randomsequence}
Any 1-random set $X$ uniformly computes a countable, disjoint sequence of sets $\{A_i\}_{i\in\omega}$ such that $A_i$ has intrinsic density $\frac{1}{2^{i+1}}$, i.e. $P(A_i)=\frac{1}{2^{i+1}}$, for each $i$. Furthermore, the $A_i$'s form a partition of $\omega$.
\end{lemma}
\begin{proof}
Recall that given a set $X$, $X^{[i]}$ denotes the $i$-th column of $X$, i.e.\ $\{n:\langle i,n\rangle\in X\}$. Let $X\subseteq\omega$ be 1-random. Then for all $i$, $X^{[i]}$ is 1-random relative to $\bigoplus_{j\neq i} X^{[j]}$. (Essentially Van Lambalgen \cite{vanlambalgen}, Downey-Hirschfeldt \cite[Corollary 6.9.6]{downeyhirschfeldt}) Proposition \ref{randomdensity} relativizes to the fact that $Z$-1-randoms have $Z$-intrinsic density $\frac{1}{2}$. In particular, taking a single 1-random automatically gives us infinitely many mutually 1-random sets, and thus infinitely many sets with intrinsic density $\frac{1}{2}$ relative to each other. Using these together with Theorem \ref{core}, we can construct the desired sequence, where the mutual randomness ensures that the conditions of Theorem \ref{core} are met.
\\
\\
Let $B_0=\omega$. Given $B_n$, let
\[A_{n}=\overline{X^{[n]}}\triangleright B_n\]
and
\[B_{n+1}=X^{[n]}\triangleright B_n\]
Note that for all $i$, $B_{i+1}\subseteq B_i$ and $A_{i}\cap B_{i+1}=\emptyset$, as $B_{i+1}=X^{[i]}\triangleright B_i$ and $A_{i}=\overline{X^{[i]}}\triangleright B_{i}$. Then for $i<j$, $A_i\cap A_j=\emptyset$ because $A_j\subseteq B_j\subseteq B_{i+1}$. Thus $\{A_i\}_{i\in\omega}$ is disjoint. Furthermore, the $A_i$'s and $B_i$'s are in fact uniformly computable in $X$, as $X$ can uniformly compute all of its columns and $X\triangleright Y$ is uniformly computable in $X\oplus Y$. We now verify that $P(A_i)=\frac{1}{2^{i+1}}$ and $P(B_i)=\frac{1}{2^i}$ by induction.
\\
\\
$P(B_0)=P(\omega)=1$, and $B_0$ is computable. Suppose that $B_i$ is $\bigoplus_{j<i} X^{[j]}$-computable and that $P(B_i)=\frac{1}{2^i}$. Then $B_{i+1}=X^{[i]}\triangleright B_i$ and $A_i=\overline{X^{[i]}}\triangleright B_i$ are both $B_i\oplus X^{[i]}$-computable, and therefore $\bigoplus_{j< i+1} X^{[j]}$-computable. By the above, both $X^{[i]}$ and $\overline{X^{[i]}}$ are 1-random relative to $B_i$ and thus have intrinsic density $\frac{1}{2}$ relative to $B_i$. Therefore $P_{B_i}(X^{[i]})=P_{B_i}(\overline{X^{[i]}})=\frac{1}{2}$ by the relativization of Proposition \ref{randomdensity}. Thus by Theorem \ref{core} and the induction hypothesis, \[P(A_i)=P(\overline{X^{[i]}}\triangleright B_i)=P(\overline{X^{[i]}})P(B_i)=\frac{1}{2}\cdot\frac{1}{2^i}=\frac{1}{2^{i+1}}\]
A nearly identical argument for $P(B_{i+1})$ verifies $P(B_{i+1})=\frac{1}{2^{i+1}}$, which completes the induction.
\\
\\
Finally, suppose for the sake of contradiction that the $A_i$'s do not form a partition of $\omega$. Since we have already shown that the sequence is disjoint, there must exist a least $m$ with $m\not\in A_i$ for any $i$. Therefore, there must be some $k$ such that $m$ is the least element of $B_n$ for all $n\geq k$. This is because every $m$ is in $B_0$ and $B_{i+1}$ is a subset of $B_i$ missing only elements of $A_i$. It follows that $0\in X^{[n]}$ for all $n>k$, as $0\in\overline{X^{[n]}}$ would imply that $m\in A_n$ since $A_n=\overline{X^{[n]}}\triangleright B_n$ and $m$ is the least, i.e. $0$-th, element of $B_n$. However, this means that $\{\langle n,0\rangle:n>k\}$ is an infinite computable subset of $X$, which contradicts the assumption that $X$ is 1-random since it is a basic fact that 1-random sets cannot have infinite computable subsets. Therefore, every $m$ must be in some $A_i$ as desired.
\end{proof}
Jockusch and Schupp \cite{JockuschSchupp} proved that asymptotic density enjoys a restricted form of countable additivity: if there is a countable sequence $\{S_i\}_{i\in\omega}$ of disjoint sets such that $\rho(S_i)$ exists for all $i$ and
\[\lim_{n\to\infty} \overline{\rho}(\bigsqcup_{i>n} S_i)=0\]
then
\[\rho(\bigsqcup_{i\in\omega}S_i)=\mathop{\Sigma}_{i=0}^\infty \rho(S_i)\]
The intrinsic density analog of this results follows immediately from the fact that permutations preserve disjoint unions. That is, if there is a countable sequence $\{S_i\}_{i\in\omega}$ of disjoint sets such that $P(S_i)$ exists for all $i$ and
\[\lim_{n\to\infty} \overline{P}(\bigsqcup_{i>n} S_i)=0\]
then
\[P(\bigsqcup_{i\in\omega}S_i)=\mathop{\Sigma}_{i=0}^\infty P(S_i)\]
Note that $\lim_{n\to\infty}\overline{P}(\bigsqcup_{i>n} A_i)=0$ must be true for any sequence satisfying Lemma \ref{randomsequence}, as $\lim_{n\to\infty} P(\bigsqcup_{i\leq n} A_i)=1$. This together with the previous lemma allows us to construct our desired set.
\begin{theorem}
\label{every}
For every $r\in(0,1)$ and any 1-random set $X$, $r\oplus X$ computes a set with intrinsic density $r$.
\end{theorem}
\begin{proof}
Let $r\in(0,1)$. Let $B_r\subseteq \omega$ be the set whose characteristic function is identified with the binary expansion that gives $r$, the set of all $n$ such that the $n$-th bit in the binary expansion for $r$ is 1. Let $\{A_i\}_{i\in\omega}$ be as in Lemma \ref{randomsequence} applied to $X$, i.e. a uniformly $X$-computable partition of $\omega$ with $A_i$ having intrinsic density $\frac{1}{2^{i+1}}$ for each $i$. Let $X_r=\bigsqcup_{n\in B_r} A_n$. We now describe the process by which $X_r$ is computable from $r\oplus X$. For a given $m$, $m\in A_n$ for some $n$ since the $A_i$'s form a partition of $\omega$. $X$ can uniformly compute the $A_i$'s and thus compute $n$ where $m\in A_n$. Then $m\in X_r$ if and only if $n\in B_r$, which $r$ can compute.
\\
\\
Now, note that
\[\lim_{n\to\infty}\overline{P}(\bigsqcup_{i\in B_r, i>n} A_i)=0\]
because $\bigsqcup_{i\in B_r, i>n} A_i\subseteq \bigsqcup_{i>n} A_i$ and $\lim_{n\to\infty}\overline{P}(\bigsqcup_{i>n} A_i)=0$. By the fact that countable unions sum intrinsic densities and the definition of $X_r$,
\[P(X_r)=\mathop{\Sigma}_{n\in B_r} P(A_n)=\mathop{\Sigma}_{n\in B_r} \frac{1}{2^{n+1}}\]
By the definition of the binary expansion,
\[P(X_r)=\mathop{\Sigma}_{n\in B_r} \frac{1}{2^{n+1}}=r\]
\end{proof}
\begin{proposition}[Reimann and Slaman \cite{ReimannSlaman}]
\label{representation}
No $r$-random set $X$ can be computable from $r$.
\end{proposition}
\begin{proof}
Reimann and Slaman \cite[Proposition 2.3]{ReimannSlaman} says that if $Y\subseteq\omega$ is a representation of some measure $\mu$ on $2^\omega$, then $Y$ computes a function $g_\mu:2^{<\omega}\times\omega\to\mathbb{Q}$ such that for all $\sigma\in2^{<\omega}$ and all $n\in\omega$, \[|g_\mu(\sigma,n)-\mu(\sigma)|\leq 2^{-n}\]
Take $\mu$ to be $\mu_r$, the $r$-Bernoulli measure. Fix $\sigma_1$ to be $1$, the binary string of length one with first bit $1$. Then we have $\mu_r(\sigma_1)=r$. Thus the $Y$-computable function $g(n)=g_{\mu_r}(\sigma_1,n)$ has the property that for all $n\in\omega$,
\[|g_{\mu_r}(\sigma_1,n)-\mu_r(\sigma_1)|=|g(n)-r|\leq 2^{-n}\]
In other words, $g$ can be used to compute $r$. As $Y$ was arbitrary, every representation of $\mu_r$ computes $r$.
\\
\\
By definition (\cite[Definition 3.2]{ReimannSlaman}), a set $X$ is $r$-random if and only if it is random with respect to some representation $Y$ of $\mu_r$. Therefore $Y$ cannot compute $X$. However, by the above argument, $Y$ computes $r$. Thus $r$ cannot compute $X$.
\end{proof}
\begin{theorem}
For almost all $r$, $r$ computes a set $A$ which has intrinsic density $r$ but cannot compute an $r$-random set.
\end{theorem}
\begin{proof}
Let $r$ be 1-random. Then we may apply Theorem \ref{every} with $r$ playing the role of both $r$ and $X$ to obtain an $r$-computable set $A$ which has intrinsic density $r$. Every $A$-computable set is $r$-computable because $A$ is $r$-computable, but by Proposition \ref{representation}, no $r$-random set can be $r$-computable. Therefore, $A$ cannot compute an $r$-random set.
\\
\\
As almost all reals are 1-random, this completes the proof.
\end{proof}
From each such set, we may use Theorem \ref{theoremwithin} to generate more examples.
\section{Closing Remarks and Future Work}
We showed a computational separation between the notions of intrinsic density and randomness by constructing examples of sets with intrinsic density $r$ which cannot compute $r$-random sets. There is room for future work in investigating whether this can be done for all $r$, or at least for all noncomputable $r$.
\\
\\
Additionally, there is room for applying these techniques to compare other notions of stochasticity with randomness. Many of the results proved for intrinsic density have analogs for MWC stochasticity, however it is currently open whether infinite unions work in that setting as they do for intrinsic density. Thus the analog of Theorem \ref{every} for MWC stochasticity remains open.
| b202d48d69969bf696cb2e5ceea7178ab8cdd747 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The field of network medicine has gained tremendous traction in recent years. With the growth of public disease datasets and development in computational methods, several recent works have attempted to use data-driven methods to study disease progression \cite{OXT18,KHO19}. Network medicine \cite{6,8,9,10,SON19} offers innovative data-driven solutions to disease modeling through analyzing complex interactions within data. For instance, \cite{10} identifies disease-related genetic variants through network-based studies of the human disease network (i.e. the ‘diseaseome’), in which diseases are linked if they share one or more associated genes. Functional brain networks for neurobiologically relevant graphical properties of Alzheimers disease are analyzed in \cite{JAL17}. However, network-based patient subtyping based not only on disease variable values but also on their trajectories, i.e., evolution of variable patterns, is relatively unexplored.
Parkinson’s Disease, a degenerative neurological disorder, is the second most common neurodegenerative disorder following Alzheimer’s disease, affecting an estimated 7-10 million people worldwide \cite{1}. It is both chronic, meaning it persists over a long duration, and progressive, meaning symptoms, such as tremor, loss of memory, impaired balance etc. worsen with time \cite{SCH06,LEW05}. In addition, it is a highly variable disease, with rate and type of progression differing significantly across the population \cite{2}. It is increasingly evident that Parkinson's disease is not a single entity but rather a heterogeneous disorder with multiple subtypes. Several studies have attempted to classify patients into subtypes \cite{KRI20, SAU16, VAN11, SEL09}.Two recent studies have developed models of PD progression based on clinical, demographic and genetic data at baseline, using hierarchical cluster analysis \cite{5} and a Bayesian multivariate predictive inference platform \cite{16}, to identify PD subtypes. However, there is no overall consensus on Parkinson's subtypes \cite{5}, and little is known about the effect of the interplay between different types of variables on Parkinson's progression. Early recognition of patient subtype allows medical workers to employ subtype-specific treatment, potentially improving and prolonging life \cite{MAR13,THE14}. However, current approaches in Parkinson's subtyping have largely not explored data-driven analyses to unravel multiple complex influences of a large number of longitudinal variables on disease progression \cite{24}.
Sophisticated data-driven methods are required to extract meaningful information effectively from medical datasets. In recent years, several deep learning approaches to disease-subtyping have emerged \cite{ZHA19,MIO18}. However, such approaches are limited in extracting interpretable information about the micro-structure of the subtypes, i.e., the variable relationships that underlie the subtype. In contrast, network science approaches offer an intuitive visualization for modeling relationships between different types of variables. Additionally, evolving interactions between variables are easily represented through a multi-layer structure.
Significant research has been done in clustering in a multi-layer network \cite{DONG12,KIM15}. However, subtyping through trajectory clustering is relatively unexplored in network medicine\cite{14,15}. This work presents a novel multi-layer-network-based Trajectory Clustering (TC) algorithm to identify disease subtypes based on similarities in trajectories through variable clusters. First, it identifies variable-communities (clusters of variables that co-express) through modeling patient-variable interactions as a bipartite network. It then tracks patient-memberships through multiple layers of variable-communities to define their trajectories. Lastly, disease subtypes are identified by clustering similar trajectories. To the best of the author's knowledge, the only other work with a trajectory-based approach to Parkinson's subtyping is the author's previous work \cite{KRI20}. While their research questions and data are the same as those in this paper, \cite{KRI20} defines trajectories as matrices (and implements a trajectory \textit{profile} clustering) which is fundamentally different than the method proposed in this paper, and yields different, although complementary, results. Additionally, it is a first-order method. In contrast, this paper defines trajectories through a multi-layer network and identifies subtypes using a graph-based second-order trajectory clustering. The novel contributions of this approach are multifold: (1) a unique subtyping approach on graphs with intuitive visualization of trajectories, (2) identification of co-expressing variable clusters in the stacked-multi-layer network offering a variable-centric perspective to disease progression (3) studying disease progression as a function of \textit{any} outcome variable (with results presented for outcome variable MDS-UPDRS3). The contributions of this work are distinct from and complementary to those in \cite{KRI20}.
The contributions of this work are (a) extracting clusters of co-expressing variables at different stages of disease, and (b) identifying subtypes based on similarities in disease trajectories. The set of variable clusters identified by our network are in agreement with Parkinson's disease domains recognized in medical literature \cite{MAR18}. Disease subtypes identified are shown to be statistically distinct and are supported by the results in \cite{KRI20}. This easily generalizable approach for multi-layer trajectory clustering presents a unique way for extracting patterns of disease progression in complex longitudinal medical data, and helps bridge the gap between data-based computational approaches and applied medicine.
\section{Data}
Data used in the preparation of this article was obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data). It consists of patient variable values across 5 timepoints: baseline year (denoted as year0) and years-1,2,3, and 4. Upon excluding patients with incomplete data, 194 patients remained in the analysis.The dataset categorizes the clinical variables into domains (Cognitive, Behavioral, Sleep, PD Severity, Autonomic, Disability) based upon functionality as shown in \ref{data}. Variable data is obtained from commonly used standard medical tests, a few of which are outlined below. JOLO is a standardized test of visuospatial skills. SDM is a test measuring neurological information processing speed and efficiency using a symbol/digit substitution tasks. HVLT is a word-learning test that measures episodic verbal memory and recall ability. SEADL assesses the difficulties patients have completing daily activities or chores. EPS test scores give a subjective measure of a patient's daytime sleepiness. STAI is used in clinical settings to diagnose anxiety and to distinguish it from depressive syndromes. The MDS - UPDRS scale includes series of ratings for typical Parkinson’s symptoms that cover all of the movement hindrances of Parkinson’s disease and consist of Mentation, Behavior and Mood, Activities of Daily Living, Motor sections etc.
According to PPMI, motor assessment for variables in this dataset was performed in a `practically defined off' state, i.e., subjects are asked to withhold medication prior to the assessment for 12 hours, practically eliminating medication effects on motor symptoms.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs1}
\end{center}
\caption{Description of the six medical domains and their variables in the PPMI clinical dataset.}
\label{data}
\end{figure}
\section{Network Architecture: Stacked Multi-layer Network}
\label{bip}
Variable-communities within each layer are identified and stacked to create a multi-layer framework across which patient trajectories are defined. This proceeds as follows:
\begin{enumerate}
\item Within each layer, individual-variable relationships for $I$ patients and $V$ relationships, are modeled as a bipartite network. Bipartite networks are graphs consisting of two sets of disjoint and independent nodes, such that there exist no edges between nodes belonging to the same set. They are a natural choice for modeling relationships between two different classes (in this case patients variables). The adjacency matrix $Z$ of the bipartite network is created as follows:
\begin{itemize}
\item The `direction' of each variable is determined. If disease progression (how severely a patient is affected) is positively correlated with values of the disease variable $v$) then the direction of the variable $d_v$ is set as +1, and if disease-affection is negatively correlated with disease variable values, then $d_v = -1$. These directions are standard and well known for variables in the PPMI clinical data. Variables ESS, RBDQ, GDS, STAI, MDS-UPDRS-1,2,3,T and Age have $d_v =1$ and HVLT, JOLO, SFT, LNS, SDM, MoCA, SEADL have $d_v =-1 $.
\item The value ($F_{iv}$) of variable $v$ in individual $i$ is then converted to a z value, $z_{iv}$ that normalizes it between $0$ and $1$,
\begin{equation}
z_{iv} =\frac{F_{iv} - \min_{j \in I}F_{jv}}{\max_{j \in I}F_{jv}-\min_{j \in I}F_{jv}}.
\end{equation}
These z values are continuous and non-thresholded.
\item The adjacency matrix $Z$ of size $I \times V$ for the patient-variable bipartite network is populated as
\begin{equation}
Z_{ij} =
\cases{
z_{iv} \quad &if $d_v=1$ \\
1-z_{iv} \quad &if $d_v=-1$.}
\end{equation}
\end{itemize}
\item An independent patient-variable bipartite network is generated for each layer (an illustrative bipartite network is shown in figure \ref{mult_bip} (left)), where a layer could represent a time-point or a range of outcome variable values. An outcome variable is simply any variable as a function of which one may desire to measure disease progression as indicated by non-outcome variables. This paper presents results for outcome variable chosen to be MDS-UPDR3. Then Louvain community detection \cite{18} is implemented on the bipartite network for each layer $l$. Louvain community detection yields the number of communities through optimization of the Newman-Girvan modularity function \cite{19} and has no hyperparameters. It identifies variable-communities $C_k^l$, where $k \in [0, \ldots, K_l-1]$, comprising of patients and variables (shown through shaded ovals in figure \ref{mult_bip} (left)), and $K_l$ is the total number of variable-communities in layer $l$.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs2}
\end{center}
\caption{(left) shows a representative bipartite patient-variable network for one layer. The highlighted ovals represent variable-communitites consisting of patients and disease variables. (right) represents a stacked multi-layer graph over three layers, where the variable-communities from (left) form the first layer. Three sample trajectories are shown, and the corresponding closeness between nodes $X$ and $Y$ is calculated.}
\label{mult_bip}
\end{figure}
The TC algorithm to cluster patient trajectories (and hence patients) is presented in sec. \ref{tc_bip}.
\item These variable-communities in each layer, variable-community represented as a node, are stacked across all layers as shown in figure~\ref{mult_bip}(right). This is the stacked multi-layer graph ($G$). Each individual belongs to a variable-community $C_k^l$ at each time-point. One can then track individual $i$ membership over time to define a patient trajectory $T_i$.
\end{enumerate}
\section{Trajectory Clustering Algorithm}
\label{tc_bip}
Each node in $G$ is a variable-community. Trajectory Clustering (TC) is a second-order algorithm identifies patient subtypes based by clustering patients with similar trajectories of disease progression on the stacked multi-layer network $G$ as follows:
\begin{enumerate}
\item `Node closeness' $cl(X,Y)$ between two nodes $X$ and $Y$ in graph $G$ is defined as the fraction of whole or partial trajectories passing through $X$ and $Y$ that overlap. Suppose $A$ and $B$ are sets of all edges (both inter-layer and intra-layer), with repetition, that belong to trajectories passing through node $X$ and node $Y$ respectively, then node closeness between is defined as
\begin{equation}
cl(X,Y)=\frac{\#(A \cap B )}{\#( A \cup B)}.
\end{equation}
where $\#$ denotes the size of a set. Note that $cl(X,X)=1.$ Figure~\ref{mult_bip} (b) presents a simple example of node closeness calculated between nodes $X$ and $Y$. There is one overlapping edge-pair {$A_2, B_2$} on trajectories passing through nodes $X,Y$, and the total size of the set of edges on these trajectories is $3$. Hence, $cl(X,Y)=1/3$.
\item Trajectory Similarity ($\eta(T_i,T_j)$) between two individual (patient) trajectories $T_i$ and $T_j$ is then given by $\Sigma_{t}cl(X(t)_{T_i},Y(t)_{T_j})$, where $X(t)_{T_i}$ is the node $X(t)$ that trajectory $T_i$ passes through at timepoint $t$. $\eta(T_i,T_j)$ measures not just first-order similarities between two trajectories, but also second-order similarities i.e., interactions between two non-overlapping trajectories that result from their mutual overlap with a third set of trajectories. Important properties of the second-order TC algorithm are outlined in \ref{timex}.
\item A patient-patient network $P$ is created whose adjacency matrix is given by
\begin{equation}
P_{i,j}= \eta(T_i,T_j).
\end{equation} .
The network P contains information of how similar patients are in terms of their disease trajectories. $P_{i,j}$ gives the trajectory similarity $\eta(T_i,T_j)$ between patients $i$ and $j$. Thus a higher value of $P_{i,j}$ indicates that patients $i$ and $j$ are connected with a higher weight in the network $P$ and are correspondingly more likely to have similar disease progression patterns. $P$ is a symmetric matrix with $P_{i,i}=1$.
\item Louvain community detection is implemented on the network P (for 100 runs), and the maximally occuring communtiy configuration over all runs was chosen to obtain $M$ trajectory-communities $S_k$ (where $k \in [0, \ldots M-1]$) with similar disease progression trajectories. In a modularity maximizing algorithm such as Louvain \cite{18}, the optimal number of trajectory-communities are unconstrained and automatically identified to correspond to the configuration with maximal modularity (a commonly used metric for community detection in networks).
\end{enumerate}
\section{Results}
\subsection{Bipartite variable-communities}
\begin{figure*}[ht!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs4}
\end{center}
\caption{Community profiles of the variable-communities in each layer. Here, the layers represent time-points. Each panel represents one variable-community comprising of individuals and variables. These communities are characterized by the variables in them, listed above each panel. The panels show the average $z$ value of all people in that variable-community, and error bars are given by their standard deviation.}
\label{multi}
\end{figure*}
Upon creating an individual-variable network for each layer (where each layer is a timepoint) as in figure \ref{bip} (left), Louvain community detection is performed on each layer. Figure ~\ref{multi} shows the communities $C_l$ identified in each layer $l$ comprising of individuals as well as dominant variable. These variable-communities are characterized by the dominant disease variables that belong to them (listed above each variable-community in figure~\ref{multi}).
The variable-communities are indicative of sets of variables that patients co-express with high severity in different stages of the disease. For instance, the last panel (bottom row) in the baseline year indicates that in early stages of disease, patients displaying severity in variable GDS for instance, are also likely to display severity in STAI.
While variables are treated independently in this analysis, the emergent variable-community structure is largely consistent with domain structure in medical literature \cite{MAR18}. The composition of variable-communities remains similar across the layers. Age and JOLO (a cognitive variable) cluster together in all layers. Another variable-community comprises of all the other cognitive variables. Disability (SEADL) and General PD Severity Variables (PD-1,2,3,T) commonly co-express with high severity together in all layers except in the baseline year where PD-1 is found to co-express with Autonomic (SCOPA-AUT) and Sleep (RBDQ) variables. As General PD Severity variables represent motor functioning, their co-occurence with disability is expected. Sleep and Autonomic variables often clustered together in several layers. Mental Health variables (GDS, STAI) initially form an independent variable-community, however, as time progresses, they co-express with high severity with PD Severity and Disability domains. Deterioration of mental health is expected with worsening motor skills and increased disability, particularly in later stages of disease progression. The clustering of variables in the same domain is further validation of our method yielding medically relevant results. The individuals within each variable-community have relatively high average severity of corresponding variables that characterize the community.
\subsubsection{Statistical analysis}
The Kruskal-Wallis statistical test \cite{KRU52} is an estimate of whether two variables are sampled from the same distribution. Kruskal-Wallis test for multiple groups was conducted to validate this approach and demonstrate some of the differences between variable-communities. This test is chosen because it is applicable in cases such as this where values for several of the variables violate the normality assumption.
\begin{table*}[htbp!]
\caption{\label{stat1} Comparing the variable-communities (described in figure~\ref{multi}) in each layers (representative of timepoints) through the Kruskal-Wallis p-values. }
\begin{indented}
\item[]
\lineup
\begin{tabular}{ ccccccc }
\br
Variable & Median & Year0 & Year1 & Year2 & Year3 & Year4 \\
\mr
Age & 71.00&1.759E-08 & 2.439E-07 & \textbf{2.263E-09} & 2.966E-06 & 5.115E-08 \\
\textbf{Cognitive} &&&&&&\\
JOLO* & 14.00& 3.687E-06 & 6.807E-06 & 8.491E-05 & 2.769E-03 & 1.526E-09 \\
SDM* & 42.00&3.951E-07 & 3.764E-09 & 7.653E-07 & 5.050E-08 & 7.435E-12 \\
SFT* & 47.50&5.151E-07 & 1.920E-06 & 1.381E-04 & 3.276E-09 & 2.753E-07 \\
HVLT* & 0.90& \textbf{3.222E-03} & 6.110E-07 & \textbf{5.731E-03} & \textbf{2.959E-01} & 7.282E-05 \\
LNS* & 11.00&1.207E-05 & 1.457E-08 & 1.347E-07 & 1.234E-11 & 5.982E-11 \\
MOCA* & 27.00& 1.767E-05 & 3.470E-05 & 1.163E-06 & 2.379E-04 & 2.928E-08 \\
\textbf{Other} &&&&&&\\
SEADL* &95.00& 4.317E-10 & 5.634E-10 & 1.039E-07 & 6.277E-07 & 3.821E-08 \\
RBDQ & 4.00&2.931E-08 & 1.247E-06 & 1.212E-09 & 2.699E-10 & 1.172E-11 \\
ESS & 6.00& \textbf{1.265E-01} & 6.166E-08 & 6.262E-08 & 4.623E-10 & 1.191E-08 \\
SCOPA-AUT & 9.00& 6.022E-09 & 4.087E-07 & 5.470E-10 & 8.695E-09 & 2.638E-10 \\
GDS & 2.00&6.678E-06 & 1.323E-08 & 7.281E-05 & 2.941E-08 & 9.002E-09 \\
STAI & 65.00&1.805E-07 & 4.617E-09 & 1.374E-03 & 6.388E-07 & 2.223E-08 \\
\textbf{General PD} &&&&&&\\
PD-1 & 5.50&3.609E-10 & 8.432E-09 & 1.789E-09 & 3.372E-12 & 4.205E-12 \\
PD-2 &5.00& 1.471E-08 & 1.969E-09 & 1.231E-08 & 1.599E-12 & 2.570E-15 \\
PD-3 & 20.00&2.587E-08 & 1.088E-08 & 1.026E-10 & 2.275E-09 & 4.794E-12 \\
PD-T &32.00& 1.674E-14 & 5.236E-14 & 1.280E-14 & 9.158E-16 & 5.981E-17\\
\br
\end{tabular}
\end{indented}
\end{table*}
Table~\ref{stat1} tabulates the medians and the p-values of the Kruskal-Wallis statistical test applied to each layer. Medians are calculated from the whole population raw data. Variables with negative directions are denoted by an asterisk (*).
The p-values are a measure of statistical differences amongst the variable values of the individuals belonging to the different variable-communities. To account for Type I errors due to multiple comparisons the Benjamini-Hochberg False Discovery Rate method \cite{BEN95} is used. This gives us an adjusted significance level for each of the p-values $\alpha_{adjusted} = \frac{\alpha \times i}{n_c}$, where $n_c$ is the total number of comparisons, and $i$ is the rank of the p-value (for example, the smallest has a rank of 1, the second smallest has a rank of 2 etc.). Significance value $\alpha = 0.05$. Total number of comparisons $n_c = T \times V = 5 \times 17 = 85$, where $T$ is the number of timepoints, and $V$ is the number of variables. For instance, in Table~\ref{stat1}, HVLT in year 3 has the largest p-value (rank one), hence it has the strictest acceptance threshold given by $\alpha_{adjusted} = \frac{0.05 \times 1}{85} = 5.882 \times 10^{-3}$. Comparisons not meeting the criteria for statistical significance are shown in bold text. A majority of the p-values are below their adjusted significance level, suggesting that there is an underlying statistical difference between the variable distributions in the variable-communitites in a layer.
\subsection{Trajectory Clustering on Temporally Stacked Multi-layer Network}
\label{timex-cluster}
\subsubsection{Trajectory communities}
Each individual is a member of one of the variable-communities at each time-point (or in each layer in this case). Tracing their membership yields their trajectory. The TC algorithm (sec. \ref{bip}) identifies patient subtypes based on similarities in their trajectories through the multi-layer network. The run time of the trajectory clustering algorithm on the temporally stacked multi-layer network is $4.9346s$. All code was run on a laptop with 2.8 GHz Intel Core i5 processor and 8 GB RAM.
\begin{figure*}[htbp!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs5}
\end{center}
\caption{Each node represents a variable-community consisting of variables and patients. The size of the node denotes the number of people in the variable-community. The variables are color-coded by domain. Patient trajectories from the baseline layer to the year4 layer are clustered using the TC algorithm. The trajectory-communities (subtypes) are color-coded and directed (from left to right). The thickness of a colored edge denotes the number of patients flowing along that edge in the corresponding trajectory-community. Number of people in each subtype are as follows: green - 56, orange - 55, red - 53, blue - 33. Edges with fewer than 3 people are not plotted.}
\label{TS_bip}
\end{figure*}
As shown in figure~\ref{TS_bip}, the TC algorithm identifies four distinct trajectory-communities (subtypes) for disease progression across temporal layers. The edges are directed from left to right (baseline year to year4), shown without arrows for ease of viewing. The thickness of the edges is a measure of the flow of people between the corresponding variable-community nodes. Thus, edge thickness is an estimate of the probability of transition of patients expressing high severity from one variable set to another as time progresses.
The green subtype $S_0$ with 56 patients is characterized by an older population and higher severity of JOLO, as well as increase in general Cognitive impairment as time progresses. The orange subtype $S_1$ with 55 patients displays high values of Cognitive variables. In addition, it progresses to high severity in General PD Severity and Sleep variables. The red subtype $S_2$ with 53 patients is characterized by high Disability and severe General PD (motor), and develops severity in Mental Health variables through progression of time. Lastly, the blue subtype $S_3$, the smallest trajectory-community with 33 patients, is characterized by severe Autonomic and Sleep variables. It shows limited Cognitive and General PD severity in earlier years, and the severity of expression of these variables reduces over time.
\subsubsection{Trajectory-community profiles}
\begin{figure*}[ht!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs6}
\end{center}
\caption{Trajectory profiles of the 4 trajectory-communities. The $i^{th}$ vertical set of panels corresponds to a trajectory-community denoted by $S_i$). Within each column, the five panels arranged from top to bottom show the average profile in years 0(baseline),1,2,3,4 respectively. Each individual panel shows the average $z$ values across all variables of $\kappa_i^l$ patients in a trajectory-community.}
\label{v_prof}
\end{figure*}
Figure ~\ref{v_prof} shows the trajectory profile (weighted average profile of all individuals belonging to a trajectory-community) across all layers (which is equivalent to timepoints in this case). The trajectory profiles ($S_i^t$) of trajectory-community $i$ at timepoint $t$ is obtained by taking the weighted average of the variable-communities that members of the trajectory-community belong to as follows:
\begin{equation}
S_i^t=\frac{\Sigma_{i} X(t)_{T_i}}{\kappa_i^t}.
\label{tc_pr}
\end{equation}
where $i \in \kappa_i^t$ and $\kappa_i^t$ is the total number of individuals in trajectory-community $S_i$ and layer $t$. $X(t)_{T_i}$ is the variable-community that individual $i$ belongs to at time $t$. As seen in figure~\ref{v_prof}, each trajectory-community has a distinct trajectory profile with different evolution patterns and high values of their corresponding dominant variables.
The green trajectory-community $S_0$ with 56 members has a relatively mild disease stage, despite having a relatively older population. Disability, Autonomic, and General PD severity start very low, and continue to stay low over time. Amongst the cognitive variables, JOLO is slightly higher than in other trajectory-communities through the layers, ESS starts high and reduces over time, whereas HVLT gets more severe.
The orange trajectory-community $S_1$ with 55 members show consistently high severity in all Cognitive variables except JOLO and MoCA. Other variables are low; Disability reduces across the layers, whereas Sleep variables show a slight increase as disease progresses across time.
The red trajectory-community $S_2$ with 53 members has relatively severe General PD Severity motor variables. As time progresses, Autonomic, Mental Health and Disability demonstrate growth in severity.
Lastly, the blue trajectory-community $S_3$, the smallest with 33 members show heterogeneity in the variable structure. General PD Severity variables remain consistently low through time, whereas Sleep variables RBDQ and ESS show growth in severity across the layers.
\subsubsection{Statistical analysis}
To validate the algorithm and demonstrate the differences in the \textit{progression} of the different trajectory-communities, the Kruskal-Wallis statistical test for multiple groups was conducted on the \textit{difference} between the consecutive year variable values. This test is chosen because it is applicable in cases such as this where values for several of the variables violate the normality assumption. We do not include the variable `age' in the statistical analysis since the difference in age between consecutive years is exactly 1 regardless of the subtype.
\begin{table}
\caption{ \label{stat2}
Comparing evolution of trajectory-communities by calculating the Kruskal-Wallis p-values for variable differences between consecutive years.}
\begin{indented}
\item[]
\lineup
\begin{tabular}{ cccc }
\br
Variable & Year1-Year0 & Year2-Year1 & Year3-Year2 \\
\mr
\textbf{Cognitive} &&&\\
JOLO* & 1.451E-06 & 7.295E-04 & 3.335E-11 \\
SDM* & 2.579E-05 & \textbf{2.556E-02} & 6.247E-05 \\
SFT* & 2.089E-05 & 9.683E-07 & 1.005E-05 \\
HVLT* & 4.225E-04 & 1.892E-05 & 1.548E-18 \\
LNS* & 3.724E-06 & \textbf{8.711E-02} & 3.486E-04 \\
MOCA* & 2.562E-04 & 5.158E-07 & 3.142E-13 \\
\textbf{Other} &&&\\
SEADL* & \textbf{1.092E-02} & 1.809E-07 & 5.618E-08 \\
RBDQ & 9.470E-04 & 3.796E-09 & \textbf{2.015E-02} \\
ESS & 6.881E-15 & 1.059E-07 & 1.708E-04 \\
SCOPA-AUT & 3.035E-03 & 1.202E-05 & 3.529E-11 \\
GDS& 4.778E-08 & 6.070E-03 & 4.111E-05 \\
STAI & 3.209E-07 & \textbf{4.996E-02} & 2.330E-05 \\
\textbf{General PD} &&&\\
MDS-UPDRS-1& 3.648E-03 & 2.545E-05 & \textbf{2.418E-02} \\
MDS-UPDRS-2 & 6.627E-03 & 3.092E-04 & 2.503E-03 \\
MDS-UPDRS-3 & 2.502E-03 & 1.405E-07 & 2.590E-06 \\
T-MDS-UPDRS& \textbf{9.561E-03} & 2.809E-04 & 3.607E-03 \\
\br
\end{tabular}
\end{indented}
\end{table}
Table~\ref{stat2} tabulates the medians and the p-values of the Kruskall-Wallis statistical test to study statistical differences in the growth of variables across layers between individuals belonging to the different variable-communities. Variables with negative directions are denoted by an asterisk (*). To account for Type I errors due to multiple comparisons the Benjamini-Hochberg False Discovery Rate method \cite{BEN95} is used. This gives us an adjusted significance level for each of the p-values $\alpha_{adjusted} = \frac{\alpha \times i}{n_c}$, where $n_c$ is the total number of comparisons, and $i$ is the rank of the p-value. Significance value $\alpha$ is chosen to be $0.05$. Total number of comparisons $n_c = N_t \times N_v = 3 \times 16 = 46$, where $N_t$ is the number of year-differences, and $N_v$ is the number of variables. Comparisons that do not meet the significance criteria are highlighted in bold in the table. A majority of the values are below their adjusted significance level, suggesting that there exist significant differences in variable progression of the different trajectory-communities.
\subsection{Trajectory Clustering Across an Independent Outcome Variable}
\label{pdx}
Often, medical practitioners may be interested in studying the evolution of disease variables as a function of an outcome variable (a clinical test or variable known to be indicative of specific aspects of disease progression that are of interest). This may, in particular, be useful when complete temporal data about a patient is unavailable, however comprehensive data of evolution of a specific variable is available instead. Defining layers through values of an outcome variable, allows us to identify subtypes based on trajectories through progression of an outcome variable.
\begin{figure*}[htbp!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs7}
\end{center}
\caption{Trajectory clustering across the outcome variable - baseline MDS-UPDRS3 / PD3 values. The layers represent quartiles of the outcome variable raw values. The boundary values that determine the layers are denoted in red. Each node represents a variable-community consisting of variables and patients. The size of the node denotes the number of patient-years associated with the community. Each patient can contribute upto 5 times (one for each year) to a variable-community. The variables are color-coded by domain in the legend. Patient trajectories directed from the baseline year to the year4 are clustered using the TC algorithm. These trajectories progress in time and hence, are not required to pass through all the outcome variable layers. The trajectory-communities (subtypes) are color-coded. The thickness of a colored edge denotes the number of patients along that edge in the corresponding trajectory-community. Number of people in each trajectory-community are as follows: green - 47, orange - 44, red - 58, blue - 48. }
\label{PD_1}
\end{figure*}
\begin{figure*}[htbp!]
\begin{center}
\includegraphics[width=0.95\linewidth]{Images/all_figs/all_figs8}
\end{center}
\caption{(a) Variable-community profiles in each layer. Here, the layers represent values of the outcome variable MDS-UPDRS-3. Each panel represents a variable-community comprising of patient-years and variables. The variables in each variable-community are listed above each panel, and are used to identify the variable-communities. The panels show the average $z$ value of all patient-years in that variable-community. (b)Trajectory profiles of the 4 trajectory-communities. The $i^{th}$ vertical set of panels corresponds to a trajectory community $S_i$. Within each column, the four panels arranged from top to bottom show the average profile in layers low, low-mid, mid-high, and high. Each individual panel shows the average $z$ values across all variables of $\kappa$ patient-years in a trajectory-community in a layer. All error bars are given by the standard deviation. }
\label{PD_2}
\end{figure*}
MDS-UPDRS-3 (divided into quartiles on the x axis) is known to be indicative of PD progression and was chosen as the outcome variable. In figure~\ref{PD_1}, four disease progression trajectory-subtypes across outcome variable layers are identified. The layers are defined by raw baseline MDS-UPDRS-3 values as follows: layer low contains values below $22.0$, layer low-mid contains values between $22.0$ and $31.0$, layer mid-high contains values between $31.0$ and $38.0$, and layer high contains values above $38.0$. Dependent variables are chosen to be the same as in section \ref{timex}, with the exception of General PD severity variables. This method is generalizable to selection of any outcome variable.
Figure~\ref{PD_1} shows the subtypes identified by the trajectory clustering algorithm. The run time of the TC algorithm on the multi-layer network stacked across outcome variables is $1.7213s$. The size of the nodes in figure~\ref{PD_1} indicate the number of patient-years in that variable-community, i.e., one patient may contribute up to $5$ times to a certain community, one for each timepoint. This implies that a patient may, in consecutive years, remain in the same variable-community, or transition to a different variable-community in the same layer, or transition to a different variable-community in a different layer.
Figure~\ref{PD_2}(a) shows the variable-community profile, i.e., average $z$ values of all the patients in each community in each layer with errors given by their standard deviation. The position of variable-community profiles corresponds to the node in figure~\ref{PD_1}. In figure~\ref{PD_2}(a), one can see that the variables denoting a variable-community have a correspondingly higher relative $z$ value. In figure~\ref{PD_2}(b) shows the average trajectory profile as calculated in \ref{tc_pr} amongst for each trajectory-community across each layer.
As seen in figure~\ref{PD_1} and figure~\ref{PD_2}(b),The green trajectory-community $S_0$ with $47$ patients largely remains in low to low-medium PD-3 layers, with $101$ patient-years in layer low, and $72$ patient-years in layer low-medium. Individuals in this trajectory-community start with low disease state (as measured by PD-3), and show slow disease progression over the years, staying in the low-medium range of disease progression.
The orange trajectory-community $S_1$ with $44$ patient-years is the least affected. Patients are most severely affected in the Cognitive domain, however they remain largely in the low disease layer.
The red trajectory-community $S_2$ with $58$ starts in a relatively high disease state, and consistently gets worse, whereas the blue trajectory-community $S_3$ with $48$ starts relatively healthy but with severe disease progression into mid-high and high layers of disease severity measured through PD-3.
\section{Conclusion}
This work introduces a novel algorithm for the identification of subtypes based on relationships between trajectories through a multi-layer network. Multidimensional clinical datasets are often not used to their full potential due to the complexities of a cohesive analysis. This work extracts disease variables that co-express with high severity at different stages of disease progression. Then it extracts trajectories of progression through these variable-communities, i.e., through sets of high severity variables. Lastly, it identifies disease subtypes through clustering patient trajectories. Additionally, it promotes second-order comparisons in the calculation of $\eta$ i.e., correlations across time between two trajectories that do not have overlapping edges but interact with common neighboring trajectories. The agreement of variable-clusters with the domain-structure e.g. Cognitive, Autonomic etc., as well as the statistical analysis, both validate the success of this approach.
Parkinson's Disease has a multitude of clinical variables that interact in complex manners that vary through stages of the disease. A number of studies have identified Parkinson's subtypes based on baseline characteristics \cite{5,20,22}. In contrast, our novel algorithm uses longitudinal data (or patient trajectory over time) to identify disease subtypes through TC. In other words, our method accounts for both disease variable values as well as their progression patterns as a patient progresses through the different layers of the multi-layer stacked-bipartite network. While Parkinson's Disease, being both multivariate and progressive, served to demonstrate the effectiveness of the TC algorithm, the algorithm can, in principle, be easily extended to include non-clinical features, such as genetic or fluid biomarkers, as well as to other datasets.
The TC algorithm identifies subtypes through clustering patient trajectories across layers. This work presents two methods of analysis and visualization of disease progression: Parkinson's disease trajectory through time, and Parkinson's disease trajectory through an outcome variable (representative of disease progression). Our second-order TC algorithm emphasizes the dynamical aspect of disease progression in addition to the static properties of the associated variables at every stage, contrary to several earlier data-driven methods in medicine that emphasize one or the other. Results are presented in an interpretable visualization that is easily accessible to and comprehensible by medical practitioners in contrast with black box methods in machine learning.
The TC algorithm is a data-driven network-based method for detection of patient subtype. Like other data-driven methods, this analysis is limited by the availability and quality of the database. The number of variables provided in the database is not exhaustive in the context of Parkinson’s disease. As larger datasets are made available, such results are likely to be more informative and robust. Additionally, in applying a data-driven approach to medical data, important medical decisions must always be made in conjunction with medical expertise. An advantage as well as caveat of this approach is that trajectories are tracked through variable sets, and not through individual variables. Future directions of such work would naturally include extension to other types of medical data (genetic \cite{17}, biomarker etc.), as well as extension to other types of time-evolving and heterogenous datasets. Additional directions of potential interest include studying disease evolution through other outcome variables, as well as treatment-based modifications to the algorithm where effects of treatment are expected during gathering of data. By the nature of such a data-driven method, this approach may not be appropriate for datasets with high variability, low quantity, or data with inconsistent temporal sampling.
Parkinson's disease is a highly variable disease with a long onset time. Knowledge of which cluster a new patient belongs to in the baseline year, would allow medical practitioners to predict their subtype and corresponding trajectory, including the type and rate of disease progression. This could open up new and exciting avenues in the field of personalized medicine. Moreover, prediction of disease progression will improve prognostic counseling, a problem commonly encountered by clinicians, by highlighting predicted disease features. It will also support them in seeking more aggressive treatment for patients predicted to display rapid disease progression. Thus, this multi-layer network-based TC approach harnesses data to provide interpretable solutions in the field of early, predictive medicine.
\section*{Data and Code Availability}
Data was obtained from PPMI, a public-private partnership funded by the Michael J. Fox Foundation for Parkinson’s Research and funding partners, including abbvie, Allergan, Avid Radiopharmaceuticals, Biogen, Biolegend, Bristol-Myers Squibb, Celgene, Denali, GE Healthcare, Genentech, gsk, Lilly, Pfizer, Merck, MSD, Lundbeck, Piramal, Prevail Therapeutics, Roche, Sanofi Genzyme, Servier, Takeda, Teva, Ucb, Verily, Voyager Therapeutics and Golub Capital. There are no patents, products in development or marketed products to declare. Python code is available at \textit{www.github.com/chimeraki/Multilayer-Trajectory-Clustering}
\ack{
The author wishes to sincerely thank Prof. Michelle Girvan, Dr. Lisa M. Shulman, MD and Dr. Rainer von Coelln, MD for helpful discussions and suggestions.}
\section*{References}
\bibliographystyle{unsrt}
| 1bdbfc60e05ae12095c3a5849dd4470cc97b0c36 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction\label{sec:intro}}
The Drell-Yan (DY) process with pions and nucleons provides
important information on the structure of pion and nucleon.
The DY differential cross section in the region of low transverse
momentum, $q_T$, of the produced lepton anti-lepton pair is
subject to the transverse momentum dependent factorization
\cite{Collins:1984kg}.
The corresponding transverse momentum dependent parton distribution functions
(TMDs) \cite{Collins:2011zzd} in the description of DY at low $q_T$ provide
essential information on correlations between transverse parton momenta and
parton or nucleon spin, and describe the three-dimensional structure of
hadrons. Early theoretical studies of TMDs
in hadron production in proton-proton processes
\cite{Sivers:1989cc,Anselmino:1994tv,Mulders:1995dh} were followed
by systematic investigations in semi-inclusive deep-inelastic
scattering (SIDIS) \cite{Cahn:1978se,Kotzinian:1994dv,
Kotzinian:1995cz,Bacchetta:2006tn} and DY
\cite{Tangerman:1994eh,Boer:1997nt,Arnold:2008kf}
(also fragmentation functions \cite{Metz:2016swz}
enter the description of SIDIS).
The basis for these descriptions are QCD factorization theorems
\cite{Collins:1981uk,Collins:1984kg,Qiu:1991pp,Ji:2004wu,Ji:2006br,Ji:2006vf,
Collins:2011zzd,Aybat:2011zv,Ma:2013aca,Collins:2014jpa,Collins:2016hqq}.
One of the challenges when interpreting pion-induced DY data is the limited
knowledge of the pion structure. At twist-2 the process is described by
the proton TMDs: unpolarized distribution $f_{1,p}^a$, transversity
distribution $h_{1,p}^a$, Sivers distribution function $f_{1T,p}^{\perp a}$,
Boer-Mulders distribution $h_{1,p}^{\perp a}$, Kotzinian-Mulders distribution
$h_{1L,p}^{\perp a}$, and ``pretzelosity'' distribution $h_{1T,p}^{\perp a}$,
and pion TMDs: unpolarized distribution $f_{1,\pi}^a$,
Boer-Mulders distribution $h_{1,\pi}^{\perp a}$.
On the proton side, for $f^a_{1,p}$ both collinear and TMD distributions
are well-known \cite{Gluck:1991ng,Gluck:1994uf,Gluck:1998xa,Martin:2009iq,
Harland-Lang:2014zoa,Dulat:2015mca,Landry:2002ix,Anselmino:2013lza,
Signori:2013mda,Bacchetta:2019sam,Scimemi:2019cmh}.
Based on global QCD analyses of data, parametrizations are available also for
$f_{1T,p}^{\perp a}$, $h_{1,p}^a$, $h_{1,p}^{\perp a}$, $h_{1T,p}^{\perp a}$ \cite{Anselmino:2011gs,Anselmino:2013vqa,Barone:2009hw,Lefky:2014eia,Cammarota:2020qcw}.
Only $h_{1L,p}^{\perp a}$ has not yet been extracted,
though it can be described based on $h_{1,p}^a$ in the
so-called Wandzura-Wilczek- (WW-)type approximation
which is compatible with available data~\cite{Bastami:2018xqd}.
On the pion side the situation is different. While extractions of $f_{1,\pi}^a$
exist \cite{Gluck:1991ey,Sutton:1991ay,Gluck:1999xe,Aicher:2010cb,Barry:2018ort,Novikov:2020snp},
no results on $h_{1,\pi}^{\perp a}$ are available. This constitutes
a ``bottleneck'' if one would like to describe the pion-induced DY data,
e.g.\ COMPASS results \cite{Aghasyan:2017jop},
based solely on phenomenological extractions since
$h_{1,\pi}^{\perp a}$ is relevant for the majority of observables
in the pion-induced polarized DY process at leading twist.
In this situation we will resort to model studies of the pion Boer-Mulders function
$h_{1,\pi}^{\perp a}$.
An important goal of theoretical studies in models is to describe
hadron structure at a low initial scale $\mu_0< 1\,{\rm GeV}$ in terms
of effective constituent quark degrees of freedom. This approach has been
successful in describing various hadronic properties in terms
of ``valence-quark degrees of freedom.'' The underlying idea is that
at a low hadronic scale $\mu_0$, for example the properties of the nucleon
can be modelled in terms of wave functions of valence $u$ and $d$ quarks,
and similarly the properties of the $\pi^-$ in terms of the wave
functions of valence $\bar u$ and $d$ quarks. It is an interesting task
in itself to apply such a framework to the description of hadronic properties
like TMDs. This has been done in a variety of complementary approaches
including chiral quark models~\cite{Weigel:1999pc} and generalizations~\cite{RuizArriola:2003bs}, spectator models (SPMs)~\cite{Jakob:1997wg,Gamberg:2007wm,Gamberg:2009uk,
Bacchetta:2008af,Lu:2004hu}, light-front constituent quark model (LFCQM)~\cite{Pasquini:2008ax,Pasquini:2010af,Lorce:2011dv,
Boffi:2009sh,Pasquini:2011tk,Pasquini:2014ppa,Lorce:2014hxa,Lorce:2016ugb}
or bag models~\cite{Yuan:2003wk,Avakian:2008dz,Courtoy:2008vi,Courtoy:2008dn,
Avakian:2010br}. Phenomenological studies in the LFCQM showed that within a model accuracy of 20-30$\%$ a good description of SIDIS and unpolarized DY data can be
obtained \cite{Boffi:2009sh,Pasquini:2011tk,Pasquini:2014ppa}.
The goal of the present work is to study the spin and azimuthal asymmetries in the DY process with pions and polarized
nucleons, and to present calculations for all twist-2 asymmetries.
We use available phenomenological extractions of TMDs and
calculations from two well-established
constituent-quark-models (CQM), the LFCQM
and the SPM. Other studies in models, perturbative QCD and lattice QCD
of the pion-induced DY or relevant TMDs have been reported
\cite{Noguera:2015iia,Engelhardt:2015xja,Lambertsen:2016wgj,Chang:2018pvk,
Bacchetta:2017vzh,Broniowski:2017gfp,Ceccopieri:2018nop}.
Several features distinguish our work from other studies. First,
we use two CQM frameworks with diverse descriptions of the
pion and nucleon structure. Second, we describe all leading-twist
observables in pion-induced polarized DY entirely in the models.
Third, we supplement our studies with ``hybrid calculations'', where we
use as much as possible information from phenomenological analyses, and
only the
Boer-Mulders function $h_{1,\pi}^{\perp a}$ is taken from models.
Overall, we present up to four different calculations
for each observable. This allows us to critically assess model dependence,
and uncertainties in our approach. Where available the results are
compared to the COMPASS DY data \cite{Aghasyan:2017jop}.
One key aspect in our study is the evolution of model results from
the low hadronic scales to experimentally relevant scales. For that
(i) knowledge of the low initial scale, and (ii) applicability of evolution
equations at low scales are crucial. Both requirements are fulfilled in
the case of parton distribution functions which depend on one scale only,
the renormalization scale~$\mu$.
First, the value of the initial quark model scale $\mu_0$ can be consistently
determined by evolving the fraction of nucleon momentum carried by valence
quarks, $M_2^{\rm val}(\mu) = \sum_a \int dx\,x(f_1^q-f_1^{\bar q})(x,\mu)$,
known from parametrizations, using DGLAP evolution down to that scale
$\mu_0$ at which valence quarks carry the entire nucleon momentum,
i.e.\ $M_2^{\rm val}(\mu_0) = 1$ \cite{Traini:1997jz}.
Numerically it is $\mu_0\sim 0.5\,{\rm GeV}$.
Second, works by the GRV and GRS groups on parametrizations of nucleon and
pion unpolarized parton distribution functions show remarkable perturbative
stability between LO and NLO fits indicating applicability of DGLAP
evolution down to initial scales as low as $\mu_0^2 = 0.26\,{\rm GeV}^2$ \cite{Gluck:1991ng,Gluck:1991ey,Gluck:1994uf,Gluck:1998xa,Gluck:1999xe}.
TMDs depend not only on the renormalization scale $\mu$ but also on
the rapidity scale $\zeta$ \cite{Collins:2011zzd}. The theoretical and
phenomenological understanding of TMDs witnessed an incredible rate of
developments in the recent years including NNLO and NNNLO calculations
of the evolution kernel of unpolarized TMDs
\cite{Gehrmann:2014yya,Echevarria:2015byo,Echevarria:2015usa,Echevarria:2016scs,Li:2016ctv,Vladimirov:2016dll,Luo:2019hmp,Luo:2019szz,Ebert:2020yqt},
NLO calculations for the quark helicity distribution
\cite{Gutierrez-Reyes:2017glx},
NLO \cite{Gutierrez-Reyes:2017glx} and NNLO \cite{Gutierrez-Reyes:2018iod}
calculations for transversity and pretzelosity,
and NLO calculations for the Sivers function
\cite{Ji:2006ub,Koike:2007dg,Sun:2013hua,Dai:2014ala,Scimemi:2019gge}.
Recently also the first non-trivial expression for the small-$b$
expansion of the pretzelosity distribution was derived \cite{Moos:2020wvd}.
However, in the context of quark model applications we face two challenges.
First, no rigorous (analog to the $\mu_0$-determination) criterion exists
to fix the value of the initial rapidity scale $\zeta_0$ of quark models,
though an educated guess may be $\zeta_0\sim\mu_0^2$.
Secondly, in the case of Collins-Soper-Sterman (CSS) or TMD evolution~\cite{Collins:1984kg,Collins:2011zzd},
no expertise is available analogous
to the GRV/GRS applications of DGLAP evolution starting from low hadronic scales.
In this situation in previous quark model studies,
TMD evolution effects were often estimated approximately
\cite{Boffi:2009sh,Pasquini:2011tk,Pasquini:2014ppa} based on
an heuristic Gaussian Ansatz for transverse parton momenta with
energy dependent Gaussian widths. While providing a useful
description of data on many processes including pion-induced Drell-Yan
\cite{Schweitzer:2010tt}, it is important to improve the simple Gaussian treatment in view of the recent progress in the TMD theory
\cite{Gehrmann:2014yya,Echevarria:2015byo,Echevarria:2015usa,Echevarria:2016scs,Li:2016ctv,Vladimirov:2016dll,Luo:2019hmp,Luo:2019szz,Ebert:2020yqt,Gutierrez-Reyes:2017glx,Gutierrez-Reyes:2018iod,Ji:2006ub,Koike:2007dg,Sun:2013hua,Dai:2014ala,Scimemi:2019gge,Moos:2020wvd}.
We will therefore use TMD evolution~\cite{Collins:2011zzd}
at Next-to-Leading Logarithmic (NLL) precision
to describe the transverse momentum dependence of the Drell-Yan process.
At present, application of TMD evolution at the low quark
model scales below 1 GeV is not known. Therefore, we shall proceed in two steps. We will evolve weighted
transverse moments of TMDs from the low initial scale $\mu_{0}^{2}$ to a scale of $Q_0^2=2.4\,{\rm GeV}^2$
where phenomenological information on transverse momentum dependence
is available from TMD fits~\cite{Su:2014wpa, Kang:2014zza,Kang:2015msa, Bacchetta:2017gcc,Scimemi:2019cmh,Bacchetta:2019sam} of polarized and unpolarized SIDIS, DY and weak boson productions data. Then we use NLL TMD evolution to evolve to the scales relevant in the
COMPASS Drell-Yan measurements, i.e., $\langle Q^2\rangle=28\,{\rm GeV}^2$.
In this way we will be able to test the $x$-dependencies of the model TMDs while the $q_T$-dependencies of the DY observables are described on the basis of TMD fits.
For completeness we remark that the importance of TMD evolution
for the description of pion-induced DY and the recent COMPASS
data was also studied in
Refs.~\cite{Wang:2017zym,Vladimirov:2019bfa,Li:2019uhj,Ceccopieri:2018nop,
Wang:2018naw,Wang:2018pmx}.
Our results serve several purposes. They help to interpret in their
full complexity the first COMPASS data \cite{Aghasyan:2017jop} on the
pion-induced polarized DY process, and in this way deepen the understanding
of the QCD description of deep-inelastic processes in terms of TMDs.
They also provide quantitative tests of the application of CQMs
to the description of pion and nucleon structure.
\section{Drell-Yan process with pions and polarized protons}
\label{sec-2}
In this section we briefly review the DY formalism, and
provide the description of the DY structure functions in our approach.
\subsection{Structure functions}
\label{sec-2.1}
In the tree-level description a dilepton $l,\,l^\prime$ is produced from
the annihilation of a quark and antiquark carrying the fractions $x_\pi$,
$x_p$ of the longitudinal momenta of respectively the pion and the proton.
The process is shown in the Collins-Soper frame in Fig.~\ref{Fig-01:kinematics}.
In the case of pions colliding with polarized protons the DY cross
section is described in terms of six structure functions \cite{Arnold:2008kf},
\begin{eqnarray}
F_{UU}^1 &=& \phantom{-}
{\cal C} \bigg[f_{1,\pi}^{\bar a} \; f_{1,p}^{a}\;\bigg], \nonumber \\
F_{UU}^{\cos 2\phi} &=& \phantom{-}
{\cal C} \bigg[ \frac{2(\hhat \cdot \vkTpi)(\hhat \cdot \vkTN)
- \vkTpi \cdot \vkTN}{M_\pi \; M_p} \;
h_{1,\pi}^{\perp \bar a} \; h_{1,p}^{\perp a} \; \bigg], \nonumber \\
F_{UL}^{\sin 2\phi} &=&
- {\cal C} \bigg[ \frac{2(\hhat \cdot \vkTpi)(\hhat \cdot \vkTN)
- \vkTpi \cdot \vkTN}{M_\pi \; M_p} \;
h_{1,\pi}^{\perp\bar a}\; h_{1L, p}^{\perp a} \; \bigg], \nonumber \\
F_{UT}^{\sin \phi_S} &=& \phantom{-}
{\cal C} \bigg[ \frac{\hhat \cdot \vkTN}{M_p} \;
f_{1,\pi}^{\bar a} \; f_{1T, p}^{\perp a} \; \bigg], \nonumber \\
F_{UT}^{\sin (2\phi - \phi_S)} &=&
- {\cal C} \bigg[ \frac{\hhat \cdot \vkTpi}{M_\pi} \;
h_{1,\pi}^{\perp\bar a} \;h_{1, p}^{a} \; \bigg], \nonumber\\
F_{UT}^{\sin (2\phi + \phi_S)} &=&
- {\cal C} \bigg[ \frac{2(\hhat \cdot \vkTN)
\big[2(\hhat \cdot \vkTpi)(\hhat \cdot \vkTN)
- \vkTpi \cdot \vkTN \big]
- \vkTN^2(\hhat \cdot \vkTpi)}{2 \; M_\pi \; M_p^2} \,
h_{1, \pi}^{\perp\bar a}\, h_{1T, p}^{\perp a} \; \bigg]. \quad \quad \label{Eq:SFs}
\end{eqnarray}
The subscripts indicate the hadron polarization which can be unpolarized $U$
(pions, protons), longitudinally $L$, or transversely $T$ polarized (protons).
The azimuthal angles $\phi$, $\phi_S$ are defined in
Fig.~\ref{Fig-01:kinematics}, where the unit vector
$\hhat=\qT/q_T$ points along the $x$-axis. Notice that in
the Collins-Soper frame the dilepton is at rest, and
each incoming hadron carries the transverse momentum
$\qT/2$, see Fig.~\ref{Fig-01:kinematics}.
The convolution integrals in Eq.~\eqref{Eq:SFs}
are defined as~\cite{Arnold:2008kf}
\begin{equation}
\label{Eq:convolution-integral}
{\cal C}\bigl[\omega \, f_{\pi}^{\bar a} \, f_p^{a}\bigr] = \frac{1}{N_c} \;
\sum_a e_a^2\int d^2\kTpi \, d^2\kTN \, \delta^{(2)}(\qT- \kTpi-\kTN) \,
\omega \, f_{\pi}^{\bar a}(x_\pi,\kTpi^2)f_p^{a}(x_p,\kTN^2)\,,
\end{equation}
where $\omega$, which is a function of the transverse momenta $\kTpi$, $\kTN$
and $\qT$, projects out the corresponding azimuthal angular dependence.
The sum over $a=u,\,\bar{u},\,d,\,\bar{d},\,\dots$ includes the active flavors.
\begin{figure}[t!]
\centering
\includegraphics[height=6cm]{\FigPath/Fig00a-CS-frame-COMPASS-3rd.pdf}
\caption{\label{Fig-01:kinematics}
The DY process in the Collins-Soper frame where the pion
and the proton come in with different momenta $P_\pi$, $P_p$, but each
carries the same transverse momentum $\frac12\,\qT$, and
the produced lepton pair is at rest. The angle $\phi$ describes the
inclination of the leptonic frame with respect to the hadronic plane,
and $\phi_S$ is the azimuthal angle of the transverse-spin vector
of the proton.
}
\end{figure}
This partonic interpretation of DY is based on a TMD factorization~\cite{Collins:1984kg,Collins:2011zzd} and applies to the region $q_T\ll Q$.
The TMDs depend on renormalization and rapidity scales which are not
indicated for brevity in (\ref{Eq:SFs}) and (\ref{Eq:convolution-integral}), and will be discussed in section~\ref{sec-2.2}.
The focus of our work is on asymmetries of the kind
\begin{equation}
\label{Eq:Asym.def}
A_{XY}^{\rm weight}(x_\pi, x_p , q_T, Q^2)=
\frac{F_{XY}^{\rm weight}(x_\pi, x_p , q_T, Q^2)}{F_{UU}^1(x_\pi, x_p , q_T, Q^2)},
\end{equation}
where various types of higher order corrections tend to largely
cancel out \cite{Ratcliffe:1982yj,Weber:1991wd,Vogelsang:1992jn,
Contogouris:1994ws,Gehrmann:1997pi,Bunce:2000uv,Shimizu:2005fp}.
The $Q^2$ dependence of the structure functions and asymmetries will often
not be explicitly indicated for brevity.
In the following we will display results for the asymmetries as functions
of one of the variables $x_\pi$, $x_p$, $q_T$. It is then understood that
the structure functions are integrated over the other variables
within the acceptance of the experiment, keeping in mind that $x_\pi, \, x_p$
are connected to each other by $x_\pi \, x_p = Q^2/s$, where $s$ is the center of mass energy squared.
\subsection{QCD evolution of Drell-Yan structure functions}
\label{sec-2.2}
The basis for the evolution are TMD factorization theorems~\cite{Collins:1981uk,Collins:1984kg,Qiu:1991pp,Ji:2004wu,Ji:2006br,Ji:2006vf,Collins:2011zzd,Aybat:2011zv,Ma:2013aca,Collins:2014jpa,Collins:2016hqq,Scimemi:2018xaf}
which constrain the operator definition
and define the QCD evolution of TMDs.
Here we will adopt the CSS framework
and use the TMD evolution formalism starting from a fixed scale $Q_0$
\cite{Collins:2014jpa} in the structure functions from Eqs.~\eqref{Eq:SFs}.
The evolution of TMDs is a double-scale problem, and
can be implemented in momentum space or impact-parameter space with examples for both approaches in the literature
\cite{Collins:1981va,Aybat:2011zv,Angeles-Martinez:2015sea,Scimemi:2019cmh,Bacchetta:2019sam,Ebert:2020dfc}.
In our work we choose to implement the TMD evolution
in the impact-parameter space with ${\boldmath b_{T}}$ the
Fourier-conjugate variable to $\kTh$
where index $h = \pi$ or $p$ refers to pion or nucleon.
The TMDs in the impact-parameter
space are generically given by $\tilde f(x_h,{ b_{T}},\mu,\zeta)$ where
$\mu\sim Q$ is the ``standard'' renormalization scale for ultraviolet logarithms,
and $\zeta\sim Q^2$ is the rapidity renormalization scale.
In principle one can solve TMD evolution equations starting from some
initial scale $Q_0$ without employing operator product expansion at low
${\boldmath b_{T}}$, Ref.~\cite{Collins:2014jpa}.
The TMD at this initial scale is then $f(x_h,{b_{T}},Q_0,Q_0^2)$.
In this formulation the unpolarized structure function is similar to
parton model result and is
expressed as~\cite{Collins:2014jpa}
\begin{align}
\label{FUUbspace} F_{UU}^1(x_\pi,x_p,q_T,Q^2) &= \frac{1}{N_c} \sum_a e_a^2 {\cal H}^{(DY)}(Q,\mu_Q)\int \frac{b_{T} d b_{T}}{2\pi} J_0 (q_T b_{T}) \nonumber \\
& \times f_{1,\pi}^{\bar a}(x_\pi,{b_{T}},Q_0,Q_0^2) \tilde f_{1,p}^a(x_p,{b_{T}},Q_0,Q_0^2)\, e^{-S(b_{T},Q_0,Q,\mu_Q)}\; ,
\end{align}
where the factor $S(b_T,Q_0,\mu_Q)$ contains important effects of gluon
radiation with $S(b_T,Q_0,Q_0) = 0$ by construction~\cite{Collins:2014jpa}. The hard factor ${\cal H}(Q,\mu_Q)$ is~\cite{Collins:2017oxh}
\begin{align}
{\cal H}^{(DY)}(Q,\mu_Q) = 1 + \frac{\alpha_s(\mu_Q)}{2 \pi} C_F \left( 3 \ln \left( \frac{Q^2}{\mu_Q^2}\right) - \ln^2 \left( \frac{Q^2}{\mu_Q^2}\right) +\frac{7\pi^2}{6} - 8\right)\; + {\cal O} (\alpha_s^2) ,
\label{eq:hard}
\end{align}
where $C_F=4/3$ and $\alpha_s$ is the strong coupling constant.
One can parametrize TMDs at initial scale $Q_0$ as
\begin{equation}
\label{Ansatzbspace}
\tilde f_{1,p}^{a}(
x_p,{ b_{T}},Q_0,Q_0^2) = f_{1,p}^{a} (
x_p,Q_0) \;
e^{-\frac14 b_{T}^2
\la k_{Tp}^2 \ra_{f_{1,p}}} \,,
\end{equation}
\begin{equation}
\label{Ansatzbspace1}
\tilde f_{1,\pi}^{a}(x_\pi,{ b_{T}},Q_0,Q_0^2) = f_{1,\pi}^{a} (x_\pi,Q_0) \; e^{-\frac14 b_{T}^2\la k_{T \pi}^2 \ra_{f_{1,\pi}}} \,,
\end{equation}
where $x$-dependent functions correspond to collinear distributions and the exponential factors are ``primordial shapes'' of TMDs at the initial scale. This particular dependence is often used in phenomenology \cite{DAlesio:2004eso,Schweitzer:2010tt},
corresponds to the Gaussian Ansatz and is supported in models
\cite{Avakian:2010br,Efremov:2010mt,Pasquini:2011tk,Pasquini:2014ppa,Schweitzer:2012hh}. The average widths of TMDs may be flavor-
and $x$-dependent and will be taken from
phenomenological parametrizations at $Q_0^2$.
Based on the $b_T$ space formalism given in Ref.~\cite{Boer:2011xd} we write down the rest of the twist-2 structure functions. We use the convenient notation from Ref.~\cite{Bacchetta:2019qkv},
\begin{align}
{\cal B}_n[\tilde f_{\pi}\; \tilde f_{p}] & \equiv \frac{1}{N_c}
\sum_a e_a^2 {\cal H}^{(DY)}(Q,\mu_Q) \int_0^\infty \frac{d b_T\, b_T}{2\pi}\; b_T^n \, J_n(q_T b_T)
\nonumber\\ &\hskip -1cm \times \tilde f_{\pi}^{\bar a}(x_\pi,{b_{T}},Q_0,Q_0^2)
\; \tilde f_{p}^{ a}(x_p,{b_{T}},Q_0,Q_0^2) \; e^{-S(b_{T},Q_0,Q,\mu_Q)}\, ,
\label{Eq:convBess}
\end{align}
which leads to the following expressions for the twist-2 structure functions,
\begin{align}
{F}_{UU}^1(x_\pi,x_p,q_T,Q^2) &=
\phantom{-} {\cal B}_0[\tilde f_{1, \pi}\; \tilde f_{1,p}] \; ,
\label{e:upb} \\
{F}_{UU}^{\cos 2\phi}(x_\pi,x_p,q_T,Q^2) &=
\phantom{-} M_\pi M_p \; {\cal B}_2[\tilde h_{1,\pi}^{\perp(1)}\; \tilde h_{1,p}^{\perp(1)}]\; ,
\label{e:bmb} \\
{F}_{UL}^{\sin 2\phi}(x_\pi,x_p,q_T,Q^2)&=
- M_\pi M_p \; {\cal B}_2[\tilde h_{1,\pi}^{\perp(1)}\; \tilde h_{1L,p}^{\perp(1)}]\; ,
\label{e:kotzb} \\
{F}_{UT}^{\sin \phi_S}(x_\pi,x_p,q_T,Q^2) &=
\phantom{-} M_p \;{\cal B}_1[\tilde f_{1,\pi}\; \tilde f_{1T,p}^{\perp(1)}]\; ,
\label{e:Sivb}\\
{F}_{UT}^{\sin (2\phi -\phi_S)}(x_\pi,x_p,q_T,Q^2) &=
-M_\pi \;{\cal B}_1[\tilde h_{1,\pi}^{\perp(1)}\; \tilde h_{1,p}]
\label{e:bmtb}\; ,\\
{F}_{UT}^{\sin (2\phi + \phi_S)}(x_\pi,x_p,q_T,Q^2) &=
-\frac{M_\pi M_p^2}{4} \; {\cal B}_3[\tilde h_{1,\pi}^{\perp(1)}\; \tilde h_{1T,p}^{\perp(2)}]
\; ,\label{Eq:sfsb}
\end{align}
where the $b_T$ space TMD moments~\cite{Boer:2011xd} are
\begin{align}
\tilde f^{(n)}(x_h,b_T,Q,Q^2) &=(-1)^n n! \left(\frac{2}{M_h^2}\frac{\partial}{\partial b_T^2}\right)^n
\tilde f(x_h,b_T,Q,Q^2) \; .
\end{align}
These moments have the important feature,
\begin{align}
\lim_{b_T \to 0} \tilde f^{(n)}(x_h,b_T,Q,Q^2) = f^{(n)}(x_h,Q)\; ,
\end{align}
where $f^{(n)}$ are conventional transverse moments of TMDs~\cite{Mulders:1995dh} defined as
\begin{equation}
f^{(n)}(x_h,Q)=\int {\rm d} ^2\kTh\,\left(\frac{\kTh^2}{2M_h^2}\right)^n\,f(x_h,\kTh^2,Q,Q^2)\, ,
\label{e:TMDmoment}
\end{equation}
and $h=\pi, p$ corresponds to pion and proton TMDs, respectively.
The evolution factor $S(b_T, Q_0, \mu_Q)$ in Eq.~\eqref{Eq:convBess},
which results from
solving the CSS evolution equation and the renormalization group
equations for the rapidity dependence of the TMDs
and for the soft factor~\cite{Aybat:2011zv,Collins:2011zzd}, is given by
\begin{eqnarray}
S(b_T, Q_0, Q, \mu_Q) = -\tilde K(b_T,Q_0) \ln\frac{Q^2}{Q_0^2} +
\int_{Q_0}^{\mu_Q}\frac{d\bar\mu}{\bar\mu}\left[\gamma_K(\alpha_s(\bar\mu))\ln\frac{Q^2}{\bar\mu^2}-2 \gamma_i(\alpha_s(\bar\mu);1)\right]\, ,
\label{e:FF_ansatz}
\end{eqnarray}
\noindent
where $\tilde K$ is the Collins-Soper evolution kernel, and the anomalous dimensions are $\gamma_i(\alpha_s(\bar\mu);1)$ and $\gamma_K(\alpha_s(\bar\mu))$~\cite{Collins:2014jpa}.
Since the integral in Eq.~\eqref{Eq:convBess} extends over all $b_T$,
one cannot avoid using $\tilde K$ in the
CSS evolution factor~\eqref{e:FF_ansatz} in the non-perturbative large $b_T$ region.
In order to combine the perturbative and non-perturbative regions,
we use the $\bstarsc$ prescription~\cite{Collins:1984kg}, namely,
\begin{align}
\label{eq:bstar}
\bstarsc = \frac{b_T }{\sqrt{ 1 + b_T^2/b_{\rm max}^2}},
\end{align}
which introduces a smooth upper cutoff $b_{\rm max}$ in the transverse distance.
Then, the perturbative part of
$\tilde{K}$ is defined by replacing $b_T$ by $\bstarsc$ and the non-perturbative part is defined by the difference
$\tilde{K}(\bstarsc,\mu)-\tilde{K}(b_T,\mu)=g_K(b_T;b_{\rm max})~$\cite{Collins:2014jpa}.
Furthermore to combine the perturbative and non-perturbative regions
using the fixed scale evolution, it is optimal to use the renormalization group running scheme for $\tilde K$ in Eq.~\eqref{e:FF_ansatz}, evolved from the fixed scale $Q_0$, i.e.
\begin{eqnarray}
\tilde K(b_T,Q_0)=\tilde K(\bstarsc,\mubstar) -\int_{\mubstar}^{Q_0}\frac{d\bar\mu}{\bar\mu}\gamma_K(\alpha_s(\bar\mu)) -g_K(b_T;b_{\rm max}) \, ,
\end{eqnarray}
where $\mubstar$ is now chosen to become a hard scale,
\begin{eqnarray}
\mubstar \equiv \frac{C_1}{b_*}\, .
\end{eqnarray}
Now Eq.~\eqref{e:FF_ansatz} reads~\cite{Collins:2014jpa}
\begin{align}
S(b_T,\bstarsc, Q_0, Q, \mu_Q) &= \left( g_K(b_T;b_{\rm max})- \tilde K(\bstarsc;\mubstar)+
\int_{\mubstar}^{Q_0}\frac{d\bar\mu}{\bar\mu}\gamma_K(\alpha_s(\bar\mu)) \right)\ln\frac{Q^2}{Q_0^2}
\nonumber \\
+ & \int_{Q_0}^{\mu_Q}\frac{d\bar\mu}{\bar\mu}\left[\gamma_K(\alpha_s(\bar\mu))\ln\frac{\mu_Q^2}{\bar\mu^2}-2 \gamma_i(\alpha_s(\bar\mu);1)\right]\, .
\label{e:final}
\end{align}
The anomalous dimensions can be expanded as
perturbative series,
$\gamma_i =\sum_{n=1}^\infty \gamma_i^{(n)} \left(\alpha_s/\pi\right)^n$, and $\gamma_K =\sum_{n=1}^\infty \gamma_K^{(n)} \left(\alpha_s/\pi\right)^n$. In our calculations we employ them to NLL accuracy:
$\gamma_K^{(1)}$, $\gamma_K^{(2)}$ and $\gamma_i^{(1)}$.
They are spin-independent~\cite{Collins:1984kg,Qiu:2000ga,Landry:2002ix,Moch:2005id,Kang:2011mr,Aybat:2011zv,Echevarria:2012pw,Grozin:2014hna,Collins:2014jpa}, and given by
\begin{eqnarray}
\gamma_K^{(1)} =2 C_F,
\quad
\gamma_K^{(2)} ={C_F}\left[C_A\left(\frac{67}{18}-\frac{\pi^2}{6}\right)-\frac{10}{9}T_F \, n_f\right],
\quad
\gamma_i^{(1)} = \frac{3}{2} C_F
\, ,
\end{eqnarray}
where $C_F=4/3$, $C_A=3$, $T_F=1/2$ and $n_f$ is the number of active flavors.
The NLL two-loop contribution for $\tilde K$~\cite{Collins:2017oxh,Echevarria:2020hpy}, valid at small values of $b_T$, is
\begin{align}
\tilde K (b_*;\mubstar) &= -2 C_F \frac{\alpha_s(\mubstar)}{\pi} \ln \left(\frac{b_* \mu_{b_*}}{2 e^{-\gamma_E}}\right)
+
\frac{C_F}{2}\left(\frac{\alpha_s(\mubstar)}{\pi}\right)^2\Bigg[ \left( \frac{2}{3} n_f - \frac{11}{3}C_A\right) \ln^2 \left(\frac{b_* \mu_{b_*}}{2 e^{-\gamma_E}}\right)
\nonumber \\
&+
\left( -\frac{67}{9} C_A + \frac{\pi^2}{3}C_A + \frac{10}{9} n_f\right) \ln \left(\frac{b_* \mu_{b_*}}{2 e^{-\gamma_E}}\right) +
\left(\frac{7}{2}\zeta_3-\frac{101}{27}\right) C_A + \frac{28}{27} T_F n_f
\Bigg] \, ,
\label{e:CS_kernel0}
\end{align}
so that for $C_1 =2 e^{-\gamma_E}$, one finds
\begin{eqnarray}
\tilde K (b_*;\mubstar) &=& \frac{C_F}{2} \left(\frac{\alpha_s(\mubstar)}{\pi}\right)^2 \left[\left(\frac{7}{2}\zeta_3-\frac{101}{27}\right) C_A + \frac{28}{27} T_F n_f
\right] \, .
\label{e:CS_kernel}
\end{eqnarray}
We will numerically calculate the integral in Eq.~\eqref{e:FF_ansatz} using the two-loop result for the strong coupling constant, tuned to the world average~\cite{Bethke:2012jm} $\alpha_s(M_Z)= 0.118$ as in the CTEQ analysis~\cite{Hou:2019efy}.
Furthermore, we adopt the functional form of $g_K(b_T;b_{\rm max})$ given by Collins and Rogers~\cite{Collins:2014jpa},
\begin{eqnarray}
g_K(b_T,b_{\rm max})&=&g_0(b_{\rm max})\left(1-\exp\left[-\frac{C_F\alpha_s(\mubstar)b_T^2}{\pi g_0(b_{\rm max})b_{\rm max}^2}\right]\right),
\end{eqnarray}
which interpolates smoothly between the small and large-$b_T$ regions, where
at small $b_T$ it approximates a power series in $b_T^2$, while at large $b_T$ the resulting value of $\tilde K$ goes to a constant~\cite{Collins:2014jpa}. We choose $g_0(b_{\rm max})=0.84$ and $b_{max} = 1\,{\rm GeV}^{-1}$ to match the non-perturbative behavior of $g_K$ used in Refs.~\cite{Kang:2014zza,Kang:2015msa} to describe the polarized SIDIS data and in Ref.~\cite{Su:2014wpa} to describe unpolarized SIDIS, DY and weak boson production data.
The scale $\mu_Q$ is usually chosen such that $\mu_Q= C_2 Q $. In the following we will use $C_2 = 1$ and $C_1 = 2 e^{-\gamma_E}$. These choices allow one to optimize the accuracy of the perturbative expansion calculations in Eqs.~(\ref{eq:hard}) and (\ref{e:CS_kernel0})~\cite{Collins:2017oxh}.
\subsection{
Input for TMDs and choice of the initial scale \boldmath{$Q_0$}}
\label{sec-2.3}
We will utilize the following parametrizations~\cite{Bastami:2018xqd}
for TMDs at the initial scale $Q_0$:
\begin{eqnarray}
f_{h}^a(x_h,\kTh,Q_0,Q_0^2) &=&
f_{h}^a(x_h,Q_0) \; \frac{e^{- \kTh^2 / \la k_{Th}^2 \ra_{f_{h}} }}{\pi \; \la k_{Th}^2 \ra_{f_{h}} }\; ,
\quad \hspace{16mm}
f_{h}^a = f_{1, p}^a, \; f_{1, \pi}^a,\; h_{1, p}^a, \nonumber\\
f_{h}^{a}(x_h,\kTh,Q_0,Q_0^2) &=&
f_{h}^{(1) a}(x_h,Q_0) \, \frac{2 M_h^2}{\pi \la k_{Th}^2 \ra_{f_{h}}^2} e^{-\kTh^2/{\la k_{Th}^2 \ra_{f_{h}}}},
\quad
f_h^a = f_{1T, p}^{\perp a},\;h_{1, p}^{\perp a},\;h_{1, \pi}^{\perp a},\;h_{1L, p}^{\perp a},\nonumber\\
f_{h}^{a}(x_h,\kTh,Q_0,Q_0^2) &=&
f_{h}^{(2) a}(x_h,Q_0) \, \frac{2 M_h^4}{\pi \la k_{Th}^2 \ra_{f_{h}}^3} e^{-\kTh^2/{\la k_{Th}^2 \ra_{f_{h}}}},
\quad
f_h^a = h_{1T, p}^{\perp a}, \label{Eq:Gauss}
\end{eqnarray}
where transverse moments of TMDs are defined in Eq.~\eqref{e:TMDmoment}. Parametrizations from Eqs.~\eqref{Eq:Gauss} are often used in phenomenology to describe polarized SIDIS and DY data.
These parametrizations correspond to the following $b_T$-space expressions
\begin{eqnarray}
\tilde f_{h}^a(x_h,b_T,Q_0,Q_0^2) &=&
f_{h}^a(x_h,Q_0) \; e^{- \frac{1}{4} \, b_T^2 \la k_{Th}^2 \ra_{f_{h}}} \; ,
\quad \hspace{2mm}
\tilde f_{h}^{a} = \tilde f_{1, p}^a, \; \tilde f_{1, \pi}^a,\; \tilde h_{1, p}^a, \nonumber\\
\tilde f_{h}^{(1) a}(x_h,b_T,Q_0,Q_0^2) &=&
f_{h}^{(1) a}(x_h,Q_0) \, e^{-\frac{1}{4} \, b_T^2 {\la k_{Th}^2 \ra_{f_{h}}}},
\quad
\tilde f_h^{(1)a} = \tilde f_{1T, p}^{\perp (1) a},\;\tilde h_{1, p}^{\perp (1) a},\;\tilde h_{1, \pi}^{\perp (1) a},\;\tilde h_{1L, p}^{\perp (1) a},\nonumber\\
\tilde f_{h}^{(2) a}(x_h,b_T,Q_0,Q_0^2) &=&
f_{h}^{(2) a}(x_h,Q_0) \, e^{-\frac{1}{4} \, b_T^2{\la k_{Th}^2 \ra_{f_{h}}}},
\quad
\tilde f_h^{(2)a} = \tilde h_{1T, p}^{\perp (2) a}, \label{Eq:Gaussb}
\end{eqnarray}
where the collinear functions are the same as in Eqs.~\eqref{Eq:Gauss}.
Using Eqs.~\eqref{Eq:Gauss} or Eqs.~\eqref{Eq:Gaussb} one obtains for the convolution integrals
in Eqs.~(\ref{Eq:SFs}) or Eqs.~\eqref{e:upb}--\eqref{Eq:sfsb} the following results at the initial scale $Q_0$,
\begin{eqnarray}
F_{UU}^1(x_{\pi},x_p,q_T,Q_0^2)
&=& \phantom{-} \frac{1}{N_c} \sum_a e_a^2 \;
\;f_{1,\pi}^{\bar{a}} (x_{\pi},Q_0)\;f_{1,p}^a(x_p,Q_0)
\;\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra}, \nonumber\\
F_{UT}^{\sin\phi_S}(x_{\pi},x_p,q_T,Q_0^2)
&=& \phantom{-} \frac{1}{N_c} \sum_a e_a^2 \;
f_{1,\pi}^{\bar{a}} (x_{\pi},Q_0) \; f_{1T,p}^{\perp (1)a}(x_p,Q_0)
\; 2 M_p \; \frac{q_T }{\la q_T^2 \ra} \;
\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra}, \nonumber\\
F_{UT}^{\sin(2\phi-\phi_S)}(x_{\pi},x_p,q_T,Q_0^2)
&=& - \frac{1}{N_c} \sum_a e_a^2 \;
h_{1,\pi}^{\perp (1) {\bar{a}}} (x_{\pi},Q_0) \; h_{1,p}^a(x_p,Q_0)
\; 2 M_{\pi} \; \frac{q_T }{\la q_T^2 \ra} \;
\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra}, \nonumber\\
F_{UU}^{\cos 2\phi}(x_{\pi},x_p,q_T,Q_0^2)
&=& \phantom{-} \frac{1}{N_c} \sum_a e_a^2 \;
h_{1,\pi}^{\perp (1) {\bar{a}}} (x_{\pi},Q_0) \; h_{1,p}^{\perp (1) a}(x_p,Q_0)\;
4 M_{\pi}M_p \; \frac{q_T^2 }{\la q_T^2 \ra^2} \;
\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra},\nonumber\\
F_{UL}^{\sin 2\phi}(x_{\pi},x_p,q_T,Q_0^2)
&=& - \frac{1}{N_c} \sum_a e_a^2 \;
h_{1,\pi}^{\perp (1){\bar{a}}} (x_{\pi},Q_0) \; h_{1L,p}^{\perp (1)a}(x_p,Q_0)
\; 4 M_{\pi}M_p\; \frac{q_T^2 }{\la q_T^2 \ra^2} \;
\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra}, \nonumber\\
F_{UT}^{\sin(2\phi+\phi_S)}(x_{\pi},x_p,q_T,Q_0^2)
&=& - \frac{1}{N_c} \sum_a e_a^2 \;
h_{1,\pi}^{\perp (1){\bar{a}}} (x_{\pi},Q_0) \; h_{1T,p}^{\perp (2) a}(x_p,Q_0)
\; 2 M_{\pi}M_p^2\;
\frac{q_T^3 }{\la q_T^2 \ra^3} \;
\frac{e^{-q_T^2/\la q_T^2 \ra}}{\pi \; \la q_T^2 \ra}, \nonumber\\ \label{Eq:all-Gauss}
\end{eqnarray}
where the index $a=u,\,\bar{u},\,d,\,\bar{d},\,\dots$
and the mean square transverse momenta $\la q_T^2 \ra$ are defined in each case as
the sum of the mean square transverse momenta of the corresponding TMDs; that is in (\ref{Eq:all-Gauss})
in the first equation $\la q_T^2 \ra = \la k_{T \pi}^2 \ra_{f_{1,\pi}}+\la k_{Tp}^2 \ra_{f_{1,p}}$,
in the second equation $\la q_T^2 \ra = \la k_{T \pi}^2 \ra_{f_{1,\pi}}+\la k_{Tp}^2 \ra_{f^\perp_{1T,p}}$, etc.
In our study we will use the corresponding
transverse moments and Gaussian widths from TMD extractions
that we will take at the initial scale
which corresponds to the $\langle Q^2\rangle$ in the HERMES experiment. We will therefore use $Q_0^2 = 2.4$ GeV$^2$ as our initial scale of TMD evolution in Eqs.~\eqref{Eq:Gaussb} for parametrization of TMDs
and in Eqs.~\eqref{e:upb}-~\eqref{Eq:sfsb} for structure functions
that we will evolve to the scale of the COMPASS Drell-Yan measurement.
The scale $Q_0^2=2.4\,{\rm GeV}^2$ is convenient because parametrizations of
many TMDs are available at this scale, and there is a great deal of
expertise how to implement CSS evolution starting from $Q_0$.
However, results from quark models refer to a lower hadronic
scale $\mu_0 \sim 0.5\,{\rm GeV} < Q_0$. Presently it is not
known how to implement CSS evolution at such low scales (see section~\ref{sec:intro}). Thus, the evolution of model results from $\mu_0$
to the initial CSS scale $Q_0$ chosen in this work, is
regarded as a part of the modelling. It will be described
in detail in section~\ref{sec-2.5}.
\subsection{TMDs extracted from experimental data}
\label{sec-2.4}
In order to compute leading-twist structure functions in pion-induced
DY the knowledge of the proton and pion TMDs $f_{1,p}^a$,
$f_{1,\pi}^a$, $f_{1T,p}^{\perp a}$, $h_{1,p}^a$, $h_{1,p}^{\perp a}$,
$h_{1T,p}^{\perp a}$, $h_{1L,p}^{\perp a}$, $h_{1,\pi}^{\perp a}$
is required, which we list here in the order from the best to the
least known TMD, see Fig.~\ref{Fig-02:knowledge-TMDs} for an overview.
\begin{figure}[b!]
\centering
\includegraphics[height=5.3cm]{\FigPath/Fig-knowledge-TMDs-NEW-NEW.pdf}
\caption{\label{Fig-02:knowledge-TMDs}
TMDs entering the pion-induced polarized DY
process at leading twist in the order from
the phenomenologically best to least known,
and the approaches used in this work, see text.}
\end{figure}
While such a classification is to some extent subjective, it is evident that the collinear proton distributions $f_{1,p}^a(x_p)$
are the best known
\cite{Gluck:1998xa,Martin:2009iq,Harland-Lang:2014zoa,Dulat:2015mca}
thanks to DIS, DY and other data. We will utilize the MSTW extraction
of $f_{1,p}^a(x_p)$ \cite{Martin:2009iq} for comparison with models
and our calculations.
The unpolarized TMDs $f_{1,p}^a(x_p,\kTN)$ have been studied
and much progress was achieved in incorporating effects of QCD evolution
\cite{Landry:2002ix,Anselmino:2013lza,Signori:2013mda,Bacchetta:2019sam,Scimemi:2019cmh}
which are taken into consideration approximately in our approach
as described in section~\ref{sec-2.2}. For the collinear pion
distribution $f_{1,\pi}^a$, listed next in Fig.~\ref{Fig-02:knowledge-TMDs},
many extractions are available
\cite{Gluck:1991ey,Sutton:1991ay,Gluck:1999xe,Aicher:2010cb,Barry:2018ort,Novikov:2020snp}. We will use the MRSS fits \cite{Sutton:1991ay}.
One of the most prominent TMDs, the Sivers distribution $f_{1T,p}^{\perp a}$
was extracted from HERMES, COMPASS, and JLab SIDIS data by several groups
with consistent results
\cite{Anselmino:2010bs,Anselmino:2005ea,Anselmino:2005an,Collins:2005ie,Vogelsang:2005cs,Anselmino:2008sga,Bacchetta:2011gx,Anselmino:2011gs,Echevarria:2014xaa,Cammarota:2020qcw,Bacchetta:2020gko}. We will use the extractions of
Ref.~\cite{Anselmino:2011gs} labelled as ``Torino'' and
Ref.~\cite{Cammarota:2020qcw} labelled as ``JAM20''.
The transversity distribution, $h_{1,p}^a$, plays a crucial role in
understanding the nucleon spin structure. It is predicted to generate a transverse single spin
asymmetry in SIDIS coupling to the Collins fragmentation function
\cite{Collins:1992kk}, which is also responsible for an azimuthal asymmetry in $e^+e^-$ annihilation into hadron pairs.
We will use the ``Torino'' parametrizations of $h_{1,p}^a$
from a global QCD analysis of SIDIS and $e^+e^-$ data
\cite{Anselmino:2013vqa} to be compared with model predictions,
and the ``JAM20'' fit from a global QCD analysis of SIDIS, DY,
$e^+e^-$, and proton-proton data \cite{Cammarota:2020qcw} for
comparisons and calculations.
The proton Boer-Mulders function $h_{1,p}^{\perp a}$
extracted from HERMES, COMPASS and DY data
in Ref.~\cite{Barone:2009hw} will be used with the label
``BMP10.'' The extraction of $h_{1,p}^{\perp a}$
\cite{Barone:2009hw} is less certain, because
in SIDIS it requires model-dependent corrections
for sizable twist-4 contamination (Cahn effect).
The so-called pretzelosity function $h_{1T,p}^{\perp a}$ was extracted
in Ref.~\cite{Lefky:2014eia}. We will label $h_{1T,p}^{\perp a}$ from Ref.~\cite{Lefky:2014eia} as ``LP15''.
Notice that large errors on extracted $h_{1T,p}^{\perp a}$ were
reported in Ref.~\cite{Lefky:2014eia}.
This is the least known proton TMD for which an extraction has been attempted.
Only the Kotzinian-Mulders distribution $h_{1L,p}^{\perp a}$ has not
yet been extracted. It was found that the data related to this TMD \cite{Airapetian:1999tv,Jawalkar:2017ube,Parsamyan:2018evv}
are compatible with the WW-type approximation
\cite{Bastami:2018xqd} which we will use to approximate
$h_{1L,p}^{\perp a}$ based on $h_{1,p}^a$
from ``Torino'' \cite{Anselmino:2013vqa} and ``JAM20'' \cite{Cammarota:2020qcw} fits.
Finally, the pion Boer-Mulders function $h_{1,\pi}^{\perp a}$
is the least known of the TMDs needed to describe the pion-proton
DY process at leading twist. No extractions
are currently available for this TMD.
\subsection{TMDs from models}
\label{sec-2.5}
In this section we briefly review the two CQM frameworks, the LFCQM
and the SPM, and compare them in Figs.~\ref{Fig:pion-models} and
\ref{protonbasis-u+d} to the available phenomenological extractions
used in this work.
Light-front models are based on the decomposition of the hadron states in
the Fock space constructed in the framework of light-front quantization.
The hadron states are then obtained as a superposition of partonic
quantum states, each one multiplied by an $N$-parton light-front wave function
which gives the probability amplitude to find the corresponding $N$-parton
state in the hadron. In the LFCQM the light-front Fock expansion is truncated
to the leading component given by the valence $3q$ and $q\bar q$ contribution
in the proton and pion, respectively.
The light-front wave functions
can be further decomposed in terms of light-front wave amplitudes
that are eigenstates of the total parton orbital angular momentum.
The TMDs can then be expressed as overlap of light-front wave amplitudes
with different orbital angular momentum~\cite{Pasquini:2008ax} which makes
very transparent the spin-orbit correlations encoded in the different TMDs~\cite{Pasquini:2008ax,Pasquini:2010af,Boffi:2009sh,Pasquini:2011tk,Pasquini:2014ppa}.
To model the $3q$ light-front wave function
of the proton, the phenomenological Ansatz
of Ref.~\cite{Schlumpf:1992pp} was used, describing the quark-momentum dependence
through a rational analytical expression with parameters fitted to the
anomalous magnetic moment of the proton and
neutron~\cite{Schlumpf:1992pp,Pasquini:2007iz}.
For the pion, the $q\bar q$
light-front wave function of Ref.~\cite{Schlumpf:1994bc} was used,
with the quark-momentum dependent part given by a Gaussian function
with parameters fitted to the pion charge radius and decay constant.
\begin{figure}[b]
\centering
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{$x \, f_{1,\, \pi^-}^{\bar u}$} &\vspace{-.5mm} \hspace{-3mm}\raisebox{-.5\height}{\includegraphics[width=0.33\textwidth]{\FigPath/Evo_xf1ubpion}} & \hspace{5mm}\rotatebox[origin=c]{90}{$x \, h_{1,\, \pi^-}^{\perp (1) \bar u}$} &\vspace{-.5mm} \hspace{-3mm}\raisebox{-.5\height}{\includegraphics[width=0.33\textwidth]{\FigPath/Evo_xh1p1pion}} \\
& $x$ & & $x$ \\
\end{tabular}
\vspace{-3mm}
\caption{\label{Fig:pion-models}
Left: $f^{\bar u}_{1,\pi^-}$ from LFCQM \cite{Pasquini:2014ppa}
and SPM~\cite{Gamberg:2009uk} LO-evolved to the scale $Q_0$ in comparison to MRSS parametrization
\cite{Sutton:1991ay}.
Right: Predictions from LFCQM \cite{Pasquini:2014ppa} and
SPM \cite{Gamberg:2009uk} for the pion Boer-Mulders function
(with the sign for DY)
for which no parametrizations are currently available.}
\end{figure}
Spectator models are based on a field theoretical description of deep
inelastic scattering in a relativistic impulse approximation.
In this parton model-like factorization, the cross section for deep
inelastic scattering processes can be expressed in terms of a Born
cross section and quark correlation functions~\cite{Gamberg:2005ip}.
In this framework, the quark correlation functions are hadronic matrix
elements expanded in Dirac and flavor structure multiplying form factors.
The essence of the SPMs is to calculate the matrix elements of the quark
correlation function by the introduction of effective hadron-spectator-quark
(e.g. nucleon-diquark-quark) vertices~\cite{Meyer:1990fr,Jakob:1997wg,Gamberg:2007wm} which in turn enable one to model essential non-perturbative
flavor and spin structure of hadrons.
\begin{figure}[t]
\centering
\vspace{-5mm}
\begin{tabular}{ccccc}
\hspace{-5mm} \rotatebox[origin=c]{90}{$x \, f_{1,p}^u$} &\vspace{-.5mm} \hspace{-2mm}\raisebox{-.5\height}{\includegraphics[width=0.3\textwidth]{\FigPath/Evo_xf1u}} & & \hspace{2mm} \rotatebox[origin=c]{90}{$x \, f_{1,p}^d$} &\vspace{-.5mm} \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=0.3\textwidth]{\FigPath/Evo_xf1d}} \\
& $x$ & & & $x$ \\
\hspace{-7mm} \rotatebox[origin=c]{90}{$x \, f_{1T,p}^{\perp (1) u}$} &\vspace{-.5mm} \hspace{-3mm}\raisebox{-.5\height}{\includegraphics[width=0.308\textwidth]{\FigPath/Evo_xf1Tp1u}} & & \rotatebox[origin=c]{90}{$x \, f_{1T,p}^{\perp (1) d}$} &\vspace{-.5mm} \hspace{-5mm}\raisebox{-.5\height}{\includegraphics[width=0.316\textwidth]{\FigPath/Evo_xf1Tp1d}} \\
& $x$ & & & $x$ \\
\hspace{-5mm} \rotatebox[origin=c]{90}{$x \, h_{1,p}^u$} &\vspace{-.5mm} \hspace{-2mm}\raisebox{-.5\height}{\includegraphics[width=0.3\textwidth]{\FigPath/Evo_xh1u}} & & \rotatebox[origin=c]{90}{$x \, h_{1,p}^d$} &\vspace{-.5mm} \hspace{-4mm}\raisebox{-.53\height}{\includegraphics[width=0.309\textwidth]{\FigPath/Evo_xh1d}} \\
& $x$ & & & $x$ \\
\hspace{-7mm} \rotatebox[origin=c]{90} {$x \, h_{1,p}^{\perp (1) u}$} &\vspace{-.5mm} \hspace{-3mm}\raisebox{-.5\height}{\includegraphics[width=0.31\textwidth]{\FigPath/Evo_xh1p1u}} & & \rotatebox[origin=c]{90}{$x \, h_{1,p}^{\perp (1) d}$} &\vspace{-.5mm} \hspace{-4mm}\raisebox{-.5\height}{\includegraphics[width=0.31\textwidth]{\FigPath/Evo_xh1p1d}} \\
& $x$ & & & $x$ \\
\hspace{-10mm} \rotatebox[origin=c]{90}{$x \, h_{1T,p}^{\perp (2) u}$} &\vspace{-.5mm} \hspace{-6mm}\raisebox{-.5\height}{\includegraphics[width=0.326\textwidth]{\FigPath/Evo_xh1Tp2u}} & & \hspace{-1mm}\rotatebox[origin=c]{90}{$x \, h_{1T,p}^{\perp (2) d}$} &\vspace{-.5mm} \hspace{-6.5mm}\raisebox{-.5\height}{\includegraphics[width=0.326\textwidth]{\FigPath/Evo_xh1Tp2d}} \\
& $x$ & & & $x$ \\
\hspace{-10mm} \rotatebox[origin=c]{90}{$x \, h_{1L,p}^{\perp (1) u}$} &\vspace{-.5mm} \hspace{-5mm}\raisebox{-.5\height}{\includegraphics[width=0.318\textwidth]{\FigPath/Evo_xh1Lp1u}} & & \rotatebox[origin=c]{90}{$x \, h_{1L,p}^{\perp (1) d}$} &\vspace{-.5mm} \hspace{-5.5mm}\raisebox{-.5\height}{\includegraphics[width=0.318\textwidth]{\FigPath/Evo_xh1Lp1d}} \\
& $x$ & & & $x$ \\
\end{tabular}
\caption{\label{protonbasis-u+d}
The proton TMDs of $u$ and $d$ quarks in LFCQM \cite{Pasquini:2008ax,Boffi:2009sh,Pasquini:2011tk}
and SPM
\cite{Gamberg:2007wm} at
the scale $Q_0$ compared to phenomenological fits for $f_{1,p}$ from MSTW2008(LO) \cite{Martin:2009iq}, $f_{1T,p}^{\perp (1) a}$ from JAM20 \cite{Cammarota:2020qcw} and Torino \cite{Anselmino:2011gs}, $h_{1,p}^a$ from JAM20 \cite{Cammarota:2020qcw} and Torino \cite{Anselmino:2013vqa},
$h_{1,p}^{\perp (1) a}$ from BMP10 \cite{Barone:2009hw}, $h_{1T,p}^{\perp (2) a}$ from LP15 \cite{Lefky:2014eia}. Sivers and Boer-Mulders TMDs are shown with the sign for DY process. The error bands show the 1-$\sigma$ uncertainty of the JAM20 extractions \cite{Cammarota:2020qcw}.}
\end{figure}
The SPMs allow one to model the dynamics of universality and
process dependence through studying the gauge-link, and phase content of
TMDs~\cite{Brodsky:2002cx,Ji:2002aa,Goldstein:2002vv,Metz:2002iz,Gamberg:2003eg,Collins:2004nx,Gamberg:2008yt}.
In turn systematic phenomenological estimates for parton distributions
and fragmentation functions for both ``T-even'' and ``T-odd'' TMDs
have been carried out~\cite{Ji:2002aa,Goldstein:2002vv,Gamberg:2003ey,Boer:2002ju,Gamberg:2007wm,Bacchetta:2008af}.
In regard to the latter, it is in this framework that the first calculations
of the Sivers and Boer-Mulders functions of the nucleon were carried
out~\cite{Brodsky:2002cx,Ji:2002aa,Goldstein:2002vv}
and shown on general grounds to contribute to semi-inclusive processes
at leading power in the hard scale. Later the Boer-Mulders function
of the pion was calculated in Ref.~\cite{Gamberg:2009uk}.
The model parameters are determined by comparing the
SPM results for $f_{1,p}^u(x)$ and $f_{1,p}^d(x)$
to the LO low-scale ($\mu^{2}_0=0.26$ GeV$^2$)
GRV98 parametrization~\cite{Gluck:1998xa}.
The proton TMDs for $u$- and $d$- quarks are given by linear
combinations of contributions from axial-vector and scalar
diquarks assuming $\rm SU(2)$ flavor symmetry
\cite{Jakob:1997wg,Gamberg:2007wm}.
We choose the scale
$Q_0^2 = 2.4\,{\rm GeV}^2$
as the initial scale for the CSS evolution.
The evolution effects between the initial model scale $\mu_0\sim 0.5\,{\rm GeV}$ and $Q_0$ cannot be determined
exactly in the CSS formalism, see section~\ref{sec-2.3}, and they also cannot be neglected.
We therefore estimate them as follows.
We start with the model predictions for the parton
distributions or transverse moments of TMDs as they
appear in Eq.~(\ref{Eq:Gauss}) at the initial quark model
scale $\mu_0$. We evolve them using LO DGLAP evolution to
the scale $Q_0$. In contrast to CSS, experience with
implementing DGLAP evolution at low scales is available \cite{Gluck:1991ng,Gluck:1991ey,Gluck:1994uf,Gluck:1998xa,Gluck:1999xe}.
Hereby we use exact DGLAP evolution for $f_{1,h}^a(x)$ and $h_{1,p}^a(x)$.
In all other cases we use approximate DGLAP evolution: for the transverse
momenta of the proton Sivers function we use the $f_{1,h}^a(x)$-nonsinglet
evolution shown to lead good results in the LFCQM model study of SIDIS
asymmetries \cite{Pasquini:2011tk}, while for all the chiral-odd TMDs we assume
the DGLAP evolution of transversity \cite{Hirai:1997mm,Boffi:2009sh}.
For the $k_T$-dependencies of the TMDs we use the same
input from TMD parametrizations as described in Eq.~(\ref{Eq:Gauss}).
The predictions from both models evolved in this way
are shown along with the available parametrizations in
Figs.~\ref{Fig:pion-models}-\ref{protonbasis-u+d} at the scale $Q_0$.
It is important to stress that in this way we are able to test the
$x$-dependencies of the model predictions against the COMPASS data.
The ultimate goal would be to test similarly also the quark model
predictions for $k_T$-dependencies. This requires an implementation of
the CSS evolution starting from low initial scales $\mu_0< 1\,{\rm GeV}$
which is beyond the scope of this work, and will be addressed in future studies.
The result from the LFCQM~\cite{Pasquini:2014ppa} and the SPM~\cite{Gamberg:2009uk} for $f_{1,\pi^-}^{\bar u}(x)$ (which coincides with $f_{1,\pi^-}^{d}(x)$ due to isospin symmetry) compare well
to the MRSS parametrization \cite{Martin:2009iq},
see Fig.~\ref{Fig:pion-models}.
In the region $0.2\lesssim x_\pi \lesssim 0.6$,
in which the COMPASS Drell-Yan data points lie,
the SPM result agrees within 20-40$\,\%$ with MRSS \cite{Martin:2009iq}.
The two models agree well with each other
in the case of the pion Boer-Mulders TMD
$h_{1,\pi^-}^{\perp(1) \bar u}(x)=h_{1,\pi^-}^{\perp(1) d}(x)$
for which no extraction is available (so far).
This robustness of the model predictions is important: the pion Boer-Mulders
function enters 4 (out of 6) twist-2 pion-nucleon DY structure functions.
The results from the LFCQM~\cite{Pasquini:2008ax,Boffi:2009sh,Pasquini:2011tk} and the SPM~\cite{Gamberg:2007wm} for the proton quark distributions are
shown in Fig.~\ref{protonbasis-u+d}.
The region $0.05\lesssim x_p\lesssim0.4$
is probed in the COMPASS DY measurements~\cite{Aghasyan:2017jop}, see section~\ref{section:kinematics}. The model results for the functions
$f_{1,p}^u(x)$, $f_{1,p}^d(x)$,
$f_{1T,p}^{(1)u}(x)$, $f_{1T,p}^{(1)d}(x)$,
$h_{1,p}^d(x)$, $h_{1,p}^{\perp(1)d}(x)$,
$h_{1L,p}^{(1)u}(x)$, $h_{1T,p}^{(2)u}(x)$
agree within 20-40$\,\%$, and for
$h_{1,p}^u(x)$,
$h_{1,p}^{\perp(1)u}(x)$,
$h_{1L,p}^{(1)d}(x)$
within 40-60$\,\%$.
Merely for $h_{1T,p}^{(2)d}(x)$ we observe a more sizable
spread of model predictions.
In all cases the models agree on the signs of the TMDs.
The model results for the unpolarized distributions agree
reasonably well with MSTW \cite{Martin:2009iq}.
The model predictions for transversity and Sivers function are compatible
with the corresponding Torino \cite{Anselmino:2011gs,Anselmino:2013vqa}
and JAM20 fits \cite{Cammarota:2020qcw}.
The 1-$\sigma$ uncertainty bands are shown for JAM20 \cite{Cammarota:2020qcw}.
The corresponding uncertainty bands of the Torino parametrizations
\cite{Anselmino:2011gs,Anselmino:2013vqa} are somewhat larger
(as more data were used in the JAM20 analysis, cf.\ section~\ref{sec-2.3})
and not displayed for better visibility. The proton Boer-Mulders
function from models is in good agreement with the BMP10
extraction~\cite{Barone:2009hw} which has significant statistical and
systematic uncertainties, as discussed in section~\ref{sec-2.3}, and
are not shown in Fig.~\ref{protonbasis-u+d}.
The model predictions for pretzelosity show little agreement with the
best fit result from LP15 \cite{Lefky:2014eia}, but are within its
1-$\sigma$ region which is not shown in the plot.
The comparison in Fig.~\ref{protonbasis-u+d} indicates an accuracy
of the CQMs which is in many cases of the order of 20--40$\,\%$.
Considering the much different physical foundations of the two models,
one may speak about an overall robust CQM picture for the TMDs
needed in our work.
\section{Results and observations}
\label{sec-3}
In this section we briefly describe the COMPASS experiment,
outline how we explore the model predictions and phenomenological
TMD fits, present our results, and compare them to the data.
\subsection{The COMPASS Drell-Yan experiment}
\label{section:kinematics}
The COMPASS 2015 data \cite{Aghasyan:2017jop} were taken with
a pion beam of $190 \; {\rm GeV}$ impinging on a transversely
polarized NH$_3$ target with a polarization of $\langle S_T\rangle \approx 73\,\%$
and a dilution factor $\langle f\rangle \approx 0.18$.
The dimuon mass range $4.3 \; {\rm GeV}< Q <8.5 \; {\rm GeV}$ above
charmonium resonance region but below $\Upsilon$ threshold was covered
with the mean value $\langle Q\rangle=5.3\,{\rm GeV}$.
Due to the fixed target kinematics the pion structure was probed at
higher $\langle x_\pi\rangle =0.50$ compared to the proton $\langle x_p\rangle = 0.17$.
The cut $q_T>0.4\,{\rm GeV}$ was imposed and $\langle q_T\rangle = 1.2\,{\rm GeV}$ \cite{Aghasyan:2017jop}. The analysis of the data collected by the
experiment in 2018 in similar conditions is currently under way~\cite{Parsamyan:2019wwd}.
\subsection{The approaches for numerical estimates}
\label{Sec3.2-approaches}
The Sivers asymmetry $A_{UT}^{\sin\phi}$ can be described completely
in terms of both, model predictions and available parametrizations, and is
the only asymmetry where the latter is possible. For the phenomenological
calculation we will use the Torino \cite{Anselmino:2013vqa} and JAM20
\cite{Cammarota:2020qcw} analysis results for $f_{1T.p}^{\perp(1)a}(x)$,
and MSTW \cite{Martin:2009iq} and MRSS \cite{Sutton:1991ay}
parametrizations for proton and pion collinear unpolarized distributions.
The other asymmetries require the knowledge of the pion Boer-Mulders
function for which no parametrization is available. In these cases
we shall adopt two different main approaches, pure and hybrid,
see Fig.~\ref{Fig-02:knowledge-TMDs} for an overview.
We will present therefore up to
four different calculations for each observable by exploring
the model results and available parametrizations discussed in sections~\ref{sec-2.2},~\ref{sec-2.3},~\ref{sec-2.4} and displayed in Figs.~\ref{Fig:pion-models}-\ref{protonbasis-u+d}.
The first approach makes a pure use of model predictions for all
pion and proton TMDs which will be labelled in the plots by the
acronyms LFCQM or SPM.
In the hybrid-approaches we will use the minimal model input,
the predictions from the LFCQM~\cite{Pasquini:2014ppa} and SPM~\cite{Gamberg:2009uk}
for the pion Boer-Mulders
function, and the maximal input from parametrizations:
JAM20 \cite{Cammarota:2020qcw} for $f_{1T,p}^{\perp a}$ and $h_{1,p}^a$,
BMP10 \cite{Barone:2009hw} for $h_{1,p}^{\perp a}$, and
LP15 \cite{Lefky:2014eia} for $h_{1T,p}^{\perp a}$.
The results will be labelled respectively
as ``LFC-JAM20'', ``LFC-LP15'', ``LFC-BMP10''
or ``SPM-JAM20,'' ``SPM-LP15,'' ``SPM-BMP10.''
For $h_{1L,p}^{\perp a}$ we make use of WW-type approximation
which allows one to approximate this TMD in terms of $h_{1,p}^a$ for
which we will use JAM20 \cite{Cammarota:2020qcw}.
WW-type approximations were explored in Ref.~\cite{Bastami:2018xqd}
and shown to work well with the available data. We will add ``WW''
in the label of calculation when WW approximation is used.
For all hybrid calculations we will use the parametrizations
\cite{Martin:2009iq,Sutton:1991ay} for $f_{1,p}^a$ and $f_{1,\pi}^a$.
\begin{figure}[t]
\centering
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{{$A_{UT}^{\sin \phi_S}$}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsinphi_xN.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsinphi_xpi.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsinphi_qT.pdf}} \\
& $x_p$ & $x_\pi$ & $q_T$ \\
\end{tabular}
\vspace{-3mm}
\caption{\label{aut1x-COM}
$A_{UT}^{\sin\phi}$ as a function of
$x_p$ (left), $x_{\pi}$ (middle) and $q_T$ (right)
vs COMPASS data \cite{Aghasyan:2017jop}.}
\vspace{7mm}
\centering
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{{$A_{UT}^{\sin (2\phi-\phi_S)}$}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_m_phiS_xN.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_m_phiS_xpi.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_m_phiS_qT.pdf}} \\
& $x_p$ & $x_\pi$ & $q_T$ \\
\end{tabular}
\vspace{-3.5mm}
\caption{\label{autsin2pmpx-COM}
$A_{UT}^{\sin(2\phi-\phi_S)}$ as a function of
$x_p$ (left), $x_{\pi}$ (middle) and $q_T$ (right)
vs COMPASS data \cite{Aghasyan:2017jop}.}
\vspace{7mm}
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{{$A_{UT}^{\sin (2\phi+\phi_S)}$}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_p_phiS_xN.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_p_phiS_xpi.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUTsin2phi_p_phiS_qT.pdf}} \\
& $x_p$ & $x_\pi$ & $q_T$ \\
\end{tabular}
\vspace{-3.5mm}
\caption{\label{autsin2pppx-COM}
$A_{UT}^{\sin(2\phi+\phi_S)}$ as a function of
$x_p$ (left), $x_{\pi}$ (middle) and $q_T$ (right)
vs COMPASS data \cite{Aghasyan:2017jop}.}
\vspace{7mm}
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{{$A_{UU}^{\cos 2\phi}$}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUUcos2phi_xN.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUUcos2phi_xpi.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AUUcos2phi_qT.pdf}} \\
& $x_p$ & $x_\pi$ & $q_T$ \\
\end{tabular}
\vspace{-3.5mm}
\caption{\label{auucos2px-COM}
$A_{UU}^{\cos 2\phi}$ as a function of
$x_p$ (left), $x_{\pi}$ (middle) and $q_T$ (right)
in the COMPASS kinematics.}
\vspace{7mm}
\begin{tabular}{cccc}
\rotatebox[origin=c]{90}{{$A_{UL}^{\sin 2\phi}$}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AULsin2phi_xN.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AULsin2phi_xpi.pdf}} & \hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=4.5cm]{\FigPath/Evo_AULsin2phi_qT.pdf}} \\
& $x_p$ & $x_\pi$ & $q_T$ \\
\end{tabular}
\vspace{-3.5mm}
\caption{\label{aulsin2px-COM}
$A_{UL}^{\sin 2\phi}$ as a function of
$x_p$ (left), $x_{\pi}$ (middle) and $q_T$ (right)
in the COMPASS kinematics. }
\end{figure}
\subsection{Discussion of the results and comparison to available data}
\label{sec-3.2}
Numerical results for the leading-twist pion-nucleon DY
asymmetries are shown in Figs.~\ref{aut1x-COM}--\ref{aulsin2px-COM}
in comparison to available COMPASS data.
Table~\ref{table-overview} gives a detailed overview
on the model results and phenomenological information.
Let us start the discussion with the Sivers asymmetry. One of the most striking features of
``naively'' T-odd (Sivers, Boer-Mulders) TMDs is the expected sign
change~\cite{Collins:2002kn} from SIDIS to DY due to the
difference of initial (DY) versus final (SIDIS) state interactions
\cite{Brodsky:2002rv,Boer:2002ju}.
Verification of the sign change of the Sivers function is one of the milestones of DY programs of COMPASS and RHIC~\cite{Aschenauer:2015eha}.
In SIDIS the proton $u$-quark Sivers function is negative,
while STAR RHIC \cite{Adamczyk:2015gyk} $W^\pm/Z$
asymmetry data favor a positive sign~\cite{Anselmino:2016uie} hinting on the predicted
process dependence of T-odd TMDs \cite{Collins:2002kn}.
The predictions for the $A_{UT}^{\sin\phi_S}$ asymmetry at COMPASS are positive,
see for instance
Refs.~\cite{Efremov:2004tp,Collins:2005rq,Anselmino:2009st}.
Our calculations confirm this expectation, see
Fig.~\ref{aut1x-COM} where we compare our results
to COMPASS data~\cite{Aghasyan:2017jop}.
The $u$-quark Sivers function in DY is expected to be positive,
see Fig.~\ref{protonbasis-u+d}.
If we disregard sea quark effects, which were shown to play a
negligible role in $\pi^-$-proton DY in the COMPASS kinematics
\cite{Collins:2005rq}, then
$A_{UT}^{\sin\phi_S} \propto f_{1T,p}^{\perp u} (x_p) > 0$.
The experimental error bars are currently sizeable, but the data
show a tendency to positive asymmetry, see Fig.~\ref{aut1x-COM},
in agreement with the expected sign change of the Sivers function.
Clearly, more experimental evidence is needed to corroborate
this finding.
In the global QCD analysis of single-spin asymmetries
\cite{Cammarota:2020qcw} the COMPASS data~\cite{Aghasyan:2017jop}
were used, such that the JAM20 result in Fig.~\ref{aut1x-COM}
is consistent with {\em all} present-day data on observables related
to Sivers functions. It is worth remarking that predictions based
on the earlier Torino extraction \cite{Anselmino:2011gs}
(which used SIDIS data only) yield a somewhat larger
asymmetry than JAM20 and are closer to the LFCQM and SPM
results in Fig.~\ref{aut1x-COM}.
This result is consistent with the different size of Sivers functions found in Ref.~\cite{Anselmino:2011gs} and Ref.~\cite{Cammarota:2020qcw}, see Fig.~\ref{protonbasis-u+d}.
Fig.~\ref{autsin2pmpx-COM} shows the asymmetry
$A_{UT}^{\sin(2\phi-\phi_S)}$ which arises from a convolution
of transversity and pion Boer-Mulders function in comparison
to COMPASS data \cite{Aghasyan:2017jop}.
In the case of this asymmetry the pure model and hybrid
calculations yield results in good mutual agreement.
Neglecting sea quarks, it is
$A_{UT}^{\sin(2\phi-\phi_S)}\propto - h_{1, \pi^-}^{\perp (1)\bar u }(x_\pi) h_{1,p}^u (x_p)<0$.
Both, $h_{1, \pi^-}^{\perp (1)\bar u }$ and $h_{1,p}^u$ are positive, see Fig~\ref{protonbasis-u+d}, and we predict a
negative asymmetry. This is consistent with the trend of the data.
We therefore conclude that the COMPASS data~\cite{Aghasyan:2017jop}
indicate a positive sign for the pion Boer-Mulders
TMD~$h_{1, \pi^-}^{\perp (1)\bar u }$.
(It~is~important to recall that absolute signs in extractions
of chiral-odd TMDs and fragmentation functions are convention-dependent
because chiral-odd functions contribute to observables always in
connection with other chiral-odd functions. The convention used
for TMD extractions is $h_{1,p}^u(x)>0$. This sign is a choice
which is well-informed by model and lattice QCD calculations
but not an experimental observation.)
The indication that $h_{1, \pi^-}^{\perp (1)\bar u }>0$ is
an important result which can be used to test the process dependence
of the proton Boer-Mulders function, see below.
Fig.~\ref{autsin2pppx-COM} shows
$A_{UT}^{\sin(2\phi+\phi_S)}$ which is due to the convolution
of pretzelosity and pion Boer-Mulders function compared to COMPASS
data \cite{Aghasyan:2017jop}. This asymmetry is proportional to
$q_T^3$ for $q_T\ll \,{1\,\rm GeV}$. This leads to a kinematic
suppression of this asymmetry as compared to the two previous
asymmetries (both proportional to $q_T$ at small transverse momenta).
As a consequence $A_{UT}^{\sin(2\phi+\phi_S)}$ is by far the
smallest of the leading-twist asymmetries in pion-nucleon DY.
Numerically it is $1\,\%$ or smaller, such that we had to include the
insets in Fig.~\ref{autsin2pppx-COM} to display the theoretical curves.
The LFCQM and the SPM are in good agreement with each other,
but not with the LP15 fit of pretzelosity \cite{Lefky:2014eia} which
suggests an opposite sign for the asymmetry.
At this point one has
to stress that the LP15 fit of \cite{Lefky:2014eia} has a large
statistical uncertainty (not displayed in Figs.~\ref{protonbasis-u+d}
and \ref{autsin2pppx-COM}) and is compatible with zero or opposite sign
within 1-$\sigma$. This TMD is difficult to measure in DY and SIDIS.
In the high luminosity SIDIS experiments at JLab 12 GeV and the
future Electron Ion Collider it may be feasible to measure pretzelosity.
The $A_{UU}^{\cos 2\phi}$ asymmetry in unpolarized DY originates
from a convolution of the Boer-Mulders functions in nucleon and pion.
Historically it was connected to the ``violation'' of the Lam-Tung
relation, see \cite{Boer:1999mm} and references therein.
A simultaneous measurement of $A_{UU}^{\cos 2\phi}$ and
$A_{UT}^{\sin(2\phi-\phi_S)}$ which we have discussed above
allows one to test the sign change of the proton Boer-Mulders
function in DY. $A_{UU}^{\cos 2\phi}$ was measured and found
positive in earlier CERN and Fermilab measurements
\cite{Guanziroli:1987rp,Conway:1989fs}.
Neglecting sea quark effects, the asymmetry is dominated by
$A_{UU}^{\cos 2\phi}\propto h_{1, \pi^-}^{\perp (1)\bar u }(x_\pi) h_{1,p}^{\perp (1) u} (x_p)$.
With the indication of the positive sign for the pion Boer-Mulders
function from the COMPASS data \cite{Aghasyan:2017jop} on
$A_{UT}^{\sin(2\phi-\phi_S)}$, we conclude a positive sign
also for the proton $u$-quark Boer-Mulders function in DY, which
is opposite to the sign seen in SIDIS analyses \cite{Barone:2010gk}
and hence in agreement with the prediction for the process dependence
property of T-odd TMDs \cite{Collins:2002kn}.
Fig.~\ref{auucos2px-COM} shows our predictions for $A_{UU}^{\cos 2\phi}$
for COMPASS kinematics. At this point no data are available from
COMPASS, but an analysis is planned~\cite{Gautheron:2010wva}
and our predictions in Fig.~\ref{auucos2px-COM} may be tested in near future.
It is worth recalling that our approach provides a good description
of the NA10 CERN \cite{Guanziroli:1987rp} and E615 Fermilab
\cite{Conway:1989fs} data. The test of our predictions
in Fig.~\ref{auucos2px-COM} will help to investigate the compatibility
of the NA10, E615 and COMPASS experiments.
Interestingly, fixed-order collinear factorized perturbative
QCD calculations, which strictly speaking require $q_T$ to be
the hard scale, can also qualitatively describe
the NA10 and E615 data \cite{Lambertsen:2016wgj,Chang:2018pvk}.
It will be interesting to confront those calculations with future
COMPASS data and TMD studies.
Notice that in the analysis \cite{Barone:2010gk}
of the proton-proton and proton-deuteron data from the FNAL
E866/NuSea experiment \cite{Zhu:2006gx,Zhu:2008sj} indications
were obtained that the proton quark and antiquark Boer-Mulders
functions (in DY) have the same signs.
With our observations
based on COMPASS data we therefore infer a first hint that
also the Boer-Mulders functions of $\bar{u}$ and $\bar{d}$
are positive in DY. Interestingly, not only valence
Boer-Mulders distributions in nucleon and pion seem ``alike''
\cite{Burkardt:2007xm}, but also the nucleon sea quark
distributions seem to have all the same sign.
This confirms an early estimate on the sign
of the anti-quark Boer Mulders function carried in the SPM in Ref.~\cite{Gamberg:2005ip}.
This is in line with
predictions from the limit of a large number of colors $N_c$ in QCD
that $h_{1,p}^{\perp u} (x_p,\kTN) = h_{1,p}^{\perp d} (x_p,\kTN)$ and
$h_{1,p}^{\perp\bar u} (x_p,\kTN) = h_{1,p}^{\perp\bar d} (x_p,\kTN)$
modulo $1/N_c$ corrections \cite{Pobylitsa:2003ty}.
Future data will provide more stringent tests of these
predictions.
Finally, it is worth pointing out that in principle one can extract the
$u$-quark transversity distribution entirely from the measurements
of $A_{UU}^{\cos 2\phi}$ and $A_{UT}^{\sin(2\phi-\phi_S)}$
in $\pi^-$-proton DY at COMPASS \cite{Sissakian:2005yp}.
While typically data available from different processes are
processed in ``global analyses,'' whenever possible it is
also valuable to extract a function from one process alone.
This would for instance allow one to test the universality
(same sign and $x$-shape in SIDIS and DY) of the $u$-quark
transversity distribution which is otherwise taken for granted.
Fig.~\ref{aulsin2px-COM} displays our predictions for the
longitudinal single-spin asymmetry $A_{UL}^{\sin 2\phi}$
in the COMPASS kinematics which is due to the
Kotzinian-Mulders TMD $h_{1L}^{\perp a}$ and the pion
Boer-Mulders function. If we disregard sea quark effects, then
$A_{UL}^{\sin 2\phi}\propto -h_{1, \pi^-}^{\perp (1)\bar u }(x_\pi) h_{1L,p}^{\perp (1) u} (x_p) > 0$. Especially the SPM predicts
a sizable and positive asymmetry. Since no parametrization on
$h_{1L}^{\perp a}$ is currently available, the hybrid calculations
make use of the WW-type approximation which is compatible with
SIDIS data \cite{Bastami:2018xqd}.
This is the only leading-twist pion-proton asymmetry in DY
which requires a longitudinal proton polarization.
We are not aware of plans to run DY experiments with longitudinal
proton polarization in the near future. Potentially $A_{UL}^{\sin2\phi}$
could be studied in DY with doubly polarized protons or deuterons in a
future NICA experiment \cite{Savin:2015paa}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Fig.\ & structure function & TMDs & LFCQM & SPM & phenomenology \\
\hline \hline
\ref{aut1x-COM}-\ref{aulsin2px-COM} & $F_{UU}^1$ & $f_{1,p}^{a}$, \ $f_{1,\pi}^a$ & \cite{Pasquini:2008ax}, \cite{Pasquini:2014ppa} & \cite{Gamberg:2007wm} \cite{Gamberg:2009uk} & \cite{Martin:2009iq}, \cite{Sutton:1991ay} \\
\hline
\ref{aut1x-COM} & $F_{UT}^{\sin\phi_S}$ & $f_{1T,p}^{\perp a}$, \ $f_{1,\pi}^a$ & \cite{Pasquini:2010af}, \cite{Pasquini:2014ppa} & \cite{Gamberg:2007wm}, \cite{Gamberg:2009uk} & \cite{Cammarota:2020qcw}, \cite{Sutton:1991ay} \\
\hline
\ref{autsin2pmpx-COM} & $F_{UT}^{\sin(2\phi-\phi_S)}$ & $h_{1,p}^{a}$, \ $h_{1,\pi}^{\perp a}$ & \cite{Pasquini:2008ax}, \cite{Pasquini:2014ppa} & \cite{Gamberg:2007wm}, \cite{Gamberg:2009uk} & \cite{Cammarota:2020qcw}, \ --- \\
\hline
\ref{autsin2pppx-COM} & $F_{UT}^{\sin(2\phi+\phi_S)}$ & $h_{1T,p}^{\perp a}$, \ $h_{1,\pi}^{\perp a}$ & \cite{Pasquini:2008ax}, \cite{Pasquini:2014ppa} & \cite{Jakob:1997wg}, \cite{Gamberg:2009uk} & \cite{Lefky:2014eia}, \ --- \\
\hline
\ref{auucos2px-COM} & $F_{UU}^{\cos 2\phi}$ & $h_{1,p}^{\perp a}$, \ $h_{1,\pi}^{\perp a}$ & \cite{Pasquini:2010af}, \cite{Pasquini:2014ppa} & \cite{Gamberg:2007wm}, \cite{Gamberg:2009uk}
& \cite{Barone:2009hw}, \ --- \\
\hline
\ref{aulsin2px-COM} & $F_{UL}^{\sin 2\phi}$ & $h_{1L,p}^{\perp a}$, \ $h_{1,\pi}^{\perp a}$ & \cite{Pasquini:2008ax}, \cite{Pasquini:2014ppa} & \cite{Gamberg:2007wm}, \cite{Gamberg:2009uk} & \cite{Bastami:2018xqd}, \ --- \\
\hline
\end{tabular}
\caption{Overview on non-perturbative input used to produce
the results in Figs.~\ref{aut1x-COM}--\ref{aulsin2px-COM} which
was taken from the LFCQM, the SPM, and phenomenological fits
(or WW-type approximation in the case of
$h_{1L,p}^{\perp a}$).
Notice that no phenomenological information is currently available
on $h_{1,\pi}^{\perp a}$, cf.\ section~\ref{sec-2.3}.
\label{table-overview}.}
\end{table}
\begin{figure}[t]
\centering
\begin{tabular}{cc}
\rotatebox[origin=c]{90}{{$A_{UU}^{\cos 2\phi}$}} &
\hspace{-2.5mm}\raisebox{-.5\height}{\includegraphics[width=5.5cm]{\FigPath/Evo_AUUcos2phi_xpi_ScaleStudy.pdf}} \\
& $x_\pi$ \\
\end{tabular}
\vspace{-3.5mm}
\caption{\label{auucos2px-COM-study}
$A_{UU}^{\cos 2\phi}$ as a function of
$x_{\pi}$, the orange region corresponds to variation of $C_1$ in the interval $ [e^{-\gamma_E}, 4\, e^{-\gamma_E}]$ and illustrates the sensitivity of our results to scale variations.}
\end{figure}
We also study the theoretical uncertainty due to the variation of $C_1$ and $C_2$ in Eqs.~(\ref{eq:hard}), (\ref{e:FF_ansatz}), and (\ref{e:CS_kernel0}) at NLL accuracy. Such studies are of importance in order to establish the control over the perturbative expansion, see e.g. \cite{Scimemi:2018xaf}. We will use $A_{UU}^{\cos 2\phi}$ asymmetry as an example. Scale dependence on $C_2$ cancels exactly at this order between the numerator and the denominator of the asymmetry. In Fig.~\ref{auucos2px-COM-study} we show the corresponding theoretical uncertainty due to variation of $C_1\in [e^{-\gamma_E}, 4\,e^{-\gamma_E}]$ for the LFCQM, notice that for $C_1= e^{-\gamma_E}$ the asymmetry becomes larger than the red curve calculated with $C_1= 2 e^{-\gamma_E}$,
while for $C_1= 4 e^{-\gamma_E}$ the value of asymmetry decreases very slightly. One can see that the theoretical uncertainty due to the scale choice of $C_1$ is not negligible and warrants the inclusion of higher order corrections in the calculations. This uncertainty is smaller than the spread of the model predictions shown in Fig.~\ref{auucos2px-COM} and therefore we expect that the future data will be able to distinguish among various models.
Before ending this section it is important to remark that the COMPASS
experiment has covered the range $0.4\,{\rm GeV} < q_T < 5\,{\rm GeV}$.
At the upper limit the condition $q_T \ll Q$ for
the applicability of the TMD factorization is not satisfied which
constitutes an uncertainty in our calculations.
However, in the experiment (and in our calculations) it is
$\langle q_T\rangle = 1.2 \,{\rm GeV}$ which is much smaller
than $\langle Q \rangle = 5.3\,{\rm GeV}$ and we verified that
the region of large $q_T$ (namely, $3\,{\rm GeV} < q_T < 5\,{\rm GeV}$)
in our calculations has a negligible impact on
the $q_T$-averaged (integrated) asymmetries in the experiment.
\section{Conclusions}
\label{sec-4}
In this work we studied the DY process with negative pions
and polarized protons with focus on the kinematics of the
COMPASS experiment. As no phenomenological extractions are
available for the Boer-Mulders TMD
function of the pion, we explored two popular and widely used
hadronic models, the LFCQM and the SPM, together with available
phenomenological information on the other TMDs. For the LFCQM and SPM the we implement TMD evolution at NLL accuracy from fixed scale according to the solution to the CSS equations in Ref.~\cite{Collins:2014jpa} and outlined in section~\ref{sec-2.2}. This approach moves beyond the approximate TMD evolution based on the Gaussian Ansatz for transverse parton momenta with
energy dependent Gaussian widths.
We presented a complete description of polarized DY at leading
twist using TMD evolution at NLL accuracy.
The required TMDs include on the nucleon side
$f_{1,p}^{a}$, $f_{1T,p}^{\perp a}$,
$h_{1,p}^{a}$, $h_{1,p}^{\perp a}$,
$h_{1T,p}^{\perp a}$, $h_{1L,p}^{\perp a}$;
and on the pion side
$f_{1,\pi}^a$, $h_{1,\pi}^{\perp a}$.
For that we compiled results from several prior LFCQM and SPM
calculations, which to the best of our knowledge have not been
presented in this completeness before
\cite{Jakob:1997wg,Pasquini:2010af,Pasquini:2008ax,Pasquini:2014ppa,Gamberg:2007wm,Gamberg:2009uk}.
Based on concise comparisons of model results with
available phenomenological information
\cite{Martin:2009iq,Sutton:1991ay,Anselmino:2011gs,Anselmino:2013vqa,Barone:2009hw,Lefky:2014eia,Cammarota:2020qcw,Bastami:2018xqd},
we estimate an accuracy of the model results of
20-40$\,\%$ for the majority of (though not all) TMDs. Similar
``model accuracies'' were found in prior phenomenological
applications of CQMs \cite{Boffi:2009sh,Pasquini:2011tk,Pasquini:2014ppa}.
Driven by the motivation to make maximal use of currently
available phenomenological information
\cite{Martin:2009iq,Sutton:1991ay,Anselmino:2011gs,Anselmino:2013vqa,Barone:2009hw,Lefky:2014eia,Cammarota:2020qcw,Bastami:2018xqd},
we also carried out ``hybrid'' calculations with a minimal model
dependence --- namely only due to the pion Boer-Mulders function
for which no extraction is currently available.
In this way we provided up to four predictions
for each DY observable, with different levels of model dependence.
The critical comparison of the various results
(pure-model and hybrid calculations in respectively LFCQM and SPM)
allows us to differentiate robust predictions from more strongly
model-dependent results.
Our study had two main goals, namely to present theoretical
calculations which help to interpret the first data from the
pion-induced DY with polarized protons measured by COMPASS,
as well as to provide quantitative tests of
the application of
CQMs to the description of pion and nucleon structure.
In regard to the interpretation of the first data from the
pion-induced DY with polarized protons, we observe a robust
picture. The pure-model and hybrid calculations from the
LFCQM and SPM are in remarkable agreement with each other
at the present stage. The theoretical spread of our
results is smaller than the present uncertainties of the
available data. Among the most interesting observations are
the encouraging indications for the change of sign of the
T-odd TMDs in DY vs SIDIS, both in the case of the proton
Sivers and proton Boer-Mulders function. These are model
independent results. Another model-independent result is
the observation that the data favor a positive (in DY)
Boer-Mulders $\bar u$-distribution in $\pi^-$.
We also report the first indication that all proton
Boer-Mulders functions for $u$, $d$, $\bar u$, $\bar d$
flavors are positive (in DY). At the present, these
observations are admittedly vague due to the
low precision of the
current data. More precise
future data from COMPASS and other facilities will allow us
to solidify the picture.
In regard to the quantitative tests of the application
of CQMs, it is important to stress that the DY process
with $\pi^-$ and proton in the COMPASS kinematics is an ideal
process for these purposes. In the COMPASS kinematics sea quarks
do not play an important role \cite{Collins:2005rq}. Due to the
$u$-quark dominance in the proton the process is strongly dominated
by annihilations of $\bar u$ from $\pi^-$ and $u$
from proton in the valence $x$-region
where CQMs can be expected to catch the main features in the hadronic structure of the
pion and nucleon.
CQMs are important qualitative tools for QCD calculations.
Within their model accuracy
and within their range of applicability
in the valence $x$-region,
we observe that CQMs yield useful results and provide
helpful guidelines for the interpretation of data. Future data
will provide more stringent tests of the CQMs, and allow
for extraction of hadron structure
by global QCD analyses.
We also provided several predictions that await experimental
confirmation.
\section*{Acknowledgments}
The authors wish to thank A.~V.~Efremov and A. Kotzinian for valuable discussions
which motivated this study and J.~Collins, T.~Rogers, and Z.~Kang for discussions on implementation of TMD evolution.
This work was supported by the National Science Foundation under the
Contracts No.\
PHY-1812423 (S.B.\ and P.S.) and No.~PHY-2012002 (A.P.),
and in part by the US Department of Energy under contracts, No.~DE-FG02-07ER41460 (L.G.)
and No.~DE-AC05-06OR23177 (A.P.) under which JSA, LLC operates JLab, the framework of the TMD Topical Collaboration (L.G. and A.P.), and by the European Union's Horizon 2020 program under grant agreement No.~824093(STRONG2020) (B.P.).
| 42775dfa720d76848f749a43db26ea451bf06389 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{}
\vspace{-1cm}
\footnotetext{\textit{$^{a}$~Fluid Mechanics Group, Carlos III University of Madrid, Legan\'es, Spain. E-mail: [email protected]}}
\footnotetext{\textit{$^{b}$~Physics of Fluids group. University of Twente, Enschede, The Nederlands. }}
\footnotetext{\dag~Electronic Supplementary Information (ESI) available: [details of any supplementary information available should be included here]. See DOI: 10.1039/b000000x/}
\section{Introduction}
The diffusion-driven dynamics of a dense cloud of bubbles (or drops) have received lately a renewed attention due to their applications in modern chemical processes such as solvent extraction \cite{Peng_etalJCIS2018} and in nanoscience \cite{Zhu_etalSoftMatter2018}. Besides these immediate applications, understanding these diffusive dynamics also sheds light on certain processes taking place in foams, such as coarsening or coalescence. Indeed, although many other physical effects are involved in the evolution of a wet foam, still diffusion-related phenomena play a central role\cite{Langevin2017}. However, despite the diffusive interaction between nearby bubbles in bubbly liquids or foams having been studied for a long time, still there is a need for quantitative predictive models\cite{Michelin_etalPRF2018}. In part, this is a consequence of the difficulties that arise in isolating purely diffusive effects in experiments. In fact, the large difference in density between the bubbles and the liquid makes gravity effects manifest rather early in the evolution of foams or bubble clouds\cite{Langevin2017}.
In foams it is hard to disentangle the role of liquid drainage (a gravity-related effect) from other effects associated to the transport of gas\cite{garcia_moreno_etalSofMatter2011} or anti-foaming agents\cite{yazhgur2015antifoams} in the rate of bubble coalescence. Moreover, drainage turns wet foams into dry ones very fast, so the former are hard to characterize experimentally\cite{Langevin2017}. Even in a situation where diffusion is the leading effect, such as the growth of a bubble cloud in a gas-supersaturated liquid, gravity effects become of comparable importance very quickly. For example, a free cloud of sub-millimetric bubbles exhibits a purely diffusion-driven growth only for a very limited time (hundreds of miliseconds), even for high gas concentrations such as those found for instance in beer or champagne, as shown by \citet{rodriguez2014}. In this study, a dense bubble cloud was produced in the bulk of a CO$_2$-saturated liquid (beer) and its evolution studied using high-speed imaging. This evolution could be divided into three stages: in a first one, which lasts only about 10 ms, the size of the bubbles increases with the square root of time, as predicted by the Epstein-Plesset equation for the case of a single isolated bubble\cite{EpsteinPlessetJCP1950}. Thus, it is reasonable to assume that in this stage bubbles grow without interacting much with each other. Then, another stage follows where bubbles moderate their growth, due to the depletion of the dissolved gas in the space between bubbles. These two stages, where the dynamics are driven by diffusive bubble growth, come to an end when gravity effects become important and make the bubble cloud rise, which occurs as early as about 100 ms even for a bubble cloud as small as about 1 mm in diameter. Of course, bubbles could be prevented from rising by keeping them pinned to a substrate. However, the pinning introduces additional complexities that obscure, to some extent, diffusive effects. Moreover, in some situations, gravity plays a role even if the bubbles remain pinned to a substrate. In the case of sessile bubbles in CO$_2$-supersaturated water, \citet{Enriquez_etalJFM2014} and \citet{Moreno_soto_etalJFM2019} have shown that, although bubbles are not allowed to move, the depletion of CO$_2$ that takes place around them induces density gradients in the liquid that ultimately trigger a convection plume, which in turn dominates the flow and overcomes the effect of diffusion.
Therefore, avoiding or minimizing gravity effects is essential to study the purely diffusion-driven dynamics of wet foams or bubble clouds. Moreover, besides this purely fundamental interest, there are situations in which bubble clouds or wet foams evolve in microgravity or, at least, in reduced gravity conditions. For example, in the field of planetary physics, the diffusive growth of bubbles is responsible for the loss of Helium in meteorites
\cite{stuart1999}. Thus, modeling the evolution of Helium content in these objects is important to understand the history of the formation of our solar system. In the Moon, rocks with a highly vesicular structure, such as the "Seatbelt Rock", have been found (see figure 4 in \citet{JolliffRobinsonPT2019}). This structure suggests that the rock has formed by solidification of a dense cloud of bubbles containing volatile gases, as a result of ancient volcanism. From a more technological point of view, the behavior of foams in reduced gravity conditions is of interest for the development of advanced manufacturing processes in space, as pointed out by \citet{Koursari_etalMGST2019}. These authors study experimentally the drying of a wet foam in the absence of gravity. Since on Earth gravity is essential to drain out the water filling the interstices between bubbles, in space they use a porous medium to soak up the liquid. Note that, if bubbles were allowed to grow by supersaturating the liquid with a gas, their growth could push out the interstitial water without the need of an external absorbing body.
In summary, studying experimentally the diffusive interaction of free growing bubbles in dense bubble clouds or in wet foams at long times requires avoiding the effect of gravity. This has been done in several studies in the case of foams\cite{garcia_moreno_etalSofMatter2011, garcia_moreno_etalSoftMatter2014,garcia_moreno_NatComm2019, Koursari_etalMGST2019} but, to the best of our knowledge, there are no studies focusing on dense bubble clouds growing diffusively in gas-supersaturated liquids in microgravity. With this motivation in mind, we have carried out experiments in which dense bubble clouds grow in CO$_2$-supersaturated water in microgravity conditions. These experiments were performed in the drop tower of the German Center of Applied Space Technology and Microgravity (ZARM) using an improved version of the facility described by \citet{VegaMartinez_etalMGST2017}. Besides, inspired by these experiments, we have put together a mathematical model of the growth of a bubble cloud in a supersaturated liquid that is able to explain, albeit in a qualitative fashion, the features of the time evolution of bubble clouds observed in the experiments.
\section{\label{sec:experiments}Experimental setup and procedure}
The experiment aims at generating a dense bubble cloud in a supersaturated liquid, in our case CO$_2$-saturated water, and then observe its evolution using high-speed imaging in microgravity conditions. Figure~\ref{fig:setup} shows the sketch of the experimental setup, which is based on that described in \citet{VegaMartinez_etalMGST2017}. This experiment is fitted into a capsule that is then dropped inside the drop tower of the German Center of Applied Space Technology and Microgravity (ZARM). This tower has total height of 146 meters and contains in its interior a 120-m-tube where the capsule falls in near vacuum for about 4 s. This generates a microgravity environment for the duration of the drop. Then the capsule decelerates and comes to a stop in a pool filled with polystyrene pellets.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figures/figure1.pdf}
\caption{Layout of the experimental setup and the drop tower facility. The two high-speed cameras form an angle of 90$^{\,\rm o}$. The chamber has a cylindrical shape, 24.4 mm of diameter and 200 mm of height. The bubble cloud is generated at the center of the chamber to avoid the effect of walls. The right panel is a sketch of the capsule inside the drop tower (not to scale)}
\label{fig:setup}
\end{figure}
The measurement chamber, where the bubble cloud will form and grow, is filled with carbonated water. This chamber is immersed in a rectangular tank filled with degassed water to minimize optical distortion. The top wall of the chamber is connected to a pressurization and depressurization system.
At the bottom of the measurement chamber there are two electrodes connected to the spark generator. The electrodes are two thin (100 $\mu$m) copper wires that touch each other near their tips. The spark generator is a discharge circuit similar to that described in \citet{Willert2010} but replacing the LEDs by the electrodes, as suggested by \citet{Goh2013}. In our experiments, the discharge has a peak between 50-100 A and lasts for about 10 $\mu$s. The discharge induces cavitation, and the collapse of the imploding cavitation bubble generates the bubble cloud, the growth of which is the target of the experiment. To measure the time evolution of the shape and size of the cloud, two high-speed cameras, suitable for use in the drop tower's capsule, image the bubble cloud from two perpendicular directions at 2000 fps. They produce images of 512$\times$512 pixels and a spatial resolution of 25 $\mu$m/pixel. As can be seen in the appendix, the information provided by the two cameras is qualitatively similar for all drops. For that reason, in the discussion of the experimental results we will only use images taken from camera 1.
The supersaturated water is prepared in a pressurized facility built to make carbonated water on-the-site which is subsequently transferred to the measurement chamber at a pressure higher than the saturation one, such as no gas is lost during the process. This procedure is similar to that followed by \citet{Enriquez2013}.
Figure~\ref{fig:panel} compares the evolution of a bubble cloud in microgravity ($10^{-6}$ $g$) and 1 $g$ conditions from its generation (the spark) to the end of the experiment, approximately 3.4 seconds later. During the first 100 ms, both tests show a similar diffusion-driven growth. Subsequently, in the 1 $g$ test, the effects of the gravity become abundantly clear: the cloud starts a buoyancy-induced rising motion as can be seen in the first row of the figure. In the microgravity test, however the bubble cloud continues growing by diffusion and no net motion is observed. For the model developed in the next section it is important to point out that, despite the high bubble density of the cloud, we observe that very few bubbles coalesce during the experiments. Thus, in a first approximation the number of bubbles may be regarded as constant.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/figure2.pdf}
\caption{Comparison of experiments where the bubble cloud develops with (upper row) and without gravity (lower row). See movie in supplemental material.}
\label{fig:panel}
\end{figure*}
\section{Mathematical model of transient mass transfer in a bubble cloud}
Let us consider a spherical bubble of radius $R_i$ surrounded by a cloud of other bubbles of radii $R_j$, with $j = 1..N$ and $j \ne i$. We introduce here three hypotheses. First, from the point of view of mass transfer, one bubble sees the others in the cloud as mass sinks of intensity $\dot{m}_j$. Second, we do not consider surface tension effects. Third, advective effects caused by bubble growth are neglected in the gas transport equation. This means that the concentration of dissolved gas obeys the non-convective heat equation. Due to the linear nature of this equation, the interaction of the $i$-th bubble with the surrounding ones can be treated by superposition. For this reason, we consider the growth of a bubble, labelled $i$, in presence of a point mass sink of intensity $\dot{m}_j$ a distance $d_{ij}$ apart from its center (see figure \ref{fig:bubble_and_sink}). The problem we will solve now is to determine the gas mass flux entering bubble $i$ in these conditions. This problem exhibits cylindrical symmetry around the line connecting the centers of both bubbles, so the concentration field obeys
\begin{equation}
\partial_t C - D\left[\frac{1}{r^2}\partial^2_{rr}\left(r^2 \, C\right) + \frac{1}{r^2\sin\theta}\partial_\theta\left(\sin\theta\partial_\theta C\right)\right] = -\dot{m}_j\delta\left[\vec{x}-d_{ij}\vec{e}_x\right].
\label{eq:time_eqnC}
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\columnwidth]{figures/figure3.pdf}
\caption{\label{fig:bubble_and_sink}Bubble of radius $R_i$ in the presence of a mass sink of intensity $\dot{m}_j$ located a distance $d_{ij}$ apart.}
\end{figure}
This equation must be subject to the boundary conditions at the bubble surface, $r=R_i$, and in the far field $r \rightarrow \infty$. At the bubble surface the concentration is given by Henry's law, in other words $C(R, t) = C_s = k_H P_s$, where $k_H$ is Henry's constant and $P_s$ the ambient pressure, whereas in the far field $C(\vec{x}, t) = C_\infty$. Note that, although equation (\ref{eq:time_eqnC}) is linear, the fact that the radius of the $i$-th bubble, $R_i$, depends on time effectively couples the computation of the mass fluxes of the different bubbles, as every bubble $j$ affects the concentration field induced by the others through the boundary condition at $r = R_i$. Nonetheless, as will be explained below, this effect can be neglected with the assumptions adopted here.
The solution of the problem will be sought by exploiting its linear nature to decompose it into a spherically-symmetric one that satisfies these boundary conditions plus another one that accounts for the effect of the point sink but that has homogeneous boundary conditions. The former solution recovers the mass flux given by the Epstein--Plesset equation\cite{EpsteinPlessetJCP1950}, $\dot{m}_{i0}$. To calculate the latter, denoted by $\dot{m}_{ij}$, we transform equation (\ref{eq:time_eqnC}) into the Laplace plane. In what folows, $\hat{C}(r, s)$ and $\hat{m}_{j}(s)$ are the Laplace transforms of $C(r, t)$ and $\dot{m}_{j}(t)$, respectively. With this notation, equation (\ref{eq:time_eqnC}) turns into
\begin{equation}
s\,\hat{C} - D\left[\frac{1}{r^2}\partial^2_{rr}\left(r^2 \, \hat{C}\right) + \frac{1}{r^2\sin\theta}\partial_\theta\left(\sin\theta\partial_\theta \hat{C}\right)\right] = -\hat{m}_{j}\delta\left[\vec{x}-d_{ij}\vec{e}_x\right].
\label{eq:laplace_eqnC}
\end{equation}
To transform the time derivative of the concentration field we take into account that, with the decomposition explained above, the concentration field induced by an individual mass point sink has homogeneous initial conditions, thus $\partial C/\partial t$ transforms to $s\,\hat{C}$.
The general solution to this equation is\cite{WuSF2015}:
\begin{equation}
\begin{split}
\hat{C} & = -\frac{\hat{m}_{j}}{4\pi D}\frac{\exp\left(-\sqrt{s/D}\sqrt{d_{ij}^2 + r^2 - 2d_{ij}r\cos\theta}\right)}{\sqrt{d_{ij}^2 + r^2 - 2d_{ij}r\cos\theta}} +\\
& + \sum_{l=0}^\infty\left[a_l\,h_l^{(1)}(\sqrt{-s/D}\,r) + b_l\,h_l^{(2)}(\sqrt{-s/D}\,r)\right]\,P_l(\cos\theta),
\end{split}
\label{eq:hC_general_solution}
\end{equation}
where $h_l^{(1)}$ and $h_l^{(2)}$ are Hankel's functions of first and second kind, respectively. Note that, since they contain exponential functions of imaginary exponent, the minus sign inside the square root of their arguments guarantees that they take real values. Although to compute the concentration field it is necessary to determine the full set of coefficients $(a_l, b_l)$, the calculation of the mass flux $\hat{m}_{ij}$ that sink $j$ induces on the bubble only requires knowledge of the coefficients $a_0$ and $b_0$. Indeed, the mass transfer due to higher-order harmonics is zero, as
\begin{equation}
\dot{m}_{ij} = -2\pi R_i^2 D \, \int_0^\pi \left.\frac{\partial C}{\partial r}\right|_{r=R}\,\sin\theta\,\mathrm{d}\theta \sim \int_0^\pi P_l(\cos\theta)\,\sin\theta\,\mathrm{d}\theta,
\label{eq:mass_flux}
\end{equation}
with the last expression being zero for $l \ge 1$. Moreover, the coefficient $b_0$ must vanish, as the function $h_0^{(2)}\left(\sqrt{-s/D}r\right)$ is unbounded as $r \rightarrow \infty$.
To obtain $a_0$ we introduce (\ref{eq:hC_general_solution}) into (\ref{eq:laplace_eqnC}) and project onto $P_0(\cos\theta)$ to find
\begin{equation}
\begin{split}
0 = & -\frac{\hat{m}_j}{4\pi D}\,\int_{-1}^1{ \frac{\exp\left(-\sqrt{\displaystyle \frac{s}{D}}\sqrt{d_{ij}^2 + R_i^2 - 2d_{ij}R_i x}\right)}{\sqrt{d_{ij}^2 + R_i^2 - 2d_{ij}R_i x}}\,\mathrm{d}x} + \\
& + 2 a_0 h_0^{(1)}\left(\sqrt{-s/D}\,R_i\right),
\end{split}
\end{equation}
where the variable change $x = \cos\theta$ has been introduced. Here it is good to note that the expression of Eq. (5) is hybrid, as it mixes expressions in Laplace space with the radius of bubble $i$, which generally is a function of time, $R_i(t)$. We will however ignore this fact and treat $R_i$ as a constant in Laplace space to make analytical progress. Physically, this amounts to saying that we assume that the time scale with which the concentration field responds to changes in the configuration is much smaller than that at which the bubble grows. This is an assumption that is quite similar to the ``frozen bubble'' assumption that was employed in solving the single bubble problem in the original paper by Epstein \& Plesset \cite{EpsteinPlessetJCP1950}.
Evaluating the integrals, and taking into account $h_0^{(1)}(\sqrt{-s/D}\,R_i) = -e^{-\sqrt{s/D}\,R_i}/\sqrt{s/D}\,R_i$,
\begin{equation}
0 = \frac{\hat{m}_j}{4\pi D}\frac{2e^{-\sqrt{s/D}\,d}}{R_i d_{ij}\sqrt{s/D}}\,\sinh\left(\sqrt{s/D}\,R_i\right) + 2 a_0 \frac{e^{-\sqrt{s/D}\,R_i}}{\sqrt{s/D}\,R_i},
\end{equation}
thus
\begin{equation}
a_0 = -\frac{\hat{m}_j}{4\pi D d_{ij}}\,e^{-\sqrt{s/D}\left(d_{ij}-R_i\right)}\,\sinh\left(\sqrt{s/D}\,R_i\right).
\end{equation}
Introducing this value of $a_0$ together with $b_0 = 0$ into (\ref{eq:hC_general_solution}) and differentiating we get
\begin{equation}
\int_0^\pi \left.\frac{\partial \hat{C}}{\partial r}\right|_{r=R_i}\,\sin\theta\,\mathrm{d}\theta = -\frac{\hat{m}_j}{2\pi D d_{ij} R_i} \exp\left(-\sqrt{\frac{s}{D}}\left(d_{ij}-R_i\right)\right).
\end{equation}
Transforming this expression back into the time domain and introducing it in the definition of the mass flux (\ref{eq:mass_flux}),
\begin{equation}
\dot{m}_{ij} = -\frac{R_i^2 D}{2\sqrt{\pi}}\int_0^t \frac{1}{R_i}\left(1-\frac{R_i}{d_{ij}}\right)\frac{\dot{m}_j(t')}{\left[D(t-t')\right]^{3/2}}\,\exp\left(-\frac{(d_{ij}-R_i)^2}{4D(t-t')}\right)\,\mathrm{d}t'.
\label{eq:mass_flux_time}
\end{equation}
To obtain the total mass flux into bubble $i$, this expression must be summed over all bubbles in the cloud (with the exception of $i$ itself) and then added to the mass flux provided by the Epstein--Plesset equation, $\dot{m}_{i0}$, to yield
\begin{equation}
\begin{split}
\dot{m}_i & = 4\pi R_i^2 D \left(C_\infty - C_s\right)\left(\frac{1}{R_i} + \frac{1}{\sqrt{\pi D t}}\right) -\\ &\frac{R_i^2 D}{2\sqrt{\pi}}\sum_{\substack{j=1 \\ j\neq i}}^N\int_0^t\frac{1}{R_i}\left(1-\frac{R_i}{d_{ij}}\right)\frac{\dot{m}_j(t')}{\left[D(t-t')\right]^{3/2}}\,\exp\left(-\frac{(d_{ij}-R_i)^2}{4D(t-t')}\right)\,\mathrm{d}t'.
\end{split}
\label{eq:eqn_mdot_dim}
\end{equation}
To make the above expressions dimensionless we introduce the following variables (suppressing indices), $a = R/R_c$, $\Delta = d/R_c$, $\tau = D t/ R_c^2$ and $\dot{M} = \mathrm{d}M/\mathrm{d}\tau = \dot{m}/(\rho_g D R_c)$, where $R_c$ is a fixed characteristic bubble diameter, and $\rho_g$ is the gas density at the ambient pressure $P_s$.
Besides, the saturation level, which indicates the amount of CO$_2$ dissolved in the liquid, is defined by $\zeta = C_\infty / C_s$. Let us define also the Ostwald coefficient as $\Lambda = C_s/\rho_g$. With these definitions, equation (\ref{eq:eqn_mdot_dim}) results in
\begin{equation}
\begin{split}
\dot{M}_i & = 4\pi a_i^2 \Lambda\left(\zeta-1\right)\left(\frac{1}{a_i} + \frac{1}{\sqrt{\pi\tau}}\right) -\\ &\frac{a_i^2}{2\sqrt{\pi}}\sum_{\substack{j=1 \\ j\neq i}}^N\int_0^\tau \frac{1}{a_i}\left(1-\frac{a_i}{\Delta_{ij}}\right) \frac{\dot{M}_j(\tau')}{\left[\tau-\tau'\right]^{3/2}}\,\exp\left(-\frac{(\Delta_{ij}-a_i)^2}{4(\tau-\tau')}\right)\,\mathrm{d}\tau'.
\end{split}
\label{eq:eqn_mdot}
\end{equation}
This mass flux must be completed with the mass conservation equation for the $i-$th bubble, namely
\begin{equation}
\dot{M}_i = 4\pi a_i^2 \frac{\mathrm{d}a_i}{\mathrm{d}\tau}.
\label{eq:eqn_massconservation}
\end{equation}
The changes in the volume of the bubble induce a velocity field which, sufficiently far away from its center, can be modelled as that of a volume point source. Thus, considering the whole cloud, each individual bubble moves as a result of the superposition of these flows. In the Stokes limit, the resulting bubble velocity can be computed using the solution provided by \citet{Michelin_etalPRF2018}
\begin{equation}
{\bar{U}_i} = \frac{1}{4\pi} \sum_{\substack{j\neq i}}^N \frac{\mathrm{\bar{e}_{ij}}}{\Delta_{ij}^2} \left( \dot{M}_j -\sum_{\substack{j\neq l}}^N \frac{a_j^3 \dot{M}_l}{\Delta_{jl}^3}\left(1-3(\mathrm{\bar{e}_{ji}} \cdot \mathrm{\bar{e}_{jl}})^2\right) \right) .
\label{eq:velocities}
\end{equation}
In this equation, $\bar{e}_{ij}$ is the unit vector pointing from the center of the $i$-th to that of the $j$-th bubble. In this work, we are going to let bubbles translate with only the first term of this expression since we expect $\left(a_{j}/\Delta_{jl}\right)^3 \ll 1$.
\subsection{Numerical method}
The computation of the time evolution of the bubble cloud is divided into two stages. First, we integrate equations (\ref{eq:eqn_mdot}) and (\ref{eq:eqn_massconservation}) together assuming that the bubbles do not displace and then, in a second stage, we compute the motion of the bubbles using the velocity field given by (\ref{eq:velocities}).
The system of integro-differential equations (\ref{eq:eqn_mdot}-\ref{eq:eqn_massconservation}) is solved using an explicit Euler method for the temporal derivatives and a trapezoid method for the integral. Note that the kernel of the integral equation tends smoothly to zero at the integration limit $\tau' = \tau$, which makes the usage of more sophisticated integration methods unnecessary. We can distinguish two terms in equation (\ref{eq:eqn_mdot}): the first one is the mass flux provided by the Epstein-Plesset equation\cite{EpsteinPlessetJCP1950}, which corresponds to the evolution of an isolated bubble. The second term, the integral, models the interaction between each individual bubble and the rest of the cloud. For the calculation of this second term we will adopt the simplifying assumption that the radius of the bubble corresponds to the initial one (frozen bubble). Note that this hypothesis is also adopted in the derivation of the Epstein--Plesset equation\cite{EpsteinPlessetJCP1950}. The time step chosen for the simulations presented in this paper is $\Delta\tau = 10^{-3}$, which has been selected after verifying that taking a smaller time step does not affect the results.
Once the time evolution of the bubble radii is computed, the bubbles are displaced using the velocity field given by (\ref{eq:velocities}) using also, for consistency with the previous stage, an Euler method with the same time step.
The start-up of the numerical computation is done by integrating solely the Epstein--Plesset equation up to the time step used for the rest of the calculation, $\Delta\tau$. Note that this is possible, since at very short times the concentration boundary layer is confined to a thin shell around the bubble of thickness $\sim \sqrt{\tau}$, which is much smaller than the typical bubble-bubble distance. This makes it possible to neglect the integral term in this first step. This strategy is also used in the computation of the bubble velocities.
\section{Results and discussion}\label{sec:results_discussion}
We split this section into two parts. First, we explore the predictions of the theoretical model just described for conditions representative of our experiments. Then, we compare these predictions, in a qualitative way, with those features of the evolution of the bubble clouds that could be determined from the experiments.
\subsection{Results of the model}
In this section we show the predictions of the model for bubble clouds representative of those observed in the experiments. Since the number of dimensionless parameters of the problem is large, we restrict our results to initially spherical bubble clouds with a number of bubbles large enough to exhibit collective effects, namely $N=100$, and evolving in liquids with different saturation levels, $\zeta$. Furthermore, all the bubbles have the same initial size, $R_0 = 30$ ${\mu}$m. This radius is consistent with the estimated size of the bubble fragments that result in a similar cavitation experiment\cite{rodriguez2014}.
Note that for bubbles of this size it is reasonable to neglect the effect of surface tension in their dynamics, as the Laplace pressure, $2\sigma/R_0$, with $\sigma \approx 0.7$ N/m the surface tension coefficient, only becomes of the order of the ambient pressure for bubbles of size $2\sigma/P_s \approx 1.4$ $\mu$m. In all the calculations we have adopted $\Lambda = 0.842$ and $D = 1.92\times 10^{-5}$ m$^2$/s, corresponding to the values of the system CO$_2$-water at 25$^\mathrm{o}$C\cite{PenasLopez_etalJFM2017}.
The initial positions of the bubbles are chosen randomly within a sphere, the cloud, of radius $R_{bc}$ such that the average bubble density is approximately uniform. To clearly identify the moment when bubbles start to interact with each other, it is convenient to seed the bubbles in the cloud with an average inter-bubble nearest neighbor distance, $d$, larger than, but still of the order of, one bubble diameter. Assuming that the volume of cloud available per bubble is $V_{\rm b} \sim \pi d^3 / 6$, setting $R_{bc} \approx 10 R_0$ yields $d \approx 4.3 R_0$, which fulfills the above criterion. In what follows, all length scales are made dimensionless with $R_0$ and time scales with $R_0^2/D$, using the same notation as in the previous section.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figures/figure4.pdf}
\caption{\label{fig:Initial_final_cloud}. Three-dimensional view of a bubble cloud generated with the model described in section 3, ($a$) at time $\tau=0$ and ($b$) at a later time, $\tau_{end}=3.2$, in dimensionless units. Panel ($c$) represents half the cloud, consisting of only those bubbles with centers obeying $x<0$, at the same time. The saturation of the liquid is $\zeta = 2$. The color represents the initial distance of each bubble to the center of cloud, growing from lighter to darker colors.}
\end{figure}
Figure \ref{fig:Initial_final_cloud} shows qualitatively the evolution of a modeled bubble cloud with a saturation $\zeta = 2$. The top left side is the initial state, $\tau=0$, whereas the top right one corresponds to the cloud at a later time $\tau = 3.2$. The color indicates the initial distance from the bubble to the center of the cloud, growing from light to dark hue. To illustrate this color coding, the bottom image represents a cut of the cloud at $\tau = 3.2$. To average results for bubbles located at a given initial distance from the center, the cloud radius in the initial state has been divided into $B=10$ layers equally spaced along the radius $R_{bc}$, with the color indicating the layer number. Using this averaging strategy we can show how bubbles grow at different depths inside the cloud.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{figures/figure5.pdf}
\caption{($a, c, e$) Dimensionless time evolution of the bubble radii for saturation levels $\zeta = 1.1, 2$ and $4$ respectively. The solid-line shows the mean radius for each cloud layer, with the color evolving from lighter to darker according to the distance of the layer to the cloud center. The error bars mark the maximum and minimum radius of the bubbles for each layer at that time instant. The red-dashed line represents the growth of an isolated bubble predicted by Epstein-Plesset equation \cite{EpsteinPlessetJCP1950}. ($b, d, f$) Final layer-averaged radius of the bubbles in the cloud as a functions of its distance to the center. Markers represent the mean final radius, whereas the vertical bars show the maximum and minimum value of the radius. The horizontal error bars correspond to the maximum and minimum final distance to the cloud center of the bubbles contained in each layer. The black horizontal dashed line is the value of $a_f$ predicted by equation (\ref{eq:af_estimation}).}
\label{fig:radii_table}
\end{figure}
Panels ($a$, $c$, $e$) of figure~\ref{fig:radii_table} show the time evolution of the radii of the bubbles in three clouds with saturation levels $\zeta = 1.1, 2$ and 4, averaged over each radial layer. Correspondingly, panels ($b$, $d$, $f$) show the average bubble radius of each layer at the end of the simulation ($\tau \approx 6.5$). The duration of the simulation has been chosen to allow for the observation of the different growth regimes. At early stages all the bubbles grow freely, that is, as isolated bubbles. In this situation, their growth is well described by the Epstein--Plesset equation \citep{EpsteinPlessetJCP1950}, corresponding to the red dashed line in the figure. Using our notation, this equation reads
\begin{equation}
\frac{\mathrm{d}a}{\mathrm{d}\tau} = \Lambda\left(\zeta-1\right)\left(\frac{1}{a} + \frac{1}{\sqrt{\pi\tau}}\right).
\label{eq:epstein_plesset_dimensionless}
\end{equation}
The departure from the growth predicted by this equation is a consequence of the fact that the thickness of the diffusive boundary layer around the bubble grows as $\delta \sim \sqrt{\tau}$. Therefore, until times of order $\tau \sim \left(\Delta/2 - 1\right)^2$ the boundary layers of the different bubbles do not interact and bubbles do not feel each other. However, from that time onwards, the diffusive boundary layers overlap and bubbles compete for the available CO$_2$. As a consequence, their growth rates diminish quickly, even becoming zero or slightly negative for bubbles in the innermost layers, where the access to the CO$_2$ present in the bulk outside the cloud is limited. Contrarily, the growth of the bubbles in the husk of the cloud continues, albeit at a much smaller pace. Note that the freezing of the growth of the innermost bubbles, which is also clearly visible in figure~\ref{fig:Initial_final_cloud}c, is consistent with the experimental observations (see figure \ref{fig:minibubbles_exp} and the associated discussion). Regarding the slight reduction of the bubble sizes for bubbles deep inside the cloud for $\tau \gtrsim 2$, observed in figure~\ref{fig:radii_table}, although it seems reminiscent of Ostwald-rippening, where surface tension causes large bubbles to grow at the expense of small ones\citep{Schmelzer1987}, surface tension is not considered here. Consequently, this bubble size reduction cannot be attributed to this effect. In fact, in the absence of surface tension, the gas concentration at the bubble surface could never be smaller than the saturation one, $C_s$, which means that bubbles should never shrink. The cause of this anomalous effect is the fact that we model nearby bubbles as point sinks. For a point sink, imposing a given mass flux means that the concentration is minus infinity at the sink. So, if the sink is very close to the bubble, it may occur that locally the concentration becomes smaller than the saturation one, which is a non-physical effect. Naturally, this would not occur were the sink replaced by a physical bubble, as in its surroundings the concentration verifies $C \ge C_s$ at all times. Nonetheless, the bubble size reduction created by this simplification is at most about 15\% in the case $\zeta = 4$ and smaller in the others.
The final radius of bubbles belonging to the innermost layers can be estimated using simple mass conservation arguments. Since these bubbles lose access to the CO$_2$ from the bulk liquid, they almost stop growing when they have absorbed all the excess CO$_2$ dissolved in the liquid volume of cloud available per bubble, which we can regard as that of a sphere of diameter $d$. Thus, the final dimensional mass, $4\pi/3 \, R_f^3 \rho_g$, must the the sum of the initial one, $4\pi/3 \, R_0^3 \rho_g$, plus the excess mass of CO$_2$ dissolved in the liquid volume with respect to the saturation one, namely $(C_\infty - C_s)\pi/6 \, d^3$. In dimensionless terms this yields for the final radius
\begin{equation}
a_f = \left(1 + \Lambda\left(\zeta-1\right)\Delta^3/8\right)^{1/3}.
\label{eq:af_estimation}
\end{equation}
For the saturation levels corresponding to the simulations in figure~\ref{fig:radii_table}, namely $\zeta = 1.1, 2$ and 4, the values of $a_f = 1.21, 2.22$ and 3.14 respectively. We must point out here that when we refer to final radius or bubble sizes, we mean at the end of the simulation. Note that, even at the center of the bubble cloud, still bubble radii keep changing with time at long times. However, they do so at a much smaller pace than that predicted by the Epstein--Plesset equation, thus we find it reasonable to regard it as quasi-frozen.
As can be observed in figures~\ref{fig:radii_table}$b$, $d$ and $f$, the agreement with the numerical results is fairly reasonable, given the assumptions adopted in this estimation. In this figure, markers denote the layer-averaged final radius, with the vertical error bars indicating the minimum and maximum bubble size for each layer. In an analogous way, the horizontal position of each marker represent the mean radial position of the bubbles in the layer at the end of the simulation, which are all contained in the segment marked by the horizontal error bars.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{figures/figure6.pdf}
\caption{Rate of change of the radius of the bubbles in the outermost layer divided by that predicted by the Epstein--Plesset equation, as a function of the inverse of the number of bubbles in this layer. The solid line is $1/N_{os}$.}
\label{fig:Nos_Rep}
\end{figure}
Once bubbles deep into the cloud reach this maximum size, it seems reasonable to assume that only those in the outer shell continue growing. Because there is almost no CO$_2$ inside the cloud, all the mass flux comes across the outer boundary of the cloud. Moreover, this mass flux must be shared between all the bubbles in the outer shell. In a first approximation, we can model this effect by assuming that each bubble does no longer intake gas from a solid angle $4\pi$ (that is, from all directions) but from a solid angle $4\pi / N_{os}$, where $N_{os}$ is the number of bubbles in the outer shell. This means that the Epstein--Plesset equation can still be used to model bubble growth in the husk, but dividing its right hand side by the number of bubbles in the outer shell. To check this hypothesis, we plot in figure~\ref{fig:Nos_Rep} the time derivative of the radius of the bubbles in the outermost layer at the end of the simulation divided by the one predicted by the Epstein--Plesset equation as a function of the inverse of the number of bubbles in this layer. These results correspond to the same saturation level, $\zeta = 2$, but to four different realizations. Because in each realization the bubbles are distributed randomly, the number of them in the outermost layer varies. The average value of this number can be estimated by assuming that the bubble density in the cloud is constant and that the outermost layer has a thickness of about $d/2$. Using simple geometrical analysis this yields $N_{os} \approx 3Nd/2R_{bc}$.
We conclude this section with a comment about the effect of the saturation level on the growth rate of the cloud. Although in the experiments the radius of the individual bubbles cannot be determined experimentally, it is possible to measure the time evolution of the projected area of the whole cloud by applying image processing to each frame of the high-speed movies. To compare these results against the simulations, we need to compute the time evolution of the projected area of the simulated cloud. To that end, we create synthetic images projecting the three-dimensional bubble cloud onto a plane, as sketched in figure~\ref{fig:sketch_projected_area} and done in figure~\ref{fig:Initial_final_cloud}. Using this technique, we are able to take into account the effect of bubbles overlapping on the projected cloud area. Note that, since the cloud is spherical and the bubbles are distributed randomly, the choice of the projection plane does not have an effect on the results.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{figures/figure7.pdf}
\caption{Schematic of the definition of the projected area.}
\label{fig:sketch_projected_area}
\end{figure}{}
The physics behind the growth of the bubble cloud becomes more clear if we focus on the time derivative of the projected area rather than on the area itself, as we explain in the next subsection. Figure~\ref{fig:dAdt_sim_different_zetas}$a$ shows how, for $\zeta \ge 2$, the evolution of both the projected area and the total gas volume is qualitatively independent of this value, just affecting the absolute value of the growth rates.
Indeed, examination of the Epstein--Plesset equation (\ref{eq:epstein_plesset_dimensionless}) suggests that, provided the radius of the bubble is sufficiently larger than the initial one, we can assume $a \sim \sqrt{\tau}$. In this case, the growth rate of $a^2$, which is roughly proportional to the projected area, scales with $\left(\zeta-1\right)$. To check this scaling, we plot in Figure~\ref{fig:dAdt_sim_different_zetas}$b$ the time derivative of the dimensional projected area divided by the saturation level, $\zeta - 1$. As predicted, the curves collapse fairly well at intermediate and long times. At short times, because the bubble radii are still close to the initial ones, the collapse is not so good. Furthermore, the curve corresponding to the smallest saturation level, $\zeta = 1.1$, does not exhibit such a good collapse at any time. This is due to the influence of the initial bubble radius, which lasts longer due to the smaller growth rate.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{figures/figure8.pdf}
\caption{($a$) Time derivative of the measured projected area and the measured 2/3 root of the volume for numerically created bubble clouds with identical initial configurations but different saturation levels, $\zeta$. ($b$) Time derivative of the projected area divided by $(\zeta-1)$. Note that, to facilitate the comparison with the experiments, we show here the numerical results in dimensional form, using $R_0 = 30 \mu$m and $R_0^2/D \approx 0.47$ s as length and time scales respectively.}
\label{fig:dAdt_sim_different_zetas}
\end{figure}
The model predicts the existence of three different stages in the evolution of the cloud which, despite their noise and variability, are also observed in the experiments. At short times, the area grows relatively fast, at a nearly constant rate. Later on, the rate of change of the area decays slightly faster than the square root of the time. Finally, at long times, the area grows again at a constant rate, albeit at a much slower pace.
The last two regimes can be easily explained with a simple heuristic model. Let us assume that the volume of the cloud is the sum of that of the $N$ bubbles plus their associated water volume (a sphere of radius $d/2$ for each bubble). Since the water volume in the cloud does not change, the radius of such a cloud would then be
\begin{equation}
a_{bc} = \left(N\left(\frac{d}{2}\right)^3 + \sum_i a_i^3\right)^{1/3} = N^{1/3}\left(\left(d/2\right)^3 + a^3\right)^{1/3}.
\end{equation}
where, for simplicity, we take all bubbles to have the same size.
The projected area, $A_{bc} = \pi a_{bc}^2$, has a rate of change
\begin{equation}
\frac{\mathrm{d}A_{bc}}{\mathrm{d}\tau} = 2\pi N^{2/3} \frac{a^2 \dot{a}}{\left(\left(d/2\right)^3 + a^3\right)^{1/3}}.
\label{eq:theoretical_rate_change_A}
\end{equation}
At short times, the radius of the bubble does not differ much from its initial one, $a \approx 1$, whereas its time derivative goes as $\dot{a} \sim \tau^{-1/2}$, as predicted by the Epstein--Plesset equation. Introducing this into Equation (\ref{eq:theoretical_rate_change_A}), we get $\mathrm{d}A_{bc}/\mathrm{d}\tau \sim \tau^{-1/2}$. This is in fairly good agreement with the behavior observed at intermediate times in both experiments and theory. Actually, the rate of change of the area seems to decay slightly faster than this prediction, which is consistent with the fact that bubbles deep inside the cloud stop growing once they deplete their surroundings from CO$_2$. In other words, the effective value of $N$ in Equation (\ref{eq:theoretical_rate_change_A}) also decays with time, until it only refers to those bubbles located in an outer shell of thickness of order $\Delta$.
At long times, the radius of the bubbles in the outer shell still grows as $a \sim \tau^{1/2}$, albeit with a prefactor smaller than that predicted by Epstein--Plesset, as discussed above. Moreover, $a \gg d/2$. In this limit, Equation (\ref{eq:theoretical_rate_change_A}) yields $\mathrm{d}A_{bc}/\mathrm{d}\tau \sim a \dot{a}$, which is constant. As we will see in the next subsection, this is indeed what is observed to occur in the simulations and is qualitatively consistent with the experiments as well, despite the noise.
At short times, the model predicts a fast growth rate, not only in absolute value, which is to be expected as a consequence of bubbles growing without competing between them, but also in terms of how slowly it decays. Indeed, although Equation (\ref{eq:theoretical_rate_change_A}) suggests that $\mathrm{d}A_{bc}/\mathrm{d}\tau \sim \tau^{-1/2}$ at short times as well, we observe that $\mathrm{d}A_{bc}/\mathrm{d}\tau$ is rather constant instead. This is a consequence of the fact that the void fraction inside the cloud grows fast at short times. Thus, the growth of the projected area is the addition of two effects: its boundaries grow, as predicted by Equation (\ref{eq:theoretical_rate_change_A}), and the interior of the projected cloud image becomes more opaque as more bubbles overlap on the projection.
Naturally, if we compute the total gas volume inside the cloud, $V$, we observe a much closer agreement with $\mathrm{d}V^{2/3}/\mathrm{d}t \sim \tau^{-1/2}$, since the fact that bubbles overlap in the image does not affect the evolution of the volume. This can be observed in figure~\ref{fig:dAdt_sim_different_zetas}, where the evolution of the time derivative of $V^{2/3}$ is compared with that of the projected area of the cloud $A$ for simulations with identical initial bubble distributions but different saturation levels.
\subsection{Comparison with experiments}
In this section we analyze the evolution of the size of the bubble clouds observed in our microgravity experiments, comparing it with that predicted by the model. Several key features of the bubble cloud, such as the number of bubbles or their initial size and spatial distributions cannot be determined experimentally. For this reason, the comparison will be predominantly of qualitative nature and restricted to the area of the bubble cloud projected on an image, which is the magnitude that can be measured experimentally. Nonetheless, despite this limitation, we will show that the different regimes and growth rate behaviors predicted by the model are consistent with those observed experimentally.
An important aspect is to determine the saturation level of CO$_2$ of the water in the experiments. Some isolated bubbles appear in almost all the experiments, and we use them to determine the saturation level, $\zeta$, as done in \cite{VegaMartinez_etalMGST2017}. To this end, we track the evolution of their radii by imaging processing. Then, we estimate $\zeta$ by fitting that evolution to that predicted by the Epstein--Plesset equation \cite{EpsteinPlessetJCP1950}, obtaining values between 3.5 and 4.3, depending on the experiment.
As stated above, the high bubble density of the cloud precludes the observation of its interior. Consequently, it is not possible to determine experimentally the number of bubbles nor their size distribution, at least not accurately. However, as soon as the capsule brakes upon reaching the bottom of the tower, the sudden appearance of a large apparent gravity allows us to instantly observe those bubbles that were at the center of the cloud (figure \ref{fig:minibubbles_exp}). This observation reveals that bubbles deep inside the cloud are much smaller than those near the outer edge, as predicted by the model (figure \ref{fig:radii_table}).
\begin{figure}[h]
\centering
\includegraphics[width =0.5\columnwidth]{figures/figure9.pdf}
\caption{Snapshot of a high-speed movie showing the cloud during the deceleration of the capsule ($\zeta \approx 4$). As the capsule comes to a stop, the large deceleration makes the bubble cloud rise fast. Because large bubbles rise faster, this allows us to see those that were inside the cloud (within the red ellipse region). See also movie in the supplemental material.}
\label{fig:minibubbles_exp}
\end{figure}
Figure~\ref{fig:Areas_drops} shows the time evolution of the projected area for the six drops. In this and the following plot, all the magnitudes are plot with dimensions, to give an idea of the orders of magnitude of the spatial and temporal scales. The jumps observed in some of the curves are due to the merging or separation of individual bubbles from the main cloud.
Besides the fact that all the curves grow monotonically, it is not obvious to extract any trend, specially not the existence of different regimes. In part, this is a consequence of the large dispersion in the growth rates of the projected area, which arises from the variability in the initial conditions of the bubble cloud caused by the collapse of the spark-induced cavitation bubble. But also, even if all the clouds were generated identical, the projected area at a given instant is an integral of all the previous growth rates. Thus, this magnitude is not suitable to identify different stages in the cloud growth rate. For this reason, we focus our analysis on the time derivative of the projected area, which is an instantaneous magnitude, i.e. not depending on the bubble cloud history or initial conditions.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\columnwidth]{figures/figure10.pdf}
\caption{Evolution of the measured projected area of six bubble clouds. The level of saturation of the water in these experiments varies between 3.5 -- 4.3. }
\label{fig:Areas_drops}
\end{figure}
Figure~\ref{fig:dAdt} shows in logarithmic scale the rate of change of the area of the cloud for the six drops carried out in the campaign (see appendix for details on how the time derivative of the area is calculated). The curves in the figure have been scaled with their value at the center of the time range where they follow the law $\mathrm{d}A/\mathrm{d}t \sim t^{-1/2}$ to be able to compare the evolution of clouds with very different initial sizes and thus growth rates. We plot also the rate of change of the area computed from a numerical simulation with $\zeta = 4$.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{figures/figure11.pdf}
\caption{Rate of change of the projected area for the six drops, compared with that predicted by a simulation with $\zeta = 4$. The curves have been scaled with their corresponding value at the center of the time range where they follow the scaling law $\mathrm{d}A/\mathrm{d}t = C t^{-1/2}$, that is $t_r = 0.4$ s. Here, $C$ is the fitting constant of each experiment.}
\label{fig:dAdt}
\end{figure}
In the experiments, the behavior at short times exhibits a larger variability than that predicted by the model. In some experiments we observe the trend $\mathrm{d}A/\mathrm{d}t \sim t^{-1/2}$ from the first instants, whereas in others the projected area shows an initially constant growth rate. We attribute this anomalous behavior to the fact that, at such short times, some of the clouds still keep some residual velocity as a result of the implosion of the cavitation bubble which gave them birth. As this initial velocity dissipates, the projected area transitions to the regime where it decays as the inverse of the square root of the time. This residual velocity can be observed in the high-speed movies (See movies in the supplemental material). Finally, despite the experimental noise, it is possible to infer that the growth rates stabilizes and decays slower than $t^{-1/2}$, consistently with the model predictions.
\section{Conclusions}
We report here experiments where dense bubble clouds grow by diffusion in a supersaturated liquid under microgravity conditions. The bubble clouds are created upon the implosion of a cavitation bubble generated by a spark. Although this way of producing the clouds introduces some variability in the experiments, the evolution of the six bubble clouds studied is qualitatively identical, which suggests that the diffusive growth mechanism that we study is fairly independent of the particular size or shape of the cloud.
To achieve microgravity conditions, the experiments have been carried out in the drop tower facility of the German Center of Applied Space Technology and Microgravity (ZARM). The absence of gravity allows us to observe the purely diffusive growth dynamics of the bubble for several seconds, more than an order of magnitude longer than what could be observed under normal gravity\cite{rodriguez2014}, where bubble buoyancy would quickly become dominant. Moreover, we avoid other effects such as natural convection \cite{Enriquez_etalJFM2014, Moreno_soto_etalJFM2019} inside the liquid phase.
Inspired by the experimental observations, we have developed a mathematical model where each bubble grows competing for the available CO$_2$ with the others in the cloud, which are treated as point mass sinks. Although, strictly speaking, the model is only valid in the limit of a dilute bubble cloud, it predicts reasonably well the qualitative evolution observed in experiments. The reasonably good agreement of the model with the observations has allowed us to explain, using simple heuristic arguments, the experimental observations. Moreover, this model has been used to investigate aspects that could not be determined experimentally, namely the final distribution of bubble sizes inside the cloud and the effect of the saturation level.
\section*{Acknowlegdements}
\textit{This work was supported by the Netherlands Center for Multiscale Catalytic Energy Conversion (MCEC), an NWO Gravitation programme funded by the Ministtry of Education, Culture and Science of the government of the Netherlands and the Spanish FEDER/Ministry of Science, Innovation and Universities--Agencia Estatal de Investigaci\'on though grants DPI2017-88201-C3-3-R and DPI2018-102829-REDT, partly funded through European Funds. We also acknowledge the support of the European Space Agency (ESA) for providing access to the drop tower through grants HSO/US/2015-29/AO and HRE/RS-PS/2018-7/AO. We thank the team from the ZARM Drop Tower Operation and Service Company (ZarM FAB mbH) for their valuable technical support during the experimental campaigns, and the technical staff of the Dept. of Thermal and Fluids Engineering of Universidad Carlos III de Madrid for their assistance in developing the experimental set up. We are indebted to Dr. L. Champougny for her valuable comments about the manuscript.}
| b8bda307b4b3d1a2ce16ef6b8eabcea04936e978 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{P}{recise} position sensing is a highly-sought ability for countless context-aware devices in our modern life, including wireless communication with new-generation protocols relying on beam-forming, high-value asset tracking and customer analytics in retail, ambient-assisted living solutions for remote health care, intruder localization in classified facilities, and victim-detection technologies for first responders. Microwave-based sensing solutions are appealing due to their ability to operate through optically opaque materials or fog, their independence of external illumination and target color, limited potential privacy infringements and the non-ionizing nature of microwaves. Moreover, existing wireless infrastructure can often be leveraged, endowing it with a dual communication and sensing functionality.
Traditional microwave position sensing relies on ballistic wave propagation and leverages ray tracing approaches, the simplest example being triangulation. Unfortunately, the above-listed applications involve irregular propagation environments which give rise to significant multi-path effects. In some cases, the position to be identified may not even be within the sensor's line of sight but hidden around a corner. In such complex environments, a propagating wave front can get completely scrambled such that its angle or time of arrival cannot be used for position sensing with conventional ray-tracing analysis.
Considerable research effort thus goes into overcoming the issues posed by multi-path effects, for instance, using distributed sensor networks encircling the region of interest in combination with a statistical analysis of shadowing effects and/or geometry-based environment models to account for reflections as virtual anchors~\cite{witrisal2016high,leitinger2019belief,mendrzik2019enabling,li2019massive,wymeersch2019radio}.
A completely different approach consists in embracing the complexity of the propagation medium as virtue rather than obstacle. An indoor environment is electrically large compared to the wavelength and can be characterized as ray-chaotic: the separation of two rays launched from the same location in slightly different directions will increase exponentially in time. A wave-chaotic field is extremely sensitive to both source location and the enclosure's geometry. Inspired by the quantum-mechanical concept of fidelity loss~\cite{kuhl2016microwave}, this sensitivity has been leveraged to distinguish nominally identical enclosures~\cite{hemmady2014apparatus}, to detect the presence or motion of small changes in the enclosure's geometry (without localizing them)~\cite{taddese2009sensor,del2018dynamic} as well as to quantify volume changing perturbations~\cite{taddese2013quantifying}. For the problem of position sensing, the wave-chaotic field's sensitivity implies that different positions are associated with distinguishable wave fields that can act like wave fingerprints (WFPs) for the positions.
WFPing can be be applied to the localization of cooperative objects (emitting a beacon signal or equipped with a tag)~\cite{jin2008position,sen2012you,he2015wi,wu2015time,chen2016achieving,ing2005solid} as well as to non-cooperative objects (no compliance with localization task)~\cite{cohen2011subwavelength,xiao2013pilot,del2018precise}. While the former leverages the sensitivity of ray chaos to the source location, the latter leverages its sensitivity to geometrical perturbations. From the wave's point of view, different object positions inevitably correspond to different geometries of the propagation environment. To ensure the distinguishability of WFPs, the chaotic wave field must be probed in a number of "independent" ways. Traditionally, this is achieved using spatial or spectral diversity with a network of sensors or broadband measurements. A more recent alternative is to use configurational diversity by reprogramming the propagation environment with a "reconfigurable intelligent surface" (RIS). Using a programmable metasurface as RIS, Ref.~\cite{del2018precise} leveraged configurational diversity to localize multiple non-cooperative objects outside the line-of-sight with single-port single-frequency measurements.
With real-life applications in mind, a fundamental challenge for indoor localization with WFPs arises: how does one handle a dynamic evolution of the propagation environment independent of the objects of concern? Indeed, given the extreme sensitivity of the chaotic wave field to geometrical details, one could expect that a perturbation not related to the object to be localized alters the wave field to an extent that makes it irrecognizable in light of a previously established WFP dictionary.
Here, we systematically study the impact of perturbations of the propagation environment on the localization accuracy, considering a frequency-diverse model system both in simulation and experiment. We investigate an interpretation of the perturber as effective source of noise and the extent to which the perturber affects the diversity of the WFP dictionary. We demonstrate that the reduction of the amount of information that can be obtained per measurement as the perturber size is increased can be compensated by taking more measurements, even in the regime where the perturber's scattering strength exceeds that of the object to be localized. Our results stress the importance of appreciating the information-theoretic encoding/decoding cycle of the sensing process in its entirety and reveal that machine-learning decoders outperform traditional decoding techniques especially in the low-SNR regime.
\section{Experimental Setup and WFP Formalism}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig1}
\caption{Experimental setup (top wall removed to show interior). The triangular object to be localized (base $9 \times 9\ \mathrm{cm}^2$, height $6.5\ \mathrm{cm}$) is placed on one of $P=5$ predefined positions (here position \#3) inside a complex scattering environment. The transmission between two monopole antennas is measured with a LimeSDR Mini. The object to be localized is outside the antenna pair's line-of-sight. A dynamic perturber consists of a metallic object of variable size (here the third largest) mounted on a stepper motor. The top inset shows the different considered sizes of the dynamic perturber in comparison to the object size. The three largest perturbers are obtained by mounting a U-shaped extension on a smaller perturber, similar to the spirit of Matryoshka dolls. The bottom inset illustrates the WFP multiplexing mechanism.}
\end{figure}
Our experimental setup is shown in Fig.~1: an object is located on one of $P=5$ possible predefined positions in an irregular metallic enclosure. $N=51$ complex-valued transmission measurements between two simple monopole antennas are taken in the interval $1\ \mathrm{GHz} < f < 2.58\ \mathrm{GHz}$ with a software-defined radio (SDR, LimeSDR Mini). Note that the predefined object positions are clearly outside the line-of-sight of the antenna pair. Dynamic perturbations of the propagation environment are introduced in our experiment with a metallic object of variable size mounted on a stepper motor which can place the object in an arbitrary angular orientation.
A measured transmission spectrum $S(f)$ can be decomposed into four contributions:
\begin{equation}
S(f) = S_{cav}(f) + S_{obj}(f) + S_{pert}(f) + \mathcal{N}(f).
\end{equation}\label{Sdecomp}
\noindent $S_{pert}(f)$ accounts for rays that encountered the perturber, $S_{obj}(f)$ accounts for rays that encountered the object but not the perturber, $S_{cav}(f)$ accounts for rays that bounced around in the cavity without encountering object or perturber, and $\mathcal{N}(f)$ denotes the measurement noise. Given the chaotic nature of the complex scattering enclosure, it is customary to assume that real and imaginary components of the entries of the first three terms are drawn from zero-mean Gaussian distributions. The measurement noise is typically also zero-mean Gaussian. The decomposition in Eq.~1 has several subtleties. First, we note that if the perturber size is increased, more rays will encounter the perturber such that not only will the elements of $S_{pert}(f)$ be drawn from a distribution with larger standard deviation, but at the same time the standard deviation of the distributions of $S_{cav}(f)$ and $S_{obj}(f)$ will decrease. In other words, $S_{cav}(f)$ and $S_{obj}(f)$ are not independent of the perturbing object. Second, since all the terms are assumed to be drawn from zero-mean distributions, in principle one would expect that by averaging over an ensemble of realizations of the perturber one can estimate $S_{cav}(f) + S_{obj}(f)$ and by additionally averaging over an ensemble of object positions one can identify $S_{cav}(f)$. In practice, proper averaging requires a sufficient number of realizations and $P=5$ may be insufficient for averaging over an ensemble of object positions.
In Eq.~1, only the term $S_{obj}(f)$ encodes information about the object position. To determine a WFP in the presence of a perturber, we therefore average $S(f)$ over an ensemble of representative perturber realizations. Here, it is relatively easy to ensure that the ensemble is sufficiently large to estimate $S_{cav}(f) + S_{obj}(f)$ properly. We can then either define the WFP as being $S_{cav}(f) + S_{obj}(f)$ or we can intend to approximate $S_{obj}(f)$ with
\begin{equation}
S^{(2)}_{obj}(f)=S_{cav}(f) + S_{obj}(f) - \langle S_{cav}(f) + S_{obj}(f)\rangle_{obj}.
\end{equation}\label{Sdecomp}
\noindent We will consider both options below and see that, counter-intuitively, the former one can be advantageous in certain cases. Moreover, Eq.~1 naturally suggests to interpret the perturber as an effective source of noise. We can quantify the scattering strength of the perturber relative to that of the object via an effective perturber-induced SNR $\rho_p$. Ideally, to that end, we would define $\sigma_{s}$ and $\sigma_{n}$ to be the standard deviation of the distributions from which the entries of $S_{obj}(f)$ and $S_{pert}(f)$, respectively, are drawn, to define $\rho_p = \sigma_{s}^2/\sigma_{n}^2$. In practice, we do not know $S_{obj}(f)$. Depending on whether we choose to use $S_{cav}(f) + S_{obj}(f)$ or $S^{(2)}_{obj}(f)$ as WFP, we can define $\sigma^{(1)}_{s}$ and $\sigma^{(2)}_{s}$ to be the respective standard deviation, yielding $\rho^{(1)}_p$ and $\rho^{(2)}_p$. These effective SNRs quantify to what extent the perturber acts as noise on our chosen WFP, but do not directly reflect the ratio of scattering strengths of object and perturber.
The $P \times N $ WFP dictionary $\mathbf{H}$ merges the $P$ WFPs (each WFP is an $N$-element vector) into a single matrix. The WFP approach can then also be framed as a multiplexing problem $Y=\mathbf{H}X+\mathcal{N}$, where $X$ is a $1\times P$ vector identifying the object position, $Y$ is the complex-valued $1\times N$ measurement vector and $\mathcal{N}$ is a $1\times N$ noise vector.
\section{Information-Theoretic Perspective}
One prerequisite for successful WFPing is the diversity of $\mathbf{H}$. In our case, the complexity of the propagation environment naturally provides this diversity. The lower the correlations between different WFPs are, the better they can be distinguished. To get a quantitative grasp of the diversity of $\mathbf{H}$, it is instructive to consider its singular value (SV) decomposition: $\mathbf{H} = \mathbf{U} \mathbf{\Sigma} \mathbf{V}^T$, where $\mathbf{\Sigma} $ is a diagonal matrix whose $i$th entry is the $i$th SV $\sigma_i$ of $\mathbf{H}$. The flatter the SV spectrum is, the more diverse is $\mathbf{H}$. A convenient metric of diversity is the effective rank of $\mathbf{H}$ which is defined as $R_{\mathrm{eff}} = \mathrm{exp}\left( -\sum_{i=1}^n \tilde{\sigma}_i \mathrm{ln}(\tilde{\sigma}_i) \right)$, where $\tilde{\sigma}_i = \sigma_i / \sum_{i=1}^n \sigma_i$ and $n=\mathrm{min}(N,P)$~\cite{roy2007effective}.
Note that only perfectly orthogonal channels with zero correlation yield $R_{\mathrm{eff}} = n$.
Unfortunately, much of the compressed sensing literature is exclusively focused on the diversity of $\mathbf{H}$ to understand the achievable performance. For instance, compression ratios are often provided without even indicating at what SNR they are valid. In principle, in the absence of any noise, the tiniest amount of diversity could be sufficient to ensure complete distinguishability even with $N=1$. Here, we argue that the achievable performance depends on the amount of (useful) information that can be extracted per measurement. In the physical layer, besides diversity the SNR is a second crucial ingredient. Moreover, high diversity and low SNR only ensure good performance if the deployed decoding method in the digital layer is capable of extracting much of the relevant encoded information from the measurement.
\begin{figure}[t]
\centering
\includegraphics[width = \columnwidth]{Fig2}
\caption{Information about the object position is (inevitably) physically encoded in the measured data via wave scattering in the irregular propagation environment. Digital data processing then seeks to retrieve the information from the measurements.}
\end{figure} \label{figIT}
WFP-based sensing in its entirety as schematically summarized in Fig.~2 can be interpreted as a process consisting of physical encoding and digital decoding of information. Wave propagation through the complex scattering environment naturally (and inevitably) encodes information about the object position in measurements of the wave field. Data processing seeks to retrieve this information. Various decoding methods exist that we will compare later on:
(i) \textit{Correlation.} Identify which row of $\mathbf{H}$ has the highest correlations with $Y$. This procedure can be interpreted as "virtual time reversal"~\cite{ing2005solid}.
(ii) \textit{Inversion.} Compute an inverse of $\mathbf{H}$, for instance, via Tikhonov regularization, and identify the entry of $\mathbf{H}^{-1}Y$ with the largest magnitude.
(iii) \textit{Optimized Inversion.} Use the result from (ii) as initial guess in a non-linear minimization of $|| Y-\mathbf{H}X ||$~\cite{redding2012using,redding2013compact}.
(iv) \textit{Learning.} Train an artificial neural network (ANN) to map $Y$ to the corresponding object position. ANN-based approaches have not been studied in the multiplexing literature to date. Besides their potential for superior decoding performance, inference is extremely fast. One forward pass through an ANN only requires a few matrix multiplications but no correlations, matrix inversions or nonlinear optimization routines.
From an information-theoretic perspective, it is important to understand fundamental bounds on the sensing performance.
A simple bound to compute is the generalized Shannon capacity
\begin{equation}
C = \sum_i \mathrm{log}_2 \left( 1+\frac{\rho}{P}\sigma_i \right)
\end{equation}\label{Sdecomp}
which has been mentioned on a few occasions in a sensing context~\cite{migliore2008electromagnetics,lorenzo2015single}. Nonetheless, the meaningfulness of $C$ for a specific sensing scheme is limited for two reasons. First, an ideal input distribution is assumed for $X$ but in reality all entries of $X$ are zero except for one which is unity. Second, an ideal decoding method is assumed. Below we will see examples where a system with nominally lower $C$ nonetheless yields a higher sensing accuracy for certain decoding methods.
It is thus essential to appreciate the sensing process in its entirety, including both encoding and decoding as illustrated in Fig.~2.
Having introduced the notion of diversity and SNR, we can now briefly comment on how faithfully the metallic enclosure in our experiment represents real-life scenarios. Without a doubt, certain cases like the inside of a vessel or a bank vault are very well represented.
Other environments like the inside of a building are less reverberant than a metallic enclosure. Essentially, the quality factor of these "cavities" is lower. This implies more correlations within a fixed frequency interval of the transmission spectrum, as well as a lower SNR due to more attenuation. Both result in a decrease of the information that can be extracted per measurement; this effect can be compensated by taking more measurements, for instance, with a wider bandwidth. Nonetheless, from a fundamental perspective, the physics of an indoor system is entirely captured by our metallic enclosure.
In scenarios with already existing wireless communication infrastructure, the beacon signals thereof could be used to implement position sensing with WFPs,
saving energy and reducing the amount of electromagnetic radiation.
\section{Semi-Analytical Simulations}
To begin with, we consider a 2D version of our experiment simulated as a 2D system of coupled dipoles~\cite{orazbayev2020label} which contains all the essential physical ingredients to simulate wave propagation, reverberation and scattering in our experiment.
These simulations offer an ideal platform to identify the effect of dynamic perturbations of the propagation environment on the sensing accuracy without any measurement noise or errors due to imperfect object positioning on the predefined positions, i.e. $\mathcal{N}(f)=0$. As shown in Fig.~3(a), a perturber of variable size with arbitrary orientation and location (within a specified area) simulates dynamic changes of the environment. Our simulation setup evaluates the transmission between an antenna pair at $25$ distinct frequencies. We use an ensemble of $150$ random perturber realizations (random orientation and random location of its center within the allowed area) to estimate $\mathbf{H}$, $R_{\mathrm{eff}}$ and $\rho_p$. The probability density function (PDF) of real and imaginary part of $S_{pert}$ is seen in Fig.~3(b-e) to be zero-mean single-peaked and tends towards a Gaussian distribution for larger perturbers.
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig3}
\caption{(a) Setup of semi-analytical coupled-dipole 2D simulations. A line-like object is placed on one of $P=5$ predefined positions in an irregularly shaped enclosure (dipole fence) of dimensions on the order of $25\times15$ wavelengths. A line-like perturbing object with variable length is randomly rotated and located such that its center lies within the indicated area. The transmission between TX and RX is evaluated. See Ref.~\cite{orazbayev2020label} for technical details on the simulation method. (b-e) PDF of real and imaginary parts of $S_{pert}$ and a Gaussian fit are shown for the smallest and largest considered perturber size.}
\end{figure} \label{figCDCsetupX}
\subsection{Impact of Perturbation on Diversity and Effective SNR}
In Fig.~4 we contrast the use of $S_{cav}(f) + S_{obj}(f)$ or $S^{(2)}_{obj}(f)$ as WFP in terms of the resulting diversity ($R_{\mathrm{eff}}$), effective SNR ($\rho_p$) and sensing capacity ($C$). As we will see below, neither of these quantities is a reliable predictor of the sensing accuracy, since they do not take the decoding method into account. For the case of using $S_{cav}(f) + S_{obj}(f)$ as WFP, the observed trend is clear: as the perturber size increases, both $R_{\mathrm{eff}}$ and $\rho_p$ as well as $C$ decrease. While the impact on $\rho_p$ was clearly expected, the reduction of diversity is more subtle. It becomes intuitive by considering the extreme case in which the perturbation alters the entire enclosure. Then, averaging over realizations yields the result that would have been obtained in an anechoic environment such that no diversity thanks to wave chaos is left.
\begin{figure}[h]
\centering
\includegraphics[width = \columnwidth]{Fig4}
\caption{(a) Effect of perturber size in the $R_{\mathrm{eff}}-\rho_p$ plane in the semi-analytical simulations. Curves for defining the WFP as $S_{cav}+S_{obj}$ or $S^{(2)}_{obj}$ are shown for three setups. All three setups are like the one in Fig.~3 but perturber area, predefined object positions and antenna positions are moved around. In all cases the objects are outside the antenna pair's line of sight. (b) Sensing capacity values corresponding to the data in (a). The results in this figure are obtained using all 25 frequency points. To ease comparison with Fig.~6, we normalized $\Sigma_{i=1}^n \sigma_i^2$ to unity; the SNR $\rho$ in Eq.~3 incorporates adverse effects on the dynamic range due to pathloss.}
\end{figure} \label{ReffrhoCDC}
Using $S_{obj}(f)$ as opposed to $S_{cav}(f) + S_{obj}(f)$ would certainly improve the diversity by removing unnecessary correlations (possibly at the expense of a better SNR such that the overall effect on capacity is unclear), but this is not possible in practice. Our closest option to that effect is to use $S^{(2)}_{obj}(f)$. Straight-forward simulations with random Gaussian matrices show that the effective rank of $S_{cav}(f) + S_{obj}(f)$ may exceed that of $S^{(2)}_{obj}(f)$ in cases where $P$ is small (preventing proper averaging over realizations of the object position) and where the ratio of the standard deviations of the distributions of $S_{obj}$ and $S_{cav}$ is large. Nonetheless, in our semi-analytical simulations, we observe in Fig.~4(a) a higher effective rank for $S^{(2)}_{obj}(f)$ than for $S_{cav}(f) + S_{obj}(f)$. Yet, since $\rho^{(2)}_p$ is substantially lower than $\rho^{(1)}_p$, the effect of using $S^{(2)}_{obj}(f)$ on the capacity is unfavorable.
Complex scattering enclosures are often seen as random field generators~\cite{corona1996reverberating}. $R_{\mathrm{eff}}$ is a measure of the number of independent samples and for $N\gg P$ one expects $R_{\mathrm{eff}} \rightarrow P$. Yet, in our simulations, $R_{\mathrm{eff}}$ saturates below $4$. This observation can be attributed to field correlations, here in the frequency domain, that prevent the field observables from being purely random variables~\cite{cozza2011skeptic}.
\subsection{Dependence of Sensing Accuracy on Perturber Size, Number of Measurements and Decoding Method}
The general trend is clear: the larger the perturbation, the less information can be extracted per measurement, as reflected by the sensing capacitance values plotted in Fig.~4(b). However, this decrease in information per measurement can be compensated with more measurements. At first sight, one may expect that such a compensation is only feasible as long as the object's scattering signature is stronger than the perturber's effect, i.e. for $\rho_p > 0\ \mathrm{dB}$. Our findings in Fig.~5, however, reveal that there is no abrupt phase change in the relation between achievable accuracy versus perturber size. Instead, using more measurements, successful position sensing is feasible at effective SNRs well below $0\ \mathrm{dB}$.
\begin{figure}[h]
\centering
\includegraphics[width = \columnwidth]{Fig5}
\caption{Localization accuracy in semi-analytical simulations. The colorscale goes from 0 (black) to 1 (white). For each choice of WFP definition (columns) and decoding method (rows), the accuracy is plotted as a function of perturber size (horizontal axis) and the number of frequency points used to ink the WFP (vertical axis). ANN results are averaged over 20 training runs with randomly initialized weights; the standard deviation is below 2 \%. The black contour line corresponds to $95\ \%$ accuracy. To aid comparison, the red contour is the same on all subfigures.}
\end{figure} \label{AccCDC}
We systematically compare the previously outlined decoding methods for both choices of WFP. For the learning-based approach, we train an artificial neural network (ANN) consisting of two fully connected layers; the first layer consists of 256 neurons and is followed by a ReLu activation, the second layer consists of $P=5$ neurons and is followed by a SoftMax activation. Using more neurons or an additional layer does not appear to notably impact the results. We consider two possibilities to provide training data from which the ANN can learn to decode the measurements. The first option is to simply use the raw data from all the perturber realizations that we generated without a need for extracting $\mathbf{H}$ or other quantities. This brute force method may prove particularly useful in cases where measurements are restricted to intensity-only information which prevents averaging as simple means to extract $\mathbf{H}$, but this scenario is outside the scope of the present paper. Note that with this approach the WFPs are never explicitly evaluated, but only implicitly contained in the ANN weights. The second option is to synthesize training data with $Y=\mathbf{H}X+\mathcal{N}$ using the estimated $\mathbf{H}$ and generating $\mathcal{N}$ with entries drawn from a Gaussian distribution whose standard deviations match those of the distribution of $S_{pert}$ extracted from the data. This second method relies on our hypothesis that $S_{pert}$ is normally distributed and offers the possibility of generating a training dataset of unlimited size. In both cases we normalize the data (zero mean, unit variance) and use the Adam method for stochastic optimization (step size $10^{-3}$) to train the ANN weights.
In Fig.~5, we show how the achieved sensing accuracy depends on the perturber size and the number of measurements. We ensure that the spacing of the utilized frequency points is always the same and that they are always centered on the same frequency. For instance, for $N=7$ measurements we pick the central frequency point out of the 25 available ones as well as its three closest neighbours to the left and right. Our results are thus for one specific system realization which explains why the contours in Fig.~5 are not perfectly smooth. Several important observations and conclusions follow from Fig.~5:
(i) WFP dictionaries with very different nominal sensing capacities can yield the same accuracy. This is the case for both ANN-based methods in which the accuracy is (almost) identical for $WFP^{(1)}$ and $WFP^{(2)}$.
(ii) The same WFP dictionary can yield very different accuracies depending on the decoding method. ANN-based decoders are seen to outperform correlation and inversion-based decoders.
(iii) The choice of WFP definition is irrelevant for the optimized inversion decoder as well as the ANN-based decoders. For correlation and inversion based decoders, however, using $WFP^{(1)}$ yields significantly better results.
(iv) Irrespective of the perturber size, we achieve an acceptable minimum accuracy (e.g. 95 \%). For larger perturbers, we need more measurements to compensate the reduction in the amount of information that can be extracted per measurement. Future information-theoretic work should seek to model the contour for a given accuracy in order to understand how the need for additional measurements scales with $\rho_p$.
(v) At low effective noise levels, some decoders achieve compression ratios above unity, that is they achieve accuracies $\geq 95\ \%$ to localize $P=5$ objects with $N<P$ measurements. For instance, the ANN (raw data) decoder with $WFP^{(2)}$ achieves 96 \% with $N=3$ at the lowest considered perturber size. However, as in any compressed sensing scenario, it is obvious that the compression ratio is heavily dependent on the noise level (here, the effective noise level due to the perturber size), the independence of different measurements (here, determined to a large extent by the interval between frequency points) and the decoding method (here, an ANN trained with raw data). Thus, a general claim of achieving a compression ratio above unity is not presented as key result of this work.
Overall, these results clearly demonstrate that it is fallacious to assume that the diversity or sensing capacity of $\mathbf{H}$ could be a reliable indicator of the sensing accuracy, hence the importance of considering the sensing process in its information-theoretic entirety as in Fig.~2.
\section{Experimental Results}
Having established an understanding of the perturber's effect under idealistic conditions in simulation, we now analyze the experimental data. Measurements with our SDR entail a few practical issues. First, there is a $\pm\pi$ uncertainty in measured phase values, originating from random phase jumps every time the Phase Locked Loop (PLL) is locked (e.g. to change the frequency). To obtain reliable data, we transform each measured complex value $z$ to $|z|\ \mathrm{exp}(2i\ \mathrm{mod}(\mathrm{arg}(z),\pi))$; the factor $2$ in the exponent ensures that the transformed variable's phase explores the entire $2\pi$ range. Second, the transmitted energy is clearly frequency-dependent, which can be caused by the frequency-dependent coupling of the monopole antennas to the cavity and/or frequency-dependent SDR components. The strong frequency-dependence means that we cannot simply model our variables as being drawn from a unique distribution, instead the distribution's standard deviation becomes frequency-dependent. To maintain the SDR's temperature constant throughout the experiment, we installed a simple CPU fan. We do not observe any significant amplitude or phase drifts over the course of the experiment.
We begin by quantifying two contributions to the $\mathcal{N}$ term in Eq.~1 that were not present in the simulations. First, we estimate the SNR due to measurement noise (by repeating the same measurement multiple times) as $\rho_1 = 25.5\ \mathrm{dB}$. Second, we estimate the SNR due to both measurement noise and imperfect positioning of the objects on the predefined locations (by repeating the same measurement multiple times after placing the object again on the same position) as $\rho_2 = 15.8\ \mathrm{dB}$.
\subsection{Impact of Perturbation on Diversity and Effective SNR}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig6}
\caption{(a-d) PDF of real and imaginary parts of $S_{pert}$ and a Gaussian fit are shown for the smallest and largest perturber size in the experiment.
(e) Effect of perturber size in the $R_{\mathrm{eff}}-\rho_p$ plane.
The blue curves only consider perturber-induced effective noise, the red curves additionally account for measurement and positioning noise.
(f) Normalized sensing capacity values corresponding to the data in (e). The results in this figure are obtained using all 51 frequency points.}
\end{figure} \label{figGNU_H_Analysis}
Based on 150 perturber realizations (random orientations) for each perturber size, in Fig.~6(a-d) we plot the PDFs of real and imaginary part of $S_{pert}$ for the smallest and largest perturber considered in our experiment. The zero-mean single-peaked distributions are identical for real and imaginary component but thinner than a Gaussian distribution. In Fig.~6(e) we plot $R_{\mathrm{eff}}(\mathbf{H})$ vs the effective SNR. Since $\mathcal{N}(f) \neq 0$ in the experiment, we plot two curves: the blue one only accounts for perturber-induced effective noise, the red one additionally accounts for measurement and positioning noise. The difference between these two curves is appreciable only for small perturber sizes since for larger perturbers $S_{pert}$ dominates over $\mathcal{N}$. Unlike in Fig.~4(a), using $S_{obj}^{(2)}$ lowers not only the effective SNR but also the effective rank. As in Fig.~4(b), we see in Fig.~6(f) that using $S_{obj}^{(2)}$ is unfavorable in terms of the (normalized) sensing capacity. The impact of $\mathcal{N}$ on $C$ is only noticeable for small perturbers.
\subsection{Dependence of Sensing Accuracy on Perturber Size, Number of Measurements and Decoding Method}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig7}
\caption{Localization accuracy in experiments. The colorscale goes from 0 (black) to 1 (white). For each choice of WFP definition (columns) and decoding method (rows), the accuracy is plotted as a function of perturber size (horizontal axis) and the number of frequency points used to ink the WFP (vertical axis). ANN results are averaged over 20 training runs with randomly initialized weights; the standard deviation does not exceed 10 \% and 3 \% for ANNs trained with synthetic and raw data, respectively. The black contour line corresponds to $95\ \%$ accuracy. To aid comparison, the red contour is the same on all subfigures.}
\end{figure} \label{AccGNU}
In Fig.~7 we compare the achievable sensing accuracy in our experiment with the two considered definitions of the WFP and different decoding methods as a function of the perturber size and number of measured frequency points. The observations already made for the corresponding simulation results in Fig.~5 about the unsuitability of $R_{\mathrm{eff}}$ or $C$ to predict the sensing accuracy are confirmed once again by Fig.~7. The most notable difference to Fig.~5 is that except for the ANN trained with raw data all decoding methods fail to achieve at least 95 \% accuracy once the perturber's surface is larger than 200 $\mathrm{cm^2}$. We attribute this to the $\pm\pi$ phase uncertainty of our SDR which introduces errors in the estimation of $\mathbf{H}$. Since the ANN trained with raw data does not rely on calculating $\mathbf{H}$, it is not affected. Interestingly, we have thus a case in which it is better to feed the ANN raw data rather than to use physical insight to pre-process the ANN's training data. The ANN decoder trained with raw data is capable of achieving high sensing accuracies despite significant amounts of noise (the effective SNR is as low as -15~dB for the largest perturber, see Fig.~6(e)) and distorted data.
Using the ANN decoder trained with raw data, we achieve 100 \% sensing accuracy with $N=3$, i.e. a compression ratio of $P/N=5/3>1$, for perturbers with a surface as large as $74\ \mathrm{cm}^2$. Again, we stress that the compression ratio depends on effective SNR, measurement independence and decoding method.
\section{Conclusion and Outlook}
From a practical point of view, our experiments, in combination with an ANN-based decoder, demonstrated the feasibility of precise position sensing with WFPs in dynamically evolving scattering enclosures using a low-cost and light-weight SDR. This capability is crucial to enable situational awareness in a plethora of emerging applications. Our technique does not rely on detailed knowledge about the environment's geometry and only requires a one-off calibration phase with multiple representative realizations of the dynamic perturbations that are expected during operation.
From a conceptual point of view, our work paves the way for a thorough information-theoretic analysis of sensing with WFPs.
The dynamic perturber's unfavorable effect on diversity and effective SNR of the WFP dictionary, resulting in the acquisition of less useful information per measurement, can be fully compensated by taking more measurements -- even in the regime in which the perturber's scattering strength clearly exceeds that of the object to be localized.
We saw that the common practice in compressed sensing to only consider the diversity or capacity of $\mathbf{H}$ is insufficient to anticipate the achievable sensing accuracy.
Our results are of very general nature: they can be applied to other types of wave phenomena (sound, light, ...) and are equally valid for WFPs established with other means such as using spatial or configurational degrees of freedom by employing a sensor network or a RIS~\cite{del2018precise}.
The importance of seeing the entirety of the information-theoretic cycle points towards jointly optimizing encoding in a programmable propagation environment and ML-based decoding, as in the recently proposed "learned sensing" paradigm~\cite{del2020learned,li2020intelligent}. In contrast to compressed sensing which indiscriminately encodes all information, learned sensing seeks to encode only task-relevant information in the measurements. For position sensing, one could carefully select the frequencies at which measurements are taken (as opposed to linear spacing) and/or engineer the propagation environment with a RIS~\cite{del2020optimal}.
Looking ahead, it appears interesting to extend the present work (i) to scenarios with multiple objects to be localized, where neglected inter-object scattering is an additional effective source of noise~\cite{del2018precise}, (ii) to deeply sub-wavelength position sensing~\cite{cohen2011subwavelength}, and (iii) to more complex tasks like image transmission~\cite{li2018deep}.
In this work, the perturber was seen as an obstacle for our task to localize an object. In other contexts, the objective may be to characterize size and motion of a perturber.
Diffuse wave spectroscopy~\cite{FishCounting,fishcountingPRL,CCSgeo,conti2006characterization} analyzes changes of the broadband impulse response over time to estimate the number or scattering cross section of objects moving through a complex medium.
Our work has evidenced that the perturber's scattering strength can also be clearly related to the capacity of a multiplexing channel matrix averaged over different realizations of the perturber's position. Considering configuration-to-configuration multiplexing with two dynamic metasurface transceivers~\cite{sleasman2019implementation} may thus enable similar characterizations of a moving perturber with single-port single-frequency measurements~\cite{del2018dynamic}.
\section*{Acknowledgment}
The author acknowledges Michael del Hougne for help with the experimental work. The code for the semi-analytical model is based on Ref.~\cite{orazbayev2020label}. The code for controlling the SDR is based on Ref.~\cite{Lime}.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\section{Introduction}
\IEEEPARstart{P}{recise} position sensing is a highly-sought ability for countless context-aware devices in our modern life, including wireless communication with new-generation protocols relying on beam-forming, high-value asset tracking and customer analytics in retail, ambient-assisted living solutions for remote health care, intruder localization in classified facilities, and victim-detection technologies for first responders. Microwave-based sensing solutions are appealing due to their ability to operate through optically opaque materials or fog, their independence of external illumination and target color, limited potential privacy infringements and the non-ionizing nature of microwaves. Moreover, existing wireless infrastructure can often be leveraged, endowing it with a dual communication and sensing functionality.
Traditional microwave position sensing relies on ballistic wave propagation and leverages ray tracing approaches, the simplest example being triangulation. Unfortunately, the above-listed applications involve irregular propagation environments which give rise to significant multi-path effects. In some cases, the position to be identified may not even be within the sensor's line of sight but hidden around a corner. In such complex environments, a propagating wave front can get completely scrambled such that its angle or time of arrival cannot be used for position sensing with conventional ray-tracing analysis.
Considerable research effort thus goes into overcoming the issues posed by multi-path effects, for instance, using distributed sensor networks encircling the region of interest in combination with a statistical analysis of shadowing effects and/or geometry-based environment models to account for reflections as virtual anchors~\cite{witrisal2016high,leitinger2019belief,mendrzik2019enabling,li2019massive,wymeersch2019radio}.
A completely different approach consists in embracing the complexity of the propagation medium as virtue rather than obstacle. An indoor environment is electrically large compared to the wavelength and can be characterized as ray-chaotic: the separation of two rays launched from the same location in slightly different directions will increase exponentially in time. A wave-chaotic field is extremely sensitive to both source location and the enclosure's geometry. Inspired by the quantum-mechanical concept of fidelity loss~\cite{kuhl2016microwave}, this sensitivity has been leveraged to distinguish nominally identical enclosures~\cite{hemmady2014apparatus}, to detect the presence or motion of small changes in the enclosure's geometry (without localizing them)~\cite{taddese2009sensor,del2018dynamic} as well as to quantify volume changing perturbations~\cite{taddese2013quantifying}. For the problem of position sensing, the wave-chaotic field's sensitivity implies that different positions are associated with distinguishable wave fields that can act like wave fingerprints (WFPs) for the positions.
WFPing can be be applied to the localization of cooperative objects (emitting a beacon signal or equipped with a tag)~\cite{jin2008position,sen2012you,he2015wi,wu2015time,chen2016achieving,ing2005solid} as well as to non-cooperative objects (no compliance with localization task)~\cite{cohen2011subwavelength,xiao2013pilot,del2018precise}. While the former leverages the sensitivity of ray chaos to the source location, the latter leverages its sensitivity to geometrical perturbations. From the wave's point of view, different object positions inevitably correspond to different geometries of the propagation environment. To ensure the distinguishability of WFPs, the chaotic wave field must be probed in a number of "independent" ways. Traditionally, this is achieved using spatial or spectral diversity with a network of sensors or broadband measurements. A more recent alternative is to use configurational diversity by reprogramming the propagation environment with a "reconfigurable intelligent surface" (RIS). Using a programmable metasurface as RIS, Ref.~\cite{del2018precise} leveraged configurational diversity to localize multiple non-cooperative objects outside the line-of-sight with single-port single-frequency measurements.
With real-life applications in mind, a fundamental challenge for indoor localization with WFPs arises: how does one handle a dynamic evolution of the propagation environment independent of the objects of concern? Indeed, given the extreme sensitivity of the chaotic wave field to geometrical details, one could expect that a perturbation not related to the object to be localized alters the wave field to an extent that makes it irrecognizable in light of a previously established WFP dictionary.
Here, we systematically study the impact of perturbations of the propagation environment on the localization accuracy, considering a frequency-diverse model system both in simulation and experiment. We investigate an interpretation of the perturber as effective source of noise and the extent to which the perturber affects the diversity of the WFP dictionary. We demonstrate that the reduction of the amount of information that can be obtained per measurement as the perturber size is increased can be compensated by taking more measurements, even in the regime where the perturber's scattering strength exceeds that of the object to be localized. Our results stress the importance of appreciating the information-theoretic encoding/decoding cycle of the sensing process in its entirety and reveal that machine-learning decoders outperform traditional decoding techniques especially in the low-SNR regime.
\section{Experimental Setup and WFP Formalism}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig1}
\caption{Experimental setup (top wall removed to show interior). The triangular object to be localized (base $9 \times 9\ \mathrm{cm}^2$, height $6.5\ \mathrm{cm}$) is placed on one of $P=5$ predefined positions (here position \#3) inside a complex scattering environment. The transmission between two monopole antennas is measured with a LimeSDR Mini. The object to be localized is outside the antenna pair's line-of-sight. A dynamic perturber consists of a metallic object of variable size (here the third largest) mounted on a stepper motor. The top inset shows the different considered sizes of the dynamic perturber in comparison to the object size. The three largest perturbers are obtained by mounting a U-shaped extension on a smaller perturber, similar to the spirit of Matryoshka dolls. The bottom inset illustrates the WFP multiplexing mechanism.}
\end{figure}
Our experimental setup is shown in Fig.~1: an object is located on one of $P=5$ possible predefined positions in an irregular metallic enclosure. $N=51$ complex-valued transmission measurements between two simple monopole antennas are taken in the interval $1\ \mathrm{GHz} < f < 2.58\ \mathrm{GHz}$ with a software-defined radio (SDR, LimeSDR Mini). Note that the predefined object positions are clearly outside the line-of-sight of the antenna pair. Dynamic perturbations of the propagation environment are introduced in our experiment with a metallic object of variable size mounted on a stepper motor which can place the object in an arbitrary angular orientation.
A measured transmission spectrum $S(f)$ can be decomposed into four contributions:
\begin{equation}
S(f) = S_{cav}(f) + S_{obj}(f) + S_{pert}(f) + \mathcal{N}(f).
\end{equation}\label{Sdecomp}
\noindent $S_{pert}(f)$ accounts for rays that encountered the perturber, $S_{obj}(f)$ accounts for rays that encountered the object but not the perturber, $S_{cav}(f)$ accounts for rays that bounced around in the cavity without encountering object or perturber, and $\mathcal{N}(f)$ denotes the measurement noise. Given the chaotic nature of the complex scattering enclosure, it is customary to assume that real and imaginary components of the entries of the first three terms are drawn from zero-mean Gaussian distributions. The measurement noise is typically also zero-mean Gaussian. The decomposition in Eq.~1 has several subtleties. First, we note that if the perturber size is increased, more rays will encounter the perturber such that not only will the elements of $S_{pert}(f)$ be drawn from a distribution with larger standard deviation, but at the same time the standard deviation of the distributions of $S_{cav}(f)$ and $S_{obj}(f)$ will decrease. In other words, $S_{cav}(f)$ and $S_{obj}(f)$ are not independent of the perturbing object. Second, since all the terms are assumed to be drawn from zero-mean distributions, in principle one would expect that by averaging over an ensemble of realizations of the perturber one can estimate $S_{cav}(f) + S_{obj}(f)$ and by additionally averaging over an ensemble of object positions one can identify $S_{cav}(f)$. In practice, proper averaging requires a sufficient number of realizations and $P=5$ may be insufficient for averaging over an ensemble of object positions.
In Eq.~1, only the term $S_{obj}(f)$ encodes information about the object position. To determine a WFP in the presence of a perturber, we therefore average $S(f)$ over an ensemble of representative perturber realizations. Here, it is relatively easy to ensure that the ensemble is sufficiently large to estimate $S_{cav}(f) + S_{obj}(f)$ properly. We can then either define the WFP as being $S_{cav}(f) + S_{obj}(f)$ or we can intend to approximate $S_{obj}(f)$ with
\begin{equation}
S^{(2)}_{obj}(f)=S_{cav}(f) + S_{obj}(f) - \langle S_{cav}(f) + S_{obj}(f)\rangle_{obj}.
\end{equation}\label{Sdecomp}
\noindent We will consider both options below and see that, counter-intuitively, the former one can be advantageous in certain cases. Moreover, Eq.~1 naturally suggests to interpret the perturber as an effective source of noise. We can quantify the scattering strength of the perturber relative to that of the object via an effective perturber-induced SNR $\rho_p$. Ideally, to that end, we would define $\sigma_{s}$ and $\sigma_{n}$ to be the standard deviation of the distributions from which the entries of $S_{obj}(f)$ and $S_{pert}(f)$, respectively, are drawn, to define $\rho_p = \sigma_{s}^2/\sigma_{n}^2$. In practice, we do not know $S_{obj}(f)$. Depending on whether we choose to use $S_{cav}(f) + S_{obj}(f)$ or $S^{(2)}_{obj}(f)$ as WFP, we can define $\sigma^{(1)}_{s}$ and $\sigma^{(2)}_{s}$ to be the respective standard deviation, yielding $\rho^{(1)}_p$ and $\rho^{(2)}_p$. These effective SNRs quantify to what extent the perturber acts as noise on our chosen WFP, but do not directly reflect the ratio of scattering strengths of object and perturber.
The $P \times N $ WFP dictionary $\mathbf{H}$ merges the $P$ WFPs (each WFP is an $N$-element vector) into a single matrix. The WFP approach can then also be framed as a multiplexing problem $Y=\mathbf{H}X+\mathcal{N}$, where $X$ is a $1\times P$ vector identifying the object position, $Y$ is the complex-valued $1\times N$ measurement vector and $\mathcal{N}$ is a $1\times N$ noise vector.
\section{Information-Theoretic Perspective}
One prerequisite for successful WFPing is the diversity of $\mathbf{H}$. In our case, the complexity of the propagation environment naturally provides this diversity. The lower the correlations between different WFPs are, the better they can be distinguished. To get a quantitative grasp of the diversity of $\mathbf{H}$, it is instructive to consider its singular value (SV) decomposition: $\mathbf{H} = \mathbf{U} \mathbf{\Sigma} \mathbf{V}^T$, where $\mathbf{\Sigma} $ is a diagonal matrix whose $i$th entry is the $i$th SV $\sigma_i$ of $\mathbf{H}$. The flatter the SV spectrum is, the more diverse is $\mathbf{H}$. A convenient metric of diversity is the effective rank of $\mathbf{H}$ which is defined as $R_{\mathrm{eff}} = \mathrm{exp}\left( -\sum_{i=1}^n \tilde{\sigma}_i \mathrm{ln}(\tilde{\sigma}_i) \right)$, where $\tilde{\sigma}_i = \sigma_i / \sum_{i=1}^n \sigma_i$ and $n=\mathrm{min}(N,P)$~\cite{roy2007effective}.
Note that only perfectly orthogonal channels with zero correlation yield $R_{\mathrm{eff}} = n$.
Unfortunately, much of the compressed sensing literature is exclusively focused on the diversity of $\mathbf{H}$ to understand the achievable performance. For instance, compression ratios are often provided without even indicating at what SNR they are valid. In principle, in the absence of any noise, the tiniest amount of diversity could be sufficient to ensure complete distinguishability even with $N=1$. Here, we argue that the achievable performance depends on the amount of (useful) information that can be extracted per measurement. In the physical layer, besides diversity the SNR is a second crucial ingredient. Moreover, high diversity and low SNR only ensure good performance if the deployed decoding method in the digital layer is capable of extracting much of the relevant encoded information from the measurement.
\begin{figure}[t]
\centering
\includegraphics[width = \columnwidth]{Fig2}
\caption{Information about the object position is (inevitably) physically encoded in the measured data via wave scattering in the irregular propagation environment. Digital data processing then seeks to retrieve the information from the measurements.}
\end{figure} \label{figIT}
WFP-based sensing in its entirety as schematically summarized in Fig.~2 can be interpreted as a process consisting of physical encoding and digital decoding of information. Wave propagation through the complex scattering environment naturally (and inevitably) encodes information about the object position in measurements of the wave field. Data processing seeks to retrieve this information. Various decoding methods exist that we will compare later on:
(i) \textit{Correlation.} Identify which row of $\mathbf{H}$ has the highest correlations with $Y$. This procedure can be interpreted as "virtual time reversal"~\cite{ing2005solid}.
(ii) \textit{Inversion.} Compute an inverse of $\mathbf{H}$, for instance, via Tikhonov regularization, and identify the entry of $\mathbf{H}^{-1}Y$ with the largest magnitude.
(iii) \textit{Optimized Inversion.} Use the result from (ii) as initial guess in a non-linear minimization of $|| Y-\mathbf{H}X ||$~\cite{redding2012using,redding2013compact}.
(iv) \textit{Learning.} Train an artificial neural network (ANN) to map $Y$ to the corresponding object position. ANN-based approaches have not been studied in the multiplexing literature to date. Besides their potential for superior decoding performance, inference is extremely fast. One forward pass through an ANN only requires a few matrix multiplications but no correlations, matrix inversions or nonlinear optimization routines.
From an information-theoretic perspective, it is important to understand fundamental bounds on the sensing performance.
A simple bound to compute is the generalized Shannon capacity
\begin{equation}
C = \sum_i \mathrm{log}_2 \left( 1+\frac{\rho}{P}\sigma_i \right)
\end{equation}\label{Sdecomp}
which has been mentioned on a few occasions in a sensing context~\cite{migliore2008electromagnetics,lorenzo2015single}. Nonetheless, the meaningfulness of $C$ for a specific sensing scheme is limited for two reasons. First, an ideal input distribution is assumed for $X$ but in reality all entries of $X$ are zero except for one which is unity. Second, an ideal decoding method is assumed. Below we will see examples where a system with nominally lower $C$ nonetheless yields a higher sensing accuracy for certain decoding methods.
It is thus essential to appreciate the sensing process in its entirety, including both encoding and decoding as illustrated in Fig.~2.
Having introduced the notion of diversity and SNR, we can now briefly comment on how faithfully the metallic enclosure in our experiment represents real-life scenarios. Without a doubt, certain cases like the inside of a vessel or a bank vault are very well represented.
Other environments like the inside of a building are less reverberant than a metallic enclosure. Essentially, the quality factor of these "cavities" is lower. This implies more correlations within a fixed frequency interval of the transmission spectrum, as well as a lower SNR due to more attenuation. Both result in a decrease of the information that can be extracted per measurement; this effect can be compensated by taking more measurements, for instance, with a wider bandwidth. Nonetheless, from a fundamental perspective, the physics of an indoor system is entirely captured by our metallic enclosure.
In scenarios with already existing wireless communication infrastructure, the beacon signals thereof could be used to implement position sensing with WFPs,
saving energy and reducing the amount of electromagnetic radiation.
\section{Semi-Analytical Simulations}
To begin with, we consider a 2D version of our experiment simulated as a 2D system of coupled dipoles~\cite{orazbayev2020label} which contains all the essential physical ingredients to simulate wave propagation, reverberation and scattering in our experiment.
These simulations offer an ideal platform to identify the effect of dynamic perturbations of the propagation environment on the sensing accuracy without any measurement noise or errors due to imperfect object positioning on the predefined positions, i.e. $\mathcal{N}(f)=0$. As shown in Fig.~3(a), a perturber of variable size with arbitrary orientation and location (within a specified area) simulates dynamic changes of the environment. Our simulation setup evaluates the transmission between an antenna pair at $25$ distinct frequencies. We use an ensemble of $150$ random perturber realizations (random orientation and random location of its center within the allowed area) to estimate $\mathbf{H}$, $R_{\mathrm{eff}}$ and $\rho_p$. The probability density function (PDF) of real and imaginary part of $S_{pert}$ is seen in Fig.~3(b-e) to be zero-mean single-peaked and tends towards a Gaussian distribution for larger perturbers.
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig3}
\caption{(a) Setup of semi-analytical coupled-dipole 2D simulations. A line-like object is placed on one of $P=5$ predefined positions in an irregularly shaped enclosure (dipole fence) of dimensions on the order of $25\times15$ wavelengths. A line-like perturbing object with variable length is randomly rotated and located such that its center lies within the indicated area. The transmission between TX and RX is evaluated. See Ref.~\cite{orazbayev2020label} for technical details on the simulation method. (b-e) PDF of real and imaginary parts of $S_{pert}$ and a Gaussian fit are shown for the smallest and largest considered perturber size.}
\end{figure} \label{figCDCsetupX}
\subsection{Impact of Perturbation on Diversity and Effective SNR}
In Fig.~4 we contrast the use of $S_{cav}(f) + S_{obj}(f)$ or $S^{(2)}_{obj}(f)$ as WFP in terms of the resulting diversity ($R_{\mathrm{eff}}$), effective SNR ($\rho_p$) and sensing capacity ($C$). As we will see below, neither of these quantities is a reliable predictor of the sensing accuracy, since they do not take the decoding method into account. For the case of using $S_{cav}(f) + S_{obj}(f)$ as WFP, the observed trend is clear: as the perturber size increases, both $R_{\mathrm{eff}}$ and $\rho_p$ as well as $C$ decrease. While the impact on $\rho_p$ was clearly expected, the reduction of diversity is more subtle. It becomes intuitive by considering the extreme case in which the perturbation alters the entire enclosure. Then, averaging over realizations yields the result that would have been obtained in an anechoic environment such that no diversity thanks to wave chaos is left.
\begin{figure}[h]
\centering
\includegraphics[width = \columnwidth]{Fig4}
\caption{(a) Effect of perturber size in the $R_{\mathrm{eff}}-\rho_p$ plane in the semi-analytical simulations. Curves for defining the WFP as $S_{cav}+S_{obj}$ or $S^{(2)}_{obj}$ are shown for three setups. All three setups are like the one in Fig.~3 but perturber area, predefined object positions and antenna positions are moved around. In all cases the objects are outside the antenna pair's line of sight. (b) Sensing capacity values corresponding to the data in (a). The results in this figure are obtained using all 25 frequency points. To ease comparison with Fig.~6, we normalized $\Sigma_{i=1}^n \sigma_i^2$ to unity; the SNR $\rho$ in Eq.~3 incorporates adverse effects on the dynamic range due to pathloss.}
\end{figure} \label{ReffrhoCDC}
Using $S_{obj}(f)$ as opposed to $S_{cav}(f) + S_{obj}(f)$ would certainly improve the diversity by removing unnecessary correlations (possibly at the expense of a better SNR such that the overall effect on capacity is unclear), but this is not possible in practice. Our closest option to that effect is to use $S^{(2)}_{obj}(f)$. Straight-forward simulations with random Gaussian matrices show that the effective rank of $S_{cav}(f) + S_{obj}(f)$ may exceed that of $S^{(2)}_{obj}(f)$ in cases where $P$ is small (preventing proper averaging over realizations of the object position) and where the ratio of the standard deviations of the distributions of $S_{obj}$ and $S_{cav}$ is large. Nonetheless, in our semi-analytical simulations, we observe in Fig.~4(a) a higher effective rank for $S^{(2)}_{obj}(f)$ than for $S_{cav}(f) + S_{obj}(f)$. Yet, since $\rho^{(2)}_p$ is substantially lower than $\rho^{(1)}_p$, the effect of using $S^{(2)}_{obj}(f)$ on the capacity is unfavorable.
Complex scattering enclosures are often seen as random field generators~\cite{corona1996reverberating}. $R_{\mathrm{eff}}$ is a measure of the number of independent samples and for $N\gg P$ one expects $R_{\mathrm{eff}} \rightarrow P$. Yet, in our simulations, $R_{\mathrm{eff}}$ saturates below $4$. This observation can be attributed to field correlations, here in the frequency domain, that prevent the field observables from being purely random variables~\cite{cozza2011skeptic}.
\subsection{Dependence of Sensing Accuracy on Perturber Size, Number of Measurements and Decoding Method}
The general trend is clear: the larger the perturbation, the less information can be extracted per measurement, as reflected by the sensing capacitance values plotted in Fig.~4(b). However, this decrease in information per measurement can be compensated with more measurements. At first sight, one may expect that such a compensation is only feasible as long as the object's scattering signature is stronger than the perturber's effect, i.e. for $\rho_p > 0\ \mathrm{dB}$. Our findings in Fig.~5, however, reveal that there is no abrupt phase change in the relation between achievable accuracy versus perturber size. Instead, using more measurements, successful position sensing is feasible at effective SNRs well below $0\ \mathrm{dB}$.
\begin{figure}[h]
\centering
\includegraphics[width = \columnwidth]{Fig5}
\caption{Localization accuracy in semi-analytical simulations. The colorscale goes from 0 (black) to 1 (white). For each choice of WFP definition (columns) and decoding method (rows), the accuracy is plotted as a function of perturber size (horizontal axis) and the number of frequency points used to ink the WFP (vertical axis). ANN results are averaged over 20 training runs with randomly initialized weights; the standard deviation is below 2 \%. The black contour line corresponds to $95\ \%$ accuracy. To aid comparison, the red contour is the same on all subfigures.}
\end{figure} \label{AccCDC}
We systematically compare the previously outlined decoding methods for both choices of WFP. For the learning-based approach, we train an artificial neural network (ANN) consisting of two fully connected layers; the first layer consists of 256 neurons and is followed by a ReLu activation, the second layer consists of $P=5$ neurons and is followed by a SoftMax activation. Using more neurons or an additional layer does not appear to notably impact the results. We consider two possibilities to provide training data from which the ANN can learn to decode the measurements. The first option is to simply use the raw data from all the perturber realizations that we generated without a need for extracting $\mathbf{H}$ or other quantities. This brute force method may prove particularly useful in cases where measurements are restricted to intensity-only information which prevents averaging as simple means to extract $\mathbf{H}$, but this scenario is outside the scope of the present paper. Note that with this approach the WFPs are never explicitly evaluated, but only implicitly contained in the ANN weights. The second option is to synthesize training data with $Y=\mathbf{H}X+\mathcal{N}$ using the estimated $\mathbf{H}$ and generating $\mathcal{N}$ with entries drawn from a Gaussian distribution whose standard deviations match those of the distribution of $S_{pert}$ extracted from the data. This second method relies on our hypothesis that $S_{pert}$ is normally distributed and offers the possibility of generating a training dataset of unlimited size. In both cases we normalize the data (zero mean, unit variance) and use the Adam method for stochastic optimization (step size $10^{-3}$) to train the ANN weights.
In Fig.~5, we show how the achieved sensing accuracy depends on the perturber size and the number of measurements. We ensure that the spacing of the utilized frequency points is always the same and that they are always centered on the same frequency. For instance, for $N=7$ measurements we pick the central frequency point out of the 25 available ones as well as its three closest neighbours to the left and right. Our results are thus for one specific system realization which explains why the contours in Fig.~5 are not perfectly smooth. Several important observations and conclusions follow from Fig.~5:
(i) WFP dictionaries with very different nominal sensing capacities can yield the same accuracy. This is the case for both ANN-based methods in which the accuracy is (almost) identical for $WFP^{(1)}$ and $WFP^{(2)}$.
(ii) The same WFP dictionary can yield very different accuracies depending on the decoding method. ANN-based decoders are seen to outperform correlation and inversion-based decoders.
(iii) The choice of WFP definition is irrelevant for the optimized inversion decoder as well as the ANN-based decoders. For correlation and inversion based decoders, however, using $WFP^{(1)}$ yields significantly better results.
(iv) Irrespective of the perturber size, we achieve an acceptable minimum accuracy (e.g. 95 \%). For larger perturbers, we need more measurements to compensate the reduction in the amount of information that can be extracted per measurement. Future information-theoretic work should seek to model the contour for a given accuracy in order to understand how the need for additional measurements scales with $\rho_p$.
(v) At low effective noise levels, some decoders achieve compression ratios above unity, that is they achieve accuracies $\geq 95\ \%$ to localize $P=5$ objects with $N<P$ measurements. For instance, the ANN (raw data) decoder with $WFP^{(2)}$ achieves 96 \% with $N=3$ at the lowest considered perturber size. However, as in any compressed sensing scenario, it is obvious that the compression ratio is heavily dependent on the noise level (here, the effective noise level due to the perturber size), the independence of different measurements (here, determined to a large extent by the interval between frequency points) and the decoding method (here, an ANN trained with raw data). Thus, a general claim of achieving a compression ratio above unity is not presented as key result of this work.
Overall, these results clearly demonstrate that it is fallacious to assume that the diversity or sensing capacity of $\mathbf{H}$ could be a reliable indicator of the sensing accuracy, hence the importance of considering the sensing process in its information-theoretic entirety as in Fig.~2.
\section{Experimental Results}
Having established an understanding of the perturber's effect under idealistic conditions in simulation, we now analyze the experimental data. Measurements with our SDR entail a few practical issues. First, there is a $\pm\pi$ uncertainty in measured phase values, originating from random phase jumps every time the Phase Locked Loop (PLL) is locked (e.g. to change the frequency). To obtain reliable data, we transform each measured complex value $z$ to $|z|\ \mathrm{exp}(2i\ \mathrm{mod}(\mathrm{arg}(z),\pi))$; the factor $2$ in the exponent ensures that the transformed variable's phase explores the entire $2\pi$ range. Second, the transmitted energy is clearly frequency-dependent, which can be caused by the frequency-dependent coupling of the monopole antennas to the cavity and/or frequency-dependent SDR components. The strong frequency-dependence means that we cannot simply model our variables as being drawn from a unique distribution, instead the distribution's standard deviation becomes frequency-dependent. To maintain the SDR's temperature constant throughout the experiment, we installed a simple CPU fan. We do not observe any significant amplitude or phase drifts over the course of the experiment.
We begin by quantifying two contributions to the $\mathcal{N}$ term in Eq.~1 that were not present in the simulations. First, we estimate the SNR due to measurement noise (by repeating the same measurement multiple times) as $\rho_1 = 25.5\ \mathrm{dB}$. Second, we estimate the SNR due to both measurement noise and imperfect positioning of the objects on the predefined locations (by repeating the same measurement multiple times after placing the object again on the same position) as $\rho_2 = 15.8\ \mathrm{dB}$.
\subsection{Impact of Perturbation on Diversity and Effective SNR}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig6}
\caption{(a-d) PDF of real and imaginary parts of $S_{pert}$ and a Gaussian fit are shown for the smallest and largest perturber size in the experiment.
(e) Effect of perturber size in the $R_{\mathrm{eff}}-\rho_p$ plane.
The blue curves only consider perturber-induced effective noise, the red curves additionally account for measurement and positioning noise.
(f) Normalized sensing capacity values corresponding to the data in (e). The results in this figure are obtained using all 51 frequency points.}
\end{figure} \label{figGNU_H_Analysis}
Based on 150 perturber realizations (random orientations) for each perturber size, in Fig.~6(a-d) we plot the PDFs of real and imaginary part of $S_{pert}$ for the smallest and largest perturber considered in our experiment. The zero-mean single-peaked distributions are identical for real and imaginary component but thinner than a Gaussian distribution. In Fig.~6(e) we plot $R_{\mathrm{eff}}(\mathbf{H})$ vs the effective SNR. Since $\mathcal{N}(f) \neq 0$ in the experiment, we plot two curves: the blue one only accounts for perturber-induced effective noise, the red one additionally accounts for measurement and positioning noise. The difference between these two curves is appreciable only for small perturber sizes since for larger perturbers $S_{pert}$ dominates over $\mathcal{N}$. Unlike in Fig.~4(a), using $S_{obj}^{(2)}$ lowers not only the effective SNR but also the effective rank. As in Fig.~4(b), we see in Fig.~6(f) that using $S_{obj}^{(2)}$ is unfavorable in terms of the (normalized) sensing capacity. The impact of $\mathcal{N}$ on $C$ is only noticeable for small perturbers.
\subsection{Dependence of Sensing Accuracy on Perturber Size, Number of Measurements and Decoding Method}
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Fig7}
\caption{Localization accuracy in experiments. The colorscale goes from 0 (black) to 1 (white). For each choice of WFP definition (columns) and decoding method (rows), the accuracy is plotted as a function of perturber size (horizontal axis) and the number of frequency points used to ink the WFP (vertical axis). ANN results are averaged over 20 training runs with randomly initialized weights; the standard deviation does not exceed 10 \% and 3 \% for ANNs trained with synthetic and raw data, respectively. The black contour line corresponds to $95\ \%$ accuracy. To aid comparison, the red contour is the same on all subfigures.}
\end{figure} \label{AccGNU}
In Fig.~7 we compare the achievable sensing accuracy in our experiment with the two considered definitions of the WFP and different decoding methods as a function of the perturber size and number of measured frequency points. The observations already made for the corresponding simulation results in Fig.~5 about the unsuitability of $R_{\mathrm{eff}}$ or $C$ to predict the sensing accuracy are confirmed once again by Fig.~7. The most notable difference to Fig.~5 is that except for the ANN trained with raw data all decoding methods fail to achieve at least 95 \% accuracy once the perturber's surface is larger than 200 $\mathrm{cm^2}$. We attribute this to the $\pm\pi$ phase uncertainty of our SDR which introduces errors in the estimation of $\mathbf{H}$. Since the ANN trained with raw data does not rely on calculating $\mathbf{H}$, it is not affected. Interestingly, we have thus a case in which it is better to feed the ANN raw data rather than to use physical insight to pre-process the ANN's training data. The ANN decoder trained with raw data is capable of achieving high sensing accuracies despite significant amounts of noise (the effective SNR is as low as -15~dB for the largest perturber, see Fig.~6(e)) and distorted data.
Using the ANN decoder trained with raw data, we achieve 100 \% sensing accuracy with $N=3$, i.e. a compression ratio of $P/N=5/3>1$, for perturbers with a surface as large as $74\ \mathrm{cm}^2$. Again, we stress that the compression ratio depends on effective SNR, measurement independence and decoding method.
\section{Conclusion and Outlook}
From a practical point of view, our experiments, in combination with an ANN-based decoder, demonstrated the feasibility of precise position sensing with WFPs in dynamically evolving scattering enclosures using a low-cost and light-weight SDR. This capability is crucial to enable situational awareness in a plethora of emerging applications. Our technique does not rely on detailed knowledge about the environment's geometry and only requires a one-off calibration phase with multiple representative realizations of the dynamic perturbations that are expected during operation.
From a conceptual point of view, our work paves the way for a thorough information-theoretic analysis of sensing with WFPs.
The dynamic perturber's unfavorable effect on diversity and effective SNR of the WFP dictionary, resulting in the acquisition of less useful information per measurement, can be fully compensated by taking more measurements -- even in the regime in which the perturber's scattering strength clearly exceeds that of the object to be localized.
We saw that the common practice in compressed sensing to only consider the diversity or capacity of $\mathbf{H}$ is insufficient to anticipate the achievable sensing accuracy.
Our results are of very general nature: they can be applied to other types of wave phenomena (sound, light, ...) and are equally valid for WFPs established with other means such as using spatial or configurational degrees of freedom by employing a sensor network or a RIS~\cite{del2018precise}.
The importance of seeing the entirety of the information-theoretic cycle points towards jointly optimizing encoding in a programmable propagation environment and ML-based decoding, as in the recently proposed "learned sensing" paradigm~\cite{del2020learned,li2020intelligent}. In contrast to compressed sensing which indiscriminately encodes all information, learned sensing seeks to encode only task-relevant information in the measurements. For position sensing, one could carefully select the frequencies at which measurements are taken (as opposed to linear spacing) and/or engineer the propagation environment with a RIS~\cite{del2020optimal}.
Looking ahead, it appears interesting to extend the present work (i) to scenarios with multiple objects to be localized, where neglected inter-object scattering is an additional effective source of noise~\cite{del2018precise}, (ii) to deeply sub-wavelength position sensing~\cite{cohen2011subwavelength}, and (iii) to more complex tasks like image transmission~\cite{li2018deep}.
In this work, the perturber was seen as an obstacle for our task to localize an object. In other contexts, the objective may be to characterize size and motion of a perturber.
Diffuse wave spectroscopy~\cite{FishCounting,fishcountingPRL,CCSgeo,conti2006characterization} analyzes changes of the broadband impulse response over time to estimate the number or scattering cross section of objects moving through a complex medium.
Our work has evidenced that the perturber's scattering strength can also be clearly related to the capacity of a multiplexing channel matrix averaged over different realizations of the perturber's position. Considering configuration-to-configuration multiplexing with two dynamic metasurface transceivers~\cite{sleasman2019implementation} may thus enable similar characterizations of a moving perturber with single-port single-frequency measurements~\cite{del2018dynamic}.
\section*{Acknowledgment}
The author acknowledges Michael del Hougne for help with the experimental work. The code for the semi-analytical model is based on Ref.~\cite{orazbayev2020label}. The code for controlling the SDR is based on Ref.~\cite{Lime}.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
| 26d308c70b53bfa412e01f87352423d76635bfd7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
adc83b19e793491b1c6ea0fd8b46cd9f32e592fc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
|
\section{Introduction}
For positive integers $k_1,\ldots,k_r$ with $k_r \ge 2$, the multiple zeta values (MZVs) are defined by
\begin{align*}
\zeta(k_1,\ldots, k_r)
:=\sum_{1\le n_1<\cdots <n_r} \frac {1}{n_1^{k_1}\cdots n_r^{k_r}}.
\end{align*}
We say that an index $(k_1,\ldots,k_r)\in\mathbb{Z}_{\ge1}^r$ is admissible if $k_r\ge2$.
For an index $\boldsymbol{k}=(k_1,\ldots,k_r)$,
$\text{wt}(\boldsymbol{k}):=k_1+\cdots +k_r$ and $\text{dep}(\boldsymbol{k}):= r$ are called
weight and depth of $\boldsymbol{k}$, respectively.
It is known that MZVs satisfy many algebraic relations over $\mathbb{Q}$.
Ohno's relation is a well-known relation among MZVs.
\begin{defn}
For an admissible index
\[
\boldsymbol{k}:=(\underbrace{1,\ldots,1}_{a_1-1},b_1+1,\dots,\underbrace{1,\ldots,1}_{a_d-1},b_d+1) \quad (a_i, b_i\ge1),
\]
we define the dual index of $\boldsymbol{k}$ by
\[
\boldsymbol{k}^\dagger :=(\underbrace{1,\ldots,1}_{b_d-1},a_d+1,\dots,\underbrace{1,\ldots,1}_{b_1-1},a_1+1).
\]
\end{defn}
\begin{thm}[Ohno's relation; Ohno \cite{Oho99}] \label{ohno}
For an admissible index $(k_1,\ldots,k_r)$ and $m\in\mathbb{Z}_{\ge 0}$, we have
\begin{align*}
\sum_{\substack{ e_1+\cdots+e_r=m \\ e_i\ge0\,(1\le i\le r) }}
\zeta (k_1+e_1,\ldots,k_r+e_r)
=\sum_{\substack{ e'_1+\cdots+e'_{r'}=m \\ e'_i\ge0\,(1\le i\le r') }}
\zeta (k'_1+e'_1,\ldots,k'_{r'}+e'_{r'}),
\end{align*}
where the index $(k'_1,\ldots,k'_{r'})$ is the dual index of $(k_1,\ldots,k_r)$.
\end{thm}
For an admissible index $\boldsymbol{k}=(k_1,\ldots,k_r)$ and $s\in\mathbb{C}$ with $\Re(s)>-1$, Hirose-Murahara-Onozuka \cite{HMO20} defined the Ohno function $I_{\boldsymbol{k}}(s)$ by
\begin{align}\label{ohnofunc}
I_{\boldsymbol{k}}(s):=\sum_{i=1}^{r}\sum_{0<n_1<\cdots<n_r} \frac{1}{n_1^{k_1}\cdots n_r^{k_r}}
\cdot \frac{1}{n_i^{s}} \prod_{j\ne i} \frac{n_j}{ n_j-n_i }.
\end{align}
This is a sum of special cases of $\zeta_{\mathfrak{sl}(r+1)}(\boldsymbol{s})$ which is called the Witten multiple zeta function associated with $\mathfrak{sl}(r+1)$.
The Witten multiple zeta function was first introduced in Matsumoto-Tsumura \cite{MaTs06}, which is also called the zeta function associated with the root system of type $A_r$ (for more details, see Komori-Matsumoto-Tsumura \cite{KMT10}), and continued meromorphically to the whole complex space $\mathbb{C}^{r(r+1)/2}$.
Hence $I_{\boldsymbol{k}}(s)$ can be continued meromorphically to $\mathbb{C}$.
When $s=m\in\mathbb{Z}_{\ge 0}$, the Ohno function is the Ohno sum, that is,
\begin{align*}
I_{\boldsymbol{k}}(m)=\sum_{\substack{ e_1+\cdots+e_r=m \\ e_i\ge0\,(1\le i\le r) }}\zeta (k_1+e_1,\ldots,k_r+e_r),
\end{align*}
and by Theorem \ref{ohno}, we have
\begin{align*}
I_{\boldsymbol{k}}(m)=I_{\boldsymbol{k}^\dagger}(m).
\end{align*}
In \cite{HMO20}, Hirose-Murahara-Onozuka gave an interpolation of Ohno's relation to complex function.
\begin{thm}[an interpolation of Ohno's relation] \label{interpolation}
For an admissible index $\boldsymbol{k}$ and $s\in\mathbb{C}$, we have
\begin{align*}
I_{\boldsymbol{k}}(s)=I_{\boldsymbol{k}^\dagger}(s).
\end{align*}
\end{thm}
In this paper, we study the Ohno function $I_{\boldsymbol{k}}(s)$. In Section 2, we give a precise region of absolute convergence of the series \eqref{ohnofunc}.
\begin{thm}\label{main1}
The series \eqref{ohnofunc} converges absolutely only for
\begin{align*}
\max_{1\leq j\leq r}\{r-2j+2-(k_j+\cdots+k_r)\}<\Re(s).
\end{align*}
\end{thm}
In Section 3, we give the following integral expression of the Ohno function;
\begin{thm}\label{thm:integral_exp}
For an admissible index
\[
\boldsymbol{k}:=(\underbrace{1,\ldots,1}_{a_1-1},b_1+1,\dots,\underbrace{1,\ldots,1}_{a_d-1},b_d+1) \quad (a_i, b_i\ge1)
\]
and $s\in\mathbb{C}$ with $\Re(s)>-1$, we have
\begin{align*}
I_{\boldsymbol{k}}(s) &=\dfrac{1}{(a_1-1)!(b_1-1)! \cdots (a_d-1)! (b_d-1)! \Gamma (s+1) }\\
&\times \int_{0<t_1<\cdots < t_{2d}<1} \dfrac{dt_1 \cdots dt_{2d}}{(1-t_1)t_2 \cdots (1-t_{2d-1})t_{2d} }\\
&\times \left( \log \dfrac{1-t_1}{1-t_2} \right)^{a_1-1}
\left( \log \dfrac{t_3}{t_2} \right)^{b_1-1}
\cdots
\left( \log \dfrac{1-t_{2d-1}}{1-t_{2d}} \right)^{a_{d}-1}
\left( \log \dfrac{1}{t_{2d}} \right)^{b_{d}-1}
\left( \log \dfrac{t_2 \cdots t_{2d} }{t_1 \cdots t_{2d-1}} \right)^{s}.
\end{align*}
\end{thm}
In Section 4, we give a new proof of Theorem \ref{interpolation}. In the original proof, we assume Ohno's relation, but our new proof gives Theorem \ref{interpolation} directly. Our method is based on Theorem \ref{thm:integral_exp} and the proof given by Ulanskii \cite{U}.
In Section 5, we consider an interpolation of $T$-interpolated sum formula.
Finally, in Section 6, we show another expression of the Ohno function.
\begin{thm}\label{main6}
For an admissible index $\boldsymbol{k}=(k_1,\ldots,k_r)$ and $s\in\mathbb{C}$ with $\max_{1\leq j\leq r}\{r-2j+2-(k_j+\cdots+k_r)\}<\Re(s)<0$, we have
\begin{align*}
I_{\boldsymbol{k}}(s)=&-\frac{\sin(\pi s)}{\pi}\sum_{0<n_1<\cdots<n_r}\frac{1}{n_1^{k_1-1}\cdots n_r^{k_r-1}}\int_0^\infty\frac{w^{-s-1}}{(w+n_1)\cdots(w+n_r)}dw.
\end{align*}
\end{thm}
By applying this theorem, we can deduce linear relations among Ohno functions.
\begin{thm}\label{main7}
Let $l$ be a positive integer with $l \le r$ and an admissible index $\boldsymbol{k}=(k_1,\ldots,k_r)$ satisfy
$$\max_{\substack{1\leq j\leq r\\|\boldsymbol{e}|=m\\e_1,\ldots,e_{r}\leq1\\e_l=0}}\{r-2j+2-(k_j+e_j+\cdots+k_r+e_r)\}<-m$$
for all $m=0,\ldots,r-1$.
For $s\in\mathbb{C}$, we have
\begin{align*}
\sum_{j=0}^{r-1}(-1)^{j}\sum_{\substack{|\boldsymbol{e}|=j\\e_1,\ldots,e_{r}\leq1\\e_l=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s-j)
=\zeta(k_1,\ldots,k_{l-1},k_l+s,k_{l+1},\ldots,k_r).
\end{align*}
\end{thm}
By theorem \ref{main6}, we can deduce the following corollary.
\begin{cor}\label{cor}
For an admissible index $\boldsymbol{k}=(k_1,\ldots,k_r)$ and a negative integer $n$ with $\max_{1\leq j\leq r}\{r-2j+2-(k_j+\cdots+k_r)\}<n<0$, we have
\begin{align*}
I_{\boldsymbol{k}}(n)=0.
\end{align*}
\end{cor}
\section{Region of absolute convergence}
We give a precise region of absolute convergence of the series \eqref{ohnofunc}.
It is enough to consider a region of absolute convergence of the Dirichlet series
\begin{align*
\begin{split}
I_{\boldsymbol{k},i}(s):=&\sum_{0<n_1<\cdots<n_r} \frac{1}{n_1^{k_1}\cdots n_r^{k_r}}
\cdot \frac{1}{n_i^{s}} \prod_{j\ne i} \frac{n_j}{ n_j-n_i }\\
=& (-1)^{i-1} \sum_{m_1,\ldots,m_r=1}^\infty
\frac{1}{ m_1^{k_1-1}(m_1+m_2)^{k_2-1}\cdots(m_1+\cdots+m_r)^{k_r-1} } \\
&\,\,\,\,\qquad \times \frac{1}{ (m_1+\cdots+m_i)^{s+1} }
\, \prod_{j<i} \frac{1}{m_{j+1}+\cdots+m_i }
\, \prod_{j>i} \frac{1}{m_{i+1}+\cdots+m_j },
\end{split}
\end{align*}
for all $1\le i\le r$.
Since the above series is a special case of the zeta function associated with the root system of type $A_r$ for each $i$,
we can apply the result given by Zhao-Zhou \cite[Proposition 2.1]{ZhZh11}.
\begin{thm}[Zhao-Zhou \cite{ZhZh11}] \label{zhzh}
The series
\begin{align}\label{ZZ}
\sum_{m_1,\ldots,m_r=1}^\infty \prod_{\boldsymbol{i}\subseteq[r]}\left(\sum_{j=1}^{{\rm lg}(\boldsymbol{i})}m_{i_j}\right)^{-\sigma_{\boldsymbol{i}}}
\end{align}
converges if and only if for all $\ell=1,\ldots,r$ and $\boldsymbol{i}=(i_1,\ldots,i_\ell)\subseteq[r]$
\begin{align}\label{ZZ1}
\sum_{\substack{\boldsymbol{j}{\rm \;contains\;at\;least\;one\;of}\\ i_1,\ldots,i_\ell}}\sigma_{\boldsymbol{j}}>\ell,
\end{align}
where the product $\prod_{\boldsymbol{i}\subseteq[r]}$ runs over all nonempty subsets of $[r]=(1,2,\ldots,r)$ as a poset and ${\rm lg}(\boldsymbol{i})$ is the length of $\boldsymbol{i}$.
\end{thm}
Note that if we put
\begin{align}\label{sigma}
\sigma_{\boldsymbol{i}}=
\begin{cases}
k_j-1&(\boldsymbol{i}=(1,2,\ldots,j)\mbox{ for } j\ne i),\\
k_i+s&(\boldsymbol{i}=(1,2,\ldots,i)),\\
1&(\boldsymbol{i}=(j+1,\ldots,i)\mbox{ for } j< i\mbox{ or }\boldsymbol{i}=(i+1\ldots,j)\mbox{ for } j> i),\\
0&(\mbox{otherwise}),
\end{cases}
\end{align}
then the series \eqref{ZZ} is $(-1)^{i-1}I_{\boldsymbol{k},i}(s)$.
\begin{proof}[Proof of Theorem \ref{main1}]
In the case $\ell=r$ for Theorem \ref{zhzh}, since $\boldsymbol{i}=(1,\ldots,r)$, the left-hand side of \eqref{ZZ1} is the sum of all $\sigma_{\boldsymbol{i}}$'s in \eqref{sigma}, so we obtain the following condition of absolute convergence
\begin{align}\label{cond1}
k_1+\cdots +k_r+\Re(s)>r.
\end{align}
When $\ell=r-1$, let us consider the case $\boldsymbol{i}=(2,\ldots,r)$ first. In this case, the left-hand side of \eqref{ZZ1} is the sum of all $\sigma_{\boldsymbol{i}}$'s but $\sigma_{(1)}$ in \eqref{sigma}, so we obtain the following conditions for $I_{\boldsymbol{k},i}(s)$:
\begin{align*}
\begin{cases}
k_2+\cdots +k_r+\Re(s)>r-2 &(i\neq1,\ \boldsymbol{i}=(2,\ldots,r)),\\
k_2+\cdots +k_r>r-1 &(i=1,\ \boldsymbol{i}=(2,\ldots,r)).
\end{cases}
\end{align*}
In the case $\boldsymbol{i}=(1,\ldots,i-1,i+1,\ldots,r)$ for $i\neq1$ or $\boldsymbol{i}=(1,\ldots,i,i+2,\ldots,r)$, the left-hand side of \eqref{ZZ1} is the sum of all $\sigma_{\boldsymbol{i}}$'s but $\sigma_{(i)}$ or $\sigma_{(i+1)}$ in \eqref{sigma}, respectively, so we obtain the following condition:
\begin{align*}
k_1+\cdots +k_r+\Re(s)>r \quad(i\neq1,\ \boldsymbol{i}=(1,\ldots,i-1,i+1,\ldots,r))\ {\rm or}\ (\boldsymbol{i}=(1,\ldots,i,i+2,\ldots,r)).
\end{align*}
Otherwise, we obtain the following condition:
\begin{align*}
k_1+\cdots +k_r+\Re(s)>r-1\quad(\rm{otherwise}).
\end{align*}
The conditions obtained by the case $\boldsymbol{i}\neq(2,\ldots,r)$ are contained in \eqref{cond1}, hence only the case $\boldsymbol{i}=(2,\ldots,r)$ deduces the region of absolute convergence when $\ell=r-1$.
(In general, it suffices to consider the case $\boldsymbol{i}=(j,j+1,\ldots,r)$ when $\ell=r-j+1$.)
By the similar way, considering all $\ell=1,\ldots,r-2$ for Theorem \ref{zhzh}, the series $I_{\boldsymbol{k},i}(s)$ converges absolutely only for
\begin{align*}
&k_1+\cdots +k_r+\Re(s)>r,\\
&k_2+\cdots +k_r+\Re(s)>r-2,\\
&\cdots\\
&k_i+\cdots+k_r+\Re(s)>r-2i+2
\end{align*}
and
\begin{align*}
&k_{i+1}+\cdots +k_r>r-i,\\
&k_{i+2}+\cdots +k_r>r-i-1,\\
&\cdots\\
&k_r>1.
\end{align*}
Hence, the series \eqref{ohnofunc} is absolutely convergent for
\begin{align*}
&k_1+\cdots +k_r+\Re(s)>r,\\
&k_2+\cdots +k_r+\Re(s)>r-2,\\
&\cdots\\
&k_r+\Re(s)>-r+2
\end{align*}
and
\begin{align*}
&k_{2}+\cdots +k_r>r-1,\\
&k_{3}+\cdots +k_r>r-2,\\
&\cdots\\
&k_r>1.
\end{align*}
Since the index $\boldsymbol{k}$ is admissible, we obtain Theorem \ref{main1}.
\end{proof}
\section{New integral expression}
We need the following lemma to prove Theorem \ref{thm:integral_exp}.
\begin{lem}\label{lamme:integral}
For $r\in \mathbb{Z}_{\ge 1}$, $c_i>0$ $(1\le i \le r)$ with $c_i\neq c_j$ $(i\neq j)$ and $s\in\mathbb{C}$ with $\Re (s) >-1$,
we have \begin{equation}\label{eq:int_lemma}
\begin{split}
& \sum_{i=1}^r \left( \dfrac{1}{c_i^{s+1}} \prod_{j\neq i} \dfrac{1}{c_j-c_i} \right) \\
&= \dfrac{1}{\Gamma(s+1)} \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+\cdots +x_r)^{s} \, dx_1 \cdots dx_r .
\end{split} \end{equation}
\end{lem}
\begin{proof}
First we prove the integral in the right-hand side of \eqref{eq:int_lemma}
converges for $\Re(s)>-1$.
Let $\sigma:=\Re(s)$.
When $-1<\sigma<0$, we have
\begin{align*}
&\left| \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+\cdots +x_r)^{s} \, dx_1 \cdots dx_r \right|\\
&\le \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+\cdots +x_r)^{\sigma} \, dx_1 \cdots dx_r \\
&\le \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} x_r^{\sigma} \, dx_1 \cdots dx_r \\
&\le \dfrac{1}{c_1\cdots c_{r-1}} \int_0^{\infty}
e^{-c_r x_r} x_r^{\sigma} \, dx_r \\
&= \dfrac{1}{c_1\cdots c_{r-1}} \dfrac{1}{c_r^{\sigma+1}}
\Gamma(\sigma +1).
\end{align*}
Hence the integral in the right-hand side of \eqref{eq:int_lemma} converges.
When $\sigma \ge 0$,
we have
\begin{align*}
&\left| \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+\cdots +x_r)^{s} \, dx_1 \cdots dx_r \right|\\
&\le \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+1)^{\sigma}\cdots (x_r+1)^{\sigma} \, dx_1 \cdots dx_r \\
&= e^{c_1} \cdots e^{c_r} \int_{1}^{\infty} \cdots \int_{1}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} x_1^{\sigma}\cdots x_r^{\sigma} \, dx_1 \cdots dx_r \\
&\le e^{c_1} \cdots e^{c_r} \dfrac{1}{c_1^{\sigma+1}\cdots c_r^{\sigma+1}}
\Gamma(\sigma+1)^r.
\end{align*}
Hence the integral in the right-hand side of \eqref{eq:int_lemma}
also converges in this case.
Next we prove the equation \eqref{eq:int_lemma} by induction on $r$.
When $r=1$, the right-hand side of \eqref{eq:int_lemma} equals
\begin{align*}
\dfrac{1}{\Gamma(s+1)}\int_0^{\infty}e^{-c_1x} x^s\, dx
= \dfrac{1}{c_1^{s+1}}\end{align*}
and the equation \eqref{eq:int_lemma} holds.
Suppose $r\ge 2$ and
\eqref{eq:int_lemma} holds for $r-1$ variables.
By the change of variables
\begin{align}
\begin{cases}\label{CoV}
x_1+\cdots +x_r=y_1, \\
x_2+\cdots +x_r=y_2, \\
\vdots \\
x_r=y_r,
\end{cases}
\end{align}
we have
\begin{align*}
& \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_r x_r} (x_1+\cdots +x_r)^{s} \, dx_1 \cdots dx_r \\
&=
\int \cdots \int_{y_1>\cdots >y_r\ge0}
e^{-c_1 y_1} e^{(c_1-c_2)y_2} \cdots e^{(c_{r-1}-c_r)y_r} y_1^{s} \, dy_1 \cdots dy_r \\
&=
\int \cdots \int_{y_1>\cdots >y_{r-1}>0}
e^{-c_1 y_1} e^{(c_1-c_2)y_2} \cdots e^{(c_{r-2}-c_{r-1})y_{r-1}} \dfrac{e^{(c_{r-1}-c_r)y_{r-1}}-1}{c_{r-1}-c_r} y_1^{s} \, dy_1 \cdots dy_{r-1} \\
&=
\int \cdots \int_{y_1>\cdots >y_{r-1}>0}
e^{-c_1 y_1} e^{(c_1-c_2)y_2} \cdots e^{(c_{r-3}-c_{r-2})y_{r-2}} \dfrac{e^{(c_{r-2}-c_r)y_{r-1}} - e^{(c_{r-2}-c_{r-1})y_{r-1}}}{c_{r-1}-c_r}
y_1^{s} \, dy_1 \cdots dy_{r-1}.
\end{align*}
By the change of variables \eqref{CoV} with $x_r=y_r=0$, we have
\begin{align*}
&\int \cdots \int_{y_1>\cdots >y_{r-1}>0}
e^{-c_1 y_1} e^{(c_1-c_2)y_2} \cdots e^{(c_{r-3}-c_{r-2})y_{r-2}} \dfrac{e^{(c_{r-2}-c_r)y_{r-1}} - e^{(c_{r-2}-c_{r-1})y_{r-1}}}{c_{r-1}-c_r}
y_1^{s} \, dy_1 \cdots dy_{r-1}\\
&=\dfrac{1}{c_{r-1}-c_r} \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_{r-2} x_{r-2}-c_r x_{r-1}} (x_1+\cdots +x_{r-1})^{s} \, dx_1 \cdots dx_{r-1} \\
&\quad-\dfrac{1}{c_{r-1}-c_r} \int_{0}^{\infty} \cdots \int_{0}^{\infty}
e^{-c_1 x_1- \cdots -c_{r-2} x_{r-2}-c_{r-1} x_{r-1}} (x_1+\cdots +x_{r-1})^{s} \, dx_1 \cdots dx_{r-1} .
\end{align*}
By the induction hypothesis, this equals
\begin{align*}
&\dfrac{\Gamma(s+1)}{c_{r-1}-c_r}
\Bigg( \sum_{i=1}^{r-2} \left( \frac{1}{c_i^{s+1}} \frac{1}{c_r-c_i} \prod_{ \substack{j\neq i \\ 1\le j \le r-2} } \frac{1}{c_j-c_i} \right)
+ \frac{1}{c_r^{s+1}} \prod_{ \substack{ 1\le j \le r-2} } \frac{1}{c_j-c_r} \\
&- \sum_{i=1}^{r-2} \left( \frac{1}{c_i^{s+1}} \frac{1}{c_{r-1}-c_i} \prod_{ \substack{j\neq i \\ 1\le j \le r-2} } \frac{1}{c_j-c_i} \right)
- \frac{1}{c_{r-1}^{s+1}} \prod_{ \substack{ 1\le j \le r-2} } \frac{1}{c_j-c_{r-1}}
\Bigg) \\
&= \Gamma(s+1) \left( \sum_{i=1}^{r-2} \frac{1}{c_i^{s+1}} \prod_{ \substack{j\neq i \\ 1\le j \le r} } \frac{1}{c_j-c_i}
+ \frac{1}{c_r^{s+1}} \prod_{ \substack{j\neq r \\ 1\le j \le r-1} } \frac{1}{c_j-c_r}
+ \frac{1}{c_{r-1}^{s+1}} \prod_{ \substack{j\neq r-1 \\ 1\le j \le r} } \frac{1}{c_j-c_{r-1}}
\right) \\
&= \Gamma(s+1) \sum_{i=1}^{r} \frac{1}{c_i^{s+1}} \prod_{ \substack{j\neq i \\ 1\le j \le r} } \frac{1}{c_j-c_i}
\end{align*}
and this proves that \eqref{eq:int_lemma} holds for $r$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:integral_exp}]
We set
\[
\boldsymbol{k}=(\underbrace{1,\ldots,1}_{a_1-1},b_1+1,\dots,\underbrace{1,\ldots,1}_{a_d-1},b_d+1) = (k_1,\ldots , k_r).
\]
Note that $a_1+\cdots +a_d=r$.
By definition, we have
\begin{align*}
I_{\boldsymbol{k}}(s)=
\sum_{0<n_1< \cdots <n_r}
\dfrac{1}{n_1^{k_1-1} n_2^{k_2-1} \cdots n_r^{k_r-1} }
\sum_{i=1}^r \left( \dfrac{1}{n_i^{s+1}} \prod_{ j\neq i }\dfrac{1}{n_j- n_i } \right).
\end{align*}
Because
$(k_1-1, \ldots ,k_r-1)=
( \underbrace{0,\ldots ,0}_{a_1-1}, b_1, \ldots , \underbrace{0,\ldots ,0}_{a_d-1}, b_d )$,
we have
\begin{align*}
I_{\boldsymbol{k}}(s)=
\sum_{0<n_1< \cdots <n_r}
\dfrac{1}{n_{a_1}^{b_1} n_{a_1+a_2}^{b_2} \cdots n_{a_1+\cdots + a_d}^{b_d} }
\sum_{i=1}^r \left( \dfrac{1}{n_i^{s+1}} \prod_{ j\neq i }\dfrac{1}{n_j- n_i } \right).
\end{align*}
By using the well-known identity
\[ \displaystyle \dfrac{1}{n^b} = \dfrac{1}{\Gamma(b)} \int_0^{\infty} e^{-ny} y^{b-1}dy\ \ \ (n, \, b>0)\]
and Lemma \ref{lamme:integral}, we get
\begin{align*}
I_{\boldsymbol{k}}(s)
&= \dfrac{1}{\Gamma(s+1) (b_1-1)! \cdots (b_d-1)!} \\
&\sum_{0<n_1< \cdots <n_r}
\int_0^{\infty} \cdots \int_0^{\infty} e^{-n_{a_1}y_1} e^{-n_{a_1+a_2}y_2}\cdots e^{-n_{a_1+\cdots +a_d}y_d}
y_1^{b_1-1} \cdots y_d^{b_d-1}\,dy_1\cdots dy_d\\
&\hspace{110pt}e^{-n_1x_1} e^{-n_2x_2}\cdots e^{-n_{r}x_r}(x_1+\cdots +x_r)^s \,dx_1\cdots dx_r.
\end{align*}
Set $n_i= m_1+\cdots +m_i$ $(1\le i \le r)$.
Then each $m_i$ runs over all positive integers and
\begin{align*}
I_{\boldsymbol{k}}(s)
=& \dfrac{1}{\Gamma(s+1) (b_1-1)! \cdots (b_d-1)!} \\
&\sum_{m_1\ge 1, \ldots , m_r\ge 1}
\int_0^{\infty} \cdots \int_0^{\infty}
\exp\left(-\sum_{i=1}^{r}m_i(X_i+Y_i)\right)\\
&\hspace{40pt} (x_1+\cdots +x_r)^s y_1^{b_1-1} \cdots y_d^{b_d-1}
\,dx_1\cdots dx_r\,dy_1\cdots dy_d\\
=& \dfrac{1}{\Gamma(s+1) (b_1-1)! \cdots (b_d-1)!} \\
&\int_0^{\infty} \cdots \int_0^{\infty}
\prod_{i=1}^r \dfrac{e^{-X_i-Y_i}}{1-e^{-X_i-Y_i}}\\
&\hspace{40pt} (x_1+\cdots +x_r)^sy_1^{b_1-1} \cdots y_d^{b_d-1}
\,dx_1\cdots dx_r\,dy_1\cdots dy_d,
\end{align*}
where
\begin{align*}
X_i&=x_i+\cdots+x_r\ \ (1\le i \le r),\\
Y_i&=\begin{cases}
y_1+\cdots +y_d &(1\le i \le a_1),\\
y_2+\cdots +y_d &(a_1< i \le a_1+a_2),\\
\vdots\\
y_d &(a_1+\cdots +a_{d-1}< i \le r).
\end{cases}
\end{align*}
We apply the following change of variables:
\[ \begin{cases}
t^{(1)}_{1}&= \exp(-X_1-Y_{a_1}),\\
\vdots &\hspace{20pt}\vdots \\
t^{(1)}_{a_1}&= \exp(-X_{a_1}-Y_{a_1}),\\
u_1&= \exp(-X_{a_1+1}-Y_{a_1}),\end{cases}\ \ \ \ \
\begin{cases}
t^{(2)}_{1}&= \exp(-X_{a_1+1}-Y_{a_1+a_2}),\\
\vdots &\hspace{20pt}\vdots \\
t^{(2)}_{a_2}&= \exp(-X_{a_1+a_2}-Y_{a_1+a_2}),\\
u_2&= \exp(-X_{a_1+a_2+1}-Y_{a_1+a_2}),\end{cases}\]
\[ \hspace{20pt} \cdots
\begin{cases}
t^{(d)}_{1}&= \exp(-X_{a_1+\cdots +a_{d-1}+1}-Y_r),\\
\vdots &\hspace{20pt}\vdots \\
t^{(d)}_{a_d}&= \exp(-X_{r}-Y_r),\\
u_d&= \exp(-Y_r).\end{cases}\]
Then it can be easily checked that
\begin{itemize}
\item $0<t^{(1)}_{1}<\cdots <t^{(1)}_{a_1}<u_1 < \cdots \cdots < t^{(d)}_{1} <\cdots < t^{(d)}_{a_d} < u_d<1$,
\item $\displaystyle \prod_{i=1}^r \dfrac{e^{-X_i-Y_i}}{1-e^{-X_i-Y_i}}
=\dfrac{t_1^{(1)}}{1-t_1^{(1)}} \cdots
\dfrac{t_{a_d}^{(d)}}{1-t_{a_d}^{(d)}}$,
\item $y_1=\log \dfrac{t_1^{(2)}}{u_1},\ \ \ldots ,\ \
y_{d-1}=\log \dfrac{t_1^{(d)}}{u_{d-1}},\ \
y_d=\log \dfrac{1}{u_d}$,
\item $x_1+\cdots +x_r= \log \dfrac{u_1}{t_1^{(1)}}
\dfrac{u_2}{t_1^{(2)}} \cdots \dfrac{u_d}{t_1^{(d)}}$,
\item $ dx_1 \cdots dx_r dy_1\cdots dy_d = \dfrac{1}{t^{(1)}_{1}\cdots t^{(d)}_{a_d} u_1 \cdots u_d} dt^{(1)}_{1}\cdots dt^{(d)}_{a_d} du_1 \cdots du_d$.
\end{itemize}
Therefore
\begin{align*}
I_{\boldsymbol{k}}(s)
=& \dfrac{1}{\Gamma(s+1) (b_1-1)! \cdots (b_d-1)!}
\int_{0<t^{(1)}_{1}<\cdots <t^{(1)}_{a_1}<u_1 < \cdots < t^{(d)}_{1} <\cdots < t^{(d)}_{a_d} < u_d<1}\\
&\dfrac{dt^{(1)}_{1}\cdots dt^{(d)}_{a_d} du_1 \cdots du_d}{ (1-t^{(1)}_{1}) \cdots (1-t^{(1)}_{a_1}) u_1 \cdots (1-t^{(d)}_{1}) \cdots
(1-t^{(1)}_{a_d})u_d }\\
&\left( \log \dfrac{t^{(2)}_{1}}{u_1} \right)^{b_1-1} \cdots \left( \log \dfrac{t^{(d)}_{1}}{u_{d-1}} \right)^{b_{d-1}-1}
\left( \log \dfrac{1}{u_{d}} \right)^{b_{d}-1}
\left( \log \dfrac{u_1}{t^{(1)}_{1}} \dfrac{u_2}{t^{(2)}_{1}} \cdots \dfrac{u_d}{t^{(d)}_{1}} \right)^{s}
\\
=& \dfrac{1}{\Gamma(s+1) (a_1-1)! \cdots (a_d-1)! (b_1-1)! \cdots (b_d-1)!}
\int_{0<t_1<u_1 < \cdots < t_{d}< u_d<1}\\
&\dfrac{dt_1\cdots dt_d du_1\cdots du_d}{ (1-t_{1}) u_1 \cdots (1-t_d) u_d }\\
& \left( \log \dfrac{1-t_{1}}{1-u_1} \right)^{a_1-1} \cdots \left( \log \dfrac{1-t_{d}}{1-u_d} \right)^{a_d-1} \\
&\left( \log \dfrac{t_{2}}{u_1} \right)^{b_1-1} \cdots \left( \log \dfrac{t_{d}}{u_{d-1}} \right)^{b_{d-1}-1}
\left( \log \dfrac{1}{u_{d}} \right)^{b_{d}-1}
\left( \log \dfrac{u_1}{t_{1}} \dfrac{u_2}{t_{2}} \cdots \dfrac{u_d}{t_{d}} \right)^{s}.
\end{align*}
Here $t^{(i)}_1$ is replaced by $t_i$ in the last equation.
This completes the proof.
\end{proof}
\section{New Proof of Theorem \ref{interpolation}}
In this section, we give another proof of $I_{\boldsymbol{k}}(s) =I_{\boldsymbol{k^{\dagger}}}(s)$
by using Theorem \ref{thm:integral_exp}.
Let $d$ be a positive integer. We consider the following change of variables:
\begin{align*}
\dfrac{1-t_{2\ell-1}}{1-t_{2\ell}} = \dfrac{u_{2(d-\ell+1)+1}} {u_{2(d-\ell+1)}}, \ \ \
\dfrac{t_{2\ell}}{t_{2\ell+1}} = \dfrac{1-u_{2(d-\ell+1)}} {1-u_{2(d-\ell+1)-1}}\ \ \ \ \ (1\le \ell \le d).
\end{align*}
Here we set $u_{2d+1}=t_{2d+1}=1$.
\begin{rem}
The change of variables above is a special case of that of Ulanskii \cite{U}.
By using this change of variables, he gave a direct proof of Ohno's relation for MZVs.
\end{rem}
This change of variables satisfies the following properties (cf. \cite[Sect.~2]{U}):
\begin{enumerate}
\item the region $0<t_1 < \cdots <t_{2d}<1$ corresponds to
the region $0<u_1 < \cdots <u_{2d}<1$,
\item $\dfrac{dt_1 \cdots dt_{2d}}{(1-t_1)t_2 \cdots (1-t_{2d-1})t_{2d} } =
\dfrac{du_1 \cdots du_{2d}}{(1-u_1)u_2 \cdots (1-u_{2d-1})u_{2d} }$,
\item $\dfrac{t_2}{t_1} \cdots \dfrac{t_{2d}}{t_{2d-1}} = \dfrac{u_2}{u_1} \cdots \dfrac{u_{2d}}{u_{2d-1}}$.
\end{enumerate}
By this change of variables, we have
\begin{align*}
&\int_{0<t_1<\cdots < t_{2d}<1} \dfrac{dt_1 \cdots dt_{2d}}{(1-t_1)t_2 \cdots (1-t_{2d-1})t_{2d} }\\
&\times \left( \log \dfrac{1-t_1}{1-t_2} \right)^{a_1-1}
\left( \log \dfrac{t_3}{t_2} \right)^{b_1-1}
\cdots
\left( \log \dfrac{1-t_{2d-1}}{1-t_{2d}} \right)^{a_{d}-1}
\left( \log \dfrac{1}{t_{2d}} \right)^{b_{d}-1}
\left( \log \dfrac{t_2 \cdots t_{2d} }{t_1 \cdots t_{2d-1}} \right)^{s}
\\
=& \int_{0<u_1<\cdots < u_{2d}<1} \dfrac{du_1 \cdots du_{2d}}{(1-u_1)u_2 \cdots (1-u_{2d-1})u_{2d} }\\
&\hspace{20pt}\times
\left( \log \dfrac{1}{u_{2d}} \right)^{a_1-1}
\left( \log \dfrac{1-u_{2d-1}}{1-u_{2d}} \right)^{b_1-1}
\cdots
\left( \log \dfrac{u_3}{u_2} \right)^{a_{d}-1}
\left( \log \dfrac{1-u_1}{1-u_2} \right)^{b_{d}-1}
\left( \log \dfrac{u_2 \cdots u_{2d} }{u_1 \cdots u_{2d-1}}\ \right)^{s}.
\end{align*}
By multiplying the gamma factors, we have
$I_{\boldsymbol{k}}(s) =I_{\boldsymbol{k^{\dagger}}}(s)$.
\section{$T$-interpolation of Ohno function}
For an admissible index $\boldsymbol{k}=(k_1,\ldots ,k_r)$,
Yamamoto \cite{Y} introduced an
interpolated multiple zeta value $\zeta^T(\boldsymbol{k})$ as
\begin{align*}
\zeta^T(\boldsymbol{k}) =
\sum_{\boldsymbol{p}} T^{r-\text{dep}(\boldsymbol{p})}
\zeta(\boldsymbol{p}) \ \in \mathbb{R}[T].
\end{align*}
Here the sum runs over all indices $\boldsymbol{p}$
such that
\[\boldsymbol{p} =(k_1 \square k_2 \square \cdots \square k_r), \]
where each $\square$ is filled by the comma $,$ or the plus $+$.
This polynomial in $T$ interpolates two kinds of multiple zeta values,
i.e.,
$\zeta^{0}(\boldsymbol{k}) =\zeta(\boldsymbol{k})$ and
$\zeta^{1}(\boldsymbol{k}) =\zeta^{\star}(\boldsymbol{k})$,
where $\zeta^{\star}(\boldsymbol{k})$ is the multiple zeta-star value
defined by
\begin{align*}
\zeta^{\star}(k_1,\ldots, k_r)
:=\sum_{1\le n_1\le \cdots \le n_r} \frac {1}{n_1^{k_1}\cdots n_r^{k_r}}.
\end{align*}
Yamamoto proved the following sum formula for $\zeta^T(\boldsymbol{k})$,
which is an interpolation of the sum formulas
for multiple zeta and zeta-star values.
\begin{thm}[{\cite[Theorem 1.1]{Y}}]\label{thm:t-sum}
For integers $m$, $a \ge 1$, we have
\begin{align*}
\sum_{\substack{ {\rm wt}(\boldsymbol{k})=m+a+1 \\ \boldsymbol{k}:{\,\rm admissible},\,{\rm dep}(\boldsymbol{k})=a }} \zeta^T(\boldsymbol{k})
&= \sum_{j=0}^{a-1} \binom{m+a}{j} T^j (1-T)^{a-1-j} \zeta(m+a+1) \\
&= \sum_{i=0}^{a-1} \binom{m+i}{i} T^i \zeta(m+a+1).
\end{align*}
\end{thm}
For $a\in \mathbb{Z}_{\ge 1}$,
let $\boldsymbol{a} = ( \underbrace{1,\ldots ,1}_{a-1}, 2)$.
We define an interpolated version of $I_{\boldsymbol{a}}(s)$ as
\begin{align*}
I^T_{\boldsymbol{a}}(s) &:=\dfrac{1}{(a-1)! \Gamma (s+1) }\\
&\times \int_{0<t_1<t_{2}<1} \dfrac{dt_1 dt_{2}}{(1-t_1)t_2 } \left( \log \dfrac{1-t_1}{1-t_2} + T \log \dfrac{t_2}{t_1} \right)^{a-1}
\left( \log \dfrac{t_2 }{t_1 } \right)^{s}\ \ \ (\Re(s)>-1). \end{align*}
When $T=0$, we have $I^0_{\boldsymbol{a}}(s) = I_{\boldsymbol{a}}(s)$.
When $s=m \in \mathbb{Z}_{\ge 0}$, the value
$I^T_{\boldsymbol{a}}(m)$ is the sum of all interpolated multiple zeta values
for fixed weight and depth:
\[ I^T_{\boldsymbol{a}}(m) =
\sum_{|\boldsymbol{e}|=m} \zeta^{T}(\boldsymbol{a}\oplus \boldsymbol{e})
=\sum_{\substack{ {\rm wt}(\boldsymbol{k})=m+a+1 \\ \boldsymbol{k}: \text{admissible}, \text{dep} (\boldsymbol{k})=a }} \zeta^T(\boldsymbol{k}).
\]
We can give the following formula,
which is an interpolation of
the sum formula for $T$-interpolated multiple zeta values.
In fact, this theorem deduces Theorem \ref{thm:t-sum} by setting $s=m$.
Since the proof is same as that in the last section, we omit it.
\begin{thm}
For $s\in \mathbb{C}$ and $a\in\mathbb{Z}_{\ge 1}$, we have
\begin{align*}
I^T_{\boldsymbol{a}}(s)=
\left( \sum_{i=0}^{a-1} \binom{s+i}{i} T^i \right) \zeta(s+a+1).
\end{align*}
\end{thm}
\section{New Relations}
We first prove Theorem \ref{main6}.
\begin{proof}[Proof of Theorem \ref{main6}]
By the partial fraction decomposition, we have
\begin{align*}
\frac{1}{(w+n_1)\cdots(w+n_r)}=\sum_{i=1}^r \frac{1}{w+n_i}\cdot\prod_{j\ne i} \frac{1}{ n_j-n_i }.
\end{align*}
Let $B(x,y)$ be the beta function. Since
$$
\int_0^\infty\frac{w^{-s}}{w+n}dw=n^{-s}B(s,1-s)=n^{-s}\frac{\pi}{\sin(\pi s)}
$$
for $0<\Re(s)<1$, we have
\begin{align*}
\int_0^\infty\frac{w^{-s-1}}{(w+n_1)\cdots(w+n_r)}dw=-\frac{\pi}{\sin(\pi s)}\sum_{i=1}^r \frac{1}{n_i^{s+1}}\prod_{j\ne i} \frac{1}{ n_j-n_i }.
\end{align*}
Therefore, the statement of theorem holds for $-1<\Re(s)<0$.
For $-r<\Re(s)<0$ and $0\le i\le r$, we have
\begin{align*}
\int_{n_i}^{n_{i+1}}\frac{w^{-\sigma-1}}{(w+n_1)\cdots(w+n_r)}dw&\le \frac{1}{n_{i+1}n_{i+2}\cdots n_r}\int_{n_i}^{n_{i+1}}w^{-\sigma-i-1}dw\\
&\ll \begin{cases}
(n_{i+1}n_{i+2}\cdots n_r)^{-1}n_{i+1}^{-\sigma-i}&(-\sigma>i),\\
(n_{i+1}n_{i+2}\cdots n_r)^{-1}\log n_{i+1}&(-\sigma=i),\\
(n_{i+1}n_{i+2}\cdots n_r)^{-1}n_{i}^{-\sigma-i}&(-\sigma<i),
\end{cases}\\
&\ll \begin{cases}
(n_{i+1}n_{i+2}\cdots n_r)^{-1}n_{i+1}^{-\sigma-i}\log (n_{i+1}+1)&(-\sigma\ge i),\\
(n_{i+1}n_{i+2}\cdots n_r)^{-1}n_{i}^{-\sigma-i}&(-\sigma<i),
\end{cases}
\end{align*}
where $n_0=0$ and $n_{r+1}=\infty$. Hence we can estimate
\begin{align}\label{bunkatsu}
\begin{split}
&\sum_{0<n_1<\cdots<n_r}\frac{1}{n_1^{k_1-1}\cdots n_r^{k_r-1}}\int_{0}^{\infty}\frac{w^{-\sigma-1}}{(w+n_1)\cdots(w+n_r)}dw\\
&\ll
\sum_{0\le i\le-\sigma}\sum_{0<n_1<\cdots<n_r}(n_1^{-k_1+1}\cdots n_i^{-k_i+1})n_{i+1}^{-k_{i+1}-\sigma-i}(n_{i+2}^{-k_{i+2}}\cdots n_r^{-k_r})\log (n_{i+1}+1)\\
&\quad+\sum_{-\sigma<i\le r}\sum_{0<n_1<\cdots<n_r}(n_1^{-k_1+1}\cdots n_{i-1}^{-k_{i-1}+1})n_{i}^{-k_{i}-\sigma-i+1}(n_{i+1}^{-k_{i+1}}\cdots n_r^{-k_r}).
\end{split}
\end{align}
Note that by Matsumoto \cite{Mat02}, the infinite series $\sum_{0< n_1<\cdots <n_r} n_1^{-\sigma_1}\cdots n_r^{-\sigma_r}$ converges for
\begin{align*}
&\sigma_{1}+\cdots +\sigma_r>r,\\
&\sigma_{2}+\cdots +\sigma_r>r-1,\\
&\cdots\\
&\sigma_r>1.
\end{align*}
Thus the first series in the right-hand side of \eqref{bunkatsu} converges when
\begin{align*}
&k_{1}+\cdots +k_r+\sigma>r,\\
&k_{2}+\cdots +k_r+\sigma>r-2,\\
&\cdots\\
&k_{i+1}+\cdots +k_r+\sigma>r-2i
\end{align*}
for $0\le i\le-\sigma$.
These conditions can be simply written as
\begin{align*}
\max_{1\leq j\leq 1-\sigma}\{r-2j+2-(k_j+\cdots+k_r)\}<\sigma.
\end{align*}
Similarly the second series of the right-hand side in \eqref{bunkatsu} converges when
\begin{align*}
\max_{1\leq j\leq r}\{r-2j+2-(k_j+\cdots+k_r)\}<\sigma.
\end{align*}
Hence by the identity theorem, we obtain the theorem.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main7}]
Let $E_j$ be the elementary symmetric polynomial of degree $j$ in\\
$(n_1,\ldots,n_{\ell-1},n_{\ell+1},\ldots,n_{r})$. Then we have
\begin{align}\label{symmetric}
\begin{split}
&-\frac{\sin(\pi s)}{\pi}\sum_{0<n_1<\cdots<n_r}\frac{n_\ell}{n_1^{k_1}\cdots n_r^{k_r}}\int_0^\infty\frac{w^{-s-1}(w+n_1)\cdots(w+n_{\ell-1})(w+n_{\ell+1})\cdots(w+n_{r})}{(w+n_1)\cdots(w+n_r)}dw\\
&=-\frac{\sin(\pi s)}{\pi}\sum_{0<n_1<\cdots<n_r}\frac{n_\ell}{n_1^{k_1}\cdots n_r^{k_r}}\int_0^\infty\frac{w^{-s-1}(w^{r-1}+E_1w^{r-2}+\cdots+E_{r-1})}{(w+n_1)\cdots(w+n_r)}dw.
\end{split}
\end{align}
The left-hand side in \eqref{symmetric} is $\zeta(k_1,\ldots,k_{\ell-1},k_\ell+s,k_{\ell+1},\ldots,k_r)$ for $-1<\sigma<0$. On the other hand, the right-hand side in \eqref{symmetric} is
\begin{align*}
&(-1)^{r-1}\sum_{\substack{|\boldsymbol{e}|=r-1\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s-r+1)+(-1)^{r-2}\sum_{\substack{|\boldsymbol{e}|=r-2\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s-r+2)\\
&+(-1)^{r-3}\sum_{\substack{|\boldsymbol{e}|=r-3\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s-r+3)\\
&+\cdots\\
&+\sum_{\substack{|\boldsymbol{e}|=0\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s)\\
=&\sum_{m=0}^{r-1}(-1)^{m}\sum_{\substack{|\boldsymbol{e}|=m\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}I_{\boldsymbol{k}+\boldsymbol{e}}(s-m).
\end{align*}
for
\begin{align}\label{last}
\max_{\substack{1\leq j\leq r\\|\boldsymbol{e}|=m\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}\{r-2j+2-(k_j+e_j+\cdots+k_r+e_r)\}<\Re(s-m)<0\quad(m=0,\ldots,r-1).
\end{align}
If $\boldsymbol{k}$ satisfies
$$\max_{\substack{1\leq j\leq r\\|\boldsymbol{e}|=m\\e_1,\ldots,e_{r}\leq1\\e_\ell=0}}\{r-2j+2-(k_j+e_j+\cdots+k_r+e_r)\}<-m$$
for all $m=0,\ldots,r-1$, the inequality \eqref{last} holds for $-1<\sigma<0$. Hence the theorem is valid for $-1<\sigma<0$. By meromorphic continuation, we obtain the result.
\end{proof}
\section*{Acknowledgements}
This work was supported by JSPS KAKENHI Grant Numbers JP20K03523 and JP19K14511.
| 8808111fa2d37ee3fa84b1f44d76efe8b5eee62b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Technology-based intimate partner surveillance (IPS) causes immense harms. A discrete form of intimate partner violence (IPV), IPS is the deliberate surveillance of an intimate partner with or without their knowledge, levied through
technical and non-technical methods. Survivors have reported their abusers use
spyware apps,
account compromise, GPS trackers, shared cellular plans, and more to monitor their digital lives and physical
locations~\cite{freed2017digital,freed2018stalker,matthews2017stories,woodlock2017abuse,southworth2007intimate}.
Prior work has also indicated that a wealth of IPS apps are
available online~\cite{chatterjee2018spyware} and in active use against
victims~\cite{roundy2020many,havron2019clinical,freed2019clinical}.
To better protect targets of abuse, we need to both improve technologies'
robustness to abuse and better inform \mbox{interventional} approaches that directly
aid victims~\cite{havron2019clinical,freed2019clinical}. To achieve these aspirations, however, we need to better understand how those interested
in perpetrating IPS learn to conduct these attacks. To date, there have been few investigations into how attackers locate the resources that help them enact abuse. Chatterjee et al.~\cite{chatterjee2018spyware} highlight that blogs, videos, and question-and-answer sites exist online that help facilitate IPS, but stop short of
investigating the communities who make use of them. There is a
methodological hurdle in discovering this information: we need a way to hear from potential attackers directly.
In this work, we provide the first study exploring how potential attackers use the Internet to learn how to enact IPS
against their victims.
We identify a set of five public, online forums where people discuss infidelity in intimate relationships and tools for monitoring cellphones. We build a crawler to
retrieve the conversations on these forums and use it to
compile a dataset containing over 200\,K posts spread across almost 20\,K threads.
This dataset contains an unprecedented amount of information about the strategies of IPS attackers, contextualized in user-generated natural language. While prior work has described the attacks experienced by victims, we present a detailed view of how these attacks are created and developed---the capabilities attackers seek, the vulnerabilities they exploit, and the community dynamics that enable them.
We analyze this data using mixed-methods. We begin with quantitative measurements of the forums, their users, and their posting behaviors.
These analyses reveal that most forums contain ``superusers'' involved with a
disproportionately large fraction of threads. For two of the five forums, we discover that
their content consists almost entirely of spam advertising particular spyware
tools, in a manner that may be consistent with search engine optimization (SEO) techniques deployed by the creators of those tools. We also
determine that many of their most highly viewed threads are about
technology-based IPS, suggesting these forums have generated a significant audience for IPS-related content.
We then perform qualitative coding of a large sample of 750 threads (250 from each of the three forums that are not \mbox{inundated} by spyware ads).
Via thematic analysis \cite{braun2006using}, we
\mbox{surface} novel insights about the online behaviors of IPS attackers. We show that potential attackers seek online support for suspicions of infidelity, and that community members respond by outlining exactly how to track, monitor, and otherwise compromise the privacy of an intimate partner. We show that discussion of IPS is prevalent in these forums, with one
forum having 78\% of sampled threads related to IPS.
We develop a taxonomy of IPS attacks surfaced from the suggestions made in these forums (Table~\ref{tab:taxonomy}). \textit{Tool-based attacks} directly weaponize technology such
as audiovisual recorders, keyloggers, backup recovery tools, and more, and can be understood in two subcategories: those requiring physical access to a partner, and those that do not.
\textit{Coercion and subterfuge} attacks manipulate a partner into unlocking their
devices or accounts. Finally, we see many suggestions to \textit{outsource
attacks} by hiring private investigators.
Although
some of these strategies have been reported by victims \cite{freed2018stalker, matthews2017stories}, our analysis
provides the complementary view of potential perpetrators.
We highlight tools, tactics, and services that have not
been reported previously, and which we believe were previously unknown to
those helping victims. We also report on the conversational patterns within these forums that enable would-be attackers (what we describe as \textit{escalation}), and, conversely, patterns in which they are discouraged away from IPS (\textit{de-escalation}). These findings suggest that public forums can serve as a source of threat intelligence informing interventions to dissuade abuse.
In fact, our work is already having impact for IPS interventions. We shared our results with the team running a clinic providing direct assistance to IPV survivors facing IPS \cite{havron2019clinical, freed2019clinical},\footnote{https://www.ipvtechresearch.org} who are working towards using our findings in their procedures. More broadly, our analyses yield insights for technologists, platforms, and advocates on how we might defend against and mitigate the spread of IPS. We close by discussing implications for platforms, future work in automated threat intelligence, and how policymakers might address the for-profit operations
that financially benefit from abuse.
\section{Background and Related Work}
\paragraph{IPV and technology abuse.} Prior work has examined the behaviors, justifications, and tactics of intimate partner abusers \cite{hearn_violences_1998, kelly_naming_2016, stark_coercive_2009}, including work identifying suspicions of infidelity as a leading trigger for IPV in heterosexual couples \cite{arnocky2015anticipated, nemeth2012sexual}. Of this literature, a growing
body of work explores the role of technology in IPV, including how abusers
exploit technology to monitor, harass, control or otherwise harm their
targets~\cite{freed2017digital,freed2018stalker,chatterjee2018spyware,matthews2017stories,dimond2011domestic,woodlock2017abuse,southworth2007intimate}. Chatterjee et al.~\cite{chatterjee2018spyware} observed that abusers are likely exploiting easy-to-find online resources, including
tutorials, question-and-answer sites, and videos explaining how
to use spyware for IPS. Roundy et al.~\cite{roundy2020many} used
datasets from a major antivirus vendor to explore a broader class of creepware
that includes spyware, but also SMS bombers, hacking tutorials, and more. These works have provided valuable intelligence for corresponding anti-IPS interventions with victims and survivors \cite{freed2019clinical, havron2019clinical}.
To date, however, less research has examined the role of online communities in IPV. Some have examined how targets experience IPV in digital media and seek support through online forums \cite{dragiewicz2018technology, leitao2019technology}, but to the best of our knowledge, ours is the first study to measure and analyze how forums lead attackers to such tools. Our work confirms that attackers are discussing and recommending IPS strategies on public forums available to any Internet user. We also identify new tactics, such as custom scripts to monitor websites visited and launch man-in-the-middle attacks.
\paragraph{Online measurement studies.} Prior work has used
measurement and analysis of online forums to shed light on communities
discussing criminal or otherwise malicious behaviors. For some of these communities, this research has led to the development of threat intelligence.
Commercially motivated criminals, such as spammers and black-hat hackers,
use online forums as marketplaces and for learning adversarial techniques from
each other~\cite{thomas2006underground,motoyama2011analysis,franklin2007inquiry}. Research on this phenomenon has identified structure, trust, and superusers in these communities~\cite{afroz2013honor, afroz2014doppelganger, garg2015fc}. Relatedly, online forums used by pedophiles and others involved in the creation and
distribution of child sexual abuse
materials have been studied to gain insights into the way participants justify
or normalize abuse, and share technical and non-technical strategies that facilitate
abuse~\cite{durkin1999propagandizing}. Similar methods have also
been used to analyze forums associated with
hate groups~\cite{chatzakou2017hate,ribeiro2018characterizing}, cyberbullying~\cite{kowalski2012cyberbullying,chatzakou2017measuring}, doxxing~\cite{snyder2017fifteen},
mis- and disinformation campaigns~\cite{starbird2019disinformation}, harassment~\cite{mariconti2018you, hua20twitter, hua20towardstwitter}, and sex
trafficking~\cite{portnoff2017backpage}. These measurement studies include ones that directly document abuse
(e.g., hate and harassment on social media), as well as
those investigating perpetrators' discussions of tactics (e.g., perpetrators sharing ways to maximize the emotional impact of harassing messages).
Our work falls in the latter category. While methodologically similar, we
differ in our focus on people who use online forums to discuss strategies for IPS. Similar to the work on cybercrime and child abuse forums, analysis of online resources such as craigslist and backpage has led to threat intelligence that helped combat human trafficking~\cite{portnoff2017backpage}. We aim to have a similar impact on IPS and IPV more broadly.
In summary, our research questions are:
\begin{newitemize}
\item What role do online forums play in surfacing IPS resources to potential attackers?
\item What role do commercially motivated entities play in these online communities?
\item What tools and tactics are being suggested to potential attackers, and at what levels of technical sophistication?
\end{newitemize}
\section{Forums and Datasets}
\label{sec:methods}
To answer these research questions,
we perform a mixed-methods analysis of a large sample of posts and threads from
public online forums with a high density of IPS-relevant conversation. In this section, we review our analysis targets and data
collection approach, as well as the resulting datasets.
\paragraph{Infidelity and IPS forums.} We identified several forums whose
content includes a large number of posts touching on IPS. These were discovered
through a combination of pointers from prior work~\cite{chatterjee2018spyware}
and online web searches using a combination of terms such as ``spyware track wife''.
While we endeavored to be exhaustive, we restricted attention only to publicly
available forums, excluding forums accessible only to registered users or users who had crossed some threshold number of active posts. We may also
have missed forums that are not easily found via search engines.
Finally, many forums have a small number of posts touching on IPS, but with the overwhelming majority of content being irrelevant to our study. We excluded those from our analysis and selected forums that seemed to have a higher concentration of IPS-related posts. Future work might explore techniques to discover IPS-related forums that are harder to access and find.
Our analyses focus primarily on three forums that aim to help people navigate infidelity. To prevent the publication of this work from advertising spyware, or from unintentionally impacting these public forums by convincing them to go private, we anonymize them.\footnote{Our data, including forum names, is available for research upon request.} The forums we study include:
\begin{newitemize}
\item \textbf{Forum~A\xspace}, a community dedicated to discussing ``\textit{investigative equipment}.'' Forum~A\xspace is a subforum of a website providing resources on resolving marital conflicts with Alexa rank approximately 500,000.
\item \textbf{Forum~B\xspace}, a community dedicated to advice on ``\textit{detecting infidelity and deception.}'' Forum~B\xspace is a subforum of a website providing resources on cheating in romantic relationships, with Alexa rank approximately 900,000.
\item \textbf{Forum~C\xspace}, a moderated Reddit subforum that bills itself as ``\textit{a safe place to ask for advice and guidance}'' for those facing infidelity in a relationship. As of February 2020, Forum~C\xspace\ had approximately 80,000 subscribers, and Reddit was the 18th most popular website in the world.\footnote{\url{https://www.alexa.com/siteinfo/reddit.com}. Alexa rankings are based on global traffic and engagement over the last 90 days.}
\end{newitemize}
We additionally investigated two subforums that focus on spyware tools: \textbf{Forums D\xspace} and \textbf{E\xspace}, both subforums of a community for cellphone advice. Forum~D\xspace\ focuses on spyware for mobile phones, while Forum~E\xspace\ focuses on spyware generally. These subforums surfaced in Internet searches for the same sets of IPS-related keywords as those used to discover the three infidelity forums above. As we discuss further in Section~\ref{sec:data-desc}, our analysis concludes
most content on these forums are spam advertisements for particular spyware tools.
For simplicity, we will use the term ``\textit{forum}'' to refer to the communities we studied in-depth (e.g., Forum~C\xspace), and ``\textit{parent forum}'' to refer to their parents where needed (e.g., Reddit).
\paragraph{Data collection.} We collected data from Forums A\xspace, B\xspace, D\xspace and E\xspace via custom crawlers built using Scrapy, an
open-source, Python-based framework for web
scraping.\footnote{\url{https://docs.scrapy.org/en/latest/}} Our crawlers preserved the threaded structure of each forum's content, as well as metadata like views and post and update timestamps where available. We did not download any media beyond text---specifically avoiding images---and stored all data in a local database on a secured server.
Our analysis covers a set of scrapes collected using this pipeline in October of 2019. For Forum~C\xspace, we used the scrape available via the Reddit corpus within ConvoKit \cite{chang2019convokit}, which was collected in October 2018.
Table ~\ref{tab:dataset} summarizes the complete dataset.
\paragraph{Limitations.}
Our study combines quantitative and qualitative methodologies to characterize a sampling of publicly available forums where discussion of IPS tactics manifests. We emphasize our work may not generalize to discussion on private forums, such as those that require the creation of an account and a threshold of posts or months active for access, or those occurring within private social groups on larger social media platforms such as Facebook and Twitter. We also focus on English-language forums, and thus our findings cannot be taken to represent the scope of IPS discussion worldwide.
We believe there is compelling future research in investigating larger public-facing communities, like other subreddits or closed communities on and off of influential social networks.
\paragraph{Ethics.} Throughout this work, we were sensitive to the ethics of
using online discussions of highly personal topics for research. Our data is from publicly available fora accessible on the Internet
without authentication.
Our IRB office reviewed the study and deemed it to not be human-subjects research, since all data was already in the public domain. Still, we took precautions to ensure our research remained safe and privacy-preserving. The only identifiers we used were the public usernames
associated with each post. We did not pursue identification of
people from their posts or usernames, or collect or store images. In reporting our work, we have scrubbed
information that might trace back to the people behind the pseudonyms, e.g.,
locations or specific narrative details.
\section{Forum Activity and Users}
\label{sec:data-desc}
\begin{table*}[ht]
\centering
\footnotesize
\begin{tabular}{@{}lrrrrr@{}}
\hline
& \multicolumn{1}{c}{\textbf{Forum~A\xspace}} &
\multicolumn{1}{c}{\textbf{Forum~B\xspace}} &
\multicolumn{1}{c}{\textbf{Forum~C\xspace}} & \multicolumn{1}{c}{\textbf{Forum~D\xspace}}
& \multicolumn{1}{c}{\textbf{Forum~E\xspace}} \\
\hline
Date of first thread & Jan 2006 & Aug 2005 & May 2013 & Oct 2008 & Feb 2013 \\
Size of forum (threads) & 268 & 1,175 & 11,291 & 3,388 & 2,788 \\
Size of forum (posts) & 1,608 & 8,932 & 183,381 & 7,540 & 4,952 \\
Unique active users in forum & 462 & 2,102 & 12,740 & 264 & 543 \\
\hline
Avg. thread views (stdev) & 3,438 (13,249) & 4,822 (12,194) & -- & 1,685 (7,634) & 6,315 (44,813) \\
Avg. thread length in posts (stdev) & 7 (17) & 4 (8) & 16 (17) & 2 (1) & 2 (2) \\
Avg. time to new thread (stdev) & 140 days (198 days) & 7 days (13 days) & 3 days (13 days) & 1 day (11 days) & 21 hrs (11 days) \\
Avg. time to new post (stdev) & 3 days (13 days) & 14 hrs (2 days) & 15 minutes (2 hrs) & 12 hrs (5 days) & 12 hrs (2 days) \\
\hline
IPS-relevant \% of threads$^\alpha$ & 78 & 51 & 18 & -- & --\\
Size of IPS-relevant sample (posts)$^\alpha$ & 1,411 & 2,011 & 1,032 & -- & -- \\
Unique users active in IPS-relevant threads$^\alpha$ & 296 & 465 & 346 & -- & -- \\
\% of IPS-relevant threads that escalate$^\alpha$ & 32 & 38 & 35 & -- & --\\
\hline
\end{tabular}
\caption{Comparison of the five forums in our dataset. Forum~C\xspace does not provide viewership
information (marked by dashes). $^\alpha$\,Calculated
via qualitative analysis of random samples of 250 threads per non-spam forum, see Section~\ref{sec:qual-methods}.}
\label{tab:dataset}
\end{table*}
We begin by measuring the nature of activity on these
forums. Later on (Sections~\ref{sec:seeker-experience},~\ref{sec:taxonomy}), we use qualitative methods to more deeply characterize their content.
\paragraph{Forum activity and viewership.} The forums in our dataset varied in their rates of activity and reported viewership.
In total, the forums
contain 18,937 threads, with Forum~C\xspace\
containing the most threads and posts (Table~\ref{tab:dataset}).
Activity data in terms of thread and post times was available for all forums. We note that activity on Forums A\xspace and B\xspace peaked between 2010 and 2015, and has significantly dropped off in the last five years, while activity on Forum~C\xspace has exploded in that time (Figure~\ref{fig:forums-over-time}). We hypothesize this may represent a shift away from niche forums focused on infidelity and towards niche subcommunities of larger social media platforms like Reddit. Despite the recent drop-offs, these forums remain publicly available resources for would-be abusers, and contain IPS tools and tactics that are still relevant today; thus we included them in our qualitative analysis (Section \ref{sec:qual-methods}).
While these three forums exhibit similar seasonal and diurnal patterns, temporal patterns for Forums E\xspace\ and D\xspace\ exhibit greater variability, as well as strong peaks in year-over-year posting activity in 2013 and 2014, respectively. As we will discuss subsequently, this reflects concentrated activity by advertisers posting spam marketing spyware products.
Across all forums, the total number of views was approximately 30\,M. This is likely a significant underestimate of total viewership given that it does not include Reddit's Forum~C\xspace, for which we do not have viewership data. Within each forum, the distribution of views per thread was dominated by one or two highly viewed threads (usually `sticky'
threads compiling forum rules or shared resources) and then a long tail of
less-viewed threads. The distributions of thread lengths for each forum followed similar long-tail patterns, with an average thread length of six posts.
\paragraph{Forum users and ``superusers''.} Table~\ref{tab:dataset} shows the number of users in each forum, identified by comparing the usernames attached to each post via case-insensitive string equality. Forums differed in the number of unique users, from 264 in Forum~D\xspace\ to 12,740 in Forum~C\xspace, but all forums have ``superusers'' that account for a disproportionate number of
posts. Figure~\ref{fig:superusers} gives (left chart) a CDF of the fraction of each forum's
posts made by users
and (right chart) a histogram of the fractions of all threads to which a user
posted. For clarity, we only show the 50 and 25 most prolific users,
respectively.
\begin{figure*}[t]
\centering
\includegraphics{figures/posts_yearly.pdf}
\caption{Histograms (normalized to maximum bin value in forum) for
\textbf{(left)} postings
per year, with shading indicating the years for which we have post data for
the forum, \textbf{(middle)} postings per month of year, and
\textbf{(right)} postings per hour of day. }
\label{fig:forums-over-time}
\end{figure*}
\begin{figure*}[t]
\centering
\hpagess{.70}{.25}{
\input{figures/posts_cdf}
}{
\footnotesize
\begin{tabular}[c]{clr}
\hline
\textbf{User} & \textbf{Forums} \\
\hline
1 & A\xspace, B\xspace & \\
2 & A\xspace, B\xspace &\\
3 & A\xspace, D\xspace, E\xspace&\\
4 & B\xspace, D\xspace&\\
5 & B\xspace, E\xspace&\\
6 & B\xspace, D\xspace, E\xspace&\\
7 & B\xspace, C\xspace&\\
8 & C\xspace, D\xspace&\\
\hline
\end{tabular}
}
\caption{\textbf{(Left)} Cumulative fraction of posts per user for the top
50 users. \textbf{(Middle)} Fraction of threads with posts from the top 25
users. Forum~E\xspace\ (resp.~Forum~D\xspace) are dominated by 1 (resp.~2) superusers, while the other three forums show a more even spread of posts and threads among users.
\textbf{(Right)} Multiple-forum users and the forums in which they posted. Excludes the 17
users found posting in Forums D\xspace\ and E\xspace, which share a parent forum.}
\label{fig:superusers}
\label{tab:mult-forum-users}
\end{figure*}
Forums E\xspace and D\xspace are clear outliers compared to the other forums; this is due to spammers, as we discuss below. While the other three forums also have superusers, they do not dominate their forums to the same degree. Additionally, cursory examination shows they are not spammers. Some are human and robot moderators, including an automatic moderator on Forum~C\xspace that posts the subreddit's rules as the first response to each thread-starting post. But most superusers appear to be humans particularly engaged in the forum, driving the culture and activity of the community with their posts.
We additionally checked whether posters were active in multiple forums in our data. Comparing usernames via case-insensitive string equality, we found just eight users recurring across forums that had no structural reason to be connected. Of these, only one user made contributions that exceeded 1\% of posts or threads in any forum. While this finding seems to indicate superusers are not cross-posting across multiple forums, we note it is simple to register accounts with different usernames in these forums. We consider the identification of users across forums to be an area of future work.
\paragraph{Spyware spam and SEO inflation in Forums D\xspace\ and E\xspace.} As mentioned above, Forums
D\xspace and E\xspace stood out along many dimensions. Most content in these two forums can be attributed to
a handful of users: notably, the top user in Forum~D\xspace\ contributed to 95\%
of threads and authored 45\% of posts, and the second-most-active user
contributed to 95\% of threads and authored 44\% of posts. Forum~E\xspace exhibits a similar pattern of dominance by a handful of users.
Inspection shows many of the threads in Forum~D\xspace constitute conversations between its top two users: one posts a spam advertisement for a spyware tool, and another follows up with a short response. We conclude this demonstrates a strategy of search engine optimization (SEO) employed by the company behind the spyware tool to boost the forum's visibility on Internet searches and attract attention to their spyware product. Specifically, 94\% of the posts made by the top user were the same message: an advertisement for the spyware tool. This user also authored nearly half (45\%) of the posts on this forum. Forum~D\xspace's
second-most-prolific user bears a username closely
associated with that spyware tool, and posts either ads for the tool or
short, meaningless messages (\textit{``Hi,''} \textit{``Hello''}) in response to the first user's ads. We found similar patterns within Forum~E\xspace. Having concluded that most activity on these two forums was spam intended to inflate SEO for specific spyware products, we excluded them from our qualitative analysis of forum content (Sections~\ref{sec:seeker-experience},~\ref{sec:taxonomy}).
\paragraph{Prevalence of IPS-related keywords.} To efficiently understand the organic content in the three infidelity forums, we sought automated ways to identify only those threads that were relevant to IPS. As a
first-cut assessment, we performed keyword-based
searches of the threads in our dataset using a small set of keywords identified from prior work \cite{chatterjee2018spyware}: ``\textit{spy}'',
``\textit{monitor}'', ``\textit{track}'', ``\textit{hack}'' and ``\textit{record}'' (Table~\ref{tab:keyword-relevance}).
This first-cut assessment showed keyword searches are, unsurprisingly,
insufficient for accurate discovery of relevant threads: for example, the keyword `\textit{record}' may be used in the context of recording someone without their consent, but also in the context of music recordings. To quantify this, we assembled a human-labeled dataset of 750 threads sampled across each of the three non-spam forums and manually coded for relevance to IPS (see Section~\ref{sec:qual-methods} for detailed methods).
We then applied a regex-based labeling method that flagged threads as relevant if any post within the thread contained any one of keywords in our seed set.
Using our 750 human-labeled threads
as ground truth, this simple approach achieves an AUC of
0.62, indicating it misses a large number of relevant threads (false negatives) and contains a
large number of irrelevant threads (false positives).
As a result, we do not rely on the regex-based approach for any of our subsequent analyses, but instead study the posts human-labeled as ground truth. The development of automated learning techniques that can efficiently flag IPS-relevant threads remains a tantalizing area of future work.
\begin{table}[t]
\centering
\footnotesize
\begin{tabular}{lrrrrr}
\hline
\textbf{Keywords} & \textbf{A\xspace} & \textbf{B\xspace} & \textbf{C\xspace} & \textbf{D\xspace} & \textbf{E\xspace} \\ \hline
spy & 26.9 & 7.1 & 1.6 & 98.4 & 42.1 \\
monitor & 8.6 & 2.2 & 3.0 & 97.8 & 27.8 \\
track & 13.4 & 5.4 & 8.8 & 25.7 & 30.0 \\
hack & 3.7 & 1.1 & 2.3 & 1.1 & 4.1 \\
record & 14.9 & 7.8 & 7.3 & 3.8 & 1.4 \\
spy, monitor, track, hack, record & 43.7 & 17.3 & 17.6 & 99.4 & 62.3 \\ \hline
\end{tabular}
\caption{Percentage of threads within each forum containing one or more of the indicated keywords.}
\label{tab:keyword-relevance}
\end{table}
\section{Understanding Forum Content}
\label{sec:qual-methods}
Our data contain rich information on attackers' strategies, interactions, and stated goals embedded in the natural language of users' posts. Here, we describe our qualitative methods for analyzing the content within Forums A\xspace, B\xspace\ \& C\xspace.
\paragraph{Establishing human ratings for IPS relevance.} Our initial measurements showed not all content on these forums is relevant to the discussion of IPS tactics: for example, while 8 of the top 10 threads by viewership on Forum~B\xspace\ contained some mention of ways to monitor an intimate partner, the other two threads discussed contraception and women's underwear. Thus, to focus our analysis, we first established human ratings for whether or not a given thread was relevant to IPS.
We began by randomly choosing 30 threads, 10 from each forum. Three coders independently rated whether each thread was IPS-relevant.
We stipulated that a relevant thread should both (1) discuss an intent to track, monitor, surveil, or otherwise compromise an intimate partner's privacy; and (2) describe doing so via technology. Inter-rater reliability showed agreement in 28/30 threads (Fleiss' kappa of 0.91 \cite{mchugh2012interrater}).
We then expanded our analysis to arrive at a set of IPS-relevant threads for further study. We randomly sampled 750 threads (250 from each forum) that we split evenly among the three coders.
As reported in Table~\ref{tab:dataset}, we ultimately found 78\% of the sampled data within Forum~A\xspace\ was relevant to IPS; 51\% within Forum~B\xspace; and 18\% within Forum~C\xspace. These figures are in line with expectations: Forum~A\xspace, which is explicitly dedicated to ``\textit{investigative equipment}'', has the highest prevalence of IPS-related content, while Forum~C\xspace, which has a more general focus on discussion of infidelity, has the lowest.
In total, 370 of the 750 randomly sampled threads were coded as IPS-relevant. We found no statistically significant correlations between thread viewership and IPS relevance in any forum, or any noteworthy patterns in seasonal, diurnal, or year-over-year posting activity within IPS-relevant data.
\paragraph{Understanding IPS-relevant content.}
We used open thematic coding \cite{braun2006using} to make sense of the 370 IPS-relevant threads.
Three researchers independently read through several threads and generated initial codes. We then met over multiple sessions to jointly develop a codebook. Through multiple rounds of iteration, we refined the codebook by applying it to additional data until we reached saturation and codebook stability.
Our final codebook contained 29 codes clustered into two high-level categories: \textit{forum culture} and \textit{tools and tactics} (see Appendix~\ref{sec:codebook}). Once the codebook was finalized, three researchers divided up the remaining threads and coded them. Our research team stayed in close correspondence throughout the analysis, repeatedly meeting to discuss threads that were unclear at first pass. We also took steps to minimize the impact of repeated readings of detailed stories of IPS and violence on our team. Researchers were encouraged to take breaks where needed, to reach out to each other regularly to process what we were reading, and to practice self-care.
We report the themes that emerged from our analysis in Sections~\ref{sec:seeker-experience} and~\ref{sec:taxonomy}. We emphasize that our analyses are qualitative: thus, we do not report raw or percentage prevalence numbers for any of our themes, except where noted and appropriately tested via inter-rater reliability measurements.
\section{Forum Interactions}
\label{sec:seeker-experience}
In this section, we give a general overview of how users interact within these forums. We begin by discussing how users self-report finding these communities, and what they seek within them. We then describe how communities respond to their requests. We identify threads in which communities encourage users to conduct IPS, either by encouraging them to carry out existing attacks, or by providing them with ideas for attacks of increased severity---what we call \textit{escalation}. We also identify threads in which communities discourage users from IPS at all, a pattern we call \textit{de-escalation}.
\paragraph{How users find these forums.} In several threads in each forum, users describe how and why they sought out these forums in the first place. Many described locating the forum via basic Google queries on topics related to infidelity and cheating. In other posts, users reported discovering the site through a trusted recommendation from a professional enlisted to help them with their relationships, such as a therapist.
Our data also show that for Forums A\xspace\ and B\xspace, users are often directed to these specific forums by moderators or users of other communities within their parent forum. For example, a moderator in Forum~A\xspace posted in response to a thread starter:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``I asked you to come here to click on the many threads and read information for yourself. There are pages and pages all about spying \dots This forum is a kind of archive where the information will be available for anyone to peruse at leisure. There is no need to wait. Look around!''}
\end{quoting}
Here, we see a forum moderator reinforce that these discussions are a resource for anyone to browse. We discovered that Forum~A\xspace\ in particular hosts several `resource threads' pinned at the top of the forum that provide primers for beginners.
\paragraph{Forum-goers' stated goals.} Once in the forum, most users make an initial post outlining a complex social situation, \mbox{usually} suspected or actualized infidelity, and ask the community for advice on what to do, e.g. ``\textit{How can I forgive him?}'' or ``\textit{How do I move on?}'' Most of these posts sought suggestions from others for `next steps', such as confronting their intimate partner or seeking legal advice. The bulk of posts within Forum~C\xspace\ followed this pattern, mimicking the advice-seeking observed broadly in forums for social support \cite{yang2019seekers, Fu-Advice:19, chen2019reddit}.
We also identified a different kind of request focused on {\it technical} support for intended or ongoing IPS, e.g. ``\textit{How can I read my wife's Facebook messages?}'' or ``\textit{How do I use this spyware tool?}'' Many of these users contextualized their asks in a detailed narrative of their situation that included an admission of past IPS, most commonly by reading a partner's text messages or emails. For example, consider this initial post, paraphrased from Forum~B\xspace:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``I caught my wife by reading her journal and emails. She does not know that I know, and I continue to monitor her email account. I don't think she knows. I haven't told a soul about this, so this is my release. I can elaborate...''}
\end{quoting}
Not all users framed their interactions as requests, however; a subset of thread starters within Forums A\xspace\ and B\xspace\ posted to share unsolicited advice on working with certain IPS tools. This advice was often couched in a personal narrative (e.g. ``\textit{Here's how I tracked his Internet history after he deleted it}'') and usually promoted the use of the tools. We consider these to be organic advertisements for these tools, the implications of which are discussed in Section~\ref{sec:discussion}.
Overall, we identified three high-level goals for users who sought IPS-related advice across these forums. Many users fixated on reading their partner's emails or text messages, or what we call
\textit{(1) investigation of a partner's prior activities}. Many were also interested in real-time access to their partner's devices, to gain information such as live updates on their partner's location and browsing history: in our view, these attackers sought \textit{(2) continuous monitoring of a partner's current and future device use}. Finally, we saw that many posters expanded the target of their IPS to include a suspected affair partner, with the goal of identifying their personal information (e.g., name, address or vehicle registration). We use the term \textit{affair partner} here to mean the person involved in an affair outside of the intimate relationship, occasionally referred to as the `other' man or woman, and we name this goal as a \textit{(3) compromise of a suspected affair partner's privacy}.
\label{sec:responses}
\paragraph{Community escalations.} Communities' responses to users' requests varied. As expected for online support forums, many responded with emotional support and advice on managing infidelity, including recommendations for looking after users' mental and physical health. In a significant body of threads, however, communities responded by encouraging thread starters to pursue their current enactment of IPS, or even to increase the severity of their attacks. We call this a pattern of \textit{escalation}: a situation in which a user begins with a relatively benign request for information, and through interactions with one or more IPS promoters is presented with ideas for enacting or increasing the severity of an IPS attack.
Consider this example from Forum~B\xspace. A user begins a thread by asking for emotional support: ``\textit{I can't believe my relationship has come to this, but I need some advice. Recently I discovered a situation that I'm not sure how to perceive...}'' An hour later, a responder offers several actionable ways for the user to invade their partner's privacy:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``There are several things you can do. Start by going into full snoop mode. Purchase a voice activated recorder and put it in his car. Snoop his phone records. Place spyware on his computer. Snoop his emails and FB account.''}
\end{quoting}
From there, a concerning dialogue unfolds. As the thread starter shares more details of their story over follow-up posts, the same responder repeatedly suggests ways to enact IPS, for example by offering:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``A voice-activated recorder is cheap, \$40 at Walmart. Stop bringing this up, and make him think everything is back to normal. Then, monitor him. Good luck.''}
\end{quoting}
To get a sense of the prevalence of this thread pattern in our corpus, we conducted an additional qualitative coding effort over our human-labeled sampling of IPS-relevant posts. Three coders first coded a random set of 30 relevant threads (10 from each forum) for whether or not they showed a pattern of escalation. Inter-rater reliability showed substantial agreement between raters: out of the 30 posts, all 3 raters agreed on 25 posts, and the remaining 5 showed 2 out of 3 raters in agreement (Fleiss's kappa of 0.77). We then split the remaining relevant threads among the three coders, finding that approximately one-third of relevant threads showed patterns of escalation (Table~\ref{tab:dataset}). This proportion remained the same in both forums explicitly focused on investigating suspected deception (A\xspace, B\xspace), and in more general support forums for those `recovering' from infidelity (C\xspace).
\paragraph{Community de-escalations.} While escalations appeared with alarming prevalence in our dataset, we also found a handful of instances of the opposite: \textit{de-escalations}, in which the community deterred a user from conducting IPS.
In many of these cases, responders reminded posters of the physical and mental impact of continuously performing IPS on a partner. These responders saw that IPS directly undermined the trust required for recovery of a healthy intimate relationship.
As this example from Forum~A\xspace shows:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``You've got to ask though, when do you stop snooping? That can't be healthy for your relationship if you're being insecure about everything.''}
\end{quoting}
De-escalating responders also often pointed out that IPS may not help people achieve their goals, and instead sabotage a relationship. For cases where IPS had already been \mbox{committed}, a small number of users pointed out how the intimate partner may be experiencing this level of privacy intrusion. We demonstrate this `pushback' against IPS through this responder on Forum~B\xspace, after a thread starter admits to monitoring his partner through a home security system:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``You sound crazy to watch her like that! The fact that you've analyzed every little detail on the system tells everyone a lot about your own insecurities ... come on dude, you're trying to make something out of nothing here.''}
\end{quoting}
Some de-escalating responders also reminded thread starters of the potential legal consequences of engaging in IPS. This included warnings that the use of some attacks could result in a criminal record, failed divorce proceedings due to misbehavior, or expulsion from social groups.
\section{Taxonomy of IPS Attacks}
\label{sec:taxonomy}
We now describe the IPS tools and tactics discussed within these forums. We present a taxonomy (Table~\ref{tab:taxonomy}) of four types of attacks: (1) \textit{tool-based attacks requiring physical access}, including installing spyware on a partner's phone and attaching GPS trackers to their person; (2) \textit{tool-based attacks not requiring physical access}, including leveraging shared cloud accounts; (3) strategies involving \textit{coercion and subterfuge}, for example convincing a partner to provide access, or tricking them into connecting with falsified social media profiles; and (4) \textit{outsourced attacks}, namely hiring private investigators.
\subsection{Tools that require physical access}
\label{sec:taxonomy-tools-physical}
\begin{table}[t]
\centering
\footnotesize
\begin{tabular}{|l|}
\hline
{\bf Tool-based attacks that require physical access} \\
Using a cellphone backup recovery tool on a partner's device \\
Installing a keylogger on a partner's device \\
Installing screen recording spyware on a partner's device \\
Installing GPS trackers on a partner's body or in their car \\
Installing audiovisual recorders in the car or the home \\ \hline
{\bf Tool-based attacks that do not require physical access} \\
Leveraging features of a shared phone plan \\
Using shared cloud tools to access a partner's personal data \\
Using router monitoring tools to track and manipulate Internet activity \\
Using reverse lookup directories to find personal information \\
\hline
{\bf Coercion and subterfuge} \\
Leveraging physical proximity to gain access \\
Convincing a partner to give total access \\
Catfishing a partner \\ \hline
{\bf Outsourced attacks} \\
Hiring a private investigator \\
\hline
\end{tabular}
\caption{Taxonomy of IPS attacks promoted on these forums.}
\label{tab:taxonomy}
\end{table}
Our analysis surfaced many attacks requiring access to a target's devices. These attacks are particularly possible in IPS, due to the proximity between intimate partners \cite{freed2018stalker}.
\paragraph{Backup recovery tools.} Recall that a common goal for attackers was the discovery of what a partner said in their texts or emails. To this end, responders promoting IPS often recommended the use of \textit{cellphone backup recovery tools}: both specific software dedicated to reading data from phones or SIM cards, and creative workarounds leveraging built-in iOS or Android features to access that same information. Some of the spyware previously reported \cite{chatterjee2018spyware} works by accessing similar data stores; our data show for the first time how attackers share these products with each other, and how they homebrew their own tools for accessing this information.
In particular, a substantial number of threads were dedicated to tools that recovered deleted texts from iPhones. Similar tools were available for Android phones, and in older threads we even surfaced evidence of responders helping attackers retrieve texts from Blackberries. While some responders in these threads advocated for the use of specific products, others presented instructions for homebrewed tools they had developed to read messages from a partner's backup files synced to shared iTunes or iCloud storage. Some responders posted code anyone could use to convert such backup files into text files for easy reading, and many also offered one-on-one technical support.
\paragraph{Keyloggers and screen recorders.} Many attackers were interested in continuous capture of their partners' digital activities, such as websites they visited or passwords they used. For these attackers, responders often recommended installing \textit{keyloggers} and \textit{screen recorders} on a partner's devices. These tools had been surfaced as potential spyware in prior work \cite{chatterjee2018spyware}, but our data highlight they are actively shared as solutions for attackers on these forums. One responder on Forum~C\xspace\ claimed he had installed keyloggers on all PCs and laptops in his home, describing the benefits of these tools:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{Great for capturing passwords \& her true thoughts when messaging (things she backspaced over and didn't send).}''
\end{quoting}
Many responders also recommended screen recorders, such as those built for companies to install on workers' devices---in fact, this use case was often invoked to prove a product's legitimacy. Responders also discussed the benefits and drawbacks of specific products, including whether the paid tiers of some tools were worth purchasing.
\paragraph{Location tracking and audiovisual recording.} We saw many instances of responders recommending tools for environmental surveillance of a partner's activities, conversations and whereabouts, e.g., \textit{voice-activated recorders} and \textit{GPS tracking devices} placed in key locations like a partner's car. Responders were quick to make recommendations about where to obtain these devices, how much one should expect to pay for them, and best practices for hiding them from targets:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{A GPS tracker can fit into a purse without them knowing. I'm positive you can figure out a place to stash one in a car. People track autistic kids and animals with them.}''
\end{quoting}
Surveillance of partners in cars was a recurring theme throughout our data. In addition to providing recommendations on the best places in a car to place a GPS tracker, several threads promoted the use of more sophisticated tools that plug into a car's on-board diagnostics (OBD) system and continuously report the car's location to a remote database, to which an attacker can then subscribe. These tools would be useful, one responder said, because ``\textit{unless a person knows to check the OBD they would never think to look for it.}''
\subsection{Tools that do not require physical access}
\label{sec:taxonomy-tools-no-physical}
For would-be attackers who were unable to access a partner's phone to install spyware or car to plant a GPS tracker, the responders in our data readily provided tools that did not require physical access to partner or device.
\paragraph{Leveraging shared phone plans.}
Many would-be attackers sought ways to leverage the fact that they shared a phone plan with their intended target. Most seemed to know that a partner's call and SMS histories were accessible on a phone bill; in fact, viewing these was often the first thing an attacker tried, and the use of these records as vectors for abuse has been documented \cite{freed2018stalker}. But the contents of messages are often left off of phone bills; in response, our data show these attackers come to the forums to find other ways to obtain more information from their service providers.
Responders regularly provided tips on how to contact service providers and obtain more detailed records: for example, in one thread on Forum~B\xspace, a responder described how to contact Verizon and set up a monthly spreadsheet dump of all call activity. Phone companies were required to provide these records to account owners, the responder claimed, as a form of consumer protection.
Attackers were also savvy to the many other ways a provider's plan management tools could be used to surveil a partner. Verizon, AT\&T, and T-Mobile were purported to have capabilities ranging from email monitoring to mobile keylogging. Consider the following exchange on Forum~A\xspace:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
{\bf Attacker:} ``\textit{We are in the process of choosing new cellphones and a new company. Which is the best company to keep tabs, records, etc? We currently have iPhones on AT\&T, and their Family Map did help me prove his affair.}'' \\
\noindent {\bf Responder:} ``\textit{If you're getting everything you need with AT\&T, I would stay with them. They have immediate online access [to phone records] and their GPS is good.}''
\end{quoting}
In this example, we see responders outline the features of shared plans that make them useful to an attacker: immediate online access to call and text histories, quality GPS for location tracking, and family sharing products that provide easy-to-use interfaces for surveillant capabilities. This last type of tool was especially common in our data, confirming prior work \cite{chatterjee2018spyware, freed2018stalker}. This example also highlights the \mbox{collaborative} nature of how attacks surface in these forums, with a responder echoing and encouraging an attacker towards IPS. Of note, the responder in this example is the third-most-prolific superuser of Forum~A\xspace, and many of their posts are similarly IPS-related.
\paragraph{Features of shared cloud services.} Many tools that did not require physical device access took advantage of the built-in features of cloud-based sharing tools. The use of cloud tools for abuse has been reported in prior work \cite{freed2018stalker}; however, our data show for the first time how attackers share these tools with each other as ways to overcome targets' defenses.
In many threads, attackers seemed aware of the ways iCloud tools in particular could be used to surveil partners who had not provided device access. One thread began:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{What is the best spyware if I can't get their phone, but have their Apple ID and password?}''
\end{quoting}
In this example and many others, our data show attackers are encouraged by the forum to use their partner's Apple ID to view their personal messages and photos from a web browser---no device access necessary. This was commonly invoked as a solution for attackers who sought more detailed information on their partner's texts than records from a service provider contained. Many of these attackers reported they arrived at this method of attack because they had seen a drop-off in their partner's texting activity as reported by their phone bills, and had inferred the partner had moved to iMessage or another messaging service that used data rather than SMS. (Messaging that uses data is not typically itemized on a phone bill.)
Some of these attacks, however, did not even require an attacker to use a partner's login, because their personal data was already syncing to a shared Apple device. For example, an attacker on Forum~C\xspace described discovering she could view a partner's messages on a family iPad, which was synced to her partner's iCloud account. Our data show attacks of this nature also levied against third-parties, namely the affair partners: in one thread on Forum~B\xspace, an attacker describes realizing her partner's affair partner was using an iPad synced to an iCloud account shared by all three-parties, making her purchasing and Internet history accessible for the attacker to browse.
Attackers were particularly eager to share how iCloud tools could be used for location tracking. One attacker on Forum~A\xspace described how to use the Significant Locations feature within iOS to examine a partner's recent location history. In another thread on Forum~A\xspace, a user shared an article on Find My Friends and called out its abusive potential:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``Interesting article about an iPhone app called `Find My Friends', which you may be able to load on your spouse's phone to track their whereabouts.''}
\end{quoting}
Cloud-based tools outside of the Apple ecosystem were also called out for similar purposes. One responder on Forum~C\xspace shared how Android users could view a ``\textit{timeline}'' of a partner's visited locations via their Google Maps account. Another shared how WhatsApp's phone-to-Web syncing features could be used in concert with one-time physical access to maintain continuous access to a partner's messages. This responder described the initial connection as a ``\textit{one-minute job}'' best done while a partner sleeps, and claimed they were ``\textit{actually shocked at what a privacy flaw this seems to be.}''
Lastly, our forums contained many suggestions for mobile spyware products that leveraged cloud-based access to a target's device, such as tools marketed for use in parental control contexts. Much of the discussion of these products also offered advice on free versus paid tiers, setup and configuration, and even best practices for contacting customer service teams.
\paragraph{Web traffic trackers on shared networks.} In several forums, we discovered threads in which responders offered advice on how to install \textit{web traffic monitoring tools} on a shared WiFi network. The scope of this attack and the level of detail in which it was described was noticeably more sophisticated than others in our data, or what has to our knowledge been previously reported.
In one thread on Forum~C\xspace, a person who described themselves as a ``\textit{heartbroken techie}'' with a background in software development started a thread detailing how they used a DNS resolution service to monitor the traffic on their home router. With their tool, the attacker said, they could record every website their partner visited, regardless of whether they deleted their Internet history, in the form of reports issued within 24 hours. The attacker shared the command line scripts and configurations they had used, and even offered to share a GitHub repository where others could retrieve their code. In addition to describing how they used the service to monitor router traffic, they went on to discuss how they used its domain blocking alerts to manipulate their partner:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``You can set up a customized message (as I call it, the `oh shit' alert) that will pop up if they try accessing a site that is blocked. It's amazing how much someone will confess if they know you're tech-savvy and you tell them you have a detailed history of their actions (even if you don't.)''}
\end{quoting}
In another case, a responder on Forum~A\xspace who claimed to be a computer security professional introduced the forum to the concept of a man-in-the-middle attack and recommended an entry-level tool
for mounting one. As they described, the tool was able to obtain not just a history of websites visited, but also copies of data sent over the network, e.g. the contents of emails and chats. Most notably, they described the tool as a way to actively manipulate a partner's activity:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``[You can also] modify the data traffic in real time. This can be used for tactics like replacing phone numbers, names and addresses as they travel over the network. Think about creative ways to change the contents of the websites/emails/chats that they're looking at.''
\end{quoting}
This last example was sourced from one of the `resource threads' in Forum~A\xspace. The responder goes on to offer his services to community members who want help mounting such attacks. We discuss the implications of these types of attacks and the role of technologists providing such support in Section \ref{sec:discussion}.
\paragraph{Reverse lookup directories.} Lastly, our data show would-be attackers seeking and receiving tips for investigating their partners' prior actions via reverse lookup tools, used most commonly to identify people from their phone numbers. Most cases presented as a thread starter finding an unknown number in a partner's texts or call records via other attacks, and then asking the forum for advice on how to discover whether it belonged to an affair partner or an escort service:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``If anyone knows a really good reverse cell lookup, please let me know. Just found a few unknown numbers on my husband's phone.''}
\end{quoting}
Many solutions offered were simple websites containing databases of people's personal information---one thread even offered tips on how to search Facebook by phone number. But responders in our data also recommended a wide array of commercial products that market themselves as collators of public information on individuals (e.g., WhitePages). Many of these tools offer a free tier enabling lookup of names, addresses, and phone numbers in addition to a paid service for more thorough background checks.
Although these tools are relatively unsophisticated from a technical perspective, they featured in several stories that resulted in an attacker confronting their partner or suspected affair partner at an address or phone number located through these services.
\subsection{Coercion and subterfuge}
\label{sec:taxonomy-coercion}
In addition to recommending specific tools,
many responders had advice for coercing or subverting a target into providing access to their data and accounts, most often passwords.
\paragraph{Leveraging physical proximity to gain access.}
Attackers frequently shared how they used their close physical proximity to their targets to overcome common defenses without specific tooling. While many of these tactics had been previously reported from victims' perspectives \cite{freed2018stalker}, we report for the first time attackers jointly developing such coercive strategies in public forums.
In many cases, attackers advised each other to manipulate a partner into `accidentally' revealing a password, as seen in the following example from Forum~A\xspace:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{Get her to send texts \dots while you are sitting next to her. Then try to make out the password as she types it in.}''
\end{quoting}
These strategies often did not require active manipulation. In some cases, gaining access was as simple as waiting for a partner to fall asleep:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
\textit{``My wife would get drunk and pass out. It was simple to just hold the iPhone up to her thumb to unlock it. Took pictures of a lot of conversations so I have a record.''}
\end{quoting}
Some would-be attackers sought help creating opportunities like these. In one thread on Forum~A\xspace, an attacker asks:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{I have wondered if there is a relaxing drug that will knock her out long enough for me to scan her texts and photos. Any suggestions on how I get that phone?}''
\end{quoting}
Once a partner slept or was otherwise unconscious, attackers and responders offered a range of strategies for exploiting their lowered defenses. These included ways to overcome two-factor authentication schemes---namely, resetting passwords on locked accounts and taking the opportunity to capture codes---as well as ways to plant monitoring tools that could track activity long-term, such as swapping their SIM card into a partner's device to capture their call and text activity.
\paragraph{Creating fake profiles to access their social media.} Coercive attacks did not necessarily need to be physical or direct: we also found evidence of manipulations via social media. Several threads showed attacks using fake social media profiles to overcome the privacy controls a target may have set. While the use of fake social media profiles to directly harass targets has been reported \cite{freed2018stalker}, we found evidence of attackers leveraging fake profiles in a new way: using second- and third-degree connections to access a target's profile.
In one thread on Forum~A\xspace, a responder details step-by-step how to fabricate a believable Facebook account, and use it to befriend accounts that are friends with the target. The attacker can thus access parts of the target's social media profiles that have been locked away from the attacker, but have remained unlocked to what they believe are friends-of-friends. The responder describes the access afforded thus:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{After your friend request has been accepted, revisit the pages where you couldn't see anything before. You'll be shocked at how much information will suddenly be available, as many MANY people set up security so it's not public, but can be viewed by friends-of-friends. In my neck of the woods there are a lot of local bars that have 1000+ friends and guess what? Every one of those 1000+ friends has now given access to those 1000+ people that allow friends-of-friends to see their info, and I'd guess over half do.}''
\end{quoting}
That responder then cautions attackers to ensure the privacy settings on their real and fake Facebook profiles are set to only show information to the account owner, because, the responder says, ``\textit{I don't believe FB is secure.}''
\paragraph{Convincing a partner to provide total access.} Our data also showed attackers and responders championing a strategy of simply convincing a partner that unfettered access to all devices and accounts should be expected in an \mbox{intimate} \mbox{partnership}. In our context, infidelity forums, this was often raised as a way to facilitate reconciliation after an affair:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{Tell her you need her iCloud password to review something. If she refuses, that's a giant red flag. Then I suggest you say `wife, I love you dearly, but if I don't see what's on that phone, then you are telling me that you're cheating. If you have nothing to hide, then let me see it.'} ''
\end{quoting}
Freed et al. \cite{freed2018stalker} previously reported that abusers often convinced IPV survivors to share their passwords during ``good'' phases of a relationship as a way to establish trust, and subsequently threatened them to continue sharing or face consequences when the relationship turned ``bad''. Our work extends this to show that attackers promote to each other the idea of privacy compromise as currency in abusive relationships, often describing this as key to overcoming the emotional toll of suspicions of infidelity. In fact, many of these threads shared stories in which a partner who was suspected of cheating in the past still shared total access months or years later.
\subsection{Outsourced attacks}
\label{sec:taxonomy-outsourcing}
The final category of IPS attacks surfaced in our data are those in which a responder recommends external resources for investigation or monitoring of an intimate partner.
\paragraph{Private investigators (PIs).} As expected for an infidelity forum context, many attackers and responders within our data referenced hiring PIs to track their partners, with the goal of finding evidence of an affair. Notably, hiring a PI was framed as a legal and ethical way to obtain information, as in this example from a responder on Forum~B\xspace:
\begin{quoting}[leftmargin = .1 in, vskip= .1 in, indentfirst = false]
``\textit{You should hire a PI, who acts within the law to obtain the confirmation you require. But that's all that it will be in the vast majority of cases...merely confirmation.}''
\end{quoting}
Many of the PIs recommended in our data offered services within their specific localities, e.g., ``\textit{He does not do surveillance unless the target originates in [specific U.S. state].}'' These recommendations typically included a phone number to call and a person to ask for---or even, in some cases, a person from the forum to claim as a referral. In addition to general recommendations to seek out PIs and referrals to specific PIs, we also found one thread in which a responder posted a link to the website of a specific national agency, as well as another thread with a directory of `\textit{vetted}' PIs.
These recommendations were often framed as more costly than other attacks, to be used as a last resort. One responder on Forum~A\xspace\ remarked PIs were expensive, but at least ``\textit{cheaper than a divorce lawyer}''.
\section{Discussion}
\label{sec:discussion}
We now synthesize takeaways from our findings for anti-IPS efforts. First, our work extends the IPS threat model outlined in this and prior works \cite{freed2018stalker, chatterjee2018spyware}. Security experts seeking to prevent their work from misuse in IPS might consider accounting for this threat model in their technology development practices---in particular, Freed et al. \cite{freed2018stalker} supplies the concept of a \textit{UI-bound adversary} as a consideration for design teams.
We additionally outline broader considerations for security experts, including the use of these online communities as a source of IPS threat intelligence and the development of countermeasures that target the commercial entities behind the spyware industry. Finally, we describe how our work is already impacting interventions, and close by posing a set of open ethical questions for the security community.
\paragraph{Online forums are a rich source of IPS threat intelligence.} Our work highlights how analysis of online communities can provide anti-IPS advocates with valuable intelligence on the motivations and tactics of intimate partner abusers. By observing how attackers interact in these forums and the specific tools they promote for use, we were able to surface new knowledge on IPS strategies that can directly inform interventional efforts.
Our results confirm findings from prior work that showed the abundance of dual-use and overt spyware apps available for attackers \cite{chatterjee2018spyware}, and highlighting victims' experiences of tech-enabled abuse \cite{freed2018stalker}. But the details of how, precisely, abusers learn to mount these attacks had not previously been reported, and attackers' levels of sophistication had not been well-understood. Our analysis highlights how attackers are collaborating on new tactics in these forums, and surfaces how these attacks are conducted at an unprecedented level of granularity. For example, we find that attackers are not just inspecting targets' call histories on shared family plans, as has been reported previously \cite{freed2018stalker}---they are also sharing strategies on how best to contact service providers and obtain more detailed records. We also surface novel attacks more sophisticated than those previously reported in the IPS context, for example the use of WiFi router tools to monitor and manipulate a partner.
Mining these forums for threat intelligence might help anti-IPS efforts stay ahead of attackers' ever-evolving techniques. We see substantial future work in creating semi-automated tools that enable analyses like ours to scale. As our initial keyword searches (Section~\ref{sec:data-desc}) showed, the way IPS manifests in user-generated natural language may be too nuanced for current automated techniques alone to reliably detect. However, human ratings are laborious and inefficient and, importantly, repeated exposure to stories of abuse can inflict harm on people analyzing large bodies of such texts \cite{figley_compassion_1995}. These problems are exacerbated on large social media platforms, where the volume and speed of conversations generated by millions of users creates urgent problems of scale.
We see a role for advanced language processing techniques in overcoming these challenges. For example, a system might quickly and reliably extract the specific strategies recommended within a post without relying on forum-specific features (as applied in cybercrime marketplaces in \cite{durrett2017identifying}). Such a system might provide a valuable pipeline for security and privacy researchers and anti-IPV advocates building frontline defenses against emergent attack strategies, for example the Coalition Against Stalkerware.\footnote{https://stopstalkerware.org}
In fact, our work has already had impact in this regard. We shared our results with the team of practitioners that runs a technology clinic providing direct interventions to IPV survivors facing IPS \cite{havron2019clinical, freed2019clinical}.\footnote{https://www.ipvtechresearch.org/} At time of writing, they are working to integrate our threat intelligence into their training materials for advocates, as well as their clinic's procedures for discovering and mitigating how abusers are enacting surveillance against their clients.
\paragraph{The for-profit industry behind IPS products uses online forums to market their tools.} Our work also shows that the online ecosystems promoting spyware feature a significant presence from companies creating and marketing their own surveillance products. At one extreme, we found entire forums bloated with spam advertisements for a single spyware tool, suggesting these forums were leveraged to manipulate that tool's SEO. We also found that recommendations for IPS tools are not just manifesting as spam advertisements and SEO for specific spyware apps, but are also shared organically among users in the forum communities.
Within these organic posts, responders engage meaningfully with forum-goers' relationship problems, but also evangelize specific IPS tools or approaches and even serve as technical support for users' spyware installations. Their posts reveal the concerns of consumers in this market: we find posts on the merits and drawbacks of a range of spyware products, both free and paid, and the market rates for voice-activated recorders, GPS trackers, and PIs. All told, our analysis suggests these forums are likely one corner of a broad industry offering `solutions' for would-be attackers to turn suspicions of infidelity into actualized IPS.
These findings suggest a role for mitigation strategies that directly target these commercial entities. Prior work has demonstrated ways to undermine commercially motivated spam and SEO attacks by working directly with banks and payment processors to make e-crime difficult to monetize \cite{kanich2011show, mccoy2012priceless}. Similar approaches may be effective in the context of IPS. Future work should investigate further the mechanics of how forums like these are leveraged as marketing tactics by spyware companies, with the goal of informing such `follow-the-money' countermeasures. Similar countermeasures might be also useful for security experts concerned about keeping their tools from being misused for IPS: it is possible, for example, that dual-use apps \cite{chatterjee2018spyware} are being advertised as spyware. Where possible, security experts should prevent their tools from being marketed in this way.
\paragraph{Platform-level defenses might mitigate the spread of IPS.} In addition, our work raises concerns for large social media platforms, where conversations escalating into IPS ``how-tos'' may be happening in spaces that are not specifically dedicated to infidelity, surveillance, or intimate relationships.
We have highlighted features of IPS-relevant conversations, and outlined an agenda for creating semi-automated techniques to extract attackers' strategies from such forums. Platforms concerned about their role in enabling the spread of IPS could use this work to develop community norms or content moderation strategies attuned to these forums' dynamics---for example, the moderators of Forum~C\xspace might consider banning posts that escalate threads into IPS, and instead seek to encourage de-escalation. Future work might investigate further the scale of the problem within popular social media networks as a first step to developing such mitigation strategies.
Social media platforms might also consider the fake profile attacks discussed in Section~\ref{sec:taxonomy-coercion}, and use the patterns we uncovered to more effectively surface falsified accounts. They might also consider de-emphasizing second- and third-degree network connections in users' experiences of their platforms, or offer privacy controls that limit users' audiences to first-degree connections by default.
\paragraph{Online communities are collaboratively creating new IPS attacks.} Our work also shows that people with significant training in computer security and privacy, while potentially well-meaning, are actually helping to further develop IPS attacks on these forums.
We were surprised to see the level of technical sophistication in some threads, particularly in contrast to the relatively unsophisticated techniques reported in prior studies with victims \cite{matthews2017stories,freed2018stalker}. Some attacks did fall into the bucket of previously known techniques, for example physical privacy violations like shoulder surfing that are addressable through existing defense strategies \cite{khan2018evaluating, davin2017baseline}. However, several involved the use of custom shell scripts and other more sophisticated techniques, including methods for extracting information from artifacts like iPhone backups and leveraging DNS resolution tools to manipulate a partner's Internet traffic.
In a sense, the IPS promoters who championed their homegrown surveillance tools in these forums were engaged in a process of collaborative innovation, working with other tech-savvy community members to create and refine new abuse tactics. This is similar to behaviors seen previously in cybercrime forums~\cite{cybercrime:weis15}, in which a handful of sophisticated users create tools and then provide or sell them to the community. These collaborative processes are not just creating more efficient attacks, they are also making attacks more accessible to more would-be attackers: much of the discussion on these forums serves as how-to guides for less tech-savvy members, and in many cases communities even provide one-on-one troubleshooting. Future research might further analyze the cooperative dynamics of how these forums develop new attacks, and compare these against known tactics used to perpetrate harm in offline settings (c.f., \cite{hearn_violences_1998}).
\paragraph{Infidelity is used as a justification for IPS.} The forums we studied were rife with emotionally vulnerable people seeking and receiving assistance with difficult interpersonal problems. But they were also rife with attackers freely admitting to and promoting the use of surveillance tools against an intimate partner, often by arguing that infidelity justifies surveillance. In this, we see that the context of infidelity both attracted people to the forums as a site for emotional support and masked them from the social exclusion they might have faced if admitting to IPS in a non-infidelity context \cite{hearn_violences_1998}. This is particularly concerning for anti-IPS efforts, as it can set a precedent of using infidelity as an excuse for abusive actions---a practice mirrored in offline discussions with abusers in IPV \cite{nemeth2012sexual}.
We see compelling areas for future work in using these forums to identify the cultural norms and justifications that encourage abusive behaviors. In concert with ongoing behavior change work with abusers \cite{kelly_naming_2016}, such work could draw on the patterns of de-escalation we uncovered to develop alternative strategies for resolving suspicions of infidelity without resorting to IPS. These alternatives could be promoted before IPS on these forums, thereby retaining the supportive community structure they provide to some forum-goers while discouraging abusive practices.
The infidelity forum setting also creates gray areas for the computer security community at large. Some of the responders providing (ab)users with strategies for IPS were self-described computer security experts who reported using the same tools they promoted for surveillance in their professional work. In the context of helping people who reported they were in toxic relationships, these experts may have felt their material support facilitating IPS was justified. What's more, in the context of publicly available forums like the ones in our study, these experts' bespoke surveillance solutions constitute a persistent record accessible to anyone on the Internet with little effort. It is possible that well-meaning computer security experts may have facilitated IPS not just for the user who posted in the forum, but also for the numerous people who would browse these threads in the future.
We raise these tensions as a set of open ethical questions for the security community. Much as medical doctors operate by a professional code of conduct to `do no harm,' should computer security experts abide by a corresponding professional ethos to wield their expertise only for good? How should judgment calls between justifiable and unjustifiable surveillance be made, and who should make them? And how can computer security experts balance publicizing attacks to support anti-IPS efforts (for example, in this work) against the possibility that doing so might inadvertently help more attackers?
\section{Conclusion}
We have provided the first measurement study of the online communities in which people enacting IPS discuss their tactics. Through a mixed-methods study of five public forums, including a Reddit subforum dedicated to infidelity, we developed a taxonomy of the IPS strategies attackers discuss online. We showed that these forums are sites for both spam advertising specific spyware products and organic discussion of surveillance tools between users. We highlighted threads in which (ab)users learn new IPS tactics from more tech-savvy forum-goers, as well as cases where forums deter them from conducting IPS. Our work is already impacting anti-IPS interventions by informing programs that directly assist victims.
\section*{Acknowledgments}
We thank Sandra Ebirim for vital contributions to the data analysis phase of our study. We are also grateful to our reviewers, whose comments greatly helped to improve our paper. This work was funded in part by NSF Awards \#1916096 and \#1916126, as well as gifts from Facebook and Google.
\bibliographystyle{plain}
| 79b3b6dc0331b43c3c35145268212022968de9d2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{SecIntro}
It is a well-known fact that the Earth's environment---its lithosphere, hydrosphere, atmosphere, and biosphere---has transformed greatly over time \citep{lun13,Knoll,KN17,SSC20}, and the same also applies to other terrestrial planets in our Solar system \citep{Eh16,KAC19}. In tandem, there is growing awareness and acceptance of the fact that habitability is a multi-faceted and dynamic concept that depends on a number of variables aside from the existence of liquid water \citep{Dole,Kast12,Co16,SBJ16,Co20,LL21}; the latter criterion has been widely employed to demarcate the limits of the so-called habitable zone and its manifold extensions \citep{Hu59,Dole,KWR93,KRK13,KRS14,Ram18,RAF19,SRO19}.
One of the most crucial environmental parameters that regulates myriad biological processes, and thus the propensity for planetary habitability, is the ambient temperature \citep{CB87,HS02,Ang09,AC17}. It is not surprising, therefore, that there exists a large corpus of work on the thermal limits of life based on comprehensive experiments on thermophiles \citep{RM01,McK14,AC14,MAB19}. In recent times, numerical models have employed the thermal limits for Earth-like complex life to assess the habitability of exoplanets for such organisms \citep{SVS17,MPV20}, and similar analyses have been undertaken for generic subsurface biospheres as well \citep{MOP13,Mana20}.
Motivated by these facts, we will study how temperature impacts the prospects for aquatic photosynthesis on Earth-analogs around stars of two noteworthy spectral types. By Earth-analogs, we refer hereafter to rocky planets that are sufficiently similar to Earth insofar as \emph{all} their geological, physical, and chemical properties are concerned. Our reasons for choosing to investigate aquatic photosynthesis are twofold. First, the importance of photosynthesis is well-established from the standpoint of physics and biochemistry as stellar radiation is the most plentiful source of thermodynamic disequilibrium \citep{DW10}, and photosynthesis represents the dominant avenue for the biosynthesis on organic compounds on Earth \citep{BPM18}.
In particular, we will focus on oxygenic photosynthesis because its electron donor (water) is available in plentiful supply, consequently ensuring that this mechanism is not stymied by the access to electron donors \citep{WRF}. Moreover, the advent of oxygenic photosynthesis is known to have profoundly altered Earth's geochemistry and biology \citep{Lan02,Jud17}. We will adopt the conventional range of $\lambda_\mathrm{min} = 400$ nm and $\lambda_\mathrm{max} = 700$ nm for oxygenic photosynthesis \citep[Chapter 1.2]{Bla14}, known as photosynthetically active radiation (PAR). To be precise, oxygenic photosynthesis can operate at wavelengths of $350$-$750$ nm \citep{CB11,NMS18,CAB21}, but the canonical choice of the PAR range delineated above does not alter our subsequent results significantly.
We will not delve into the feasibility of multi-photon schemes that might elevate $\lambda_\mathrm{max}$ to longer wavelengths, because their efficacy has not been adequately established. On the one hand, it is plausible that the upper bound (namely $\lambda_\mathrm{max}$) for PAR could be boosted to wavelengths of $\gtrsim 1000$ nm \citep{WoRa,TRY06,KST07,Man19,Man20}. However, on the other hand, these multi-photon schemes may be more fragile and susceptible to low efficiencies due to side reactions \citep{KST07,Kume19,LiM20}. Moreover, recent numerical modeling based on empirical data indicates that, while photosynthesis in the near-infrared is feasible, oxygenic photosynthesis on M-dwarfs may eventually revert to the conventional PAR range described in the preceding paragraph \citep{GW17,TMT17}.
The second reason why we opt to investigate the prospects for aquatic photosynthesis stems from the basic datum that the oceans contribute nearly half to the overall NPP of modern Earth \citep{FB98}. In fact, Earth was almost exclusively composed of oceans (i.e., virtually devoid of large landmasses) for a certain fraction of its history \citep{IK10,AN12}, implying that aquatic photosynthesis may have played an even more significant role in those periods. A few theoretical models have even proposed that continents only emerged in late-Archean era in the neighborhood of $2.5$ Gya \citep{FCR08,LL21}; this conjecture seems to be compatible with the recent analysis of oxygen-18 isotope data from the Pilbara Craton of Western Australia \citep{JW20}.
Looking beyond Earth, statistical analyses of exoplanets indicate that a substantial fraction of super-Earths are rich in volatiles \citep{Rog15,WL15,ChKi17,ZJS18,JM18}. In particular, some of the Earth-sized planets in the famous TRAPPIST-1 system \citep{GT17} may fall under this category, with the water fraction potentially reaching values as high as $\sim 10\%$ by mass \citep{GDG18,UHD18,DMG18}. The habitability of ocean planets (also called water worlds), which are wholly devoid of continents \citep{Kuc03,LS04}, has been analyzed from multiple standpoints \citep{ACC12,KSR13,CA14,Gold15,NSR17,KF18,RL18,LiMa}. In recent times, increasing attention is being directed toward oceanographic phenomena such as salinity, circulation patterns and nutrient upwelling \citep{HY14,CS16,LL18,YAK19,COJ19,DWK19,OJA,SOK20} on such worlds. However, a detailed treatment of the salient characteristics of aquatic photosynthesis remains missing for the most part.
It is important to recognize that we will deal with aquatic environments, but this does not necessarily imply that all worlds under consideration must be solely composed of oceans. The outline of the paper is as follows. We commence with a description of some of the basic tools needed to facilitate our analysis in Sec. \ref{SSecPrelim}. We proceed thereafter by calculating how the properties of aquatic photosynthesis such as the compensation depth and the net primary productivity (NPP) vary with the PAR flux and ocean temperature in Sec. \ref{SecChar}. Next, we explain the salient model limitations in Sec. \ref{SecModLim}. Subsequently, we delineate the ramifications arising from our modeling for Earth-like exoplanets in Sec. \ref{SecDisc}, and we conclude with a synopsis of our central findings in Sec. \ref{SecConc}.
\section{Mathematical preliminaries}\label{SSecPrelim}
In order to study the basic characteristics of aquatic photosynthesis and their dependence on the average ocean temperature ($T_W$), we hold all parameters (biological, geological and astrophysical) identical to that of Earth. We consider two different Earth-analogs hereafter: one around a solar twin (Planet G) and the other around a late-type M-dwarf (Planet M) with effective temperatures of $T_\odot = 5780$ K and $T = 2500$ K, respectively. Planet M is taken to be \emph{tidally locked} \citep{Barn17}, and the star that it orbits has a temperature closely resembling that of TRAPPIST-1 \citep{DGT18}. The reason for doing so is that Sun-like stars are considered ``safe'' targets for biosignature searches \citep{KWR93,HA14,LiL18}, whereas the habitability of M-dwarf exoplanets, especially those orbiting active stars, remains subject to many ambiguities \citep{TB07,SKS07,SBJ16,LL19}.
In what follows, we draw upon two major simplifying assumptions. First, we model the star as an idealized black body with an effective temperature of $T$. Second, we account for the attenuation of PAR after the passage through the atmosphere by introducing a fudge factor. While neither of these simplifications are entirely realistic, the global results are known to deviate from more realistic models and data by a factor of only $< 1.5$ for the most part \citep{LL20}.\footnote{In fact, the spatial heterogeneity inherent to oceans are known to introduce local variations that are more than an order of magnitude greater than this factor \citep{BBS05}, owing to which the estimates in \citet{LL20} can be regarded as fairly accurate global values.} The reason for this reasonable degree of accuracy stems from the fact that most of the basic quantities we compute hereafter exhibit a weak (i.e., semi-logarithmic) dependence on the two assumptions outlined above.
As we are dealing with Earth-analogs, the stellar flux at the planet's location is taken to be $S_\oplus \approx 1360$ W/m$^2$. At the substellar point on the planet, the photon flux density ($\mathcal{N}_\mathrm{max}$) at the top of the atmosphere is given by
\begin{equation}\label{SpecRadDef}
\mathcal{N}_\mathrm{max}(\lambda) \approx n_\lambda \left(\frac{R_\star}{d_\star}\right)^2,
\end{equation}
with $R_\star$ and $d_\star$ constituting the stellar and orbital radius, respectively, whereas $n_\lambda$ is the photon flux density of the star at its surface. The black body brightness $B_\lambda$ is invoked to yield
\begin{equation}
n_\lambda = \frac{B_\lambda}{(hc/\lambda)} = \frac{2c}{\lambda^4}\left[\exp\left(\frac{h c}{\lambda k_B T}\right)-1\right]^{-1},
\end{equation}
where $\lambda$ is the photon wavelength. As we have assumed the stellar flux is equal to $S_\oplus$ for the Earth-analogs, we can express $d_\star$ as follows:
\begin{equation}
d_\star = \sqrt{\frac{L_\star}{4\pi S_\oplus}}
\end{equation}
where the stellar luminosity ($L_\star$) is given by $L_\star = 4\pi \sigma R_\star^2 T^4$. After employing this relation in (\ref{SpecRadDef}), we find that $ \mathcal{N}_\mathrm{max}$ transforms into
\begin{equation}\label{SpecRadMax}
\mathcal{N}_\mathrm{max}(\lambda) \approx \frac{n_\lambda S_\oplus}{\sigma T^4}.
\end{equation}
It is, however, necessary to recognize that $\mathcal{N}_\mathrm{max}$ constitutes an upper bound for the photon flux density at the surface for two reasons. First, this photon flux density is calculated at the zenith, and therefore ignores the fact that a given location will not always correspond to the substellar point. Second, the effects of clouds and atmospheric attenuation are neglected. Hence, a more viable expression for the photon flux density at the planet's surface ($\mathcal{N}_\mathrm{avg}$) is given by
\begin{equation}\label{SpecRadavg}
\mathcal{N}_\mathrm{avg}(\lambda) \approx \mathcal{N}_\mathrm{max}(\lambda) \cdot f_\mathrm{I} \cdot f_\mathrm{CL},
\end{equation}
with $f_\mathrm{CL}$ embodying the total atmospheric attenuation \citep[Chapter 4.2]{SG06}, and $f_\mathrm{I}$ quantifying the average intensity of light at a given location as a fraction of the intensity at the substellar point. Henceforth, we adopt $f_A \equiv f_\mathrm{I} \cdot f_\mathrm{CL} \approx 0.2$ for reasons elucidated further in \citet{LL20} and to maintain compatibility with Earth's global parameters \citep[Chapter 4.3]{SG06}; altering this fraction by a factor of order unity does not change our results significantly due to the logarithmic dependence alluded to earlier in this section. With this choice of $f_A$, it should be noted that $\mathcal{N}_\mathrm{avg}(\lambda) \approx 0.2 \mathcal{N}_\mathrm{max}(\lambda)$.
Although the above choice has been motivated in \citet{LL20}, a recapitulation is warranted at this stage. On the one hand, $f_\mathrm{I}$ is higher for M-dwarf exoplanets due to the fact that the tidally locked dayside does not experience nights and is bathed in continual illumination \citep{GW17}. On the other hand, $f_\mathrm{CL}$ is reduced as a consequence of the potentially higher atmospheric absorptivity and increased cloud clover, among other factors \citep{KWR93,YCA13,KWH16}. Therefore, by specifying $f_A$ to be constant (as we did in the previous paragraph), we are effectively already ensuring that the atmospheric attenuation experienced by M-dwarf Earth-analogs is a few times higher than their counterparts around Sun-like stars, in line with prior theoretical predictions.
We have verified that quantities such as the compensation depth, the critical depth, and the net primary productivity (all of which are defined later) decrease by a factor of $\lesssim 2$, \emph{ceteris paribus}, even up to nearly an order of magnitude increase in the degree of atmospheric attenuation. Lastly, we remark that the interplay of all the aforementioned variables is further complicated by the presence of climate feedback mechanisms as well as the atmospheric and surface compositions that may collectively yield different values from one climate model to another, even for the same setup, which makes estimating them challenging \citep{ZSDS,SBJ16,CS16}. Assessing the properties of aquatic photosynthesis is a complicated task, as elucidated in Sec. \ref{SecModLim}, owing to which our goal herein is to primarily focus on understanding how the salient characteristics vary as a function of key \emph{physical} parameters that can be constrained by present-day or forthcoming observations \citep{FAD18}.
Given the photon flux density, denoted by $\mathcal{N}_0(\lambda)$, at the surface, we are in a position to calculate the photon flux $\mathcal{F}$ at a depth $z$ below the surface of the ocean. This quantity is found by convolving $\mathcal{N}_0(\lambda)$ and the vertical attenuation coefficient $K$ in the oceans, thus yielding
\begin{equation}\label{IntRel}
\mathcal{F}(z) \approx \int_{\lambda_\mathrm{min}}^{\lambda_\mathrm{max}} \mathcal{N}_0(\lambda) \exp\left(-K z\right)\,d\lambda.
\end{equation}
It should be noted that $\mathcal{N}_0(\lambda)$ is equal to $\mathcal{N}_\mathrm{max}$ or $\mathcal{N}_\mathrm{avg}$, depending on what scenario we wish to analyze. Now, let us turn our attention to $K$, which we shall rewrite as $K = K_W + K_I$ \citep[Chapter 9.5]{Kirk11}. The first term ($K_W$) is the attenuation coefficient associated with water whereas $K_I$ accounts for the attenuation stemming from impurities as well as biota. In order to tackle $K_W$, we begin by noting that it has been tabulated as a function of $\lambda$ in many sources \citep{HQ73,SB81,KLC93,LQF99,MGC07}. Based on the data taken from \citet[Table 3]{PF97}, which is consistent with later studies over the PAR range \citep{LWV15}, the following simple exponential fit was employed by \citet{LL20} across the PAR range:
\begin{equation}
K_{W,22} \approx 1.4 \times 10^{-5}\,\mathrm{m}^{-1}\,\exp\left(\lambda \cdot 1.54 \times 10^7\,\mathrm{m}^{-1}\right),
\end{equation}
although it is essential to recognize that the data had been collected at $22$ $^\circ$C \citep[Table 3]{PF97}. In general, $K_W$ is not only dependent on $\lambda$ but also on $T_W$. The ocean temperature in turn varies with depth, but it only changes by a few K in the zone where the bulk of photosynthesis occurs \citep{Paw13}. Hence, we shall treat $T_W$ as being roughly constant, thereby enabling us to model it as a free parameter in the model. In order to account for the dependence on $T_W$, we employ the linear temperature scaling that has been confirmed by a number of empirical studies \citep{LMQ01,STZ06,RRM14}, thereupon enabling us to write
\begin{equation}
K_W(T_W,\lambda) = K_{W,22} + \alpha(\lambda) \Delta T_{22},
\end{equation}
where $\alpha(\lambda)$ represents the wavelength-dependent temperature coefficient (units of m$^{-1}$ K$^{-1}$), and $\Delta T_{22} = T_W - 295$ is a measure of the deviation from the standard water temperature of $22$ $^\circ$C employed in calculating $K_{W,22}$.
For the PAR range considered herein, the second term on the right-hand-side of the above expression is always much smaller than the first term provided that $\Delta T_{22}$ is $\mathcal{O}(10)$ K. This condition arises because $\alpha$ is nearly zero across the PAR range ($\lesssim 10^{-3}$ m$^{-1}$ K$^{-1}$), as can be verified by comparing \citet[Figure 5]{RRM14} and \citet[Table 1]{STZ06} with \citet[Table 3]{PF97}. Apart from the temperature dependence, we remark that $K_W$ also exhibits a dependence on the salinity, which is naturally expected to vary from one ocean to another \citep{CS16,OJA}. However, we have implicitly held the salinity fixed to that of the global value of Earth's oceans. More importantly, the salinity dependence is weak across the PAR range, as shown by experimental studies \citep{STZ06,RRM14}.
\begin{figure*}
$$
\begin{array}{cc}
\includegraphics[width=7.3cm]{IMaxG.pdf} &
\includegraphics[width=7.5cm]{IMaxM.pdf}\\
\end{array}
$$
\caption{In both panels, the photon flux in units of the compensation flux is shown as a function of the ocean depth for the idealized case described in Sec. \ref{SSecPrelim}; the compensation flux specifies the photon flux at which net growth of the organism is not feasible. The various curves correspond to different choices of the global ocean temperature, and the intersection points of the various curves with the dashed horizontal line yield the compensation depths. The left and right panels correspond to Planets G \& M, respectively, introduced in Sec. \ref{SSecPrelim}.}
\label{FigCDepG}
\end{figure*}
\begin{figure*}
$$
\begin{array}{cc}
\includegraphics[width=7.5cm]{IRealG.pdf} &
\includegraphics[width=7.4cm]{IRealM.pdf}\\
\end{array}
$$
\caption{In both panels, the photon flux in units of the compensation flux is shown as a function of the ocean depth for the global average case described in Sec. \ref{SSecPrelim}; the compensation flux specifies the photon flux at which net growth of the organism is not feasible. The various curves correspond to different choices of the global ocean temperature, and the intersection points of the various curves with the dashed horizontal line yield the compensation depths. The left and right panels correspond to Planets G \& M, respectively, introduced in Sec. \ref{SSecPrelim}; for Planet M, the global average encapsulates the properties on the dayside.}
\label{FigCDepM}
\end{figure*}
Next, we turn our attention to the other attenuation coefficient $K_I$. If one considers the case with pure water, i.e., amounting to $K_I \rightarrow 0$, it follows that $\mathcal{F}$ is maximized for a given depth \emph{ceteris paribus}. In a more realistic setting, however, we shall adopt $K_I \approx 0.08$ m$^{-1}$ to maintain consistency with the typical diffuse attenuation coefficient in the PAR range for Earth's oceans \citep[Chapter 4.2]{SG06}; this choice is also compatible with the coefficients deduced from measurements of clear ocean waters \citep{LDM05,SHG13,SW15}. In actuality, $K_I$ will also be a function of wavelength and temperature, but the exact dependence is dictated by several complex oceanographic and biological (e.g., density of photoautotrophs) factors \citep{MGC07}, owing to which we have opted to work with a constant value. The wavelength variation, in particular, is rather weak because $K_I$ changes by only a factor of $\sim 2$ across the PAR range \citep[Figure 4]{MM01}.
At this stage, it is worth recapitulating the two broad scenarios we shall be considering. The first corresponds to the so-called idealized case where the star is located at the substellar point, and there is no attenuation because of the atmosphere and oceanic impurities. In other words, we employ $\mathcal{N}_0(\lambda) = \mathcal{N}_\mathrm{max}$ and $K = K_W$, and introduce the superscript ``I'' (for idealized). This outcome was studied extensively by \citet{RLR18}, albeit with an exclusive focus on Earth and Proxima b. The second case accounts for time-averaged stellar flux and the existence of biological attenuation. Here, we select $\mathcal{N}_0(\lambda) = \mathcal{N}_\mathrm{avg}$, $K = K_W + K_I$ and the non-zero value of $K_I$ defined in the prior paragraph, and label it using the superscript ``A'', to wit, the ``global average'' case. For each of these two scenarios, we consider two different Earth-analogs (Planets G \& M) delineated at the beginning of this section. For Planet M, however, the ``global average'' refers to the \emph{dayside} parameters, e.g., the variable $T_W$ represents the average temperature on the dayside of the tidally locked M-dwarf exoplanet.
\section{Characteristics of aquatic biospheres}\label{SecChar}
In this section, we examine how certain the salient properties of aquatic biospheres depend on the ocean temperature; in some cases, we investigate the joint dependence on stellar and ocean temperatures. Before embarking on the discussion, we will define the quantities of interest that appear herein; for a historical treatment, we defer to \citet{Mi12,BB18}.
The first concept that we introduce from biological oceanography is the euphotic zone depth: the location where the photon flux becomes $1\%$ of its surface value. As one can see from the definition, it is divorced from biological properties for the most part. The euphotic zone depth is commonly interpreted as a measure of the photosynthesis zone on Earth, but it does not constitute a reliable metric in actuality \citep{Ban04,MLV14}. Before proceeding ahead, we note that the depth of the euphotic zone decreases from $\sim 10$-$100$ m for Earth-analogs orbiting Sun-like stars to $\mathcal{O}(1)$ m for tidally locked late-type M-dwarf exoplanets \citep{RLR18,Kal19,LL20}.
Next, we consider the compensation depth ($\mathcal{Z}_\mathrm{CO}$), which is determined by calculating the location at which $\mathcal{F}(z)$ is equal to the compensation flux ($\mathcal{F}_C$). The latter is roughly defined as the flux at which the rate of growth via photosynthesis becomes equal to the rate of respiration \citep{GG27,MO28}; in other words, at this depth, the net growth rate of the photoautotroph under consideration is equal to zero at $\mathcal{Z}_\mathrm{CO}$. As per the definition, the compensation depth is regulated by $\mathcal{F}(z)$, which in turn, \emph{inter alia}, depends on the parameter $f_\mathrm{I}$ introduced in Sec. \ref{SSecPrelim}. Here, it is important to appreciate that $f_\mathrm{I}$ is \emph{different} for M- and G-type exoplanets \citep{LL20}, due to the fact the dayside of the former receives permanent illumination when tidally locked (amounting to higher $f_\mathrm{I}$ broadly speaking). Although $f_\mathrm{I}$ functions as a fudge factor to an extent, we acknowledge that it does not fully capture the distinct spatiotemporal differences in light distribution, or oceanic properties like nutrient upwelling \citep{LL18,OJA}, associated with tidally locked M-dwarf exoplanets.
The other quantity of interest is the critical depth ($\mathcal{Z}_\mathrm{CR}$), which was elucidated by \citet{GB35} and placed on a quantitative footing by \citet{Ril46} and \citet{Sver53}. It can be envisioned as the integrated (i.e., global) version of the compensation depth. The critical depth is the location where the \emph{vertically integrated} net growth rate becomes zero, i.e., the integrated photosynthetic growth rate is equal to the integrated depletion rate arising from respiration and other factors \citep[Chapter 3]{ML06}. The critical depth is relevant from an observational standpoint because it may regulate the feasibility of phytoplankton blooms \citep{FR07}, which have been posited as an example of temporal biosignatures \citep{LL18,SKP18}. If the ocean mixed layer depth is greater than the critical layer depth,\footnote{As the name indicates, the mixed layer refers to the region of the ocean that is characterized by nearly uniform characteristics (e.g., temperature and salinity), and is governed by the vertical potential density gradient \citep{Kirk11,Mid19}.} the initiation of phytoplankton blooms is rendered unlikely, and vice-versa \citep[pg. 94]{ML06}. Although the critical depth concept remains influential and useful to this day \citep{NS91,OIE,SDY02,Chi11,Kirk11,FMA14,SJB15}, it has been subjected to some criticism \citep{SP90,Behr10,BB18}.
Thus, broadly speaking, the compensation depth and the critical depth represent important concepts inasmuch as exo-oceanography is concerned because they enable us to gauge the depths at which photosynthetic organisms can exist and/or give rise to tangible biosignatures \citep{SJF42,Sver53}. We refer to \citet[Figure 9.5]{FR07} for a schematic overview of these two quantities along with the euphotic zone depth.
\subsection{Compensation depth}\label{SSecCoD}
The key point worth appreciating when it comes to the compensation depth is that the compensation flux ($\mathcal{F}_C$) is \emph{not} constant even for a given organism because it is intrinsically temperature-dependent. Thus, our chief objective is to find a suitable function that will adequately describe the behavior of $\mathcal{F}_C$ with respect to $T_W$.
In the classical model for the compensation flux, it is proportional to $\Gamma_R/\Gamma_P$---see \citet{Ril46}, \citet{Sver53}, \citet[Equation 2]{SDY02} and \citet[Chapter 3]{ML06}---where $\Gamma_R$ and $\Gamma_P$ signify the rates of respiration and photosynthesis, respectively. Thus, if we know how these rates vary with temperature, one can duly formulate the expression for $\mathcal{F}_C$. The temperature dependence of these rates is subject to uncertainty and many different fitting functions have been considered. However, both the well-known metabolic theory of ecology \citep{GBW01,BGA04,DPS11,BCC15,AC17} and recent analyses of empirical data from Earth's oceans \citep[Figure 3.3]{Kir18} predict that these rates are fairly well described by the classic Arrhenius equation.
Hence, by utilizing the respective activation energies for these two processes \citep[Section 3]{RD12}, we end up with
\begin{equation}\label{GamRat}
\frac{\Gamma_R}{\Gamma_P} \propto \exp\left(-\frac{\Delta E}{k_B T_W}\right),
\end{equation}
where $\Delta E \approx 0.34$ eV constitutes the ``net'' activation energy, i.e., the difference between the corresponding activation energies \citep{YDC12}. An important point worth noting is that the above ansatz for $\Gamma_R/\Gamma_P$ is monotonically increasing with temperature. It is very unlikely that this behavior would be obeyed \emph{ad infinitum} because the Arrhenius equation breaks down beyond a certain temperature \citep{King09,Ang09,Sch15}. The issue, however, is that the optimum temperature, after which the trend reverses, is species-dependent \citep{AC17,CMB18}, and is modulated to a substantial degree by the environment(s) of the putative organisms. We will restrict ourselves to $273 < T_W < 323$ K, as this interval roughly overlaps with the temperature range of $280 < T_W < 322$ K studied in \citet[pg. 724]{BJ20}. In that study, diverse marine phytoplankton were shown to obey (\ref{GamRat}) for a broad thermal range.
By utilizing the above relationships, the temperature dependence of $\mathcal{F}_C$ is modeled as
\begin{equation}\label{Fcomp}
\mathcal{F}_C \approx 10\,\mathrm{\mu mol\,m^{-2}\,s^{-1}}\, \mathcal{G}(T_W),
\end{equation}
where we have introduced the auxiliary function
\begin{equation}\label{AuxF}
\mathcal{G}(T_W) \equiv \exp\left[13.6\left(1-\frac{T_0}{T_W}\right)\right],
\end{equation}
with $T_0 \approx 289$ K representing the global surface temperature of Earth's oceans.\footnote{\url{https://www.ncdc.noaa.gov/sotc/global/201913}} The constant of proportionality in (\ref{Fcomp}) has been chosen as it represents the compensation flux for phytoplankton in Earth's oceans within a factor of $\sim 2$ \citep{NS91,Mar04,ML06,RD10}. By solving for $\mathcal{F}(z) = \mathcal{F}_C$, we are now equipped to calculate the compensation depth $\mathcal{Z}_\mathrm{CO}$ as a function of both the stellar temperature and ocean temperature.
In Fig. \ref{FigCDepG}, the photon flux normalized by the compensation flux is plotted as a function of the depth $z$ for the idealized case delineated in Sec. \ref{SSecPrelim}, where the star is at the substellar point and the attenuation in water is assumed to be minimal. The left panel corresponds to Planet G, while the right panel depicts the results for Planet M. By inspecting both panels, we find that $\mathcal{Z}_\mathrm{CO}$ decreases with the temperature along expected lines. The physical reason for this trend is that the increase in the rate of respiration outpaces that of photosynthesis when the temperature is elevated, thereby ensuring that the location at which the two processes balance each other is shifted closer to the surface of the ocean, i.e., leading to a reduction in $\mathcal{Z}_\mathrm{CO}$.
\begin{figure*}
$$
\begin{array}{cc}
\includegraphics[width=7.1cm]{SurfFluxST.pdf} &
\includegraphics[width=7.5cm]{SurfFluxOT.pdf}\\
\end{array}
$$
\caption{In both panels, the ratio of the photon flux at the surface to that of the compensation flux (denoted by $\zeta$) is depicted. Regions lying below the horizontal dashed line are relatively unlikely to host Earth-like biota in the oceans. Left panel: variation of $\zeta$ with stellar temperature (in K) for different ocean temperatures. Right panel: variation of $\zeta$ with ocean temperature (in $^\circ$C) for different stellar temperatures.}
\label{FigSurf}
\end{figure*}
We observe that the ocean temperature exerts a fairly significant effect on the magnitude of $\mathcal{Z}_\mathrm{CO}$ for both worlds. As far as Planet G (orbiting a solar twin) is concerned, the compensation depth changes from $\mathcal{Z}_\mathrm{CO}^{(I)} \approx 300$ m at $T_W = 5$ $^\circ$C to $\mathcal{Z}_\mathrm{CO}^{(I)} \approx 130$ m at $T_W = 45$ $^\circ$C. On the other hand, when we consider Planet M, situated around a late-type M-dwarf closely resembling TRAPPIST-1, the compensation depth morphs from $\mathcal{Z}_\mathrm{CO}^{(I)} \approx 26.5$ m at $T_W = 5$ $^\circ$C to $\mathcal{Z}_\mathrm{CO}^{(I)} \approx 3.5$ m at $T_W = 45$ $^\circ$C. Hence, for the idealized case studied in this figure, we predict that the ocean temperature might cause $\mathcal{Z}_\mathrm{CO}$ to change by nearly an order of magnitude for Planet M; the variation associated with Planet G is smaller, but still non-negligible.
Fig. \ref{FigCDepM} is analogous to that of Fig. \ref{FigCDepG}, except that we consider the so-called global average case described at the end of Sec. \ref{SSecPrelim} in lieu of the idealized scenario. When it comes to Planet G (left panel), the compensation depth evolves from $\mathcal{Z}_\mathrm{CO}^{(A)} \approx 24$ m at $T_W = 5$ $^\circ$C to $\mathcal{Z}_\mathrm{CO}^{(A)} \approx 8.5$ m at $T_W = 45$ $^\circ$C. However, a striking result is manifested vis-\`a-vis Planet M (right panel). At $T_W = 5$ $^\circ$C, we obtain $\mathcal{Z}_\mathrm{CO}^{(A)} \approx 3$ m, but we end up with $\mathcal{Z}_\mathrm{CO}^{(A)} = 0$ at $T_W = 45$ $^\circ$C. The null value arises because the temperature elevates the compensation point to such an extent that it overshoots the incident photon flux at the ocean's surface.\footnote{We reiterate that our analysis deals with Earth-like biota, and the results are not necessarily applicable to putative oxygenic photoautotrophs in the oceans of M-dwarf exoplanets.} In fact, we determine that $\mathcal{F}_0 \equiv \mathcal{F}^{(A)}(z=0) < \mathcal{F}_C$ is fulfilled when $T_W > 24$ $^\circ$C, implying that ocean temperatures above this value appear to be relatively unsuitable for supporting phytoplankton-like biota on tidally locked Earth-analogs orbiting stars akin to TRAPPIST-1.
Motivated by the above finding, we define $\zeta \equiv \mathcal{F}_0/\mathcal{F}_C$ and study the regimes in which $\zeta < 1$ is valid. This criterion enables us to gauge the conditions under which Earth-like oxygenic photoautotrophs may have a low likelihood of existing. We only tackle the global average case herein, as it permits $\zeta < 1$ to occur in the parameter space. From examining Fig. \ref{FigSurf}, where the results are depicted, it is apparent that some tidally locked late-type M-dwarf exoplanets might be incapable of hosting phytoplankton-like biota. In particular, for the upper bound of $T_W = 50$ $^\circ$C, we surmise that stars with $T < 3150$ K can be ruled out in this category. Thus, if the oceans are sufficiently warm, tidally locked Earth-analogs around late-type M-dwarfs could encounter difficulties in sustaining marine photosynthetic organisms analogous to modern Earth.
Lastly, before proceeding further, there is one other point worth mentioning. As the depth of the photosynthesis zone grows more shallower, whether it be due to oceanic temperature or stellar spectral type, the photoautotrophs are expected to live closer to the surface. In doing so, they incur a greater risk of damage by ultraviolet radiation and energetic particles from flares and superflares, the latter of which could deposit high doses \citep{LL17,YMA,At20}. However, experiments and numerical models suggest that hazes (in)organic films \citep{CM98,EV18,LL19}, along with biogenic ultraviolet screening compounds and evolutionary adaptations \citep{CK99,ALO}, may suffice to protect them.
\subsection{Critical depth}
In order to calculate the critical depth ($\mathcal{Z}_\mathrm{CR}$), a number of different formulae have been delineated in the literature \citep{Sver53,Kirk11,Mid19}. Most of the simpler models reduce to \citep[equation 9.7]{FR07}:
\begin{equation}
K \mathcal{Z}_\mathrm{CR} \approx \frac{\Gamma_P}{\Gamma_R},
\end{equation}
but they are correct only in the limiting case of $K = \mathrm{const}$, which is manifestly invalid. The generalization of the above equation was adumbrated in \citet{LL20}, who eventually obtained
\begin{equation}\label{CritD}
\mathcal{Z}_\mathrm{CR} \approx \left(\frac{\Gamma_R}{\Gamma_P}\right)^{-1} \frac{\int_{\lambda_\mathrm{min}}^{\lambda_\mathrm{max}} \left[\mathcal{N}_0(\lambda)/K(\lambda)\right]\,d\lambda}{\int_{\lambda_\mathrm{min}}^{\lambda_\mathrm{max}} \mathcal{N}_0(\lambda)\,d\lambda}.
\end{equation}
It is, however, necessary to recognize that $\Gamma_R/\Gamma_P$ has an intrinsic temperature dependence, as seen from (\ref{GamRat}). Hence, we combine (\ref{CritD}) with (\ref{GamRat}), thereby yielding
\begin{equation}\label{CritF}
\mathcal{Z}_\mathrm{CR} \approx \frac{3.36 \times 10^{-2}}{\mathcal{G}(T_W)} \frac{\int_{\lambda_\mathrm{min}}^{\lambda_\mathrm{max}} \left[\mathcal{N}_0(\lambda)/K(\lambda)\right]\,d\lambda}{\int_{\lambda_\mathrm{min}}^{\lambda_\mathrm{max}} \mathcal{N}_0(\lambda)\,d\lambda},
\end{equation}
where the normalization has been adopted based on the global value for phytoplankton in Earth's oceans \citep[Chapter 4.3]{SG06}. The parameters pertaining to the ``A'' scenario are adopted for the sake of comparison with prior empirical studies.
The temperature dependence of the critical depth is illustrated in Fig. \ref{FigCritD}. Two points that emerge from scrutinizing this figure. From the left panel, we notice that the dependence on the stellar temperature is weak at any given ocean temperature. This result is consistent with \citet{LL20}, and is mostly attributable to the fact that net growth primarily occurs in the upper layers and thus compensates for the regions with $z > \mathcal{Z}_\mathrm{CO}$. However, when it comes to the ocean temperature, a much stronger variation of $\mathcal{Z}_\mathrm{CR}$ is discerned. As we cover the entire ocean temperature range considered herein, we find that $\mathcal{Z}_\mathrm{CR}$ changes by nearly an order of magnitude for any given stellar temperature (right panel). For instance, after we specify $T = T_\odot$, the critical depth evolves from $\mathcal{Z}_\mathrm{CR}^{(A)} \approx 416$ m at $T_W \approx 0$ $^\circ$C to $\mathcal{Z}_\mathrm{CR}^{(A)} \approx 45$ m at $T_W \approx 50$ $^\circ$C.
Therefore, it is conceivable that the ocean temperature plays a major role in regulating the critical depth on other worlds. In turn, this development suggests that $T_W$ also acts as a key determinant of phenomena analogous to phytoplankton blooms, which may constitute viable temporal biosignatures as noted earlier.
\begin{figure*}
$$
\begin{array}{cc}
\includegraphics[width=7.3cm]{CritDST.pdf} &
\includegraphics[width=7.5cm]{CritDOT.pdf}\\
\end{array}
$$
\caption{In both panels, the critical depth ($\mathcal{Z}_\mathrm{CR}$)---to wit, the location where the vertically integrated net growth rate is zero---is plotted. Left panel: variation of $\mathcal{Z}_\mathrm{CR}$ with stellar temperature (in K) depicted for different ocean temperatures. Right panel: variation of $\mathcal{Z}_\mathrm{CR}$ with ocean temperature (in $^\circ$C) illustrated for different stellar temperatures.}
\label{FigCritD}
\end{figure*}
\subsection{Net primary productivity}\label{SSecNPP}
\begin{figure}
\includegraphics[width=7.5cm]{NPP.pdf} \\
\caption{The oceanic NPP relative to that of modern Earth as a function of the ocean temperature for Planet G (stellar temperature of $T = 5780$ K) and Planet M (stellar temperature of $T = 2500$ K). }
\label{FigNPP}
\end{figure}
\begin{table*}
\caption{Net primary productivity for the Earth-analogs as a function of the mean oceanic temperature}
\label{TabNPP}
\begin{tabular}{|c|c|c|}
\hline
Ocean temperature ($^\circ$ C) & Relative NPP of Planet G & Relative NPP of Planet M \tabularnewline
\hline
\hline
$5$ & $0.4$ & $9 \times 10^{-3}$\tabularnewline
\hline
$10$ & $0.6$ & $10^{-2}$\tabularnewline
\hline
$15$ & $0.9$ & $10^{-2}$\tabularnewline
\hline
$20$ & $1.4$ & $8 \times 10^{-3}$\tabularnewline
\hline
$25$ & $2.0$ & $0$\tabularnewline
\hline
$30$ & $2.7$ & $0$\tabularnewline
\hline
$35$ & $1.2$ & $0$\tabularnewline
\hline
$40$ & $6 \times 10^{-2}$ & $0$\tabularnewline
\hline
$45$ & $2 \times 10^{-3}$ & $0$\tabularnewline
\hline
\end{tabular}
\medskip
{\bf Notes:} The NPP is expressed in terms of the temporally averaged value associated with modern Earth's oceans, namely, $1.5 \times 10^{-2}\,\mathrm{g\, C\, m^{-2}\, h^{-1}}$. The NPP for these two Earth-analogs is calculated by deploying (\ref{NPPRes}). Planet G orbits a solar twin whereas Planet M is situated near a late-type M-dwarf akin to TRAPPIST-1; the other properties of the two planets are otherwise identical.
\end{table*}
The NPP is arguably one of the most crucial and informative property of a biosphere as it quantifies the net amount of organic carbon synthesized via biological pathways after accounting for losses dues to respiration and other factors; we will express our results in units of g C m$^{-2}$ h$^{-1}$ for the NPP. The NPP is a reliable measure of the amount of organic C generated via photosynthesis, as the latter constitutes the dominant carbon fixation pathway \citep{Berg11,Knoll,Jud17,BPM18}. A wide spectrum of models have been developed to model NPP, and comprehensive reviews can be found in \citet{BF97} and \citet[Chapter 9]{FR07}.
We make use of \citet[Equation 3]{FB98} to calculate the NPP, because this outwardly simple expression accounts for a number of environmental factors:
\begin{equation}\label{NPPDef}
\mathrm{NPP} = C_\mathrm{sur} \cdot \mathcal{Z}_\mathrm{CO}^{(A)} \cdot f(\mathrm{PAR}) \cdot P_\mathrm{opt}(T_W),
\end{equation}
where $C_\mathrm{sur}$ is the chlorophyll concentration at the surface, $f(\mathrm{PAR})$ embodies the fraction of the water column up to $\mathcal{Z}_\mathrm{CO}^{(A)}$ where photosynthesis is light saturated, and $P_\mathrm{opt}(T_W)$ is the optimal carbon fixation rate. There exists, however, an inherent crucial subtlety that needs to be spelt out here. In canonical versions of the above formula, $\mathcal{Z}_\mathrm{CO}^{(A)}$ is replaced by the euphotic zone depth. However, as noted in \citet[pg. 237]{FB98}, the proper variable that ought to be deployed is the depth of the zone where positive NPP is feasible, which is congruent with the definition of the compensation depth. On Earth, the euphotic zone depth \citep{LWK07} and the compensation depth \citep{SJF42,Mid19} are roughly equal to one another, but the same relationship is not necessarily valid \emph{a priori} for other worlds; even on Earth, the reliability of the euphotic zone as a measure of the photosynthesis zone has been called into question \citep{Ban04,MLV14}.
The NPP will depend not only on the stellar and ocean temperatures but also on inherent biological factors such as $C_\mathrm{sur}$ that are spatially and temporally very heterogeneous. As the goal of the paper is to construct heuristic global estimates, we rewrite (\ref{NPPDef}) so that it yields the average global value for the Earth at $T = T_\odot$ and $T_W = T_0$ (i.e., the parameters for Earth). By adopting the normalization from \citet{FB98},\footnote{We note that some subsequent estimates for the oceanic NPP have revised the classic analysis of \citet{FB98} by $\mathcal{O}(10\%)$ \citep{WBS08}, but this has a minimal impact on both our subsequent qualitative and quantitative results.} we obtain
\begin{eqnarray}\label{NPPRes}
&& \mathrm{NPP} \approx 1.5 \times 10^{-2}\,\mathrm{g\, C\, m^{-2}\, h^{-1}} \left(\frac{\mathcal{Z}_\mathrm{CO}^{(A)}}{\mathcal{Z}_0}\right) \nonumber \\
&& \hspace{0.5in} \times \,\mathcal{P}(T_W) \left(\frac{\mathcal{D}}{0.5}\right) \left(\frac{G(T)}{G(T_\odot)}\right),
\end{eqnarray}
where $\mathcal{Z}_0 \approx 19$ m represents the compensation depth calculated at the fiducial ocean temperature of $T_0$ using the methodology in Sec. \ref{SSecCoD}, while $\mathcal{D}$ denotes the fraction of time that a given location receives stellar illumination. For planets like Earth, we expect $\mathcal{D} \approx 0.5$ (i.e., equipartition of day and night) whereas tidally locked planets ought to have $\mathcal{D} \approx 1$ on the day side because they receive stellar radiation \emph{in perpetuo}. The auxiliary functions $G(T)$ and $\mathcal{P}(T_W)$ are defined as follows:
\begin{equation}\label{Gdef}
G(T) = \frac{\mathcal{F}_0}{\mathcal{F}_0 + \mathcal{F}_S},
\end{equation}
where $\mathcal{F}_S \approx 1.1 \times 10^3$ $\mu$mol m$^{-2}$ s$^{-1}$, and the stellar temperature is implicitly present via $\mathcal{F}_0$.
\begin{eqnarray}
&& \mathcal{P}(T_W) = \left[ \frac{1 + \exp\left[\frac{E_h}{k_B}\left(\frac{1}{T_h} - \frac{1}{T_0}\right)\right]}{1 + \exp\left[\frac{E_h}{k_B}\left(\frac{1}{T_h} - \frac{1}{T_W}\right)\right]}\right] \nonumber \\
&& \hspace{0.6in} \times \exp\left[\frac{E_a}{k_B}\left(\frac{1}{T_0} - \frac{1}{T_W}\right)\right],
\end{eqnarray}
where $E_a \approx 0.74$ eV, $E_h \approx 6.10$ eV and $T_h \approx 34$ $^\circ$C are adopted for our putative biota from \citet[pg. 726]{BJ20}.\footnote{It goes without saying that all of these parameters exhibit substantive variation across species. For instance, the thermal performance curves for certain species of marine phytoplankton reveal optimal temperatures of $\sim 20$-$25$ $^\circ$C \citep{BRAF}, in which case $T_h$ is lowered by several $^\circ$C.} Here, we have constructed (\ref{NPPRes}) and (\ref{Gdef}) based on \citet[Section 2.4]{BBS05} and \citet[Equation 10]{BF97}, but one point of divergence is that a modified Sharpe–Schoolfield equation \citep{SD77,SSM81} was utilized as a proxy for $P_\mathrm{opt}(T_W)$, following \citet{BYD19,BJ20} in lieu of \citet[Equation 11]{BF97}, as the latter becomes invalid for $T_W > 30$ $^\circ$C. The precise expression for $P_\mathrm{opt}(T_W)$ for phytoplankton is challenging to accurately pin down, owing to the panoply of expressions used to model phytoplankton growth \citep{GMS17}. In consequence, a diverse array of functions, some exhibiting exactly opposite trends with temperature, have been employed for this purpose \citep[Figure 4]{MBPF}; see also \citet{GMS17}. Hence, the ensuing results should be interpreted with due caution.
We have presented the NPP for the two Earth-analogs (Planet G and Planet M) in Table \ref{TabNPP} and Figure \ref{FigNPP}, which were calculated by using the global average case as seen from (\ref{NPPDef}). There are several interesting results that emerge from inspecting these two items. We begin by considering Planet G (orbiting a solar twin) to gauge the role of $T_W$. We notice that the NPP increases with ocean temperature until $\sim 30$ $^\circ$C, but the growth is relatively modest. It is primarily driven by the rise in the rate of carbon fixation, as encapsulated by $\mathcal{P}(T_W)$, with temperature in this regime. In some controlled experiments and modeling where the temperature was steadily elevated, the photosynthetic capacity has been found to increase up to a point \citep{LBH14,PYD16,SBB17}.\footnote{It is important to recognize, however, that the variation of NPP with temperature is subject to much uncertainty due to the large number of coupled variables and nonlinear feedback mechanisms \citep{TO11,LVG15}.} As per our simple model, once the peak temperature of the thermal performance curve is exceeded ($T_\mathrm{pk}$), the rate of carbon fixation falls sharply thereafter, and consequently drives the steep decline in NPP when $T_W > 35$ $^\circ$C.
Now, we turn our attention to Planet M around a late-type M-dwarf similar to TRAPPIST-1. For any fixed temperature, say $T_W = 5$ $^\circ$C, we notice that the NPP is lower than Planet G by roughly two orders of magnitude. The reasons for the diminished NPP are twofold: (i) the compensation depth is greatly reduced as pointed out in Sec. \ref{SSecCoD}, and (ii) the flux of PAR is corresponding lower at the surface, thereby making the last term on the right-hand-side of (\ref{NPPRes}) smaller than unity. The next major feature we notice is that the NPP vanishes at $T_W \sim 24$ $^\circ$C. This result is a direct consequence of the fact that the compensation depth becomes zero above a threshold temperature for reasons explained in Sec. \ref{SSecCoD}. Hence, tidally locked exoplanets around late-type M-dwarfs may evince a low likelihood of large-scale carbon fixation by phytoplankton-like biota. Needless to say, the NPP is not anticipated to be zero \emph{sensu stricto}, because anoxygenic photoautotrophs are capable of carbon fixation by definition \citep{Ko07,SB13}, and so are many microbial taxa in the deep biosphere \citep{OSK11,EBC12,MP14,CPS17}.
\section{Limitations of the model}\label{SecModLim}
It is worth emphasizing at the outset that the productivity of biospheres is constrained by a number of factors including water, electron donors, temperature, PAR flux and nutrients \citep{LL21}. Our analysis tackles the modulation of the productivity of biospheres by PAR and ambient ocean temperature \emph{ceteris paribus}. In doing so, we follow the likes of \citet{LCPH,LM19} in setting aside the constraints imposed by the access to nutrients and some of the other variables.
In the case of Earth's terrestrial (land-based) NPP, the NPP for $> 80\%$ of the area is limited by water and temperature \citep{CR98}. In contrast, Earth's oceanic NPP---both the globally averaged value and the spatiotemporal variations---is governed by the prevalence of nutrients, especially the ultimate limiting nutrient phosphate \citep{Ty99,Fil08,SB13}. Ocean planets, in particular, may be impacted due to their potentially lower rates of weathering and delivery of nutrients to the oceans \citep{WP13,LM18,KF18,LiMa}. The key caveat in this paper, therefore, is that the oceanic NPP is not constrained by nutrients, but is instead regulated the two factors adumbrated in the preceding paragraph. The ensuing results might comprise upper bounds for the NPP because the abundance and distribution of nutrients could introduce additional limits.
As we shall demonstrate in Sec. \ref{SSecNPP}, the oceanic NPP for tidally locked Earth-like exoplanets around late-type M-dwarfs is severely constrained by the paucity of PAR photons, and is orders of magnitude smaller than Earth's oceanic NPP. Hence, the prior assumption might not pose a major problem for these worlds because the most dominant bottleneck on the oceanic NPP may prove to be the PAR flux. However, when it comes to Earth-analogs around Sun-like stars, PAR flux is not a major limiting factor and the thermal effects on NPP might become prominent only at high temperatures. Thus, a brief discussion of nutrient limitation and how it could impact the oceanic NPP of other worlds is apropos.
The first and foremost point that needs to be appreciated is that modeling nutrient limitation even on Earth is a complicated endeavor. The reason is that the nutrient concentration in the ocean depends on a variety of factors such as the remineralization efficiency \citep{KS17,LS18}, hydrothermal activity \citep{WFM96}, submarine weathering \citep{Hao20}, and mineral solubility in seawater \citep{Derr15}, to name just a few. Moreover, each of these quantities has fluctuated over time and witnessed shifts in magnitude, sometimes up to a factor of $\sim 10$ as may have occurred vis-\`a-vis the remineralization efficiency during the Ediacaran period \citep{LSJ20}. For these reasons, theoretical models for the biogeochemical cycles of the bioessential elements have yielded very different results \citep{Len20}; see also \citet{Hao20} for an exposition of this issue.
As the prior discussion suggests, there are numerous mechanisms that control the nutrient concentration in oceans. In consequence, it is not inconceivable that some Earth-like planets could bypass or mitigate the issue of nutrient limitation. Geological processes that have been proposed hitherto for counteracting the nutrient deficiency to varying degrees include elevated nutrient upwelling \citep{LL18,OJA,SOK20}, submarine basalt weathering \citep{SRI20} and serpentinization \citep{POL20}. As noted earlier, we will presume hereafter that nutrient abundance is not the chief limitation, perhaps via some of the above channels coming into play. To reiterate, we suppose that either photon flux and/or temperature act to throttle the productivity. We will demonstrate hereafter that these factors become exceedingly important for planets around late-type M-dwarfs and/or with high ambient ocean temperatures; in particular, the NPP might become orders of magnitude smaller relative to Earth.
In relation to the preceding points, we note that the constraints imposed by the ambient photon flux, temperature and nutrients do not act independently of one another. In fact, a multitude of experiments and field studies have established that these environmental parameters are non-linearly coupled to one another \citep{ETK16,SMS16,GMS17,TAK17,MLC18}. For instance, the value of $E_a$ introduced previously may vary significantly in some species depending on the availability of nitrogen \citep{MLC18} and the ocean temperature \citep{MBM20}. While such effects are indubitably important, they are not well understood even on Earth and exhibit considerable intra- and inter-species variability. Hence, given that the implicit goal of this paper was to construct heuristic models that provide rough estimates for future observations and modeling, we have not taken these subtle processes into account.
Lastly, in our subsequent analysis, we will draw upon the basic physiological properties of the dominant phytoplankton species on Earth. While this line of reasoning is undoubtedly parochial, we note that Earth-based organisms are commonly used as proxies in numerous astrobiological contexts \citep{MCK17}, ranging from extremophiles and microbial ecosystems in the oceans of icy moons \citep{CH01,RM01,CKB17,MCK17,MAB19,LMa19} to the limits of complex multicellular life on exoplanets \citep{SVS17,SRO19,Li20,Ram20}. Furthermore, the choice of phytoplankton as putative biota is motivated by the fact that they are the major source of carbon fixation in the oceans of modern Earth \citep{Har86,Fal04,CKT05,Rav09,UCG10}. Hence, by utilizing the prior framework, we are now equipped to analyze the prospects for Earth-like aquatic photosynthesis on other worlds characterized by different ocean temperatures.
\section{Discussion}\label{SecDisc}
We will discuss some of the implications of our work in connection with mapping the trajectories of the Earth as well as tidally locked M-dwarf exoplanets.
\subsection{Potential future evolution of Earth}
We begin by tackling the ramifications of the preceding analysis for the Earth's aquatic biosphere, with respect to its potential future.
Before doing so, it is worth briefly highlighting the inherent spatiotemporal variability of Earth's oceanic NPP. To begin with, let us recall that a global sea surface temperature (SST) of $T_0 \approx 16$ $^\circ$C was chosen herein based on satellite data. However, in reality, the SST of Earth is characterized by distinct heterogeneity, ranging from $35$ $^\circ$C to below-freezing temperatures.\footnote{\url{https://earthobservatory.nasa.gov/global-maps/MYD28M}} Moreover, the Earth's NPP is modulated by the access to not only light and temperature (both of which are present in our model) but also nutrients \citep{BBS05}; the latter may play a crucial role as noted in Sec. \ref{SecModLim}. Collectively, these factors engender variations in the oceanic NPP across both the spatial and temporal domains \citep{WBS08}, sometimes by roughly an order of magnitude. Thus, we reiterate that our model only seeks to extract globally averaged values for the relevant variables from a heuristic standpoint.
There is a sharp downswing in NPP shortly after the peak temperature $T_\mathrm{pk}$ is attained, which becomes evident upon inspecting Fig. \ref{FigNPP}. While there are grounds for contending that $T_\mathrm{pk} \sim 30$ $^\circ$C \citep{BJ20}, this matter is admittedly not conclusively settled. Now, let us suppose that the Earth's temperature was raised by $\sim 10$ $^\circ$C abruptly. In large swathes of the ocean, it is conceivable that $T_W > T_\mathrm{pk}$, thereby triggering a sharp downswing in the NPP in these regions. In turn, given that phytoplankton are the foundation of oceanic food webs and trophic interactions \citep{BH99,Val15,Kir18}, this rapid decline in NPP ought to have adverse consequences for marine ecosystems and could thus potentially drive large-scale extinctions of marine biota.
As the Sun continues to grow brighter, the surface temperature will also increase commensurately because of the greenhouse effect until the Earth is eventually rendered uninhabitable \citep{CK92,GW12,RCOW}. Based on \citet[Section 3.1]{WT15}, we note that a global temperature of $312$ K is predicted when the solar luminosity is $1.1$ times the present-day value. By utilizing \citet[Equation 1]{Go81}, the stellar luminosity associated with this temperature is expected to occur $\sim 1.2$ Gyr in the future. It is important to note, however, that climate models do not fully agree on the critical flux at which the greenhouse state is likely to be activated, implying that a timescale of $< 1$ Gyr ought not be ruled out \citep{GRZ13,LFC13,KCK15,PSM16,WSK17}.
If we suppose that the global ocean temperature tracks the average surface temperature, the above analysis suggests that $T_W \sim 39$ $^\circ$C would occur $\sim 1.2$ Gyr hereafter. After examining Fig. \ref{FigNPP}, we find that the oceanic NPP at this $T_W$ might be $< 10\%$ of modern Earth. Due to the diminished NPP, eventual depletion of atmospheric O$_2$ is plausible for reasons adumbrated in Sec. \ref{SSecMDExo}, namely, when the sinks for oxygen outpace the sources. A decline in atmospheric O$_2$ could, in turn, drive the extinction of motile macroscopic organisms, as their long-term survival customarily necessitates oxygen levels $\sim 10\%$ of their present value \citep{CGZM,Wil09,ZC16,RPO16}.\footnote{In contrast, relatively sessile animals, such as the demosponge \emph{Halichondria panicea} \citep{MWJ14}, are capable of surviving at oxygen levels around $2$-$3$ orders of magnitude smaller than today \citep{SKG15,LK18}.} Thus, \emph{in toto}, the biosphere is unlikely to exhibit the same complexity as that of present-day Earth: this qualitative result is broadly consistent with earlier predictions by \citet{OMJ13,OMJ14}.
\subsection{Tidally locked M-dwarf exoplanets}\label{SSecMDExo}
We turn our attention to Planet M, i.e., the putative tidally locked exoplanet around a late-type M-dwarf similar to TRAPPIST-1.
It is instructive to compare our results against prior analyses of related topics. \citet[Table A9]{WoRa} calculated the oceanic NPP, albeit at a fixed depth of $10$ m using a simple model based on the photon flux, and estimated that it was $\sim 5$ times lower for an Earth-analog around an M0 star. In a similar vein, \citet{LCPH} and \citet{LM19} employed simple models for the NPP that were linearly proportional to the incident photon flux and determined that planets orbiting late-type M-dwarfs are unlikely to host biospheres with the same NPP as modern Earth and build up atmospheric O$_2$ to detectable levels. Thus, by and large, our work maintains consistency with earlier studies, but it has taken several other environmental and physiological variables into account that were missing in previous analyses.
We have previously calculated that the NPP for Planet M is, at most, only a few percent of the Earth's current oceanic NPP. Hence, because of the low NPP, unless the burial efficiency of organic carbon is unusually high, it seems likely that the flux of O$_2$ contributed by oxygenic photosynthesis will be correspondingly small. Hence, it ought to become more feasible for the sinks of atmospheric O$_2$ (e.g., continental weathering and volcanic outgassing) to dominate this source (which is a major player on Earth). The end result is that O$_2$ has a low likelihood of accumulating to detectable levels in the atmosphere \citep{CK17}.
This potential effect has two consequences in turn. First, O$_2$ has been conjectured to be an essential prerequisite for complex life insofar as metabolism is concerned \citep{Kno85,McK96,CGZM,LL21}, at least up to a certain threshold after which oxygen toxicity may set in \citep{Li20}. Hence, the evolution of complex life, and potentially technological intelligence, might be suppressed on this category of worlds. Second, and more importantly, the absence of detectable atmospheric O$_2$ or O$_3$ for the aforementioned reasons despite the existence of a biosphere is an archetypal example of a ``false negative'' that can hinder or complicate the search for extraterrestrial life \citep{ROS17,MRA18}.
\subsection{Build-up of atmospheric oxygen on ocean worlds}
It is worth quantifying the above qualitative treatment to gain further insights for ocean planets that are otherwise akin to present-day Earth. We will adopt the prescription laid out in \citet{LCPH}. We begin the analysis by noting that Earth's current oceanic NPP translates to an O$_2$ production flux of $5 \times 10^{-4}$ mol m$^{-2}$ h$^{-1}$, because the simplified reaction scheme for oxygenic photosynthesis takes the form
\begin{equation}\label{PhotOx}
\mathrm{CO_2} + 2 \mathrm{H_2O} \, \xrightarrow[\text{pigments}]{h \nu}\, \mathrm{CH_2O} + \mathrm{H_2O} + \mathrm{O_2},
\end{equation}
where CH$_2$O embodies the synthesis of organic compounds, and H$_2$O appears in both sides of the equation as reactant and product, respectively. However, only a minuscule fraction of this O$_2$ is deposited in the atmosphere, since the vast majority is consumed by respiration and oxidative decay. If we denote the burial fraction by $\phi$, the flux of O$_2$ produced is then given by
\begin{equation}
\dot{S} \sim 1.5 \times 10^{-6}\,\mathrm{mol\,m^{-2}\,h^{-1}}\,\left(\frac{\mathrm{NPP}}{\mathrm{NPP}_\oplus}\right)\left(\frac{\phi}{\phi_\oplus}\right),
\end{equation}
where $\dot{S}$ is the O$_2$ flux generated from organic carbon burial, $\mathrm{NPP}_\oplus$ is the globally averaged oceanic NPP of the Earth, and $\phi_\oplus \approx 3 \times 10^{-3}$ is the fraction of organic carbon (fixed by photosynthesis) subjected to burial on present-day Earth \citep{Holl02,LCPH}. In order for O$_2$ to build up on anoxic worlds, the above source must exceed the primary sink, namely, reducing gases arising from a mixture of surface and submarine volcanism, metamorphism and serpentinization to name a few \citep{CK17}. We introduced $\dot{D}$, the depletion flux of O$_2$ associated with reducing gases, and specify a fiducial value of $\dot{D}_\oplus \sim 1.3 \times 10^{-6}\,\mathrm{mol\,m^{-2}\,h^{-1}}$ for modern Earth \citep{CK17}. The criterion for O$_2$ accumulation in the atmosphere is thus expressible as $\dot{S} > \dot{D}$, which simplifies to
\begin{equation}\label{NPPcond}
\mathrm{NPP} > 0.9 \mathrm{NPP}_\oplus \left(\frac{\phi}{\phi_\oplus}\right)^{-1} \left(\frac{\dot{D}}{\dot{D}_\oplus}\right).
\end{equation}
Hence, the above relation suggests that the oceanic NPP must be close to its present-day value in order for the build up of atmospheric O$_2$ to potentially take place, if all other parameters are held fixed. In contrast, if the burial of carbon is very efficient or the flux of reducing gases is extremely low, even a NPP that is much smaller than that of modern Earth may support the accumulation of atmospheric O$_2$. We run into an immediate difficulty here since both $\phi$ and $\dot{D}$ are not tightly constrained for Earth-like worlds in general.
However, if we interpret Earth-analogs to include only those worlds with all geochemical parameters similar to Earth, we can make headway. In such cases, (\ref{NPPcond}) reduces to the simpler $\mathrm{NPP} > 0.9 \mathrm{NPP}_\oplus$. By comparing this criterion with Fig. \ref{FigNPP}, it is possible to deduce the conditions that permit the build up of atmospheric O$_2$ if only light and temperature constitute the sole limiting factors (cf. the next paragraph). For Earth-analogs around solar twins, we find that a temperature range of $\sim 15$-$35$ $^\circ$C might permit the build up of atmospheric O$_2$. At much higher and lower temperatures, the NPP is accordingly diminished, owing to which the reducing gases could overwhelm the O$_2$ generated from carbon burial. When we consider tidally locked Earth-analogs around late-type M-dwarfs, Fig. \ref{FigNPP} and Table \ref{TabNPP} suggest that the NPP would be only a few percent of Earth's current oceanic NPP. In that event, it might not be feasible for the accumulation of atmospheric O$_2$ to occur, as per the simplified formalism we have adopted.
\subsection{Observational tests for the future}
It is helpful to examine the prospects for testing our results by means of future observations at this juncture. A number of publications mentioned at the beginning of Sec. \ref{SSecMDExo} have already propounded strategies to gauge whether the stellar spectral type affects the NPP and the accompanying rise in atmospheric O$_2$ levels. The basic idea is to search for correlations between the spectral type of the host star on the one hand and the presence/absence of O$_2$ on the other. However, these putative correlations need to be weighed carefully because of the presence of major sources and sinks of O$_2$ not prevalent on Earth; for instance, the abiotic build-up of O$_2$ may be driven by electromagnetic radiation \citep{WP14,LB15,HSS15,KWF18} or its depletion may be effectuated by intense stellar winds and space weather events \citep{GGD17,DLMC,DHL17,DJL18,DLY18,DHL19,DJL20,ABC20}.
Therefore, we will restrict ourselves to assessing the observational implications insofar as worlds with varying ocean temperature are concerned. A scrutiny of Fig. \ref{FigNPP} reveals that a fairly steep decline in the NPP is potentially anticipated above a certain cutoff temperature. On the other hand, the surface density of chlorophyll ($C_\mathrm{sur}$) appearing in (\ref{NPPDef}) might not be affected to the same degree; in fact, we have held it fixed for the sake of simplicity. Hence, at least in principle, the detection of photoautotrophs ought to be feasible via the photosynthetic red edge \citep{STSF}, especially in the event that the organisms cover a large fraction of the surface \citep{OK19}.
If we can therefore sample enough planets and discern a critical mean ocean temperature (the surface temperature might comprise a rough proxy for $T_W$) for a particular spectral type above which no biogenic O$_2$ and O$_3$ are detected but a tangible photosynthetic red edge is identified, such a distinct correlation could provide an avenue for falsifying our hypothesis. However, we caution that this strategy is not easily implementable in the near-future because it necessitates access to a sufficiently large sample of worlds with confirmed reliable biosignatures, oceans, and surface temperature measurements \citep{Kal17,SKP18,FAD18,Mad19}.
\section{Conclusions}\label{SecConc}
In this paper, we investigated how the ambient ocean temperature $T_W$ and the spectral type of the host star may influence the characteristics of aquatic biospheres on Earth-like worlds, albeit under a set of key assumptions that were expounded in Sec. \ref{SecModLim}.
In spite of the underlying simplifications and the ensuing limitations, there are several new results that were presented in this work, of which some of the major ones are outlined below and described in more detail later.
\begin{itemize}
\item The compensation depth and critical depth is calculated as a function of the mean ocean temperature for Earth-analogs around a late-type M-dwarf (Planet M) and a Sun-like star (Planet G); it should be noted that Planet M is modeled as being tidally locked.
\item The \emph{vertically integrated} average oceanic NPP is estimated for Planet G and Planet M. In other words, the procedure for determining the oceanic NPP as a function of the spectral type and the ocean temperature was explicated.
\item The criterion for the accumulation of oxygen (O$_2$) in the atmospheres of Planets G and M is derived; this criterion is dependent not only on the spectral type but also on the ocean temperature.
\end{itemize}
We began by estimating the compensation depth and critical depth, as they serve to quantify the depths at which the net growth rate and vertically integrated net growth rate become zero, respectively. We showed that the ocean temperature has a relatively moderate influence on the compensation depth for an Earth-analog around a solar twin, as $\mathcal{Z}_\mathrm{CO}$ varies only by a factor of $< 3$. In contrast, when it comes to an Earth-like tidally locked world orbiting a late-type M-dwarf akin to TRAPPIST-1, we found that $T_W$ causes $\mathcal{Z}_\mathrm{CO}$ to vary by at least an order of magnitude. Furthermore, sufficiently warm oceans may preclude phytoplankton-like biota from existing on these worlds altogether. We calculated the critical depth and showed that it is sensitive to $T_W$, and varies by nearly an order of magnitude for the temperature range considered herein.
Next, we examined the oceanic NPP of Planet G and Planet M as a function of the ocean temperature. The NPP constitutes one of the most vital metrics for a biosphere, and it has practical consequences that are delineated in the next paragraph. This calculation entailed the estimation of several variables, of which a few have not been robustly determined as empirical functions of $T_W$. Bearing this caveat in mind, we found that the NPP on Planet G was not very sensitive to $T_W$ until it exceeded a certain threshold after which the rate of carbon fixation dropped precipitously and drove a corresponding decline in the NPP. For the case of Planet M, the NPP was determined to be $\lesssim 1\%$ that of modern Earth, primarily on account of the shallowness of the photosynthesis zone in tandem with the lower PAR fluxes. When the ocean temperatures were raised sufficiently, the conditions for phytoplankton-like biota became untenable as noted in the earlier paragraph, and consequently resulted in the NPP approaching zero.
Lastly, we analyzed the ramifications of our work in the context of our planet as well as tidally locked Earth-like exoplanets orbiting late-type M-dwarfs. We discussed how an increase of $\sim 10$ $^\circ$C in the ocean temperature, such as what is expected to happen $\lesssim 1$ Gyr in Earth's future due to the growing solar luminosity, could radically transform the aquatic biosphere of Earth-analogs around G-type stars and diminish the NPP to $< 10\%$ of the Earth's current oceanic NPP in large swathes of the oceans. In a similar vein, we surmised that the aquatic biospheres of tidally locked Earth-like worlds around late-type M-dwarfs may evince NPPs that are $\lesssim 1\%$ of our planet's oceanic NPP today. If this prediction is correct, these worlds would be unlikely to accumulate atmospheric O$_2$---except in circumstances where they have much higher carbon burial and lower outgassing of reducing gases---but signatures of life are nonetheless potentially detectable through the photosynthetic red edge if the coverage and density of photoautotrophs is high enough. We concluded our discussion by sketching rubrics which might enable the behavior of NPP with spectral type and temperature to be gauged by future observations.
\section*{Data Availability Statement}
No new data were generated or analysed in support of this research.
\section*{Acknowledgements}
We thank the reviewer for the helpful and constructive report, which helped improve the quality of the paper. This research was supported in part by the Breakthrough Prize Foundation, Harvard University's Faculty of Arts and Sciences, and the Institute for Theory and Computation (ITC) at Harvard University.
| 197a70d6c00a988f3b807ac314421cbf899d15b0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\medskip
\section{Introduction and preliminaries}
\medskip
Let $\mathcal{A}$ be the class of functions $f$ which are analytic in the open unit disc ${\mathbb D}=\{z:|z|<1\}$ of the form
\begin{equation}\label{e1}
f(z)=z+a_2z^2+a_3z^3+\cdots,
\end{equation}
and let $\mathcal{S}$ be the subclass of $\mathcal{A}$ consisting of functions that are univalent in ${\mathbb D}$.
\medskip
Important milestone in study of univalent functions was the proof of the famous Bieberbach conjecture $|a_n|\le n$ for $n\ge2$ by Lewis de Branges in 1985 \cite{Bra85}.
This ended an era, but a great many other problems concerning the coefficients $a_n$ remain open. One such is the Zalcman conjecture,
\[ |a_n^2-a_{2n-1}|\le (n-1)^2 \quad\quad (n\in {\mathbb N}, n\ge2), \]
posed in 1960 and proven in 2014 by Krushkal (\cite{krushkal}) for the whole class ${\mathcal S}$ by using complex geometry of the universal Teichm\"{u}ller space. In 1999, Ma (\cite{ma}) proposed a generalized Zalcman conjecture,
\[ |a_m a_n-a_{m+n-1}|\le (m-1)(n-1) \quad\quad (m,n\in {\mathbb N}, m\ge2, n\ge2),\]
which is still an open problem, closed by Ma for the class of starlike functions and for the class of univalent functions with real coefficients. Ravichandran and Verma in \cite{ravi} closed it for the classes of starlike and convex functions of given order and for the class of functions with bounded turning.
\medskip
In this paper we study the generalized Zalcman conjecture for the class
\[{\mathcal U}=\left\{f\in{\mathcal A}:\left| \left(\frac{z}{f(z)}\right)^2 f'(z) -1\right|<1, \, z\in{\mathbb D} \right\}.\]
Functions from this class are proven to be univalent but do not follow the traditional patterns of other univalent functions. For example, they are not starlike which makes them interesting since the class of starlike functions is very wide. So, class ${\mathcal U}$ attracts significant attention in the past decades and an overview of the most valuable results is given in Chapter 12 from \cite{DTV-book}.
\medskip
Here we will prove of the generalized Zalcman conjecture for the class ${\mathcal U}$ and for the cases $m=2$, $n=3$; $m=2$, $n=4$; and $m=n=3$.
\medskip
We also give direct proof and sharpness of the inequality
\[ |a_n^p-a_2^{p(n-1)}|\le 2^{p(n-1)}-n^p\]
over the class ${\mathcal U}$ for the cases $n=4$, $p=1$ and $n=5$, $p=1$. This inequality introduced by Krushkal and proven for the whole class of univalent functions in \cite{krushkal}.
\medskip
For the study and the proofs we will use the following useful property of functions in ${\mathcal U}$.
\begin{lem}[\cite{OB-NT-2018}] \label{lem-1}
For each function $f$ in ${\mathcal U}$, there exists function $\omega_1$, analytic in the unit disk, such that $|\omega_1(z)|\le|z|<1$ and $|\omega_1'(z)|\le1$ for all $z\in{\mathbb D}$, with
\begin{equation}\label{u1}
\frac{z}{f(z)} = 1-a_2z-z\omega_1(z).
\end{equation}
Additionally, for $\omega_1(z) = c_1z+c_2z^2+\cdots$,
\[|c_1|\le 1,\quad\quad |c_2|\le \frac12(1-|c_1|^2) \quad \mbox{and}\quad |c_3|\le \frac13 \left( 1-|c_1|^2-\frac{4|c_2|^2}{1+|c_1|} \right). \]
\end{lem}
\medskip
Let note that for functions $f$ from ${\mathcal U}$, of form \eqref{e1}, from Lemma \ref{lem-1} we have
\[ z = [1-a_2z-z\omega_1(z)]\cdot f(z), \]
and after equating the coefficients,
\[
\begin{split}
a_3 &= c_1+a_2^2,\\
a_4 &= c_2+2a_2c_1+a_2^3,\\
a_5 &= c_3+2a_2c_2+c_{1}^{2}+3a_2^2 c_{1}+a_{2}^{4}.
\end{split}
\]
\medskip
\section{Zalcman conjecture for the class ${\mathcal U}$}
\medskip
We now give direct proof of the Zalcman conjecture for the class ${\mathcal U}$ for the cases when $n=2$ and $n=3$.
\medskip
\begin{thm}\label{th-1}
Let $f\in{\mathcal U}$ be of the form \eqref{e1}. Then
\begin{itemize}
\item[$(i)$] $|a_2^2-a_3|\le1$;
\item[$(ii)$] $|a_3^2-a_5|\le4$.
\end{itemize}
These inequalities are sharp with equality for the Koebe function $k(z)=\frac{z}{(1-z)^2}=z+\sum_{n=2}nz^n$ and its rotations.
\end{thm}
\medskip
\begin{proof}$ $
\begin{itemize}
\item[$(i)$] From $a_3 = c_1+a_2^2$ we have $|a_2^2-a_3| = |-c_1|\le1$.
%
\item[$(ii)$] From the Bieberbach conjecture, $|a_3| = |c_1+a_2^2|\le 3$, and further calculations show that
\[
\begin{split}
\left|a_3^2-a_5 \right|
&= \left|c_3+2a_2c_2+a_2^2c_1 \right| \\
&= \left|c_3+2a_2c_2 - c_1^2+c_1(c_1 +a_2^2) \right|\\
&\le |c_3|+2|a_2||c_2| + |c_1|^2+|c_1||c_1 +a_2^2| \\
&\le |c_3|+2|a_2||c_2| + |c_1|^2+ 3|c_1| \\
&\le \frac13 \left( 1-|c_1|^2-\frac{4|c_2|^2}{1+|c_1|} \right) + 4|c_2| + |c_1|^2+ 3|c_1|\\
&:= f_1(|c_1|,|c_2|),
\end{split}
\]
where
\[ f_1(x,y) = \frac13 \left( 1-x^2-\frac{4y^2}{1+x} \right) + 4y + x^2+ 3x, \]
$0\le x=|c_1|\le1$ and $0\le y=|c_2|\le \frac12(1-x^2)$, i.e., $(x,y)\in G:=[0,1]\times[0,(1-x^2)/2]$.
\medskip
Since, $\frac{\partial f_1}{\partial y}(x,y) = \frac43 \left( \frac{y}{1+x} \right)^2+\frac43x+3 >0$ for all $(x,y)\in G$, we have that there are no singular points in the interior of $G$ and $f_1$ attains its maximum on the boundary of $G$.
\medskip
Further, for $x=0$ we have $0\le y\le \frac12$ and $f_1(0,y)=\frac13(1-4y^2)+4y\le2$.
\medskip
Also, for $0\le x\le1$ and $y=0$ we have $f_1(x,0)=\frac13(1-x^2)+x^2+3x\le4$.
\medskip
Finally, for $0\le x\le1$ and $y=\frac12(1-x^2)$ we have $f_1(x,\frac12(1-x^2))=2+\frac{10}{3}x-x^2-\frac13x^3 \le4$, since the last function is an increasing one on $[0,1]$.
\end{itemize}
\end{proof}
\medskip
\section{Generalized Zalcman conjecture for the class ${\mathcal U}$}
\medskip
In this section we give direct proof of the generalized Zalcman conjecture for the class ${\mathcal U}$ for the cases $m=2$, $n=3$; and $m=2$, $n=4$.
\medskip
\begin{thm}
Let $f\in{\mathcal U}$ be of the form \eqref{e1}. Then
\begin{itemize}
\item[$(i)$] $|a_2a_3-a_4|\le2$;
\item[$(ii)$] $|a_2a_4-a_5|\le3$.
\end{itemize}
These inequalities are sharp with equality for the Koebe function $k(z)=\frac{z}{(1-z)^2}=z+\sum_{n=2}nz^n$ and its rotations.
\end{thm}
\medskip
\begin{proof}$ $
\begin{itemize}
\item[$(i)$] In this case we have
\[
\begin{split}
|a_2a_3-a_4|
&= |c_2+a_2c_1| \le |c_2|+|a_2||c_1| \le |c_2|+2|c_1| \\
&\le \frac12 (1-|c_1|^2)+ 2|c_1| \le \frac12 (1-|c_1|^2+4|c_1| )\le 2.
\end{split}
\]
\item[$(ii)$] In a similar way as in the proof of Theorem \ref{th-1}($ii$),
\[
\begin{split}
\left|a_4a_2-a_5 \right|
&= \left|c_3 + a_2c_2 + a_2^2c_1 + c_1^2 \right| \\
&\le |c_3| + |a_2||c_2| + |c_1| |a_2^2+c_1| \\
&\le |c_3| + |a_2||c_2| + 3|c_1| \\
&\le \frac13 \left( 1-|c_1|^2-\frac{4|c_2|^2}{1+|c_1|} \right) + 2|c_2| + 3|c_1|\\
&:= f_2(|c_1|,|c_2|),
\end{split}
\]
where
\[ f_2(x,y) = \frac13 \left( 1-x^2-\frac{4y^2}{1+x} \right) + 2y + 3x, \]
$0\le x=|c_1|\le1$ and $0\le y=|c_2|\le \frac12(1-x^2)$, i.e., $(x,y)\in G:=[0,1]\times[0,(1-x^2)/2]$.
\medskip
Again, $\frac{\partial f_2}{\partial y}(x,y) = \frac43 \left( \frac{y}{1+x} \right)^2-\frac23x+3 >0$ for all $(x,y)\in G$, so $f_2$ attains its maximum on the boundary of $G$.
\medskip
The conclusion follows since on the edges of $G$ we have:
\begin{itemize}
\item[-] $x=0$, $0\le y\le \frac12$ and $f_2(0,y)=\frac13(1-4y^2)+2y\le1$;
\item[-] $y=0$, $0\le x\le1$ and $f_2(x,0)=\frac13(1-x^2)+3x\le3$;
\item[-] $y=\frac12(1-x^2)$, $0\le x\le1$ and $f_2(x,\frac12(1-x^2))=1+\frac{10}{3}x-x^2-\frac13x^3 \le3$.
\end{itemize}
\end{itemize}
\end{proof}
\medskip
\section{Krushkal inequality for the class ${\mathcal U}$}
\medskip
In this section we give direct proof of the Krushkal inequality for the class ${\mathcal U}$ in the cases when $n=4$, $p=1$ and $n=5$, $p=1$.
\medskip
\begin{thm}
Let $f\in{\mathcal U}$ be of the form \eqref{e1}. Then
\begin{itemize}
\item[$(i)$] $|a_4-a_2^3|\le4$;
\item[$(ii)$] $|a_5-a_2^4|\le11$.
\end{itemize}
These inequalities are sharp with equality for the Koebe function $k(z)=\frac{z}{(1-z)^2}=z+\sum_{n=2}nz^n$ and its rotations.
\end{thm}
\medskip
\begin{proof}$ $
\begin{itemize}
\item[$(i)$] It is easy to verify that
\[
\begin{split}
|a_4-a_2^3| &= |c_2+2a_2c_1| \le |c_2|+2|a_2||c_1| \\
&\le \frac12 (1-|c_1|^2)+ 4|c_1| = \frac12\left(1+8|c_1|-|c_1|^2\right)\le 4.
\end{split}
\]
\medskip
\item[$(ii)$] We will again use that $|a_3| = |c_1+a_2^2|\le 3$ and receive
\[
\begin{split}
|a_5-a_2^4| &= |c_3+2a_2c_2+c_1^2+3a_2^2c_1| \\
& = |c_3+2a_2c_2-2c_1^2+3c_1(c_1+a_2^2)| \\
&\le |c_3|+2|a_2||c_2|+2|c_1|^2+9|c_1| \\
&\le \frac13 \left( 1-|c_1|^2-\frac{4|c_2|^2}{1+|c_1|} \right) + 4|c_2| + 2|c_1|^2 + 9|c_1|\\
&:= g(|c_1|,|c_2|),
\end{split}
\]
where
\[ g(x,y) = \frac13 \left( 1-x^2-\frac{4y^2}{1+x} \right) + 4y + 2x^2+9x, \]
$0\le x=|c_1|\le1$ and $0\le y=|c_2|\le \frac12(1-x^2)$, i.e., $(x,y)\in G:=[0,1]\times[0,(1-x^2)/2]$.
\medskip
Since, $\frac{\partial g}{\partial y}(x,y) = \frac{10}{3}x + \frac43 \left( \frac{y}{1+x} \right)^2 + 9 >0$ for all $(x,y)\in G$, so $g$ has no critical points in the interior of $G$ and attains its maximum on the boundary:
\begin{itemize}
\item[-] $x=0$, $0\le y\le \frac12$ and $g(0,y)=\frac13(1+12y-4y^2)\le2$;
\item[-] $y=0$, $0\le x\le1$ and $g(x,0)=\frac53x^2+9x+\frac13\le11$;
\item[-] $y=\frac12(1-x^2)$, $0\le x\le1$ and $g(x,\frac12(1-x^2))=2+\frac{28}{3}x-\frac13x^3 \le11$.
\end{itemize}
\medskip
The statement ($ii$) follows directly.
\end{itemize}
\end{proof}
\medskip
| f22c4aa6486d88e5afee45270fce7904d98d5221 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{The ALICE Collaboration}
\begingroup
\small
\begin{flushleft}
S.~Acharya\Irefn{org141}\And
D.~Adamov\'{a}\Irefn{org95}\And
A.~Adler\Irefn{org74}\And
J.~Adolfsson\Irefn{org81}\And
M.M.~Aggarwal\Irefn{org100}\And
G.~Aglieri Rinella\Irefn{org34}\And
M.~Agnello\Irefn{org30}\And
N.~Agrawal\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
Z.~Ahammed\Irefn{org141}\And
S.~Ahmad\Irefn{org16}\And
S.U.~Ahn\Irefn{org76}\And
Z.~Akbar\Irefn{org51}\And
A.~Akindinov\Irefn{org92}\And
M.~Al-Turany\Irefn{org107}\And
S.N.~Alam\Irefn{org40}\textsuperscript{,}\Irefn{org141}\And
D.S.D.~Albuquerque\Irefn{org122}\And
D.~Aleksandrov\Irefn{org88}\And
B.~Alessandro\Irefn{org59}\And
H.M.~Alfanda\Irefn{org6}\And
R.~Alfaro Molina\Irefn{org71}\And
B.~Ali\Irefn{org16}\And
Y.~Ali\Irefn{org14}\And
A.~Alici\Irefn{org10}\textsuperscript{,}\Irefn{org26}\textsuperscript{,}\Irefn{org54}\And
N.~Alizadehvandchali\Irefn{org125}\And
A.~Alkin\Irefn{org2}\textsuperscript{,}\Irefn{org34}\And
J.~Alme\Irefn{org21}\And
T.~Alt\Irefn{org68}\And
L.~Altenkamper\Irefn{org21}\And
I.~Altsybeev\Irefn{org113}\And
M.N.~Anaam\Irefn{org6}\And
C.~Andrei\Irefn{org48}\And
D.~Andreou\Irefn{org34}\And
A.~Andronic\Irefn{org144}\And
M.~Angeletti\Irefn{org34}\And
V.~Anguelov\Irefn{org104}\And
C.~Anson\Irefn{org15}\And
T.~Anti\v{c}i\'{c}\Irefn{org108}\And
F.~Antinori\Irefn{org57}\And
P.~Antonioli\Irefn{org54}\And
N.~Apadula\Irefn{org80}\And
L.~Aphecetche\Irefn{org115}\And
H.~Appelsh\"{a}user\Irefn{org68}\And
S.~Arcelli\Irefn{org26}\And
R.~Arnaldi\Irefn{org59}\And
M.~Arratia\Irefn{org80}\And
I.C.~Arsene\Irefn{org20}\And
M.~Arslandok\Irefn{org104}\And
A.~Augustinus\Irefn{org34}\And
R.~Averbeck\Irefn{org107}\And
S.~Aziz\Irefn{org78}\And
M.D.~Azmi\Irefn{org16}\And
A.~Badal\`{a}\Irefn{org56}\And
Y.W.~Baek\Irefn{org41}\And
S.~Bagnasco\Irefn{org59}\And
X.~Bai\Irefn{org107}\And
R.~Bailhache\Irefn{org68}\And
R.~Bala\Irefn{org101}\And
A.~Balbino\Irefn{org30}\And
A.~Baldisseri\Irefn{org137}\And
M.~Ball\Irefn{org43}\And
S.~Balouza\Irefn{org105}\And
D.~Banerjee\Irefn{org3}\And
R.~Barbera\Irefn{org27}\And
L.~Barioglio\Irefn{org25}\And
G.G.~Barnaf\"{o}ldi\Irefn{org145}\And
L.S.~Barnby\Irefn{org94}\And
V.~Barret\Irefn{org134}\And
P.~Bartalini\Irefn{org6}\And
C.~Bartels\Irefn{org127}\And
K.~Barth\Irefn{org34}\And
E.~Bartsch\Irefn{org68}\And
F.~Baruffaldi\Irefn{org28}\And
N.~Bastid\Irefn{org134}\And
S.~Basu\Irefn{org143}\And
G.~Batigne\Irefn{org115}\And
B.~Batyunya\Irefn{org75}\And
D.~Bauri\Irefn{org49}\And
J.L.~Bazo~Alba\Irefn{org112}\And
I.G.~Bearden\Irefn{org89}\And
C.~Beattie\Irefn{org146}\And
C.~Bedda\Irefn{org63}\And
N.K.~Behera\Irefn{org61}\And
I.~Belikov\Irefn{org136}\And
A.D.C.~Bell Hechavarria\Irefn{org144}\And
F.~Bellini\Irefn{org34}\And
R.~Bellwied\Irefn{org125}\And
V.~Belyaev\Irefn{org93}\And
G.~Bencedi\Irefn{org145}\And
S.~Beole\Irefn{org25}\And
A.~Bercuci\Irefn{org48}\And
Y.~Berdnikov\Irefn{org98}\And
D.~Berenyi\Irefn{org145}\And
R.A.~Bertens\Irefn{org130}\And
D.~Berzano\Irefn{org59}\And
M.G.~Besoiu\Irefn{org67}\And
L.~Betev\Irefn{org34}\And
A.~Bhasin\Irefn{org101}\And
I.R.~Bhat\Irefn{org101}\And
M.A.~Bhat\Irefn{org3}\And
H.~Bhatt\Irefn{org49}\And
B.~Bhattacharjee\Irefn{org42}\And
A.~Bianchi\Irefn{org25}\And
L.~Bianchi\Irefn{org25}\And
N.~Bianchi\Irefn{org52}\And
J.~Biel\v{c}\'{\i}k\Irefn{org37}\And
J.~Biel\v{c}\'{\i}kov\'{a}\Irefn{org95}\And
A.~Bilandzic\Irefn{org105}\And
G.~Biro\Irefn{org145}\And
R.~Biswas\Irefn{org3}\And
S.~Biswas\Irefn{org3}\And
J.T.~Blair\Irefn{org119}\And
D.~Blau\Irefn{org88}\And
C.~Blume\Irefn{org68}\And
G.~Boca\Irefn{org139}\And
F.~Bock\Irefn{org96}\And
A.~Bogdanov\Irefn{org93}\And
S.~Boi\Irefn{org23}\And
J.~Bok\Irefn{org61}\And
L.~Boldizs\'{a}r\Irefn{org145}\And
A.~Bolozdynya\Irefn{org93}\And
M.~Bombara\Irefn{org38}\And
G.~Bonomi\Irefn{org140}\And
H.~Borel\Irefn{org137}\And
A.~Borissov\Irefn{org93}\And
H.~Bossi\Irefn{org146}\And
E.~Botta\Irefn{org25}\And
L.~Bratrud\Irefn{org68}\And
P.~Braun-Munzinger\Irefn{org107}\And
M.~Bregant\Irefn{org121}\And
M.~Broz\Irefn{org37}\And
E.~Bruna\Irefn{org59}\And
G.E.~Bruno\Irefn{org33}\textsuperscript{,}\Irefn{org106}\And
M.D.~Buckland\Irefn{org127}\And
D.~Budnikov\Irefn{org109}\And
H.~Buesching\Irefn{org68}\And
S.~Bufalino\Irefn{org30}\And
O.~Bugnon\Irefn{org115}\And
P.~Buhler\Irefn{org114}\And
P.~Buncic\Irefn{org34}\And
Z.~Buthelezi\Irefn{org72}\textsuperscript{,}\Irefn{org131}\And
J.B.~Butt\Irefn{org14}\And
S.A.~Bysiak\Irefn{org118}\And
D.~Caffarri\Irefn{org90}\And
M.~Cai\Irefn{org6}\And
A.~Caliva\Irefn{org107}\And
E.~Calvo Villar\Irefn{org112}\And
J.M.M.~Camacho\Irefn{org120}\And
R.S.~Camacho\Irefn{org45}\And
P.~Camerini\Irefn{org24}\And
F.D.M.~Canedo\Irefn{org121}\And
A.A.~Capon\Irefn{org114}\And
F.~Carnesecchi\Irefn{org26}\And
R.~Caron\Irefn{org137}\And
J.~Castillo Castellanos\Irefn{org137}\And
A.J.~Castro\Irefn{org130}\And
E.A.R.~Casula\Irefn{org55}\And
F.~Catalano\Irefn{org30}\And
C.~Ceballos Sanchez\Irefn{org75}\And
P.~Chakraborty\Irefn{org49}\And
S.~Chandra\Irefn{org141}\And
W.~Chang\Irefn{org6}\And
S.~Chapeland\Irefn{org34}\And
M.~Chartier\Irefn{org127}\And
S.~Chattopadhyay\Irefn{org141}\And
S.~Chattopadhyay\Irefn{org110}\And
A.~Chauvin\Irefn{org23}\And
C.~Cheshkov\Irefn{org135}\And
B.~Cheynis\Irefn{org135}\And
V.~Chibante Barroso\Irefn{org34}\And
D.D.~Chinellato\Irefn{org122}\And
S.~Cho\Irefn{org61}\And
P.~Chochula\Irefn{org34}\And
T.~Chowdhury\Irefn{org134}\And
P.~Christakoglou\Irefn{org90}\And
C.H.~Christensen\Irefn{org89}\And
P.~Christiansen\Irefn{org81}\And
T.~Chujo\Irefn{org133}\And
C.~Cicalo\Irefn{org55}\And
L.~Cifarelli\Irefn{org10}\textsuperscript{,}\Irefn{org26}\And
L.D.~Cilladi\Irefn{org25}\And
F.~Cindolo\Irefn{org54}\And
M.R.~Ciupek\Irefn{org107}\And
G.~Clai\Irefn{org54}\Aref{orgI}\And
J.~Cleymans\Irefn{org124}\And
F.~Colamaria\Irefn{org53}\And
D.~Colella\Irefn{org53}\And
A.~Collu\Irefn{org80}\And
M.~Colocci\Irefn{org26}\And
M.~Concas\Irefn{org59}\Aref{orgII}\And
G.~Conesa Balbastre\Irefn{org79}\And
Z.~Conesa del Valle\Irefn{org78}\And
G.~Contin\Irefn{org24}\textsuperscript{,}\Irefn{org60}\And
J.G.~Contreras\Irefn{org37}\And
T.M.~Cormier\Irefn{org96}\And
Y.~Corrales Morales\Irefn{org25}\And
P.~Cortese\Irefn{org31}\And
M.R.~Cosentino\Irefn{org123}\And
F.~Costa\Irefn{org34}\And
S.~Costanza\Irefn{org139}\And
P.~Crochet\Irefn{org134}\And
E.~Cuautle\Irefn{org69}\And
P.~Cui\Irefn{org6}\And
L.~Cunqueiro\Irefn{org96}\And
D.~Dabrowski\Irefn{org142}\And
T.~Dahms\Irefn{org105}\And
A.~Dainese\Irefn{org57}\And
F.P.A.~Damas\Irefn{org115}\textsuperscript{,}\Irefn{org137}\And
M.C.~Danisch\Irefn{org104}\And
A.~Danu\Irefn{org67}\And
D.~Das\Irefn{org110}\And
I.~Das\Irefn{org110}\And
P.~Das\Irefn{org86}\And
P.~Das\Irefn{org3}\And
S.~Das\Irefn{org3}\And
A.~Dash\Irefn{org86}\And
S.~Dash\Irefn{org49}\And
S.~De\Irefn{org86}\And
A.~De Caro\Irefn{org29}\And
G.~de Cataldo\Irefn{org53}\And
J.~de Cuveland\Irefn{org39}\And
A.~De Falco\Irefn{org23}\And
D.~De Gruttola\Irefn{org10}\And
N.~De Marco\Irefn{org59}\And
S.~De Pasquale\Irefn{org29}\And
S.~Deb\Irefn{org50}\And
H.F.~Degenhardt\Irefn{org121}\And
K.R.~Deja\Irefn{org142}\And
A.~Deloff\Irefn{org85}\And
S.~Delsanto\Irefn{org25}\textsuperscript{,}\Irefn{org131}\And
W.~Deng\Irefn{org6}\And
P.~Dhankher\Irefn{org49}\And
D.~Di Bari\Irefn{org33}\And
A.~Di Mauro\Irefn{org34}\And
R.A.~Diaz\Irefn{org8}\And
T.~Dietel\Irefn{org124}\And
P.~Dillenseger\Irefn{org68}\And
Y.~Ding\Irefn{org6}\And
R.~Divi\`{a}\Irefn{org34}\And
D.U.~Dixit\Irefn{org19}\And
{\O}.~Djuvsland\Irefn{org21}\And
U.~Dmitrieva\Irefn{org62}\And
A.~Dobrin\Irefn{org67}\And
B.~D\"{o}nigus\Irefn{org68}\And
O.~Dordic\Irefn{org20}\And
A.K.~Dubey\Irefn{org141}\And
A.~Dubla\Irefn{org90}\textsuperscript{,}\Irefn{org107}\And
S.~Dudi\Irefn{org100}\And
M.~Dukhishyam\Irefn{org86}\And
P.~Dupieux\Irefn{org134}\And
R.J.~Ehlers\Irefn{org96}\And
V.N.~Eikeland\Irefn{org21}\And
D.~Elia\Irefn{org53}\And
B.~Erazmus\Irefn{org115}\And
F.~Erhardt\Irefn{org99}\And
A.~Erokhin\Irefn{org113}\And
M.R.~Ersdal\Irefn{org21}\And
B.~Espagnon\Irefn{org78}\And
G.~Eulisse\Irefn{org34}\And
D.~Evans\Irefn{org111}\And
S.~Evdokimov\Irefn{org91}\And
L.~Fabbietti\Irefn{org105}\And
M.~Faggin\Irefn{org28}\And
J.~Faivre\Irefn{org79}\And
F.~Fan\Irefn{org6}\And
A.~Fantoni\Irefn{org52}\And
M.~Fasel\Irefn{org96}\And
P.~Fecchio\Irefn{org30}\And
A.~Feliciello\Irefn{org59}\And
G.~Feofilov\Irefn{org113}\And
A.~Fern\'{a}ndez T\'{e}llez\Irefn{org45}\And
A.~Ferrero\Irefn{org137}\And
A.~Ferretti\Irefn{org25}\And
A.~Festanti\Irefn{org34}\And
V.J.G.~Feuillard\Irefn{org104}\And
J.~Figiel\Irefn{org118}\And
S.~Filchagin\Irefn{org109}\And
D.~Finogeev\Irefn{org62}\And
F.M.~Fionda\Irefn{org21}\And
G.~Fiorenza\Irefn{org53}\And
F.~Flor\Irefn{org125}\And
A.N.~Flores\Irefn{org119}\And
S.~Foertsch\Irefn{org72}\And
P.~Foka\Irefn{org107}\And
S.~Fokin\Irefn{org88}\And
E.~Fragiacomo\Irefn{org60}\And
U.~Frankenfeld\Irefn{org107}\And
U.~Fuchs\Irefn{org34}\And
C.~Furget\Irefn{org79}\And
A.~Furs\Irefn{org62}\And
M.~Fusco Girard\Irefn{org29}\And
J.J.~Gaardh{\o}je\Irefn{org89}\And
M.~Gagliardi\Irefn{org25}\And
A.M.~Gago\Irefn{org112}\And
A.~Gal\Irefn{org136}\And
C.D.~Galvan\Irefn{org120}\And
P.~Ganoti\Irefn{org84}\And
C.~Garabatos\Irefn{org107}\And
J.R.A.~Garcia\Irefn{org45}\And
E.~Garcia-Solis\Irefn{org11}\And
K.~Garg\Irefn{org115}\And
C.~Gargiulo\Irefn{org34}\And
A.~Garibli\Irefn{org87}\And
K.~Garner\Irefn{org144}\And
P.~Gasik\Irefn{org105}\textsuperscript{,}\Irefn{org107}\And
E.F.~Gauger\Irefn{org119}\And
M.B.~Gay Ducati\Irefn{org70}\And
M.~Germain\Irefn{org115}\And
J.~Ghosh\Irefn{org110}\And
P.~Ghosh\Irefn{org141}\And
S.K.~Ghosh\Irefn{org3}\And
M.~Giacalone\Irefn{org26}\And
P.~Gianotti\Irefn{org52}\And
P.~Giubellino\Irefn{org59}\textsuperscript{,}\Irefn{org107}\And
P.~Giubilato\Irefn{org28}\And
A.M.C.~Glaenzer\Irefn{org137}\And
P.~Gl\"{a}ssel\Irefn{org104}\And
A.~Gomez Ramirez\Irefn{org74}\And
V.~Gonzalez\Irefn{org107}\textsuperscript{,}\Irefn{org143}\And
\mbox{L.H.~Gonz\'{a}lez-Trueba}\Irefn{org71}\And
S.~Gorbunov\Irefn{org39}\And
L.~G\"{o}rlich\Irefn{org118}\And
A.~Goswami\Irefn{org49}\And
S.~Gotovac\Irefn{org35}\And
V.~Grabski\Irefn{org71}\And
L.K.~Graczykowski\Irefn{org142}\And
K.L.~Graham\Irefn{org111}\And
L.~Greiner\Irefn{org80}\And
A.~Grelli\Irefn{org63}\And
C.~Grigoras\Irefn{org34}\And
V.~Grigoriev\Irefn{org93}\And
A.~Grigoryan\Irefn{org1}\And
S.~Grigoryan\Irefn{org75}\And
O.S.~Groettvik\Irefn{org21}\And
F.~Grosa\Irefn{org30}\textsuperscript{,}\Irefn{org59}\And
J.F.~Grosse-Oetringhaus\Irefn{org34}\And
R.~Grosso\Irefn{org107}\And
R.~Guernane\Irefn{org79}\And
M.~Guittiere\Irefn{org115}\And
K.~Gulbrandsen\Irefn{org89}\And
T.~Gunji\Irefn{org132}\And
A.~Gupta\Irefn{org101}\And
R.~Gupta\Irefn{org101}\And
I.B.~Guzman\Irefn{org45}\And
R.~Haake\Irefn{org146}\And
M.K.~Habib\Irefn{org107}\And
C.~Hadjidakis\Irefn{org78}\And
H.~Hamagaki\Irefn{org82}\And
G.~Hamar\Irefn{org145}\And
M.~Hamid\Irefn{org6}\And
R.~Hannigan\Irefn{org119}\And
M.R.~Haque\Irefn{org63}\textsuperscript{,}\Irefn{org86}\And
A.~Harlenderova\Irefn{org107}\And
J.W.~Harris\Irefn{org146}\And
A.~Harton\Irefn{org11}\And
J.A.~Hasenbichler\Irefn{org34}\And
H.~Hassan\Irefn{org96}\And
Q.U.~Hassan\Irefn{org14}\And
D.~Hatzifotiadou\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
P.~Hauer\Irefn{org43}\And
L.B.~Havener\Irefn{org146}\And
S.~Hayashi\Irefn{org132}\And
S.T.~Heckel\Irefn{org105}\And
E.~Hellb\"{a}r\Irefn{org68}\And
H.~Helstrup\Irefn{org36}\And
A.~Herghelegiu\Irefn{org48}\And
T.~Herman\Irefn{org37}\And
E.G.~Hernandez\Irefn{org45}\And
G.~Herrera Corral\Irefn{org9}\And
F.~Herrmann\Irefn{org144}\And
K.F.~Hetland\Irefn{org36}\And
H.~Hillemanns\Irefn{org34}\And
C.~Hills\Irefn{org127}\And
B.~Hippolyte\Irefn{org136}\And
B.~Hohlweger\Irefn{org105}\And
J.~Honermann\Irefn{org144}\And
D.~Horak\Irefn{org37}\And
A.~Hornung\Irefn{org68}\And
S.~Hornung\Irefn{org107}\And
R.~Hosokawa\Irefn{org15}\textsuperscript{,}\Irefn{org133}\And
P.~Hristov\Irefn{org34}\And
C.~Huang\Irefn{org78}\And
C.~Hughes\Irefn{org130}\And
P.~Huhn\Irefn{org68}\And
T.J.~Humanic\Irefn{org97}\And
H.~Hushnud\Irefn{org110}\And
L.A.~Husova\Irefn{org144}\And
N.~Hussain\Irefn{org42}\And
S.A.~Hussain\Irefn{org14}\And
D.~Hutter\Irefn{org39}\And
J.P.~Iddon\Irefn{org34}\textsuperscript{,}\Irefn{org127}\And
R.~Ilkaev\Irefn{org109}\And
H.~Ilyas\Irefn{org14}\And
M.~Inaba\Irefn{org133}\And
G.M.~Innocenti\Irefn{org34}\And
M.~Ippolitov\Irefn{org88}\And
A.~Isakov\Irefn{org95}\And
M.S.~Islam\Irefn{org110}\And
M.~Ivanov\Irefn{org107}\And
V.~Ivanov\Irefn{org98}\And
V.~Izucheev\Irefn{org91}\And
B.~Jacak\Irefn{org80}\And
N.~Jacazio\Irefn{org34}\textsuperscript{,}\Irefn{org54}\And
P.M.~Jacobs\Irefn{org80}\And
S.~Jadlovska\Irefn{org117}\And
J.~Jadlovsky\Irefn{org117}\And
S.~Jaelani\Irefn{org63}\And
C.~Jahnke\Irefn{org121}\And
M.J.~Jakubowska\Irefn{org142}\And
M.A.~Janik\Irefn{org142}\And
T.~Janson\Irefn{org74}\And
M.~Jercic\Irefn{org99}\And
O.~Jevons\Irefn{org111}\And
M.~Jin\Irefn{org125}\And
F.~Jonas\Irefn{org96}\textsuperscript{,}\Irefn{org144}\And
P.G.~Jones\Irefn{org111}\And
J.~Jung\Irefn{org68}\And
M.~Jung\Irefn{org68}\And
A.~Jusko\Irefn{org111}\And
P.~Kalinak\Irefn{org64}\And
A.~Kalweit\Irefn{org34}\And
V.~Kaplin\Irefn{org93}\And
S.~Kar\Irefn{org6}\And
A.~Karasu Uysal\Irefn{org77}\And
D.~Karatovic\Irefn{org99}\And
O.~Karavichev\Irefn{org62}\And
T.~Karavicheva\Irefn{org62}\And
P.~Karczmarczyk\Irefn{org142}\And
E.~Karpechev\Irefn{org62}\And
A.~Kazantsev\Irefn{org88}\And
U.~Kebschull\Irefn{org74}\And
R.~Keidel\Irefn{org47}\And
M.~Keil\Irefn{org34}\And
B.~Ketzer\Irefn{org43}\And
Z.~Khabanova\Irefn{org90}\And
A.M.~Khan\Irefn{org6}\And
S.~Khan\Irefn{org16}\And
A.~Khanzadeev\Irefn{org98}\And
Y.~Kharlov\Irefn{org91}\And
A.~Khatun\Irefn{org16}\And
A.~Khuntia\Irefn{org118}\And
B.~Kileng\Irefn{org36}\And
B.~Kim\Irefn{org61}\And
B.~Kim\Irefn{org133}\And
D.~Kim\Irefn{org147}\And
D.J.~Kim\Irefn{org126}\And
E.J.~Kim\Irefn{org73}\And
H.~Kim\Irefn{org17}\And
J.~Kim\Irefn{org147}\And
J.S.~Kim\Irefn{org41}\And
J.~Kim\Irefn{org104}\And
J.~Kim\Irefn{org147}\And
J.~Kim\Irefn{org73}\And
M.~Kim\Irefn{org104}\And
S.~Kim\Irefn{org18}\And
T.~Kim\Irefn{org147}\And
T.~Kim\Irefn{org147}\And
S.~Kirsch\Irefn{org68}\And
I.~Kisel\Irefn{org39}\And
S.~Kiselev\Irefn{org92}\And
A.~Kisiel\Irefn{org142}\And
J.L.~Klay\Irefn{org5}\And
C.~Klein\Irefn{org68}\And
J.~Klein\Irefn{org34}\textsuperscript{,}\Irefn{org59}\And
S.~Klein\Irefn{org80}\And
C.~Klein-B\"{o}sing\Irefn{org144}\And
M.~Kleiner\Irefn{org68}\And
T.~Klemenz\Irefn{org105}\And
A.~Kluge\Irefn{org34}\And
M.L.~Knichel\Irefn{org34}\And
A.G.~Knospe\Irefn{org125}\And
C.~Kobdaj\Irefn{org116}\And
M.K.~K\"{o}hler\Irefn{org104}\And
T.~Kollegger\Irefn{org107}\And
A.~Kondratyev\Irefn{org75}\And
N.~Kondratyeva\Irefn{org93}\And
E.~Kondratyuk\Irefn{org91}\And
J.~Konig\Irefn{org68}\And
S.A.~Konigstorfer\Irefn{org105}\And
P.J.~Konopka\Irefn{org34}\And
G.~Kornakov\Irefn{org142}\And
L.~Koska\Irefn{org117}\And
O.~Kovalenko\Irefn{org85}\And
V.~Kovalenko\Irefn{org113}\And
M.~Kowalski\Irefn{org118}\And
I.~Kr\'{a}lik\Irefn{org64}\And
A.~Krav\v{c}\'{a}kov\'{a}\Irefn{org38}\And
L.~Kreis\Irefn{org107}\And
M.~Krivda\Irefn{org64}\textsuperscript{,}\Irefn{org111}\And
F.~Krizek\Irefn{org95}\And
K.~Krizkova~Gajdosova\Irefn{org37}\And
M.~Kr\"uger\Irefn{org68}\And
E.~Kryshen\Irefn{org98}\And
M.~Krzewicki\Irefn{org39}\And
A.M.~Kubera\Irefn{org97}\And
V.~Ku\v{c}era\Irefn{org34}\textsuperscript{,}\Irefn{org61}\And
C.~Kuhn\Irefn{org136}\And
P.G.~Kuijer\Irefn{org90}\And
L.~Kumar\Irefn{org100}\And
S.~Kundu\Irefn{org86}\And
P.~Kurashvili\Irefn{org85}\And
A.~Kurepin\Irefn{org62}\And
A.B.~Kurepin\Irefn{org62}\And
A.~Kuryakin\Irefn{org109}\And
S.~Kushpil\Irefn{org95}\And
J.~Kvapil\Irefn{org111}\And
M.J.~Kweon\Irefn{org61}\And
J.Y.~Kwon\Irefn{org61}\And
Y.~Kwon\Irefn{org147}\And
S.L.~La Pointe\Irefn{org39}\And
P.~La Rocca\Irefn{org27}\And
Y.S.~Lai\Irefn{org80}\And
M.~Lamanna\Irefn{org34}\And
R.~Langoy\Irefn{org129}\And
K.~Lapidus\Irefn{org34}\And
A.~Lardeux\Irefn{org20}\And
P.~Larionov\Irefn{org52}\And
E.~Laudi\Irefn{org34}\And
R.~Lavicka\Irefn{org37}\And
T.~Lazareva\Irefn{org113}\And
R.~Lea\Irefn{org24}\And
L.~Leardini\Irefn{org104}\And
J.~Lee\Irefn{org133}\And
S.~Lee\Irefn{org147}\And
S.~Lehner\Irefn{org114}\And
J.~Lehrbach\Irefn{org39}\And
R.C.~Lemmon\Irefn{org94}\And
I.~Le\'{o}n Monz\'{o}n\Irefn{org120}\And
E.D.~Lesser\Irefn{org19}\And
M.~Lettrich\Irefn{org34}\And
P.~L\'{e}vai\Irefn{org145}\And
X.~Li\Irefn{org12}\And
X.L.~Li\Irefn{org6}\And
J.~Lien\Irefn{org129}\And
R.~Lietava\Irefn{org111}\And
B.~Lim\Irefn{org17}\And
V.~Lindenstruth\Irefn{org39}\And
A.~Lindner\Irefn{org48}\And
C.~Lippmann\Irefn{org107}\And
M.A.~Lisa\Irefn{org97}\And
A.~Liu\Irefn{org19}\And
J.~Liu\Irefn{org127}\And
S.~Liu\Irefn{org97}\And
W.J.~Llope\Irefn{org143}\And
I.M.~Lofnes\Irefn{org21}\And
V.~Loginov\Irefn{org93}\And
C.~Loizides\Irefn{org96}\And
P.~Loncar\Irefn{org35}\And
J.A.~Lopez\Irefn{org104}\And
X.~Lopez\Irefn{org134}\And
E.~L\'{o}pez Torres\Irefn{org8}\And
J.R.~Luhder\Irefn{org144}\And
M.~Lunardon\Irefn{org28}\And
G.~Luparello\Irefn{org60}\And
Y.G.~Ma\Irefn{org40}\And
A.~Maevskaya\Irefn{org62}\And
M.~Mager\Irefn{org34}\And
S.M.~Mahmood\Irefn{org20}\And
T.~Mahmoud\Irefn{org43}\And
A.~Maire\Irefn{org136}\And
R.D.~Majka\Irefn{org146}\Aref{org*}\And
M.~Malaev\Irefn{org98}\And
Q.W.~Malik\Irefn{org20}\And
L.~Malinina\Irefn{org75}\Aref{orgIII}\And
D.~Mal'Kevich\Irefn{org92}\And
P.~Malzacher\Irefn{org107}\And
G.~Mandaglio\Irefn{org32}\textsuperscript{,}\Irefn{org56}\And
V.~Manko\Irefn{org88}\And
F.~Manso\Irefn{org134}\And
V.~Manzari\Irefn{org53}\And
Y.~Mao\Irefn{org6}\And
M.~Marchisone\Irefn{org135}\And
J.~Mare\v{s}\Irefn{org66}\And
G.V.~Margagliotti\Irefn{org24}\And
A.~Margotti\Irefn{org54}\And
A.~Mar\'{\i}n\Irefn{org107}\And
C.~Markert\Irefn{org119}\And
M.~Marquard\Irefn{org68}\And
C.D.~Martin\Irefn{org24}\And
N.A.~Martin\Irefn{org104}\And
P.~Martinengo\Irefn{org34}\And
J.L.~Martinez\Irefn{org125}\And
M.I.~Mart\'{\i}nez\Irefn{org45}\And
G.~Mart\'{\i}nez Garc\'{\i}a\Irefn{org115}\And
S.~Masciocchi\Irefn{org107}\And
M.~Masera\Irefn{org25}\And
A.~Masoni\Irefn{org55}\And
L.~Massacrier\Irefn{org78}\And
E.~Masson\Irefn{org115}\And
A.~Mastroserio\Irefn{org53}\textsuperscript{,}\Irefn{org138}\And
A.M.~Mathis\Irefn{org105}\And
O.~Matonoha\Irefn{org81}\And
P.F.T.~Matuoka\Irefn{org121}\And
A.~Matyja\Irefn{org118}\And
C.~Mayer\Irefn{org118}\And
F.~Mazzaschi\Irefn{org25}\And
M.~Mazzilli\Irefn{org53}\And
M.A.~Mazzoni\Irefn{org58}\And
A.F.~Mechler\Irefn{org68}\And
F.~Meddi\Irefn{org22}\And
Y.~Melikyan\Irefn{org62}\textsuperscript{,}\Irefn{org93}\And
A.~Menchaca-Rocha\Irefn{org71}\And
E.~Meninno\Irefn{org29}\textsuperscript{,}\Irefn{org114}\And
A.S.~Menon\Irefn{org125}\And
M.~Meres\Irefn{org13}\And
S.~Mhlanga\Irefn{org124}\And
Y.~Miake\Irefn{org133}\And
L.~Micheletti\Irefn{org25}\And
L.C.~Migliorin\Irefn{org135}\And
D.L.~Mihaylov\Irefn{org105}\And
K.~Mikhaylov\Irefn{org75}\textsuperscript{,}\Irefn{org92}\And
A.N.~Mishra\Irefn{org69}\And
D.~Mi\'{s}kowiec\Irefn{org107}\And
A.~Modak\Irefn{org3}\And
N.~Mohammadi\Irefn{org34}\And
A.P.~Mohanty\Irefn{org63}\And
B.~Mohanty\Irefn{org86}\And
M.~Mohisin Khan\Irefn{org16}\Aref{orgIV}\And
Z.~Moravcova\Irefn{org89}\And
C.~Mordasini\Irefn{org105}\And
D.A.~Moreira De Godoy\Irefn{org144}\And
L.A.P.~Moreno\Irefn{org45}\And
I.~Morozov\Irefn{org62}\And
A.~Morsch\Irefn{org34}\And
T.~Mrnjavac\Irefn{org34}\And
V.~Muccifora\Irefn{org52}\And
E.~Mudnic\Irefn{org35}\And
D.~M{\"u}hlheim\Irefn{org144}\And
S.~Muhuri\Irefn{org141}\And
J.D.~Mulligan\Irefn{org80}\And
A.~Mulliri\Irefn{org23}\textsuperscript{,}\Irefn{org55}\And
M.G.~Munhoz\Irefn{org121}\And
R.H.~Munzer\Irefn{org68}\And
H.~Murakami\Irefn{org132}\And
S.~Murray\Irefn{org124}\And
L.~Musa\Irefn{org34}\And
J.~Musinsky\Irefn{org64}\And
C.J.~Myers\Irefn{org125}\And
J.W.~Myrcha\Irefn{org142}\And
B.~Naik\Irefn{org49}\And
R.~Nair\Irefn{org85}\And
B.K.~Nandi\Irefn{org49}\And
R.~Nania\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
E.~Nappi\Irefn{org53}\And
M.U.~Naru\Irefn{org14}\And
A.F.~Nassirpour\Irefn{org81}\And
C.~Nattrass\Irefn{org130}\And
R.~Nayak\Irefn{org49}\And
T.K.~Nayak\Irefn{org86}\And
S.~Nazarenko\Irefn{org109}\And
A.~Neagu\Irefn{org20}\And
R.A.~Negrao De Oliveira\Irefn{org68}\And
L.~Nellen\Irefn{org69}\And
S.V.~Nesbo\Irefn{org36}\And
G.~Neskovic\Irefn{org39}\And
D.~Nesterov\Irefn{org113}\And
L.T.~Neumann\Irefn{org142}\And
B.S.~Nielsen\Irefn{org89}\And
S.~Nikolaev\Irefn{org88}\And
S.~Nikulin\Irefn{org88}\And
V.~Nikulin\Irefn{org98}\And
F.~Noferini\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
P.~Nomokonov\Irefn{org75}\And
J.~Norman\Irefn{org79}\textsuperscript{,}\Irefn{org127}\And
N.~Novitzky\Irefn{org133}\And
P.~Nowakowski\Irefn{org142}\And
A.~Nyanin\Irefn{org88}\And
J.~Nystrand\Irefn{org21}\And
M.~Ogino\Irefn{org82}\And
A.~Ohlson\Irefn{org81}\And
J.~Oleniacz\Irefn{org142}\And
A.C.~Oliveira Da Silva\Irefn{org130}\And
M.H.~Oliver\Irefn{org146}\And
C.~Oppedisano\Irefn{org59}\And
A.~Ortiz Velasquez\Irefn{org69}\And
A.~Oskarsson\Irefn{org81}\And
J.~Otwinowski\Irefn{org118}\And
K.~Oyama\Irefn{org82}\And
Y.~Pachmayer\Irefn{org104}\And
V.~Pacik\Irefn{org89}\And
S.~Padhan\Irefn{org49}\And
D.~Pagano\Irefn{org140}\And
G.~Pai\'{c}\Irefn{org69}\And
J.~Pan\Irefn{org143}\And
S.~Panebianco\Irefn{org137}\And
P.~Pareek\Irefn{org50}\textsuperscript{,}\Irefn{org141}\And
J.~Park\Irefn{org61}\And
J.E.~Parkkila\Irefn{org126}\And
S.~Parmar\Irefn{org100}\And
S.P.~Pathak\Irefn{org125}\And
B.~Paul\Irefn{org23}\And
J.~Pazzini\Irefn{org140}\And
H.~Pei\Irefn{org6}\And
T.~Peitzmann\Irefn{org63}\And
X.~Peng\Irefn{org6}\And
L.G.~Pereira\Irefn{org70}\And
H.~Pereira Da Costa\Irefn{org137}\And
D.~Peresunko\Irefn{org88}\And
G.M.~Perez\Irefn{org8}\And
S.~Perrin\Irefn{org137}\And
Y.~Pestov\Irefn{org4}\And
V.~Petr\'{a}\v{c}ek\Irefn{org37}\And
M.~Petrovici\Irefn{org48}\And
R.P.~Pezzi\Irefn{org70}\And
S.~Piano\Irefn{org60}\And
M.~Pikna\Irefn{org13}\And
P.~Pillot\Irefn{org115}\And
O.~Pinazza\Irefn{org34}\textsuperscript{,}\Irefn{org54}\And
L.~Pinsky\Irefn{org125}\And
C.~Pinto\Irefn{org27}\And
S.~Pisano\Irefn{org10}\textsuperscript{,}\Irefn{org52}\And
D.~Pistone\Irefn{org56}\And
M.~P\l osko\'{n}\Irefn{org80}\And
M.~Planinic\Irefn{org99}\And
F.~Pliquett\Irefn{org68}\And
M.G.~Poghosyan\Irefn{org96}\And
B.~Polichtchouk\Irefn{org91}\And
N.~Poljak\Irefn{org99}\And
A.~Pop\Irefn{org48}\And
S.~Porteboeuf-Houssais\Irefn{org134}\And
V.~Pozdniakov\Irefn{org75}\And
S.K.~Prasad\Irefn{org3}\And
R.~Preghenella\Irefn{org54}\And
F.~Prino\Irefn{org59}\And
C.A.~Pruneau\Irefn{org143}\And
I.~Pshenichnov\Irefn{org62}\And
M.~Puccio\Irefn{org34}\And
J.~Putschke\Irefn{org143}\And
S.~Qiu\Irefn{org90}\And
L.~Quaglia\Irefn{org25}\And
R.E.~Quishpe\Irefn{org125}\And
S.~Ragoni\Irefn{org111}\And
S.~Raha\Irefn{org3}\And
S.~Rajput\Irefn{org101}\And
J.~Rak\Irefn{org126}\And
A.~Rakotozafindrabe\Irefn{org137}\And
L.~Ramello\Irefn{org31}\And
F.~Rami\Irefn{org136}\And
S.A.R.~Ramirez\Irefn{org45}\And
R.~Raniwala\Irefn{org102}\And
S.~Raniwala\Irefn{org102}\And
S.S.~R\"{a}s\"{a}nen\Irefn{org44}\And
R.~Rath\Irefn{org50}\And
V.~Ratza\Irefn{org43}\And
I.~Ravasenga\Irefn{org90}\And
K.F.~Read\Irefn{org96}\textsuperscript{,}\Irefn{org130}\And
A.R.~Redelbach\Irefn{org39}\And
K.~Redlich\Irefn{org85}\Aref{orgV}\And
A.~Rehman\Irefn{org21}\And
P.~Reichelt\Irefn{org68}\And
F.~Reidt\Irefn{org34}\And
X.~Ren\Irefn{org6}\And
R.~Renfordt\Irefn{org68}\And
Z.~Rescakova\Irefn{org38}\And
K.~Reygers\Irefn{org104}\And
A.~Riabov\Irefn{org98}\And
V.~Riabov\Irefn{org98}\And
T.~Richert\Irefn{org81}\textsuperscript{,}\Irefn{org89}\And
M.~Richter\Irefn{org20}\And
P.~Riedler\Irefn{org34}\And
W.~Riegler\Irefn{org34}\And
F.~Riggi\Irefn{org27}\And
C.~Ristea\Irefn{org67}\And
S.P.~Rode\Irefn{org50}\And
M.~Rodr\'{i}guez Cahuantzi\Irefn{org45}\And
K.~R{\o}ed\Irefn{org20}\And
R.~Rogalev\Irefn{org91}\And
E.~Rogochaya\Irefn{org75}\And
D.~Rohr\Irefn{org34}\And
D.~R\"ohrich\Irefn{org21}\And
P.F.~Rojas\Irefn{org45}\And
P.S.~Rokita\Irefn{org142}\And
F.~Ronchetti\Irefn{org52}\And
A.~Rosano\Irefn{org56}\And
E.D.~Rosas\Irefn{org69}\And
K.~Roslon\Irefn{org142}\And
A.~Rossi\Irefn{org28}\textsuperscript{,}\Irefn{org57}\And
A.~Rotondi\Irefn{org139}\And
A.~Roy\Irefn{org50}\And
P.~Roy\Irefn{org110}\And
O.V.~Rueda\Irefn{org81}\And
R.~Rui\Irefn{org24}\And
B.~Rumyantsev\Irefn{org75}\And
A.~Rustamov\Irefn{org87}\And
E.~Ryabinkin\Irefn{org88}\And
Y.~Ryabov\Irefn{org98}\And
A.~Rybicki\Irefn{org118}\And
H.~Rytkonen\Irefn{org126}\And
O.A.M.~Saarimaki\Irefn{org44}\And
R.~Sadek\Irefn{org115}\And
S.~Sadhu\Irefn{org141}\And
S.~Sadovsky\Irefn{org91}\And
K.~\v{S}afa\v{r}\'{\i}k\Irefn{org37}\And
S.K.~Saha\Irefn{org141}\And
B.~Sahoo\Irefn{org49}\And
P.~Sahoo\Irefn{org49}\And
R.~Sahoo\Irefn{org50}\And
S.~Sahoo\Irefn{org65}\And
P.K.~Sahu\Irefn{org65}\And
J.~Saini\Irefn{org141}\And
S.~Sakai\Irefn{org133}\And
S.~Sambyal\Irefn{org101}\And
V.~Samsonov\Irefn{org93}\textsuperscript{,}\Irefn{org98}\And
D.~Sarkar\Irefn{org143}\And
N.~Sarkar\Irefn{org141}\And
P.~Sarma\Irefn{org42}\And
V.M.~Sarti\Irefn{org105}\And
M.H.P.~Sas\Irefn{org63}\And
E.~Scapparone\Irefn{org54}\And
J.~Schambach\Irefn{org119}\And
H.S.~Scheid\Irefn{org68}\And
C.~Schiaua\Irefn{org48}\And
R.~Schicker\Irefn{org104}\And
A.~Schmah\Irefn{org104}\And
C.~Schmidt\Irefn{org107}\And
H.R.~Schmidt\Irefn{org103}\And
M.O.~Schmidt\Irefn{org104}\And
M.~Schmidt\Irefn{org103}\And
N.V.~Schmidt\Irefn{org68}\textsuperscript{,}\Irefn{org96}\And
A.R.~Schmier\Irefn{org130}\And
J.~Schukraft\Irefn{org89}\And
Y.~Schutz\Irefn{org136}\And
K.~Schwarz\Irefn{org107}\And
K.~Schweda\Irefn{org107}\And
G.~Scioli\Irefn{org26}\And
E.~Scomparin\Irefn{org59}\And
J.E.~Seger\Irefn{org15}\And
Y.~Sekiguchi\Irefn{org132}\And
D.~Sekihata\Irefn{org132}\And
I.~Selyuzhenkov\Irefn{org93}\textsuperscript{,}\Irefn{org107}\And
S.~Senyukov\Irefn{org136}\And
D.~Serebryakov\Irefn{org62}\And
A.~Sevcenco\Irefn{org67}\And
A.~Shabanov\Irefn{org62}\And
A.~Shabetai\Irefn{org115}\And
R.~Shahoyan\Irefn{org34}\And
W.~Shaikh\Irefn{org110}\And
A.~Shangaraev\Irefn{org91}\And
A.~Sharma\Irefn{org100}\And
A.~Sharma\Irefn{org101}\And
H.~Sharma\Irefn{org118}\And
M.~Sharma\Irefn{org101}\And
N.~Sharma\Irefn{org100}\And
S.~Sharma\Irefn{org101}\And
O.~Sheibani\Irefn{org125}\And
K.~Shigaki\Irefn{org46}\And
M.~Shimomura\Irefn{org83}\And
S.~Shirinkin\Irefn{org92}\And
Q.~Shou\Irefn{org40}\And
Y.~Sibiriak\Irefn{org88}\And
S.~Siddhanta\Irefn{org55}\And
T.~Siemiarczuk\Irefn{org85}\And
D.~Silvermyr\Irefn{org81}\And
G.~Simatovic\Irefn{org90}\And
G.~Simonetti\Irefn{org34}\And
B.~Singh\Irefn{org105}\And
R.~Singh\Irefn{org86}\And
R.~Singh\Irefn{org101}\And
R.~Singh\Irefn{org50}\And
V.K.~Singh\Irefn{org141}\And
V.~Singhal\Irefn{org141}\And
T.~Sinha\Irefn{org110}\And
B.~Sitar\Irefn{org13}\And
M.~Sitta\Irefn{org31}\And
T.B.~Skaali\Irefn{org20}\And
M.~Slupecki\Irefn{org44}\And
N.~Smirnov\Irefn{org146}\And
R.J.M.~Snellings\Irefn{org63}\And
C.~Soncco\Irefn{org112}\And
J.~Song\Irefn{org125}\And
A.~Songmoolnak\Irefn{org116}\And
F.~Soramel\Irefn{org28}\And
S.~Sorensen\Irefn{org130}\And
I.~Sputowska\Irefn{org118}\And
J.~Stachel\Irefn{org104}\And
I.~Stan\Irefn{org67}\And
P.J.~Steffanic\Irefn{org130}\And
E.~Stenlund\Irefn{org81}\And
S.F.~Stiefelmaier\Irefn{org104}\And
D.~Stocco\Irefn{org115}\And
M.M.~Storetvedt\Irefn{org36}\And
L.D.~Stritto\Irefn{org29}\And
A.A.P.~Suaide\Irefn{org121}\And
T.~Sugitate\Irefn{org46}\And
C.~Suire\Irefn{org78}\And
M.~Suleymanov\Irefn{org14}\And
M.~Suljic\Irefn{org34}\And
R.~Sultanov\Irefn{org92}\And
M.~\v{S}umbera\Irefn{org95}\And
V.~Sumberia\Irefn{org101}\And
S.~Sumowidagdo\Irefn{org51}\And
S.~Swain\Irefn{org65}\And
A.~Szabo\Irefn{org13}\And
I.~Szarka\Irefn{org13}\And
U.~Tabassam\Irefn{org14}\And
S.F.~Taghavi\Irefn{org105}\And
G.~Taillepied\Irefn{org134}\And
J.~Takahashi\Irefn{org122}\And
G.J.~Tambave\Irefn{org21}\And
S.~Tang\Irefn{org6}\textsuperscript{,}\Irefn{org134}\And
M.~Tarhini\Irefn{org115}\And
M.G.~Tarzila\Irefn{org48}\And
A.~Tauro\Irefn{org34}\And
G.~Tejeda Mu\~{n}oz\Irefn{org45}\And
A.~Telesca\Irefn{org34}\And
L.~Terlizzi\Irefn{org25}\And
C.~Terrevoli\Irefn{org125}\And
D.~Thakur\Irefn{org50}\And
S.~Thakur\Irefn{org141}\And
D.~Thomas\Irefn{org119}\And
F.~Thoresen\Irefn{org89}\And
R.~Tieulent\Irefn{org135}\And
A.~Tikhonov\Irefn{org62}\And
A.R.~Timmins\Irefn{org125}\And
A.~Toia\Irefn{org68}\And
N.~Topilskaya\Irefn{org62}\And
M.~Toppi\Irefn{org52}\And
F.~Torales-Acosta\Irefn{org19}\And
S.R.~Torres\Irefn{org37}\And
A.~Trifir\'{o}\Irefn{org32}\textsuperscript{,}\Irefn{org56}\And
S.~Tripathy\Irefn{org50}\textsuperscript{,}\Irefn{org69}\And
T.~Tripathy\Irefn{org49}\And
S.~Trogolo\Irefn{org28}\And
G.~Trombetta\Irefn{org33}\And
L.~Tropp\Irefn{org38}\And
V.~Trubnikov\Irefn{org2}\And
W.H.~Trzaska\Irefn{org126}\And
T.P.~Trzcinski\Irefn{org142}\And
B.A.~Trzeciak\Irefn{org37}\textsuperscript{,}\Irefn{org63}\And
A.~Tumkin\Irefn{org109}\And
R.~Turrisi\Irefn{org57}\And
T.S.~Tveter\Irefn{org20}\And
K.~Ullaland\Irefn{org21}\And
E.N.~Umaka\Irefn{org125}\And
A.~Uras\Irefn{org135}\And
G.L.~Usai\Irefn{org23}\And
M.~Vala\Irefn{org38}\And
N.~Valle\Irefn{org139}\And
S.~Vallero\Irefn{org59}\And
N.~van der Kolk\Irefn{org63}\And
L.V.R.~van Doremalen\Irefn{org63}\And
M.~van Leeuwen\Irefn{org63}\And
P.~Vande Vyvre\Irefn{org34}\And
D.~Varga\Irefn{org145}\And
Z.~Varga\Irefn{org145}\And
M.~Varga-Kofarago\Irefn{org145}\And
A.~Vargas\Irefn{org45}\And
M.~Vasileiou\Irefn{org84}\And
A.~Vasiliev\Irefn{org88}\And
O.~V\'azquez Doce\Irefn{org105}\And
V.~Vechernin\Irefn{org113}\And
E.~Vercellin\Irefn{org25}\And
S.~Vergara Lim\'on\Irefn{org45}\And
L.~Vermunt\Irefn{org63}\And
R.~Vernet\Irefn{org7}\And
R.~V\'ertesi\Irefn{org145}\And
L.~Vickovic\Irefn{org35}\And
Z.~Vilakazi\Irefn{org131}\And
O.~Villalobos Baillie\Irefn{org111}\And
G.~Vino\Irefn{org53}\And
A.~Vinogradov\Irefn{org88}\And
T.~Virgili\Irefn{org29}\And
V.~Vislavicius\Irefn{org89}\And
A.~Vodopyanov\Irefn{org75}\And
B.~Volkel\Irefn{org34}\And
M.A.~V\"{o}lkl\Irefn{org103}\And
K.~Voloshin\Irefn{org92}\And
S.A.~Voloshin\Irefn{org143}\And
G.~Volpe\Irefn{org33}\And
B.~von Haller\Irefn{org34}\And
I.~Vorobyev\Irefn{org105}\And
D.~Voscek\Irefn{org117}\And
J.~Vrl\'{a}kov\'{a}\Irefn{org38}\And
B.~Wagner\Irefn{org21}\And
M.~Weber\Irefn{org114}\And
S.G.~Weber\Irefn{org144}\And
A.~Wegrzynek\Irefn{org34}\And
S.C.~Wenzel\Irefn{org34}\And
J.P.~Wessels\Irefn{org144}\And
J.~Wiechula\Irefn{org68}\And
J.~Wikne\Irefn{org20}\And
G.~Wilk\Irefn{org85}\And
J.~Wilkinson\Irefn{org10}\And
G.A.~Willems\Irefn{org144}\And
E.~Willsher\Irefn{org111}\And
B.~Windelband\Irefn{org104}\And
M.~Winn\Irefn{org137}\And
W.E.~Witt\Irefn{org130}\And
J.R.~Wright\Irefn{org119}\And
Y.~Wu\Irefn{org128}\And
R.~Xu\Irefn{org6}\And
S.~Yalcin\Irefn{org77}\And
Y.~Yamaguchi\Irefn{org46}\And
K.~Yamakawa\Irefn{org46}\And
S.~Yang\Irefn{org21}\And
S.~Yano\Irefn{org137}\And
Z.~Yin\Irefn{org6}\And
H.~Yokoyama\Irefn{org63}\And
I.-K.~Yoo\Irefn{org17}\And
J.H.~Yoon\Irefn{org61}\And
S.~Yuan\Irefn{org21}\And
A.~Yuncu\Irefn{org104}\And
V.~Yurchenko\Irefn{org2}\And
V.~Zaccolo\Irefn{org24}\And
A.~Zaman\Irefn{org14}\And
C.~Zampolli\Irefn{org34}\And
H.J.C.~Zanoli\Irefn{org63}\And
N.~Zardoshti\Irefn{org34}\And
A.~Zarochentsev\Irefn{org113}\And
P.~Z\'{a}vada\Irefn{org66}\And
N.~Zaviyalov\Irefn{org109}\And
H.~Zbroszczyk\Irefn{org142}\And
M.~Zhalov\Irefn{org98}\And
S.~Zhang\Irefn{org40}\And
X.~Zhang\Irefn{org6}\And
Z.~Zhang\Irefn{org6}\And
V.~Zherebchevskii\Irefn{org113}\And
Y.~Zhi\Irefn{org12}\And
D.~Zhou\Irefn{org6}\And
Y.~Zhou\Irefn{org89}\And
Z.~Zhou\Irefn{org21}\And
J.~Zhu\Irefn{org6}\textsuperscript{,}\Irefn{org107}\And
Y.~Zhu\Irefn{org6}\And
A.~Zichichi\Irefn{org10}\textsuperscript{,}\Irefn{org26}\And
G.~Zinovjev\Irefn{org2}\And
N.~Zurlo\Irefn{org140}\And
\renewcommand\labelenumi{\textsuperscript{\theenumi}~}
\section*{Affiliation notes}
\renewcommand\theenumi{\roman{enumi}}
\begin{Authlist}
\item \Adef{org*}Deceased
\item \Adef{orgI}Italian National Agency for New Technologies, Energy and Sustainable Economic Development (ENEA), Bologna, Italy
\item \Adef{orgII}Dipartimento DET del Politecnico di Torino, Turin, Italy
\item \Adef{orgIII}M.V. Lomonosov Moscow State University, D.V. Skobeltsyn Institute of Nuclear, Physics, Moscow, Russia
\item \Adef{orgIV}Department of Applied Physics, Aligarh Muslim University, Aligarh, India
\item \Adef{orgV}Institute of Theoretical Physics, University of Wroclaw, Poland
\end{Authlist}
\section*{Collaboration Institutes}
\renewcommand\theenumi{\arabic{enumi}~}
\begin{Authlist}
\item \Idef{org1}A.I. Alikhanyan National Science Laboratory (Yerevan Physics Institute) Foundation, Yerevan, Armenia
\item \Idef{org2}Bogolyubov Institute for Theoretical Physics, National Academy of Sciences of Ukraine, Kiev, Ukraine
\item \Idef{org3}Bose Institute, Department of Physics and Centre for Astroparticle Physics and Space Science (CAPSS), Kolkata, India
\item \Idef{org4}Budker Institute for Nuclear Physics, Novosibirsk, Russia
\item \Idef{org5}California Polytechnic State University, San Luis Obispo, California, United States
\item \Idef{org6}Central China Normal University, Wuhan, China
\item \Idef{org7}Centre de Calcul de l'IN2P3, Villeurbanne, Lyon, France
\item \Idef{org8}Centro de Aplicaciones Tecnol\'{o}gicas y Desarrollo Nuclear (CEADEN), Havana, Cuba
\item \Idef{org9}Centro de Investigaci\'{o}n y de Estudios Avanzados (CINVESTAV), Mexico City and M\'{e}rida, Mexico
\item \Idef{org10}Centro Fermi - Museo Storico della Fisica e Centro Studi e Ricerche ``Enrico Fermi', Rome, Italy
\item \Idef{org11}Chicago State University, Chicago, Illinois, United States
\item \Idef{org12}China Institute of Atomic Energy, Beijing, China
\item \Idef{org13}Comenius University Bratislava, Faculty of Mathematics, Physics and Informatics, Bratislava, Slovakia
\item \Idef{org14}COMSATS University Islamabad, Islamabad, Pakistan
\item \Idef{org15}Creighton University, Omaha, Nebraska, United States
\item \Idef{org16}Department of Physics, Aligarh Muslim University, Aligarh, India
\item \Idef{org17}Department of Physics, Pusan National University, Pusan, Republic of Korea
\item \Idef{org18}Department of Physics, Sejong University, Seoul, Republic of Korea
\item \Idef{org19}Department of Physics, University of California, Berkeley, California, United States
\item \Idef{org20}Department of Physics, University of Oslo, Oslo, Norway
\item \Idef{org21}Department of Physics and Technology, University of Bergen, Bergen, Norway
\item \Idef{org22}Dipartimento di Fisica dell'Universit\`{a} 'La Sapienza' and Sezione INFN, Rome, Italy
\item \Idef{org23}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Cagliari, Italy
\item \Idef{org24}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Trieste, Italy
\item \Idef{org25}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Turin, Italy
\item \Idef{org26}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Bologna, Italy
\item \Idef{org27}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Catania, Italy
\item \Idef{org28}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Padova, Italy
\item \Idef{org29}Dipartimento di Fisica `E.R.~Caianiello' dell'Universit\`{a} and Gruppo Collegato INFN, Salerno, Italy
\item \Idef{org30}Dipartimento DISAT del Politecnico and Sezione INFN, Turin, Italy
\item \Idef{org31}Dipartimento di Scienze e Innovazione Tecnologica dell'Universit\`{a} del Piemonte Orientale and INFN Sezione di Torino, Alessandria, Italy
\item \Idef{org32}Dipartimento di Scienze MIFT, Universit\`{a} di Messina, Messina, Italy
\item \Idef{org33}Dipartimento Interateneo di Fisica `M.~Merlin' and Sezione INFN, Bari, Italy
\item \Idef{org34}European Organization for Nuclear Research (CERN), Geneva, Switzerland
\item \Idef{org35}Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split, Split, Croatia
\item \Idef{org36}Faculty of Engineering and Science, Western Norway University of Applied Sciences, Bergen, Norway
\item \Idef{org37}Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czech Republic
\item \Idef{org38}Faculty of Science, P.J.~\v{S}af\'{a}rik University, Ko\v{s}ice, Slovakia
\item \Idef{org39}Frankfurt Institute for Advanced Studies, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org40}Fudan University, Shanghai, China
\item \Idef{org41}Gangneung-Wonju National University, Gangneung, Republic of Korea
\item \Idef{org42}Gauhati University, Department of Physics, Guwahati, India
\item \Idef{org43}Helmholtz-Institut f\"{u}r Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universit\"{a}t Bonn, Bonn, Germany
\item \Idef{org44}Helsinki Institute of Physics (HIP), Helsinki, Finland
\item \Idef{org45}High Energy Physics Group, Universidad Aut\'{o}noma de Puebla, Puebla, Mexico
\item \Idef{org46}Hiroshima University, Hiroshima, Japan
\item \Idef{org47}Hochschule Worms, Zentrum f\"{u}r Technologietransfer und Telekommunikation (ZTT), Worms, Germany
\item \Idef{org48}Horia Hulubei National Institute of Physics and Nuclear Engineering, Bucharest, Romania
\item \Idef{org49}Indian Institute of Technology Bombay (IIT), Mumbai, India
\item \Idef{org50}Indian Institute of Technology Indore, Indore, India
\item \Idef{org51}Indonesian Institute of Sciences, Jakarta, Indonesia
\item \Idef{org52}INFN, Laboratori Nazionali di Frascati, Frascati, Italy
\item \Idef{org53}INFN, Sezione di Bari, Bari, Italy
\item \Idef{org54}INFN, Sezione di Bologna, Bologna, Italy
\item \Idef{org55}INFN, Sezione di Cagliari, Cagliari, Italy
\item \Idef{org56}INFN, Sezione di Catania, Catania, Italy
\item \Idef{org57}INFN, Sezione di Padova, Padova, Italy
\item \Idef{org58}INFN, Sezione di Roma, Rome, Italy
\item \Idef{org59}INFN, Sezione di Torino, Turin, Italy
\item \Idef{org60}INFN, Sezione di Trieste, Trieste, Italy
\item \Idef{org61}Inha University, Incheon, Republic of Korea
\item \Idef{org62}Institute for Nuclear Research, Academy of Sciences, Moscow, Russia
\item \Idef{org63}Institute for Subatomic Physics, Utrecht University/Nikhef, Utrecht, Netherlands
\item \Idef{org64}Institute of Experimental Physics, Slovak Academy of Sciences, Ko\v{s}ice, Slovakia
\item \Idef{org65}Institute of Physics, Homi Bhabha National Institute, Bhubaneswar, India
\item \Idef{org66}Institute of Physics of the Czech Academy of Sciences, Prague, Czech Republic
\item \Idef{org67}Institute of Space Science (ISS), Bucharest, Romania
\item \Idef{org68}Institut f\"{u}r Kernphysik, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org69}Instituto de Ciencias Nucleares, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org70}Instituto de F\'{i}sica, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil
\item \Idef{org71}Instituto de F\'{\i}sica, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org72}iThemba LABS, National Research Foundation, Somerset West, South Africa
\item \Idef{org73}Jeonbuk National University, Jeonju, Republic of Korea
\item \Idef{org74}Johann-Wolfgang-Goethe Universit\"{a}t Frankfurt Institut f\"{u}r Informatik, Fachbereich Informatik und Mathematik, Frankfurt, Germany
\item \Idef{org75}Joint Institute for Nuclear Research (JINR), Dubna, Russia
\item \Idef{org76}Korea Institute of Science and Technology Information, Daejeon, Republic of Korea
\item \Idef{org77}KTO Karatay University, Konya, Turkey
\item \Idef{org78}Laboratoire de Physique des 2 Infinis, Ir\`{e}ne Joliot-Curie, Orsay, France
\item \Idef{org79}Laboratoire de Physique Subatomique et de Cosmologie, Universit\'{e} Grenoble-Alpes, CNRS-IN2P3, Grenoble, France
\item \Idef{org80}Lawrence Berkeley National Laboratory, Berkeley, California, United States
\item \Idef{org81}Lund University Department of Physics, Division of Particle Physics, Lund, Sweden
\item \Idef{org82}Nagasaki Institute of Applied Science, Nagasaki, Japan
\item \Idef{org83}Nara Women{'}s University (NWU), Nara, Japan
\item \Idef{org84}National and Kapodistrian University of Athens, School of Science, Department of Physics , Athens, Greece
\item \Idef{org85}National Centre for Nuclear Research, Warsaw, Poland
\item \Idef{org86}National Institute of Science Education and Research, Homi Bhabha National Institute, Jatni, India
\item \Idef{org87}National Nuclear Research Center, Baku, Azerbaijan
\item \Idef{org88}National Research Centre Kurchatov Institute, Moscow, Russia
\item \Idef{org89}Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark
\item \Idef{org90}Nikhef, National institute for subatomic physics, Amsterdam, Netherlands
\item \Idef{org91}NRC Kurchatov Institute IHEP, Protvino, Russia
\item \Idef{org92}NRC \guillemotleft Kurchatov\guillemotright~Institute - ITEP, Moscow, Russia
\item \Idef{org93}NRNU Moscow Engineering Physics Institute, Moscow, Russia
\item \Idef{org94}Nuclear Physics Group, STFC Daresbury Laboratory, Daresbury, United Kingdom
\item \Idef{org95}Nuclear Physics Institute of the Czech Academy of Sciences, \v{R}e\v{z} u Prahy, Czech Republic
\item \Idef{org96}Oak Ridge National Laboratory, Oak Ridge, Tennessee, United States
\item \Idef{org97}Ohio State University, Columbus, Ohio, United States
\item \Idef{org98}Petersburg Nuclear Physics Institute, Gatchina, Russia
\item \Idef{org99}Physics department, Faculty of science, University of Zagreb, Zagreb, Croatia
\item \Idef{org100}Physics Department, Panjab University, Chandigarh, India
\item \Idef{org101}Physics Department, University of Jammu, Jammu, India
\item \Idef{org102}Physics Department, University of Rajasthan, Jaipur, India
\item \Idef{org103}Physikalisches Institut, Eberhard-Karls-Universit\"{a}t T\"{u}bingen, T\"{u}bingen, Germany
\item \Idef{org104}Physikalisches Institut, Ruprecht-Karls-Universit\"{a}t Heidelberg, Heidelberg, Germany
\item \Idef{org105}Physik Department, Technische Universit\"{a}t M\"{u}nchen, Munich, Germany
\item \Idef{org106}Politecnico di Bari, Bari, Italy
\item \Idef{org107}Research Division and ExtreMe Matter Institute EMMI, GSI Helmholtzzentrum f\"ur Schwerionenforschung GmbH, Darmstadt, Germany
\item \Idef{org108}Rudjer Bo\v{s}kovi\'{c} Institute, Zagreb, Croatia
\item \Idef{org109}Russian Federal Nuclear Center (VNIIEF), Sarov, Russia
\item \Idef{org110}Saha Institute of Nuclear Physics, Homi Bhabha National Institute, Kolkata, India
\item \Idef{org111}School of Physics and Astronomy, University of Birmingham, Birmingham, United Kingdom
\item \Idef{org112}Secci\'{o}n F\'{\i}sica, Departamento de Ciencias, Pontificia Universidad Cat\'{o}lica del Per\'{u}, Lima, Peru
\item \Idef{org113}St. Petersburg State University, St. Petersburg, Russia
\item \Idef{org114}Stefan Meyer Institut f\"{u}r Subatomare Physik (SMI), Vienna, Austria
\item \Idef{org115}SUBATECH, IMT Atlantique, Universit\'{e} de Nantes, CNRS-IN2P3, Nantes, France
\item \Idef{org116}Suranaree University of Technology, Nakhon Ratchasima, Thailand
\item \Idef{org117}Technical University of Ko\v{s}ice, Ko\v{s}ice, Slovakia
\item \Idef{org118}The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland
\item \Idef{org119}The University of Texas at Austin, Austin, Texas, United States
\item \Idef{org120}Universidad Aut\'{o}noma de Sinaloa, Culiac\'{a}n, Mexico
\item \Idef{org121}Universidade de S\~{a}o Paulo (USP), S\~{a}o Paulo, Brazil
\item \Idef{org122}Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil
\item \Idef{org123}Universidade Federal do ABC, Santo Andre, Brazil
\item \Idef{org124}University of Cape Town, Cape Town, South Africa
\item \Idef{org125}University of Houston, Houston, Texas, United States
\item \Idef{org126}University of Jyv\"{a}skyl\"{a}, Jyv\"{a}skyl\"{a}, Finland
\item \Idef{org127}University of Liverpool, Liverpool, United Kingdom
\item \Idef{org128}University of Science and Technology of China, Hefei, China
\item \Idef{org129}University of South-Eastern Norway, Tonsberg, Norway
\item \Idef{org130}University of Tennessee, Knoxville, Tennessee, United States
\item \Idef{org131}University of the Witwatersrand, Johannesburg, South Africa
\item \Idef{org132}University of Tokyo, Tokyo, Japan
\item \Idef{org133}University of Tsukuba, Tsukuba, Japan
\item \Idef{org134}Universit\'{e} Clermont Auvergne, CNRS/IN2P3, LPC, Clermont-Ferrand, France
\item \Idef{org135}Universit\'{e} de Lyon, Universit\'{e} Lyon 1, CNRS/IN2P3, IPN-Lyon, Villeurbanne, Lyon, France
\item \Idef{org136}Universit\'{e} de Strasbourg, CNRS, IPHC UMR 7178, F-67000 Strasbourg, France, Strasbourg, France
\item \Idef{org137}Universit\'{e} Paris-Saclay Centre d'Etudes de Saclay (CEA), IRFU, D\'{e}partment de Physique Nucl\'{e}aire (DPhN), Saclay, France
\item \Idef{org138}Universit\`{a} degli Studi di Foggia, Foggia, Italy
\item \Idef{org139}Universit\`{a} degli Studi di Pavia, Pavia, Italy
\item \Idef{org140}Universit\`{a} di Brescia, Brescia, Italy
\item \Idef{org141}Variable Energy Cyclotron Centre, Homi Bhabha National Institute, Kolkata, India
\item \Idef{org142}Warsaw University of Technology, Warsaw, Poland
\item \Idef{org143}Wayne State University, Detroit, Michigan, United States
\item \Idef{org144}Westf\"{a}lische Wilhelms-Universit\"{a}t M\"{u}nster, Institut f\"{u}r Kernphysik, M\"{u}nster, Germany
\item \Idef{org145}Wigner Research Centre for Physics, Budapest, Hungary
\item \Idef{org146}Yale University, New Haven, Connecticut, United States
\item \Idef{org147}Yonsei University, Seoul, Republic of Korea
\end{Authlist}
\endgroup
\section{Introduction} \label{section:introduction}
Ultra-relativistic heavy-ion collisions are the means to create under laboratory conditions the deconfined state of strongly-interacting matter called quark--gluon plasma (QGP). This state behaves like an ideal fluid with a shear viscosity to entropy ratio approaching the conjectured lowest possible value of $\hslash/(4\pi k_{\rm B})$ \cite{Kovtun:2004de,Voloshin:2008dg,Romatschke:2009im}.
One of the most important observables for studying the properties of the QGP is the azimuthal dependence of particle production, also called anisotropic flow, quantified in terms of a Fourier expansion with respect to the azimuthal angle of the initial state symmetry plane for the $n$-th harmonic $\Psi_{\mathrm n}$ as
\begin{equation}
\frac{{\mathrm d}N}{{\mathrm d} \varphi} \propto 1 + 2 \sum_{n=1}^{+ \infty} \ensuremath{v_{\mathrm n}}\xspace\cos \left[n(\varphi-\Psi_{\mathrm n}) \right],
\label{eq:fourier_def_vn}
\end{equation}
where \ensuremath{v_{\mathrm n}}\xspace is the $n$-th order harmonic coefficient and $\varphi$ is the azimuthal angle of the particles.
The initial state spatial anisotropy of the collision overlap region is transformed into a momentum anisotropy of the produced final state particles~\cite{Ollitrault:1992bk,Voloshin:1994mz,Qiu:2011iv,Teaney:2009qa}. The medium response to the initial state anisotropy ($\varepsilon_{\mathrm n}$), which is transformed into the \ensuremath{v_{\mathrm n}}\xspace coefficients, strongly depends on the macroscopic properties of the fireball, like the temperature dependent equation of state and the shear and bulk viscosity.
The dominant source of anisotropy is the ellipsoidal shape of the overlap region in non-central collisions that have a non-zero finite impact parameter (transverse distance separating the centers of the two nuclei), which gives rise to a large second order harmonic coefficient, \ensuremath{v_{\mathrm 2}}\xspace, also known as elliptic flow. Fluctuations in the initial energy-density profile within the overlap region are thought to be the origin of the triangular flow, \ensuremath{v_{\mathrm 3}}\xspace~\cite{Luzum:2008cw,Alver:2010gr,Teaney:2010vd}. Higher order harmonics are strongly damped, do not depend linearly on the initial anisotropy, and have significant contributions from the interplay of lower order harmonics~\cite{Niemi:2012aj,Gardim:2011xv,Gardim:2014tya,Acharya:2017zfg,Acharya:2020taj}.
The ALICE Collaboration published extensive studies of anisotropic flow measurements for identified light and strange particles~\cite{Abelev:2014pua,Acharya:2018zuq}. Flow coefficients for all particles show, in the low \pt range, an increasing trend with \pt mainly attributed to the radial hydrodynamic expansion of the QGP, reach a maximum in the \pt range 3--5 GeV/$c$\xspace depending on the particle mass and species, and finally drop towards higher \pt. The behavior in the high-\pt region is commonly attributed to path-length dependent effects like energy loss~\cite{Betz:2016ayq,CMS:2012aa,Armesto:2005iq}.
At both RHIC and LHC energies, an approximate scaling of the flow coefficients with the number of valence quarks is observed for light and strange particles~\cite{Adams:2003am,Afanasiev:2007tv,Adamczyk:2015fum,Abelev:2014pua,Acharya:2018zuq}. In the low to moderate \pt range (approximately $3 < \pt < 8$~GeV/$c$\xspace), this scaling is hypothesized to be the consequence of the hadronization process via quark coalescence and of a common underlying partonic flow during the hydrodynamic stage of the collision~\cite{Molnar:2003ff,Lin:2003jy,Adams:2005zg,Fries:2008hs,Adams:2004bi}.
The production of charmonia, and especially of \mbox{J/$\psi$}\xspace, is one of the first proposed probes of the QGP properties, in particular the deconfinement~\cite{Matsui:1986dk}. Since charm quarks are produced during the early hard partonic collisions, they experience the entire evolution of the fireball. At the same time, their initial production cross section can be calculated in perturbative quantum chromodynamics (QCD). The suppression of the production of bound charmonium states by the free color charges of the dense deconfined medium is sensitive to both the medium bulk characteristics~\cite{Digal:2001ue,Rothkopf:2019ipj} and to the microscopic ones, like the charm-quark diffusion coefficient~\cite{Riek:2010fk,Scardina:2017ipo}.
Measurements of the \mbox{J/$\psi$}\xspace nuclear modification factor \ensuremath{R_{\mathrm{AA}}}~at RHIC in Au--Au collisions at $\ensuremath{\sqrt{s_{\mathrm{NN}}}}=200$ GeV~\cite{Adare:2006ns} indicated a strong nuclear suppression especially for the most central collisions. At the LHC, in Pb--Pb collisions at $\ensuremath{\sqrt{s_{\mathrm{NN}}}}=2.76$ and 5.02~TeV, the ALICE Collaboration reported a much larger \ensuremath{R_{\mathrm{AA}}}~compared to the one observed at RHIC~\cite{Abelev:2012rv,Abelev:2013ila,Adam:2016rdg}, despite the higher energy density present in the system. This effect is concentrated in the low-\pt region, which is consistent with charmonium regeneration by recombination of charm quarks, either at the QGP phase boundary via statistical hadronization~\cite{BraunMunzinger:2000px} or continuously throughout the fireball evolution~\cite{Du:2015wha,Du:2017qkv,Zhou:2014kka}.
Within the statistical hadronization scenario, charm quarks thermalize in the QGP and all of the charmed bound hadrons are created at the phase boundary assuming chemical equilibration~\cite{BraunMunzinger:2000px,Andronic:2019wva}, except a small fraction created in the fireball corona that escape the medium. In transport model approaches, where charm quarks reach only a partial thermalization, roughly 50\% of the produced \mbox{J/$\psi$}\xspace originate from the recombination process, while the rest comes from primordial production~\cite{Du:2015wha,Du:2017qkv,Zhou:2014kka}. In both phenomenological approaches, it is expected that charm quarks will inherit some of the medium radial and anisotropic flow. Indeed, a significant D-meson~\cite{Sirunyan:2017plt,Acharya:2018bxo,Acharya:2017qps} and \mbox{J/$\psi$}\xspace elliptic flow~\cite{Khachatryan:2016ypw,Acharya:2017tgv,Acharya:2018pjd,Aaboud:2018ttm} was already observed at the LHC, indicating a hierarchy between the flow of charged particles, D and \mbox{J/$\psi$}\xspace mesons, with the \mbox{J/$\psi$}\xspace flow being the smallest. A positive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace observed also at high \pt, typically underestimated by transport model calculations, might suggest the presence of important path length dependent effects like energy loss and the survival probability in the medium~\cite{Arleo:2017ntr,Spousta:2016agr}. In addition to \ensuremath{v_{\mathrm 2}}\xspace, the ALICE Collaboration also published in Ref.~\cite{Acharya:2018pjd} an evidence of a positive \mbox{J/$\psi$}\xspace\ensuremath{v_{\mathrm 3}}\xspace with a statistical significance of 3.7$\sigma$.
In this paper, the measurements of inclusive \mbox{J/$\psi$}\xspace\ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace at forward rapidity (2.5 $<y<$ 4) and \ensuremath{v_{\mathrm 2}}\xspace at midrapidity ($|y|<$ 0.9) in Pb--Pb collisions at \fivenn are discussed. Inclusive \mbox{J/$\psi$}\xspace mesons include both a prompt component from direct \mbox{J/$\psi$}\xspace production and decays of excited charmonium states and a non-prompt component from weak decays of beauty hadrons. The results are presented as a function of \pt in several collision centrality classes, expressed in percentages of the total hadronic cross section, and are compared with calculations from a microscopic transport model. The analyzed data include the full LHC Run 2 Pb--Pb data set, which improves the statistical precision with respect to the previous results by approximately a factor of two at forward rapidity~\cite{Acharya:2018pjd}, and a factor of nine (four) in central (semi-central) collisions at midrapidity~\cite{Acharya:2017tgv} allowing the experimental evidence of a statistically significant non-zero \mbox{J/$\psi$}\xspace\ensuremath{v_{\mathrm 2}}\xspace at midrapidity.
\section{Experimental setup, data samples and event selection} \label{section:experimental_setups}
A detailed description of the ALICE apparatus and its performance can be found in Refs.~\cite{Aamodt:2008zz,Abelev:2014ffa}.
At forward rapidity, \mbox{J/$\psi$}\xspace are reconstructed in the $\ensuremath{\rm{\mu}^{+}\rm{\mu}^{-}}\xspace$ decay channel with the muon spectrometer which covers the pseudorapidity range $-4 <\eta< -2.5$~\footnote{In the ALICE reference frame, the muon spectrometer covers a negative $\eta$ range and consequently a negative $y$ range. We have chosen to present our results with a positive $y$ notation, due to the symmetry of the collision system.}.
The spectrometer includes five tracking stations, each composed of two planes of cathode pad chambers. The third station is placed inside a dipole magnet with a 3 Tm field integral.
Two trigger stations, containing two planes of resistive plate chambers each, provide single and dimuon triggers with a programmable single-muon \pt threshold.
A front absorber, made of carbon, concrete, and steel, is placed in between the primary interaction point (IP) and the first tracking station to remove primary hadrons from the collision. A second absorber, made of iron, is placed in front of the trigger chambers to further reject secondary hadrons escaping the front absorber and low-\pt muons, mainly from pion and kaon decays.
An additional conical absorber surrounds the beam pipe to protect the muon spectrometer against secondary particles produced by the interaction of large-$\eta$ particles with the beam pipe.
At midrapidity, J/$\psi$ mesons are reconstructed in the \ee\ decay channel using the Inner Tracking System (ITS)~\cite{Aamodt:2010aa} and the Time Projection Chamber (TPC)~\cite{Alme:2010ke} in the rapidity range $|y|<0.9$. The ITS is a cylindrical-shaped detector, consisting of 6 layers of silicon detectors used for precision tracking, reconstruction of the primary vertex of the event and event selection. The innermost two layers consists of pixels (SPD), the middle two are drift (SDD), while the two outermost layers are equipped with strip detectors (SSD). The tracklets, track segments reconstructed as pairs of hits in the SPD layers pointing to the primary vertex, are used for the determination of the event flow vector.
The TPC is the main detector used for tracking and particle identification and consists of a cylindrical-shaped gas-filled active volume placed around the ITS. Radially, it extends between an inner radius of 0.85 m and an outer radius of 2.5 m, with a total length of 5 m along the beam axis. Particle identification in the TPC is performed via the measurement of the specific energy loss, ${\rm d}E/{\rm d}x$.
Besides the muon spectrometer and the central barrel detectors, a set of detectors for global event characterization are also used. Two arrays of 32 scintillator counters each, covering $2.8 <\eta< 5.1$ (V0A) and $-3.7 <\eta< -1.7$ (V0C)~\cite{Abbas:2013taa}, are used for triggering, beam induced background rejection, and for the determination of the collision centrality. The 32 channels are arranged in four concentric rings with full azimuthal coverage allowing for the calculation of the event flow vector. The centrality of the events, expressed in fractions of the total inelastic hadronic cross section, is determined via a Glauber fit to the V0 amplitude as described in Refs.~\cite{centrality276,centralityNote}. In addition, two neutron Zero Degree Calorimeters~\cite{ALICE:2012aa}, installed at $\pm 112.5$ m from the nominal IP along the beam axis, are used to remove beam induced background events and electromagnetic interactions.
The analyzed data samples were collected by ALICE during the 2015 and 2018 LHC Pb--Pb runs at \fivenn using different trigger strategies for the forward muon spectrometer and the midrapidity detectors.
At forward rapidity, data were collected requiring the coincidence of the minimum bias (MB) and unlike-sign dimuon triggers. The former is defined by the coincidence of signals in the V0A and V0C arrays while the latter requires at least a pair of opposite-sign track segments in the muon trigger stations. The programmable threshold of the muon trigger algorithm was set so that the trigger efficiency for muon tracks with $\pt = 1$~GeV/$c$\xspace is 50\% and reaches a plateau value of about 98\% at \pt $\approx$ 2.5~GeV/$c$\xspace.
In order to study the background, additional samples of single muon and like-sign dimuon events were also collected by requiring, in addition to the MB condition and the low-\pt threshold, at least one or a pair of same-sign track segments in the trigger system, respectively.
At midrapidity, data were collected using the MB trigger during the 2015 data taking period, and the MB, central, and semi-central triggers in the 2018 period. The central and semi-central triggers require the MB trigger to be fired but, in addition, a condition on the total signal amplitude in the V0 detectors, corresponding to collision centralities of 0--10\% and 30--50\%, respectively, was applied.
Both forward and midrapidity analyses require to have a primary vertex position within $\pm 10$ cm from the nominal IP along the beam axis.
Events containing more than one collision (pile-up) are removed by exploiting the correlations between the number of clusters in the SPD, the number of reconstructed SPD tracklets, and the total signal in the V0A and V0C detectors.
At midrapidity, events with pile-up occurring during the drift time of the TPC are rejected in the offline analysis based on the correlation between the number of SDD and SSD clusters and the total number of clusters in the TPC. The beam-induced background is filtered out offline by applying a selection based on the V0 and the ZDC timing information~\cite{Abelev:2013qoq}.
The integrated luminosity of the analyzed data samples is about 750 $\mathrm{\mu}$b$^{-1}$ for the dimuon analysis. For the measurements at midrapidity, the total luminosity recorded depends on the centrality range due to the centrality triggers, and amounts to 93 $\mathrm{\mu}$b$^{-1}$, 41 $\mathrm{\mu}$b$^{-1}$, and 20 $\mathrm{\mu}$b$^{-1}$ for the central, semi-central, and MB triggers, respectively.
\section{Data analysis} \label{section:analysis_details}
The \ensuremath{v_{\mathrm n}}\xspace coefficients are obtained using the scalar product (SP) method~\cite{Adler:2002pu,Voloshin:2008dg}. This is a two-particle correlation technique based on the scalar product between the unit flow vector for a given harmonic $n$, ${\bf u}_{\rm n} = e^{in\varphi }$, of the particle of interest (here a dilepton) and the complex conjugate of the event flow vector in a subdetector A, ${\bf Q}_{\rm n}^{\rm A*}$. The flow coefficients are thus defined as
\begin{equation}
v_{\rm n}\{{\rm SP}\}= \Bigg\langle {{\bf u}_{\rm n} {\bf Q}_{\rm n}^{\rm A *}} \Bigg/ \sqrt{ \frac{\langle {\bf Q}_{\rm n}^{\rm A} {\bf Q}_{\rm n}^{\rm B *} \rangle \langle {\bf Q}_{\rm n}^{\rm A} {\bf Q}_{\rm n}^{\rm C *} \rangle} { \langle {\bf Q}_{\rm n}^{\rm B} {\bf Q}_{\rm n}^{\rm C *} \rangle } } \Bigg\rangle_{\ell\ell},
\label{eq:sp_method}
\end{equation}
where ${\bf Q}_{\rm n}^{\rm B}$ and ${\bf Q}_{\rm n}^{\rm C}$ are the $n$-th harmonic event flow vectors measured in two additional subdetectors, B and C, respectively, which are used to correct the event flow vector via the three sub-event technique~\cite{Luzum:2012da}. The star ($*$) represents the complex conjugate and the bracket $\langle ... \rangle_{\ell\ell}$ indicates the average over dileptons from all events in a given \pt range, dilepton invariant mass (\ensuremath{m_{\rm \ell \ell}}\xspace), and centrality interval.
The brackets $\langle ... \rangle$ in the denominator denote the average over all events in a narrow centrality interval containing the event under consideration. The V0A and V0C detectors are used in the analysis at both rapidities, while the analysis at forward rapidity uses the SPD as the third subdetector, and the analysis at midrapidity uses the TPC. As detector A, the SPD is chosen for the forward analysis and the V0C for the midrapidity one.
The V0A and V0C event flow vectors are calculated using the energy deposition measured in the individual channels. For the SPD and TPC event flow vectors, the reconstructed tracklets and the tracks are used, respectively.
\begin{table}[h!]
\begin{center}
\caption{\label{tab:table-detectors}Summary of the details concerning the dimuon and dielectron analyses, corresponding to the forward and midrapidity region, respectively. The detectors cited in this table are described in Sec.~\ref{section:experimental_setups}, and the details concerning the three sub-event technique is presented in Sec.~\ref{section:analysis_details}.}
\begin{tabular}{cccccc}
\hline
\hline
\multicolumn{2}{c}{Dilepton analysis} & \multicolumn{3}{c}{Three sub-event technique, detectors used} & Corresponding gap between \\
\multicolumn{2}{c}{\mbox{J/$\psi$}\xspace$\rightarrow l^{+}l^{-} $} & A & B & C & ${\bf u}_{\rm n}$ and ${\bf Q}_{\rm n}^{\rm A}$ \\
\hline
$\mu^{+}\mu^{-}$ & $2.5 < y^{\mu\mu} < 4$ & SPD & V0A & V0C & $|\Delta \eta|>$1.1 \\
\hline
$\rm{e}^{+}\rm{e}^{-}$ & $|y^{\rm{ee}}| < 0.9 $ & V0C & TPC & V0A & $|\Delta \eta|>$0.8 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
The effects of non-uniform acceptance of the detectors used for the flow vector determination are corrected through the procedure described in Ref.~\cite{Selyuzhenkov:2007zi}.
As was discussed in Sec.~\ref{section:experimental_setups}, the three detectors used for the event flow determination cover distinct pseudorapidity ranges, allowing for pseudorapidity gaps $\Delta \eta$ between the sub-events used for flow vector determination and dilepton reconstruction.
The pseudorapidity gap between ${\bf u}_{\rm n}$ and ${\bf Q}_{\rm n}^{\rm A}$, corresponding to $|\Delta \eta| > 1.1$ and $|\Delta \eta| > 0.8$ for the dimuon and dielectron analysis, respectively, suppresses the short-range correlations originating from resonance decays or jets (non-flow effects), not related to the global azimuthal anisotropy.
In the dimuon analysis, \mbox{J/$\psi$}\xspace candidates are formed by combining pairs of opposite-sign tracks reconstructed in the geometrical acceptance of the muon spectrometer using the tracking algorithm described in Ref.~\cite{Aamodt:2011gj}. The same single-muon and dimuon selection criteria used in previous analyses~\cite{Adam:2015isa,Acharya:2018pjd} are applied.
Namely, each muon track candidate should have $-4<\eta_{\rm \mu}<-2.5$, a radial transverse position at the end of the front absorber in the range $17.6<R_{\mathrm{abs}}<89.5\;\mathrm{cm}$, and must match a track segment in the muon trigger chambers above the 1~GeV/$c$\xspace \pt threshold.
The rapidity of the muon pair should be within the acceptance of the muon spectrometer ($2.5 < y < 4.0$).
At midrapidity, \mbox{J/$\psi$}\xspace mesons are reconstructed in the dielectron decay channel. Electron candidates are required to be good quality tracks matched in both the ITS and the TPC, and to have a $\pt > 1$~GeV/$c$\xspace and $|\eta|<0.9$. Tracks are selected to have at least 70 space points in the TPC, out of a maximum of 159, and a $\chi^{2}/N_{\rm dof} <$ 2 for the track fit quality. At least one hit in either of the two SPD layers is required to reject secondary electrons from photons converted in the detector material and to improve the tracking resolution. Secondary electrons are further rejected by requiring the distance-of-closest-approach (DCA) to the collision vertex to be smaller than 1~cm and 3~cm in the transverse and longitudinal directions, respectively. Electrons are identified via their specific energy loss in the TPC gas, d$E$/d$x$, by selecting a band of $\pm 3\sigma$ around the expectation value, with $\sigma$ being the d$E$/d$x$ measurement resolution. To reduce further the hadronic contamination, candidate tracks compatible within $\pm 3.5 \sigma$ with the pion or proton hypothesis are rejected.
\begin{figure}[b!]
\begin{center}
\includegraphics[width = 1\textwidth]{figures/Fig1.pdf}
\end{center}
\caption{(Color online) Invariant mass distribution (top panels (a1), (b1)) and $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm \ell \ell}}\xspace)$ (bottom panels (a2), (b2)) for dimuons in the ranges $2 < \pt < 3$~GeV/$c$\xspace (top left) and for dielectrons in $0 < \pt < 4$~GeV/$c$\xspace (top right), for the 30--50\% centrality interval. Fit functions of the invariant mass distributions and $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm \mu \mu}}\xspace)$, as discussed in Sec.~\ref{section:analysis_details}, are also shown.
Bottom panel, invariant mass ((c1)) and $\ensuremath{v_{\mathrm 3}}\xspace(\ensuremath{m_{\rm \mu \mu}}\xspace)$ ((c2)) distributions for dimuons in the \pt range $2 < \pt < 5$~GeV/$c$\xspace for the 0--50\% centrality interval. The $\ensuremath{v_{\mathrm n}}\xspace(\ensuremath{m_{\rm \mu \mu}}\xspace)$ and $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm ee}}\xspace)$ distributions are plotted with the background flow obtained from the event-mixing procedure and the fit function, as discussed in the text. Only statistical uncertainties are shown.}
\label{fig:figures_extraction}
\end{figure}
The flow coefficients are extracted from sequential fits to the dilepton invariant mass distribution, \ensuremath{m_{\rm \ell \ell}}\xspace, and the \ensuremath{v_{\mathrm n}}\xspace as a function of \ensuremath{m_{\rm \ell \ell}}\xspace, which include the superposition of a \mbox{J/$\psi$}\xspace signal and a background contribution, using the function
\begin{equation}
\ensuremath{v_{\mathrm n}}\xspace(\ensuremath{m_{\rm \ell \ell}}\xspace) = \alpha(\ensuremath{m_{\rm \ell \ell}}\xspace) \hspace{0.1cm} \ensuremath{v_{\mathrm n}}\xspace^{{\rm J}/\psi} + [1-\alpha(\ensuremath{m_{\rm \ell \ell}}\xspace)] \hspace{0.1cm} \ensuremath{v_{\mathrm n}}\xspace^{\rm bkg}(\ensuremath{m_{\rm \ell \ell}}\xspace).
\label{eq:vn_mass}
\end{equation}
Here, $\ensuremath{v_{\mathrm n}}\xspace^{{\rm J}/\psi}$ denotes the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace or \ensuremath{v_{\mathrm 3}}\xspace and $\alpha(\ensuremath{m_{\rm \ell \ell}}\xspace)$ is the signal fraction defined as ${\rm S}/{\rm (S+B)}$.
The latter is extracted from fits to the dilepton invariant mass distribution as described below.
The $\ensuremath{v_{\mathrm n}}\xspace^{{\rm bkg}}(\ensuremath{m_{\rm \ell \ell}}\xspace)$ corresponds to the dilepton background \ensuremath{v_{\mathrm 2}}\xspace or \ensuremath{v_{\mathrm 3}}\xspace.
In the dimuon analysis, the \mbox{J/$\psi$}\xspace signal is parameterized using an extended Crystal Ball (CB2) function and the background with a Variable Width Gaussian (VWG) function~\cite{ALICE-PUBLIC-2015-006}.
In the fit, the \mbox{J/$\psi$}\xspace peak position and width are left free, while the CB2 tail parameters are fixed to the values reported in Ref.~\cite{Acharya:2017hjh}.
The signal of the $\psi(2S)$ is not included in the fit of the \ensuremath{v_{\mathrm n}}\xspace coefficients due to its marginal significance.
At midrapidity, the signal fraction is obtained from the dielectron invariant mass distribution in two steps. First, the combinatorial background is estimated using an event mixing technique, where pairs are built from different events with similar collision centrality, flow-vector orientation, and longitudinal position of the event vertex, and then subtracted from the same-event dielectron invariant mass distribution. The combinatorial background normalization is obtained from the ratio of the number of same-event to mixed-event like-sign pairs. Second, the remaining distribution is fitted using a component for the signal and one for the residual background. For the \mbox{J/$\psi$}\xspace signal shape, the dielectron invariant mass distribution obtained from Monte Carlo simulations is used. The residual background, originating mainly from semileptonic decays of ${\rm c\overline{c}}$ and ${\rm b\overline{b}}$ pairs (correlated background) and imperfect matching between the same-event and mixed-event distributions, is parameterized using either a third order polynomial function at low \pt or an exponential function at high \pt.
The \ensuremath{v_{\mathrm n}}\xspace extraction method employed in this work is described in detail in Ref.~\cite{Acharya:2018pjd}, where the $\ensuremath{v_{\mathrm n}}\xspace^{\rm bkg}(\ensuremath{m_{\rm \ell \ell}}\xspace)$ distribution is obtained using an event mixing technique.
There, it was first demonstrated that the flow coefficients of the background can be obtained from the flow coefficients of the single leptons used to form the background dileptons as
\begin{equation}
v_{\rm n}^{\rm bkg}(\ensuremath{m_{\rm \ell \ell}}\xspace) = \frac{\langle v_{\rm n}^{(1)}\cos[{\rm n}(\varphi^{(1)}-\varphi)]+v_{\rm n}^{(2)}\cos[{\rm n}(\varphi^{(2)}-\varphi)]\rangle_{\ensuremath{m_{\rm \ell \ell}}\xspace}}{\langle 1+2\sum\limits_{{\rm m}=1}^{\infty}{v_{\rm m}^{(1)}v_{\rm m}^{(2)}\cos[{\rm m}(\varphi^{(1)}-\varphi^{(2)}]}\rangle_{\ensuremath{m_{\rm \ell \ell}}\xspace}},
\label{eq:vn_bkg}
\end{equation}
where $v_{\rm n}^{(1)}$ ($\varphi^{(1)}$) and $v_{\rm n}^{(2)}$ ($\varphi^{(2)}$) are the flow coefficients (azimuthal angles) of the two leptons, respectively, and $\varphi$ is the dilepton azimuthal angle.
The brackets $\langle \cdots \rangle_{\ensuremath{m_{\rm \ell \ell}}\xspace}$ denote an average over all dileptons belonging to the given $\ensuremath{m_{\rm \ell \ell}}\xspace$ interval.
Here, it is worth to note that the denominator in Eq.~\ref{eq:vn_bkg} represents the modification of the dilepton yields induced by the flow of single leptons.
Then, when background dileptons are built using the event mixing technique, the numerator in Eq.~\ref{eq:vn_bkg} is given by
\begin{equation}
\Big\langle \frac{\langle {{\bf u}_{\rm n}}^{(1)} {{\bf Q}_{\rm n}}^{(1),{\rm A*}} \rangle }{ R_{\rm n}^{(1)}} \cos [{\rm n}(\varphi^{(1)} - \varphi)] + \frac{\langle {{\bf u}_{\rm n}}^{(2)} {{\bf Q}_{\rm n}}^{(2),{\rm A*}} \rangle }{ R_{\rm n}^{(2)}} \cos [{\rm n}(\varphi^{(2)} - \varphi)]
\Big\rangle_{\ensuremath{m_{\rm \ell \ell}}\xspace} .
\label{eqn:numv2SPmethodME}
\end{equation}
Here, ${\bf u}_{\rm n}^{(1)}$ and ${\bf u}_{\rm n}^{(2)}$ are the unit vector of the two leptons, ${\bf Q}_{\rm n}^{(1),{\rm A}}$ and ${\bf Q}_{\rm n}^{(2),{\rm A}}$ are the event flow vectors, reconstructed in detector A, of the events containing the two leptons, and $R_{\rm n}^{(1)}$ and $R_{\rm n}^{(2)}$ their respective event flow factors (corresponding to the denominator of Eq.~\ref{eq:sp_method}).
Since the event flow vectors of the mixed events are not correlated, the mixed-event dilepton yield is not modified by the flow of the single leptons.
Examples of fits to the invariant mass distribution (top panels corresponding to (a1), (b1), (c1)) and to $\ensuremath{v_{\mathrm n}}\xspace(\ensuremath{m_{\rm \ell \ell}}\xspace)$ (bottom panels related to (a2) for $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm \mu \mu}}\xspace)$, (b2) for $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm ee}}\xspace)$, and (c2) for $\ensuremath{v_{\mathrm 3}}\xspace(\ensuremath{m_{\rm \mu \mu}}\xspace)$) are shown in Fig.~\ref{fig:figures_extraction} for the dimuon and dielectron analyses.
The background, which is mostly combinatorial, especially in central events, is well reproduced with the event mixing technique.
In the absence of correlated background, the background flow $\ensuremath{v_{\mathrm n}}\xspace^{\rm bkg}$ is directly given by the mixed-event flow.
At forward rapidity, the effect of the unknown flow contribution of the correlated background and residual mismatches between the same-event and mixed-event background flow, is considered as a systematic uncertainty and is discussed in Sec.~\ref{section:systematic_uncertainties}.
In the default approach, the flow of the correlated background is assumed to be negligible, and thus the denominator of Eq.~\ref{eq:vn_bkg} is given by the ratio $N^{\rm bkg}_{+-}/N^{\rm mix}_{+-}$ between the number of background unlike-sign dileptons $N^{\rm bkg}_{+-}$ and the number of unlike-sign dileptons from mixed events $N^{\rm mix}_{+-}$, which is obtained after a proper normalization involving like-sign dileptons as described in Ref.~\cite{Acharya:2018pjd}.
At midrapidity, due to the smaller signal-to-background ratio, the difference between mixed and same event background flow is taken into account by considering in the fit function an additional term which accounts for the flow of the correlated background and imperfections of the mixed event procedure. This term is parameterized using a second order polynomial and acts as a correction to the background flow obtained from the mixed event procedure.
\section{Systematic uncertainties} \label{section:systematic_uncertainties}
The systematic uncertainties related to the $v_{n}$ extraction procedure, the track and event selection criteria, residual detector effects, and non-flow contributions are evaluated as described below and summarized in Tab.\ref{tab:table-syst}. A quadratic sum of the systematic uncertainties from the independent sources is used as final systematic uncertainty on the measurements.
In the dimuon analysis, the signal fraction $\alpha(\ensuremath{m_{\rm \mu \mu}}\xspace)$ is estimated by fitting the invariant mass distribution with standard signal and background functions.
The systematic uncertainty on the determination of $\alpha(\ensuremath{m_{\rm \mu \mu}}\xspace)$ is estimated by varying the signal and background functions, as well as the mass fit range.
For the signal, in addition to a CB2, a pseudo-Gaussian with a mass-dependent width~\cite{ALICE-PUBLIC-2015-006} is also used. The tail parameters were fixed to the values obtained in Monte Carlo simulations or in other analyses with better signal significance~\cite{Acharya:2017hjh,Adam:2016rdg}.
For the background, the VWG function was changed to a fourth order Chebyshev polynomial. The invariant mass fit range is varied from the standard $2-4$~\GeVmass to $2.6-4.6$~\GeVmass in steps of 200~\MeVmass. The corresponding systematic uncertainty for each \pt bin, evaluated as the RMS of the results of the various tests, does not exceed 0.003 for \ensuremath{v_{\mathrm 2}}\xspace and 0.002 for \ensuremath{v_{\mathrm 3}}\xspace. In the dielectron analysis, the fit ranges of the residual background fit are varied. No significant changes of the extracted elliptic flow are observed and no uncertainty due to the \mbox{J/$\psi$}\xspace signal extraction is assigned.
The non-uniformity in the detector acceptance could lead to a residual effect in the calibration of the event flow vector ${\bf Q}_{\rm n}$.
The cross-term products of the event flow vector, $\langle Q_{\rm x,A}\times Q_{\rm y,B}\rangle$, are evaluated to verify that values are negligible compared to the linear products. In addition, possible impacts on the \ensuremath{v_{\mathrm n}}\xspace are checked by calculating the cross-term products between the components of the ${\bf Q}_{\rm n}$ vector and the unitary vector ${\bf u}_{\rm n}$ of the \mbox{J/$\psi$}\xspace candidates.
No clear \pt or centrality dependence is found for this contribution, and the corresponding systematic uncertainty is estimated to be less than 1\%.
Additional uncertainties related to the calculation of the reference flow vector are evaluated as the difference between the event flow factor $R_{\rm n}$ obtained using MB events or dimuon-triggered events.
For the dimuon analysis it amounts to 1\% for $R_{\rm 2}$ and up to 3\% for $R_{\rm 3}$.
The variation of the \mbox{J/$\psi$}\xspace reconstruction efficiency with the local occupancy of the detector could bias the measured \ensuremath{v_{\mathrm n}}\xspace. At forward rapidity, this effect is evaluated using azimuthally isotropic simulated $\mbox{J/$\psi$}\xspace \rightarrow \ensuremath{\rm{\mu}^{+}\rm{\mu}^{-}}\xspace$ decays embedded into real Pb--Pb events. A maximum effect of 0.002 for \ensuremath{v_{\mathrm 2}}\xspace and 0.001 for \ensuremath{v_{\mathrm 3}}\xspace is observed in non-central collisions with no clear \pt dependence. At midrapidity, the strongest dependence of reconstruction performance on the local detector occupancy is caused by the TPC particle identification (PID). A data driven study, using a clean electron sample from photon conversions, shows that the largest variation of the TPC electron PID response between the region along the event flow vector and the region orthogonal to it is approximately 2\% of the d$E$/d$x$ resolution. This leads to a decrease of the observed \ensuremath{v_{\mathrm 2}}\xspace by less than 1\% and is thus neglected.
\begin{table}[t]
\begin{center}
\caption{\label{tab:table-syst}Summary of absolute and relative (in \% of \ensuremath{v_{\mathrm n}}\xspace) systematic uncertainties of the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace coefficients, for the dimuon and dielectron analyses. The uncertainties vary within the indicated ranges depending on the \pt bin, or centrality interval.}
\begin{tabular}{cccccc}
\hline
\hline
& \multicolumn{4}{c}{$\mu^{+}\mu^{-}$} & $\rm{e}^{+}\rm{e}^{-}$ \\
Sources & \ensuremath{v_{\mathrm 2}}\xspace(\pt) & \ensuremath{v_{\mathrm 3}}\xspace(\pt) & \ensuremath{v_{\mathrm 2}}\xspace(Centrality) & \ensuremath{v_{\mathrm 3}}\xspace(Centrality) & \ensuremath{v_{\mathrm 2}}\xspace(\pt) \\
\hline
Extraction method & 0--0.003 &0--0.002 & 0.001--0.004 & 0.001--0.006 & negl \\
Centrality-$R_{\rm n}$ determination & 1\% & 3\% & 2\% & 3\% & negl \\
Non-flow estimation & $<$1\% & negl & $<$1\% & negl & \\
Reconstruction efficiency & 0.001--0.002 & 0--0.001 & 0--0.002 & 0--0.001 & negl \\
Correlated background shape & 0--0.009 & 0--0.015 & 0--0.010 & 0--0.011 & \\
TPC electron & & & & & 0.010 \\
identification selection & & & & & to 0.023 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
The presence of a correlated background and its unknown flow contribution can affect the \ensuremath{v_{\mathrm n}}\xspace extraction. The contribution of the correlated background to the flow of the background can be introduced in Eq.~\ref{eq:vn_bkg} by replacing the denominator $N^{\rm bkg}_{+-}/N^{\rm mix}_{+-}$ by $N^{\rm bkg}_{+-}/(N^{\rm mix}_{+-} + \beta (N^{\rm bkg}_{+-} - N^{\rm mix}_{+-}))$, where $\beta$ represents the relative strength of the correlated background flow with respect to the combinatorial background flow.
The systematic uncertainty is defined as the difference between the default fit, equivalent to $\beta = 0$, and the modified fit with $\beta$ left as a free parameter.
This uncertainty is, as expected, negligible for central collisions and low $p_{\rm T}$ but becomes significant for peripheral collisions and high $p_{\rm T}$.
The estimated systematic uncertainty for the $v_{2}$ and $v_{3}$ extraction reaches a maximum of about 0.01 for peripheral collisions and at high $p_{\rm T}$.
In the dielectron analysis, the signal-to-background ratio can vary significantly depending on the TPC electron identification selection and centrality, which may impact the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace fits. Thus, the \ensuremath{v_{\mathrm 2}}\xspace was extracted for a set of nine electron PID cuts where both the electron selection and the hadron rejection were varied such that the \mbox{J/$\psi$}\xspace efficiency is changed by approximately 50\%. The RMS of the \ensuremath{v_{\mathrm 2}}\xspace obtained from all of these selections is assigned as a systematic uncertainty, which ranges between 0.010 and 0.023 depending on the centrality and \pt interval, while the average value is taken as central value.
In addition, the fit range of the $\ensuremath{v_{\mathrm 2}}\xspace(\ensuremath{m_{\rm ee}}\xspace)$ is varied by either making it narrower or wider, but no significant systematic effects are observed.
\section{Results and discussions} \label{section:results}
The \mbox{J/$\psi$}\xspace elliptic flow coefficient measured by ALICE in Pb--Pb collisions at \fivenn at forward and central rapidity is shown in Fig.~\ref{fig:figure_v2vsPT_050} as a function of \pt, for the centrality intervals 0--10\%, 10--30\%, 30--50\% and 0--50\%.
Systematic uncertainties, obtained as described in the previous section, are shown as boxes around the data points, while the statistical uncertainties are shown as error bars.
Here, and in all figures as a function of $p_{\rm T}$, the \mbox{J/$\psi$}\xspace\ data points are located at the average \pt of the reconstructed \mbox{J/$\psi$}\xspace.
These results are compared with the midrapidity \ensuremath{v_{\mathrm 2}}\xspace measurements for charged pions by ALICE~\cite{Acharya:2018zuq} and prompt D mesons by ALICE~\cite{Acharya:2020zzz} and CMS~\cite{Sirunyan:2017plt}. At forward rapidity and for all centrality intervals, the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace values increase with \pt, possibly reaching a maximum at intermediate values of \pt, and decreasing or saturating towards high \pt.
\begin{figure}[h!]
\centering
\includegraphics[width = 0.95\textwidth]{figures/Fig2.pdf}
\caption{(Color online) Inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace as function of \pt in different centrality intervals (0--10\%, 10--30\%, 30--50\% and 0--50\%) in Pb--Pb collisions at \fivenn. Both midrapidity and forward rapidity \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace measurements are shown. The results are compared with the \ensuremath{v_{\mathrm 2}}\xspace coefficients at midrapidity for charged pions~\cite{Acharya:2018zuq} and prompt ${\rm D}^{0}$ mesons~\cite{Sirunyan:2017plt,Acharya:2020zzz}. The statistical and systematic uncertainties are shown as bars and boxes, respectively. The shaded cyan boxes represent the systematic uncertainties from the contribution of non-prompt ${\rm D}^{0}$ mesons.}
\label{fig:figure_v2vsPT_050}
\end{figure}
Also, the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace values increase when decreasing centrality from the 0--10\% to 10--30\%, then to 30--50\%. This behavior is qualitatively similar to the one for light hadrons and D mesons.
The \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace measurement at midrapidity is statistically compatible to the one at forward rapidity in both centrality intervals within uncertainties.
Considering all the midrapidity data points as statistically independent measurements, it was found that the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace is larger than zero with a significance of approximately 2.5 standard deviations in both centrality intervals.
It is worth to remark that the ALICE apparatus is undergoing an ambitious upgrade programme in preparation of Runs 3 and 4 of the LHC that will enable the separation of the prompt and non-prompt J/$\psi$ contributions to the measured flow coefficients, thus providing valuable information on both the charmonium and open beauty hadron production dynamics.
\begin{figure}[tb]
\centering
\includegraphics[width = 0.95\textwidth]{figures/Fig3.pdf}
\caption{(Color online) Inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace at forward rapidity as function of \pt in different centrality intervals (0--10\%, 10--30\%, 30--50\% and 0--50\%) in Pb--Pb collisions at \fivenn. The results are compared to the \ensuremath{v_{\mathrm 3}}\xspace coefficients at midrapidity for charged pions~\cite{Acharya:2018zuq} and prompt ${\rm D}^{0}$ mesons~\cite{Sirunyan:2017plt, Acharya:2020zzz}. The statistical and systematic uncertainties are shown as bars and boxes, respectively. The shaded bands represent the systematic uncertainties from the contribution of non-prompt ${\rm D}^{0}$ mesons.}
\label{fig:figure_v3vsPT_050}
\end{figure}
As also noted previously~\cite{Acharya:2018pjd}, a clear mass hierarchy of the \ensuremath{v_{\mathrm 2}}\xspace values is seen in the low-\pt region ($\pt<~6$~GeV/$c$\xspace) for the light hadrons and D mesons measured at midrapidity and inclusive \mbox{J/$\psi$}\xspace, with the \mbox{J/$\psi$}\xspace exhibiting the lowest elliptic flow. Here, it is important to note that in the considered $\eta$ range, the $\eta$ dependence of the \ensuremath{v_{\mathrm 2}}\xspace at a given \pt is expected to be negligible, as shown by the CMS measurement for charged particles~\cite{Sirunyan:2017igb}, albeit in a somewhat narrower $\eta$ range.
At high \pt ($\pt>8 $~GeV/$c$\xspace), the \ensuremath{v_{\mathrm 2}}\xspace coefficients from all species converge into a single curve suggesting that, in this kinematic range, the anisotropy for all particles arises dominantly from path-length dependent energy-loss effects~\cite{Abelev:2012di}.
However, in the case of the much heavier \mbox{J/$\psi$}\xspace, one may also consider that the hydrodynamic flow, which arises from a common velocity field, still contributes significantly even at high \pt, as can be expected from the particle mass dependence of the \pt range where the flow reaches its maximum.
In Fig.~\ref{fig:figure_v3vsPT_050}, the \pt-dependent inclusive \mbox{J/$\psi$}\xspace triangular flow coefficient measured at forward rapidity is shown in each of the considered centrality intervals.
For most of the centrality and \pt intervals, the measured inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace is positive and with no significant centrality dependence.
In the 0--50\% centrality range, the triangular flow coefficient is larger than zero (0.0250 $\pm$ 0.0045 (stat.) $\pm$ 0.0020 (syst.) in $2<\pt<5$ GeV/$c$) corresponding to a significance of 5.1$\sigma$, calculated adding quadratically the statistical and systematic uncertainties.
The positive \ensuremath{v_{\mathrm 3}}\xspace indicates that the initial state energy-density fluctuations, the dominant source of \ensuremath{v_{\mathrm 3}}\xspace, are reflected also in the anisotropic flow of charm quarks.
Also shown in Fig.~\ref{fig:figure_v3vsPT_050} are similar measurements for charged pions~\cite{Acharya:2018zuq} and D mesons~\cite{Sirunyan:2017plt,Acharya:2020zzz} obtained at midrapidity.
The mass hierarchy observed for \ensuremath{v_{\mathrm 2}}\xspace holds also in the case of \ensuremath{v_{\mathrm 3}}\xspace.
Together with the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace, these observations provide a strong support for the hypothesis of charm quark being, at least partially, kinetically equilibrated in the dense and deconfined QGP medium.
\begin{figure}[tb]
\begin{center}
\includegraphics[width = 0.95\textwidth]{figures/Fig4.pdf}
\end{center}
\caption{(Color online) Ratio of \ensuremath{v_{\mathrm 3}}\xspace to \ensuremath{v_{\mathrm 2}}\xspace of inclusive \mbox{J/$\psi$}\xspace (left panel) and $v_{3}/v_{2}^{3/2}$ (right panel) at forward rapidity as a function of \pt for the 0--50\% centrality interval in Pb--Pb collisions at \fivenn. The results are compared with the flow coefficients of charged pions~\cite{Acharya:2018zuq} and prompt ${\rm D}^{0}$ mesons at midrapidity \cite{Sirunyan:2017plt}. The statistical and systematic uncertainties are shown as bars and boxes. The shaded bands represent the systematic uncertainties from the contribution of non-prompt ${\rm D}^{0}$ mesons.}
\label{fig:figure_v2v3RatiovsPT}
\end{figure}
The ratio of the triangular to elliptic flow coefficients, \ensuremath{v_{\mathrm 3}}\xspace/\ensuremath{v_{\mathrm 2}}\xspace, as a function of \pt is shown in the left panel of Fig.~\ref{fig:figure_v2v3RatiovsPT} for the inclusive \mbox{J/$\psi$}\xspace at forward rapidity, D mesons and charged pions at midrapidity.
In this ratio, the statistical uncertainties are considered to be uncorrelated due to the weak correlation between the orientation of the ${\textbf Q}_{2}$ and ${\textbf Q}_{3}$ flow vectors~\cite{Aad:2014fla}, while the systematic uncertainties related to $\alpha(\ensuremath{m_{\rm \mu \mu}}\xspace)$ and to the reconstruction efficiency discussed in Sec.~\ref{section:systematic_uncertainties}, cancel in the ratio. The same hierarchy observed for the individual \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace measurements is also observed in the \ensuremath{v_{\mathrm 3}}\xspace/\ensuremath{v_{\mathrm 2}}\xspace ratio, which suggests that higher harmonics are damped faster for heavy quarks than for the light ones.
At RHIC~\cite{Adams:2003zg,Adare:2010ux} and LHC~\cite{ATLAS:2012at,Acharya:2018lmh}, it was observed that the flow coefficients of light particles from different harmonics follow a power-law scaling as $v_{\rm n}^{1/\rm{n}} \propto v_{\rm m}^{1/\rm{m}}$ up to about 6~GeV/$c$\xspace, for most centrality ranges, but the 0--5\%, independently of the harmonics n and m. The ratio $v_3/v_2^{3/2}$ in the right panel of Fig.~\ref{fig:figure_v2v3RatiovsPT} illustrates such a scaling.
Furthermore, the $v_3/v_2^{3/2}$ for pions, D and \mbox{J/$\psi$}\xspace mesons tend to converge, although the \mbox{J/$\psi$}\xspace values are systematically lower than the ones of pions.
\begin{figure}[tb]
\begin{center}
\includegraphics[width = 0.6\textwidth]{figures/Fig5.pdf}
\end{center}
\caption{(Color online) Inclusive J/$\psi$ $v_{2}$ as function of $p_{\rm T}$ at forward rapidity for semi-central (20--40\%) Pb--Pb collisions at \fivenn. Calculations from a transport model \cite{Du:2015wha,Du:2017qkv} are also shown.}
\label{fig:figure_v2vsPT_models}
\end{figure}
In Fig.~\ref{fig:figure_v2vsPT_models}, the inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace as a function of \pt in the 20--40\% centrality interval is compared with the microscopic transport calculations by Du et al.~\cite{Du:2015wha,Du:2017qkv}.
In this model, the \mbox{J/$\psi$}\xspace are created both from the primordial hard partonic interactions but also from the recombination of thermalized charm quarks in the medium, which accounts for roughly 50\% of all \mbox{J/$\psi$}\xspace at low \pt.
Non-prompt \mbox{J/$\psi$}\xspace mesons, created in the weak decays of beauty hadrons, are also included in the model.
The amplitude of the inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace in the calculations is in good agreement with the experimental measurements for \pt $<$ 4~GeV/$c$\xspace.
However, the overall trend of the model calculation does not describe the data well, especially in the intermediate \pt range, $4<\pt<10$~GeV/$c$\xspace, where the \mbox{J/$\psi$}\xspace flow is largely underestimated.
The primordial \mbox{J/$\psi$}\xspace component, which is sensitive mainly to path length dependent effects, like survival probability, exhibits a monotonically increasing trend from low towards high \pt, with this mechanism becoming the dominant source of the anisotropic flow for \pt larger than 8~GeV/$c$\xspace. Path length dependent energy loss, widely seen as a major source of anisotropy at large \pt, is not implemented for \mbox{J/$\psi$}\xspace mesons in this calculation. It is worth noting that this model provides a qualitative good description of the centrality and transverse momentum of the \mbox{J/$\psi$}\xspace nuclear modification factor~\cite{Adam:2016rdg,Acharya:2019iur}.
Figure~\ref{fig:figure_v2v3vsCentrality} shows the centrality dependence of the inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace (top panels) and \ensuremath{v_{\mathrm 3}}\xspace (bottom panels) for a low-\pt interval ($0 < \pt < 5$~GeV/$c$\xspace) on the left, and a high-\pt one ($5 < \pt < 20$~GeV/$c$\xspace) on the right.
Here, due to the large integrated \pt range, the \ensuremath{v_{\mathrm n}}\xspace coefficients are corrected for the \mbox{J/$\psi$}\xspace acceptance and efficiency $A\times \varepsilon$.
Each dimuon pair is weighted using the inverse of the \pt and $y$ dependent $A\times \varepsilon$ factor before filling the invariant mass and $v_{\rm n}(\ensuremath{m_{\rm \mu \mu}}\xspace)$ distributions.
The \pt and $y$ dependent $A\times \varepsilon$ map was obtained from the embedded simulations discussed in Section~\ref{section:systematic_uncertainties}.
Any possible systematic uncertainty related to the $A\times \varepsilon$ corrections is already included in the systematic uncertainty due to the dependence of the reconstruction efficiency on the local detector occupancy.
The \mbox{J/$\psi$}\xspace results are compared with the flow coefficients of charged pions for a \pt value similar to the corrected \mbox{J/$\psi$}\xspace $\langle \pt \rangle$, published by ALICE in Ref.~\cite{Acharya:2018zuq}. In addition, the ratio $v_{\rm 2}^{{\rm \pi}}/v_{\rm 2}^{{\rm J}/\psi}$ is computed and shown in the bottom sub-panels.
Both at low \pt ($1.75<\pt<2$~GeV/$c$\xspace) and high \pt ($6<\pt<7$~GeV/$c$\xspace), the \ensuremath{v_{\mathrm 2}}\xspace of $\pi^{\pm}$ increases from central to semi-central collisions, reaching a maximum at 40--50\% centrality, and then decreases towards peripheral collisions.
For the \mbox{J/$\psi$}\xspace at low \pt, while the centrality trend is qualitatively similar, the maximum (or even saturation) of \ensuremath{v_{\mathrm 2}}\xspace seems to be reached for more central collisions than for the pions. This is more clearly emphasized by the increasing trend of the ratio $v_{\rm 2}^{{\rm \pi}}/v_{\rm 2}^{{\rm J}/\psi}$, from central to peripheral collisions, which deviates from unity by a significance of 8.5$\sigma$.
In the framework of transport models, this could be understood by the increasing fraction of regenerated J/$\psi$ at low $p_{\rm T}$ when moving from peripheral to central collisions.
Alternatively, and independently of the regeneration scenario, the increase of the $v_{\rm 2}^{{\rm \pi}}/v_{\rm 2}^{{\rm J}/\psi}$ from central to peripheral collisions, could also be understood in terms of partial or later thermalization of the charm quarks compared to light quarks.
The decrease in energy density and lifetime of the system is counterbalanced by the increase of the initial spatial anisotropy towards peripheral collisions.
The $v_{2}$ of the \mbox{J/$\psi$}\xspace will therefore reach its maximum at more central collisions compared to light particles because charm quarks require larger energy densities to develop flow~\cite{Scardina:2017ipo,Rapp:2018qla,Song:2015sfa,Cao:2011et}.
At high \pt, \mbox{J/$\psi$}\xspace mesons and charged pions seem to exhibit the same centrality dependence, although the $v_{\rm 2}$ coefficients are systematically lower for the \mbox{J/$\psi$}\xspace mesons than for the pions.
Such a similar centrality dependence could indicate a similar mechanism at the origin of the flow for both \mbox{J/$\psi$}\xspace mesons and pions at high \pt.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width = 0.95\textwidth]{figures/Fig6.pdf}
\end{center}
\caption{The inclusive J/$\psi$ $v_{2}$ and $v_{3}$ as function of the centrality of the collision, at forward rapidity, for the low-$p_{\rm T}$ range $0<p_{\rm T}<5$ GeV/$c$ (left panel) and high-$p_{\rm T}$ range $5<p_{\rm T}<20$ GeV/$c$ (right panel). The results are compared to the $v_{\rm n}$ coefficients of midrapidity $\pi^{\pm}$~\cite{Acharya:2018zuq} at low and high $p_{\rm T}$ corresponding to $1.75<p_{\rm T}<2$ GeV/$c$ and $6<p_{\rm T}<7$ GeV/$c$, respectively. The ratio of midrapidity $\pi^{\pm}$ $v_{2}$ to inclusive J/$\psi$ $v_{2}$ is also shown. }
\label{fig:figure_v2v3vsCentrality}
\end{figure}
The centrality dependence of the \ensuremath{v_{\mathrm 3}}\xspace coefficient at low \pt is less pronounced than that of the \ensuremath{v_{\mathrm 2}}\xspace for both pions and \mbox{J/$\psi$}\xspace, as expected since initial state fluctuations only weakly depend on centrality. Also, the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace is smaller relative to the one of charged pions, in both the \pt intervals considered.
\begin{figure}[tb]
\begin{center}
\includegraphics[width = 0.95\textwidth]{figures/Fig7.pdf}
\end{center}
\caption{(Color online) Elliptic (left panels) and triangular (right panels) flow of inclusive \mbox{J/$\psi$}\xspace , D mesons~\cite{Sirunyan:2017plt} and charged pions as a function of \pt for the centrality intervals 0--10\% (top), 10--30\% (middle) and 30--50\% (bottom). The continuous curves show the calculated D-meson flow based on different values of the \pt fraction carried by the light quark (see text). The red dashed curves show the fits to the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm n}}\xspace using ad-hoc functions (see text).}
\label{fig:figure_scalingCharmFlow}
\end{figure}
The flow of light and strange particles was shown to approximately scale with the number of constituent quarks (NCQ scaling) at both RHIC and LHC energies \cite{Zheng:2016iia,Singha:2016aim}. This was typically interpreted to arise naturally in hadronization scenarios based on quark coalescence in which the flow of bound mesons and baryons depends solely on the collective flow of light and strange quarks (assumed to be identical) and the number of valence quarks \cite{Molnar:2003ff,Adams:2004bi}. In the case of charmed hadrons, the NCQ scaling assuming a flavor independent flow would obviously not work due to the large observed differences between the flow of light-flavor particles, D and \mbox{J/$\psi$}\xspace mesons. However, this scaling can be extended by assuming that the much heavier charm quark has a different flow magnitude \cite{Lin:2003jy} and that it can be derived from the flow of the \mbox{J/$\psi$}\xspace via the usual NCQ formula, $v_{\rm n}^{\rm J/\psi}(\pt^{\rm J/\psi}) = 2 \cdot v_{\rm n}^{\rm c}(\pt^{\rm J/\psi}/2)$. Then it is straightforward to show that the flow of the D meson can be constructed as the sum of the flow coefficients for light and charm quarks as
\begin{equation}
v_{\rm n}^{\rm D}(p_{\rm T}^{\rm D}) = v_{\rm n}^{\rm q}(p_{\rm T}^{\rm q}) + v_{\rm n}^{\rm c}(p_{\rm T}^{\rm c}),
\label{eq:quarkScaling}
\end{equation}
where $p_{\rm T}^{\rm q}$ and $p_{\rm T}^{\rm c}$ are the \pt of the light and charm quarks, respectively, corresponding to the D-meson \pt, $p_{\rm T}^{\rm D}$. The light quark flow is obtained by interpolating the measured charged pions flow using $v_{\rm n}^{\rm \pi}(\pt^{\rm \pi}) = 2 \cdot v_{\rm n}^{\rm q}(\pt^{\rm \pi}/2)$.
Figure~\ref{fig:figure_scalingCharmFlow} shows a comparison of the D-meson \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace as a function of \pt, derived assuming the above described procedure, to the D-meson \ensuremath{v_{\mathrm n}}\xspace measured by the CMS Collaboration~\cite{Sirunyan:2017plt}.
The red dashed curves show fits to the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm n}}\xspace employing an ad-hoc function, a third order polynomial at low \pt and a linear function at high \pt, used to extract the flow of charm quarks needed to obtain the scaled D-meson flow according to Eq.~\ref{eq:quarkScaling}.
The scaled D-meson flow is found to be very sensitive to the fraction of \pt carried by each of the constituent quarks. In coalescence-like models, constituent quarks must have equal velocities which leads to a sharing of the D-meson \pt proportional to the effective quark masses. This implies that by far the largest fraction of \pt should be carried by the charm quark. Based on the simplistic and naive approach described here, a \pt sharing between light and charm quarks~\cite{Jia:2006vj,Lin:2003jy} where the ratio $p_{\rm T}^{\rm q}/p_{\rm T}^{\rm D} = 0.2$ (black curve), is clearly disfavored by the data. Surprisingly, it was found that a good description of the D-meson flow measurement, as illustrated by the blue curves in Fig.~\ref{fig:figure_scalingCharmFlow}, is obtained when the light quark carries a relatively large fraction of the D-meson \pt (dark blue and green curves).
The best agreement with the D-meson data of the CMS Collaboration~\cite{Sirunyan:2017plt} is obtained when the light-quark \pt fraction has a value of $p_{\rm T}^{\rm q}/p_{\rm T}^{\rm D}$ = 0.4 (dark blue curve), but a rather good description of the data is observed also when assuming that the light and charm quarks share equally the D-meson \pt (green curve).
Within uncertainties, the scaling seems to work well for both \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace over the entire covered \pt range and in all centrality intervals.
\section{Conclusion}
In summary, the inclusive \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace at forward and midrapidity and the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace at forward rapidity were measured in Pb--Pb collisions at \fivenn using the scalar product method.
In non-central collisions, the \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 2}}\xspace values are found to be positive up to the last interval corresponding to $12<\pt<20$~GeV/$c$\xspace and reach a maximum of approximately 0.1 around a \pt of 5~GeV/$c$\xspace.
The \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace values at forward rapidity reach 0.04 around a \pt of 4~GeV/$c$\xspace and are positive in the 0--50\% centrality interval for $2<\pt<5$~GeV/$c$\xspace with a significance of 5.1$\sigma$.
The mass hierarchy observed for \ensuremath{v_{\mathrm 2}}\xspace, $v_{\rm 2,\pi}>v_{\rm 2,D}>v_{\rm 2,J/\psi}$, seems to also hold in the case of \ensuremath{v_{\mathrm 3}}\xspace and will be the subject of more detailed studies with the Run 3 and Run 4 data.
At high \pt, the \ensuremath{v_{\mathrm 2}}\xspace for all particles converge to similar values, suggesting that path-length dependent effects become dominant there.
The measured \mbox{J/$\psi$}\xspace \ensuremath{v_{\mathrm 3}}\xspace/\ensuremath{v_{\mathrm 2}}\xspace ratios exhibits the same hierarchy indicating that higher harmonics are damped faster for charmonia compared to lighter particles.
The \pt-integrated $v_2$ coefficient in a low and a high-\pt region is in both cases dependent on centrality and reaches a maximum value of about 0.1, while the $v_3$ has no clear centrality dependence.
Both J/$\psi$ \pt-integrated $v_2$ and $v_3$ coefficients, either at low $p_{\rm T}$ or at high $p_{\rm T}$ are found to be lower than the ones of charged pions at a \pt similar to the \mbox{J/$\psi$}\xspace average \pt.
At low \pt, the ratio of the charged pions $v_2$ to those of \pt-integrated J/$\psi$ increase from central to peripheral collisions, compatible with a scenario in which charm quarks thermalize later than the light ones. At high \pt, this ratio is compatible with unity without any statistically significant centrality dependence.
Using an extension of the well known number of constituent quark scaling, the measured charged pion and J/$\psi$ $v_{\rm n}$ can be used as proxies in order to derive the D-meson \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace as a combination of the flow of light and charm quarks. Within this procedure, it is surprising to observe that the measured D-meson \ensuremath{v_{\mathrm 2}}\xspace and \ensuremath{v_{\mathrm 3}}\xspace can be described if one considers that the light and charm quarks share similar fractions of the D-meson $p_{\rm T}$, which is counterintuitive in a coalescence approach.
The fact that such a simple scaling works suggests that the flow of charmonia and open charm mesons can be effectively explained assuming a common underlying charm quark flow in addition to the flow of light quarks.
\newenvironment{acknowledgement}{\relax}{\relax}
\begin{acknowledgement}
\section*{Acknowledgements}
\input{fa_2020-05-17.tex}
\end{acknowledgement}
\bibliographystyle{utphys}
| 421bd6632c21639787020b6732f4824c6410d2e1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Success of a secure quantum network depends on quantum correlations distributed and shared among different parties over many sites \cite{Kim}. Different kind of quantum correlations, for instance, multipartite entanglement\cite{Hor,Guh} and multipartite nonlocality \cite{Bru} have been extensively used as a resource to perform many task in such networks. A key property of these quantum correlations used to secure those quantum networks is that they have limited shareability properties and sometimes can even be monogamous. For example, when a quantum system $A$ is entangled with another system $B$ then this entanglement puts a constraint on the amount of entanglement that can exist between one of those parties ($B$, say) and a third party, $C$. This limited shareability phenomenon is termed as monogamy. This is one of the fundamental differences between quantum entanglement and classical correlations, where all classical probability distributions can be shared \cite{Ton}. Monogamy of entanglement was first quantified by Coffman, Kundu, and Wootters (CKW) in \cite{Cof}, where it was shown that the sum of the individual pairwise entanglement between A and B and C cannot exceed the entanglement between A and the remaining parties together. Since then many research work have been done on such monogamy relations of quantum entanglement \cite{Osb,Yco,Bai,Zhu,Reg,Teh}. This characteristic of quantum entanglement has found potential applications in various quantum information processing tasks such as quantum key distribution \cite{Pal,Gis}, classification of quantum states \cite{Dur,Gio,Pra}, study of black-hole physics \cite{Llo}, and frustrated spin systems \cite{Rao}, etc. Similar to monogamy of entanglement, if any two quantum systems $A$ and $B$ are correlated in such a way that they violate Bell-CHSH inequality \cite{Cla} then neither of $A$ nor $B$ can be Bell-CHSH nonlocal with the other system $C$. In the last few years, several fundamental results on shareability of nonolocal correlations have been proven that constrain the distribution of nonlocal correlations in terms of violation of some Bell-type inequalities among the subsystems of a multipartite system\cite{Ton,Sca,Scn,Tne,Bar,Mas,See,Paw,Kur,Qin,Che,Tra,Ram} and they play a key role in the applications of quantum nonlocal correlations to cryptography\cite{Pal,Gis}. Monogamy relations have also been studied for quantum discord \cite{Ste}, indistinguishability \cite{Kar}, coherence \cite{Rad} , measurement induced nonlocality \cite{Chn} and other nonclassical correlations\cite{Chn}.\\
Despite the importance of shareability in quantum information, the knowledge of shareability for quantum steering is so far rather limited \cite{Rei,Xia,Chg}. The objective of this paper is to understand more about the shareability associated with the quantum steering. The notion of steering was introduced by Schr$\mathrm{o}$dinger in 1935 \cite {Sch} and the effect was recently formalised from foundational as well as quantum information perspective \cite{Wis,Jon}. Considering two distant observers Alice and Bob sharing an entangled state, steering captures the fact that Alice, by performing a local measurement on her subsystem, can remotely steer Bob's state. This is not possible if the shared state is only classically correlated. This kind of quantum correlation is known as steering \cite{Uol}.
It can be understood as a form of quantum nonlocality intermediate between entanglement and Bell nonlocality \cite{Qui}. Quantum steering is certified by the violation of steering inequalities. A number of steering inequalities have been designed to observe steering \cite{Red,Cav,Wal,Sco,Jlc,Kog,Cva,Zuk,Jev,Cos}.
Violation of such steering inequalities certify the presence of entanglement in one-sided device-independent way. Steerable states were shown to be beneficial for tasks involving randomness generation \cite{Law}, subchannel discrimination
\cite{Pia}, quantum information processing \cite{Bra}, and one-sided device-independent processing in quantum key distribution \cite{Ban}. However, comparatively little is known about the shareability of this type of nonlocality. By deriving shareability relations, one can understand how this special type of nonlocal correlation (steering) can be distributed over different subsystems. In this paper, by using the three settings CJWR linear steering inequality \cite{Cav,Cos}, we will derive different kind of trade-off relations that quantify the amount of bipartite steering that can be shared among the three qubit systems. In turn, these trade-off relations help us to prove that at most two of three reduced states of an arbitrary three qubit state can violate the three settings CJWR linear steering inequality contrary to two settings CJWR linear steering inequality or Bell-CHSH inequality, where at most one of the reduced states can violate those inequalities. Consequently, in general, steering correlations turn out to be non-monogamous.\\
Over the past few years it has become clear that correlation statistics of two-body subsystems can be very fruitful in inferring
the multipartite properties of a composite quantum system \cite{Bne,Wni,Tur,Tot,Kor,Kna,Las}. In this context, we have also studied how the reduced bipartite steering of a three qubit state depends on the bipartite and genuine tripartite entanglement of the three-qubit states. Interestingly, criteria for detecting different kind of entanglement of pure three qubit state are obtained based on these shareability relations. We illustrate the relevance of our results with different examples.
\section{PRELIMINARIES}\label{sec2}
In this section, we briefly discuss the concept of steering and the three settings CJWR linear steering inequality that we use in this work.
\subsection{Steering}
Steering is usually formulated by considering a quantum information task \cite{Wis,Jon}. Suppose two spatially separated observers, say Alice and Bob share a bipartite state $\rho_{AB}$ and they can perform measurements in the sets $M_{A}$ and $M_{B}$,
respectively. In a steering test, Bob, who does trust his own but not Alice's apparatus, wants to verify whether the shared state between them is entangled. He will be convinced that the shared state $\rho_{AB}$ is entangled only if his system is genuinely influenced by Alice's measurement, instead of some preexisting local hidden states (LHS) which Alice may have access to. To make sure that Bob must exclude the LHS model
\begin{equation}\label{mon1}
P(a,b | A, B, \rho_{AB})=\sum_\lambda p_\lambda P(a|A, \lambda)P_{Q}(b|B, \rho_{\lambda}),
\end{equation}
in which $P(a,b|A,B, \rho_{AB}) = Tr(A_{a} \otimes B_{b}\, \rho_{AB})$ is the probability
of getting outcomes $a$ and $b$ when measurements $A$ and $B $ are performed on $\rho_{AB}$ by Alice and Bob respectively, $A_{a}$ and $B_{b}$ are their corresponding measurement
operators;
$ \lambda $ is the hidden variable, $\rho_{\lambda}$ is the state that Alice sends with probability $p_{\lambda}$($ \sum_{\lambda}p_{\lambda} =1 $); $P(a|A, \lambda)$ is the conditioned probability of Alice obtaining outcome $a$ under $\lambda$ , $P_{Q}(b|B, \rho_{\lambda})$ denotes the quantum probability of outcome $b$ given by measuring $B$ on the local hidden state $\rho_{\lambda}$.
Now, if Bob determines that such correlation $P(a,b | A, B, \rho_{AB})$ cannot be explained by any LHS models, then he will be convinced that Alice can steer his state, and thus the corresponding bipartite state is entangled. In short, the bipartite state $\rho_{AB}$ is unsteerable by Alice to Bob if and only if the joint probability distributions satisfy the Eq.(\ref{mon1}) for all measurements $A$ and $B$. The assumption of such LHS model leads to certain steering inequalities, violation of which indicates the occurrence of steering.\\
The simplest way of constructing steering inequality is to find constraint for the correlations between Alice's and Bob's measurement statistics. In this work, we are interested in using such type of linear steering inequality formulated by
Cavalcanti, Jones, Wiseman, and Reid(CJWR) \cite{Cav}. They proposed the following series of steering inequalities to check whether a bipartite state is steerable from Alice to Bob when both the parties are allowed to perform $n$ dichotomic measurements on their respective subsystems:
\begin{equation}\label{mon2}
F_{n}(\rho_{AB},\mu) = \frac{1}{\sqrt{n}}|\sum_{k=1}^{n} \langle A_{k} \otimes B_{k} \rangle | \leq 1
\end{equation}
where $A_{k} = \hat{a}_{k}\cdot \overrightarrow{\sigma}$, $B_{k} = \hat{b}_{k} \cdot \overrightarrow{\sigma}$, $\overrightarrow{\sigma} = (\sigma_{1},\sigma_{2},\sigma_{3})$ is a vector composed of the Pauli matrices, $\hat{a}_{k} \in \mathbb{R}^{3}$ are unit vectors, $\hat{b}_{k} \in \mathbb{R}^{3}$ are orthonormal vectors, $\mu =\{\hat{a}_{1},\hat{a}_{2},....\hat{a}_{n}, \hat{b}_{1},\hat{b}_{2},...,\hat{b}_{n} \}$ is the set of measurement directions, $\langle A_{k} \otimes B_{k} \rangle = Tr(\rho_{AB} (A_{k} \otimes B_{k}))$ and $\rho_{AB} \in \mathbb{H_{A}} \otimes \mathbb{H_{B}}$ is any bipartite quantum state.\\
Here our attention is confined to the qubit case. In Hilbert-Schmidt representation any two qubit state can be expressed as,
\begin{equation}\label{mon3}
\rho_{AB} = \frac{1}{4}[\mathbf{ I} \otimes \mathbf{ I} + \vec{a } \cdot \vec{\sigma} \otimes \mathbf{ I} + \mathbb{I} \otimes \vec{b} \cdot \vec{\sigma} + \sum_{i,j} t^{AB}_{ij} \sigma_{i} \otimes \sigma_{j}]
\end{equation}
$\vec{a}$, $\vec{b}$ being the local bloch vectors and $T_{AB} = [t^{AB}_{ij}]$ is the correlation matrix. The components $t^{AB}_{ij}$ are given by $t^{AB}_{ij} = Tr[\rho_{AB} {\sigma}_{i} \otimes {\sigma}_{j}]$ and $\vec{a } ^{2}+ \vec{b }^{2} + \sum_{i,j} {(t^{AB}_{ij})} ^{2}\leq 3$.
In \cite{Luo}, Luo showed that any two-qubit state
can be reduced, by local unitary equivalence, to
\begin{equation}\label{mon4}
\rho'_{AB} = \frac{1}{4}[\mathbf{ I} \otimes \mathbf{ I} + \vec{a' } \cdot \vec{\sigma} \otimes \mathbf{ I} + \mathbb{I} \otimes \vec{b'} \cdot \vec{\sigma} + \sum_{i} u'_{i} \, {\sigma}_{i} \otimes {\sigma}_{i}]
\end{equation}
where the correlation matrix of $\rho'_{AB} $ is $T'_{AB} = diag(u'_{1},u'_{2},u'_{3})$.
In \cite{Cos}, for any two qubit state $\rho'_{AB}$, the authors derived an analytical expression for the maximum value of the two settings and three settings CJWR linear steering inequality in terms of diagonal elements of the correlation matrix $T'_{AB} = diag(u'_{1},u'_{2},u'_{3})$. \\
Specifically, $\max_{\mu} F_{2}(\rho'_{AB},\mu)$ and $\max_{\mu} F_{3}(\rho'_{AB},\mu)$ have been evaluated to be the following :
\begin{equation}\label{mon5}
\max_{\mu} F_{2}(\rho'_{AB},\mu) = \sqrt{u'^{2}_{1} + u'^{2}_{2} },
\end{equation}
and
\begin{equation}\label{mon6}
\max_{\mu} F_{3}(\rho'_{AB},\mu) = \sqrt{Tr[T'^{T}_{AB} T'_{AB}] },
\end{equation}
where ${u'}_{1}^{2}$ and ${u'}_{2}^{2}$ are two largest diagonal elements of $T'^{2}_{AB}$.
Here we do consider only the three settings linear steering inequality as under two measurement settings the notion of steering and Bell-CHSH nonlocality are indistinguishable.
Since the states given in Eq.(\ref{mon3}) and Eq.(\ref{mon4}) are local unitary equivalent, we must have,
\begin{eqnarray*}
\max_{\mu} F_{3}(\rho'_{AB},\mu) = \sqrt{Tr[T'^{T}_{AB} T'_{AB}] } &&=\sqrt{Tr[T^{T}_{AB} T_{AB}] } \\
&&=\max_{\mu} F_{3}(\rho_{AB},\mu).
\end{eqnarray*}
Consequently, the linear inequality(\ref{mon2}) (for three measurement settings) implies that any state $\rho_{AB}$ is $F_{3}$ steerable if and only if
\begin{equation}\label{mon8}
S_{AB} = Tr[T^{T}_{AB} T_{AB}] >1.
\end{equation}
Note that this condition is just a sufficient criterion to check steerability. There exist steerable states which satisfy $S_{AB} \leq 1$.
\section{Shareability and monogamy of Steering Correlations}\label{monsec1}
Consider a scenario in which Alice, Bob, and Charlie share a three qubit state $\rho_{ABC}.$ Let $\rho_{AB}$, $\rho_{AC}$, $\rho_{BC}$ denote the three bipartite reduced states of $\rho_{ABC}$.
In general, for tripartite states, monogamy relations have the following form:
\begin{equation}\label{mon9}
Q(\rho_{AB}) + Q(\rho_{AC}) \leq Q(\rho_{A|BC})
\end{equation}
or
\begin{equation}\label{mon10}
Q(\rho_{AB}) + Q(\rho_{AC}) \leq K
\end{equation}
for some bipartite quantum measure $Q$ and positive real number $K$. Here $Q(\rho_{A|BC})$ represents the correlation between subsystems $A$ and $BC$.
Entanglement, Bell-CHSH nonlocality, and steering (via two settings linear steering inequality) are examples of such correlation measures satisfying this monogamy relation(Eq.(\ref{mon10})).
Particularly, for Bell-CHSH inequality and $F_{2}$ inequality, monogamy relation(\ref{mon10}) takes the form \cite{Ton,Tne,See,Che}
\begin{equation}\label{mon11}
Q(\rho_{AB}) + Q(\rho_{AC}) \leq 2.
\end{equation}
Thus, at most one bipartite reduced state with respect to a certain observer (say, $A$) can violate the linear steering $F_{2}$ inequality. This feature is known as ``\textit{monogamy of steering correlations}''.\\
It is a known fact that entanglement is a property of a quantum state, now correlations generated due to measurements performed on any entangled quantum state are not solely determined by the state of the system under consideration. It is also dependent on the specific setup used to determine the correlations. Consequently, in general, steerability of a quantum state varies from one measurement scenario to another. In this context, an obvious question arises: \textit{can addition of one more observable per party change the monogamous nature of steering?}
Affirmative answer of this query is given by the following theorem.
\begin{theorem}\label{mont1}
For any three qubit state $\rho_{ABC} \in \mathbb{H}^{A} \otimes \mathbb{H}^{B} \otimes \mathbb{H}^{C}$, at most two reduced states can violate the three settings CJWR linear steering inequality, i.e steering can be non-monogamous when each party measures three dichotomic observables.
\end{theorem}
\begin{proof}
Any three qubit state $\rho_{ABC}$ can be represented as
\begin{align}\label{mon12}
&\rho_{ABC} = \frac{1}{8}[\mathbb{I} \otimes \mathbb{I} \otimes \mathbb{I} + \vec{a} \cdot \vec{\sigma} \otimes \mathbb{I} \otimes \mathbb{I} + \mathbb{I} \otimes \vec{b} \cdot \vec{\sigma} \otimes \mathbb{I}\nonumber\\
& +\mathbb{I} \otimes \mathbb{I} \otimes \vec{c} \cdot \vec{\sigma} + \sum_{ij} t^{AB}_{ij} {\sigma}_{i} \otimes {\sigma}_{j} \otimes \mathbb{I} + \sum_{ik} t^{AC}_{ik} {\sigma}_{i} \otimes \mathbb{I} \otimes {\sigma}_{k} \nonumber \\
& + \sum_{jk} t^{BC}_{jk} \mathbb{I} \otimes {\sigma}_{j} \otimes {\sigma}_{k} + \sum_{ijk} t^{ABC}_{ijk} {\sigma}_{i} \otimes {\sigma}_{j} \otimes {\sigma}_{k} ]
\end{align}
In the following we denote $\rho_{i}$ as the reduced density matrices for the subsystem $i = A,B,C$. One computes from Eq.(\ref{mon12}), that
\begin{equation}\label{mon13}
tr(\rho_{A}^{2}) = \frac{1+ \vec{a}^{2}}{2}, Tr(\rho_{BC}^{2}) = \frac{1}{4}(1+ \vec{b}^{2} + \vec{c}^{2} + S_{BC}).
\end{equation}
Similarly,
\begin{eqnarray}\label{mon14}
&&tr(\rho_{B}^{2}) = \frac{1+ \vec{b}^{2}}{2}, Tr(\rho_{AC}^{2}) = \frac{1}{4}(1+ \vec{a}^{2} + \vec{c}^{2} +S_{AC}),\nonumber\\
&&tr(\rho_{C}^{2}) = \frac{1+ \vec{c}^{2}}{2}, Tr(\rho_{AB}^{2}) = \frac{1}{4}(1+ \vec{a}^{2} + \vec{b}^{2} + S_{AB}).
\end{eqnarray}
First consider $\rho_{ABC}$ is a pure state. Then from Schimdt decomposition, we have $Tr(\rho_{i}^{2}) = Tr(\rho_{jk}^{2})$ for $i \neq j \neq k$, $i, j, k = A,B,C$.
Using these relations and Eqs.(\ref{mon13},\ref{mon14}), it is straightforward to calculate $S_{ij}$ of each pair of qubits, yielding:
\begin{equation}\label{mon15}
S_{AB} = 1 + 2 \vec{c}^{2} - \vec{a}^{2} - \vec{b}^{2},
\end{equation}
\begin{equation}\label{mon16}
S_{AC} = 1 + 2 \vec{b}^{2} - \vec{a}^{2} - \vec{c}^{2},
\end{equation}
and
\begin{equation}\label{mon17}
S_{BC} = 1 + 2 \vec{a}^{2} - \vec{b}^{2} - \vec{c}^{2}.
\end{equation}
Adding these three relations and simplifying it, one obtains the following relation:
\begin{equation}\label{mon18}
S_{AB} + S_{AC} + S_{BC} = 3.
\end{equation}
This relation is derived by the similar method used in \cite{Ken} for developing Bell monogamy relations.\\
Now, taking mixed state $\rho_{ABC}$ as $\sum_{n} p_{n}|\psi_{n}\rangle \langle \psi_{n}|$, one has
$S_{AB} \leq \sum_{n} p_{n} S^{n}_{AB}$, and similar relations for $S_{AC}$, $S_{BC}$.
Adding these relations and using Eq.(\ref{mon18}), we obtain
\begin{equation}\label{mon19}
S_{AB} + S_{AC} + S_{BC} \leq 3
\end{equation}
This is a trade off relation among two qubits of any three qubit state $\rho_{ABC}$. Now $S_{AB} >1$ is sufficient for Alice and bob to witness violation of $F_{3}$ inequality. Hence, inequality Eq.(\ref{mon19}) imposes constraint on quantum steering: it is impossible that all pair of qubits violate $F_{3}$ inequality. \\
But the trade off relation (\ref{mon19}) is unable to assure us about the number of two qubit reduced states that can violate $F_{3}$ inequality.
To complete the proof, we still have to find two reduced states of $\rho_{ABC}$ which violate $F_{3}$ inequality. \\
Using Eqs.(\ref{mon15},\ref{mon16},\ref{mon17}), one can easily find that the reduced states $\rho_{AB}$ and $\rho_{AC}$ of the pure three qubit state $\rho_{ABC}$ will violate $F_{3}$ inequality iff the following inequality is satisfied :
\begin{equation}\label{mon20}
\vec{c}^{2} > \frac{\vec{a}^{2} + \vec{b}^{2}}{2},
\vec{b}^{2} > \frac{\vec{a}^{2} + \vec{c}^{2}}{2},
\end{equation}
One can similarly obtain condition of violation for other pairs of reduced states.
Now, consider the fully entangled three qubit state,
\begin{equation}\label{mon21}
|\psi_{ABC}\rangle = \frac{1}{2}(|100\rangle +|010\rangle + \sqrt{2} |001\rangle).
\end{equation}
By using the above conditions, one can find that bipartite correlations between party A and C of subsystem AC and between B and C of subsystem BC violate the $F_{3}$ inequality: $S_{BC} = S_{AC} = 1+ \frac{1}{4}>1$. This shows that some of the steering correlations between party A and C can thus be shared with party B and C. Thus, under some conditions (for example, Eq. (\ref{mon20}) and its permutations), steering is non-monogamous with respect to $F_{3}$ inequality.
\end{proof}
The above result on symmetric states leads to the following corollary.
\begin{corollary}\label{monc1}
None of the three reduced states of any three qubit symmetric state $\rho_{ABC}$ violates $F_{3}$ inequality i.e steering is monogamous for such states with respect to $F_{3}$ inequality.
\end{corollary}
\begin{figure}
\includegraphics[scale=0.32]{monop.png}
\caption{Steering Graphs: Here each circle represents physical system and a solid line connecting two systems describes the bipartite steering correlation between them. Different possibilities of sharing bipartite steering among three distant physical systems are depicted in this figure. (a) No bipartite steering is detected between individual parties. For example, the tripartite GHZ state \cite{Ghz} $|\phi_{GHZ}\rangle = \frac{|000\rangle + |111\rangle}{\sqrt{2}}$ has no bipartite steering. (b) bipartite steering of one reduced state is detected. One such kind of state is pure biseparable state. (c) Two bipartite reduced states are steerable. As we have shown in sec.(\ref{monsec1}), the state $|\psi_{ABC}\rangle$ belongs to this group. (d) Trade-off relation Eq.(\ref{mon19}) prevents bipartite steering between every pair of systems.}\label{monp}
\end{figure}
Theorem\ref{mont1} guarantees existence of three qubit states for which all two-party
reduced states with respect to a certain observer violate the $F_{3}$ inequality(Fig.\ref{monp}). This non-monogamous nature of steering allows one to reveal shareability(non-monogamous nature) of the entanglement of bipartite mixed states. As far as the shareability of quantum correlations is concerned, quantum entanglement is strictly speaking only monogamous in the case of pure entangled states. If the state of two systems, say $\rho_{AB}$ is a mixed entangled state, then it is possible that both of the two systems $A$ and $B$ are entangled to a third system, say $C$. For example, the so-called $W$ state \cite{Dur} $|W\rangle = \frac{(|001\rangle + |010\rangle + |100\rangle)}{\sqrt{3}}$ has bipartite reduced states that are all identical and entangled. Thus, entanglement of these reduced bipartite mixed states is sharable (non-monogamous), however, the steering correlations obtainable from these states follow the monogamy inequality Eq.(\ref{mon11}). So, by considering $F_{2}$ inequality, one cannot reveal shareability of entanglement of bipartite mixed states. To reveal this, steering correlations obtainable from these states must be non-monogamous. As shown above in Theorem\ref{mont1}, the state $|\psi_{ABC}\rangle$ (Eq.(\ref{mon21})) provides steerable bipartite reduced states between subsystems $AC$ and $BC$. Therefore the corresponding reduced mixed states $\rho_{AC}$ and $\rho_{BC}$ are also entangled and the two qubit mixed entangled state $\rho_{AC}$ is shareable to at least one other qubit. This in turn indicates that $F_{3}$ inequality is an appropriate ingredient to reveal shareability of entanglement of mixed states.\\
Unlike the standard $|W\rangle$ state, the state $|\psi_{ABC}\rangle$ can be used as a resource for deterministic teleportation and dense coding \cite{Pat}. As another application of the non-monogamous nature of steering correlations, consider that a pure three qubit state is provided to experimentalists which they have to use as a resource in deterministic teleportation or dense coding. They are also provided with the information that the state is either $|\psi_{ABC}\rangle$ or $|W\rangle.$ We show that the non-monogamy phenomenon as described in theorem\ref{mont1} can be used to determine the desired state. For $|W\rangle$ state, $S_{ij} = 1$ for all reduced states, so monogamy is preserved. On the other hand, the state $|\psi_{ABC}\rangle$ does not follow the monogamy as shown in theorem \ref{mont1}. Thus, the above result distinguishes the two types of states though they belong to the same class ($W$-like states \cite{Dur}). \\
Keeping in mind the usefulness of shareability relations, one naturally would be interested to know which of the three qubit states obey monogamy(or non-monogamy) of steering. The explicit evaluation of the number of reduced steerable states along with monogamy(or non-monogamy) in each class of three qubit pure states as classified by Sab\'{i}n and Garc\'{i}a-Alcaine\cite{Sab} is reported in Appendix \ref{Classi}, where we see that only star shaped states and $W$-like states can be non-monogamous. Next we ask whether non-monogamous behavior of those two classes of pure states is robust against white noise admixture. The results are presented in Appendix \ref{Classi}, where it is shown that less entangled states can be more robust against white noise admixture in comparison to higher entangled states. \\
Other than constraint given by Eq. (\ref{mon20}) and its permutations, few other conditions are also derived in the following sections under which $F_{3}$ steering is non-monogamous.
\section{Reduced Steering Versus Entanglement}\label{monsec3}
In two qubit systems, the more entangled a pure state is, the more it can violate the Bell-CHSH inequality. In this context, a relevant study is to find the relation between violation of $F_{3}$ inequality by the reduced bipartite states of a pure state and their corresponding entanglement(with respect to some measure).
The relation between $S_{AB}$ and concurrence $\mathcal{C}_{AB}$(a measure of entanglement)\cite{Hil,Wot} can be derived with similar methods used in \cite{Ves}. For pure bipartite states the relation is $S_{AB} = 1+2 \, \mathcal{C}_{AB}^{2}$. Hence, for pure states, more entanglement generates larger violation of $F_{3}$ inequality. However from this relation we cannot infer anything about mixed bipartite reduced states of a three qubit pure state. In the theorem below we derive a relation justifying our claim.
\begin{theorem}\label{mont4}
The triples $(S_{AB},S_{AC},S_{BC})$ of three reduced states obtained from a pure three qubit state and $(\mathcal{C}_{AB},\mathcal{C}_{AC},\mathcal{C}_{BC})$ maintain the same ordering i.e.,
\begin{equation}\label{mon24}
S_{AB}> S_{AC} > S_{BC}\,\,\,\,\, \text{iff} \,\,\,\, \mathcal{C}_{AB} > \mathcal{C}_{AC} > \mathcal{C}_{BC}.
\end{equation}
\end{theorem}
\begin{proof}
By eliminating $\vec{a}$ from Eq.(\ref{mon15}) and Eq.(\ref{mon16}), we have
\begin{equation}\label{mon25}
S_{AB} - S_{AC} = 3 (\vec{c}^{2} - \vec{b}^{2}).
\end{equation}
Now, three tangle $\tau$ \cite{Cof}, for three qubit pure state, is given by\cite{Che}
\begin{align}\label{mon26}
&&\tau= 1 - \vec{a}^{2} - \mathcal{C}_{AB}^{2} - \mathcal{C}_{AC}^{2}\nonumber \\
&&= 1 - \vec{b}^{2} - \mathcal{C}_{AB}^{2} - \mathcal{C}_{BC}^{2}\nonumber \\
&&= 1 - \vec{c}^{2} - \mathcal{C}_{AC}^{2} - \mathcal{C}_{BC}^{2}
\end{align}
Comparing these equalities, we obtain
\begin{equation}\label{mon27}
\mathcal{C}_{AB}^{2} - \mathcal{C}_{AC}^{2} = \vec{c}^{2} - \vec{b}^{2}
\end{equation}
and its permutations, which immediately lead to
\begin{equation}\label{mon28}
S_{AB} - S_{AC} = 3 (\mathcal{C}_{AB}^{2} - \mathcal{C}_{AC}^{2})
\end{equation}
and its permutations. Thus, we have developed the ordering relation as per Eq.(\ref{mon24}).
\end{proof}
It is interesting to note that $(S_{AB},S_{AC},S_{BC})$ and $(\vec{c}^{2}, \vec{b}^{2}, \vec{a}^{2} )$ follow the same ordering for all pure three qubit state.\\
Distribution of bipartite quantum entanglement (i.e., entanglement of reduced bipartite states) of any pure three qubit state is subjected to certain shareability laws. In particular, addition of squared concurrence of all bipartite reduced states cannot be greater than $\frac{4}{3}$\cite{Dur},
\begin{equation}\label{mon100}
\mathcal{C}_{AB}^{2} + \mathcal{C}_{AC}^{2} + \mathcal{C}_{BC}^{2} \leq \frac{4}{3}.
\end{equation}
This shareability constraint indicates that shareability of reduced bipartite steerability as well as individual bipartite steerability of any pure three qubit state might depend on concurrence of reduced bipartite states. This is in fact the case. We next discuss few results in this direction.
\begin{theorem}\label{mont4}
If the squared concurrence of any bipartite reduced state for a pure three qubit state is greater than $\frac{4}{9}$, then the corresponding reduced state is $F_{3}$ steerable i.e.,
if $\mathcal{C}_{ij}^{2} > \frac{4}{9}$ $(i \neq j, i, j = A, B, C)$, then the corresponding reduced state $\rho_{ij}$ is $F_{3}$ steerable.
\end{theorem}
\begin{proof}
By using Eqs. (\ref{mon27}) and (\ref{mon15},\ref{mon16},\ref{mon17}), each of $S_{ij}$ can be expressed in terms of $\mathcal{C}_{ij}$,
\begin{equation}\label{mon99}
S_{AB} = 1 + 2 \, \mathcal{C}_{AB}^{2} - \mathcal{C}_{AC}^{2} - \mathcal{C}_{BC}^{2}
\end{equation}
and its permutations.
Let, $\mathcal{C}_{AB}^{2} = \frac{4}{9} + \epsilon$, where $\epsilon$ is sufficiently small positive number. This immediately restricts the sum of squared concurrence of other two bipartite reduced states,
\begin{equation}\label{mon98}
\mathcal{C}_{AC}^{2} + \mathcal{C}_{BC}^{2} \leq \frac{8}{9} - \epsilon.
\end{equation}
Applying these to the expression of $S_{AB}$, this leads to the sharp inequality $S_{AB} \geq 1 + \epsilon$. So, if $\mathcal{C}_{AB}^{2} > \frac{4}{9}$, the $F_{3}$ inequality is violated. Similarly, it can be proved for other bipartite reduced states.
\end{proof}
This result holds for all three qubit pure states. As an example, consider the pure state $|\psi_{ABC}\rangle$ which has two $F_{3}$ steerable reduced states $\rho_{AC}$ and $\rho_{BC}$ with $\mathcal{C}_{AC}^{2} = \mathcal{C}_{BC}^{2} = \frac{1}{2} > \frac{4}{9}$.
However, one should note that the above inequality $\mathcal{C}_{ij}^{2} > \frac{4}{9}$ is only a sufficient condition for $F_{3}$ steerability of reduced bipartite state $\rho_{ij}$, because there are reduced states which violate the inequality $\mathcal{C}_{ij}^{2} > \frac{4}{9}$, but still give rise to $F_{3}$ steerability. One such example is $|\phi_{con}\rangle = \frac{\sqrt{3}}{2} |000\rangle + \frac{1}{2\sqrt{2}} |101\rangle + \frac{1}{2\sqrt{2}} |110\rangle$. For this state, from the above formulae one can obtain $\mathcal{C}_{AB}^{2} = \frac{3}{8} < \frac{4}{9}$ and $S_{AB} = 1 + \frac{5}{16}$. Clearly, the reduced state $\rho_{AB}$ violates $F_{3}$ inequality, while it violates the inequality $\mathcal{C}_{AB}^{2} > \frac{4}{9}$.
Although, an obvious necessary and sufficient condition can be derived from Eq.(\ref{mon99}) and its permutations.
\begin{corollary}\label{monc2}
Any reduced state $\rho_{ij}$ of a three qubit pure state will violate $F_{3}$ inequality if and only if squared concurrence of the corresponding reduced state is greater than the average of the squared concurrence of the remaining two reduced states,i.e.,\\
$S_{ij}>1$ if and only if $\mathcal{C}_{ij}^{2} > \frac{\mathcal{C}_{ik}^{2} + \mathcal{C}_{jk}^{2}}{2}$, where $ i \neq j \neq k$ and $i, j, k = A, B, C$.
\end{corollary}
Due to the shareability constraint Eq.(\ref{mon100}), violation of one of the reduced states (say, $\rho_{AB}$) puts strong restriction on the
average of squared concurrences of the remaining reduced states.
\begin{corollary}\label{monc3}
For any $F_{3}$ steerable reduced state $\rho_{ij}$, the following inequality holds : \\
$\frac{\mathcal{C}_{ik}^{2} + \mathcal{C}_{jk}^{2}}{2} < \frac{4}{9}$, where $ i \neq j \neq k$ and $i, j, k = A, B, C$.
\end{corollary}
As shown in \cite{Cof}, sum of squared concurrence between $i$ and $k$, and the squared concurrence between $j$ and $k$, cannot be greater than $1$, i.e., $\mathcal{C}_{ik}^{2} + \mathcal{C}_{jk}^{2} \leq 1$. Hence, from the above corollary it is observed that this restriction is further strengthened if one consider $F_{3}$ steerability of $\rho_{ij}$. \\
Since the last corollary (\ref{monc3}) puts more stringent restriction, using it we get the following sufficient condition for monogamy of $F_{3}$ steering:
\begin{corollary}\label{monc4}
For any pure three qubit state $\rho_{ABC}$, steering correlations will obey monogamy if $\mathcal{C}_{ik}^{2} + \mathcal{C}_{jk}^{2} \geq \frac{8}{9}$, where $ i \neq j \neq k$ and $i, j, k = A, B, C$ holds for at least two of three possible cases.
\end{corollary}
It may be noted that theorem \ref{mont4}, gives rise to a sufficient condition for non-monogamy of $F_{3}$ steerability.
\begin{corollary}\label{monc5}
$F_{3}$ steering is non-monogamous if $\mathcal{C}_{ij}^{2} > \frac{4}{9}$ ($ i \neq j$ and $i, j = A, B, C$) for any two pairs of $i$, $j$.
\end{corollary}
Now, we discuss how the $F_{3}$ inequality violation by the reduced bipartite states depends on the genuine entanglement of the three qubit state. As shown in Sec.(\ref{monsec1}), maximum two bipartite reduced states of $\rho_{ABC}$ can violate the $F_{3}$ inequality, so the bipartite steering of $\rho_{ABC} $ implies that it comes from one component of either this triple $(S_{AB}, S_{AC}, S_{BC})$ or $((S_{AB}, S_{AC}),(S_{AB}, S_{BC}),(S_{AC},S_{BC}))$. Considering both the possibilities, we adopt two different measures: $S^{\max}(\rho_{ABC})$ and $S_{total}^{max}(\rho_{ABC})$, where
$S^{\max}(\rho_{ABC}) = \max{\{S_{AB},S_{AC},S_{BC}\}}$ and $S_{total}^{\max}(\rho_{ABC}) = \max{\{S_{AB}+ S_{AC}, S_{AB} + S_{BC}, S_{AC}+ S_{BC}\}}$.\\
In each case, we will now derive a relation with tripartite entanglement of $\rho_{ABC}$.
\begin{theorem}\label{mont5}
For an arbitrary three qubit state $\rho_{ABC}$, the three tangle $\tau(\rho_{ABC})$ and maximum bipartite steering($S^{\max}(\rho_{ABC})$) with respect to $F_{3}$ inequality obeys the following complementary relation :
\begin{equation}\label{mon29}
S^{max}(\rho_{ABC}) + 2 \tau(\rho_{ABC}) \leq 3.
\end{equation}
\end{theorem}
\begin{proof}
Note that for pure three qubit state Eq.(\ref{mon15}) provides $S_{AB} = 1 + 2 \vec{c}^{2} - \vec{a}^{2} - \vec{b}^{2}$.
Incorporating this with the third equality of three tangle in Eq.(\ref{mon26}), we obtain
\begin{align}\label{mon30}
S_{AB} + 2 \tau(\rho_{ABC}) &= 3 - \vec{a}^{2} - \vec{b}^{2} - 2 \mathcal{C}_{AC}^{2} - 2 \mathcal{C}_{BC}^{2} \nonumber\\
& \leq 3
\end{align}
Similarly, one has $S_{AC} + 2 \tau(\rho_{ABC}) \leq 3 $ and $S_{BC} + 2 \tau(\rho_{ABC}) \leq 3$.
Hence for pure state $S^{max}(\rho_{ABC}) + 2 \tau(\rho_{ABC}) \leq 3.$
As the three tangle $\tau$ and $S^{max}(\rho_{ABC})$ both are convex under mixing, it implies that the relation in Eq.(\ref{mon29}) holds for all three qubit states.
\end{proof}
This complementary relation suggests that $F_{3}$ inequality violation by the reduced bipartite states depends on the tripartite entanglement present in the tripartite system.
We determine a class of three qubit genuinely entangled states which saturates the above-mentioned relation. This single parameter class of states is given by $|\phi_{m}\rangle = \frac{|000\rangle +m(|101\rangle + |010\rangle) + |111\rangle}{\sqrt{2 + 2m^{2}}}$, where $m \in [0,1]$. The above class of states has been identified in \cite{Nep} as the maximum dense-coding capable class of states. For this class of states, $S^{max}(|\phi_{m}\rangle)= 1 + \frac{8 m^{2}}{(1+m^{2})^{2}}$ and $\tau(|\phi_{m}\rangle) = 1 - \frac{4 m^{2}}{(1+m^{2})^{2}}$. Hence, for this class of states, one can show the follwing relation : $S^{max}(|\phi_{m}\rangle) + 2 \tau(|\phi_{m}\rangle) = 3$.
\begin{theorem}\label{mont6}
For an arbitrary three qubit state $\rho_{ABC}$, the three tangle $\tau(\rho_{ABC})$ and maximum bipartite steering($S^{\max}_{total}(\rho_{ABC})$) satisfy the following complementary relation:
\begin{equation}\label{mon80}
S^{max}_{total}(\rho_{ABC}) + \tau(\rho_{ABC}) \leq 3.
\end{equation}
\end{theorem}
\begin{proof}
Combining Eqs.(\ref{mon15},\ref{mon16}) and last two equalities of Eq.(\ref{mon26}), we get:
\begin{align*}
S_{AB} + S_{AC} + 2 \tau(\rho_{ABC}) & = 4 - 2 \vec{a}^{2} - \mathcal{C}_{AC}^{2} - \mathcal{C}_{AC}^{2} - 2\, \mathcal{C}_{BC}^{2} \nonumber\\
&= 3 + \tau - \vec{a}^{2} - 2\, \mathcal{C}_{BC}^{2}.\nonumber\\
\end{align*}
Thus,
\begin{equation}\label{mon31}
S_{AB} + S_{AC} + \tau(\rho{ABC}) \leq 3.
\end{equation}
Considering all permutation of parties, we get $S_{AB} + S_{BC} + \tau(\rho{ABC}) \leq 3 $ and $S_{AC} + S_{BC} + \tau(\rho_{ABC}) \leq 3$.\\
Now, by using the convexity property of the left hand sides of these inequalities, we claim that the relation(Eq.(\ref{mon29})) holds for all three-qubit states.
\end{proof}
We have identified a class of genuinely entangled states which saturates the afore-mentioned relation. This class of states is given by $|\phi_{q}\rangle = \frac{1}{\sqrt{2}}|000\rangle + \sqrt{\frac{1}{2}-q^{2}} |101\rangle + q |111\rangle$ where $q \in (0, \frac{1}{\sqrt{2}})$.
For $|\phi_{q}\rangle$, $S_{total}^{max} = 3 - 2q^{2}$ and $\tau = 2 q^{2}$. Hence, $S^{max}_{total}(\rho_{ABC}) + \tau(\rho_{ABC})= 3$. However, $|\phi_{q}\rangle$ has only one reduced state which violate $F_{3}$ inequality. Since, among all pure three qubit GHZ class of states only star shaped states can have two reduced steerable states (see Appendix \ref{Classi}) and for this class of states, $S^{max}_{total}(\rho_{ABC}) + \tau(\rho_{ABC}) < 3$, so there is no three qubit pure state with $\tau \neq 0$ having two reduced bipartite steerable states which saturates the above inequality.\\
All the above-mentioned relations are obtained with respect to three tangle. However, three tangle is not a good measure of genuine tripartite entanglement even for pure states as there exists a large number of pure states ($W$-like states\cite{Dur}) for which it becomes zero. Hence, none of the relations are meaningful for those $W$-like states.\\
To obtain such relations for $W$-like states, we consider the measure for $W$ entanglement introduced by Dur et. al. \cite{Dur}, defined as
$E_{W} = \min\{\mathcal{C}_{AB}^{2},\mathcal{C}_{AC}^{2},\mathcal{C}_{BC}^{2}\}$.
Any pure state $\rho_{ABC}$ contains $W$ entanglement if $E_{W} > 0$. The $W$ entanglement $E_{W}$ achieves its maximum value $\frac{4}{9}$ in the $|W\rangle$ state.
\begin{theorem}\label{mont7}
For an arbitrary three qubit pure state $|\phi_{ABC}\rangle$, the $W$ entanglement $(E_{W})$ and maximum bipartite steering($S^{\max}_{total}(\rho_{ABC})$) satisfies the following complementary relation:
\begin{equation}\label{mon32}
S^{max}_{total}(|\phi_{ABC}\rangle) + 3 E_{W}(|\phi_{ABC}\rangle) \leq \frac{10}{3}.
\end{equation}
\end{theorem}
\begin{proof}
Using Eq.(\ref{mon99}) and its permutations, we have
\begin{equation}\label{mon33}
S_{AB} + S_{AC} + 3\, \mathcal{C}_{BC}^{2} = 2 + \mathcal{C}_{AB}^{2} + \mathcal{C}_{AC}^{2} + \mathcal{C}_{BC}^{2}.
\end{equation}
If one uses Eq.(\ref{mon100}), the above equality immediately leads to
\begin{equation}\label{mon34}
S_{AB} + S_{AC} + 3 \, \mathcal{C}_{BC}^{2} \leq \frac{10}{3}.
\end{equation}
Similarly, permutation of parties gives $S_{AB} + S_{BC} + 3\, \mathcal{C}_{AC}^{2} \leq \frac{10}{3}$, and $S_{AC} + S_{BC} + 3 \, \mathcal{C}_{AB}^{2} \leq \frac{10}{3}$.
The above equations confirm the validity of the claim made in Eq.(\ref{mon32}).
\end{proof}
This relation imposes a restriction on the bipartite steering for a given amount of $W$ entanglement and it is saturated by $|W\rangle$ state. \\
We have also investigated such complementary relations for bipartite nonlocality(with respect to Bell-CHSH violation), bipartite steering and three tangle. Following the same procedure as before, a similar trade-off relation can be obtained for them:
\begin{equation}\label{mon34}
S^{\max}(\rho_{ABC}) + \mathcal{M}^{\max}(\rho_{ABC}) + 3 \,\tau(\rho_{ABC}) \leq 5
\end{equation}
where $\mathcal{M}^{\max}(\rho_{ABC}) = \max \{\mathcal{M}_{AB}, \mathcal{M}_{AC}, \mathcal{M}_{BC}\}$ and $\mathcal{M} = u_{1}^{2} + u_{2}^{2}$ is the Horodecki parameter \cite{Hro} used for measuring the degree of Bell-CHSH violation. $u_{1}^{2}, u_{2}^{2}$ are being the largest two eigen values of $T^{T}_{AB} T_{AB}$.
\section{Complementary relations for local and nonlocal information contents}
Total information content of a three qubit state can be divided into two forms: local and nonlocal information contents. Local information can be defined as \cite{Zha} :
\begin{equation}\label{mon35}
I_{local} = \vec{a}^{2} + \vec{b}^{2} + \vec{c}^{2}.
\end{equation}
To derive the complementary relation between local and nonlocal information contents, we consider only bipartite nonlocal information present in the three qubit state. Bipartite nonlocal information content can be defined as,
\begin{equation}\label{mon36}
I_{nonlocal} = \max\{N_{AB} + N_{AC}, N_{AB} + N_{BC}, N_{AC} + N_{BC}\}
\end{equation}
where $N_{ij} = \max\{0, S_{ij} - 1\}$, $i \neq j$ and $i, j = A, B, C$, quantifies the amount of $F_{3}$ inequality violation and hence the steering nonlocal correlations of the two qubit state $\rho_{ij}$ .
\begin{theorem}
For an arbitrary three qubit state $\rho_{ABC}$,
\begin{equation}\label{mon37}
I_{local} + I_{nonlocal} \leq 3.
\end{equation}
\end{theorem}
\begin{proof}
For pure three qubit states, it is straightforward to check that
\begin{align}\label{mon38}
I_{local} + (S_{AB} -1) + (S_{AC} - 1) &= 2(\vec{b}^{2} + \vec{c}^{2})- \vec{a}^{2}\nonumber \\
&\leq 2(1+\vec{a}^{2}) - \vec{a}^{2}\nonumber \\
& \leq 3
\end{align}
where, in the first inequality we have used the fact that relation $\vec{b}^{2} + \vec{c}^{2} \leq 1 + \vec{a}^{2}$ holds for all pure three qubit states \cite{Hig}.
Since $I_{local} \leq 3$, the above inequality(Eq.(\ref{mon38})) also holds when both of $N_{AB}$ and $N_{AC}$ are equal to zero. Hence $I_{local} + N_{AB} + N_{AC} \leq 3$. Similarly, one gets $I_{local} + N_{AB} + N_{BC} \leq 3$ and $I_{local} + N_{AC} + N_{BC} \leq 3$.
Note that the left hand sides of these inequalities are convex under mixing. This confirms the relation presented in Eq.(\ref{mon37}).
\end{proof}
The above trade-off relation links between local information and bipartite steering. One can easily show that $I_{local} = 3$, and $I_{nonlocal} = 0$ for the product state. On the other hand, in order to exist bipartite steering, $I_{local}$ must be less than $3$. For $|\psi_{ABC}\rangle$(Eq.(\ref{mon21})), $I_{local} = \frac{1}{2}$, $I_{nonlocal} = 2 + \frac{1}{2}$ and it is the state which saturates this trade-off.
In this context, it may be noted that to get larger violation of $F_{3}$ inequality(characterizing larger amount of steering), the amount of local information content must be reduced. This fact will be confirmed in the next section, where we will show that amount of local information content must be less than one for any three qubit pure state to have two $F_{3}$ steerable bipartite reduced states.
\section{Entanglement detection}\label{monsec2}
We now illustrate the relevance of the above results with
some applications. By using the shareability relations, we will derive criteria of detecting different types of tripartite entanglement.
\begin{theorem}\label{mont2}
For any three qubit pure state $|\phi_{ABC}\rangle$$ \in$$ \mathbb{H}^{A} \otimes \mathbb{H}^{B} \otimes \mathbb{H}^{C}$, if at least one of the following conditions holds:
\begin{equation}\label{mon22}
(i)\vec{a}^{2} \neq \frac{\vec{b}^{2} + \vec{c}^{2}}{2},
(ii)\vec{b}^{2} \neq \frac{\vec{a}^{2} + \vec{c}^{2}}{2},
(iii)\vec{c}^{2} \neq \frac{\vec{a}^{2} + \vec{b}^{2}}{2}
\end{equation}
then the state is entangled.
\end{theorem}
\begin{proof}
Let $|\phi_{ABC}\rangle$ be a separable state, then all bipartite reduced states are also separable and $S_{AB}, S_{AC}, S_{BC} \leq 1.$ Hence violation of $F_{3}$ inequality by any bipartite reduced state entails entanglement of $|\phi_{ABC}\rangle$.
It is clear from Eq.(\ref{mon15})-Eq.(\ref{mon17} that if $S_{AB}$, $S_{AC}$ and $S_{BC}>1$ then
$\vec{c}^{2} > \frac{\vec{a}^{2} + \vec{b}^{2}}{2}$, $\vec{b}^{2} > \frac{\vec{a}^{2} + \vec{c}^{2}}{2}$ and $\vec{a}^{2} > \frac{\vec{b}^{2} + \vec{c}^{2}}{2}$ hold respectively.
Again by adding Eq.(\ref{mon15}) and Eq.(\ref{mon16}), we have $S_{AB} + S_{AC} = 2 + \vec{b}^{2} + \vec{c}^{2} - 2 \vec{a}^{2} $. By noting that $S_{AB} + S_{AC}> 2$ implies steerability of at least one of $\rho_{AB}$ or $\rho_{AC}$, $|\phi_{ABC} \rangle$ is entangled if $\vec{a}^{2} < \frac{\vec{b}^{2} + \vec{c}^{2}}{2}$. Similarly permutation of the parties gives $\vec{b}^{2} < \frac{\vec{a}^{2} + \vec{c}^{2}}{2}$ and $\vec{c}^{2} < \frac{\vec{a}^{2} + \vec{b}^{2}}{2}$. Combining all these expressions, we arrive at Eq.(\ref{mon22})
\end{proof}
Now one may enquire whether condition (\ref{mon22}) is also necessary for entanglement. Unfortunately, this is not the case. For example, consider the $|W\rangle$ state, which does not satisfy (\ref{mon22}), but is entangled.
\begin{theorem}\label{mont3}
For any three qubit pure state $|\phi_{ABC}\rangle $$\in$$ \mathbb{H}^{A} \otimes \mathbb{H}^{B} \otimes \mathbb{H}^{C}$, if at least one of the following conditions holds:
\begin{align}\label{mon23}
(i) && \vec{a}^{2} > \frac{\vec{b}^{2} + \vec{c}^{2}}{2},
\vec{b}^{2} > \frac{\vec{a}^{2} + \vec{c}^{2}}{2},\nonumber \\
(ii) && \vec{a}^{2} > \frac{\vec{b}^{2} + \vec{c}^{2}}{2},
\vec{c}^{2} > \frac{\vec{a}^{2} + \vec{b}^{2}}{2},\nonumber \\
(iii) && \vec{b}^{2} > \frac{\vec{a}^{2} + \vec{c}^{2}}{2},
\vec{c}^{2} > \frac{\vec{a}^{2} + \vec{b}^{2}}{2} \nonumber \\
\end{align}
then the state is genuinely entangled.
\end{theorem}
\begin{proof}
Let $|\phi_{ABC}\rangle$ be any bi-separable state in which $AB$ is independent of $C$, then it can be expressed as $(\cos\theta |00\rangle + \sin\theta |11\rangle)_{AB} \otimes |0\rangle_{C}$ where $0\leq \theta \leq \frac{\pi}{4}$. For this state, $\vec{c}^{2} = 1$ and $\vec{a}^{2} = \vec{b}^{2}$. Using Eq.(\ref{mon15})-Eq.(\ref{mon17}), one can find that $S_{AB} = 3 - 2 \vec{a}^{2}, S_{AC} = \vec{a}^{2}, S_{BC} = \vec{a}^{2}$. So, only $S_{AB}$ can be greater than $1$. Similarly, one can show that only one reduced state will violate the $F_{3}$ inequality in which another system other than C system factorizes. This immediately leads to a simple sufficient condition for genuinely entangled pure states: Violation of $F_{3}$ inequality by two reduced states indicates genuine entanglement of $|\phi_{ABC}\rangle$. Then, from
Eq.(\ref{mon15})-Eq.(\ref{mon17}), we obtain the conditions (\ref{mon23}).
\end{proof}
It is important to note that for a pure biseparable state $\vec{a}^{2} + \vec{b}^{2} + \vec{c}^{2} \geq 1$ and exactly one of the reduced bipartite states is $F_{3}$ steerable. Therefore, for the existence of two $F_{3}$ steerable bipartite reduced states of a three qubit pure state, $\vec{a}^{2} + \vec{b}^{2} + \vec{c}^{2} < 1$ must hold. This condition can be treated as a necessary condition for a three qubit pure state to have two $F_{3}$ steerable bipartite reduced states. However, this is not sufficient, for example, $\vec{a}^{2} + \vec{b}^{2} + \vec{c}^{2} = \frac{1}{3} < 1$ for $|W\rangle$ state, but no reduced bipartite state of this state is $F_{3}$ steerable.\\
At this stage a pertinent question would be whether there exists any biseparable mixed state which has more than one reduced steerable states. Let us consider the following example:
\begin{align}\label{mon90}
&|\phi_{b}\rangle = \frac{4}{9}(1 + \epsilon) {|\phi^{+}\rangle \langle \phi^{+}|}_{AB} \otimes |0\rangle \langle 0|_{C} +\frac{4}{9}(1 + \epsilon) \times \nonumber \\
& {|\phi^{+}\rangle \langle \phi^{+}|}_{AC} \otimes |0\rangle \langle 0|_{B} + \frac{1}{9}(1 - 8\epsilon) {|\phi^{+}\rangle \langle \phi^{+}|}_{BC} \otimes |0\rangle \langle 0|_{A} \nonumber\\
\end{align}
where $0 \leq \epsilon \leq 1$ and $|\phi^{+}\rangle = \frac{|00\rangle + |11\rangle}{\sqrt{2}}.$ For this biseparable mixed state, the bipartite reduced states $\rho_{AB}$ and $\rho_{AC}$ are $F_{3}$ steerable if $\epsilon > \frac{9}{4\sqrt{3}} - 1$. Thus, genuine entanglement is not necessary to reveal non-monogamous nature of steering correlations.
\section{Discussions}\label{sec4}
Analysis of shareability of correlations between parties sharing a quantum system is an effective way of interpreting quantum theory. In this paper, we have investigated the shareability properties of quantum steering correlations. For our purpose, we have considered the three settings linear steering ($F_{3}$) inequality. Interestingly it is observed that at most two reduced states of any arbitrary three qubit state can violate the $F_{3}$ inequality. This in turn reveal non monogamous nature of steering correlations. Such an observation is however in contrary to the monogamous nature of steering obtained while using two setting linear steering inequality or Bell-CHSH inequality. This indicates that steering correlations can be non-monogamous depending on the measurement scenario.
Now steering correlations in a setup with two settings per party cannot be shared, whereas the same is possible when a setup with three settings per party is considered. So it might be tempting to think that increase of more settings per party could provide more steerable reduced states. Consequently, it will be interesting if one investigate this shareability phenomenon for more than three settings scenario.\\
We have also addressed the question how different measures of genuine entanglement and also entanglement of reduced states relate with reduced bipartite steering of three qubit states. We have established several relations between reduced bipartite steering and different measures of entanglement. Relation existing between bipartite steering, Bell-CHSH nonlocality, and genuine entanglement for three qubit states has also been analyzed.\\
Then, we have determined the complementarity relation between the local information content and bipartite steering. We believe that this will be helpful in designing some appropriate information-theoretic measures of steering. Moreover, we have shown that the shareability constraints allow us to detect different types of tripartite entanglement. Now, monogamy is the essential part in ensuring the security of quantum cryptographic protocols \cite{Pal}. For this reason, it is beneficial to capture precisely under what condition steering correlations is monogamous. So, our observations may be used in framing some more secured quantum cryptographic protocols.\\
We hope that our results will be useful for further understanding of formalism underlying steering correlations and their distribution in multipartite states. Apart from investigating our work for more than three setting scenario, it will be interesting to generalize the shareability concept of steering correlations and relations between different quantum correlations for more than two party reduced states. Also, investigation of the same beyond qubit systems is a source of potential future research.
| 3f7d38b1e6295ae818592aaaa6bb6d06143d04a5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Sample for first level head}\label{sec1}
Lorem ipsum dolor sit amet, consectetuer adipiscing elit~\citet{Rothermel1997} Ut purus elit, vestibulum ut, placerat ac, adipiscing vitae,
felis. Curabitur dictum gravida mauris. Nam arcu libero, nonummy eget, consectetuer id, vulputate a, magna. Donec
vehicula augue eu neque. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas.
Mauris ut leo. Cras viverra metus rhoncus sem. Nulla et lectus vestibulum urna fringilla ultrices. Phasellus eu tellus
sit amet tortor gravida placerat. Integer sapien est, iaculis in, pretium quis, viverra ac, nunc. Praesent eget sem vel
leo ultrices bibendum. Aenean faucibus. Morbi dolor nulla, malesuada eu, pulvinar at, mollis ac, nulla. Curabitur
auctor semper nulla. Donec varius orci eget risus. Duis nibh mi, congue eu, accumsan eleifend, sagittis quis, diam.
Duis eget orci sit amet orci dignissim rutrum.
\begin{eqnarray}
s(nT_{s}) &= &s(t)\times \sum\limits_{n=0}^{N-1} \delta (t-nT_{s}) \xleftrightarrow{\mathrm{DFT}} S \left(\frac{m}{NT_{s}}\right) \nonumber\\
&= &\frac{1}{N} \sum\limits_{n=0}^{N-1} \sum\limits_{k=-N/2}^{N/2-1} s_{k} e^{\mathrm{j}2\pi k\Delta fnT_{s}} e^{-j\frac{2\pi}{N}mn}
\end{eqnarray}
\section{Sample for another first level head}\label{sec2}
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus viverra
fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa ac
quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo~\cite{Rothermel1998} Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
Example for bibliography citations cite~\citep{Elbaum2002}, cites~\cite{Allen2011,Yoo2007}
Quisque ullamcorper placerat ipsum. Cras nibh~\cite{Yoo2007,Schulz2012} Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{figure}[t]
\centerline{\includegraphics[width=342pt,height=9pc,draft]{blankfig}}
\caption{This is the sample figure caption.\label{fig1}}
\end{figure}
Suspendisse vel felis. Ut lorem lorem, interdum eu, tincidunt sit amet, laoreet vitae, arcu. Aenean faucibus pede eu
ante. Praesent enim elit, rutrum at, molestie non, nonummy vel, nisl. Ut lectus eros, malesuada sit amet, fermentum
eu, sodales cursus, magna. Donec eu purus. Quisque vehicula, urna sed ultricies auctor, pede lorem egestas dui, et
convallis elit erat sed nulla. Donec luctus. Curabitur et nunc. Aliquam dolor odio, commodo pretium, ultricies non,
pharetra in, velit. Integer arcu est, nonummy in, fermentum faucibus, egestas vel, odio.
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor.
Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
\begin{figure*}
\centerline{\includegraphics[width=342pt,height=9pc,draft]{blankfig}}
\caption{This is the sample figure caption.\label{fig2}}
\end{figure*}
\subsection{Example for second level head}
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit, dictum
in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam vestibulum
fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi. Pellentesque
eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut mi mi, lacinia
sit amet, placerat et, mollis vitae, dui. Sed ante tellus, tristique ut, iaculis eu, malesuada ac, dui. Mauris nibh leo,
facilisis non, adipiscing quis, ultrices a, dui.
Morbi luctus, wisi viverra faucibus pretium, nibh est placerat odio, nec commodo wisi enim eget quam. Quisque
libero justo, consectetuer a, feugiat vitae, porttitor eu, libero. Suspendisse sed mauris vitae elit sollicitudin malesuada.
Maecenas ultricies eros sit amet ante. Ut venenatis velit. Maecenas sed mi eget dui varius euismod. Phasellus aliquet
volutpat odio. Vestibulum ante ipsum primis in faucibus orci luctus et ultrices posuere cubilia Curae; Pellentesque sit
amet pede ac sem eleifend consectetuer. Nullam elementum, urna vel imperdiet sodales, elit ipsum pharetra ligula,
ac pretium ante justo a nulla. Curabitur tristique arcu eu metus. Vestibulum lectus. Proin mauris. Proin eu nunc eu
urna hendrerit faucibus. Aliquam auctor, pede consequat laoreet varius, eros tellus scelerisque quam, pellentesque
hendrerit ipsum dolor sed augue. Nulla nec lacus.
\begin{quote}
This is an example~\citep{Elbaum2002,Blanchard2015} for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text~\cite{Blanchard2015}. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text~\cite{Blanchard2015}. This is an example for quote text. This is an example for quote text. This is an example for quote text.
\end{quote}
\section{Sample for next first level head}\label{sec3}
\subsection{Example for another second level head}
Suspendisse vitae elit. Aliquam arcu neque, ornare in, ullamcorper quis, commodo eu, libero. Fusce sagittis erat at
erat tristique mollis. Maecenas sapien libero, molestie et, lobortis in, sodales eget, dui. Morbi ultrices rutrum lorem.
Nam elementum ullamcorper leo. Morbi dui. Aliquam sagittis. Nunc placerat. Pellentesque tristique sodales est.
Maecenas imperdiet lacinia velit. Cras non urna. Morbi eros pede, suscipit ac, varius vel, egestas non, eros. Praesent
malesuada, diam id pretium elementum, eros sem dictum tortor, vel consectetuer odio sem sed wisi.
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
\subsection{Second level head text}
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\subsubsection{Third level head text}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra tortor. Sed fringilla
justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue est scelerisque
sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas urna mi, suscipit
in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
Nulla in ipsum. Praesent eros nulla, congue vitae, euismod ut, commodo a, wisi. Pellentesque habitant morbi
tristique senectus et netus et malesuada fames ac turpis egestas. Aenean nonummy magna non leo. Sed felis erat,
ullamcorper in, dictum non, ultricies ut, lectus. Proin vel arcu a odio lobortis euismod. Vestibulum ante ipsum primis
in faucibus orci luctus et ultrices posuere cubilia Curae; Proin ut est. Aliquam odio. Pellentesque massa turpis, cursus
eu, euismod nec, tempor congue, nulla. Duis viverra gravida mauris. Cras tincidunt. Curabitur eros ligula, varius ut,
pulvinar in, cursus faucibus, augue.
\begin{boxtext}
\section*{Example of Boxtext}%
This is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext.
\end{boxtext}
\paragraph{Fourth level head text}
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\subparagraph{Fifth level head text}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra
tortor. Sed fringilla justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue
est scelerisque sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas
urna mi, suscipit in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
in faucibus orci luctus et ultrices posuere cubilia Curae; Proin ut est. Aliquam odio. Pellentesque massa turpis, cursus
eu, euismod nec, tempor congue, nulla. Duis viverra gravida mauris. Cras tincidunt. Curabitur eros ligula, varius ut,
pulvinar in, cursus faucibus, augue.
Curabitur tellus magna, porttitor a, commodo a, commodo in, tortor. Donec interdum. Praesent scelerisque. Mae-
cenas posuere sodales odio. Vivamus metus lacus, varius quis, imperdiet quis, rhoncus a, turpis. Etiam ligula arcu,
elementum a, venenatis quis, sollicitudin sed, metus. Donec nunc pede, tincidunt in, venenatis vitae, faucibus vel,
nibh. Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo. Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
Nulla non mauris vitae wisi posuere convallis. Sed eu nulla nec eros scelerisque pharetra. Nullam varius. Etiam
dignissim elementum metus. Vestibulum faucibus, metus sit amet mattis rhoncus, sapien dui laoreet odio, nec ultricies
nibh augue a enim. Fusce in ligula. Quisque at magna et nulla commodo consequat. Proin accumsan imperdiet sem.
Nunc porta. Donec feugiat mi at justo. Phasellus facilisis ipsum quis ante. In ac elit eget ipsum pharetra faucibus.
Maecenas viverra nulla in massa.
Nulla in ipsum. Praesent eros nulla, congue vitae, euismod ut, commodo a, wisi. Pellentesque habitant morbi
tristique senectus et netus et malesuada fames ac turpis egestas. Aenean nonummy magna non leo. Sed felis erat,
ullamcorper in, dictum non, ultricies ut, lectus. Proin vel arcu a odio lobortis euismod. Vestibulum ante ipsum primis
\begin{center}
\begin{table*}[t]%
\caption{This is sample table caption.\label{tab1}}
\centering
\begin{tabular*}{500pt}{@{\extracolsep\fill}lccD{.}{.}{3}c@{\extracolsep\fill}}
\toprule
&\multicolumn{2}{@{}c@{}}{\textbf{Spanned heading\tnote{1}}} & \multicolumn{2}{@{}c@{}}{\textbf{Spanned heading\tnote{2}}} \\\cmidrule{2-3}\cmidrule{4-5}
\textbf{col1 head} & \textbf{col2 head} & \textbf{col3 head} & \multicolumn{1}{@{}l@{}}{\textbf{col4 head}} & \textbf{col5 head} \\
\midrule
col1 text & col2 text & col3 text & 12.34 & col5 text\tnote{1} \\
col1 text & col2 text & col3 text & 1.62 & col5 text\tnote{2} \\
col1 text & col2 text & col3 text & 51.809 & col5 text \\
\bottomrule
\end{tabular*}
\begin{tablenotes
\item Source: Example for table source text.
\item[1] Example for a first table footnote.
\item[2] Example for a second table footnote.
\end{tablenotes}
\end{table*}
\end{center}
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla
facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{center}
\begin{table}[t]%
\centering
\caption{This is sample table caption.\label{tab2}}%
\begin{tabular*}{500pt}{@{\extracolsep\fill}lcccc@{\extracolsep\fill}}
\toprule
\textbf{col1 head} & \textbf{col2 head} & \textbf{col3 head} & \textbf{col4 head} & \textbf{col5 head} \\
\midrule
col1 text & col2 text & col3 text & col4 text & col5 text\tnote{$\dagger$} \\
col1 text & col2 text & col3 text & col4 text & col5 text \\
col1 text & col2 text & col3 text & col4 text & col5 text\tnote{$\ddagger$} \\
\bottomrule
\end{tabular*}
\begin{tablenotes}
\item Source: Example for table source text.
\item[$\dagger$] Example for a first table footnote.
\item[$\ddagger$] Example for a second table footnote.
\end{tablenotes}
\end{table}
\end{center}
Below is the example~\cite{Rothermel1998,Yoo2007,Schulz2012} for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list\footnote{This is an example for footnote.}:
\begin{itemize}
\item bulleted list entry sample bulleted list entry~\cite{Allen2011} sample list entry text.
\item bulleted list entry sample bulleted list entry. bulleted list entry sample bulleted list entry. bulleted list entry sample bulleted list entry.
\item bulleted list entry sample bulleted list entry~\cite{Ballen2011} bulleted list entry sample bulleted list entry~\citet{Allen2011} sample list entry text. bulleted list entry sample bulleted list entry.
\item sample list entry text. sample list entry text.
\end{itemize}
Suspendisse vel felis. Ut lorem lorem, interdum eu, tincidunt sit amet, laoreet vitae, arcu. Aenean faucibus pede eu
ante. Praesent enim elit, rutrum at, molestie non, nonummy vel, nisl. Ut lectus eros, malesuada sit amet, fermentum
eu, sodales cursus, magna. Donec eu purus. Quisque vehicula, urna sed ultricies auctor, pede lorem egestas dui, et
convallis elit erat sed nulla. Donec luctus. Curabitur et nunc. Aliquam dolor odio, commodo pretium, ultricies non,
pharetra in, velit. Integer arcu est, nonummy in, fermentum faucibus, egestas vel, odio.
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor. Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
Below is the sample for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list:\\[12pt]
\noindent\textbf{Description sample:}
\begin{description}
\item[first entry] description text. description text. description text. description text. description text. description text. description text.
\item[second long entry] description text. description text. description text. description text. description text. description text. description text.
\item[third entry] description text. description text. description text. description text. description text.
\item[fourth entry] description text. description text.
\end{description}
\noindent\textbf{Numbered list items sample:}
\begin{enumerate}[1.]
\item First level numbered list entry. sample numbered list entry.
\item First numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry.
\begin{enumerate}[a.]
\item Second level alpabetical list entry. Second level alpabetical list entry. Second level alpabetical list entry~\citet{Allen2011} Second level alpabetical list entry.
\item Second level alpabetical list entry. Second level alpabetical list entry~\citet{Schulz2012,Allen2011,Ballen2011}.
\begin{enumerate}[ii.]
\item Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry.
\item Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry~\cite{Yoo2007}.
\end{enumerate}
\item Second level alpabetical list entry. Second level alpabetical list entry~\cite{Elbaum2002}.
\end{enumerate}
\item First level numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry.
\item Another first level numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry.
\end{enumerate}
\noindent\textbf{un-numbered list items sample:}
\begin{enumerate}[]
\item Sample unnumberd list text..
\item Sample unnumberd list text.
\item sample unnumberd list text.
\item Sample unnumberd list text.
\end{enumerate}
\section{Examples for enunciations}\label{sec4}
\begin{theorem}[Theorem subhead]\label{thm1}
Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text.
\end{theorem}
Quisque ullamcorper placerat ipsum. Cras nibh. Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla
facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{proposition}
Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text.
\end{proposition}
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus
viverra fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa
ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
Quisque ullamcorper placerat ipsum. Cras nibh. Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
\begin{definition}[Definition sub head]
Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text.
\end{definition}
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor.
Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit,
dictum in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam
vestibulum fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi.
Pellentesque eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut
mi mi, lacinia sit amet, placerat et, mollis vitae, dui. Sed ante tellus, tristique ut, iaculis eu, malesuada ac, dui.
Mauris nibh leo, facilisis non, adipiscing quis, ultrices a, dui.
\begin{proof}
Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text.
\end{proof}
Nam dui ligula, fringilla a, euismod sodales, sollicitudin vel, wisi. Morbi auctor lorem non justo. Nam lacus libero,
pretium at, lobortis vitae, ultricies et, tellus. Donec aliquet, tortor sed accumsan bibendum, erat ligula aliquet magna,
vitae ornare odio metus a mi. Morbi ac orci et nisl hendrerit mollis. Suspendisse ut massa. Cras nec ante. Pellentesque
a nulla. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Aliquam tincidunt
urna. Nulla ullamcorper vestibulum turpis. Pellentesque cursus luctus mauris.
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus
viverra fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa
ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
\begin{proof}[Proof of Theorem~\ref{thm1}]
Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text.
\end{proof}
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra tortor. Sed fringilla
justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue est scelerisque
sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas urna mi, suscipit
in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit,
dictum in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam
vestibulum fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi.
Pellentesque eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut Curabitur tellus magna, porttitor a, commodo a, commodo in, tortor. Donec interdum. Praesent scelerisque. Mae-
cenas posuere sodales odio. Vivamus metus lacus, varius quis, imperdiet quis, rhoncus a, turpis. Etiam ligula arcu,
elementum a, venenatis quis, sollicitudin sed, metus. Donec nunc pede, tincidunt in, venenatis vitae, faucibus vel,
\begin{sidewaystable
\caption{Sideways table caption. For decimal alignment refer column 4 to 9 in tabular* preamble.\label{tab3}}%
\begin{tabular*}{\textheight}{@{\extracolsep\fill}lccD{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}@{\extracolsep\fill}}%
\toprule
& \textbf{col2 head} & \textbf{col3 head} & \multicolumn{1}{c}{\textbf{10}} &\multicolumn{1}{c}{\textbf{20}} &\multicolumn{1}{c}{\textbf{30}} &\multicolumn{1}{c}{\textbf{10}} &\multicolumn{1}{c}{\textbf{20}} &\multicolumn{1}{c}{\textbf{30}} \\
\midrule
&col2 text &col3 text &0.7568&1.0530&1.2642&0.9919&1.3541&1.6108 \\
& &col2 text &12.5701 &19.6603&25.6809&18.0689&28.4865&37.3011 \\
3 &col2 text & col3 text &0.7426&1.0393&1.2507&0.9095&1.2524&1.4958 \\
& &col3 text &12.8008&19.9620&26.0324&16.6347&26.0843&34.0765 \\
& col2 text & col3 text &0.7285&1.0257&1.2374&0.8195&1.1407&1.3691\tnote{*} \\
& & col3 text &13.0360&20.2690&26.3895&15.0812&23.4932&30.6060\tnote{\dagger} \\
\bottomrule
\end{tabular*}
\begin{tablenotes
\item[*] First sideways table footnote. Sideways table footnote. Sideways table footnote. Sideways table footnote.
\item[$\dagger$] Second sideways table footnote. Sideways table footnote. Sideways table footnote. Sideways table footnote.
\end{tablenotes}
\end{sidewaystable}
\begin{sidewaysfigure}
\centerline{\includegraphics[width=542pt,height=9pc,draft]{blankfig}}
\caption{Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption.\label{fig3}}
\end{sidewaysfigure}
nibh. Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
\begin{algorithm}
\caption{Pseudocode for our algorithm}\label{alg1}
\begin{algorithmic}
\For each frame
\For water particles $f_{i}$
\State compute fluid flow~\cite{Rothermel1997}
\State compute fluid--solid interaction~\cite{Allen2011}
\State apply adhesion and surface tension~\cite{Ballen2011}
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute virtual water film \\(see Section~\ref{sec3})
\EndFor
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute growth direction vector \\(see Section~\ref{sec2})
\EndFor
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute $F_{\theta}$ (see Section~\ref{sec1})
\State compute $CE(s_{i},f_{j})$ \\(see Section~\ref{sec3})
\If $CE(b_{i}, f_{j})$ $>$ glaze threshold
\State $j$th water particle's phase $\Leftarrow$ ICE
\EndIf
\If $CE(c_{i}, f_{j})$ $>$ icicle threshold
\State $j$th water particle's phase $\Leftarrow$ ICE
\EndIf
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa~\cite{Rothermel1997,Elbaum2002}. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo~\cite{Elbaum2002} Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo. Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
\begin{equation}\label{eq23}
\|\tilde{X}(k)\|^2
=\frac{\left\|\sum\limits_{i=1}^{p}\tilde{Y}_i(k)+\sum\limits_{j=1}^{q}\tilde{Z}_j(k) \right\|^2}{(p+q)^2}
\leq\frac{\sum\limits_{i=1}^{p}\left\|\tilde{Y}_i(k)\right\|^2+\sum\limits_{j=1}^{q}\left\|\tilde{Z}_j(k)\right\|^2 }{p+q}.
\end{equation}
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\begin{equation}\label{eq24}
\|\tilde{X}(k)\|^2
=\frac{\left\|\sum\limits_{i=1}^{p}\tilde{Y}_i(k)+\sum\limits_{j=1}^{q}\tilde{Z}_j(k) \right\|^2}{(p+q)^2}
\leq\frac{\sum\limits_{i=1}^{p}\left\|\tilde{Y}_i(k)\right\|^2+\sum\limits_{j=1}^{q}\left\|\tilde{Z}_j(k)\right\|^2 }{p+q}.
\end{equation}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra
tortor. Sed fringilla justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue
est scelerisque sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas
urna mi, suscipit in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
\section{Conclusions}\label{sec5}
Lorem ipsum dolor sit amet, consectetuer adipiscing elit. Ut purus elit, vestibulum ut, placerat ac, adipiscing vitae,
felis. Curabitur dictum gravida mauris. Nam arcu libero, nonummy eget, consectetuer id, vulputate a, magna. Donec
vehicula augue eu neque. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas.
Mauris ut leo. Cras viverra metus rhoncus sem. Nulla et lectus vestibulum urna fringilla ultrices. Phasellus eu tellus
sit amet tortor gravida placerat. Integer sapien est, iaculis in, pretium quis, viverra ac, nunc. Praesent eget sem vel
leo ultrices bibendum. Aenean faucibus. Morbi dolor nulla, malesuada eu, pulvinar at, mollis ac, nulla. Curabitur
auctor semper nulla. Donec varius orci eget risus. Duis nibh mi, congue eu, accumsan eleifend, sagittis quis, diam.
Duis eget orci sit amet orci dignissim rutrum.
Nam dui ligula, fringilla a, euismod sodales, sollicitudin vel, wisi. Morbi auctor lorem non justo. Nam lacus libero,
pretium at, lobortis vitae, ultricies et, tellus. Donec aliquet, tortor sed accumsan bibendum, erat ligula aliquet magna,
vitae ornare odio metus a mi. Morbi ac orci et nisl hendrerit mollis. Suspendisse ut massa. Cras nec ante. Pellentesque
a nulla. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Aliquam tincidunt
urna. Nulla ullamcorper vestibulum turpis. Pellentesque cursus luctus mauris.
\section*{Acknowledgments}
This is acknowledgment text~\cite{Elbaum2002}. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here.
\subsection*{Author contributions}
This is an author contribution text. This is an author contribution text. This is an author contribution text. This is an author contribution text. This is an author contribution text.
\subsection*{Financial disclosure}
None reported.
\subsection*{Conflict of interest}
The authors declare no potential conflict of interests.
\section*{Supporting information}
The following supporting information is available as part of the online article:
\noindent
\textbf{Figure S1.}
{500{\uns}hPa geopotential anomalies for GC2C calculated against the ERA Interim reanalysis. The period is 1989--2008.}
\noindent
\textbf{Figure S2.}
{The SST anomalies for GC2C calculated against the observations (OIsst).}
\section{Sample for first level head}\label{sec1}
Lorem ipsum dolor sit amet, consectetuer adipiscing elit~\citet{Rothermel1997} Ut purus elit, vestibulum ut, placerat ac, adipiscing vitae,
felis. Curabitur dictum gravida mauris. Nam arcu libero, nonummy eget, consectetuer id, vulputate a, magna. Donec
vehicula augue eu neque. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas.
Mauris ut leo. Cras viverra metus rhoncus sem. Nulla et lectus vestibulum urna fringilla ultrices. Phasellus eu tellus
sit amet tortor gravida placerat. Integer sapien est, iaculis in, pretium quis, viverra ac, nunc. Praesent eget sem vel
leo ultrices bibendum. Aenean faucibus. Morbi dolor nulla, malesuada eu, pulvinar at, mollis ac, nulla. Curabitur
auctor semper nulla. Donec varius orci eget risus. Duis nibh mi, congue eu, accumsan eleifend, sagittis quis, diam.
Duis eget orci sit amet orci dignissim rutrum.
\begin{eqnarray}
s(nT_{s}) &= &s(t)\times \sum\limits_{n=0}^{N-1} \delta (t-nT_{s}) \xleftrightarrow{\mathrm{DFT}} S \left(\frac{m}{NT_{s}}\right) \nonumber\\
&= &\frac{1}{N} \sum\limits_{n=0}^{N-1} \sum\limits_{k=-N/2}^{N/2-1} s_{k} e^{\mathrm{j}2\pi k\Delta fnT_{s}} e^{-j\frac{2\pi}{N}mn}
\end{eqnarray}
\section{Sample for another first level head}\label{sec2}
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus viverra
fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa ac
quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo~\cite{Rothermel1998} Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
Example for bibliography citations cite~\citep{Elbaum2002}, cites~\cite{Allen2011,Yoo2007}
Quisque ullamcorper placerat ipsum. Cras nibh~\cite{Yoo2007,Schulz2012} Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{figure}[t]
\centerline{\includegraphics[width=342pt,height=9pc,draft]{blankfig}}
\caption{This is the sample figure caption.\label{fig1}}
\end{figure}
Suspendisse vel felis. Ut lorem lorem, interdum eu, tincidunt sit amet, laoreet vitae, arcu. Aenean faucibus pede eu
ante. Praesent enim elit, rutrum at, molestie non, nonummy vel, nisl. Ut lectus eros, malesuada sit amet, fermentum
eu, sodales cursus, magna. Donec eu purus. Quisque vehicula, urna sed ultricies auctor, pede lorem egestas dui, et
convallis elit erat sed nulla. Donec luctus. Curabitur et nunc. Aliquam dolor odio, commodo pretium, ultricies non,
pharetra in, velit. Integer arcu est, nonummy in, fermentum faucibus, egestas vel, odio.
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor.
Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
\begin{figure*}
\centerline{\includegraphics[width=342pt,height=9pc,draft]{blankfig}}
\caption{This is the sample figure caption.\label{fig2}}
\end{figure*}
\subsection{Example for second level head}
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit, dictum
in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam vestibulum
fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi. Pellentesque
eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut mi mi, lacinia
sit amet, placerat et, mollis vitae, dui. Sed ante tellus, tristique ut, iaculis eu, malesuada ac, dui. Mauris nibh leo,
facilisis non, adipiscing quis, ultrices a, dui.
Morbi luctus, wisi viverra faucibus pretium, nibh est placerat odio, nec commodo wisi enim eget quam. Quisque
libero justo, consectetuer a, feugiat vitae, porttitor eu, libero. Suspendisse sed mauris vitae elit sollicitudin malesuada.
Maecenas ultricies eros sit amet ante. Ut venenatis velit. Maecenas sed mi eget dui varius euismod. Phasellus aliquet
volutpat odio. Vestibulum ante ipsum primis in faucibus orci luctus et ultrices posuere cubilia Curae; Pellentesque sit
amet pede ac sem eleifend consectetuer. Nullam elementum, urna vel imperdiet sodales, elit ipsum pharetra ligula,
ac pretium ante justo a nulla. Curabitur tristique arcu eu metus. Vestibulum lectus. Proin mauris. Proin eu nunc eu
urna hendrerit faucibus. Aliquam auctor, pede consequat laoreet varius, eros tellus scelerisque quam, pellentesque
hendrerit ipsum dolor sed augue. Nulla nec lacus.
\begin{quote}
This is an example~\citep{Elbaum2002,Blanchard2015} for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text~\cite{Blanchard2015}. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text. This is an example for quote text~\cite{Blanchard2015}. This is an example for quote text. This is an example for quote text. This is an example for quote text.
\end{quote}
\section{Sample for next first level head}\label{sec3}
\subsection{Example for another second level head}
Suspendisse vitae elit. Aliquam arcu neque, ornare in, ullamcorper quis, commodo eu, libero. Fusce sagittis erat at
erat tristique mollis. Maecenas sapien libero, molestie et, lobortis in, sodales eget, dui. Morbi ultrices rutrum lorem.
Nam elementum ullamcorper leo. Morbi dui. Aliquam sagittis. Nunc placerat. Pellentesque tristique sodales est.
Maecenas imperdiet lacinia velit. Cras non urna. Morbi eros pede, suscipit ac, varius vel, egestas non, eros. Praesent
malesuada, diam id pretium elementum, eros sem dictum tortor, vel consectetuer odio sem sed wisi.
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
\subsection{Second level head text}
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\subsubsection{Third level head text}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra tortor. Sed fringilla
justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue est scelerisque
sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas urna mi, suscipit
in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
Nulla in ipsum. Praesent eros nulla, congue vitae, euismod ut, commodo a, wisi. Pellentesque habitant morbi
tristique senectus et netus et malesuada fames ac turpis egestas. Aenean nonummy magna non leo. Sed felis erat,
ullamcorper in, dictum non, ultricies ut, lectus. Proin vel arcu a odio lobortis euismod. Vestibulum ante ipsum primis
in faucibus orci luctus et ultrices posuere cubilia Curae; Proin ut est. Aliquam odio. Pellentesque massa turpis, cursus
eu, euismod nec, tempor congue, nulla. Duis viverra gravida mauris. Cras tincidunt. Curabitur eros ligula, varius ut,
pulvinar in, cursus faucibus, augue.
\begin{boxtext}
\section*{Example of Boxtext}%
This is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext this is sample for boxtext.
\end{boxtext}
\paragraph{Fourth level head text}
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\subparagraph{Fifth level head text}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra
tortor. Sed fringilla justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue
est scelerisque sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas
urna mi, suscipit in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
in faucibus orci luctus et ultrices posuere cubilia Curae; Proin ut est. Aliquam odio. Pellentesque massa turpis, cursus
eu, euismod nec, tempor congue, nulla. Duis viverra gravida mauris. Cras tincidunt. Curabitur eros ligula, varius ut,
pulvinar in, cursus faucibus, augue.
Curabitur tellus magna, porttitor a, commodo a, commodo in, tortor. Donec interdum. Praesent scelerisque. Mae-
cenas posuere sodales odio. Vivamus metus lacus, varius quis, imperdiet quis, rhoncus a, turpis. Etiam ligula arcu,
elementum a, venenatis quis, sollicitudin sed, metus. Donec nunc pede, tincidunt in, venenatis vitae, faucibus vel,
nibh. Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo. Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
Nulla non mauris vitae wisi posuere convallis. Sed eu nulla nec eros scelerisque pharetra. Nullam varius. Etiam
dignissim elementum metus. Vestibulum faucibus, metus sit amet mattis rhoncus, sapien dui laoreet odio, nec ultricies
nibh augue a enim. Fusce in ligula. Quisque at magna et nulla commodo consequat. Proin accumsan imperdiet sem.
Nunc porta. Donec feugiat mi at justo. Phasellus facilisis ipsum quis ante. In ac elit eget ipsum pharetra faucibus.
Maecenas viverra nulla in massa.
Nulla in ipsum. Praesent eros nulla, congue vitae, euismod ut, commodo a, wisi. Pellentesque habitant morbi
tristique senectus et netus et malesuada fames ac turpis egestas. Aenean nonummy magna non leo. Sed felis erat,
ullamcorper in, dictum non, ultricies ut, lectus. Proin vel arcu a odio lobortis euismod. Vestibulum ante ipsum primis
\begin{center}
\begin{table*}[t]%
\caption{This is sample table caption.\label{tab1}}
\centering
\begin{tabular*}{500pt}{@{\extracolsep\fill}lccD{.}{.}{3}c@{\extracolsep\fill}}
\toprule
&\multicolumn{2}{@{}c@{}}{\textbf{Spanned heading\tnote{1}}} & \multicolumn{2}{@{}c@{}}{\textbf{Spanned heading\tnote{2}}} \\\cmidrule{2-3}\cmidrule{4-5}
\textbf{col1 head} & \textbf{col2 head} & \textbf{col3 head} & \multicolumn{1}{@{}l@{}}{\textbf{col4 head}} & \textbf{col5 head} \\
\midrule
col1 text & col2 text & col3 text & 12.34 & col5 text\tnote{1} \\
col1 text & col2 text & col3 text & 1.62 & col5 text\tnote{2} \\
col1 text & col2 text & col3 text & 51.809 & col5 text \\
\bottomrule
\end{tabular*}
\begin{tablenotes
\item Source: Example for table source text.
\item[1] Example for a first table footnote.
\item[2] Example for a second table footnote.
\end{tablenotes}
\end{table*}
\end{center}
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla
facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{center}
\begin{table}[t]%
\centering
\caption{This is sample table caption.\label{tab2}}%
\begin{tabular*}{500pt}{@{\extracolsep\fill}lcccc@{\extracolsep\fill}}
\toprule
\textbf{col1 head} & \textbf{col2 head} & \textbf{col3 head} & \textbf{col4 head} & \textbf{col5 head} \\
\midrule
col1 text & col2 text & col3 text & col4 text & col5 text\tnote{$\dagger$} \\
col1 text & col2 text & col3 text & col4 text & col5 text \\
col1 text & col2 text & col3 text & col4 text & col5 text\tnote{$\ddagger$} \\
\bottomrule
\end{tabular*}
\begin{tablenotes}
\item Source: Example for table source text.
\item[$\dagger$] Example for a first table footnote.
\item[$\ddagger$] Example for a second table footnote.
\end{tablenotes}
\end{table}
\end{center}
Below is the example~\cite{Rothermel1998,Yoo2007,Schulz2012} for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list. Below is the example for bulleted list\footnote{This is an example for footnote.}:
\begin{itemize}
\item bulleted list entry sample bulleted list entry~\cite{Allen2011} sample list entry text.
\item bulleted list entry sample bulleted list entry. bulleted list entry sample bulleted list entry. bulleted list entry sample bulleted list entry.
\item bulleted list entry sample bulleted list entry~\cite{Ballen2011} bulleted list entry sample bulleted list entry~\citet{Allen2011} sample list entry text. bulleted list entry sample bulleted list entry.
\item sample list entry text. sample list entry text.
\end{itemize}
Suspendisse vel felis. Ut lorem lorem, interdum eu, tincidunt sit amet, laoreet vitae, arcu. Aenean faucibus pede eu
ante. Praesent enim elit, rutrum at, molestie non, nonummy vel, nisl. Ut lectus eros, malesuada sit amet, fermentum
eu, sodales cursus, magna. Donec eu purus. Quisque vehicula, urna sed ultricies auctor, pede lorem egestas dui, et
convallis elit erat sed nulla. Donec luctus. Curabitur et nunc. Aliquam dolor odio, commodo pretium, ultricies non,
pharetra in, velit. Integer arcu est, nonummy in, fermentum faucibus, egestas vel, odio.
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor. Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
Below is the sample for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list. Below is the example for description list:\\[12pt]
\noindent\textbf{Description sample:}
\begin{description}
\item[first entry] description text. description text. description text. description text. description text. description text. description text.
\item[second long entry] description text. description text. description text. description text. description text. description text. description text.
\item[third entry] description text. description text. description text. description text. description text.
\item[fourth entry] description text. description text.
\end{description}
\noindent\textbf{Numbered list items sample:}
\begin{enumerate}[1.]
\item First level numbered list entry. sample numbered list entry.
\item First numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry.
\begin{enumerate}[a.]
\item Second level alpabetical list entry. Second level alpabetical list entry. Second level alpabetical list entry~\citet{Allen2011} Second level alpabetical list entry.
\item Second level alpabetical list entry. Second level alpabetical list entry~\citet{Schulz2012,Allen2011,Ballen2011}.
\begin{enumerate}[ii.]
\item Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry.
\item Third level lowercase roman numeral list entry. Third level lowercase roman numeral list entry~\cite{Yoo2007}.
\end{enumerate}
\item Second level alpabetical list entry. Second level alpabetical list entry~\cite{Elbaum2002}.
\end{enumerate}
\item First level numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry.
\item Another first level numbered list entry. sample numbered list entry. Numbered list entry. sample numbered list entry. Numbered list entry.
\end{enumerate}
\noindent\textbf{un-numbered list items sample:}
\begin{enumerate}[]
\item Sample unnumberd list text..
\item Sample unnumberd list text.
\item sample unnumberd list text.
\item Sample unnumberd list text.
\end{enumerate}
\section{Examples for enunciations}\label{sec4}
\begin{theorem}[Theorem subhead]\label{thm1}
Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text.
\end{theorem}
Quisque ullamcorper placerat ipsum. Cras nibh. Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
Fusce mauris. Vestibulum luctus nibh at lectus. Sed bibendum, nulla a faucibus semper, leo velit ultricies tellus, ac
venenatis arcu wisi vel nisl. Vestibulum diam. Aliquam pellentesque, augue quis sagittis posuere, turpis lacus congue
quam, in hendrerit risus eros eget felis. Maecenas eget erat in sapien mattis porttitor. Vestibulum porttitor. Nulla
facilisi. Sed a turpis eu lacus commodo facilisis. Morbi fringilla, wisi in dignissim interdum, justo lectus sagittis dui, et
vehicula libero dui cursus dui. Mauris tempor ligula sed lacus. Duis cursus enim ut augue. Cras ac magna. Cras nulla.
Nulla egestas. Curabitur a leo. Quisque egestas wisi eget nunc. Nam feugiat lacus vel est. Curabitur consectetuer.
\begin{proposition}
Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text.
\end{proposition}
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus
viverra fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa
ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
Quisque ullamcorper placerat ipsum. Cras nibh. Morbi vel justo vitae lacus tincidunt ultrices. Lorem ipsum dolor sit
amet, consectetuer adipiscing elit. In hac habitasse platea dictumst. Integer tempus convallis augue. Etiam facilisis.
Nunc elementum fermentum wisi. Aenean placerat. Ut imperdiet, enim sed gravida sollicitudin, felis odio placerat
quam, ac pulvinar elit purus eget enim. Nunc vitae tortor. Proin tempus nibh sit amet nisl. Vivamus quis tortor
vitae risus porta vehicula.
\begin{definition}[Definition sub head]
Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text.
\end{definition}
Sed commodo posuere pede. Mauris ut est. Ut quis purus. Sed ac odio. Sed vehicula hendrerit sem. Duis non
odio. Morbi ut dui. Sed accumsan risus eget odio. In hac habitasse platea dictumst. Pellentesque non elit. Fusce
sed justo eu urna porta tincidunt. Mauris felis odio, sollicitudin sed, volutpat a, ornare ac, erat. Morbi quis dolor.
Donec pellentesque, erat ac sagittis semper, nunc dui lobortis purus, quis congue purus metus ultricies tellus. Proin
et quam. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Praesent sapien
turpis, fermentum vel, eleifend faucibus, vehicula eu, lacus.
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit,
dictum in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam
vestibulum fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi.
Pellentesque eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut
mi mi, lacinia sit amet, placerat et, mollis vitae, dui. Sed ante tellus, tristique ut, iaculis eu, malesuada ac, dui.
Mauris nibh leo, facilisis non, adipiscing quis, ultrices a, dui.
\begin{proof}
Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text.
\end{proof}
Nam dui ligula, fringilla a, euismod sodales, sollicitudin vel, wisi. Morbi auctor lorem non justo. Nam lacus libero,
pretium at, lobortis vitae, ultricies et, tellus. Donec aliquet, tortor sed accumsan bibendum, erat ligula aliquet magna,
vitae ornare odio metus a mi. Morbi ac orci et nisl hendrerit mollis. Suspendisse ut massa. Cras nec ante. Pellentesque
a nulla. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Aliquam tincidunt
urna. Nulla ullamcorper vestibulum turpis. Pellentesque cursus luctus mauris.
Nulla malesuada porttitor diam. Donec felis erat, congue non, volutpat at, tincidunt tristique, libero. Vivamus
viverra fermentum felis. Donec nonummy pellentesque ante. Phasellus adipiscing semper elit. Proin fermentum massa
ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. Sed lacinia
nulla vitae enim. Pellentesque tincidunt purus vel magna. Integer non enim. Praesent euismod nunc eu purus. Donec
bibendum quam in tellus. Nullam cursus pulvinar lectus. Donec et mi. Nam vulputate metus eu enim. Vestibulum
pellentesque felis eu massa.
\begin{proof}[Proof of Theorem~\ref{thm1}]
Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text.
\end{proof}
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra tortor. Sed fringilla
justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue est scelerisque
sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas urna mi, suscipit
in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Donec odio elit,
dictum in, hendrerit sit amet, egestas sed, leo. Praesent feugiat sapien aliquet odio. Integer vitae justo. Aliquam
vestibulum fringilla lorem. Sed neque lectus, consectetuer at, consectetuer sed, eleifend ac, lectus. Nulla facilisi.
Pellentesque eget lectus. Proin eu metus. Sed porttitor. In hac habitasse platea dictumst. Suspendisse eu lectus. Ut Curabitur tellus magna, porttitor a, commodo a, commodo in, tortor. Donec interdum. Praesent scelerisque. Mae-
cenas posuere sodales odio. Vivamus metus lacus, varius quis, imperdiet quis, rhoncus a, turpis. Etiam ligula arcu,
elementum a, venenatis quis, sollicitudin sed, metus. Donec nunc pede, tincidunt in, venenatis vitae, faucibus vel,
\begin{sidewaystable
\caption{Sideways table caption. For decimal alignment refer column 4 to 9 in tabular* preamble.\label{tab3}}%
\begin{tabular*}{\textheight}{@{\extracolsep\fill}lccD{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}D{.}{.}{4}@{\extracolsep\fill}}%
\toprule
& \textbf{col2 head} & \textbf{col3 head} & \multicolumn{1}{c}{\textbf{10}} &\multicolumn{1}{c}{\textbf{20}} &\multicolumn{1}{c}{\textbf{30}} &\multicolumn{1}{c}{\textbf{10}} &\multicolumn{1}{c}{\textbf{20}} &\multicolumn{1}{c}{\textbf{30}} \\
\midrule
&col2 text &col3 text &0.7568&1.0530&1.2642&0.9919&1.3541&1.6108 \\
& &col2 text &12.5701 &19.6603&25.6809&18.0689&28.4865&37.3011 \\
3 &col2 text & col3 text &0.7426&1.0393&1.2507&0.9095&1.2524&1.4958 \\
& &col3 text &12.8008&19.9620&26.0324&16.6347&26.0843&34.0765 \\
& col2 text & col3 text &0.7285&1.0257&1.2374&0.8195&1.1407&1.3691\tnote{*} \\
& & col3 text &13.0360&20.2690&26.3895&15.0812&23.4932&30.6060\tnote{\dagger} \\
\bottomrule
\end{tabular*}
\begin{tablenotes
\item[*] First sideways table footnote. Sideways table footnote. Sideways table footnote. Sideways table footnote.
\item[$\dagger$] Second sideways table footnote. Sideways table footnote. Sideways table footnote. Sideways table footnote.
\end{tablenotes}
\end{sidewaystable}
\begin{sidewaysfigure}
\centerline{\includegraphics[width=542pt,height=9pc,draft]{blankfig}}
\caption{Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption. Sideways figure caption.\label{fig3}}
\end{sidewaysfigure}
nibh. Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
\begin{algorithm}
\caption{Pseudocode for our algorithm}\label{alg1}
\begin{algorithmic}
\For each frame
\For water particles $f_{i}$
\State compute fluid flow~\cite{Rothermel1997}
\State compute fluid--solid interaction~\cite{Allen2011}
\State apply adhesion and surface tension~\cite{Ballen2011}
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute virtual water film \\(see Section~\ref{sec3})
\EndFor
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute growth direction vector \\(see Section~\ref{sec2})
\EndFor
\EndFor
\For solid particles $s_{i}$
\For neighboring water particles $f_{j}$
\State compute $F_{\theta}$ (see Section~\ref{sec1})
\State compute $CE(s_{i},f_{j})$ \\(see Section~\ref{sec3})
\If $CE(b_{i}, f_{j})$ $>$ glaze threshold
\State $j$th water particle's phase $\Leftarrow$ ICE
\EndIf
\If $CE(c_{i}, f_{j})$ $>$ icicle threshold
\State $j$th water particle's phase $\Leftarrow$ ICE
\EndIf
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa~\cite{Rothermel1997,Elbaum2002}. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo~\cite{Elbaum2002} Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
Pellentesque wisi. Nullam malesuada. Morbi ut tellus ut pede tincidunt porta. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Etiam congue neque id dolor.
Donec et nisl at wisi luctus bibendum. Nam interdum tellus ac libero. Sed sem justo, laoreet vitae, fringilla at,
adipiscing ut, nibh. Maecenas non sem quis tortor eleifend fermentum. Etiam id tortor ac mauris porta vulputate.
Integer porta neque vitae massa. Maecenas tempus libero a libero posuere dictum. Vestibulum ante ipsum primis in
faucibus orci luctus et ultrices posuere cubilia Curae; Aenean quis mauris sed elit commodo placerat. Class aptent
taciti sociosqu ad litora torquent per conubia nostra, per inceptos hymenaeos. Vivamus rhoncus tincidunt libero.
Etiam elementum pretium justo. Vivamus est. Morbi a tellus eget pede tristique commodo. Nulla nisl. Vestibulum
sed nisl eu sapien cursus rutrum.
\begin{equation}\label{eq23}
\|\tilde{X}(k)\|^2
=\frac{\left\|\sum\limits_{i=1}^{p}\tilde{Y}_i(k)+\sum\limits_{j=1}^{q}\tilde{Z}_j(k) \right\|^2}{(p+q)^2}
\leq\frac{\sum\limits_{i=1}^{p}\left\|\tilde{Y}_i(k)\right\|^2+\sum\limits_{j=1}^{q}\left\|\tilde{Z}_j(k)\right\|^2 }{p+q}.
\end{equation}
Sed feugiat. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Ut pellentesque
augue sed urna. Vestibulum diam eros, fringilla et, consectetuer eu, nonummy id, sapien. Nullam at lectus. In sagittis
ultrices mauris. Curabitur malesuada erat sit amet massa. Fusce blandit. Aliquam erat volutpat. Aliquam euismod.
Aenean vel lectus. Nunc imperdiet justo nec dolor.
Etiam euismod. Fusce facilisis lacinia dui. Suspendisse potenti. In mi erat, cursus id, nonummy sed, ullamcorper
eget, sapien. Praesent pretium, magna in eleifend egestas, pede pede pretium lorem, quis consectetuer tortor sapien
facilisis magna. Mauris quis magna varius nulla scelerisque imperdiet. Aliquam non quam. Aliquam porttitor quam
a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
\begin{equation}\label{eq24}
\|\tilde{X}(k)\|^2
=\frac{\left\|\sum\limits_{i=1}^{p}\tilde{Y}_i(k)+\sum\limits_{j=1}^{q}\tilde{Z}_j(k) \right\|^2}{(p+q)^2}
\leq\frac{\sum\limits_{i=1}^{p}\left\|\tilde{Y}_i(k)\right\|^2+\sum\limits_{j=1}^{q}\left\|\tilde{Z}_j(k)\right\|^2 }{p+q}.
\end{equation}
Aliquam lectus. Vivamus leo. Quisque ornare tellus ullamcorper nulla. Mauris porttitor pharetra
tortor. Sed fringilla justo sed mauris. Mauris tellus. Sed non leo. Nullam elementum, magna in cursus sodales, augue
est scelerisque sapien, venenatis congue nulla arcu et pede. Ut suscipit enim vel sapien. Donec congue. Maecenas
urna mi, suscipit in, placerat ut, vestibulum ut, massa. Fusce ultrices nulla et nisl.
Etiam ac leo a risus tristique nonummy. Donec dignissim tincidunt nulla. Vestibulum rhoncus molestie odio. Sed
lobortis, justo et pretium lobortis, mauris turpis condimentum augue, nec ultricies nibh arcu pretium enim. Nunc
purus neque, placerat id, imperdiet sed, pellentesque nec, nisl. Vestibulum imperdiet neque non sem accumsan laoreet.
In hac habitasse platea dictumst. Etiam condimentum facilisis libero. Suspendisse in elit quis nisl aliquam dapibus.
Pellentesque auctor sapien. Sed egestas sapien nec lectus. Pellentesque vel dui vel neque bibendum viverra. Aliquam
porttitor nisl nec pede. Proin mattis libero vel turpis. Donec rutrum mauris et libero. Proin euismod porta felis.
Nam lobortis, metus quis elementum commodo, nunc lectus elementum mauris, eget vulputate ligula tellus eu neque.
Vivamus eu dolor.
\section{Conclusions}\label{sec5}
Lorem ipsum dolor sit amet, consectetuer adipiscing elit. Ut purus elit, vestibulum ut, placerat ac, adipiscing vitae,
felis. Curabitur dictum gravida mauris. Nam arcu libero, nonummy eget, consectetuer id, vulputate a, magna. Donec
vehicula augue eu neque. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas.
Mauris ut leo. Cras viverra metus rhoncus sem. Nulla et lectus vestibulum urna fringilla ultrices. Phasellus eu tellus
sit amet tortor gravida placerat. Integer sapien est, iaculis in, pretium quis, viverra ac, nunc. Praesent eget sem vel
leo ultrices bibendum. Aenean faucibus. Morbi dolor nulla, malesuada eu, pulvinar at, mollis ac, nulla. Curabitur
auctor semper nulla. Donec varius orci eget risus. Duis nibh mi, congue eu, accumsan eleifend, sagittis quis, diam.
Duis eget orci sit amet orci dignissim rutrum.
Nam dui ligula, fringilla a, euismod sodales, sollicitudin vel, wisi. Morbi auctor lorem non justo. Nam lacus libero,
pretium at, lobortis vitae, ultricies et, tellus. Donec aliquet, tortor sed accumsan bibendum, erat ligula aliquet magna,
vitae ornare odio metus a mi. Morbi ac orci et nisl hendrerit mollis. Suspendisse ut massa. Cras nec ante. Pellentesque
a nulla. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Aliquam tincidunt
urna. Nulla ullamcorper vestibulum turpis. Pellentesque cursus luctus mauris.
\section*{Acknowledgments}
This is acknowledgment text~\cite{Elbaum2002}. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here. This is acknowledgment text. Provide text here.
\subsection*{Author contributions}
This is an author contribution text. This is an author contribution text. This is an author contribution text. This is an author contribution text. This is an author contribution text.
\subsection*{Financial disclosure}
None reported.
\subsection*{Conflict of interest}
The authors declare no potential conflict of interests.
\section*{Supporting information}
The following supporting information is available as part of the online article:
\noindent
\textbf{Figure S1.}
{500{\uns}hPa geopotential anomalies for GC2C calculated against the ERA Interim reanalysis. The period is 1989--2008.}
\noindent
\textbf{Figure S2.}
{The SST anomalies for GC2C calculated against the observations (OIsst).}
\section{Introduction}\label{sec1}
\footnotetext{\textbf{Abbreviations:} SSAs, Selective solar absorbers}
Solar energy is abundant, striking our earth at a rate of 90,000 TW, which is 5,000 times of our current global power consumption \cite{abbott2010keeping}. The conversion of solar energy is considered as a promising approach to address the energy and environmental crisis. To date, various solar energy conversion technologies have been developed including photovoltaics\cite{polman2012photonic,lin2012small,schmidt2001self}, photo-thermal methods\cite{tian2013review,garg2012solar,tritt2008thermoelectrics}, and artificial photosynthesis\cite{meyer1989chemical,imahori2003nanostructured,gust2009solar}. Among those technologies, solar thermal power systems, such as concentrated solar power (CSP) via Rankine cycle \cite{mills2004advances}, solar thermoelectric generators (STEGs) \cite{kraemer2011high}, and solar thermophotovoltaics (STPVs) \cite{wang2011novel,tian2018tunable,tian2019near} have attracted increasing attention recently for various environmentally sustainable applications in industrial heating\cite{kaempener2015solar}, air conditioning\cite{lim2017heat}, and electricity generation\cite{kraemer2011high,kraemer2016concentrating}. As a crucial element, selective solar absorbers (SSAs), which are capable of converting solar radiation into heat, have a considerable influence on the overall performance of solar thermal systems. However, it is challenging to achieve high thermal resistance of the SSAs, which operates under a sever high-temperature working environment. Therefore, it is crucial to boost the performance of solar thermal systems by enhancing the sunlight absorption capability by depositing the infrared anti-reflection layer \cite{barshilia2012nanometric,li2018broadband,cao2017high} and improved the resistance under high-temperature environment \cite{wang2018spectrally,wang2018design}. Furthermore, SSAs are required to be omnidirectional and polarization-insensitive in the solar irradiance regime due to the randomly distributed solar irradiation \cite{wang2015highly}. Additionally, the cut-off wavelength of a selective absorber, at which its absorptance spectrum changes steeply, is highly dependent on its operational temperature because of the shifting nature of blackbody radiation according to Wien's displacement law \cite{siegel2001thermal}. However, existing technologies are either relied on high-cost nanofabrications methods or suffer from poor thermal resistance in the ambient environment \cite{cao2014review} and the omnidirectional and polarization insensitivity are scarcely achieved. Hence, SSAs with high-temperature ambient resistance and omnidirectional and polarization insensitivity are highly demanded in solar thermal applications.
Recently, advances in the small-scale fabrications have enabled manipulation of sunlight trapping in micro/nanostructures. Excitation of plasmonic resonance can produce selective and tunable absorption peaks at different wavelengths \cite{zhao2017design}. Simultaneously, the alternation from high UV/visible absorptance to low mid-infrared emittance is sharp as compared with natural materials based SSAs. Abundant metamaterial based selective absorbers have been investigated, such as 1-D or 2-D surface gratings \cite{ghanekar2018optimal,wang2013perfect,lee2014wavelength,han2016broadband,wang2017broadband,marques2002left,aydin2007capacitor,shchegolkov2010perfect,liu2007plasmon,tao2008metamaterial}, nanoparticles embedded dielectrics \cite{tian2019near,tian2018tunable,ghanekar2017mie,ghanekar2016novel}, crossbar or nano-disk arrays \cite{chen2012dual,Nielsen2012Efficient}, and photonic crystals \cite{stelmakh2013high,celanovic2008two,rinnerbauer2013large,li2015large,jiang2016refractory}. Nevertheless, these absorbers depend on time-consuming nanofabrication technologies, such as photolithography and electron beam lithography, which makes them unrealistic to be scalable-manufactured. Moreover, the high temperature will yield permanent damage to SSAs' spectral selectivity.
Metal-insulator-metal (MIM) \cite{chirumamilla2016multilayer,langlais2014high,nuru2014heavy} based novel absorbers consist of two metallic layers with a thin dielectric layer sandwiched in between. The incident light is absorbed due to strong optical interaction at the resonant frequencies and converted to heat in lossy metallic nanostructures \cite{zhang2017ultra}. They are feasible to be scale-manufactured with vacuum deposition methods like sputtering\cite{teixeira2001spectrally,fan1977selective,zhang2004high,zhang2000recent}, evaporation \cite{zhang1992new,niklasson1983solar,craighead1981graded} or chemical vapor deposition \cite{gogova1997optical,seraphin1976chemical}. Though high wavelength selectivity has been demonstrated on the multilayer-based SSAs structures, their high-temperature resistance is far from satisfying, mainly due to the lack of materials optimization. The refractory metals or metal oxides, including tungsten (W) \cite{zhang2004high,zhang1997direct}, nickel (Ni) \cite{sathiaraj1990ni,craighead1977optical}, chromium (Cr) \cite{teixeira2001spectrally,fan1977selective}, silica (SiO$_2$) \cite{wang2011optical,esposito2009fabrication,farooq1998high}, alumina (Al$_2$O$_3$) \cite{barshilia2011structure,xinkang2008microstructure,cheng2013improvement,antonaia2010stability}, and chromia (Cr$_2$O$_3$) \cite{nunes2003graded,yin2009direct} have been applied to manufacture SSAs. Specifically, W, Ni, and Cr thin film display high reflectivity in the mid-infrared wavelength region \cite{ghanekar2018optimal}, while maintaining the high absorptance in UV, visible, and the near-infrared regime when they are in the form of nanoparticles \cite{ghanekar2017mie,ghanekar2016novel}. On the other hand, oxides such as SiO$_2$, Al$_2$O$_3$, and Cr$_2$O$_3$ have high-temperature resistance even in the ambient environment. Here, we report a MIM based (W-SiO$_2$-W) SSAs that show high spectral selectivity (solar absorptance/thermal emittance = 13.9) and high thermal efficiency (80.03\% without optical concentration, compared to the state of the art 78\% \cite{wang2015highly}.). The spectral selectivity of fabricated SSAs is angular independent (\textasciitilde75$^{\circ}$, compared to the state of the art \textasciitilde60$^{\circ}$\cite{wang2018design}) and polarization insensitive (0$^{\circ}$ or 90$^{\circ}$.). The optical performance remains 93.6\% and 94.1\% after 1-hour thermal annealing up to 500$^{\circ}$C and 96-hour thermal treatment cycle at 400$^{\circ}$C, respectively. A high stagnation temperature of 193.5$^{\circ}$C was achieved under unconcentrated solar irradiance, which indicates its wide feasibility of low-temperature solar thermal applications without solar concentrators. The sandpaper abrasion robustness tests are also conducted in this work to demonstrate the mechanical abrasion resistance of SSAs' surface.
\section{Results and Discussions} \label{sec2}
\subsection{Energy conversion efficiency analysis of the SSAs}
\begin{figure*}[ht]
\centering
\includegraphics[width=1.0\textwidth]{sem_variangle}
\caption{\label{fig:sem_variangle} (\textbf{A}) Solar spectral irradiance (AM 1.5, global tilt), radiative heat flux of blackbody thermal radiation at various temperature, and reflectivity spectra of ideal SSAs and black surfaces. (\textbf{B}) and (\textbf{C}) The 3-D plot of solar to heat energy conversion efficiency for SSAs' contour plotted against concentration factor and cut-off wavelength at different operational temperature of 500$^{\circ}$C and 800$^{\circ}$C, respectively. (\textbf{D}) 3D schematic of proposed multilayer stack consisting of W, Al$_2$O$_3$, and SiO$_2$, the incidence angle, $\theta$, is defined as the angle between solar incident radiation and the surface normal. (\textbf{E}) A photo of sample fabrication on a 2 inch silicon wafer compared with a quarter coin and black painted paper. (\textbf{F}) A cross section SEM micrograph of the fabricated sample, the 2D schematic shows the thickness of each layer for the multilayer stack. (\textbf{F}) Normalized spectral distribution for radiative heat flux of solar (ASTM G173 AM 1.5) and blackbody thermal radiation (400$^{\circ}$C), as well as simulated and measured reflectivity spectrum of multilayer solar absorber. (\textbf{H}) The SSAs' high absorptance $\epsilon_{solar}$ $(\theta)$ and (\textbf{I}) $R_{IR}$ ($\theta$) across various angle of incident (AOI) result in excellent hemispherical $\epsilon_{solar}$ and $R_{IR}$. (\textbf{J}) The SSAs' high absorptance $\epsilon_{solar}$ $(\theta)$ and (\textbf{K}) $R_{IR}$ ($\theta$) across different angle of polarizations.
}
\end{figure*}
According to the Planck's formula of blackbody radiation intensity distribution, we know the solar intensity and the blackbody intensity for mid temperatures (400$^{\circ}$C\textasciitilde700$^{\circ}$C) are distributed over different wavelength regimes (Fig. \ref{fig:sem_variangle}(\textbf{A})). The spectral irradiance of a 400$^{\circ}$C blackbody is at the same order of solar heat flux, while it will be twice as the solar intensity when the temperature of blackbody increases to 500$^{\circ}$C. Meanwhile, the expanded regime of blackbody radiation moves to a shorter wavelength as its temperature increases, as illustrated by peaks of the blackbody irradiance curves from 100$^{\circ}$C to 500$^{\circ}$C (Fig. \ref{fig:sem_variangle}(\textbf{A})). An ideal spectral selective absorber has zero reflectance at solar radiation regime and unity at the blackbody thermal radiation region, with a sharp transition between these two regions. The cut-off wavelength, $\lambda_{cut-off}$, at which the transition happens, is where the blackbody radiation begins to exceed the solar intensity. The ideal transition wavelength will shift to shorter wavelengths, called blue-shift, for higher-temperature SSAs. Consequently, it is a trade-off to design perfect SSAs at different engineering applications. The efficiency of an SSAs, $\eta_{abs}$, can be simplified as: $ \eta_{\mathrm{abs}}=\alpha_{\mathrm{abs}}-\epsilon_{\mathrm{abs}} (\sigma\left(T_{\mathrm{abs}}^{4}-T_{\mathrm{amb}}^{4}\right))/(C F \cdot Q_{\mathrm{abs}})$. Figure \ref{fig:sem_variangle}(\textbf{B}) and (\textbf{C}) are the 3D contour plots which illustrate the energy conversion efficiency, $\eta_{abs}$, as a function of the cut-off wavelength, $\lambda_{abs,cut}$ and concentration factors, CF, at different operational temperature, $T_{abs}$, of 500$^{\circ}$C, and 800$^{\circ}$C, in which the red color represents higher efficiency, and the blue means lower efficiency (plots by analyzing formulae section 1 in supplementary materials). The efficiency curves share the same trend of increasing to maximum from 0 and then decrease as $\lambda_{cut-off}$ sweeps from 0.25 $\mu$m to 4.0 $\mu$m (Fig. \ref{fig:sem_variangle}(\textbf{B}) and (\textbf{C})). The maximum efficiency of the 500$^{\circ}$C SSAs (Fig. \ref{fig:sem_variangle}(\textbf{B})), increases with the concentration factors varying from 1 to 100 at a fixed cut-off wavelength, which can be seen that the red color is getting deeper. The efficiency, $\eta_{abs}$, will increase when the concentration factor, CF, increases since the CF is in the denominator. Additionally, the corresponding cut-off wavelength will shift to a longer wavelength with the increasing of concentration factors, since the red-shift ensures that the SSAs absorb more solar energy. By comparing Fig. \ref{fig:sem_variangle}(\textbf{B}) and \ref{fig:sem_variangle}(\textbf{C}), it shows that the maximum energy efficiency of the same corresponding concentration factor drops with the increasing of operational temperature (black points in Fig. \ref{fig:sem_variangle}(\textbf{B}) and \ref{fig:sem_variangle}(\textbf{C})). It is reasonable that the energy efficiency, $\eta_{abs}$, decreases with the increasing of absorber operational temperature, $T_{abs}$, because $T_{abs}$ is in the numerator.
\subsection{Characterizations of sample radiative properties}
Figure \ref{fig:sem_variangle}(\textbf{D}) illustrates the 3D schematic of a proposed multilayered stack consisting of five layers. Figure \ref{fig:sem_variangle}(\textbf{E}) manifests that the fabricated SSA is black at various angles with visual observation, which elucidates its high absorptance in the visible wavelength. The hemispherical reflectivity which exhibits spectral selectivity with solar reflectivity $R_{abs}$ $<$ 0.06 in the solar spectrum and thermal reflectively $R_{abs}$ $>$ 0.94 in the mid-infrared region.
The thickness of different layers for the fabricated sample is confirmed by the cross-section view of FE-SEM shown in Fig. \ref{fig:sem_variangle}(\textbf{F}). The bottom W layer is thick enough to block all the incident light and the solar absorptance of the absorber, $\alpha$ = 1 $-$ $R_{abs}$. Obviously, the solar irradiance randomly distributed in a one-day cycle, it is consequential to make the SSA facing to the sun all the time with the help of a dual-axis solar tracker system. The fabricated SSA has angular-independent hemispherical $\epsilon_{solar} (\theta)$ $>$ 0.90 (from 0.3$\mu$m to 2.5$\mu$m) (Fig. \ref{fig:sem_variangle}(\textbf{H})) and $R_{IR} (\theta)$ $>$ 0.93 (from 2.5$\mu$m to 20$\mu$m) (Fig. \ref{fig:sem_variangle}(\textbf{I}) (AOI, from 6$^{\circ}$ to 75$^{\circ}$), and it enables the SSAs latitude-insensitive and to work all day without solar-track systems.
In contrast, other SSAs report in literature often have directionally orientated structure, which enhances $\epsilon_{solar}, (0^{\circ})$, but reduces $\epsilon_{solar} (> 60^{\circ})$\cite{zhu2009optical,pettit1978solar}. As shown in Fig.\ref{fig:sem_variangle}({\textbf{H}}), for $\theta$ > 60$^{\circ}$, the SSAs have a higher $\epsilon_{solar}(\theta)$ than other SSAs. Besides, the hemispherical $\epsilon_{solar} (\theta)$ $>$ 0.89 (Fig.\ref{fig:sem_variangle}(\textbf{J}); AOI, 6$^{\circ}$) and $R_{IR} (\theta)$ $>$ 0.95 (Fig.\ref{fig:sem_variangle}(\textbf{K}); AOI, 12$^{\circ}$) for various polarization angles (from 6$^{\circ}$ to 75$^{\circ}$). Here, Figure \ref{fig:TE_TM} shows both the 3D contour plot of reflectivity spectra for the designed SSAs as functions of incident angle $\theta$, and wavelength $\lambda$, transverse electric (TE), transverse magnetic (TM), and unpolarized waves. The simulated reflectivity remains low within visible and near-infrared and high reflectivity ($>$ 0.95) in the mid-infrared region (from 3 $\mu$m to 15 $\mu$m) for unpolarized waves (Fig. \ref{fig:TE_TM}(\textbf{C})). Table S1 in supplementary materials lists the reflectivity at 0.55 $\mu$m, which the irradiance peak of solar lies, incident light at various angles. For both TE and TM waves, the reflectivity of the designed absorber is low from 0$^{\circ}$ to 75$^{\circ}$. However, the measured reflectivity of SSAs at 60$^{\circ}$ and 75$^{\circ}$ is higher than the simulated results, this is due to the fabrication process: the chemical bond (Si-O and Al-O) of SiO$_{2}$ and Al$_{2}$O$_{3}$ can be broken when collided with the charged energetic ions beams, resulting in the change of atomic ratio of Si/O and Al/O. Note that, the reflectivity spectra show two dips around $\lambda$ = 8.0 (7.8) $\mu$m and $\lambda$ = 10.8 (10.5) $\mu$m (Fig. \ref{fig:TE_TM}(\textbf{B}) and \ref{fig:TE_TM}(\textbf{C})), which are due to the surface phonon resonance of SiO$_2$ and Al$_2$O$_3$, respectively, this is also confirmed by the specular reflectivity measurements. However, the irradiance intensity peak of 500$^{\circ}$C blackbody thermal radiation lies at 3.8 $\mu$m (Fig. \ref{fig:sem_variangle}(\textbf{A})), while the blackbody emission power of 500$^{\circ}$C blackbody radiation at $\lambda$ = 8 $\mu$m and $\lambda$ = 11 $\mu$m only equals to 30\% and 11\% of the peak value. For higher temperature applications ($> 500^{\circ}$C), the blue-shifts of blackbody radiation will make the proportion of thermal radiation at around $\lambda$ = 8 $\mu$m and $\lambda$ = 11 $\mu$m even smaller and these dips are prominent only for higher incident angles, so it can be concluded that these two dips in reflectivity do not affect the selectivity performance of the SSAs. The overall absorptance of the fabricated samples at solar irradiance wavelength regime is 92\%, while it shows a low emittance of 6.6\% at a thermal wavelength with the measured spectral reflectivity under unconcentrated sunlight by solving Eq. S2 and S3 (the integral interval of solar radiation is from 0.25 $\mu$m to 2.5 $\mu$m, and the integral interval of blackbody thermal radiation is in the range of 2.5 $\mu$m - 15 $\mu$m).
\begin{figure*}[ht]
\centering
\includegraphics[width=1.0\textwidth]{TE_TM}
\caption{ \label{fig:TE_TM} Simulated angle dependent TE polarization (\textbf{A}), TM polarization (\textbf{B}), and unpolarized (\textbf{C}) specular reflectivity spectra of SSAs' 3-D plot and projected 2-D contour plot as functions of wavelength and angle of incidence, $\theta$. Measured specular reflectivity spectra of SSAs across angles at TE polarization (\textbf{D}), TM polarization (\textbf{E}), and unpolarized light (\textbf{F}).
}
\end{figure*}
\subsection{Ambient thermal resistance test and thermal degradation mechanism from SEM topography characterizations}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.0\textwidth]{thermaltest}
\caption{\label{fig:thermaltest} Reflectivity spectra (\textbf{A}) and overall $\epsilon_{solar}$, $\epsilon_{IR}$, and $\eta_{abs}$ (\textbf{B}) of the SSAs measured by FTIR spectrometer after one-hour thermal treatment at different temperatures under ambient environment. Reflectivity spectra (\textbf{C}) and overall $\epsilon_{solar}$, $\epsilon_{IR}$, and $\eta_{abs}$ (\textbf{D}) of the SSAs at 400$^{\circ}$C after long-time thermal treatment. SEM topographic image of the fabricated SSAs sample as fabricated and after one-hour thermal annealing at various temperatures (\textbf{E}) and long-time thermal treatment (\textbf{F}). (The blackbody temperature is fixed to be 100$^{\circ}$ when calculate the $\eta_{abs}$) (\textbf{G}) Real-time measurement of $R_{IR}$ for SSAs at elevated temperature under high vacuum ($<$ 10$^{-3}$Pa).
}
\end{figure*}
To maintain high energy efficiency, the CSP system always works under the concentrated sunlight to obtain high operational temperature, so it is consequential to have a constant spectral performance for SSAs at high temperatures. For the sake of evaluating the radiative properties of SSAs after thermal annealing at elevated temperatures and long-time thermal resistance tests, we measure the reflectivity spectrum of the fabricated SSAs after each thermal test for one-hour thermal annealing (Fig. \ref{fig:thermaltest}(\textbf{A})) and long-time thermal cycle (Fig. \ref{fig:thermaltest}(\textbf{C})). Figure \ref{fig:thermaltest} (\textbf{B}) and (\textbf{D}) show the overall of $\epsilon_{solar}$, $R_{IR}$, and $\eta_{abs}$ of SSAs after elevated temperature and long-time thermal tests. The spectral reflectivity of the tested sample reduces from 25$^{\circ}$C to 200$^{\circ}$C (Fig. \ref{fig:thermaltest}(\textbf{A}) and (\textbf{B})), since the $\epsilon_{solar}$ decreases from 92.0\% to 88.5\% (95.3\% remain) and thermal emittance, $\epsilon_{IR}$ increase from 6.6\% to 12.53\% (89.8\% increase) and the $\eta_{abs}$ reduce by 3.31\%. Though the SEM topography of SSAs shows little change, the dielectric functions of SiO$_{2}$ and Al$_{2}$O$_{3}$ might change at elevated temperatures. The $\epsilon_{solar}$ and $\epsilon_{IR}$ keep changing from 200$^{\circ}$C to 500$^{\circ}$C, which results in a decrease in $\eta_{abs}$ from 75.7\% to 74.5\%. This is possibly caused by the SSAs' topography damage, as shown in Fig. \ref{fig:thermaltest}(\textbf{E}) for 400$^{\circ}$C and 500$^{\circ}$C. Additionally, the cut-off wavelength of SSAs shifts to shorter wavelengths to maintain a high efficiency at high temperatures, as discussed in Sec \ref{sec2}.1.
Figure \ref{fig:thermaltest}(\textbf{C}) elucidates the reflectivity spectra of the fabricated samples after long-time thermal resistance tests. The reflectivity spectrum of SSAs changes severely after 400$^{\circ}$C thermal resistance test over 12h and
$\epsilon_{solar}$ decreases from 92.0\% to 85.2\% (92.6\% remain) and thermal emittance $\epsilon_{IR}$ increases from 6.6\% to 12.5\% (89.8\% increase) and $\eta_{abs}$ only reduces by 3.39\%. After that, from 12h to 48h, $\epsilon_{solar}$ and $\epsilon_{IR}$ barely change (Fig.\ref{fig:thermaltest}(\textbf{D})). However, $\epsilon_{IR}$ increases from 48h to 72h and causes a drop of $\eta_{abs}$.
To understand the mechanism that causes the degradation above 400$^{\circ}$C and after 72h thermal treatment, the samples are characterized using FE-SEM (SIGMA VP). Figure \ref{fig:thermaltest}(\textbf{E}) shows the SEM topographic images of the SSAs surface before and after thermal annealing for one hour. From 25$^{\circ}$C to 400$^{\circ}$C, it is difficult to see any apparent changes. When the temperature keeps going up from 400$^{\circ}$C to 500$^{\circ}$C, small show up, which is also demonstrated in the reflectivity spectra in Fig. \ref{fig:thermaltest}(\textbf{A}), due to the thermal stress coming from the difference in thermal expansion coefficient between SiO$_2$ (0.55 $\times$ 10$^{-6}$ m/m$\cdot$K) and W (4.2 $\times$ 10$^{-6}$ m/m$\cdot$K) \cite{chirumamilla2016multilayer}. Figure \ref{fig:thermaltest}(\textbf{F}) shows the SEM topographic images of the SSAs surface before and after long-time thermal treatment. Little topographic change of SSAs can be seen from 12h to 48h, while small round cracks like burst bubbles appear after 72h and become bigger after 96h. It agrees with the reflectivity spectra change of Fig. \ref{fig:thermaltest}(\textbf{C}). It provides a guideline for improvement approaching a perfect SSAs by the utilization of other materials with a similar thermal expansion coefficient to avoid the thermal stress at high temperature and long-time heating.
Figure \ref{fig:thermaltest}(\textbf{G}) shows that the $R_{IR}$ of SSAs gradually drops heating under elevated temperatures. Each stepped drops from 100$^{\circ}$C to 400$^{\circ}$C is larger than that from 400$^{\circ}$C to 800$^{\circ}$C. It proves that most of the thermal stress between different layers of SSAs is released when heated from 100$^{\circ}$C to 400 $^{\circ}$C.
Combined with the SEM topographic image of Fig. \ref{fig:thermaltest}(\textbf{E}), we conclude that the formation of cracks contributes more than the development of cracks to the decrease of $R_{IR}$. Even heated up to 900$^{\circ}$C, the $R_{IR}$ of SSAs still remain 90.15\% compared with the value at ambient temperature, this indicates that the combinations of SiO$_{2}$ and Al$_{2}$O$_{3}$ has the potential applications under high-temperature solar thermal systems.
\begin{figure}[!ht]
\centering
\includegraphics[width=1.0\textwidth]{eff_mechanical_field}
\caption{ \label{fig:eff_mechanical_field} (\textbf{A}) Solar to heat conversion efficiency as a function of absorber operational temperature, $T_{abs}$, for an ideal selective absorber, SSAs with reflectivity spectrum of measured or simulated, and a black surface, under the unconcentrated solar light (1 sun); (\textbf{B}) Solar to heat conversion efficiency for abovementioned four surfaces as a function of concentration factor, CF, at the operational temperature of $T_{abs}$ = 400$^{\circ}$C. (\textbf{C}) Thermal performance of the selective absorber (orange curve) and the black surface (yellow curve) over a one-day solar cycle from sunrise (5:00 a.m.) to one hour after sunset (8:00 p.m.) at varying ambient temperature (blue curve) under 1 sun. 3D schematic (\textbf{D}) of the real (\textbf{E}) outdoor experimental set-up. (\textbf{F}) The temperature variations of the sample's surface, vacuum chamber, and outdoor ambient environment. $\epsilon_{solar}$ (\textbf{G}), $R_{IR}$ (\textbf{H}), and $\eta_{abs}$ (\textbf{I}) after the different abrasion tests under different weight (50g/100g) as a function of cycles.
}
\end{figure}
\subsection{Solar conversion efficiency investigations of fabricated multilayer SSAs}
As discussed above, the designed structure of SSAs is demonstrated to be angular-insensitive as well as thermal stable up to 400$^{\circ}$C. We employ the reflectivity measured by the UV/VIS/NIR spectrophotometer and FTIR spectrometer at room temperature (25$^{\circ}$C), as shown in Fig. \ref{fig:sem_variangle}(\textbf{G}). The $\alpha_{\lambda,abs}^{\prime}$ and $\epsilon_{\lambda,abs}^{\prime}$ are assumed to be independent of temperature as observed in Fig. \ref{fig:thermaltest}(\textbf{A}). The spectral integration for $\alpha_{\lambda,abs}^{\prime}$ and $\epsilon_{\lambda,abs}^{\prime}$ is performed over wavelengths from 0.3 $\mu$m to 16 $\mu$m, which covers 99.9\% of the solar radiation and only 4.7\% of thermal radiation energy outside this spectral region for a 400$^{\circ}$C blackbody. The reflectivity of the ideal absorber is zero below the cut-off wavelength, while its reflectivity is unity beyond the cut-off wavelength to minimize thermal leakage from blackbody radiation (Fig. \ref{fig:sem_variangle} (\textbf{A}), the dark-orange dot-dash line). The cut-off wavelength is optimized according to the shifts of blackbody thermal radiation at different operational temperatures to maximize the energy conversion efficiency, which defines an upper limit of system efficiency. The reflectivity of black surface, $R_{abs,black}$, is zero over the entire wavelength region ($\alpha$ = 1 -- $R_{abs,black}$ = 1) and shows no spectral selectivity (Fig. \ref{fig:sem_variangle}(\textbf{A}), the black long dash line).
The conversion efficiencies (by analyzing formulae section 1 in supplementary materials) of SSAs are 84.28\% and 86.32\% for the simulated and measured spectral reflectivity at $T_{abs}$ = 100$^{\circ}$C, respectively (Fig. \ref{fig:eff_mechanical_field}(\textbf{A})). The efficiency drops gradually to zero at the stagnation temperature of 396$^{\circ}$C and 486$^{\circ}$C using simulated and measured optical properties, respectively. The absorbed solar energy equals the blackbody re-emission energy (i.e., no solar thermal energy is actually harvested) at this point. The stagnation temperature of SSAs using the simulated reflectivity spectrum is higher than the one using the measured reflectivity spectrum since the reflectivity curve of the simulated one is steeper than the measured one over the transition wavelength region. The efficiency curves of SSAs with measured and simulated radiative properties intersect at 210$^{\circ}$C. The fabricated SSA has a higher efficiency than the designed SSA below 210$^{\circ}$C, because the measured spectrum has a higher reflectivity than the simulated one from 7.7 $\mu$m to 20 $\mu$m, (as marked by red solid pentagram in Fig. \ref{fig:sem_variangle}(\textbf{G})). Half of the blackbody thermal radiation spreads in this wavelength region (the thermal radiation of a 210$^{\circ}$C blackbody is within 3.0 $\mu$m to 15 $\mu$m). When the temperature goes above 210$^{\circ}$C, the efficiency of SSAs using simulated reflectivity spectrum exhibits an advantage than the measured one since the blackbody radiation blue-shifts to the shorter wavelength, where the simulated spectrum has a higher reflectivity than the measured.
As a reference, the black surface can only convert 34.8\% solar energy to heat at $T_{abs}$ = 100$^{\circ}$C, and its efficiency goes down to zero at 126$^{\circ}$C, which further demonstrates the significance of spectral selectivity in enhancing the solar to heat conversion efficiency. On the other hand, the efficiency of the multilayered absorber is 17\% and 20\% lower than the ideal surface at $T_{abs}$ = 100$^{\circ}$C with measured optical properties and simulated radiative spectrum data, respectively. This gap becomes large when the stagnation temperature increases. It mainly results from the larger thermal emittance (around 0.09) in the mid-infrared regime and the cut-off wavelength of the ideal surface is optimized at each temperature according to Wien's displacement law, while the cut-off wavelength of selective absorber keeps unchanged at around 1.2 $\mu$m. Simultaneously, the reflectivity spectrum of the ideal surface changes much sharply at the cut-off point than the multilayered SSAs, comparing Fig. \ref{fig:sem_variangle}(\textbf{A}) and Fig. \ref{fig:sem_variangle}(\textbf{G}). Therefore, the geometry parameters of the multilayered SSAs need to be optimized to make the cut-off wavelength perfectly matched the operational temperature.
Figure \ref{fig:eff_mechanical_field}(\textbf{B}) shows the efficiency as a function of concentration factors, CF, from 1 to 1,000, when $T_{abs}$ = 400$^{\circ}$C, aiming at a mid-temperature application. The cut-off wavelength of the ideal surface is optimized according to different concentration factors. This indicates an upper limit for the SSAs' performance under different concentration factors. The energy efficiency of SSAs with measured or simulated optical properties and the black surface keeps going up with the increasing of CF. The efficiency of SSAs with the measured reflectivity spectrum is lower than the simulated one because the simulated spectrum has a higher solar absorptance than the measured one. For the black surface, its energy conversion efficiency becomes greater than zero at around 12 suns and climbs close to the ideal surface when the CF = 1,000 since the solar radiation heat flux is much larger than the 400$^{\circ}$C blackbody thermal radiation under 1000 suns.
\subsection{Thermal performance investigations under unconcentrated solar radiation}
Figure \ref{fig:eff_mechanical_field}(\textbf{C}) shows the calculated transient temperature variations of SSAs and the black surface under one sun over a solar cycle in Boston, Massachusetts \cite{Weatherboston} using the ambient temperature \cite{Weatherboston} and the solar illumination data \cite{solarangleboston} of July 10, 2018 (see formulae section 2). The highest temperature of the SSAs is 273$^{\circ}$C, while the highest temperature of the black surface is 72$^{\circ}$C. The thermal performance of the selective solar absorber is overwhelmingly better than the black surface at any time under sunshine. The highest temperature difference between SSAs and the black surface is 311$^{\circ}$C under 20 suns (see supplementary Fig. S1), which is higher than the temperature difference under one sun (201$^{\circ}$C). We evaluate the outdoor test performance by measuring the stagnation temperature under unconcentrated solar radiation in Kingston, Rhode Island on May 27, 2019. A SSAs sample was placed in a 100 cm diameter vacuum chamber which was equipped with a Potassium bromide (KBr) window (70 cm in diameter) (Fig. \ref{fig:eff_mechanical_field}(\textbf{D}) and (\textbf{E})). The KBr window is highly transparent (\textgreater 93\%) from UV to the mid-infrared range, which allows sunlight to come in and the thermal radiation of the heated SSAs re-emit to the sky. The vacuum chamber was set on a mounting frame with an angle of $\sim$ 32$^{\circ}$ to ensure the absorber was normal to the incoming sunlight. A K-type thermocouple was secured to the back of the SSAs sample with Kapton tape. Another K-type couple was taped to the bottom of the radiation shield to measure the local temperature of the vacuum chamber. In order to eliminate the conductive and radiative heat losses, we supported the absorber sample with low thermal conductivity (29 mW/m$\cdot$K) aerogel tiles that were on top of the radiation shield made of several double-side polished aluminum sheets. The vacuum chamber was pumped continuously to be $\sim$ 5 $\times$ 10$^{-3}$ Pa with the turbomolecular pump to eliminate the convective and conductive thermal losses from the air. The warmed up to the maximum stagnation temperature of 193.5$^{\circ}$C at around 11:30 since the weather was partly cloudy, the temperature of the vacuum chamber and the absorber sample showed strong fluctuations during the test and was lower than the calculated maximum temperature. The stagnation temperature is enough for the rooftop solar thermal systems even under partly cloudy weather.
\subsection{SSAs’ surface abrasion robustness tests}
Figure \ref{fig:eff_mechanical_field}(\textbf{G}, \textbf{H}, and \textbf{I}) show $\epsilon_{solar}$ (\textbf{G}), $R_{IR}$ (\textbf{H}), and $\eta_{abs}$ (\textbf{I}) of SSAs after each abrasion cycle (the definition of one abrasion cycle is in supplementary Fig. S2 ). It can be found that the $\epsilon_{solar}$, $R_{IR}$, and $\eta_{abs}$ remain 94.6\%, 83.5\%, and 86.9\% after 50 abrasion cycles under 100g weight (2040 Pa). The $R_{IR}$ of SSAs under 50g weight (1020 Pa) first drops and then increases, since the pressure of 50g weight can remove part of the top deposited SiO$_2$ and Al$_2$O$_3$ layer and make the bottom W layer exposed out in 5 cycles. While the 100g weight can make part of the bottom W layer exposed out only in 1 cycle. Since the abrasion resistance tests will break the fabricated SSAs layer, the $\epsilon_{solar}$ keeps decreasing whichever 50g or 100g weight is placed above. The $\eta_{abs}$ of 50g weight tests first decreases and goes up after 10 cycles, due to the $R_{IR}$ fluctuations of 50g weight tests.
\section{Discussion}
\label{Sec3}
The spectral selectivity and thermal performance of multilayered SSAs consisting of Al$_2$O$_3$-W-SiO$_2$-W stacks are investigated analytically and experimentally. It demonstrates 92\% solar absorptance and 6\% reflectivity in the blackbody thermal radiation region. Oblique reflectivity is also experimentally characterized to show high spectral selectivity for both TE and TM polarizations, which demonstrates its angular and polarization insensitivity. High-temperature thermal annealing performed at elevated temperatures for one hour shows the fabricated SSAs maintain its wavelength selectivity at even 400$^{\circ}$C and 500$^{\circ}$C. The high-temperature thermal resistance test is investigated under 400$^{\circ}$C for 12h, 24h, 48h, 72h, and 96h cycles and it demonstrates SSAs keep good selectivity within 48h, while slight degradation of spectral selectivity happens after the 72h and 96h thermal treatment. Real-time reflectivity measurements show that the material displays only a minor change in optical properties above 800$^{\circ}$C. Our study indicates that the proposed metamaterial exhibits thermal resistance and practically retains its optical proprieties even after 96h operation at 400$^{\circ}$C. Retention of optical properties after high temperature annealing can be attributed to pattern-free design of the material as cracks can be significantly more detrimental to optical properties of surface designs such as gratings. The outdoor tests show a peak stagnation temperature of 193.5$^{\circ}$C, and thus indicate its tremendous potential for low temperature solar thermal applications in the absence of solar concentrator. Additionally, the surface abrasion test yields that SSAs have a robust resistance to sandpaper abrasion attack for a long-duration practical application, Abrasion resistance can also be attributed to this pattern-free design.
\section{Materials and Methods}
\label{Sec4}
\subsection{Materials}
SiO$_2$ (99.995\%, 4N5), Al$_2$O$_3$(99.995\%, 4N5), and W (99.995\%, 3N5) are purchased from Angstrom Sciences, Inc. 2 inches wafers (Type: P; dopant: B; orientation: <1-0-0>; resistivity: 10-20 ohm/cm; thickness: 279$\pm$25 $\mu$m) are purchased from WaferPro.
\subsection{Methods}
\subsubsection{Sample fabrication and SEM topography characterizations}
SSAs samples are deposited with a magnetron sputtering technique within a home-built high-vacuum sputtering system (see supplementary Fig. S3). The fabrication process has been described in detail in this publication \cite{tian2019near}. The fabrication parameters, such as deposition rate, base pressure, and sputtering power for each deposition procedure are specified in Table S2 in supplementary materials. The base pressure before sputtering is 3.4 $\times$ 10$^{-7}$ Torr. The Al$_2$O$_3$ and W layer are both deposited directly from Al$_2$O$_3$ and W target, respectively, while the SiO$_2$ layer is oxidized from the silicon (Si) with radio frequency (RF) sputtering. The cross-section and topography of absorber samples after various temperature thermal annealing and diverse long-time thermal resistance tests are characterized by SIGMA VP Field Emission-Scanning Electron Microscope (FE-SEM).
\subsubsection{Optical properties measurements}
Hemispherical reflectance measurements in the UV, visible, and near-infrared regions (from 0.3 $\mu$m to 2.5 $\mu$m) are performed on a Jasco V770 spectrophotometer with a 60 mm PTFE based integrating sphere. A wavelength scan step is fixed as 1 nm with AOI of 6$^{\circ}$ and normalized to a labsphere spectralon reflectance standard. The mid-infrared (2.5 $\mu$m to 15 $\mu$m) reflectance measurements are executed on Jasco 6600 FTIR spectrometer together with a Pike 3-inches golden integrating sphere. The AOI of the incident beam from the FTIR spectrometer is fixed at 6$^{\circ}$ and the diffused golden reference is selected as a normalized standard. These spectra are taken at a scan rate of 64 with a wavelength resolution of 0.4 cm$^{-1}$. Details about FTIR spectrum measurements can be referred to the recent article \cite{tian2019near}.
The hemispherical reflectance measurement for different angles of incident (AOI) is conducted by using home-built wedge blocks (see supplementary Fig. S4) with (18$^{\circ}$, 30$^{\circ}$, 33$^{\circ}$, 33$^{\circ}$, 51$^{\circ}$, 66$^{\circ}$, and 76$^{\circ}$) combined with Jasco 60mm PTFE based integrating sphere and Pike 3-inches golden integrating sphere. The wedge blocks for PTFE integrating sphere are put faced to the sample ports of the Jasco and Pike integrating sphere. SSAs samples and the reference standard are put faced to the sample ports on designed wedge blocks. Background spectrum and the reflectance spectrum are taken as usual.
The specular reflectance measurements of variable angle in the UV, visible, and near-infrared regions (from 0.3 $\mu$m to 2.5 $\mu$m) are performed on Jasco V770 spectrophotometer with Harrick variable angle reflection accessory. The variable angle specular reflectance measurements (from 2.5 $\mu$m to 15 $\mu$m) are executed on Jasco 6600 FTIR spectrometer together with Harrick Seagull$^{TM}$ Variable Angle Reflection Accessory. Both measurements are taken for angles of 12$^{\circ}$, 30$^{\circ}$, 45$^{\circ}$, 60$^{\circ}$, and 75$^{\circ}$.
The hemispherical reflectance measurements of variable angle polarization from 0.3 $\mu$m to 2.5 $\mu$m are done on Jasco V770 spectrophotometer with GPH-506 polarizer and 60 mm integrating sphere. The hemispherical reflectance measurements of variable angle polarization from 2.5 $\mu$m to 15 $\mu$m are executed on Jasco 6600 FTIR spectrometer together with PL-82 polarizer and Pike 3-inches golden integrating sphere.
\subsubsection{High-temperature thermal annealing and resistance tests}
High-temperature thermal annealing and resistance tests are both carried out in a tube oven at an ambient atmosphere with an alumina tube of 5 cm diameter and 80 cm length. Samples are put in an alumina crucible boat (100 mm $\times$ 30 mm $\times$ 20 mm) that is placed in the center of the tube. The increasing rate of temperature is set to be at 15$^{\circ}$C/min for 100$^{\circ}$C 200$^{\circ}$C 300$^{\circ}$C 400$^{\circ}$C, and 500$^{\circ}$C for one hour. The hemispherical reflectance measurements are carried out after each thermal treatment. For the high-temperature thermal resistance tests, the temperature controller of the tube oven is set to be 400$^{\circ}$C for 12h, 24h, 48h, 72h, and 96h thermal cycles. The samples are taken out after each thermal cycle for reflectance spectrum measurements and a small piece of the sample ($\sim$ 3mm $\times$ 3mm) is cut for SEM topography characterization.
\subsubsection{Real-time reflectivity measurement of SSAs under high vacuum}
The real-time measurement of reflectivity for SSAs is conducted on Jasco FTIR 6600 spectrometer within the Harrick high-temperature reaction chamber. The chamber can be heated by an electrical heater up to 910
$^{\circ}$C and vacuumed down to < 10$^{-3}$ Pa connected to a Pfeiffer TMH 260 turbopump. The temperature of the SSAs is measured by a K-type thermocouple and controlled by the Waterloo PID temperature controller.
\subsubsection{SSAs’ surface abrasion robustness tests}
A sandpaper abrasion test was carried out using 400 grit SiC sandpaper as an abrasion surface, as shown in supplementary Fig. S2. The SSAs samples with a weight of 50/100 g above it are put face-down to sandpaper and move 10 cm along the ruler (see supplementary Fig. S2 B); then, the sample is then rotated counterclockwise by 90$^{\circ}$ (face to the sandpaper) and move 10 cm along the ruler. The three steps shown are in supplementary Fig. S2 B is defined as one abrasion cycle. This guarantees that the SSAs’ surface is abraded longitudinally and transversely in each cycle. The reflectivity spectra are measured after 1, 3, 5, 10, 20, 30, and 50 cycles.
\section{Acknowledgements} This project is supported in part by the National Science Foundation through grant numbers CBET-1941743, and National Aeronautics and Space Administration through grant number NNX15AK52A.
\section{Author Contributions}
Y.T. and X.L. develop the theoretical calculations and carry out spectrum measurements and field tests. L.Q. fabricates SSAs samples. A.G. and J.L. contribute to the interpretation of the results. All authors provide critical feedback and help revise the final version of the manuscript. T.T., G.X., and Y.Z. supervise the project.
\section{Conflicts of interest}
The authors declare no conflict of interest.
\section{Bibliography}
\section{Introduction}\label{sec1}
\footnotetext{\textbf{Abbreviations:} SSAs, Selective solar absorbers}
Solar energy is abundant, striking our earth at a rate of 90,000 TW, which is 5,000 times of our current global power consumption \cite{abbott2010keeping}. The conversion of solar energy is considered as a promising approach to address the energy and environmental crisis. To date, various solar energy conversion technologies have been developed including photovoltaics\cite{polman2012photonic,lin2012small,schmidt2001self}, photo-thermal methods\cite{tian2013review,garg2012solar,tritt2008thermoelectrics}, and artificial photosynthesis\cite{meyer1989chemical,imahori2003nanostructured,gust2009solar}. Among those technologies, solar thermal power systems, such as concentrated solar power (CSP) via Rankine cycle \cite{mills2004advances}, solar thermoelectric generators (STEGs) \cite{kraemer2011high}, and solar thermophotovoltaics (STPVs) \cite{wang2011novel,tian2018tunable,tian2019near} have attracted increasing attention recently for various environmentally sustainable applications in industrial heating\cite{kaempener2015solar}, air conditioning\cite{lim2017heat}, and electricity generation\cite{kraemer2011high,kraemer2016concentrating}. As a crucial element, selective solar absorbers (SSAs), which are capable of converting solar radiation into heat, have a considerable influence on the overall performance of solar thermal systems. However, it is challenging to achieve high thermal resistance of the SSAs, which operates under a sever high-temperature working environment. Therefore, it is crucial to boost the performance of solar thermal systems by enhancing the sunlight absorption capability by depositing the infrared anti-reflection layer \cite{barshilia2012nanometric,li2018broadband,cao2017high} and improved the resistance under high-temperature environment \cite{wang2018spectrally,wang2018design}. Furthermore, SSAs are required to be omnidirectional and polarization-insensitive in the solar irradiance regime due to the randomly distributed solar irradiation \cite{wang2015highly}. Additionally, the cut-off wavelength of a selective absorber, at which its absorptance spectrum changes steeply, is highly dependent on its operational temperature because of the shifting nature of blackbody radiation according to Wien's displacement law \cite{siegel2001thermal}. However, existing technologies are either relied on high-cost nanofabrications methods or suffer from poor thermal resistance in the ambient environment \cite{cao2014review} and the omnidirectional and polarization insensitivity are scarcely achieved. Hence, SSAs with high-temperature ambient resistance and omnidirectional and polarization insensitivity are highly demanded in solar thermal applications.
Recently, advances in the small-scale fabrications have enabled manipulation of sunlight trapping in micro/nanostructures. Excitation of plasmonic resonance can produce selective and tunable absorption peaks at different wavelengths \cite{zhao2017design}. Simultaneously, the alternation from high UV/visible absorptance to low mid-infrared emittance is sharp as compared with natural materials based SSAs. Abundant metamaterial based selective absorbers have been investigated, such as 1-D or 2-D surface gratings \cite{ghanekar2018optimal,wang2013perfect,lee2014wavelength,han2016broadband,wang2017broadband,marques2002left,aydin2007capacitor,shchegolkov2010perfect,liu2007plasmon,tao2008metamaterial}, nanoparticles embedded dielectrics \cite{tian2019near,tian2018tunable,ghanekar2017mie,ghanekar2016novel}, crossbar or nano-disk arrays \cite{chen2012dual,Nielsen2012Efficient}, and photonic crystals \cite{stelmakh2013high,celanovic2008two,rinnerbauer2013large,li2015large,jiang2016refractory}. Nevertheless, these absorbers depend on time-consuming nanofabrication technologies, such as photolithography and electron beam lithography, which makes them unrealistic to be scalable-manufactured. Moreover, the high temperature will yield permanent damage to SSAs' spectral selectivity.
Metal-insulator-metal (MIM) \cite{chirumamilla2016multilayer,langlais2014high,nuru2014heavy} based novel absorbers consist of two metallic layers with a thin dielectric layer sandwiched in between. The incident light is absorbed due to strong optical interaction at the resonant frequencies and converted to heat in lossy metallic nanostructures \cite{zhang2017ultra}. They are feasible to be scale-manufactured with vacuum deposition methods like sputtering\cite{teixeira2001spectrally,fan1977selective,zhang2004high,zhang2000recent}, evaporation \cite{zhang1992new,niklasson1983solar,craighead1981graded} or chemical vapor deposition \cite{gogova1997optical,seraphin1976chemical}. Though high wavelength selectivity has been demonstrated on the multilayer-based SSAs structures, their high-temperature resistance is far from satisfying, mainly due to the lack of materials optimization. The refractory metals or metal oxides, including tungsten (W) \cite{zhang2004high,zhang1997direct}, nickel (Ni) \cite{sathiaraj1990ni,craighead1977optical}, chromium (Cr) \cite{teixeira2001spectrally,fan1977selective}, silica (SiO$_2$) \cite{wang2011optical,esposito2009fabrication,farooq1998high}, alumina (Al$_2$O$_3$) \cite{barshilia2011structure,xinkang2008microstructure,cheng2013improvement,antonaia2010stability}, and chromia (Cr$_2$O$_3$) \cite{nunes2003graded,yin2009direct} have been applied to manufacture SSAs. Specifically, W, Ni, and Cr thin film display high reflectivity in the mid-infrared wavelength region \cite{ghanekar2018optimal}, while maintaining the high absorptance in UV, visible, and the near-infrared regime when they are in the form of nanoparticles \cite{ghanekar2017mie,ghanekar2016novel}. On the other hand, oxides such as SiO$_2$, Al$_2$O$_3$, and Cr$_2$O$_3$ have high-temperature resistance even in the ambient environment. Here, we report a MIM based (W-SiO$_2$-W) SSAs that show high spectral selectivity (solar absorptance/thermal emittance = 13.9) and high thermal efficiency (80.03\% without optical concentration, compared to the state of the art 78\% \cite{wang2015highly}.). The spectral selectivity of fabricated SSAs is angular independent (\textasciitilde75$^{\circ}$, compared to the state of the art \textasciitilde60$^{\circ}$\cite{wang2018design}) and polarization insensitive (0$^{\circ}$ or 90$^{\circ}$.). The optical performance remains 93.6\% and 94.1\% after 1-hour thermal annealing up to 500$^{\circ}$C and 96-hour thermal treatment cycle at 400$^{\circ}$C, respectively. A high stagnation temperature of 193.5$^{\circ}$C was achieved under unconcentrated solar irradiance, which indicates its wide feasibility of low-temperature solar thermal applications without solar concentrators. The sandpaper abrasion robustness tests are also conducted in this work to demonstrate the mechanical abrasion resistance of SSAs' surface.
\section{Results and Discussions} \label{sec2}
\subsection{Energy conversion efficiency analysis of the SSAs}
\begin{figure*}[ht]
\centering
\includegraphics[width=1.0\textwidth]{sem_variangle}
\caption{\label{fig:sem_variangle} (\textbf{A}) Solar spectral irradiance (AM 1.5, global tilt), radiative heat flux of blackbody thermal radiation at various temperature, and reflectivity spectra of ideal SSAs and black surfaces. (\textbf{B}) and (\textbf{C}) The 3-D plot of solar to heat energy conversion efficiency for SSAs' contour plotted against concentration factor and cut-off wavelength at different operational temperature of 500$^{\circ}$C and 800$^{\circ}$C, respectively. (\textbf{D}) 3D schematic of proposed multilayer stack consisting of W, Al$_2$O$_3$, and SiO$_2$, the incidence angle, $\theta$, is defined as the angle between solar incident radiation and the surface normal. (\textbf{E}) A photo of sample fabrication on a 2 inch silicon wafer compared with a quarter coin and black painted paper. (\textbf{F}) A cross section SEM micrograph of the fabricated sample, the 2D schematic shows the thickness of each layer for the multilayer stack. (\textbf{F}) Normalized spectral distribution for radiative heat flux of solar (ASTM G173 AM 1.5) and blackbody thermal radiation (400$^{\circ}$C), as well as simulated and measured reflectivity spectrum of multilayer solar absorber. (\textbf{H}) The SSAs' high absorptance $\epsilon_{solar}$ $(\theta)$ and (\textbf{I}) $R_{IR}$ ($\theta$) across various angle of incident (AOI) result in excellent hemispherical $\epsilon_{solar}$ and $R_{IR}$. (\textbf{J}) The SSAs' high absorptance $\epsilon_{solar}$ $(\theta)$ and (\textbf{K}) $R_{IR}$ ($\theta$) across different angle of polarizations.
}
\end{figure*}
According to the Planck's formula of blackbody radiation intensity distribution, we know the solar intensity and the blackbody intensity for mid temperatures (400$^{\circ}$C\textasciitilde700$^{\circ}$C) are distributed over different wavelength regimes (Fig. \ref{fig:sem_variangle}(\textbf{A})). The spectral irradiance of a 400$^{\circ}$C blackbody is at the same order of solar heat flux, while it will be twice as the solar intensity when the temperature of blackbody increases to 500$^{\circ}$C. Meanwhile, the expanded regime of blackbody radiation moves to a shorter wavelength as its temperature increases, as illustrated by peaks of the blackbody irradiance curves from 100$^{\circ}$C to 500$^{\circ}$C (Fig. \ref{fig:sem_variangle}(\textbf{A})). An ideal spectral selective absorber has zero reflectance at solar radiation regime and unity at the blackbody thermal radiation region, with a sharp transition between these two regions. The cut-off wavelength, $\lambda_{cut-off}$, at which the transition happens, is where the blackbody radiation begins to exceed the solar intensity. The ideal transition wavelength will shift to shorter wavelengths, called blue-shift, for higher-temperature SSAs. Consequently, it is a trade-off to design perfect SSAs at different engineering applications. The efficiency of an SSAs, $\eta_{abs}$, can be simplified as: $ \eta_{\mathrm{abs}}=\alpha_{\mathrm{abs}}-\epsilon_{\mathrm{abs}} (\sigma\left(T_{\mathrm{abs}}^{4}-T_{\mathrm{amb}}^{4}\right))/(C F \cdot Q_{\mathrm{abs}})$. Figure \ref{fig:sem_variangle}(\textbf{B}) and (\textbf{C}) are the 3D contour plots which illustrate the energy conversion efficiency, $\eta_{abs}$, as a function of the cut-off wavelength, $\lambda_{abs,cut}$ and concentration factors, CF, at different operational temperature, $T_{abs}$, of 500$^{\circ}$C, and 800$^{\circ}$C, in which the red color represents higher efficiency, and the blue means lower efficiency (plots by analyzing formulae section 1 in supplementary materials). The efficiency curves share the same trend of increasing to maximum from 0 and then decrease as $\lambda_{cut-off}$ sweeps from 0.25 $\mu$m to 4.0 $\mu$m (Fig. \ref{fig:sem_variangle}(\textbf{B}) and (\textbf{C})). The maximum efficiency of the 500$^{\circ}$C SSAs (Fig. \ref{fig:sem_variangle}(\textbf{B})), increases with the concentration factors varying from 1 to 100 at a fixed cut-off wavelength, which can be seen that the red color is getting deeper. The efficiency, $\eta_{abs}$, will increase when the concentration factor, CF, increases since the CF is in the denominator. Additionally, the corresponding cut-off wavelength will shift to a longer wavelength with the increasing of concentration factors, since the red-shift ensures that the SSAs absorb more solar energy. By comparing Fig. \ref{fig:sem_variangle}(\textbf{B}) and \ref{fig:sem_variangle}(\textbf{C}), it shows that the maximum energy efficiency of the same corresponding concentration factor drops with the increasing of operational temperature (black points in Fig. \ref{fig:sem_variangle}(\textbf{B}) and \ref{fig:sem_variangle}(\textbf{C})). It is reasonable that the energy efficiency, $\eta_{abs}$, decreases with the increasing of absorber operational temperature, $T_{abs}$, because $T_{abs}$ is in the numerator.
\subsection{Characterizations of sample radiative properties}
Figure \ref{fig:sem_variangle}(\textbf{D}) illustrates the 3D schematic of a proposed multilayered stack consisting of five layers. Figure \ref{fig:sem_variangle}(\textbf{E}) manifests that the fabricated SSA is black at various angles with visual observation, which elucidates its high absorptance in the visible wavelength. The hemispherical reflectivity which exhibits spectral selectivity with solar reflectivity $R_{abs}$ $<$ 0.06 in the solar spectrum and thermal reflectively $R_{abs}$ $>$ 0.94 in the mid-infrared region.
The thickness of different layers for the fabricated sample is confirmed by the cross-section view of FE-SEM shown in Fig. \ref{fig:sem_variangle}(\textbf{F}). The bottom W layer is thick enough to block all the incident light and the solar absorptance of the absorber, $\alpha$ = 1 $-$ $R_{abs}$. Obviously, the solar irradiance randomly distributed in a one-day cycle, it is consequential to make the SSA facing to the sun all the time with the help of a dual-axis solar tracker system. The fabricated SSA has angular-independent hemispherical $\epsilon_{solar} (\theta)$ $>$ 0.90 (from 0.3$\mu$m to 2.5$\mu$m) (Fig. \ref{fig:sem_variangle}(\textbf{H})) and $R_{IR} (\theta)$ $>$ 0.93 (from 2.5$\mu$m to 20$\mu$m) (Fig. \ref{fig:sem_variangle}(\textbf{I}) (AOI, from 6$^{\circ}$ to 75$^{\circ}$), and it enables the SSAs latitude-insensitive and to work all day without solar-track systems.
In contrast, other SSAs report in literature often have directionally orientated structure, which enhances $\epsilon_{solar}, (0^{\circ})$, but reduces $\epsilon_{solar} (> 60^{\circ})$\cite{zhu2009optical,pettit1978solar}. As shown in Fig.\ref{fig:sem_variangle}({\textbf{H}}), for $\theta$ > 60$^{\circ}$, the SSAs have a higher $\epsilon_{solar}(\theta)$ than other SSAs. Besides, the hemispherical $\epsilon_{solar} (\theta)$ $>$ 0.89 (Fig.\ref{fig:sem_variangle}(\textbf{J}); AOI, 6$^{\circ}$) and $R_{IR} (\theta)$ $>$ 0.95 (Fig.\ref{fig:sem_variangle}(\textbf{K}); AOI, 12$^{\circ}$) for various polarization angles (from 6$^{\circ}$ to 75$^{\circ}$). Here, Figure \ref{fig:TE_TM} shows both the 3D contour plot of reflectivity spectra for the designed SSAs as functions of incident angle $\theta$, and wavelength $\lambda$, transverse electric (TE), transverse magnetic (TM), and unpolarized waves. The simulated reflectivity remains low within visible and near-infrared and high reflectivity ($>$ 0.95) in the mid-infrared region (from 3 $\mu$m to 15 $\mu$m) for unpolarized waves (Fig. \ref{fig:TE_TM}(\textbf{C})). Table S1 in supplementary materials lists the reflectivity at 0.55 $\mu$m, which the irradiance peak of solar lies, incident light at various angles. For both TE and TM waves, the reflectivity of the designed absorber is low from 0$^{\circ}$ to 75$^{\circ}$. However, the measured reflectivity of SSAs at 60$^{\circ}$ and 75$^{\circ}$ is higher than the simulated results, this is due to the fabrication process: the chemical bond (Si-O and Al-O) of SiO$_{2}$ and Al$_{2}$O$_{3}$ can be broken when collided with the charged energetic ions beams, resulting in the change of atomic ratio of Si/O and Al/O. Note that, the reflectivity spectra show two dips around $\lambda$ = 8.0 (7.8) $\mu$m and $\lambda$ = 10.8 (10.5) $\mu$m (Fig. \ref{fig:TE_TM}(\textbf{B}) and \ref{fig:TE_TM}(\textbf{C})), which are due to the surface phonon resonance of SiO$_2$ and Al$_2$O$_3$, respectively, this is also confirmed by the specular reflectivity measurements. However, the irradiance intensity peak of 500$^{\circ}$C blackbody thermal radiation lies at 3.8 $\mu$m (Fig. \ref{fig:sem_variangle}(\textbf{A})), while the blackbody emission power of 500$^{\circ}$C blackbody radiation at $\lambda$ = 8 $\mu$m and $\lambda$ = 11 $\mu$m only equals to 30\% and 11\% of the peak value. For higher temperature applications ($> 500^{\circ}$C), the blue-shifts of blackbody radiation will make the proportion of thermal radiation at around $\lambda$ = 8 $\mu$m and $\lambda$ = 11 $\mu$m even smaller and these dips are prominent only for higher incident angles, so it can be concluded that these two dips in reflectivity do not affect the selectivity performance of the SSAs. The overall absorptance of the fabricated samples at solar irradiance wavelength regime is 92\%, while it shows a low emittance of 6.6\% at a thermal wavelength with the measured spectral reflectivity under unconcentrated sunlight by solving Eq. S2 and S3 (the integral interval of solar radiation is from 0.25 $\mu$m to 2.5 $\mu$m, and the integral interval of blackbody thermal radiation is in the range of 2.5 $\mu$m - 15 $\mu$m).
\begin{figure*}[ht]
\centering
\includegraphics[width=1.0\textwidth]{TE_TM}
\caption{ \label{fig:TE_TM} Simulated angle dependent TE polarization (\textbf{A}), TM polarization (\textbf{B}), and unpolarized (\textbf{C}) specular reflectivity spectra of SSAs' 3-D plot and projected 2-D contour plot as functions of wavelength and angle of incidence, $\theta$. Measured specular reflectivity spectra of SSAs across angles at TE polarization (\textbf{D}), TM polarization (\textbf{E}), and unpolarized light (\textbf{F}).
}
\end{figure*}
\subsection{Ambient thermal resistance test and thermal degradation mechanism from SEM topography characterizations}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.0\textwidth]{thermaltest}
\caption{\label{fig:thermaltest} Reflectivity spectra (\textbf{A}) and overall $\epsilon_{solar}$, $\epsilon_{IR}$, and $\eta_{abs}$ (\textbf{B}) of the SSAs measured by FTIR spectrometer after one-hour thermal treatment at different temperatures under ambient environment. Reflectivity spectra (\textbf{C}) and overall $\epsilon_{solar}$, $\epsilon_{IR}$, and $\eta_{abs}$ (\textbf{D}) of the SSAs at 400$^{\circ}$C after long-time thermal treatment. SEM topographic image of the fabricated SSAs sample as fabricated and after one-hour thermal annealing at various temperatures (\textbf{E}) and long-time thermal treatment (\textbf{F}). (The blackbody temperature is fixed to be 100$^{\circ}$ when calculate the $\eta_{abs}$) (\textbf{G}) Real-time measurement of $R_{IR}$ for SSAs at elevated temperature under high vacuum ($<$ 10$^{-3}$Pa).
}
\end{figure*}
To maintain high energy efficiency, the CSP system always works under the concentrated sunlight to obtain high operational temperature, so it is consequential to have a constant spectral performance for SSAs at high temperatures. For the sake of evaluating the radiative properties of SSAs after thermal annealing at elevated temperatures and long-time thermal resistance tests, we measure the reflectivity spectrum of the fabricated SSAs after each thermal test for one-hour thermal annealing (Fig. \ref{fig:thermaltest}(\textbf{A})) and long-time thermal cycle (Fig. \ref{fig:thermaltest}(\textbf{C})). Figure \ref{fig:thermaltest} (\textbf{B}) and (\textbf{D}) show the overall of $\epsilon_{solar}$, $R_{IR}$, and $\eta_{abs}$ of SSAs after elevated temperature and long-time thermal tests. The spectral reflectivity of the tested sample reduces from 25$^{\circ}$C to 200$^{\circ}$C (Fig. \ref{fig:thermaltest}(\textbf{A}) and (\textbf{B})), since the $\epsilon_{solar}$ decreases from 92.0\% to 88.5\% (95.3\% remain) and thermal emittance, $\epsilon_{IR}$ increase from 6.6\% to 12.53\% (89.8\% increase) and the $\eta_{abs}$ reduce by 3.31\%. Though the SEM topography of SSAs shows little change, the dielectric functions of SiO$_{2}$ and Al$_{2}$O$_{3}$ might change at elevated temperatures. The $\epsilon_{solar}$ and $\epsilon_{IR}$ keep changing from 200$^{\circ}$C to 500$^{\circ}$C, which results in a decrease in $\eta_{abs}$ from 75.7\% to 74.5\%. This is possibly caused by the SSAs' topography damage, as shown in Fig. \ref{fig:thermaltest}(\textbf{E}) for 400$^{\circ}$C and 500$^{\circ}$C. Additionally, the cut-off wavelength of SSAs shifts to shorter wavelengths to maintain a high efficiency at high temperatures, as discussed in Sec \ref{sec2}.1.
Figure \ref{fig:thermaltest}(\textbf{C}) elucidates the reflectivity spectra of the fabricated samples after long-time thermal resistance tests. The reflectivity spectrum of SSAs changes severely after 400$^{\circ}$C thermal resistance test over 12h and
$\epsilon_{solar}$ decreases from 92.0\% to 85.2\% (92.6\% remain) and thermal emittance $\epsilon_{IR}$ increases from 6.6\% to 12.5\% (89.8\% increase) and $\eta_{abs}$ only reduces by 3.39\%. After that, from 12h to 48h, $\epsilon_{solar}$ and $\epsilon_{IR}$ barely change (Fig.\ref{fig:thermaltest}(\textbf{D})). However, $\epsilon_{IR}$ increases from 48h to 72h and causes a drop of $\eta_{abs}$.
To understand the mechanism that causes the degradation above 400$^{\circ}$C and after 72h thermal treatment, the samples are characterized using FE-SEM (SIGMA VP). Figure \ref{fig:thermaltest}(\textbf{E}) shows the SEM topographic images of the SSAs surface before and after thermal annealing for one hour. From 25$^{\circ}$C to 400$^{\circ}$C, it is difficult to see any apparent changes. When the temperature keeps going up from 400$^{\circ}$C to 500$^{\circ}$C, small show up, which is also demonstrated in the reflectivity spectra in Fig. \ref{fig:thermaltest}(\textbf{A}), due to the thermal stress coming from the difference in thermal expansion coefficient between SiO$_2$ (0.55 $\times$ 10$^{-6}$ m/m$\cdot$K) and W (4.2 $\times$ 10$^{-6}$ m/m$\cdot$K) \cite{chirumamilla2016multilayer}. Figure \ref{fig:thermaltest}(\textbf{F}) shows the SEM topographic images of the SSAs surface before and after long-time thermal treatment. Little topographic change of SSAs can be seen from 12h to 48h, while small round cracks like burst bubbles appear after 72h and become bigger after 96h. It agrees with the reflectivity spectra change of Fig. \ref{fig:thermaltest}(\textbf{C}). It provides a guideline for improvement approaching a perfect SSAs by the utilization of other materials with a similar thermal expansion coefficient to avoid the thermal stress at high temperature and long-time heating.
Figure \ref{fig:thermaltest}(\textbf{G}) shows that the $R_{IR}$ of SSAs gradually drops heating under elevated temperatures. Each stepped drops from 100$^{\circ}$C to 400$^{\circ}$C is larger than that from 400$^{\circ}$C to 800$^{\circ}$C. It proves that most of the thermal stress between different layers of SSAs is released when heated from 100$^{\circ}$C to 400 $^{\circ}$C.
Combined with the SEM topographic image of Fig. \ref{fig:thermaltest}(\textbf{E}), we conclude that the formation of cracks contributes more than the development of cracks to the decrease of $R_{IR}$. Even heated up to 900$^{\circ}$C, the $R_{IR}$ of SSAs still remain 90.15\% compared with the value at ambient temperature, this indicates that the combinations of SiO$_{2}$ and Al$_{2}$O$_{3}$ has the potential applications under high-temperature solar thermal systems.
\begin{figure}[!ht]
\centering
\includegraphics[width=1.0\textwidth]{eff_mechanical_field}
\caption{ \label{fig:eff_mechanical_field} (\textbf{A}) Solar to heat conversion efficiency as a function of absorber operational temperature, $T_{abs}$, for an ideal selective absorber, SSAs with reflectivity spectrum of measured or simulated, and a black surface, under the unconcentrated solar light (1 sun); (\textbf{B}) Solar to heat conversion efficiency for abovementioned four surfaces as a function of concentration factor, CF, at the operational temperature of $T_{abs}$ = 400$^{\circ}$C. (\textbf{C}) Thermal performance of the selective absorber (orange curve) and the black surface (yellow curve) over a one-day solar cycle from sunrise (5:00 a.m.) to one hour after sunset (8:00 p.m.) at varying ambient temperature (blue curve) under 1 sun. 3D schematic (\textbf{D}) of the real (\textbf{E}) outdoor experimental set-up. (\textbf{F}) The temperature variations of the sample's surface, vacuum chamber, and outdoor ambient environment. $\epsilon_{solar}$ (\textbf{G}), $R_{IR}$ (\textbf{H}), and $\eta_{abs}$ (\textbf{I}) after the different abrasion tests under different weight (50g/100g) as a function of cycles.
}
\end{figure}
\subsection{Solar conversion efficiency investigations of fabricated multilayer SSAs}
As discussed above, the designed structure of SSAs is demonstrated to be angular-insensitive as well as thermal stable up to 400$^{\circ}$C. We employ the reflectivity measured by the UV/VIS/NIR spectrophotometer and FTIR spectrometer at room temperature (25$^{\circ}$C), as shown in Fig. \ref{fig:sem_variangle}(\textbf{G}). The $\alpha_{\lambda,abs}^{\prime}$ and $\epsilon_{\lambda,abs}^{\prime}$ are assumed to be independent of temperature as observed in Fig. \ref{fig:thermaltest}(\textbf{A}). The spectral integration for $\alpha_{\lambda,abs}^{\prime}$ and $\epsilon_{\lambda,abs}^{\prime}$ is performed over wavelengths from 0.3 $\mu$m to 16 $\mu$m, which covers 99.9\% of the solar radiation and only 4.7\% of thermal radiation energy outside this spectral region for a 400$^{\circ}$C blackbody. The reflectivity of the ideal absorber is zero below the cut-off wavelength, while its reflectivity is unity beyond the cut-off wavelength to minimize thermal leakage from blackbody radiation (Fig. \ref{fig:sem_variangle} (\textbf{A}), the dark-orange dot-dash line). The cut-off wavelength is optimized according to the shifts of blackbody thermal radiation at different operational temperatures to maximize the energy conversion efficiency, which defines an upper limit of system efficiency. The reflectivity of black surface, $R_{abs,black}$, is zero over the entire wavelength region ($\alpha$ = 1 -- $R_{abs,black}$ = 1) and shows no spectral selectivity (Fig. \ref{fig:sem_variangle}(\textbf{A}), the black long dash line).
The conversion efficiencies (by analyzing formulae section 1 in supplementary materials) of SSAs are 84.28\% and 86.32\% for the simulated and measured spectral reflectivity at $T_{abs}$ = 100$^{\circ}$C, respectively (Fig. \ref{fig:eff_mechanical_field}(\textbf{A})). The efficiency drops gradually to zero at the stagnation temperature of 396$^{\circ}$C and 486$^{\circ}$C using simulated and measured optical properties, respectively. The absorbed solar energy equals the blackbody re-emission energy (i.e., no solar thermal energy is actually harvested) at this point. The stagnation temperature of SSAs using the simulated reflectivity spectrum is higher than the one using the measured reflectivity spectrum since the reflectivity curve of the simulated one is steeper than the measured one over the transition wavelength region. The efficiency curves of SSAs with measured and simulated radiative properties intersect at 210$^{\circ}$C. The fabricated SSA has a higher efficiency than the designed SSA below 210$^{\circ}$C, because the measured spectrum has a higher reflectivity than the simulated one from 7.7 $\mu$m to 20 $\mu$m, (as marked by red solid pentagram in Fig. \ref{fig:sem_variangle}(\textbf{G})). Half of the blackbody thermal radiation spreads in this wavelength region (the thermal radiation of a 210$^{\circ}$C blackbody is within 3.0 $\mu$m to 15 $\mu$m). When the temperature goes above 210$^{\circ}$C, the efficiency of SSAs using simulated reflectivity spectrum exhibits an advantage than the measured one since the blackbody radiation blue-shifts to the shorter wavelength, where the simulated spectrum has a higher reflectivity than the measured.
As a reference, the black surface can only convert 34.8\% solar energy to heat at $T_{abs}$ = 100$^{\circ}$C, and its efficiency goes down to zero at 126$^{\circ}$C, which further demonstrates the significance of spectral selectivity in enhancing the solar to heat conversion efficiency. On the other hand, the efficiency of the multilayered absorber is 17\% and 20\% lower than the ideal surface at $T_{abs}$ = 100$^{\circ}$C with measured optical properties and simulated radiative spectrum data, respectively. This gap becomes large when the stagnation temperature increases. It mainly results from the larger thermal emittance (around 0.09) in the mid-infrared regime and the cut-off wavelength of the ideal surface is optimized at each temperature according to Wien's displacement law, while the cut-off wavelength of selective absorber keeps unchanged at around 1.2 $\mu$m. Simultaneously, the reflectivity spectrum of the ideal surface changes much sharply at the cut-off point than the multilayered SSAs, comparing Fig. \ref{fig:sem_variangle}(\textbf{A}) and Fig. \ref{fig:sem_variangle}(\textbf{G}). Therefore, the geometry parameters of the multilayered SSAs need to be optimized to make the cut-off wavelength perfectly matched the operational temperature.
Figure \ref{fig:eff_mechanical_field}(\textbf{B}) shows the efficiency as a function of concentration factors, CF, from 1 to 1,000, when $T_{abs}$ = 400$^{\circ}$C, aiming at a mid-temperature application. The cut-off wavelength of the ideal surface is optimized according to different concentration factors. This indicates an upper limit for the SSAs' performance under different concentration factors. The energy efficiency of SSAs with measured or simulated optical properties and the black surface keeps going up with the increasing of CF. The efficiency of SSAs with the measured reflectivity spectrum is lower than the simulated one because the simulated spectrum has a higher solar absorptance than the measured one. For the black surface, its energy conversion efficiency becomes greater than zero at around 12 suns and climbs close to the ideal surface when the CF = 1,000 since the solar radiation heat flux is much larger than the 400$^{\circ}$C blackbody thermal radiation under 1000 suns.
\subsection{Thermal performance investigations under unconcentrated solar radiation}
Figure \ref{fig:eff_mechanical_field}(\textbf{C}) shows the calculated transient temperature variations of SSAs and the black surface under one sun over a solar cycle in Boston, Massachusetts \cite{Weatherboston} using the ambient temperature \cite{Weatherboston} and the solar illumination data \cite{solarangleboston} of July 10, 2018 (see formulae section 2). The highest temperature of the SSAs is 273$^{\circ}$C, while the highest temperature of the black surface is 72$^{\circ}$C. The thermal performance of the selective solar absorber is overwhelmingly better than the black surface at any time under sunshine. The highest temperature difference between SSAs and the black surface is 311$^{\circ}$C under 20 suns (see supplementary Fig. S1), which is higher than the temperature difference under one sun (201$^{\circ}$C). We evaluate the outdoor test performance by measuring the stagnation temperature under unconcentrated solar radiation in Kingston, Rhode Island on May 27, 2019. A SSAs sample was placed in a 100 cm diameter vacuum chamber which was equipped with a Potassium bromide (KBr) window (70 cm in diameter) (Fig. \ref{fig:eff_mechanical_field}(\textbf{D}) and (\textbf{E})). The KBr window is highly transparent (\textgreater 93\%) from UV to the mid-infrared range, which allows sunlight to come in and the thermal radiation of the heated SSAs re-emit to the sky. The vacuum chamber was set on a mounting frame with an angle of $\sim$ 32$^{\circ}$ to ensure the absorber was normal to the incoming sunlight. A K-type thermocouple was secured to the back of the SSAs sample with Kapton tape. Another K-type couple was taped to the bottom of the radiation shield to measure the local temperature of the vacuum chamber. In order to eliminate the conductive and radiative heat losses, we supported the absorber sample with low thermal conductivity (29 mW/m$\cdot$K) aerogel tiles that were on top of the radiation shield made of several double-side polished aluminum sheets. The vacuum chamber was pumped continuously to be $\sim$ 5 $\times$ 10$^{-3}$ Pa with the turbomolecular pump to eliminate the convective and conductive thermal losses from the air. The warmed up to the maximum stagnation temperature of 193.5$^{\circ}$C at around 11:30 since the weather was partly cloudy, the temperature of the vacuum chamber and the absorber sample showed strong fluctuations during the test and was lower than the calculated maximum temperature. The stagnation temperature is enough for the rooftop solar thermal systems even under partly cloudy weather.
\subsection{SSAs’ surface abrasion robustness tests}
Figure \ref{fig:eff_mechanical_field}(\textbf{G}, \textbf{H}, and \textbf{I}) show $\epsilon_{solar}$ (\textbf{G}), $R_{IR}$ (\textbf{H}), and $\eta_{abs}$ (\textbf{I}) of SSAs after each abrasion cycle (the definition of one abrasion cycle is in supplementary Fig. S2 ). It can be found that the $\epsilon_{solar}$, $R_{IR}$, and $\eta_{abs}$ remain 94.6\%, 83.5\%, and 86.9\% after 50 abrasion cycles under 100g weight (2040 Pa). The $R_{IR}$ of SSAs under 50g weight (1020 Pa) first drops and then increases, since the pressure of 50g weight can remove part of the top deposited SiO$_2$ and Al$_2$O$_3$ layer and make the bottom W layer exposed out in 5 cycles. While the 100g weight can make part of the bottom W layer exposed out only in 1 cycle. Since the abrasion resistance tests will break the fabricated SSAs layer, the $\epsilon_{solar}$ keeps decreasing whichever 50g or 100g weight is placed above. The $\eta_{abs}$ of 50g weight tests first decreases and goes up after 10 cycles, due to the $R_{IR}$ fluctuations of 50g weight tests.
\section{Discussion}
\label{Sec3}
The spectral selectivity and thermal performance of multilayered SSAs consisting of Al$_2$O$_3$-W-SiO$_2$-W stacks are investigated analytically and experimentally. It demonstrates 92\% solar absorptance and 6\% reflectivity in the blackbody thermal radiation region. Oblique reflectivity is also experimentally characterized to show high spectral selectivity for both TE and TM polarizations, which demonstrates its angular and polarization insensitivity. High-temperature thermal annealing performed at elevated temperatures for one hour shows the fabricated SSAs maintain its wavelength selectivity at even 400$^{\circ}$C and 500$^{\circ}$C. The high-temperature thermal resistance test is investigated under 400$^{\circ}$C for 12h, 24h, 48h, 72h, and 96h cycles and it demonstrates SSAs keep good selectivity within 48h, while slight degradation of spectral selectivity happens after the 72h and 96h thermal treatment. Real-time reflectivity measurements show that the material displays only a minor change in optical properties above 800$^{\circ}$C. Our study indicates that the proposed metamaterial exhibits thermal resistance and practically retains its optical proprieties even after 96h operation at 400$^{\circ}$C. Retention of optical properties after high temperature annealing can be attributed to pattern-free design of the material as cracks can be significantly more detrimental to optical properties of surface designs such as gratings. The outdoor tests show a peak stagnation temperature of 193.5$^{\circ}$C, and thus indicate its tremendous potential for low temperature solar thermal applications in the absence of solar concentrator. Additionally, the surface abrasion test yields that SSAs have a robust resistance to sandpaper abrasion attack for a long-duration practical application, Abrasion resistance can also be attributed to this pattern-free design.
\section{Materials and Methods}
\label{Sec4}
\subsection{Materials}
SiO$_2$ (99.995\%, 4N5), Al$_2$O$_3$(99.995\%, 4N5), and W (99.995\%, 3N5) are purchased from Angstrom Sciences, Inc. 2 inches wafers (Type: P; dopant: B; orientation: <1-0-0>; resistivity: 10-20 ohm/cm; thickness: 279$\pm$25 $\mu$m) are purchased from WaferPro.
\subsection{Methods}
\subsubsection{Sample fabrication and SEM topography characterizations}
SSAs samples are deposited with a magnetron sputtering technique within a home-built high-vacuum sputtering system (see supplementary Fig. S3). The fabrication process has been described in detail in this publication \cite{tian2019near}. The fabrication parameters, such as deposition rate, base pressure, and sputtering power for each deposition procedure are specified in Table S2 in supplementary materials. The base pressure before sputtering is 3.4 $\times$ 10$^{-7}$ Torr. The Al$_2$O$_3$ and W layer are both deposited directly from Al$_2$O$_3$ and W target, respectively, while the SiO$_2$ layer is oxidized from the silicon (Si) with radio frequency (RF) sputtering. The cross-section and topography of absorber samples after various temperature thermal annealing and diverse long-time thermal resistance tests are characterized by SIGMA VP Field Emission-Scanning Electron Microscope (FE-SEM).
\subsubsection{Optical properties measurements}
Hemispherical reflectance measurements in the UV, visible, and near-infrared regions (from 0.3 $\mu$m to 2.5 $\mu$m) are performed on a Jasco V770 spectrophotometer with a 60 mm PTFE based integrating sphere. A wavelength scan step is fixed as 1 nm with AOI of 6$^{\circ}$ and normalized to a labsphere spectralon reflectance standard. The mid-infrared (2.5 $\mu$m to 15 $\mu$m) reflectance measurements are executed on Jasco 6600 FTIR spectrometer together with a Pike 3-inches golden integrating sphere. The AOI of the incident beam from the FTIR spectrometer is fixed at 6$^{\circ}$ and the diffused golden reference is selected as a normalized standard. These spectra are taken at a scan rate of 64 with a wavelength resolution of 0.4 cm$^{-1}$. Details about FTIR spectrum measurements can be referred to the recent article \cite{tian2019near}.
The hemispherical reflectance measurement for different angles of incident (AOI) is conducted by using home-built wedge blocks (see supplementary Fig. S4) with (18$^{\circ}$, 30$^{\circ}$, 33$^{\circ}$, 33$^{\circ}$, 51$^{\circ}$, 66$^{\circ}$, and 76$^{\circ}$) combined with Jasco 60mm PTFE based integrating sphere and Pike 3-inches golden integrating sphere. The wedge blocks for PTFE integrating sphere are put faced to the sample ports of the Jasco and Pike integrating sphere. SSAs samples and the reference standard are put faced to the sample ports on designed wedge blocks. Background spectrum and the reflectance spectrum are taken as usual.
The specular reflectance measurements of variable angle in the UV, visible, and near-infrared regions (from 0.3 $\mu$m to 2.5 $\mu$m) are performed on Jasco V770 spectrophotometer with Harrick variable angle reflection accessory. The variable angle specular reflectance measurements (from 2.5 $\mu$m to 15 $\mu$m) are executed on Jasco 6600 FTIR spectrometer together with Harrick Seagull$^{TM}$ Variable Angle Reflection Accessory. Both measurements are taken for angles of 12$^{\circ}$, 30$^{\circ}$, 45$^{\circ}$, 60$^{\circ}$, and 75$^{\circ}$.
The hemispherical reflectance measurements of variable angle polarization from 0.3 $\mu$m to 2.5 $\mu$m are done on Jasco V770 spectrophotometer with GPH-506 polarizer and 60 mm integrating sphere. The hemispherical reflectance measurements of variable angle polarization from 2.5 $\mu$m to 15 $\mu$m are executed on Jasco 6600 FTIR spectrometer together with PL-82 polarizer and Pike 3-inches golden integrating sphere.
\subsubsection{High-temperature thermal annealing and resistance tests}
High-temperature thermal annealing and resistance tests are both carried out in a tube oven at an ambient atmosphere with an alumina tube of 5 cm diameter and 80 cm length. Samples are put in an alumina crucible boat (100 mm $\times$ 30 mm $\times$ 20 mm) that is placed in the center of the tube. The increasing rate of temperature is set to be at 15$^{\circ}$C/min for 100$^{\circ}$C 200$^{\circ}$C 300$^{\circ}$C 400$^{\circ}$C, and 500$^{\circ}$C for one hour. The hemispherical reflectance measurements are carried out after each thermal treatment. For the high-temperature thermal resistance tests, the temperature controller of the tube oven is set to be 400$^{\circ}$C for 12h, 24h, 48h, 72h, and 96h thermal cycles. The samples are taken out after each thermal cycle for reflectance spectrum measurements and a small piece of the sample ($\sim$ 3mm $\times$ 3mm) is cut for SEM topography characterization.
\subsubsection{Real-time reflectivity measurement of SSAs under high vacuum}
The real-time measurement of reflectivity for SSAs is conducted on Jasco FTIR 6600 spectrometer within the Harrick high-temperature reaction chamber. The chamber can be heated by an electrical heater up to 910
$^{\circ}$C and vacuumed down to < 10$^{-3}$ Pa connected to a Pfeiffer TMH 260 turbopump. The temperature of the SSAs is measured by a K-type thermocouple and controlled by the Waterloo PID temperature controller.
\subsubsection{SSAs’ surface abrasion robustness tests}
A sandpaper abrasion test was carried out using 400 grit SiC sandpaper as an abrasion surface, as shown in supplementary Fig. S2. The SSAs samples with a weight of 50/100 g above it are put face-down to sandpaper and move 10 cm along the ruler (see supplementary Fig. S2 B); then, the sample is then rotated counterclockwise by 90$^{\circ}$ (face to the sandpaper) and move 10 cm along the ruler. The three steps shown are in supplementary Fig. S2 B is defined as one abrasion cycle. This guarantees that the SSAs’ surface is abraded longitudinally and transversely in each cycle. The reflectivity spectra are measured after 1, 3, 5, 10, 20, 30, and 50 cycles.
\section{Acknowledgements} This project is supported in part by the National Science Foundation through grant numbers CBET-1941743, and National Aeronautics and Space Administration through grant number NNX15AK52A.
\section{Author Contributions}
Y.T. and X.L. develop the theoretical calculations and carry out spectrum measurements and field tests. L.Q. fabricates SSAs samples. A.G. and J.L. contribute to the interpretation of the results. All authors provide critical feedback and help revise the final version of the manuscript. T.T., G.X., and Y.Z. supervise the project.
\section{Conflicts of interest}
The authors declare no conflict of interest.
\section{Bibliography}
| bce80dc9557b8d2bcefaac2b43c29e89efc67e29 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A \emph{signed graph} is a graph in which every edge has an associated sign. We write a signed graph $\Sigma$ as the triple $(V, E, \sigma)$ where $V$ is the vertex set, $E$ is the edge set, and $\sigma: E \to \{+,-\}$ is the \emph{signature}. Our graphs are \emph{signed simple graphs}: with no loops and no multiple edges.
We define signed-graph coloring as in \cite{zaslavsky}, and chromatic number as in \cite{macajova}. A \emph{proper coloration} of a signed graph $\Sigma$ is a function, $\kappa : V \to \{\pm1, \pm 2, \ldots, \allowbreak \pm k, 0\},$ such that for any edge $e_{ab} \in E$, $\kappa(a) \neq \sigma(e) \kappa(b).$ The \emph{chromatic number} of $\Sigma$, written $\chi(\Sigma)$, is the size of the smallest set of colors which can be used to properly color $\Sigma.$ A graph with chromatic number $k$ is called \emph{k-chromatic}. A coloration is \emph{minimal} if it is proper and uses a set of colors of size $\chi(\Sigma).$ If $\chi = 2k$ then a minimal \emph{color set} is $\{\pm1, \pm2, \ldots, \pm k\},$ and if $\chi = 2k+1$ then a minimal color set is $\{\pm1, \pm2, \ldots, \pm k, 0\}.$ If $\chi = 2k+1$, there must be at least one vertex colored 0 in every minimal coloration.
The \emph{deficiency} of a coloration, $\operatorname{def}(\kappa)$, is the number of unused colors from the color set of $\kappa.$ The \emph{deficiency set}, $\operatorname{D}(\kappa)$, is the set of unused colors. The \emph{maximum deficiency} of a graph, $\operatorname{M}(\Sigma) $, is $\max\{\operatorname{def}(\kappa) \mid \kappa$ is a minimal proper coloration of $\Sigma \}.$ For more about deficiency, see \cite{mattern}.
Let $\Sigma_1$ and $\Sigma_2$ be signed graphs. The \emph{$\sigma^*$-join} of $\Sigma_1$ and $\Sigma_2$, written $\Sigma_1 \vee_{\sigma^*} \Sigma_2$, is the signed graph with
\begin{align*}
V &= V(\Sigma_1) \cup V(\Sigma_2), \\
E &= E(\Sigma_1) \cup E(\Sigma_2) \cup \{e_{vw} \mid v \in V(\Sigma_1), w \in V(\Sigma_2)\}, \\
\text{and } \sigma(e) &= \begin{cases}
\sigma_1(e) &\text{ if } e \in E(\Sigma_1), \\
\sigma_2(e) &\text{ if } e \in E(\Sigma_2), \\
\sigma^*(e) &\text{ otherwise,}
\end{cases}
\end{align*}
where $\sigma^*$ is a function from the new join edges to the set $\{+,-\}.$
We explore joins of signed graphs and prove an analogue to the theorem that the chromatic number of the join of two graphs equals the sum of their chromatic numbers. In this case, the chromatic number of the join of two signed graphs depends on both the chromatic numbers and the maximum deficiencies of the two graphs.
\section{Joins of Signed Graphs}
\label{joins of signed graphs}
Unlike in ordinary graph theory, there are many ways to join two signed graphs. The result of joining two signed graphs depends on the signs of the new edges. In this paper we focus on the \emph{all-positive join} of two signed graphs, where each new join edge has positive sign. We write the all-positive join of $\Sigma_1$ and $\Sigma_2$ as $\Sigma_1 \vee_+ \Sigma_2.$
Let $\Sigma$ be a signed graph with even chromatic number and maximum deficiency $M$. Then $\Sigma$ is an {\it exceptional graph} if in every proper coloration using $\chi(\Sigma) - M$ colors, every color that is used appears at both ends of some negative edge. Figure 1 shows two examples of exceptional graphs. Graph A is a 6-chromatic graph with maximum deficiency 3. Graph B is a 4-chromatic graph with maximum deficiency 2. In both colorations, every color used appears on both endpoints of some edge. In depictions we use solid lines for positive edges and dashed lines for negative edges.
\begin{figure}[H]
\centering
\label{exceptional}
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}
]
\node at (0,0) [acteur,label = below:1]{};
\node at (3,0) [acteur,label = below:-2]{};
\node at (0,4) [acteur,label = above:3]{};
\node at (3,4) [acteur,label = above:3]{};
\node at (-1,2) [acteur,label = left:1]{};
\node at (4,2) [acteur,label = right: -2]{};
\node at (8,0) [acteur, label = below: 1]{};
\node at (10,0) [acteur, label = below: 2]{};
\node at (12,0) [acteur, label = below: 2]{};
\node at (8,2) [acteur, label = above: 1]{};
\node at (10,2) [acteur, label = above: 1]{};
\node at (12,2) [acteur, label = above: 1]{};
\draw[thick] (0,0) -- (3,0) -- (0,4) -- (-1,2) -- (4,2) -- (3,4) -- cycle;
\draw[thick, dashed] (0,4) -- (3,4) -- (-1,2) -- (0,0) -- cycle;
\draw[thick, dashed] (3,0) -- (4,2) -- (0,0);
\draw[thick, dashed] (3,0) -- (3,4);
\draw[thick, dashed] (3,0) -- (-1,2);
\draw[thick, dashed] (0,4) -- (4,2);
\draw[thick] (10,0) -- (10,2);
\draw[thick] (12,0) -- (12,2);
\draw[thick, dashed] (10,0) -- (8,0) -- (8,2) -- (10,2) -- (8,0);
\draw[thick, dashed] (10,0) -- (8,2);
\draw[thick, dashed] (10,0) -- (12,0) -- (12,2) -- (10,2) -- (12,0);
\draw[thick, dashed] (10,0) -- (12,2);
\node at (1.5, -2) [label = Graph A]{};
\node at (10,-2) [label = Graph B]{};
\end{tikzpicture}
\caption[Two properly colored exceptional graphs]{Two properly colored exceptional graphs.}
\end{figure}
There exist infinitely many exceptional graphs. For example, for non-negative $k,$ if $\chi = 2k,$ the complete graph on $4k$ vertices with a negative perfect matching and all other edges positive is an exceptional graph with maximum deficiency 0. It remains an open problem to characterize exceptional graphs in terms of their structure.
\begin{thm}
Let $\Sigma_1$ and $\Sigma_2$ be signed graphs with maximum deficiencies $M_1$ and $M_2,$ respectively. Assume that $M_1 \geq M_2.$ Then, with one exception,
$$\chi(\Sigma_1 \vee_+ \Sigma_2) = \max\{\chi_1 + \chi_2 - M_1 - M_2, \chi_1\}.$$
Exception: If $\Sigma_1$ and $\Sigma_2$ both have even chromatic number, exactly one of $M_1$ and $M_2$ is odd, and both $\Sigma_1$ and $\Sigma_2$ are exceptional graphs, then
$$\chi(\Sigma_1 \vee_+ \Sigma_2) = \max\{\chi_1 + \chi_2 - M_1 - M_2 +1, \chi_1\}.$$
\label{main thm}
\end{thm}
Note that in Theorem \ref{main thm}, $\chi_2$ can never be larger than $\chi_1 + \chi_2 - M_1 - M_2.$ If $\chi_2 > \chi_1 + \chi_2 - M_1 - M_2,$ this would imply that $M_2 > \chi_1 - M_1.$ Since $M_1 \leq \frac{1}{2}\chi_1,$ we then would have $M_2 > \frac{1}{2} \chi_1 \geq M_1$. This contradicts our assumption that $M_1 \geq M_2.$
Let $A$ be a set of vertices of a signed graph $\Sigma.$ \emph{Switching} $A$ is negating the signs of the edges with exactly one endpoint in $A$. Two signed graphs are \emph{switching equivalent} if they are related by switching. The chromatic number of a signed graph is the same as the chromatic number of the switched graph. A minimal coloration of the switched graph is simply a byproduct of switching the graph; given a minimal coloration of $\Sigma$, the signs of the colors on $A$ are negated during the switching process. Two colorations of $\Sigma, \kappa$ and $\kappa^*,$ are \emph{switching equivalent} if they are related by switching.
\begin{rmk}
Theorem \ref{main thm} also holds for the all-negative join of signed graphs. The possible joins of $\Sigma_1$ and $\Sigma_2$ come in switching equivalent pairs. For a signature $\sigma,$ let $-\sigma$ be the signature in which the sign of an edge $e$ is $-\sigma(e).$
Then $\Sigma_1 \vee_{\sigma^*} \Sigma_2$ switches to $\Sigma_1 \vee_{-\sigma^*} \Sigma_2$. Thus, $\Sigma_1 \vee_- \Sigma_2$ switches to $\Sigma_1 \vee_+ \Sigma_2$ by switching the vertices of $\Sigma_1.$ Furthermore, switching all the vertices of $\Sigma_1$ does not change the signs of the edges in $\Sigma_1$ or $\Sigma_2$, and thus does not change the maximum deficiencies.
\end{rmk}
\section{Preliminaries}
\label{preliminaries}
When coloring the graph $\Sigma_1 \vee_+ \Sigma_2,$ one cannot use the same color on both the vertices of $\Sigma_1$ and $\Sigma_2$. This gives a straightforward lower bound for the chromatic number of the all-positive join.
\begin{lem}
Let $\Sigma_1$ and $\Sigma_2$ be signed graphs with maximum deficiencies $M_1$ and $M_2$, respectively. Then $\chi(\Sigma_1 \vee_+ \Sigma_2) \geq \chi_1 + \chi_2 - M_1 - M_2.$
\label{lower bound}
\end{lem}
Throughout the proofs of Theorem \ref{main thm}, we use several recoloration tools in order to give a proper coloration of the correct size. Define the following replacement types:
{\bf Type 1:} Let $r \in \mathbb Z \setminus \{0\}.$ For $i$ in the color set of $\kappa,$ recolor the $i$-color set with the color $r.$
{\bf Type 2:} For $i \neq 0$ in the color set of $\kappa,$ such that $i$ does not appear on both endpoints of an edge, recolor the $i$-color set with the color 0.
{\bf Type 3:} Let $r \in \mathbb Z \setminus \{0\}.$ For $i$ and $-i$ in the color set of $\kappa,$ such that $i \neq 0$, recolor the $i$-color set with the color $r$ and the $(-i)$-color set with the color $-r$.
{\bf Type 4:} Let $r_1, r_2 \in \mathbb Z \setminus \{0\}$ such that $r_1 \neq -r_2.$ For $i$ and $-i$ in the color set of $\kappa,$ such that $i \neq 0$, recolor the $i$-color set with the color $r_1$ and the $(-i)$-color set with the color $r_2$.
\begin{proposition}
Let $\Sigma$ be a signed graph with even chromatic number and non-zero maximum deficiency. Then in every minimal coloration with maximum deficiency, the negative of every color in the deficiency set appears on both endpoints of some negative edge.
\label{even chromatic}
\end{proposition}
\begin{proof}
Let $\Sigma$ be a signed graph with even chromatic number $\chi$ and non-zero maximum deficiency $M$. Let $\kappa$ be a coloration of $\Sigma$ such that $\operatorname{def}(\kappa) = M.$ Suppose to the contrary that there exists some $i \in \operatorname{D}(\kappa)$ such that $-i$ does not appear on both endpoints of a negative edge. Thus, if $\kappa(v) = -i$, then no neighbor of $v$ is colored $-i.$ Recoloring all vertices colored $-i$ with the color 0 yields a proper coloration since $\kappa$ did not use the color 0 and no two vertices colored 0 are adjacent. But, the size of the new color set is $\chi - 1.$
\end{proof}
\section{Chromatic Number $\chi_1$ and the Exceptional Case}
\label{chi_1 and exception}
\begin{lem}
If $\chi_1 \geq \chi_1 + \chi_2 - M_1 -M_2,$ then $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1.$
\label{exception 1}
\end{lem}
\begin{proof}
Let $\kappa$ be a coloration of $\Sigma_1 \vee_+ \Sigma_2.$ Then $\kappa$ restricted to $\Sigma_1$ is a proper coloration of $\Sigma_1.$ Therefore, the size of the color set must be at least $\chi_1.$
Now we show there exists a proper coloration of $\Sigma_1 \vee_+ \Sigma_2$ using a color set of size $\chi_1$. Color $\Sigma_1 \vee_+ \Sigma_2$ in the following way.
\begin{enumerate}
\item Properly color $\Sigma_1$ using $\chi_1 - M_1$ colors. Call this coloration $\kappa$.
\item Properly color $\Sigma_2$ using the colors in the deficiency set of $\kappa$. This is possible since the deficiency set is made up of $M_1$ colors with distinct absolute values, and $M_1 > \chi_2 - M_2$, because $\chi_1 > \chi_1 + \chi_2 - M_1 - M_2.$
\end{enumerate}
This coloration is proper on $\Sigma_1$ and $\Sigma_2$ by definition. Since every join edge is positive, and every color used on $\Sigma_2$ does not appear on a vertex of $\Sigma_1$, the coloration is proper on $\Sigma_1 \vee_+ \Sigma_2.$ Furthermore, the size of the color set is $\chi_1,$ and therefore, $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1.$
\end{proof}
\begin{lem}[Exception]
Assume that $\chi_1$ and $\chi_2$ are even, $M_1 > M_2,$ and $\chi_1 \leq \chi_1 + \chi_2 - M_1 - M_2 + 1.$ If exactly one of $M_1$ and $M_2$ is odd, and both $\Sigma_1$ and $\Sigma_2$ are exceptional, then $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1 + \chi_2 - M_1 - M_2 +1.$
\label{exception 2}
\end{lem}
\begin{proof}
Let $\chi_1 = 2k_1$ and $\chi_2 = 2k_2$ for some positive integers $k_1$ and $k_2.$ By Lemma \ref{lower bound} we know that $\chi(\Sigma_1 \vee_+ \Sigma_2) \geq \chi_1 + \chi_2 - M_1 - M_2.$
Suppose that $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1 + \chi_2 - M_1 - M_2.$ Note that $\chi_1 + \chi_2 - M_1 - M_2$ is odd. Thus a minimal coloration of $\Sigma_1 \vee_+ \Sigma_2$ must use 0 as a color. Let $\kappa$ be such a coloration. Because all join edges are positive, no color can be used on both $\Sigma_1$ and the vertices of $\Sigma_2.$ Thus, $\kappa$ must use $2k_1 - M_1$ colors on $\Sigma_1$ and $2k_2 - M_2$ different colors on the vertices of $\Sigma_2.$ Therefore $\kappa$ restricted to $\Sigma_i$ is a proper coloration using $\chi_i - m_i$ colors. Since both $\Sigma_1$ and $\Sigma_2$ are exceptional graphs, every color must appear on both endpoints of some negative edge. But the color 0 cannot appear on both endpoints of an edge in a proper coloration. Therefore, $\chi(\Sigma_1 \vee_+ \Sigma_2) > \chi_1 + \chi_2 - M_1 - M_2.$
Now we show that there exists a proper coloration of $\Sigma_1 \vee_+ \Sigma_2$ using a color set of size $\chi_1 + \chi_2 - M_1 - M_2 + 1$. Although the size of the color set will be $\chi_1 + \chi_2 - M_1 - M_2 + 1$, we will only use $\chi_1 + \chi_2 - M_1 - M_2$ colors. Color $\Sigma_1 \vee_+ \Sigma_2$ in the following way.
\begin{enumerate}
\item Properly color $\Sigma_1$ with colors $\pm 1, \pm 2, \ldots, \pm k_1$ using $2k_1 - m_1$ colors. Let the $M_1$ unused colors be $x_1, x_2, \ldots, x_{M_1}$.
\item Properly color $\Sigma_2$ with colors $\pm 1, \pm 2, \ldots, \pm k_2$ using $2k_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, y_2, \ldots, y_{M_2}$.
\item Make the following Type 1 color replacements on the vertices of $\Sigma_2:$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Using Type 4 replacements, recolor $\frac{M_1 - M_2 - 1}{2}$ more pairs of color in $\Sigma_2$ using colors $x_{M_2 + 1}, \ldots, x_{M_1 - 1}.$ Note that we only go up to $x_{M_1 - 1}$ since we need an even number of colors for this step.
\item Replace the remaining pairs of colors in $\Sigma_2$ using Type 3 replacements and colors $\pm(k_1 + 1), \ldots, \pm (k_1 + k_2 - \frac{M_1 - M_2 - 1}{2}).$
\end{enumerate}
Call the new coloration $\kappa.$ Since no colors were changed in $\Sigma_1$, $\kappa$ is proper on $\Sigma_1.$ In $\Sigma_2$ all of the color sets were recolored using Type 1, Type 3, and Type 4 replacements, and the original partition of $V(\Sigma_2)$ into color sets was maintained. Thus, the subgraph induced by every color set of $\kappa$ on $\Sigma_2$ is all-negative. Furthermore, if $r$ is a color of $\kappa$ resulting from a Type 1 or Type 4 replacement, then there are no vertices colored $-r$ in $\Sigma_2.$ If $r$ and $-r$ are colors resulting from a Type 3 replacement, then there are no negative edges between the two color sets.
Therefore, $\kappa$ is proper on $\Sigma_2.$ Finally, no color is used on the vertices of both $\Sigma_1$ and $\Sigma_2$. Thus, $\kappa$ is proper on $\Sigma_1 \vee_+ \Sigma_2.$ Furthermore, the color set is
$\{\pm1, \ldots, \pm (k_1 + k_2 - \frac{M_1-M_2-1}{2})\}.$
Therefore, $$\chi(\Sigma_1 \vee_+ \Sigma_2) = 2\left(k_1 + k_2 - \frac{M_1-M_2-1}{2}\right) = \chi_1 + \chi_2 - M_1 - M_2 + 1. \qedhere$$
\end{proof}
\begin{rmk}
In the case where $M_2 = 0,$ one would simply skip step 3 in the recoloration procedure.
\end{rmk}
Because it is similar to the proof in the exceptional case, for the remaining cases we leave the proof that the new coloration is both proper and of correct size as an exercise for the reader.
\section{Non-Exceptional Cases}
\label{Non-exceptions}
\begin{lem}[$M_2 = 0$]
Let $M_2 = 0.$ Assume $\Sigma_1$ and $\Sigma_2$ do not satisfy the conditions of the exception and that $\chi_1 < \chi_1 + \chi_2 - M_1 - M_2$. Then $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1 + \chi_2 - M_1 - M_2.$
\end{lem}
\label{main lemma 0}
\begin{proof}
We need only show a proper coloration using a color set of size $\chi_1 + \chi_2 - M_1 - M_2.$ We have two cases to consider.
{\bf Case 1:} Suppose $M_1$ is even. We have two subcases.
{\bf Case 1.1:} Suppose at least one of $\chi_1$ and $\chi_2$ is even. Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way.
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1,$ and $0$ if $\chi_1$ is odd, using $\chi_1 -M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ If $\chi_1$ is odd, 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2,$ and $0$ if $\chi_1$ is even and $\chi_2$ is odd, using $\chi_2$ colors.
\item Make the following Type 1 and Type 3 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{ | c | c || c | c | }
\hline
Old Color & New Color & Old Color & New Color\\ \hline
1 & $x_1$ & $\frac{M_1}{2} + 1$ & $k_1 +1$ \\ \hline
-1 & $x_2$ & $-(\frac{M_1}{2} + 1)$ & $-(k_1 +1)$ \\ \hline
$\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ \hline
$\frac{M_1}{2}$ & $x_{M_1 -1}$ & $k_2$ & $k_1 + k_2 - \frac{M_1}{2}$ \\ \hline
$-\frac{M_1}{2}$ & $x_{M_1}$ & $-k_2$ & $-(k_1 + k_2 - \frac{M_1}{2})$ \\
\hline
\end{tabular}
\end{center}
\end{enumerate}
{\bf Case 1.2:} Suppose both $\chi_1$ and $\chi_2$ are odd. That is, $\chi_1 = 2k_1 + 1$ and $\chi_2 = 2k_2 + 1$ for some positive integers $k_1$ and $k_2.$ Then create the coloration $\kappa$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1, 0$ using $\chi_1 -M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ Note that 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2, 0$ using $\chi_2$ colors.
\item Make the following Type 1 and Type 3 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{ | c | c || c | c | }
\hline
Old Color & New Color & Old Color & New Color \\ \hline
1 & $x_1$ & $\frac{M_1}{2} + 1$ & $k_1 + 1$ \\ \hline
-1 & $x_2$ & $-(\frac{M_1}{2} + 1)$ & $-(k_1 + 1)$ \\ \hline
$\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ \hline
$\frac{M_1}{2}$ & $x_{M_1 -1}$ & $k_2$ & $k_1 + k_2 - \frac{M_1}{2}$ \\ \hline
$-\frac{M_1}{2}$ & $x_{M_1}$ & $-k_2$ & $-(k_1 + k_2 - \frac{M_1}{2})$ \\
\hline
\end{tabular}
\end{center}
\item Use Type 1 replacements to recolor the 0-color set in $\Sigma_1$ with $k_1 + k_2 - \frac{M_1}{2} + 1$ and to recolor the 0-color set in $\Sigma_2$ with $-(k_1 + k_2 - \frac{M_1}{2} + 1).$
\end{enumerate}
{\bf Case 2:} Suppose $M_1$ is odd. We have several cases.
{\bf Case 2.1:} Suppose both $\chi_1$ and $\chi_2$ are even. Let $\chi_1 = 2k_1$ and $\chi_2 = 2k_2$ for some positive integers $k_1$ and $k_2$. Then either $\Sigma_1$ or $\Sigma_2$ is not exceptional.
{\bf Case: 2.1a:} Suppose $\Sigma_1$ is not exceptional. Then there exists a coloration using $\chi_1 - M_1$ colors such that there is no edge with both endpoints colored $a,$ for some $a$ in the set of colors used. Call this coloration $\kappa_1.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Color $\Sigma_1$ with $\kappa_1$. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}$. By choice of notation, let $a = 1.$
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2$ using $\chi_2$ colors.
\item Use a Type 2 replacement to replace $a$ with 0 in $\Sigma_1.$
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-1$ & $x_1$ \\ \hline
$2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$\frac {M_1+1}{2}$ & $x_{M_1-1}$ \\ \hline
$-\frac{M_1+1}{2}$ & $x_{M_1}$ \\
\hline
\end{tabular}
\end{center}
\item Use Type 3 replacements to recolor the remaining $k_2 - \frac{M_1 + 1}{2}$ pairs of colors in $\Sigma_2$ with colors $\pm (k_1+1), \ldots, \pm (k_1 + k_2 -\frac{M_1 + 1}{2}) .$
\end{enumerate}
{\bf Case 2.1b:}
Suppose $\Sigma_2$ is not exceptional. Then there exists a proper coloration such that no edge has both endpoints colored $a$ for some $a$ in the color set. Call this coloration $\kappa_2.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1$ and using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$
\item Color $\Sigma_2$ with $\kappa_2$ using colors $\pm 1, \ldots, \pm k_2$ using $\chi_2$ colors. By choice of notation, let $a = 1.$
\item Make the following Type 1, Type 2, and Type 3 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c | | c | c |}
\hline
Old Color & New Color & Old Color & New Color \\ \hline
1 & 0 & $\frac{M_1+1}{2}$ & $k_1 + 1$ \\ \hline
-1 & $x_{M_1}$ & $-\frac{M_1+1}{2}$ & $-(k_1 + 1)$ \\ \hline
2 & $x_1$ & $\vdots$ & $\vdots$ \\ \hline
-2 & $x_2$ & $-k_2$ & $-(k_1 + k_2 - \frac{M_1+1}{2})$ \\ \hline
$\vdots$ & $\vdots$ & & \\ \hline
$-\left(\frac{M_1-1}{2}\right)$ & $x_{M_1 - 1}$ & & \\
\hline
\end{tabular}
\end{center}
\end{enumerate}
{\bf Case 2.2:} Suppose $\chi_1 = 2k_1 + 1$ and $\chi_2 = 2k_2$ for some positive integers $k_1$ and $k_2$. Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1, 0$ using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ Note that $0$ must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2$ using $\chi_2$ colors.
\item In $\Sigma_1$ use a Type 1 replacement to replace 0 with $-(k_1+1).$
\item Make the following Type 1 and Type 3 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c || c | c |}
\hline
Old Color & New Color & Old Color & New Color \\ \hline
1 & $x_1$ & $-\frac{M_1+1}{2}$ & $k_1 + 1$ \\ \hline
-1 & $x_2$ & $\frac{M_1 + 2}{2}$ & $k_1 + 2$ \\ \hline
$\vdots$ & $\vdots$ & $-\frac{M_1 + 2}{2}$ & $-(k_1 + 2)$ \\ \hline
$-\left(\frac{M_1-1}{2}\right)$ & $x_{M_1 - 1}$ & $\vdots$ & $\vdots$ \\ \hline
$\frac{M_1+1}{2}$ & $x_{M_1}$ & $-k_2$ & $-(k_1 + k_2 - \frac{M_1-1}{2})$ \\
\hline
\end{tabular}
\end{center}
\end{enumerate}
{\bf Case 2.3:} Suppose $\chi_2 = 2k_2 + 1$ for some positive integer $k_2$. Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1,$ and 0 if $\chi_1$ is also odd, using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ If $\chi_1$ is odd, 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2, 0$ using $\chi_2$ colors.
\item Make the following Type 1 and Type 3 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c || c | c |}
\hline
Old Color & New Color & Old Color & New Color \\ \hline
1 & $x_1$ & $\frac{M_1+1}{2}$ & $k_1 + 1$ \\ \hline
-1 & $x_2$ & $-\frac{M_1+1}{2}$ & $-(k_1 + 1)$ \\ \hline
$\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ \hline
$-\left(\frac{M_1-1}{2}\right)$ & $x_{M_1 - 1}$ & $-k_2$ & $-(k_1 + k_2 - \frac{M_1-1}{2})$ \\ \hline
$0$ & $x_{M_1}$ & & \\
\hline
\end{tabular}
\end{center}
\end{enumerate}
This concludes the non-exceptional cases where $M_2 = 0.$.
\end{proof}
\begin{lem} [$M_2 > 0$]
Let $M_2 > 0.$ Assume that $M_1 \geq M_2,$ $\Sigma_1$ and $\Sigma_2$ are not exceptional, and $\chi_1 < \chi_1 + \chi_2 - M_1 - M_2$. Then $\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi_1 + \chi_2 - M_1 - M_2.$
\end{lem}
\label{main lemma non-0}
\begin{proof}
We need only show a coloration using a color set of size $\chi_1 + \chi_2 - M_1 - M_2.$ We have two cases to consider.
{\bf Case 1:} Suppose $M_1$ and $M_2$ are either both even or both odd. This implies $M_1 - M_2$ is even.
{\bf Case 1.1:} Suppose $\chi_1 = 2k_1 + 1$ and $\chi_2 = 2k_2 + 1$ for some positive integers $k_1$ and $k_2.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1, 0$ using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ Note that 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2, 0$ using $\chi_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$ Note that $0$ must be used.
\item In $\Sigma_1$ perform a Type 1 replacement and replace 0 with $k_1 + 1.$
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$0$ & $-(k_1 +1)$ \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Replace $\frac{M_1 - M_2}{2}$ more pairs of colors in $\Sigma_2$ using Type 4 replacements and colors $x_{M_2 + 1}, \ldots, x_{M_1}$.
\item Use Type 3 replacements to recolor the remaining pairs of colors in $\Sigma_2$ using $\pm (k_1 + 2), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2}{2} + 1).$
\end{enumerate}
{\bf Case 1.2:} Suppose at least one of $\chi_1$ and $\chi_2$ is even. Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1,$ and possibly 0 using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ If $\chi_1$ is odd, 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2,$ and possibly 0 using $\chi_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$ If $\chi_2$ is odd, $0$ must be used.
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Use Type 4 replacements to recolor $\frac{M_1 - M_2}{2}$ more pairs of colors in $\Sigma_2$ using $x_{M_2 + 1}, \ldots, x_{M_1}$.
\item Replace the remaining pairs of colors in $\Sigma_2$ using Type 3 replacements and colors $\pm (k_1 + 1), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2}{2}).$
\end{enumerate}
{\bf Case 2:} Now suppose exactly one of $M_1$ and $M_2$ is odd. This implies that $M_1 - M_2$ is odd.
{\bf Case 2.1:} Suppose $\chi_1 = 2k_1$ and $\chi_2 = 2k_2$ for some positive integers $k_1$ and $k_2.$ Then either $\Sigma_1$ or $\Sigma_2$ is not exceptional.
{\bf Case 2.1a:} Suppose $\Sigma_1$ is not exceptional. Then there exists a coloration using $\chi_1 - M_1$ colors such that there is no edge with both endpoints colored $a,$ for some $a$ in the set of colors used. Call this coloration $\kappa_1.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Color $\Sigma_1$ using $\kappa_1.$ Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ By choice of notation, let $a = 1.$
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2$ using $\chi_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$
\item In $\Sigma_1$ perform a Type 2 replacement to replace $a$ with 0.
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2 -1}$ & $x_{M_2 -1}$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Use Type 4 replacements to recolor $\frac{M_1 - M_2 + 1}{2}$ more pairs of colors in $\Sigma_2$ using $x_{M_2+1}, \ldots, x_{M_1}$ and $a.$
\item Replace the remaining pairs of colors in $\Sigma_2$ using Type 3 replacements and colors $\pm (k_1 +1), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2 +1}{2}).$
\end{enumerate}
{\bf Case 2.1b:} Suppose $\Sigma_2$ is not exceptional. Then there exists a coloration using $\chi_2 - M_2$ colors and such that no edge has both endpoints colored $a,$ for some $a$ in the set of used colors. Call this coloration $\kappa_2.$ By Lemma \ref{even chromatic} we know that $-a$ cannot be in the deficiency set of $\kappa.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm 1, \ldots, \pm k_1$ using $\chi_1 - M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$
\item Color $\Sigma_2$ with $\kappa_2$ and colors $\pm 1, \ldots, \pm k_2$ Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$
\item In $\Sigma_2$ perform a Type 2 replacement to replace $a$ with 0.
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2 -1}$ & $x_{M_2-1}$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\ \hline
$-a$ & $x_{M_1}$ \\
\hline
\end{tabular}
\end{center}
\item Replace $\frac{M_1 - M_2 - 1}{2}$ more pairs of colors in $\Sigma_2$ using Type 4 replacements and colors $x_{M_2+1}, \ldots, x_{M_1 - 1}.$
\item Use Type 3 replacements to replace the remaining pairs of colors in $\Sigma_2$ using $\pm (k_1 +1), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2 -1}{2} - 1).$
\end{enumerate}
{\bf Case 2.2:} Suppose $\chi_1 = 2k_1 + 1$ and $\chi_2 = 2k_2$ for some positive integers $k_1$ and $k_2$. Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1,$ and 0 using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ Note that 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2$ using $\chi_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Use Type 4 replacements to recolor $\frac{M_1 - M_2 - 1}{2}$ more pairs of colors in $\Sigma_2$ using $x_{M_2 + 1}, \ldots, x_{M_1 - 1}$.
\item Of the remaining pairs of colors in $\Sigma_2$, let $c,-c$ be one. Use a Type 4 replacement to replace $c$ with $x_{M_1}$ and $-c$ with $k_1 + 1.$
\item In $\Sigma_1$ use a Type 1 replacement to replace 0 with $-(k_1 +1).$
\item Replace the remaining pairs of colors in $\Sigma_2$ using Type 3 replacements and colors $\pm (k_1 + 2), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2-1}{2}).$
\end{enumerate}
{\bf Case 2.3:} Suppose $\chi_2 = 2k_2 + 1$ for some positive integer $k_2.$ Then color $\Sigma_1 \vee_+ \Sigma_2$ in the following way:
\begin{enumerate}
\item Properly color $\Sigma_1$ with $\pm1, \ldots, \pm k_1,$ and possibly 0 using $\chi_1-M_1$ colors. Let the $M_1$ unused colors be $x_1, \ldots, x_{M_1}.$ If $\chi_1$ is odd, 0 must be used.
\item Properly color $\Sigma_2$ with $\pm 1, \ldots, \pm k_2,$ and 0 using $\chi_2 - M_2$ colors. Let the $M_2$ unused colors be $y_1, \ldots, y_{M_2}.$ Note that $0$ must be used.
\item Make the following Type 1 replacements on $\Sigma_2.$
\begin{center}
\begin{tabular}{| c | c |}
\hline
Old Color & New Color \\ \hline
$0$ & $x_{M_1}$ \\ \hline
$-y_1$ & $x_1$ \\ \hline
$-y_2$ & $ x_2$ \\ \hline
$\vdots$ & $\vdots$ \\ \hline
$-y_{M_2}$ & $x_{M_2}$ \\
\hline
\end{tabular}
\end{center}
\item Replace $\frac{M_1 - M_2 - 1}{2}$ more pairs of colors in $\Sigma_2$ using Type 4 replacements and colors $x_{M_2 + 1}, \ldots, x_{M_1 -1}$.
\item Use Type 3 replacements to recolor the remaining pairs of colors in $\Sigma_2$ using $\pm (k_1 + 1), \ldots, \pm(k_1 + k_2 - M_2 - \frac{M_1 - M_2 - 1}{2}).$
\end{enumerate}
This concludes the non-exceptional cases where $M_2 > 0$.
\end{proof}
Lemmas \ref{exception 1}, \ref{exception 2}, \ref{main lemma 0}, and \ref{main lemma non-0} together prove Theorem \ref{main thm}. \hfill\qedsymbol
| db3f1a6ec20d4199961db851bd92a01b1883049c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
%
Since the beginning of the digital control era, periodic sampling has been the standard choice for practitioners on all sorts of applications, due to its simple implementation and the existence of many analysis and design results and tools. However, with the replacement of point-to-point communication channels with networked control systems (NCSs), wireless networks in particular, minimizing control data generation and transmission becomes crucial. Because of this and the fundamental limitations of periodic control, aperiodic methods such as event-triggered control (ETC) have been proposed \cite{tabuada2007event} and gained enormous traction since then.
In ETC, instead of time determining when a sensor should send the data, this is dictated by a triggering condition, typically a function of current and held measurements. From the beginning, ETC has shown immense promise in reducing control communications, and subsequent work was dedicated, among other objectives, to further decrease the number of transmissions generated \cite{wang2008event, girard2015dynamic, heemels2012introduction}. While most event-triggered mechanisms assume that a smart sensory system continuously monitors the designed triggering condition, periodic event-triggered control (PETC, \cite{heemels2013periodic}) considers the case where such sensory system is also digital, and the condition checking is periodic. For this practical reason, this paper focuses on this class of ETC.
Even though numerical simulations provided in much of the literature provide evidence that the promised traffic reduction is substantial \cite{wang2008event, heemels2012introduction, heemels2013periodic}, little is formally known about the traffic patterns generated by (P)ETC. Existing work can be divided into two categories: the first involves creating abstract models of the generated traffic, such as \cite{kolarijani2016formal, mazo2018abstracted, fu2018traffic, gleizer2020scalable}; the second involves inferring asymptotic properties of the trajectories that inter-event times describe \cite{postoyan2019interevent}. In the latter, planar linear systems using the triggering condition from \cite{tabuada2007event} are investigated, and the authors show that, under some conditions on the triggering function, inter-event times either converge to a constant value or to a periodic pattern. Our ultimate goal, however, is designing tools for scheduling, and thus we are more concerned with precise short-term predictions than longer-term tendencies. Consequently, the current piece of work \rev{belongs to} the first category. The work in \cite{kolarijani2016formal, mazo2018abstracted} is dedicated to approximate similar models \cite{tabuada2009verification} of (continuous) ETC traffic using the novel notion of power quotient systems: in this case, the abstraction's states represent regions of the state-space of the ETC system, and the output associated to a discrete state of the abstraction is a time interval, to which the actual inter-event time is guaranteed to belong. This modeling strategy is extended to PETC in \cite{fu2018traffic}. To obtain an exact simulation, \cite{gleizer2020scalable} explores the discrete nature of inter-event times of PETC, leading to a novel quotient system that predicts the exact inter-sample time from any given state. Additionally, the main advantage of the latter in comparison to its predecessors is scalability w.r.t.~state-space dimension. On the downside, the generated models exhibit severe non-determinism, likely due to the small number of states of the abstraction and the relaxations used when computing the transition relation. Therefore, even though its one-step ahead predictions are exact, after a couple of steps it loses its prediction capability is severely limited.
The present work tackles precisely this longer-term predictability issue. Building upon the quotient model from \cite{gleizer2020scalable}, we develop new abstractions that enumerate all possible sequences of inter-event times until \rev{a Lyapunov sublevel set is reached. Based on this, we propose a modified PETC mixed with periodic control, hereafter denoted MPETC, that initiates with PETC sampling and switches to periodic sampling when the states lie inside the aforementioned sublevel set. The MPETC retains the practical benefits from PETC, while improving traffic predictability; for it, our abstraction constitutes a bisimulation.}
The abstraction is computable because the number of possible sampling sequences generated during the PETC \rev{phase} is finite, and checking whether PETC generates a given sampling sequence is decidable: it is equivalent to the satisfiability of a conjunction of non-convex quadratic inequalities, to which solvers exist (e.g.,~\rev{the satisfiability-modulo-theories (SMT) solver }Z3~\cite{demoura2008z3}). In our symbolic model, each state is associated with a sequence of inter-event times, which is similar in spirit to \cite{lecorronc2013mode}. When generating these finitely many discrete states, exhaustive search can be avoided by employing a recursive algorithm.
The clear advantages of our new model are the exact enumeration of sampling sequences that can be generated by PETC on a significant future horizon and the establishment of tight bounds on the Lyapunov function convergence speed. Naturally, a disadvantage of our presented method is a substantial growth in the number of discrete states when compared to \cite{gleizer2020scalable}.
Finally, we show how to modify our bisimilar model to obtain a tight traffic model simulating an unmodified PETC system, \rev{presenting two derived results: tighter decay rate estimation (compared to, e.g., \cite{heemels2013periodic}), and maximum average triggering frequency computation.}
%
\subsection{Notation}
We denote by $\mathbb{N}_{0}$ the set of natural numbers including zero, $\mathbb{N} \coloneqq \mathbb{N}_{0} \setminus \{0\}$, $\mathbb{N}_{\leq n} \coloneqq \{1,2,...,n\}$, and $\mathbb{R}_+$ the set of non-negative reals. We denote by $\ceil{a}$ the smallest integer not smaller than $a\in\mathbb{R}$. We denote by $|\boldsymbol{x}|$ the norm of a vector $\boldsymbol{x} \in \mathbb{R}^n$, but if $s$ is a sequence or set, $|s|$ denotes its length or cardinality, respectively. For a square matrix $\boldsymbol{A} \in \mathbb{R}^{n \times n},$ we write $\boldsymbol{A} \succ \mathbf{0}$ ($\boldsymbol{A} \succeq \mathbf{0}$) if $\boldsymbol{A}$ is positive definite (semi-definite). The set $\mathbb{S}^n$ denotes the set of symmetric matrices in $\mathbb{R}^n$. For $\boldsymbol{P} \in \mathbb{S}^n$, $\lambda_{\max}(\boldsymbol{P})$ ($\lambda_{\min}(\boldsymbol{P})$) denotes its maximum (minimum) eigenvalue. \rev{For a set $\mathcal{X}\subseteq\Omega$, we denote by $\bar{\mathcal{X}}$ its complement: $\Omega \setminus \mathcal{X}$}.
\rev{For a relation $\mathcal{R} \subseteq \mathcal{X}_a \times \mathcal{X}_b$, its inverse is denoted as} $\mathcal{R}^{-1} = \{(x_b, x_a) \in \mathcal{X}_b \times \mathcal{X}_a : (x_a, x_b) \in \mathcal{R}\}$. Finally, we denote by $\pi_\mathcal{R}(\mathcal{X}_a) \coloneqq \{ x_b \in \mathcal{X}_b | (x_a, x_b) \in \mathcal{R} \text{ for some } x_a \in \mathcal{X}_b\}$ the natural projection $\mathcal{X}_a$ onto $\mathcal{X}_b$.
We say that an autonomous system $\dot{\boldsymbol{\xi}}(t) = f(\boldsymbol{\xi}(t))$ is globally exponentially stable (GES) if there exist $M < \infty$ and $b > 0$ such that every solution of the system satisfies $|\boldsymbol{\xi}(t)| \leq M\mathrm{e}^{-bt}|\boldsymbol{\xi}(0)|$ for every initial state $\boldsymbol{\xi}(0)$; moreover, we call $b$ a decay rate estimate of the system. When convenient, we use $\boldsymbol{\xi}_{\boldsymbol{x}}(t)$ to denote a trajectory from initial state $\boldsymbol{\xi}(0) = \boldsymbol{x}.$
\section{Preliminaries}
\subsection{Periodic event-triggered control}
Consider a linear plant controlled with sample-and-hold state feedback described by
%
\begin{align}
\dot{\boldsymbol{\xi}}(t) &= \boldsymbol{A}\boldsymbol{\xi}(t) + \boldsymbol{B}\boldsymbol{K}\hat{\boldsymbol{\xi}}(t),\label{eq:plant}\\
\boldsymbol{\xi}(0) &= \hat{\boldsymbol{\xi}}(0) = \rev{\boldsymbol{x}_0}, \nonumber
\end{align}
%
where $\boldsymbol{\xi}(t) \in \mathbb{R}^{n_{\mathrm{x}}}$ is the plant's state with initial value $\rev{\boldsymbol{x}_0}$, $\hat{\boldsymbol{\xi}}(t) \in \mathbb{R}^{n_{\mathrm{x}}}$ is the measurement of the state available to the controller, $\boldsymbol{K}\hat{\boldsymbol{\xi}}(t) \in \mathbb{R}^{n_{\mathrm{u}}}$ is the control input, ${n_{\mathrm{x}}}$ and ${n_{\mathrm{u}}}$ are the state-space and input-space dimensions, respectively, and $\boldsymbol{A}, \boldsymbol{B}, \boldsymbol{K}$ are matrices of appropriate dimensions.
%
The holding mechanism is zero-order: let $t_i \in \mathbb{R}_+, i \in \mathbb{N}_0$ be a sequence of sampling times, with $t_0 = 0$ and $t_{i+1} - t_i > \varepsilon$ for some $\varepsilon > 0$; then $\hat{\boldsymbol{\xi}}(t) = \boldsymbol{\xi}(t_i), \forall t \in [t_i, t_{i+1})$.
In ETC, a \emph{triggering condition} determines the sequence of times $t_i$. In PETC, this condition is checked only periodically, with a fundamental checking period $h$. We consider the family of static \emph{quadratic triggering conditions} from \cite{heemels2013periodic} with an additional maximum inter-event time condition below:
%
\begin{equation}\label{eq:quadtrig}
t_{i+1} = \inf\left\{\!kh>t_i, k \in \mathbb{N} \middle|
\!\begin{array}{c}
\begin{bmatrix}\boldsymbol{\xi}(kh) \\ \rev{\boldsymbol{\xi}(t_i)}\end{bmatrix}\tran
\!\boldsymbol{Q} \begin{bmatrix}\boldsymbol{\xi}(kh) \\ \rev{\boldsymbol{\xi}(t_i)}\end{bmatrix} > 0\! \\
\rev{\text{ or }} \ kh-t_i \leq \bar{k}h\phantom{\dot{\hat{I}}}
\end{array}\!
\right\}\!,
\end{equation}
where $\boldsymbol{Q} \in \mathbb{S}^{2{n_{\mathrm{x}}}}$ is the designed triggering matrix, and $\bar{k}$ is the chosen maximum (discrete) inter-event time.\footnote{Often a maximum inter-event time arises naturally from the closed-loop system itself (see \cite{gleizer2018selftriggered}). Still, one may want to set a smaller maximum inter-event time so as to establish a ``heart beat'' of the system.} Many of the triggering conditions available in the literature can be written as in Eq.~\eqref{eq:quadtrig}; the interested reader may refer to \cite{heemels2013periodic} for a comprehensive list of triggering and stability conditions.
As noted in \cite{gleizer2020scalable}, the inter-event time $t_{i+1} - t_{i}$ is a function of \rev{$\boldsymbol{x}_i \coloneqq \boldsymbol{\xi}(t_i)$}; denoting $\kappa \coloneqq (t_{i+1}-t_i)/h$ as the discrete inter-event time, from Eq.~\eqref{eq:quadtrig} it follows that
%
\begin{gather}
\kappa(\rev{\boldsymbol{x}_i}) = \min\left\{k \in \{1, 2, ...\bar{k}\}: \rev{\boldsymbol{x}_i}\tran\boldsymbol{N}(k)\rev{\boldsymbol{x}_i} > 0 \rev{\text{ or }} k=\bar{k}\right\}, \nonumber\\
\boldsymbol{N}(k) \coloneqq \begin{bmatrix}\boldsymbol{M}(k) \\ \mathbf{I}\end{bmatrix}\tran
\boldsymbol{Q} \begin{bmatrix}\boldsymbol{M}(k) \\ \mathbf{I}\end{bmatrix}, \label{eq:petc_time}\\
\!\!\boldsymbol{M}(k) \coloneqq \boldsymbol{A}_\ensuremath{\mathrm{d}}(k) + \boldsymbol{B}_\ensuremath{\mathrm{d}}(k)\boldsymbol{K} \coloneqq \mathrm{e}^{\boldsymbol{A} hk} + \int_0^{hk}\mathrm{e}^{\boldsymbol{A}\tau}\ensuremath{\mathrm{d}}\tau \boldsymbol{B}\boldsymbol{K}.\!\!\nonumber
\end{gather}
%
where $\mathbf{I}$ denotes the identity matrix.
\subsection{Transition systems}
We use the framework of \cite{tabuada2009verification} to formally relate systems of different natures, e.g., those described by differential equations with those described by finite-state machines. First, a generalized notion of transition system is given:
%
\begin{defn}[Transition System \cite{tabuada2009verification}]\label{def:system}
A system $\mathcal{S}$ is a tuple $(\mathcal{X},\mathcal{X}_0,\mathcal{U},\Es,\mathcal{Y},H)$ where:
\begin{itemize}
\item $\mathcal{X}$ is the set of states,
\item $\mathcal{X}_0 \subseteq \mathcal{X}$ is the set of initial states,
\item $\mathcal{U}$ is the set of inputs,
\item $\Es \subseteq \mathcal{X} \times \mathcal{U} \times \mathcal{X}$ is the set of edges (or transitions),
\item $\mathcal{Y}$ is the set of outputs, and
\item $H: \mathcal{X} \to \mathcal{Y}$ is the output map.
\end{itemize}
\end{defn}
%
A system is said to be finite (infinite) state when the cardinality of $\mathcal{X}$ is finite (infinite). A system is autonomous if $\mathcal{U} = \emptyset$, in which case a transition is denoted by a pair $(x, x') \in \mathcal{X}\times\mathcal{X}$ instead of a triplet. Hereafter, we focus on autonomous systems. For these cases, we define $\Post_\mathcal{S}(x) \coloneqq \{x'| (x,x') \in \Es\}$.
We call $x_0 \to x_1 \to x_2 \to ...$ an \emph{infinite internal behavior} of $\mathcal{S}$ if $x_0 \in \mathcal{X}_0$ and $(x_i,x_{i+1}) \in \Es$, and $y_0 \to y_1 \to ...$ its corresponding \emph{infinite external behavior}, or \emph{trace}, if $H(x_i) = y_i$.
The concept of simulation relation is fundamental for relating two transition systems.
\begin{defn}[Simulation Relation \cite{tabuada2009verification}]\label{def:sim}
Consider two systems $\mathcal{S}_a$ and $\mathcal{S}_b$ with $\mathcal{Y}_a$ = $\mathcal{Y}_b$. A relation $\mathcal{R} \subseteq \mathcal{X}_a \times \mathcal{X}_b$ is a simulation relation from $\mathcal{S}_a$ to $\mathcal{S}_b$ if the following conditions are satisfied:
\begin{enumerate}
\item[i)] for every $x_{a0} \in \mathcal{X}_{a0}$, there exists $x_{b0} \in \mathcal{X}_{b0}$ with $(x_{a0}, x_{b0}) \in \mathcal{R};$
\item[ii)] for every $(x_a, x_b) \in \mathcal{R}, H_a(x_a) = H_b(x_b);$
\item[iii)] for every $(x_a, x_b) \in \mathcal{R},$ we have that $(x_a, x_a') \in \Es_a$ implies the existence of $(x_b, x_b') \in \Es_b$ satisfying $(x_a', x_b') \in \mathcal{R}.$
\end{enumerate}
\end{defn}
%
We say $\mathcal{S}_a \preceq \mathcal{S}_b$ when $\mathcal{S}_b$ simulates $\mathcal{S}_a$.
If $\mathcal{S}_a \preceq \mathcal{S}_b$, from the definition above, it becomes clear that any sequence of outputs from $\mathcal{S}_a$ can be generated by $\mathcal{S}_b$; the converse is not true, unless there is in fact a bisimulation
%
\begin{defn}[Bisimulation \cite{tabuada2009verification}]
Consider two systems $\mathcal{S}_a$ and $\mathcal{S}_b$ with $\mathcal{Y}_a$ = $\mathcal{Y}_b$.
$\mathcal{S}_a$ is said to be bisimilar to $\mathcal{S}_b$, denoted $\mathcal{S}_a \ensuremath{\cong} \mathcal{S}_b$, if there exists a relation
$\mathcal{R}$ such that:
\begin{itemize}
\item $\mathcal{R}$ is a simulation relation from $\mathcal{S}_a$ to $\mathcal{S}_b$;
\item $\mathcal{R}^{-1}$ is a simulation relation from $\mathcal{S}_b$ to $\mathcal{S}_a$.
\end{itemize}
\end{defn}
\section{Problem formulation}
Consider system \eqref{eq:plant}--\eqref{eq:quadtrig}, a quadratic Lyapunov function $V(\boldsymbol{x}) = \boldsymbol{x}\tran\boldsymbol{P}\boldsymbol{x}, \boldsymbol{P} \succ \mathbf{0}$, and the following assumptions:
%
\begin{assum}\label{assum:stable} System \eqref{eq:plant}--\eqref{eq:quadtrig} is GES, and there exists a known constant $0 \leq a < 1$ such that every solution of the system satisfies $V(\boldsymbol{\xi}(t_{i+1})) \leq aV(\boldsymbol{\xi}(t_i))$.
\end{assum}
\begin{rem}\label{rem:a_computation}
To compute $a$, one can verify the implication $$\forall \boldsymbol{x} \!\in\! \mathbb{R}^{n_{\mathrm{x}}} (\forall i \in \mathbb{N}_{< k} \, \boldsymbol{x}\tran\boldsymbol{N}(i)\boldsymbol{x} \leq 0) \wedge (\boldsymbol{x}\tran\boldsymbol{N}(k)\boldsymbol{x} > 0) \implies V(\boldsymbol{M}(k)\boldsymbol{x}) \leq a V(\boldsymbol{x})$$ for every $k\in\{1,...\bar{k}\}$. This can be cast as a set of LMIs through the S-procedure.
\end{rem}
Note that, from Eq.~\eqref{eq:petc_time}, no triggering can occur at $k$ if $\boldsymbol{N}(k) \preceq 0$. Thus, we can determine the global minimum inter-sample time as $h\underline{k}$, with $\underline{k} \!\coloneqq\! \min\{k\! \in\! \{1,...,\bar{k}\}: \boldsymbol{N}(k)\! \npreceq\! \mathbf{0}\}$.
\begin{assum}\label{assum:periodicstable} For system \eqref{eq:plant}, there exists some $h_{{\mathrm{P}}} > 0$ such that the periodic sampling sequence with $t_{i+1} = t_i + h_{{\mathrm{P}}}$ ensures $V(\boldsymbol{\xi}(t_{i+1}) \leq V(\boldsymbol{\xi}(t_i))$.
\end{assum}
This will not necessarily follow from Assumption \ref{assum:stable}; however, ETC is typically designed based on a continuous-time Lyapunov function, and for small enough values of $h$, the same Lyapunov function will work for periodic control.%
%
\footnote{This is easy to see when one considers the first order approximation of the discrete-time transition matrix $\mathrm{e}^{\boldsymbol{A} h} \approxeq \mathbf{I} + \boldsymbol{A} h$. If the continuous-time Lyapunov inequality $\boldsymbol{A}\tran\boldsymbol{P} + \boldsymbol{P}\boldsymbol{A} \preceq -\epsilon\mathbf{I}$ holds for some $\epsilon > 0$ then: $h(\boldsymbol{A}\tran\boldsymbol{P} + \boldsymbol{P}\boldsymbol{A}) \preceq -h\epsilon\mathbf{I}
\iff (\mathbf{I} + \boldsymbol{A} h)\tran\boldsymbol{P}(\mathbf{I} + \boldsymbol{A} h) - \boldsymbol{P} \preceq -h\epsilon\mathbf{I} + h^2\boldsymbol{A}\tran\boldsymbol{P}\boldsymbol{A}$, which for sufficiently small $h$ results in $\mathrm{e}^{\boldsymbol{A}\tran h}\boldsymbol{P}\mathrm{e}^{\boldsymbol{A} h} - \boldsymbol{P} \preceq -h\epsilon\mathbf{I} \prec \mathbf{0}$, i.e. the discrete-time Lyapunov inequality.}
\begin{assum}\label{assum:compact} A value $V_0 > 0$ is known such that $\boldsymbol{\xi}(0) \in \mathcal{X}_0 = \{\boldsymbol{x}\in\mathbb{R}^{n_\plant}: V(\boldsymbol{x}) \leq V_0\}$.
\end{assum}
Now let us propose a modification to the PETC system. \rev{Since ETC can reduce communication frequency while ensuring a fast decay rate, it makes practical sense to focus on ETC during the transient phase. However, once states are close enough to the origin, decay rates have disputable practical relevance. Therefore, we admit that, when $\hat{\boldsymbol{\xi}}(t)$ enters a small sublevel set $\mathcal{X}_{{\mathrm{P}}} \coloneqq \{\boldsymbol{x}\in\mathbb{R}^{n_\plant}| V(x) \leq rV_0\}, r < 1$, the controller can switch to periodic sampling, \rev{with $h_{{\mathrm{P}}}$ significantly bigger than $h$}; in fact, it can be as big as possible, provided that it preserves Assumption \ref{assum:periodicstable}. This results in more predictable (hence schedulable) traffic while retaining a reduction of traffic.}
Let us denote by $t_{i+1}(t_i, \boldsymbol{\xi}(t_i))|_{\mathrm{PETC}}$ the solution of Eq.~\eqref{eq:quadtrig}. Mathematically, the mixed sampling strategy\rev{, hereafter denoted MPETC,} dictates the sampling times as follows:
%
\begin{equation}\label{eq:mixedtiming}
\begin{aligned}
t_{i+1} &= t_{i+1}(t_i, \boldsymbol{\xi}(t_i))|_{\mathrm{PETC}}, && V(\boldsymbol{\xi}(t_i)) > rV_0 \\
t_{i+1} &= t_i + h_{{\mathrm{P}}}, && V(\boldsymbol{\xi}(t_i)) \leq rV_0.
\end{aligned}
\end{equation}
Hereafter, denote $\mathcal{X}_{{\mathrm{P}}} \coloneqq \{\boldsymbol{x} \in \mathbb{R}^{n_{\mathrm{x}}}|V(\boldsymbol{x}) \leq rV_0\} = r\mathcal{X}_0$.
This system has the following infinite-state traffic model: $\mathcal{S}_{{\mathrm{E}}} \coloneqq (\mathcal{X}, \mathcal{X}_0, \emptyset, \Es_{\mathrm{PETC}} \cup \Es_{\mathrm{P}}, \mathcal{Y}_{{\mathrm{E}}}, H_{{\mathrm{E}}})$ where
%
\begin{multline}\label{eq:original}
\begin{aligned}
\mathcal{X} &= \mathcal{X}_0; \\
\Es_{\mathrm{PETC}} &= \{(\boldsymbol{x}, \boldsymbol{x}') \in (\mathcal{X} \setminus \mathcal{X}_{{\mathrm{P}}}) \times \mathcal{X}: \boldsymbol{x}' = \boldsymbol{\xi}_{\boldsymbol{x}}(h\kappa(\boldsymbol{x}))\}; \\
\Es_{\mathrm{P}} &= \{(\boldsymbol{x}, \boldsymbol{x}') \in \mathcal{X}_{{\mathrm{P}}} \times \mathcal{X}_{{\mathrm{P}}}: \boldsymbol{x}' = \boldsymbol{\xi}_{\boldsymbol{x}}(h_{{\mathrm{P}}})\}; \\
\mathcal{Y}_{{\mathrm{E}}} &= \{h, 2h, ..., \bar{k}h, h_{{\mathrm{P}}}\};\\
H_{{\mathrm{E}}}(\boldsymbol{x}) &= \begin{cases} h\kappa(x), & \boldsymbol{x} \in \mathcal{X} \setminus \mathcal{X}_{{\mathrm{P}}}, \\
h_{{\mathrm{P}}}, & \boldsymbol{x} \in \mathcal{X}_{{\mathrm{P}}}.
\end{cases}
\end{aligned}
\raisetag{3\baselineskip}
\end{multline}
%
For states starting outside $\mathcal{X}_{{\mathrm{P}}}$, transitions and outputs (the inter-sample times) are dictated by the PETC strategy; for states inside $\mathcal{X}_{{\mathrm{P}}}$, transitions and outputs are dictated by periodic sampling. Note that $\Es_{\mathrm{P}}$ is defined over $\mathcal{X}_{{\mathrm{P}}} \times \mathcal{X}_{{\mathrm{P}}}$, i.e., states starting in $\mathcal{X}_{{\mathrm{P}}}$ always land in $\mathcal{X}_{{\mathrm{P}}}$: this comes from the fact that the periodic \rev{phase} is forward-invariant due to Assumption \ref{assum:periodicstable}.
%
We are ready to define our main problem:
%
\begin{prob}\label{prob:bisim} Considering Assumptions \ref{assum:stable}--\ref{assum:compact}, determine if $\mathcal{S}_{{\mathrm{E}}}$ admits a computable finite-state bisimulation. If so, provide an algorithm to compute it.
\end{prob}
\section{Main result}
To build a bisimilar model of $\mathcal{S}_{{\mathrm{E}}}$, the main observation is that eventually all trajectories of the system \eqref{eq:plant}, \eqref{eq:mixedtiming} enter $\mathcal{X}_{{\mathrm{P}}}$, which follows from Assumption \ref{assum:stable}. Clearly, when in $\mathcal{X}_{{\mathrm{P}}}$ the system admits a trivial, single-state traffic bisimulation:
%
\begin{prop}\label{prop:periodic_bisim}
Define $H_{\mathrm{P}}: \mathbb{R}^{n_{\mathrm{x}}} \to \mathbb{R}$ such that $H_{\mathrm{P}} \equiv h_{{\mathrm{P}}}$. The system $$\mathcal{S}_{\mathrm{P}}^{\mathrm{B}} = (\{\mathcal{X}_{{\mathrm{P}}}\}, \{\mathcal{X}_{{\mathrm{P}}}\}, \emptyset, \{(\mathcal{X}_{{\mathrm{P}}}, \mathcal{X}_{{\mathrm{P}}})\}, \{h_{{\mathrm{P}}}\}, H_{\mathrm{P}})$$ is a bisimilar quotient system of $(\mathcal{X}_{{\mathrm{P}}}, \mathcal{X}_{{\mathrm{P}}}, \emptyset, \Es_{\mathrm{PETC}} \cup \Es_{\mathrm{P}}, \mathcal{Y}_{{\mathrm{E}}}, H_{{\mathrm{E}}})$.
\end{prop}
Another important observation is that, since the PETC \rev{phase} is asymptotically stable (Assumption \ref{assum:stable}), states from $\mathcal{X}_0$ reach $\mathcal{X}_{{\mathrm{P}}}$ in finite time. Thus, for any state in $\mathcal{X}_0$, there is a finite number of PETC-generated samples, after which all samples are periodically taken. Let $K \coloneqq \{\underline{k}, \underline{k}+1,...,\bar{k}\}$; since at each step there are finitely many ($|K|$) inter-sample time possibilities, we can state the following:
%
\begin{lem}\label{lem:finite} Let Assumptions \ref{assum:stable}--\ref{assum:compact} hold, define $N \coloneqq \ceil{\log_a(r)}$. Then system \eqref{eq:original} can produce at most $|K|(\left(|K|-1\right)^N-1)(|K|-1)^{-1}$ different traces.
\end{lem}
%
\begin{proof}
Using assumption \ref{assum:stable}, recursively apply $V(\boldsymbol{\xi}(t_{i+1}) \leq aV(\boldsymbol{\xi}(t_i))$ to get $V(\boldsymbol{\xi}(t_N) \leq a^NV(\rev{\boldsymbol{x}_0}) \leq a^NV_0.$ Then, $N > \log_a(r)$ implies $a^NV_0 \leq rV_0$; thus, it takes at most $N$ steps to enter $\mathcal{X}_{{\mathrm{P}}}$. After this, from Proposition \ref{prop:periodic_bisim}, the remaining trace is $h_{{\mathrm{P}}},h_{{\mathrm{P}}},...$. This is the trace if $\rev{\boldsymbol{x}_0}\in\mathcal{X}_{{\mathrm{P}}}$, which accounts for one trace; $\mathcal{S}_{{\mathrm{E}}}$ has at most $|K|$ traces for which it takes one step to reach $\mathcal{X}_{{\mathrm{P}}}$ from $\rev{\boldsymbol{x}_0}$, at most $|K|^2$ traces for which it takes two steps to reach $\mathcal{X}_{{\mathrm{P}}}$, and etc., up to $|K|^N$ for the maximum number of steps. Summing up this geometric series gives $|K|(\left(|K|-1\right)^N-1)(|K|-1)^{-1}$.
\end{proof}
Lemma \ref{lem:finite} permits the construction of a rather straightforward finite-state model similar to $\mathcal{S}_{{\mathrm{E}}}$. Denote by $K^m$ the set of all sequences of length $m$ of the form $(k_i)_{i=0}^m, k_i \in K.$ We create one state for each sequence in $K^m$. The state ${k_1k_2...k_m}$ is associated with the trace $hk_1, hk_2, ..., hk_m, h_{{\mathrm{P}}}, h_{{\mathrm{P}}}, ...$, thus taking $m$ samples to enter the periodic \rev{phase}. By definition, its successor must be ${k_2...k_m}$. Finally, let $\varepsilon$ denote the empty sequence; a state ${k}$ generates the trace $hk, h_{{\mathrm{P}}}, h_{{\mathrm{P}}}, ...$, and thus its successor is $\varepsilon$, associated with the periodic \rev{phase}. Hence, $\Post(\varepsilon) = \varepsilon$ and $H^{{\mathrm{S}}}(\varepsilon) = h_{{\mathrm{P}}}$. Let $\mathbb{K}_N \coloneqq \cup_{i=1}^NK^i \cup \{\varepsilon\}$; we consolidate this modeling strategy with the following result:
%
\begin{prop}\label{prop:simtrivial}
Let Assumptions \ref{assum:stable}--\ref{assum:compact} hold and $N \coloneqq \ceil{\log_a(r)}$. Consider $\mathcal{S}^{{\mathrm{S}}} \!\coloneqq\! \left(\mathbb{K}_N, \mathbb{K}_N, \emptyset, \Es^{{\mathrm{S}}}, \mathcal{Y}_{{\mathrm{E}}}, H^{{\mathrm{S}}}\right)\!$ with
\begin{itemize}
\item $\Es^{{\mathrm{S}}} = \{({k\sigma}, \sigma)|{k\sigma}\in\mathbb{K}_N\} \cup \{(\varepsilon, \varepsilon)\}$;
\item $H^{{\mathrm{S}}}({k\sigma}) = hk$ and $H^{{\mathrm{S}}}(\varepsilon) = h_{{\mathrm{P}}}$.
\end{itemize}
%
Then $\mathcal{S}^{{\mathrm{S}}} \succeq \mathcal{S}_{{\mathrm{E}}}$.
\end{prop}
\begin{proof}
System $\mathcal{S}^{{\mathrm{S}}}$ generates all possible traces of type $hk_1, hk_2, ..., hk_m,\allowbreak h^*, h^*, ...,$ for $0 \leq m \leq N$, which, according to Lemma \ref{lem:finite}, include all possible traces of $\mathcal{S}_{{\mathrm{E}}}$; thus, the behavior of $\mathcal{S}^{{\mathrm{S}}}$ contains that of $\mathcal{S}_{{\mathrm{E}}}$. Because both systems $\mathcal{S}_{{\mathrm{E}}}$ and $\mathcal{S}^{{\mathrm{S}}}$ are deterministic and non-blocking, this implies that $\mathcal{S}^{{\mathrm{S}}} \succeq \mathcal{S}_{{\mathrm{E}}}$ \cite[Proposition 4.11]{tabuada2009verification}.
\end{proof}
The set $\mathbb{K}_N$ includes sequences that may not be generated by the PETC \rev{phase}. To trim off these spurious sequences, let us define the following relation:
%
\begin{defn}[Inter-sample sequence relation]\label{def:bisimrel} We denote by $\mathcal{R}_{{\mathrm{B}}} \subseteq \mathcal{X} \times \mathbb{K}_N$ the relation satisfying
\begin{equation}\label{eq:Xempty}
(\boldsymbol{x},\varepsilon) \in \mathcal{R}_{{\mathrm{B}}} \text{ iff } \boldsymbol{x} \in \mathcal{X}_{{\mathrm{P}}},
\end{equation}
%
and
$(\boldsymbol{x},{k_1k_2...k_m}) \in \mathcal{R}_{{\mathrm{B}}}$ if and only if
%
%
\begin{subequations}\label{eq:sequence}
\begin{align}
\boldsymbol{x} &\in \mathcal{X}_0, \label{eq:sequence_x0}\\
\boldsymbol{x} &\in \mathcal{Q}_{k_1}, \label{eq:sequence_k1}\\
\boldsymbol{M}(k_1)\boldsymbol{x} &\in \mathcal{Q}_{k_2}, \label{eq:sequence_k2}\\
\boldsymbol{M}(k_2)\boldsymbol{M}(k_1)\boldsymbol{x} &\in \mathcal{Q}_{k_3}, \label{eq:sequence_k3}\\
& \vdots \nonumber\\
\boldsymbol{M}(k_{m-1})...\boldsymbol{M}(k_1)\boldsymbol{x} &\in \mathcal{Q}_{k_m}, \label{eq:sequence_km-1}\\
\boldsymbol{x} &\notin \mathcal{X}_{{\mathrm{P}}}, \label{eq:sequence_x0_not_periodic}\\
\boldsymbol{M}(k_1)\boldsymbol{x} &\notin \mathcal{X}_{{\mathrm{P}}}, \label{eq:sequence_x1_not_periodic}\\
& \vdots \nonumber\\
\boldsymbol{M}(k_{m-1})...\boldsymbol{M}(k_1)\boldsymbol{x} &\notin \mathcal{X}_{{\mathrm{P}}}, \label{eq:sequence_xm-1_not_periodic} \\
\boldsymbol{M}(k_m)...\boldsymbol{M}(k_1)\boldsymbol{x} &\in \mathcal{X}_{{\mathrm{P}}}, \label{eq:sequence_km}
\end{align}
\end{subequations}
%
where
%
\begin{equation}\label{eq:setq}
\begin{gathered}
\mathcal{Q}_k \coloneqq \mathcal{K}_k \setminus \left(\bigcap_{j=\underline{k}}^{k-1} \mathcal{K}_{j}\right) = \mathcal{K}_k \cap \bigcap_{j=\underline{k}}^{k-1} \bar{\mathcal{K}}_{j}, \\
\mathcal{K}_k \coloneqq \begin{cases}
\{\boldsymbol{x} \in \mathcal{X}| \boldsymbol{x}\tran\boldsymbol{N}(k)\boldsymbol{x} > 0\}, & k < \bar{k}, \\
\mathcal{X}, & k = \bar{k}.
\end{cases}
\end{gathered}
\end{equation}
%
\end{defn}
Eq.~\eqref{eq:setq}, taken from \cite{gleizer2020scalable}, defines the sets $\mathcal{Q}_k$, containing the states that trigger exactly with inter-sample time $hk$. Eq.~\eqref{eq:Xempty} determines that states $\boldsymbol{x} \in \mathcal{X}_{\mathrm{P}}$ are related to the state $\varepsilon$. Finally, a state $\boldsymbol{x} \in \mathbb{R}^n$ is related to a state $k_1k_2...k_m$ of the abstraction if the following are satisfied: 1) it belongs to the compact set of interest (Eq.~\eqref{eq:sequence_x0}), 2) the inter-sample time sequence that it generates up until it enters $\mathcal{X}_{\mathrm{P}}$ is $hk_1,hk_2,...,kh_m$ (Eqs.~\eqref{eq:sequence_k1}--\eqref{eq:sequence_km-1}), and 3) the sampled states $\boldsymbol{\xi}_{\boldsymbol{x}}(k_1h), \boldsymbol{\xi}_{\boldsymbol{x}}((k_1+k_2)h), ...$ of the trajectory starting from $\boldsymbol{x}$ do not belong to $\mathcal{X}_{\mathrm{P}}$ (Eqs.~\eqref{eq:sequence_x0_not_periodic}--\eqref{eq:sequence_xm-1_not_periodic}), while the m-th sampled state does belong to $\mathcal{X}_{\mathrm{P}}$ (Eq.~\eqref{eq:sequence_km}).
We now employ the relation $\mathcal{R}_{{\mathrm{B}}}$ to derive a finite model bisimilar to $\mathcal{S}_{{\mathrm{E}}}$ as follows:
\begin{defn}\label{def:bisim} The MPETC traffic model is the system $$\mathcal{S}^{{\mathrm{B}}} \coloneqq \left(\mathcal{X}^{{\mathrm{B}}}, \mathcal{X}^{{\mathrm{B}}}, \emptyset, \Es^{{\mathrm{S}}}, \mathcal{Y}_{{\mathrm{E}}}, H^{{\mathrm{S}}}\right)$$ with $\mathcal{X}^{{\mathrm{B}}} \coloneqq \pi_{\mathcal{R}_{{\mathrm{B}}}}(\mathcal{X})$.
\end{defn}
This model is a subset of $\mathcal{S}^{{\mathrm{S}}}$, generating only inter-sample sequences that can be produced by the concrete system $\mathcal{S}_{{\mathrm{E}}}$. Topologically, it is still a tree, such as $\mathcal{S}^{{\mathrm{S}}}$, but with fewer states (see Figure \ref{fig:abstraction}). Our main result follows:
\begin{figure}[tb]
\begin{center}
\begin{footnotesize}
\begin{tikzpicture}[->,>=stealth',shorten >=1pt, auto, node distance=1.5cm,
semithick, xscale=0.8, yscale=0.9]
\tikzset{every state/.style={minimum size=2em, inner sep=2pt}}
\node[state] (0) {$\varepsilon$};
\node[state] (1) at ([shift=({135:1.2 cm})]0) {1};
\node[state] (2) at ([shift=({45:1.2 cm})]0) {2};
\node[state] (3) at ([shift=({-45:1.2 cm})]0) {3};
\node[state] (4) at ([shift=({-135:1.2 cm})]0) {4};
\node[state] (11) at ([shift=({195:1.2 cm})]1) {1,1};
\node[state] (21) at ([shift=({155:1.2 cm})]1) {2,1};
\node[state] (31) at ([shift=({115:1.2 cm})]1) {3,1};
\node[state] (41) at ([shift=({75:1.2 cm})]1) {4,1};
\node[state] (12) at ([shift=({105:1.2 cm})]2) {1,2};
\node[state] (22) at ([shift=({65:1.2 cm})]2) {2,2};
\node[state] (32) at ([shift=({25:1.2 cm})]2) {3,2};
\node[state, dashed] (42) at ([shift=({-15:1.2 cm})]2) {4,2};
\node[state] (13) at ([shift=({15:1.2 cm})]3) {1,3};
\node[state, dashed] (23) at ([shift=({-25:1.2 cm})]3) {2,3};
\node[state] (33) at ([shift=({-65:1.2 cm})]3) {3,3};
\node[state, dashed] (43) at ([shift=({-105:1.2 cm})]3) {4,3};
\node[state, dashed] (14) at ([shift=({-75:1.2 cm})]4) {1,4};
\node[state, dashed] (24) at ([shift=({-115:1.2 cm})]4) {2,4};
\node[state, dashed] (34) at ([shift=({-155:1.2 cm})]4) {3,4};
\node[state] (44) at ([shift=({-195:1.2 cm})]4) {4,4};
\path (11) edge (1) (1) edge (0)
(21) edge (1)
(31) edge (1)
(41) edge (1)
(12) edge (2) (2) edge (0)
(22) edge (2)
(32) edge (2)
(42) edge (2)
(13) edge (3) (3) edge (0)
(23) edge (3)
(33) edge (3)
(43) edge (3)
(14) edge (4) (4) edge (0)
(24) edge (4)
(34) edge (4)
(44) edge (4)
(0) edge [loop above] (0);
\end{tikzpicture} \quad \quad \quad %
%
\begin{tikzpicture}[->,>=stealth',shorten >=0pt, auto, node distance=1.0cm,
semithick, xscale=1.0, yscale=1.0]
\tikzset{every state/.style={minimum size=2em, inner sep=2pt}}
\node[state] (44) {4,4};
\node[state] (41) [right of=44] {4,1};
\node[state] (11) [below right of=41] {1,1};
\node[state] (12) [right of=11] {1,2};
\node[state] (13) [below of=11] {1,3};
\node[state] (21) [right of=13] {2,1};
\node[state] (31) [left of=13] {3,1};
\node[state] (33) [below of=13] {3,3};
\node[state] (32) [right of=33] {3,2};
\node[state] (22) [right of=21] {2,2};
\path (44) edge [loop above] (44) (44) edge (41)
(41) edge (11) (41) edge [bend right] (13) (41) edge [bend left] (12)
(11) edge (13) (11) edge (12) (11) edge [loop above] (11)
(12) edge (21) (12) edge (22)
(22) edge [loop above] (22) (22) edge (21)
(31) edge (11) (31) edge (12) (31) edge [bend right] (13)
(13) edge (31) (13) edge (33)
(21) edge [bend left] (12) (21) edge (13) (21) edge (11)
(33) edge (31) (33) edge [loop below] (33) (33) edge (32)
(32) edge (21) (32) edge (22);
\end{tikzpicture}
\end{footnotesize}
\caption{\label{fig:abstraction} On the left, an illustration of $\mathcal{S}^{{\mathrm{S}}}$ (all states) and $\mathcal{S}^{{\mathrm{B}}}$ (only solid-line states). On the right, a depiction of $\mathcal{S}^{{\mathrm{S'}}}$ with $\mathcal{X}^{{\mathrm{S'}}} = \{\sigma \in \mathcal{X}^{{\mathrm{B}}}: |\sigma|=2\}$.}
\vspace{-1.5em}
\end{center}
\end{figure}
\begin{thm}\label{thm:bisim}
Let Assumptions \ref{assum:stable}--\ref{assum:compact} hold and $N \coloneqq \ceil{\log_a(r)}$. Then, $\mathcal{S}^{{\mathrm{E}}} \approxeq \mathcal{S}^{{\mathrm{B}}}$.
\end{thm}
%
\begin{proof}
We show that $\mathcal{R}_{{\mathrm{B}}}$ is a simulation relation from $\mathcal{S}_{{\mathrm{E}}}$ to $\mathcal{S}^{{\mathrm{B}}}$ and $\mathcal{R}^{-1}_{{\mathrm{B}}}$ is a simulation relation from $\mathcal{S}^{{\mathrm{B}}}$ to $\mathcal{S}_{{\mathrm{E}}}$, checking each of the conditions of Definition \ref{def:sim}.
{\bf Step 1:} $\mathcal{R}_{{\mathrm{B}}}$ is a simulation relation from $\mathcal{S}_{{\mathrm{E}}}$ to $\mathcal{S}^{{\mathrm{B}}}$.
For condition (i), take a point $\boldsymbol{x}_0 \in \mathcal{X}_0 = \mathcal{X}$. It either belongs to $\mathcal{X}_{{\mathrm{P}}}$, for which Eq.~\eqref{eq:Xempty} provides its related state; or it takes $m$ PETC steps to reach $\mathcal{X}_{{\mathrm{P}}}$. In this latter case, it generates some trace $hk_1, hk_2, ..., hk_m, h^*, h^*, ...$ and therefore, by definition, it satisfies Eq.~\eqref{eq:sequence}. Hence, the related state ${k_1k_2...k_m}$ belongs to $\mathcal{X}^{{\mathrm{B}}}.$
%
Condition (ii) trivially holds by the definition of $\mathcal{R}_{{\mathrm{B}}}$, and in particular by \eqref{eq:Xempty} and \eqref{eq:sequence_k1}.
Finally, for condition (iii), take $(\boldsymbol{x}, \sigma) \in \mathcal{R}_{{\mathrm{B}}}.$ If $\boldsymbol{x} \in \mathcal{X}_{{\mathrm{P}}}$, then $\sigma = \varepsilon$. From Assumption \ref{assum:periodicstable}, $\Post_{\mathcal{S}_{{\mathrm{E}}}}(\boldsymbol{x}) \in \mathcal{X}_{{\mathrm{P}}}$, which is related to $\varepsilon = \Post_{\mathcal{S}^{\mathrm{B}}}(\varepsilon)$. If $\boldsymbol{x} \notin \mathcal{X}_{{\mathrm{P}}}$, then $\sigma=k_1\sigma' \in \mathbb{K}_N$. Therefore, $\Post_{\mathcal{S}_{{\mathrm{E}}}}(\boldsymbol{x}) = \boldsymbol{M}(k_1)\boldsymbol{x}.$ From Assumption \ref{assum:stable}, $\boldsymbol{M}(k_1)\boldsymbol{x} \in \mathcal{X}_0$; also, by inspecting Eq.~\eqref{eq:sequence}, $\boldsymbol{M}(k_1)\boldsymbol{x}$ satisfies Eqs.~\eqref{eq:sequence_k2}--\eqref{eq:sequence_km} and Eqs.~\eqref{eq:sequence_x1_not_periodic}--\eqref{eq:sequence_xm-1_not_periodic}: this implies that $k_2...k_m = \sigma' = \Post_{\mathcal{S}^{\mathrm{B}}}(k_1\sigma')$ is related to $\boldsymbol{M}(k_1)\boldsymbol{x}$.
{\bf Step 2:} $\mathcal{R}^{-1}_{{\mathrm{B}}}$ is a simulation relation from $\mathcal{S}^{{\mathrm{B}}}$ to $\mathcal{S}_{{\mathrm{E}}}$.
For condition (i), if ${k_1k_2...k_m} \in \mathcal{X}_0^{{\mathrm{B}}}$, then there exists a related initial state $\boldsymbol{x}_0$ which satisfies Eq.~\eqref{eq:sequence}; hence, from Eq.~\eqref{eq:sequence_x0}, $\boldsymbol{x}_0 \in \mathcal{X}_0$. For $\varepsilon$, any related state $\boldsymbol{x}_0$ belongs to $\mathcal{X}_{{\mathrm{P}}} \subset \mathcal{X} = \mathcal{X}_0$.
%
Condition (ii) is the same as in Step 1.
%
Finally, condition (iii) is verified because the reasoning in Step 1 applies to every $\boldsymbol{x} \in \mathcal{X}$ satisfying $(\sigma, \boldsymbol{x}) \in \mathcal{R}^{-1}_{{\mathrm{B}}}\!$.
\end{proof}
\begin{rem}\label{rem:decidable}
Determining %
if there exists $\boldsymbol{x}$ satisfying Eq.~\eqref{eq:sequence}
is a problem of checking non-emptiness of a semi-algebraic set, which has been proven to be decidable \cite{basu2006}. One tool that can be used to solve it is the SMT solver Z3 \cite{demoura2008z3}.
\end{rem}
\begin{prop}[Complexity]
The state set $\mathcal{X}^{{\mathrm{B}}}$ can be computed with $$\bigO\left(|K|^N (N|K|)^{{n_{\mathrm{x}}}}\right)\cdot2^{\bigO({n_{\mathrm{x}}})}$$ operations.
\end{prop}
\begin{proof}
From Lemma \ref{lem:finite}, we have seen that there can be at most $$|K|(\left(|K|-1\right)^N-1)(|K|-1)^{-1}\in \bigO(|K|^N)$$ sampling sequences. Determining the state set $\mathcal{X}^{{\mathrm{B}}}$ requires, in the worst case, checking the existence of all of those sequences. For a sequence of length $m$, Eq.~\eqref{eq:sequence} has one membership in $\mathcal{X}_0$ and $m$ memberships in $\mathcal{X}_{{\mathrm{P}}}$, each corresponding to one quadratic inequality; and $m$ memberships in $\mathcal{Q}_k$, each corresponding to $k-\underline{k}+1$ quadratic inequalities. Therefore, in the worst case, Eq.~\eqref{eq:sequence} has $m+1+m|K|$ inequalities, or $1+N+N|K| \in \bigO(N|K|)$ for the longest sequence. The best known bound for deciding the existence of a real solution to a conjunction of $s$ polynomial inequalities of ${n_{\mathrm{x}}}$ variables and maximum degree $d$ is $s^{{n_{\mathrm{x}}}+1}d^{\bigO({n_{\mathrm{x}}})}$ \cite{basu1996combinatorial}. Replacing $s$ by $1+N+N|K|$ and $d$ by 2, multiplying by the number of checks and working out the limits for big-O notation concludes the proof.
\end{proof}
\begin{rem}\label{rem:properalgorithm} While all sequences of length $N$ must be checked in the worst case, for other cases it is more efficient to employ a recursive algorithm, i.e., verifying Eq.~\eqref{eq:sequence} for sequences from length 1 to $N$. If a sequence $\sigma$ shorter than $N$ does not verify Eq.~\eqref{eq:sequence}, then no sequence $k\sigma$ can do. Hence, many checks can be discarded using this simple observation.
\end{rem}
\begin{rem}\label{rem:homo} Due to characteristics of the inequalities associated to Definition \ref{def:bisimrel}, one can set $V_0 = 1$ without loss of generality, with the only input to the model being the ratio of contraction $r$. For $V_0 = c > 0$, the model is the same: replace $\boldsymbol{x}$ by $\sqrt{c}\boldsymbol{x}$ in Eq.~\eqref{eq:sequence}, and $\sqrt{c}$ can be canceled out. \end{rem}
\subsection{Derived results for the original PETC}
With a few changes to $\mathcal{S}^{{\mathrm{B}}}$, we can build a similar model of the PETC traffic that generates fewer spurious traces than, e.g., \cite{gleizer2020scalable}. This is because the PETC section of the MPETC trace is of course generated by the pure PETC system \eqref{eq:plant}--\eqref{eq:petc_time}. Hence, to simulate the PETC traffic, one could do the following: for a given state $\boldsymbol{x}\in\mathbb{R}^{n_{\mathrm{x}}}$, take $V_0 = V(\boldsymbol{x})$ and determine its related state ${k\sigma}$ from Eq.~\eqref{eq:sequence}. Now take its successor $\boldsymbol{M}(k)\boldsymbol{x}$. Again, set $V_0 = V(\boldsymbol{M}(k)\boldsymbol{x})$ and determine its related state: it has to take the form ${\sigma\sigma'}$, i.e., its first inter-sample times should be all but the first inter-sample times of its predecessor.
This idea is depicted in Fig.~\ref{fig:idea_PETC}.
%
\begin{figure}
\begin{center}
\input{ell_1.tex}
\caption{\label{fig:idea_PETC} Depiction of the strategy used to build the PETC traffic abstraction: the trajectory $\hat{\boldsymbol{\xi}}(t)$ is in orange with samples marked, and $\mathcal{X}_{{\mathrm{P}}}$ is the blue ellipse. The tail of the sequence from $t_0=0$ is the head of the sequence from the following sample $t_1$.}
\vspace{-1.5em}
\end{center}
\end{figure}
%
Let us formalize this procedure.
%
\begin{defn}[PETC inter-sample sequence relation]\label{def:petcrel} Let $V_0 = 1.$ The relation $\mathcal{R}_{{\mathrm{S'}}} \subseteq \mathbb{R}^{{n_{\mathrm{x}}}} \times \mathbb{K}_N$ is given by: $(\boldsymbol{x}, {k_1k_2...k_m}) \in \mathcal{R}_{{\mathrm{S'}}}$ iff $\boldsymbol{x}/\sqrt{V(\boldsymbol{x})}$ satisfies Eq.~\eqref{eq:sequence}.
\end{defn}
%
\begin{thm}
Let Assumption \ref{assum:stable} hold. Then, the system $$\mathcal{S}_{{\mathrm{S'}}} \coloneqq (\mathcal{X}^{{\mathrm{S'}}}, \mathcal{X}^{{\mathrm{S'}}}, \emptyset, \Es^{{\mathrm{S'}}}, \mathcal{Y}_{{\mathrm{E}}}, H^{{\mathrm{S}}}),$$ with $\mathcal{X}^{{\mathrm{S'}}} \coloneqq \pi_{\mathcal{R}_{{\mathrm{S'}}}}(\mathbb{R}^{{n_{\mathrm{x}}}})$ and
$\Es^{{\mathrm{S'}}} = \{({k\sigma}, {\sigma\sigma'}) | {k\sigma}, {\sigma\sigma'} \in \mathcal{X}^{{\mathrm{S'}}}\},$
simulates the traffic generated by System \eqref{eq:plant}--\eqref{eq:quadtrig}.
\end{thm}
%
\begin{proof}
\rev{Take an initial state $\boldsymbol{x}\in\mathbb{R}^{n_{\mathrm{x}}}$,} a PETC trajectory $\boldsymbol{\xi}_{\boldsymbol{x}}(t)$ and its associated inter-sample sequence $k_1k_2...k_m$, after which $V(\boldsymbol{\xi}_{\boldsymbol{x}}(t_m)) \leq rV(\boldsymbol{x})$ but $V(\boldsymbol{\xi}_{\boldsymbol{x}}(t_{m-1})) > rV(\boldsymbol{x})$. This implies that $\boldsymbol{x}/\sqrt{V(\boldsymbol{x})}$ satisfies Eq.~\eqref{eq:sequence}. Hence, $(\boldsymbol{x}/\sqrt{V(\boldsymbol{x})}, {k_1k_2...k_m}) \in \mathcal{R}_{{\mathrm{S'}}}$, and condition (i) of Def.~\ref{def:sim} holds. Condition (ii) is trivially satisfied, as $H(\boldsymbol{x}) = hk_1 = H^{{\mathrm{S}}}(k_1k_2...k_m)$. For condition (iii), take its related sequence $k_1k_2...k_m$. The successor of $\boldsymbol{x}$ is $\boldsymbol{x}' \coloneqq \boldsymbol{\xi}_{\boldsymbol{x}}(hk_1) = \boldsymbol{M}(k_1)\boldsymbol{x}$, which satisfies Eqs.~\eqref{eq:sequence_k2}--\eqref{eq:sequence_km} and Eqs.~\eqref{eq:sequence_x1_not_periodic}--\eqref{eq:sequence_xm-1_not_periodic}; from homogeneity of $\mathcal{Q}_k$, $\boldsymbol{M}(k_1)\boldsymbol{x}/\sqrt{V(\boldsymbol{M}(k_1)\boldsymbol{x})}$ also satisfies Eqs.~\eqref{eq:sequence_k2}--\eqref{eq:sequence_km-1}. Additionally, because of Assumption \ref{assum:stable}, we have that $V(\boldsymbol{x}') < V(\boldsymbol{x})$; hence, Eqs.~\eqref{eq:sequence_x1_not_periodic}--\eqref{eq:sequence_xm-1_not_periodic} holding for $\boldsymbol{x}/\sqrt{V(\boldsymbol{x})}$ imply that $V(\boldsymbol{\xi}_{\boldsymbol{x}}(t_i)) > rV(\boldsymbol{x}) > rV(\boldsymbol{x}')$ for all $1 \leq i \leq m$, and therefore Eqs.~\eqref{eq:sequence_x1_not_periodic}--\eqref{eq:sequence_xm-1_not_periodic} also hold for $\boldsymbol{x}'/\sqrt{V(\boldsymbol{x}')}$. This shows that the prefix of the sequence related to $\boldsymbol{x}'$ is $k_2...k_m$. Finally, $\boldsymbol{x}'/\sqrt{V(\boldsymbol{x}')}$ satisfies Eq.~\eqref{eq:sequence} for some sequence in $\mathcal{S}_{{\mathrm{S'}}}$; combining with the conclusion about the prefix above, $(\boldsymbol{x}',\sigma\sigma') \in \mathcal{R}_{{\mathrm{S'}}}$. The related transition exists because $(k\sigma, \sigma\sigma') \in \Es^{{\mathrm{S'}}}$ for every $\sigma\sigma' \in \mathcal{X}^{{\mathrm{S'}}}$.
%
\end{proof}
Note that $\mathcal{S}_{{\mathrm{S'}}}$ is, in general, nondeterministic. A depiction of such construction is seen in Fig.~\ref{fig:abstraction}. Some useful verification applications can be derived from the model $\mathcal{S}_{{\mathrm{S'}}}$:
\begin{prop}\label{prop:freq} An upper bound for the average triggering frequency \rev{of system \eqref{eq:plant}, \eqref{eq:quadtrig}} is $f^* = \max_{\sigma\in\mathcal{X}^{{\mathrm{S'}}}}(|\sigma|/(h{\sum_{k_i\in\sigma}k_i}))$. \end{prop}
\begin{proof} In the worst case, $\mathcal{S}_{{\mathrm{S'}}}$ generates $\sigma^* \coloneqq \arg\!\max_{\sigma\in\mathcal{X}^{{\mathrm{S'}}}}(|\sigma|/(h{\sum_{k_i\in\sigma}k_i}))$ repeatedly.
\end{proof}
\begin{prop}\label{prop:ges} Let $T^* = h\max_{\sigma\in\mathcal{X}^{{\mathrm{S'}}}}(\sum_{k_i\in\sigma}(k_i))$ be the longest (time-wise) sequence in $\mathcal{X}^{{\mathrm{S'}}}$. Then $b^*\! =\! -\log(r)/2T^*$ is an upper bound for the GES decay rate of system \eqref{eq:plant},\eqref{eq:quadtrig}.\end{prop}
%
\begin{proof}
Take an initial state $\boldsymbol{x}\in\mathbb{R}^{{n_{\mathrm{x}}}}$, its related sequence $\sigma \in \mathcal{X}^{{\mathrm{S'}}}$, and set $T = h\sum_{k_i\in\sigma}(k_i)$. From Def.~\ref{def:petcrel}, $V(T) = V(\boldsymbol{\xi}_{\boldsymbol{x}}(T)) \leq rV_0 = \mathrm{e}^{\log(r)}V(0) = \mathrm{e}^{-2b^*T}V(0).$
From GES of the PETC (Assumption \ref{assum:stable}), $V(t) \leq M\mathrm{e}^{-2bt}V(0)$ for some $b > 0$ and $M < \infty$. Consider two cases.
{\bf Case 1:} $t<T$. Combining the inequalities above gives
%
\begin{multline}\label{eq:t<T}
V(t) \leq M\mathrm{e}^{-2b(t-T)}V(T) \leq \mathrm{e}^{-2b^*T}M\mathrm{e}^{-2b(t-T)}V(0) \\ \leq M\mathrm{e}^{2bT}\mathrm{e}^{-2b^*T}V(0) \leq M'\mathrm{e}^{-2b^*t}V(0),
\end{multline}
%
where $M' \coloneqq M\mathrm{e}^{2bT^*} \geq M\mathrm{e}^{2bT}$.
{\bf Case 2:} $t>T$; then we can partition the trajectory $\boldsymbol{\xi}_{\boldsymbol{x}}(t)$ in intervals $[0, t_{m_1}],$ $[t_{m_1}, t_{m_2}],$ $...,$ $[t_{m_n}, t]$ satisfying $V(t_{m_i}) \leq rV(t_{m_{i-1}})$ and $V(t_{m_i}) > rV(t_{m_{i-1}-1}).$ Each interval but the last is associated with a sequence $\sigma_i \in \mathcal{X}^{{\mathrm{S'}}}$, and therefore its duration is $T_i \leq T^*$. Thus, with $t' = t - t_{m_n},$
%
\begin{multline*}
V(t) = V\left(\textstyle\sum_{i=1}^nT_i + t'\right) \leq r^nV(t') \\ \stackrel{{\text{Eq.~\eqref{eq:t<T}}}}{\leq} r^nM'\mathrm{e}^{-2b^*t'}V(0)
= M'\mathrm{e}^{-2b^*\!nT^*}\mathrm{e}^{-2b^*t'}V(0) \\
= M'\mathrm{e}^{-2b^*(nT^*+t')}V(0) \leq M'\mathrm{e}^{-2b^*\!t}V(0).
\end{multline*}
In the two cases, we have $V(t) \leq M'\!\mathrm{e}^{-2b^*\!t}V(0)$, which implies $$|\boldsymbol{\xi}(t)| \leq \sqrt{M'\lambda_{\max}(\boldsymbol{P})/\lambda_{\min}(\boldsymbol{P})}\mathrm{e}^{-b^*\!t}|\boldsymbol{\xi}(0)|.$$
\end{proof}
We conjecture that Proposition \ref{prop:ges} provides a better estimate of the convergence rate of System \eqref{eq:plant}--\eqref{eq:quadtrig} than what can be obtained by, e.g., the theorems in \cite{heemels2013periodic}. The reason behind this conjecture is that, as $N\to\infty$ (or $r\to 0$), our bound approaches what would be the joint spectral radius of the associated discrete-time system (see, e.g.,~\cite{ahmadi2014joint}).
\section{Numerical results}\label{sec:num}
Consider a plant and controller of the form \eqref{eq:plant} from \cite{tabuada2007event}, and the Lyapunov function $V(\boldsymbol{x}) = \boldsymbol{x}\tran\Pm_{\mathrm{Lyap}}\boldsymbol{x}$ such that the continuous-time closed-loop system satisfies $\ensuremath{\mathrm{d}} V(\boldsymbol{\xi}(t))/\ensuremath{\mathrm{d}} t = -\boldsymbol{\xi}(t)\tran\Qm_{\mathrm{Lyap}}\boldsymbol{\xi}(t)$, determined by the following matrices:
%
\begin{gather*} \boldsymbol{A} = \begin{bmatrix}0 & 1 \\ -2 & 3\end{bmatrix}, \ \boldsymbol{B} = \begin{bmatrix}0 \\ 1\end{bmatrix}, \ \boldsymbol{K} = \begin{bmatrix}1 & -4\end{bmatrix}, \\
\Pm_{\mathrm{Lyap}} = \begin{bmatrix}1 & 0.25 \\ 0.25 & 1\end{bmatrix}, \ \Qm_{\mathrm{Lyap}} = \begin{bmatrix}0.5 & 0.25 \\ 0.25 & 1.5\end{bmatrix}.
\end{gather*}
%
For the PETC implementation, we use a predictive Lyapunov-based triggering condition of the form
%
$ V(\boldsymbol{\zeta}(t)) > -\rho \boldsymbol{\zeta}(t)\tran\Qm_{\mathrm{Lyap}}\boldsymbol{\zeta}(t), $
%
where $\boldsymbol{\zeta}(t) \coloneqq \boldsymbol{A}_\ensuremath{\mathrm{d}}(1)\boldsymbol{\xi}(t) + \boldsymbol{B}_\ensuremath{\mathrm{d}}(1)\boldsymbol{K}\hat{\boldsymbol{\xi}}(t)$ is the next-sample prediction of the state and $0 < \rho < 1$ is the triggering parameter, here set to $\rho=0.8$. Setting $h = 0.1$ and $\bar{k} = 6$, we put it in the form \eqref{eq:quadtrig}, and obtained $a=0.952$ using LMIs based on Remark \ref{rem:a_computation}.
For the periodic \rev{phase}, the maximum period that verifies Assumption \ref{assum:periodicstable} is $h_{{\mathrm{P}}} = 0.4$ (with resolution of 0.01). Finally, we verified that $\underline{k}=1$ and set $r=0.1$.
%
\begin{figure}
\begin{center}
\input{mpetc.tex}
\vspace{-2em}
\caption{\label{fig:lyap} {Trajectory of the Lyapunov function for 10 different initial conditions under MPETC, with PETC samples marked. The maximum time $T^*$ it takes to reach $\mathcal{X}_{{\mathrm{P}}} = r\mathcal{X}_0$ is highlighted.} }
\vspace{-1.5em}
\end{center}
\end{figure}
%
Lemma \ref{lem:finite} gives $N=47$ and a worst-case value of $8.5\cdot10^{32}$ bisimulation states. We computed the bisimilar state set $\mathcal{X}^{{\mathrm{B}}}$ implementing a recursive algorithm (as discussed in Remark \ref{rem:properalgorithm}), obtaining a total of 219 states, out of which 109 belong to $\mathcal{X}^{{\mathrm{S'}}}$. The Python implementation, using Z3 to solve Eq.~\eqref{eq:sequence}, took 21 min to generate these state sets.
Using MPETC, the maximum time it takes for $\mathcal{X}_{{\mathrm{P}}}$ to be reached is $$T^* = h\max_{\sigma\in\mathcal{X}^{{\mathrm{B}}}}(\sum_{k_i\in\sigma}(k_i)) = 2.3.$$ This is highlighted in Fig.~\ref{fig:lyap}, which shows simulations from 10 different initial conditions.
%
For PETC, applying Proposition \ref{prop:ges} gives $b^* = 0.5,$ while the best GES rate that can be obtained using the LMI approaches from \cite{heemels2013periodic} is $b = 0.23$, using Theorem III.4. For the average PETC sampling frequency, Proposition \ref{prop:freq} gives $f^* = 20/3$ (compared to $1/h = 10$), corresponding to the sequence $\sigma = (4,1,1,1,1,1).$
\section{Conclusions}
We have presented a practical alternative to ETC, the MPETC, \rev{which provides the benefits of PETC during transients and the traffic predictability of periodic sampling when close to steady state. Furthermore, we have presented a method to compute a bisimilar traffic model for MPETC.} In addition, we have presented some verification applications of the (bi)similar models that can be used for both PETC and MPETC. This is an important step towards understanding traffic characteristics of ETC, and it may support its applicability in real NCSs, since the traffic benefits are among the main motivations for the usage of ETC. Future work shall focus on expanding these models to scheduling \cite{mazo2018abstracted, gleizer2020scalable} and other triggering conditions, as well as using efficient relaxations to solve the satisfiability of Eq.~\eqref{eq:sequence}, such as $\delta$-SMT \cite{gao2013dreal} and semi-definite relaxations.
%
\bibliographystyle{ieeetr}
| da17be0a2dcb61586e05c31b8e3619316f155a5e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the companion letter \cite{C-DC-I-M-S-2020-01}, the problem of the definition of a covariant bracket on the space of functions on the solutions of a suitable variational principle has been addressed from the point of view of the geometry of contact manifolds.
Specifically, in the case of non-relativistic Hamiltonian mechanics, and of the relativistic particle, it was proved how the covariant bracket may be read in terms of a Poisson subalgebra of the algebra of smooth functions on a contact manifold endowed with its natural Jacobi bracket\cite{AsoCiagliaDCosmoIbortMarmo2017-Covariant_Jacobi_brackets,AsoCiagliaDCosmoIbort2017-Covariant_brackets,C-C-M-2018}.
In this letter, we push further this analysis by considering the case of Klein-Gordon theory, and the case of the multisymplectic formulation of Schr\"{o}dinger equation.
The main difference with the cases considered in the companion letter\cite{C-DC-I-M-S-2020-01} is given by the fact that these systems describe partial differential equations.
Consequently, we will see that it is possible to describe them in terms of two different but equivalent action principles taking place on two different spaces of sections.
On the one hand, it is possible to formulate an action principle for sections of a suitable finite-dimensional fiber bundle on the spacetime manifold in the spirit of the multisymplectic formalism\cite{CarinCrampIbort1991-Multisymplectic,IbSpiv2017-Covariant_Hamiltonian_boundary, KijowskiTulczyjew1979, BinzSniatyckiFisher1988, GoldschmidtSternberg}.
In this case, it is not possible to express the covariant bracket in terms of a Poisson subalgebra of a suitable contact manifold (see, for instance, \cite{Cr87,Cr88,MarsdMongMorrThom1986-Covariant_Poisson_bracket} for historical attempts in writing a covariant bracket for Field Theory within the canonical formalism and \cite{ForgRomero2005-Covariant_poisson_brackets,Fredenhagen2015-Algebraic_quantum_field_theory}, and references therein, for a more modern geometric approach).
However, this point of view may be considered as the starting point for the generalization of the notion of contact manifold and contact structure to framework which is more suitable for field theories along the lines presented in \cite{Vinogradov-Contact_Monge}.
In this sense, we expect that an approach in terms of multicontact geometry could help for a better understanding of the covariant formalism, in the same way as multisymplectic formalism has generalized symplectic geometry, and leave the discussion of these aspects to a future work.
On the other hand, it is possible to formulate an action principle for sections of a suitable \grit{infinite-dimensional} fiber bundle over the time manifold $\mathbb{R}$, in analogy with the particle case.
In this context, the infinite-dimensional nature of the fiber bundle will make possible to read the Poisson bracket on the space of functionals on the solutions of the action principle in terms of a Poisson subalgebra of the algebra of smooth functions endowed with a Jacobi bracket.
\section{Multisymplectic formulation of free Klein-Gordon theory}\label{sec: KG}
We consider the covariant Hamiltonian description of Klein-Gordon theory on the Minkowski spacetime $\mathcal{M} = (\mathbb{R}^4,\eta)$, where $\eta$ is the metric tensor
\begin{equation}
\eta = \eta_{\mu \nu} {\rm d} x^{\mu} \otimes {\rm d} x^{\nu} = -{\rm d} x^0 \otimes {\rm d} x^0 + \delta_{jk}{\rm d} x^j \otimes {\rm d} x^k ,
\end{equation}
and $(x^{\mu})$ is a global set of Cartesian coordinates.
Let $\pi_0\colon E=\mathbb{R}\times\mathcal{M}\,\rightarrow \,\mathcal{M}$ be a bundle with $\pi_{0}$ the standard projection on the second factor.
This bundle represents the analogue of the extended configuration space $\mathcal{Q}\times \mathbb{R}$ of non-relativistic Hamiltonian mechanics considered in the companion letter \cite{C-DC-I-M-S-2020-01}.
Klein-Gordon fields are sections $\phi$ of the bundle $\pi_0\colon E=\mathbb{R}\times\mathcal{M}\,\rightarrow \,\mathcal{M}$.
The extended phase space $\mathbf{T}^{*}\mathcal{Q}\times \mathbb{R}$ of non-relativistic particle mechanics is here replaced by the so called covariant phase space $\mathcal{P}$ which is a fibre bundle over both $E$ and the spacetime $\mathcal{M}$ \cite{CarinCrampIbort1991-Multisymplectic,IbSpiv2017-Covariant_Hamiltonian_boundary, BinzSniatyckiFisher1988, KijowskiTulczyjew1979}.
Technically speaking, $\mathcal{P}$ is the (affine) dual bundle of the first jet bundle $J^1(E)$ that can be identified with the space of 1-semibasic $4$-forms on $E$.
In the Klein-Gordon case on Minkowski spacetime we are considering, $\mathcal{P}$ is diffeomorphic to $E\times\mathbb{R}^{4}$.
If $(x^{\mu})$, $(u,x^{\mu})$, and $\left( \rho^{\mu},u,x^{\mu} \right)$ are, respectively, Cartesian coordinates on $\mathcal{M}$, bundle coordinates on $E$, and bundle coordinates on $\mathcal{P}$, the projection $\pi_{E}\colon \mathcal{P}\,\rightarrow \, E$ is locally given by
\begin{equation}
\pi_{E}(\rho^{\mu},u,x^{\mu}) = (u,x^{\mu}) \,,
\end{equation}
while the projection $\pi\colon \mathcal{P}\,\rightarrow \, \mathcal{M}$ is given by
\begin{equation}
\pi(\rho^{\mu},u,x^{\mu}) = (x^{\mu})\,.
\end{equation}
Let $\mathscr{F}_{\mathcal{P}}$ denote the space of sections $\chi\colon \mathcal{M}\,\rightarrow\,\mathcal{P}$ of the form:
\begin{equation}
\chi(x^{\mu}) = (x^{\mu}, \phi(x^{\mu}), P^{\nu}(x^{\mu}))\,, \quad \mu,\nu = 0,1,2,3\,.
\end{equation}
Elements in $\mathscr{F}_{\mathcal{P}}$ will be the fields of the Hamiltonian theory where $P^\nu$ are identified with the momenta fields of the theory.
A rigorous analytical framework can be provided by introducing some regularity conditions on the various spaces of fields, for instance, we may consider the fields $\phi$ in the Sobolev space $\mathcal{V}= \mathcal{H}^1(\mathcal{M},\mathrm{vol}_{\mathcal{M}})$ of square integrable functions on $\mathcal{M}$ with respect to the Lebesgue measure on $\mathcal{M}$, and the momenta fields $P = (P^{\mu})$ in the Hilbert space $\mathcal{W} = L^2(\phi^*\mathcal{P})$ of square integrable sections of the pullback of the bundle $\mathcal{P} \to E$ along $\phi$.
With these choices, the space $\mathscr{F}_{\mathcal{P}}$ becomes the Hilbert space $\mathscr{F}_{\mathcal{P}}\,=\,\mathcal{V}\,\oplus\,\mathcal{W}$.
A variation for $\chi\in\mathscr{F}_{\mathcal{P}}$, commonly denoted as $\delta \chi$, is a tangent vector at $\chi$, which may be identified with a vector field $U_{\chi}$ along $\chi$ on $\mathcal{P}$, which is vertical with respect to the fibration $\pi\colon\mathcal{P}\rightarrow\mathcal{M}$.
The tangent space $\mathbf{T}_{\chi}\mathscr{F}_{\mathcal{P}}$ is given by all variations $\delta\chi_U = U_{\chi}$.
In the following, it will be useful to extend $U_{\chi}$ to a vertical vector field $\widetilde{U}$ on a neighbourhood of the image of $\chi$ within $\mathcal{P}$, as:
\begin{equation}
\tilde{U} = U_{\phi} \frac{\partial }{\partial u} + U_{P}^{\mu}\frac{\partial }{\partial \rho^{\mu}}\,.
\end{equation}
Now, we pass to describe the dynamics in terms of the (Schwinger-Weiss) action principle.
Given a Hamiltonian function $H\colon \mathcal{P}\,\rightarrow\,\mathbb{R}$ and volume form $\mathrm{vol}_\mathcal{M} = {\rm d} x^0\wedge {\rm d} x^1 \wedge {\rm d} x^2 \wedge {\rm d} x^3$ on the base manifold $\mathcal{M}$, we consider the 4-form\footnote{It is possible to define this form in an intrinsic way, see for instance \cite{IbSpiv2017-Covariant_Hamiltonian_boundary}.} $\theta_H$ on $\mathcal{P}$ given by
\begin{equation}
\theta_H = \rho^{\mu}{\rm d} u \wedge i_{\frac{\partial}{\partial x^{\mu}}}\mathrm{vol}_\mathcal{M} - H \mathrm{vol}_\mathcal{M}\,.
\end{equation}
In the particular instance of free dynamics, the Hamiltonian function is
\begin{equation}\label{eqn: relativistic Hamiltonian}
H = \frac{1}{2} \left( \eta_{\mu \nu} \rho^{\mu}\rho^{\nu} - \mathfrak{m}^2 u^2 \right)\,.
\end{equation}
The action functional $S\colon\mathscr{F}_{\mathcal{P}}\, \rightarrow\, \mathbb{R}$, of the theory can be written as:
\begin{equation}
S[\chi] = \int_{\mathcal{M}}\chi^*\left( \theta_H \right) = \int_{\mathcal{M}}\left( P^{\mu}\partial_\mu \phi - H \right) \mathrm{vol}_\mathcal{M}\,.
\label{action_klein-gordon}
\end{equation}
Given $U_{\chi}\in\mathbf{T}_{\chi}\mathscr{F}_{\mathcal{P}}$, the variation $\mathrm{d}S[\chi](U_{\chi})$ of $S$ is
\begin{equation}
\begin{split}
\mathrm{d}S[\chi](U_{\chi})&= \int_{\mathcal{M} }\chi^*\left( \mathrm{L}_{\tilde{U}}\theta_H \right) = \int_{\mathcal{M} }\chi^*\left( i_{\tilde{U}}{\rm d} \theta_H \right) + \int_{\partial \mathcal{M} } \chi_{\partial\mathcal{M}}^*\left( i_{\tilde{U}}\theta_H \right),
\end{split}
\end{equation}
where $\tilde{U}$ is any extension of $U_{\chi}$,
and, if $\partial\mathcal{M}=\emptyset$, we get:
$$
\mathrm{d}S[\chi](U_{\chi}) = \,\int_{\mathcal{M} }\chi^*\left( i_{\tilde{U}}{\rm d} \theta_H \right) = \int_{\mathcal{M} } \left(\frac{\partial \phi}{\partial x^\mu} - \frac{\partial H}{\partial P^\mu} \right) U_P^\mu + \left(\frac{\partial P^\mu}{\partial x^\mu} + \frac{\partial H}{\partial \phi} \right) U_\phi^\mu \, \mathrm{vol}_\mathcal{M}\, .
$$
The Schwinger-Weiss action principle states that the variations of the action depend solely on the variations of the fields at the boundary, hence the actual dynamical configurations of the fields of the theory must satisfy the Euler-Lagrange equations:
$$
\frac{\partial \phi}{\partial x^\mu} = \frac{\partial H}{\partial P^\mu} \, , \qquad \frac{\partial P^\mu}{\partial x^\mu} = - \frac{\partial H}{\partial \phi} \, ,
$$
which can be geometrically interpreted as the zeroes of a 1-form $\mathbb{EL}$ on the space of fields $\mathscr{F}_{\mathcal{P}}$ given by:
\begin{equation}\label{eqn: Schwinger-Weiss action principle 1}
\mathbb{EL}_{\chi}(U_{\chi}) := \int_{\mathcal{M} }\chi^*\left( i_{\tilde{U}}{\rm d} \theta_H \right) = 0 \,,\quad \forall U_{\chi}\in\mathbf{T}_{\chi}\mathscr{F}_{\mathcal{P}}\, .
\end{equation}
We note that such a geometrical reformulation of the Schwinger-Weiss variational principle should be used to introduce a Quantum Action Principle in the groupoid reformulation of Schwinger's algebra of selective measurements which has been recently proposed by some of the authors (see \cite{C-I-M-2018,C-I-M-02-2019,C-I-M-03-2019,C-I-M-05-2019,C-DC-I-M-2020,C-DC-I-M-02-2020} for more details).
We will denote by $\mathcal{EL}_{\mathcal{M}}\subset \mathscr{F}_{\mathcal{P}}$ the space of solutions of Euler-Lagrange equations, that is:
\begin{equation}
\mathcal{EL}_{\mathcal{M}}\,:=\,\left\{\chi\in \mathscr{F}_{\mathcal{P}}\,\colon\,\mathbb{EL}_{\chi}(U_\chi) = 0\,,\quad \forall U_{\chi} \in \mathbf{T}_{\chi}\mathscr{F}_{\mathcal{P}}\right\}.
\end{equation}
Hence, we obtain the Euler-Lagrange equations (also known as the de Donder-Weyl equations) for the Klein-Gordon theory:
\begin{equation}
\frac{\partial \phi}{\partial x^{\mu}} = \eta_{\mu \nu} P^{\nu}\,,\qquad\, \frac{\partial P^{\mu}}{\partial x^{\mu}} = \mathfrak{m}^2 \phi\,.
\end{equation}
Analogously to what happens in non-relativistic Hamiltonian mechanics \cite{C-DC-I-M-S-2020-01}, if we select a codimension-one, spacelike submanifold $\Sigma \subset \mathcal{M}$, for instance
\begin{equation}
\Sigma=\{m\in\mathcal{M}\,|\,x^{0}(m)= \tau^0 \},
\end{equation}
and denote by $i_{\Sigma}:\Sigma\rightarrow\mathcal{M}$ the canonical immersion of $\Sigma$ in $\mathcal{M}$, the space $\mathcal{EL}_{\mathcal{M}}$ is equipped with a 2-form $\Omega_{\chi}$ given by:
\begin{equation}\label{eqn: omega on solutions of de Donder Weyl for KG}
\Omega_{\chi} (U_{\chi},V_{\chi}) = \int_{\Sigma}i_{\Sigma}^*\left( \chi^*\left( i_{\tilde{V}}i_{\tilde{U}} {\rm d} \theta_H \right) \right) = \int_{\Sigma} i_{\Sigma}^*\left( \delta P^{0}_U\, \delta\phi_V - \delta\phi_U \, \delta P^{0}_V \right) \mathrm{vol}_\Sigma \ , \nonumber
\end{equation}
with $U_\chi = \delta \phi_U \partial /\partial u + \delta P^\mu_U \partial /\partial \rho^\mu$ and $V_\chi$ similarly.
Following the procedure outlined in \cite{C-DC-I-M-S-2020-01}, provided that $\chi \in \mathcal{EL}_{\mathcal{M}}$, it is possible to prove that $\Omega_{\chi}$ is independent of the choice of $\Sigma$, and it defines a canonical symplectic structure on the space of solutions of Klein-Gordon equation.
\vspace{0.4cm}
Now, we will provide an evolution description of Klein-Gordon theory in terms of a vector field on an infinite-dimensional manifold.
The main idea is to provide a framework in which the role of a contact structure is manifestly evident, and which allows us to write the Poisson bracket associated with $\Omega$ in terms of the Jacobi bracket defined by such structure as done in the companion letter\cite{C-DC-I-M-S-2020-01} for the case of non-relativistic Hamiltonian mechanics and the relativistic particle.
To define the infinite-dimensional manifold which will be the carrier space of the dynamics, we fix a space-like, codimension-one submanifold $\Sigma\subset\mathcal{M}$ as before.
Then, we consider the pullback bundle $i_{\Sigma}^{*}\mathcal{P}$ whose sections $\sigma$ are just compositions of sections $\chi$ of $\mathcal{P}$ with $i_{\Sigma}$, $\sigma = \chi \circ i_\Sigma$ and, in local coordinates, we have:
$$
\sigma(x^{j})=\left(x^{j},\tau^0, \varphi(x^{j}), \, p(x^{j}) ,\, \beta^k(x^{j}) \, \right),
$$
where
\begin{equation}
\varphi(x^{j}) =(\phi(x^{j},\tau^{0})) |_{\Sigma}\, , \quad
p(x^{j}) =(P^{0}(x^{j},\tau^{0})) |_{\Sigma} \, , \quad
\beta^k(x^{j}) =(P^k(x^{j},\tau^{0})) |_{\Sigma}\,.
\end{equation}
As before, we will focus on those sections of $i_{\Sigma}^{*}\mathcal{P}$ satisfying some additional regularity conditions.
Specifically, we will consider $\varphi\in\mathscr{V}_\Sigma$, the space of sections of Sobolev class 1 of the bundle $i_\Sigma^*E$, $p\in\mathscr{W}_\Sigma =\mathcal{L}^2(\Sigma, \mathrm{vol}_{\Sigma})$, and $\beta = (\beta^{j}) \in\mathscr{B}_\Sigma$, the space of square integrable sections of the tangent bundle $\mathbf{T}\Sigma$, so that we have the Hilbert space $ \mathcal{F}_{\Sigma}$ of fields at $\Sigma$ given by $\mathcal{F}_{\Sigma}\,:=\,\mathscr{V}_\Sigma \,\oplus\,\mathscr{W}_\Sigma\,\oplus\,\mathscr{B}_\Sigma$.
Then, we consider the Hilbert bundle $\tau\colon\mathcal{F}_{\Sigma}\times\mathbb{R}\rightarrow\mathbb{R}$, where $\tau$ is the projection on the second factor, which plays the role of the extended phase space in \cite{C-DC-I-M-S-2020-01}, and we denote by $\Gamma(\mathcal{F}_\Sigma)$ the space of sections of this bundle.
The key observation is that, under suitable regularity conditions, the sections in $\Gamma(\mathcal{F}_\Sigma)$, i.e., curves $\sigma \colon \mathbb{R} \to \mathcal{F}_{\Sigma}$, are in one-to-one correspondence with fields in $\mathscr{F}_{\mathcal{P}}$.
Indeed, given $\chi\in\mathscr{F}_{\mathcal{P}}$, we can define the section $\sigma_{\chi}\in\Gamma$ setting
\begin{equation}
\sigma_{\chi}(s)\,:=\,\left(s,x^{j}, \tau^{0},\varphi_{s}(x^{j}),p_{s}(x^{j}),\beta^{k}_{s}(x^{j})\right),
\end{equation}
where $s=x^{0}$, and
\begin{equation}
\varphi_{s}(x^{j}) =\phi(x^{j},s) \, , \qquad
p_{s}(x^{j}) =P^{0}(x^{j},s) \, , \qquad
\beta^{k}_{s}(x^{j}) = P^k(x^{j},s) \,.
\end{equation}
On the other hand, given $\sigma \in\Gamma(\mathcal{F}_\Sigma)$, we can define $\chi_{\sigma}\in\mathscr{F}_{\mathcal{P}}$ by reading the previous equations in the other direction.
The regularity conditions for $\sigma\in\Gamma(\mathcal{F}_\Sigma)$ alluded to above are precisely those assuring that $\phi(x^{j},s):=\varphi_{s}(x^{j})$ is in $\mathcal{V}$, and that $(P^{0}(x^{j},s):=p_{s}(x^{j}),P^k(x^{j},s):=\beta^{k}_{s}(x^{j}))$ are in $\mathcal{W}$.
\vspace{0.4cm}
A variation $U_{\gamma}$ at $\gamma\in \Gamma (\mathcal{F}_\Sigma)$ is then a vector field along $\gamma$ which is vertical with respect to the fibration $\tau\colon\mathcal{F}_{\Sigma}\times\mathbb{R}\rightarrow\mathbb{R}$.
The space of all $U_{\gamma}$ is denoted by $\mathbf{T}_{\gamma}\Gamma_{\gamma}$.
If necessary, $U_{\gamma}$ can be extended to a vertical vector field $\tilde{U}$ in an open neighbourhood of the image of $\gamma$ written as
\begin{equation}
\tilde{U}\,=\,U_{\varphi}\frac{\delta}{\delta \varphi} + U_{p}\frac{\delta}{\delta p} + U_{\beta }^{j} \frac{\delta}{\delta \beta^{j}},
\end{equation}
where $U_{\varphi}\in\mathscr{V}_\Sigma$, $U_{p}\in\mathscr{W}_\Sigma$, $U_{\beta}^{j}\in\mathscr{B}_\Sigma$.
The symbols $\frac{\delta}{\delta \varphi}$ have a double interpretation.
On one side, they represent the canonical basis of tangent vectors on a linear space associated to the natural chart provided by the linear space itself.
On the other, they represent the standard functional derivatives\cite{AbraMars-Foundations_of_Mechanics}, i.e., they act on a function $F$ as the functional derivative $\frac{\delta F}{\delta \varphi (x)}$, and similarly for $\frac{\delta}{\delta p (x)}$ and $\frac{\delta}{\delta \beta^{j}(x)}$.
In a similar way, the symbols $\delta \varphi$ and $\delta p$ can be understood as a basis of covectors on the manifold $\mathcal{F}_\Sigma$, and $\delta F = \frac{\delta F}{\delta \sigma} \delta \sigma$ is understood as the differential of the function $F$ given by
$$
\langle \delta F, U_\sigma \rangle = \int_\Sigma \frac{\delta F}{\delta \sigma (x)} \, U_\sigma (x) \, \mathrm{vol}_\Sigma (x) \, ,
$$
Now, we define the Hamiltonian function
\begin{equation}
\mathcal{H}(\varphi,p,\beta;s) = \int_{\Sigma} \dfrac{1}{2}\left( p^2 + \delta^{jk} \frac{\partial \varphi}{\partial x^{j}}\frac{\partial \varphi}{\partial x^k} - \mathfrak{m}^2 \varphi^2 \right)\mathrm{vol}_{\Sigma}\,=:\,\int_{\Sigma}\,H\,\mathrm{vol}_{\Sigma}\,,
\end{equation}
and we consider the one-form $ \mathcal{F}_{\Sigma}\times\mathbb{R}$ given by
\begin{equation}\label{eqn: infinite theta}
\Theta_{\mathcal{H}}\,=\,\int_{\Sigma}\,\left(p\,\delta\varphi\right)\,\mathrm{vol}_{\Sigma} - \mathcal{H}{\rm d} t\,.
\end{equation}
This form plays a role analogous to that of the contact one-form on the extended phase space $\mathbf{T}^{*}\mathcal{Q}\times\mathbb{R}$ in the case of non-relativistic Hamiltonian dynamics, and to that of the contact one-form on the mass-shell for the relativistic particle considered in the companion letter \cite{C-DC-I-M-S-2020-01}.
Then, we write the action functional $S$ on $\Gamma(\mathcal{F}_{\Sigma})$ given by
\begin{equation}
S[\gamma]\,=\,\int_{\mathbb{R}}\,\gamma^{*}\Theta_{\mathcal{H}},
\end{equation}
and we compute its variation along $U_{\gamma}$ as
\begin{equation}
\mathrm{d}S_{\gamma}(U_{\gamma})\,=\,\int_{\mathbb{R}}\,\gamma^{*}\left(i_{\tilde{U}}{\rm d}\Theta_{\mathcal{H}} + {\rm d}(i_{\tilde{U}}\Theta_{\mathcal{H}})\right)\,.
\end{equation}
Exploiting Stokes' theorem and the fact that $\partial\mathbb{R}=\emptyset$, we obtain
\begin{equation}
\mathrm{d}S_{\gamma}(U_{\gamma})\,=\,\int_{\mathbb{R}}\,\gamma^{*}\left(i_{\tilde{U}}{\rm d}\Theta_{\mathcal{H}} \right).
\end{equation}
Following the Schwinger-Weiss action principle, the dynamical trajectories are given by all those $\gamma$ such that
\begin{equation}
\mathbb{EL}_{\gamma}(U_{\gamma})\,:=\,\int_{\mathbb{R}}\,\gamma^{*}\left(i_{\tilde{U}}{\rm d}\Theta_{\mathcal{H}} \right)\,=\,0 \quad\forall\,U_{\gamma}\,\,\mathbf{T}_{\gamma}\Gamma_{\gamma}\,.
\end{equation}
A direct computation shows that all such $\gamma$ must satisfy the constraint relations
\begin{equation}
\delta_{jk}\beta^k = \frac{\partial \varphi}{\partial x^j} ,
\label{constraints klein-gordon}
\end{equation}
and the ``evolution equations''
\begin{equation}
\frac{{\rm d} \varphi}{{\rm d} s} \,= p\, ,\quad\,\frac{{\rm d} p}{{\rm d} s} = - \Delta_{\Sigma} \varphi - \mathfrak{m}^2 \varphi .
\label{Cauchy-form klein gordon equation}
\end{equation}
The constraint conditions in Eq. \eqref{constraints klein-gordon} being linear determine a submanifold $\mathcal{C}\subset \mathcal{F}_{\Sigma}$ which is a linear subspace.
We define the trivial bundle $\tau_{\mathcal{C}}\colon\mathcal{C}\times\mathbb{R}\rightarrow\mathbb{R}$, where $ \tau_{\mathcal{C}}$ is the projection on the second factor, and we denote by $\Gamma_{\mathcal{C}}$ the space of sections of this bundle.
Moreover, we write, with an evident abuse of notation, $\Theta_{\mathcal{H}}$ for the pullback to $\mathcal{C}\times\mathbb{R}$ of the one-form in equation \eqref{eqn: infinite theta}.
Then, it is clear that Eq.\eqref{Cauchy-form klein gordon equation} may be read as defining the integral curves of the (densely defined) vector field
\begin{equation}
X_{H} = \frac{\partial}{\partial t} + \frac{\delta H}{\delta p}\,\frac{\delta}{\delta \varphi} - \frac{\delta H}{\delta \varphi}\,\frac{\delta}{\delta p}\,
\end{equation}
which is in the kernel of the two-form
\begin{equation}
{\rm d} \Theta_{\mathcal{H}}\, =\, \int_{\Sigma}\left( \delta p \wedge \delta\varphi\right)\mathrm{vol}_{\Sigma} - {\rm d}\mathcal{H} \wedge {\rm d} t \,.
\end{equation}
This two-form plays the role of the contact two-form on $\mathbf{T}^{*}\mathcal{Q}\times\mathbb{R}$ in the case of non-relativistic Hamiltonian mechanics, and of the contact two-form on the mass-shell in the case of the relativistic particle considered in \cite{C-DC-I-M-S-2020-01}.
Eventually, we obtained the de Donder-Weyl equations of the free Klein-Gordon theory as a Hamiltonian system on an infinite-dimensional manifold, in such a way that a section $\gamma$ satisfying Eq.\eqref{Cauchy-form klein gordon equation} provides a representation in terms of Cauchy data on $\Sigma$ of the points in $\mathcal{EL}_{\mathcal{M}}$.
We denote by $\mathcal{EL}_{\mathcal{C}}$ the space of all $\gamma\in\Gamma_{\mathcal{C}}$ satisfying Eq.\eqref{Cauchy-form klein gordon equation}.
Just as it happens for non-relativistic Hamiltonian mechanics and for the relativistic particle\cite{C-DC-I-M-S-2020-01}, any vector field of the form $f\,X_{H}$, with $f$ a smooth, non-vanishing function on $\mathcal{C}\times\mathbb{R}$ is again in the kernel of ${\rm d} \theta_{\mathcal{H}}$, and the support of its integral curves coincide with the support of the integral curves of $X_{H}$.
Accordingly, we may interpret the integral curves of $f\,X_{H}$ as reparametrizations of the dynamical trajectories.
In particular, the vector field $\Gamma_H$ satisfying $i_{\Gamma_H}\theta_\mathcal{H} = 1$ is called Reeb vector field.
Under suitable regularity properties for $X_H$, the family of its integral curves defines a regular foliation of the manifold $\mathcal{C}\times \mathbb{R}$, and every point in the quotient manifold, say $\mathrm{Q}$, associated with the foliation can be identified with one and only one element in $\mathcal{EL}_{\mathcal{C}}$.
Accordingly, we may look at the space of smooth functions on $\mathrm{Q}$ as the subalgebra $C^{\infty}_H(\mathcal{C}\times \mathbb{R})\subset C^{\infty} (\mathcal{C}\times \mathbb{R})$ of smooth functions such that $\mathrm{L}_{X_H}f=0$.
Then, we may look for a bivector $\Lambda$ on $\mathcal{C}\times \mathbb{R}$ satisfying the relations
\begin{equation}
\left[\Lambda,\Lambda\right]_{S}\,=\,2\,\Gamma_H\,\wedge\,\Lambda\,,\quad\,\mathrm{L}_{\Gamma_H}\Lambda\,=\,0,
\end{equation}
where $[,]_{S}$ denotes the Schouten-Nijenhuis bracket, in order to define a Jacobi bracket\cite{AsoCiagliaDCosmoIbortMarmo2017-Covariant_Jacobi_brackets,C-C-M-2018} $[,]_{J}$ on $ C^{\infty} (\mathcal{C}\times \mathbb{R})$ by means of
\begin{equation}
[f,g]_{J}\,:=\,\Lambda(\mathrm{d}f,\mathrm{d}g) + f\,\mathrm{L}_{\Gamma_H}g - g\,\mathrm{L}_{\Gamma_H}f\,.
\end{equation}
Recalling that $\Gamma_{H}$ annihilates the elements in $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$, it is immediate to check that $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$ becomes a Poisson subalgebra of $C^{\infty}(\mathcal{C}\times \mathbb{R})$ with respect to the Jacobi bracket defined above.
Upon identifying $\mathcal{EL}_{\mathcal{M}}$ with $\mathcal{EL}_{\mathcal{C}}$, and then $\mathcal{EL}_{\mathcal{C}}$ with $\mathrm{Q}$, the Poisson bracket on $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$ is precisely the Poisson bracket associated with the two-form $\Omega$ in Eq.\eqref{eqn: omega on solutions of de Donder Weyl for KG}.
To explicitely write the Jacobi and Poisson bracket, we must be able to find an analogue of the generalized Darboux coordinates used in the particle case in the companion letter \cite{C-DC-I-M-S-2020-01}.
At this purpose, we first move from $\mathcal{C}$ to $\mathcal{C}'$ introducing the Fourier transforms
\begin{equation}
\varphi(x)\,=\,\int_{\overline{\Sigma}}\,\hat{\varphi}(k)\,\mathrm{e}^{i k\cdot x}\,\,\mathrm{vol}_{\overline{\Sigma}},\quad\,p(x)\,=\,\int\,\hat{p}(k)\,\mathrm{e}^{i k\cdot x}\,\mathrm{vol}_{\overline{\Sigma}}.
\end{equation}
Note that the fact that $\varphi$ and $p$ are real-valued imposes the constraints $\overline{\hat{\varphi}}(k)=\hat{\varphi}(-k)$ and $\overline{\hat{p}}(k)=\hat{p}(-k)$.
The equations of motion become
\begin{equation}
\frac{{\rm d} \hat{\varphi}}{{\rm d} s}\,=\,-\hat{p},\quad\,\frac{{\rm d}\hat{p}}{{\rm d} s}\,=\, \omega_{k}\,\hat{\varphi}\,,
\end{equation}
with $\omega_{k}=(k^{2}+\mathfrak{m}^{2})$, and we may read them as a superposition in $k$ of harmonic oscillators.
Now, we consider the ``change of coordinates'' in $\mathcal{C}'\times\mathbb{R}$ given by
\begin{equation}
\begin{split}
W\,=\,\frac{1}{2}\int_{\overline{\Sigma}}\,\biggl( \left(\hat{p}\,\overline{\hat{p}} - \omega_{k}^{2}\,\hat{\varphi}\,\overline{\hat{\varphi}}\right) & \frac{\cos(\omega_{k}s)\,\sin(\omega_{k}s)}{\omega_{k}} + 2\left(\hat{p}\overline{\hat{\varphi}} + \overline{\hat{p}}\,\hat{\varphi}\right)\sin^{2}(\omega_{k}s) \biggr) \,\mathrm{vol}_{\overline{\Sigma}} \\
\hat{\Phi}&\,=\,\hat{\varphi}\,\cos(\omega_{k}s) - \frac{\hat{p}}{\omega_{k}}\,\sin(\omega_{k}s) \\
\hat{P}&\,=\,\hat{p}\cos(\omega_{k}s) + \omega_{k}\hat{\varphi}\,\sin(\omega_{k}s)\,,
\end{split}
\end{equation}
where, again, it is $\overline{\hat{\Phi}}(k)=\hat{\Phi}(-k)$ and $\overline{\hat{P}}(k)=\hat{P}(-k)$.
In this coordinate system, it is possible to see that the one-form $\Theta_{\mathcal{H}}$ in equation \eqref{eqn: infinite theta} becomes
\begin{equation}
\Theta_{\mathcal{H}}\,=\,\int_{\overline{\Sigma}} \,\frac{1}{2}\,\left(\overline{\hat{P}}\,\delta\hat{\Phi} + \hat{P}\,\delta\overline{\hat{\Phi}}\right) \,\mathrm{vol}_{\overline{\Sigma}} + {\rm d} W\,.
\end{equation}
The vector field $\Gamma_{H}$ becomes $\Gamma_{H}\,=\,\frac{\partial}{\partial W}$, and, in analogy with what we did in the particle case in the companion letter \cite{C-DC-I-M-S-2020-01}, the bivector field $\Lambda$ in the definition of the Jacobi bracket reads
\begin{equation}
\Lambda\,=\,\int_{\overline{\Sigma}}\,\left(\frac{\delta}{\delta \hat{\Phi} } - \hat{P} \,\frac{\partial}{\partial W}\right)\,\wedge\,\frac{\delta}{\delta \overline{\hat{P}} }\,\mathrm{vol}_{\overline{\Sigma}} \,.
\end{equation}
Clearly, elements in $C^{\infty}_H(\mathcal{C}'\times \mathbb{R})$ are just those functionals which do not depend on $W$, and thus the Jacobi bracket among them becomes the Poisson bracket given by
\begin{equation}
\Lambda({\rm d} F,{\rm d} G)\,=\,\int_{\overline{\Sigma}}\, \frac{\delta F}{\delta \hat{\Phi} } \,\wedge\,\frac{\delta G}{\delta \overline{\hat{P}} }\,\mathrm{vol}_{\overline{\Sigma}} \,.
\end{equation}
\section{Multisymplectic formulation of free Schr\"{o}dinger equation}
In this section, we will consider the Schr\"{o}dinger equation for a free particle\footnote{The case of unitary evolutions of a finite-level quantum system\cite{C-DC-I-L-M-2017,C-DC-L-M-2017} may be consistently dealt with within the formalism described in the companion letter \cite{C-DC-I-M-S-2020-01}.} in $\mathbb{R}^{3}$.
The analysis of the Schr\"{o}dinger equation with a potential may be given following the lines presented here, bearing in mind, however, that the presence of the potential may affect the choice of the appropriate Sobolev and Hilbert spaces for the fields considered.
For the Schr\"{o}dinger equation, our spacetime manifold will be $\mathcal{M}\cong\mathbb{R}^{4}$.
In this case, we are in a non-relativistic context, where the notion of {\itshape absolute simultaneity} is available\cite{deritis_marmo_preziosi-a_new_look_at_relativity_transformations,MarmoPreziosi}, and thus the spacetime may be splitted in the product of space and time.
The notion of absolute simultaneity is encoded in the existence of an exact one-form $\vartheta$ on $\mathcal{M}$ the kernel of which determines an integrable foliation whose leaves are diffeomorphic with $\mathbb{R}^{3}$.
The foliation associated with $\vartheta$ defines absolute simultaneity, and the leaves define the simultaneity surfaces.
Since $\vartheta$ is exact, there will be a global time function $t\colon\mathcal{M}\rightarrow\mathbb{R}$ such that $\vartheta\,=\,{\rm d} t$, and the simultaneity surfaces are identified with the level sets of the time function.
In the following, we will always choose Cartesian coordinates $(x^{1},x^{2},x^{3},t)$ reflecting the fact that the $t$-coordinate has a clear and definite physical interpretation being the global time function defining absolute simultaneity.
A simultaneity surface will be denoted by $\Sigma_{\bar{t}}$, where $\bar{t}$ is the value of the time function characterizing the simultaneity surface.
Differently from Klein-Gordon theory, the wave function $\psi$ of Schr\"{o}dinger equation is a complex-valued function on $\mathcal{M}$, which in addition does not transform as a function but as a section of a $U(1)$-bundle \cite{ECM}.
However, we will describe $\psi$ in a real-valued context by considering a pair of fields $(\phi^R, \phi^I)$ associated with any wave function, where $\phi^{R}$ denote the real part of the wave function, and $\phi^{I}$ the imaginary part.
Any field will have its own momenta, say $(P^{0}_R, P^{j}_R)$ and $(P^{0}_I, P^{j}_I)$, and the covariant phase space is $\pi\colon\mathcal{P}=\mathbb{R}^{10}\times \mathcal{M}\rightarrow \mathcal{M}$, where $\pi$ is the projection on the second factor, which is a vector bundle over $\mathcal{M}$ as in the case of Klein-Gordon theory.
Coordinate functions on $\mathcal{P}$ are $(u^a,\rho^{j}_a;\rho^{0}_{a}, x^{j},t)$, with $a=R,I$ and $j= 1,2,3$.
A reference frame on $\mathcal{M}$ is defined by the choice of a nowhere-vanishing vector field $\Gamma$ on $\mathcal{M}$ such tha $\vartheta(\Gamma)=1$ \cite{AuchmannKurz,deritis_marmo_preziosi-a_new_look_at_relativity_transformations,Fecko,MarmoPreziosi, CiagliaDCMarmoSchiavone}.
As we are dealing with Schr\"{o}dinger equation, it seems natural to limit ourselves to inertial Galilean frames.
These are those reference frames characterized by $\Gamma=\frac{\partial}{\partial t} + v^{j}\frac{\partial}{\partial x^{j}}$.
Given $\Gamma$, we can complete it to a frame of vector fields on $\mathcal{M}$ by introducing the vector fields $\frac{\partial}{\partial x^{j}}$ with $j=1,2,3$.
These ``auxiliary'' vector fields are taken to be in the kernel of $\vartheta={\rm d} t$ so that they are tangent to the simultaneity surfaces determined by $\vartheta={\rm d} t$.
Then, we determine the associated dual frame given by $\{\vartheta={\rm d} t,\,{\rm d} x^{j} - v^{j}{\rm d} t\}$ with $j=1,2,3$ and we build the frame-dependent covariant tensor $G_{\Gamma}$ given by
\begin{equation}
G_{\Gamma}\,=\,\delta_{jk}\,\left({\rm d} x^{j} - v^{j}{\rm d} t\right)\otimes \left({\rm d} x^{k} - v^{k}{\rm d} t \right).
\end{equation}
It is straightforward to notice that it determines an Euclidean metric tensor on every simultaneity leaf associated with $\vartheta={\rm d} t$.
Now, by means of $G_{\Gamma}$ we can define the frame-dependent Hamiltonian function given by
\begin{equation}
H_{\Gamma}\,=\,-\frac{\delta_{ab}}{2}\left(\delta_{jk}v^{j}v^{k}\rho^{0}_{a}\rho^{0}_{b} - \delta_{jk}v^{k}(\rho^{j}_{a}\rho^{0}_{b} + \rho^{j}_{b}\rho^{0}_{a}) + \delta_{jk}\rho^{j}_{a}\rho^{k}_{b}\right).
\end{equation}
We stress that, in the non-relativistic case, the Hamiltonian function used for the description of a free quantum particle depends on the choice of a reference frame on $\mathcal{M}$.
This is due to the fact that the covariant tensor $G_{\Gamma}$ is only defined once we determine the frame of vector fields $\{\Gamma,\frac{\partial}{\partial x^{j}}\}$, with $j=1,2,3$, and its dual frame $\{\vartheta, \,{\rm d} x^{j} - v^{j}{\rm d} t\}$ with $j=1,2,3$.
On the other hand, in the relativistic case, we already have the spacetime Lorentzian metric $\eta$ and we do not need the choice of a reference frame to define the Hamiltonian function (see equation \eqref{eqn: relativistic Hamiltonian}).
In the following, we are going to perform our analysis in the inertial reference frame identified by $\vartheta={\rm d} t$ and $\Gamma=\frac{\partial}{\partial t}$, so that $G_{\Gamma}\,=\,\delta_{jk}{\rm d} x^{j}\otimes{\rm d} x^{k}$, and the Hamiltonian function reads
\begin{equation}
H = -\delta_{ab}\delta_{jk} \frac{\rho^{j}_{a}\rho^{k}_{b} }{2},
\end{equation}
where we have set $H_{\Gamma}\equiv H$.
However, everything that will be said below could be adapted to a different choice for $\Gamma$ thus obtaining a description in a different inertial Galilean reference frame. Now, we can define the 4-form $\tilde{\theta}_H$ on $\mathcal{P}$ given by
\begin{equation}
\tilde{\theta}_H = \rho^{0}_a {\rm d} u^a\wedge i_{\Gamma}\mathrm{vol}_\mathcal{M} + \rho^{j}_a {\rm d} u^a\wedge i_{\frac{\partial}{\partial x^{j}}}\mathrm{vol}_\mathcal{M} - H\mathrm{vol}_\mathcal{M} \, .
\label{4-form schrodinger}
\end{equation}
Note that the vector field $\Gamma$ defines the Galielan reference frame, while the vector fields $\frac{\partial}{\partial x^{j}}$ transform as the component of a vector with respect to the action of the Galilei group.
Unlike the relativistic case dealt with in section \ref{sec: KG}, the 4-form $\tilde{\theta}_H$ on $\mathcal{P}$ depends on the choice of a reference frame on $\mathcal{M}$ because the Hamiltonian function does so.
To be able to recover Schr\"{o}dinger equation, we must impose two constraints and select a sub-bundle of $\mathcal{P}$.
Specifically, we consider the vector sub-bundle $\pi\colon\mathcal{Q}\rightarrow\mathcal{M}$ singled out by the constraints
\begin{equation}
\rho^0_R - u^I = 0 ,\quad \rho^0_I + u^R = 0\,.
\end{equation}
The origin of these constraints is related to the circumstance that wave-functions under Galilei transformations behave like sections of a $U(1)$ bundle \cite{ECM}.
Moreover, they appear naturally when the Schr\"{o}dinger equation is written by means of a suitable reduction procedure of a 5-dimensional equation.
This aspect will be considered elsewhere. Let $\chi$ be a section of $\pi\colon\mathcal{Q}\rightarrow\mathcal{M}$ given by
\begin{equation}
\chi(x^{\mu}) = (x^{j},t, \phi^{R}(x^{j},t), P^{k}_{R}(x^{j},t),\phi^{I}(x^{j},t), P^{k}_{I}(x^{j},t))\,.
\end{equation}
As before, we assume some regolarity conditions on the sections we actually consider.
Specifically, we take $\left( \phi^a \right) \in \mathcal{V}=\mathcal{H}^1(\mathcal{M}, \mathrm{vol}_M)$ and $\left( P^{j}_a \right) \in \mathcal{W}= \mathcal{L}^2(\mathcal{M}, \mathrm{vol}_M)$.
In this way, the space of fields of the theory is the Hilbert space $\mathscr{F}_{\mathcal{Q}}\,=\,\mathcal{V}\oplus\mathcal{W}$.
A variation for $\chi\in\mathscr{F}_{\mathcal{Q}}$ is a tangent vector at $\chi$, which may be identified witha vector field $U_{\chi}$ along along $\chi$ on $\mathcal{Q}$, which is vertical with respect to the fibration $\pi\colon\mathcal{Q}\rightarrow\mathcal{M}$.
The tangent space $\mathbf{T}_{\chi}\mathscr{F}_{\mathcal{Q}}$ is given by all the $U_{\chi}$.
In the following, it will be useful to extend $U {\chi}$ to a vertical vector field $\widetilde{U}$ in a neighbourhood of the image of $\chi$ inside $\mathcal{P}$ given by
\begin{equation}
\tilde{U} = U^{\phi}_{R} \frac{\partial }{\partial u^{R}} + U_{R}^{j}\frac{\partial }{\partial \rho^{j}_{R}} + U^{\phi}_{I} \frac{\partial }{\partial u^{I}} + U_{I}^{j}\frac{\partial }{\partial \rho^{j}_{I}}\,.
\end{equation}
As before, the dynamics may be described in terms of the Schwinger-Weiss action principle for sections $\chi\in\mathscr{F}_{\mathcal{Q}}$.
First of all, we consider the pullback $\theta_H $ to $\mathcal{Q}$ of the form $\tilde{\theta}_H $ in Eq.\eqref{4-form schrodinger} given by
\begin{equation}
\theta_H = \left( u^I{\rm d} u^R - u^R{\rm d} u^I \right)\wedge i_{\Gamma}\mathrm{vol}_{\mathcal{M}} + \rho^j_a {\rm d} u^a \wedge i_{\frac{\partial}{\partial x^{j}}}\mathrm{vol}_\mathcal{M} - H\mathrm{vol}_\mathcal{M} \, .
\end{equation}
Then, we define the action functional $S$ on $\mathscr{F}_{\mathcal{Q}}$ as
\begin{equation}
S[\chi]= \int_{\mathcal{M}}\chi^*\left( \theta_H \right)\,,
\end{equation}
with $\chi\in\mathscr{F}_{\mathcal{Q}}$.
At this point, we may proceed as in the previous section and compute the variation $\mathrm{d}S[\chi](U_{\chi})$ of $S$ to be
\begin{equation}
\begin{split}
\mathrm{d}S[\chi](U_{\chi})&= \int_{\mathcal{M} }\chi^*\left( \mathrm{L}_{\tilde{U}}\theta_H \right) = \int_{\mathcal{M} }\chi^*\left( i_{\tilde{U}}{\rm d} \theta_H \right) + \int_{\mathcal{M} } {\rm d} \chi^*\left( i_{\tilde{U}}\theta_H \right),
\end{split}
\end{equation}
where $\tilde{U}$ is any extension of $U_{\chi}$.
Again, we use Stokes' theorem and see that the boundary term vanishes because $\partial\mathcal{M}=\emptyset$.
The action principle states that the dynamical trajectories satisfy the Euler-Lagrange equations for the action functional
\begin{equation}\label{eqn: Schwinger-Weiss action principle 2}
\mathbb{EL}_{\chi}(U_{\chi}) := \int_{\mathcal{M} }\chi^*\left( i_{\tilde{U}}{\rm d} \theta_H \right) = 0 \,,\quad \forall U_{\chi}\in\mathbf{T}_{\chi}\mathscr{F}_{\mathcal{P}}\,,
\end{equation}
which are nothing but the de Donder-Weyl equations
\begin{equation}
\begin{split}
\frac{\partial \phi^I}{\partial t} = - \frac{1}{2}\frac{\partial P^j_R}{\partial x^j} \,, \qquad \frac{\partial \phi^I}{\partial x^j} = - \delta_{jk}P^k_I \\
\frac{\partial \phi^R}{\partial t} = \frac{1}{2}\frac{\partial P^j_R}{\partial x^j} \,, \qquad \frac{\partial \phi^R}{\partial x^j} = - \delta_{jk}P^k_R\,.
\end{split}
\label{deDonder-Weyl-equations Schrodinger}
\end{equation}
It is a matter of straightforward computation to see that the free Schr\"{o}dinger equation for $\psi$ follows from equation \eqref{deDonder-Weyl-equations Schrodinger} upon writing $\psi= \phi^R + i\phi^I$.
The space of solutions is denoted by $\mathcal{EL}_{\mathcal{M}}$, and it is equipped with the two-form
\begin{equation}\label{eqn: omega on solutions of de Donder Weyl for Schroedinger}
\Omega^{\Sigma}_{\chi}(U_{\chi},V_{\chi}) = \int_{\Sigma}i_{\Sigma}^*\chi^*(i_{\tilde{V}}i_{\tilde{U}}{\rm d} \theta_H)\,.
\end{equation}
Once again, following the steps outlined in the companion letter \cite{C-DC-I-M-S-2020-01}, it is possible to show that $\Omega^{\Sigma}$ is actually independent of the simultaneity surface
and frame of reference, and we can simply write $\Omega$.
In the remainder of the section, we want to formulate the dynamics in terms of an action principle for sections of an infinite-dimensional bundle which replaces $\mathcal{Q}$.
In this way, we will be able to read the bracket associated with $\Omega$ in terms of a Poisson subalgebra of the algebra of smooth functions on a suitable manifold endowed with a Jacobi bracket.
We proceed in analogy with what is done in the previous section.
Therefore, we fix a time-slice $\Sigma$ and we consider the pullback bundle $i_{\Sigma}^{*}\mathcal{Q}$, where $i_{\Sigma}$ is the immersion map of $\Sigma$ inside $\mathcal{M}$.
A section $\sigma$ of this bundle is given by
$$
\sigma(x^{j})=\left(x^{j},t^{0}, \varphi(x^{j})^{a},\, \beta^k_{a}(x^{j}) \right),
$$
where $j=1,2,3$ and $a=R,I$, and where
\begin{equation}
\varphi^{a}(x^{j})=(\phi^{a}(x^{j},t^{0})) |_{\Sigma},\quad \beta^k_{a}(x^{j})=(P^{k}_{a}(x^{j},t^{0})) |_{\Sigma} \,,
\end{equation}
for some section $\chi(x^{j},t)\,=(x^{j},t,\phi^{a}(x^{j},t), P^{k}(x^{j},t)$ in $\mathscr{F}_{\mathcal{Q}}$.
As before, we impose some regularity conditions on the admissible sections $\chi$.
Specifically, we will consider $(\varphi^{a}) \in\mathscr{V}=\mathcal{H}^1(\Sigma, \mathrm{vol}_{\Sigma})$, and $(\beta^{j}_{a}) \in\mathscr{B}=\mathcal{L}^2(\Sigma, \mathrm{vol}_{\Sigma})$, so that we have the Hilbert space of fields given by $\mathcal{F}_{\Sigma}\,:=\,\mathscr{V}\oplus \mathscr{B}$.
Then, we consider the Hilbert bundle $\tau\colon\mathcal{F}_{\Sigma}\times\mathbb{R}\rightarrow\mathbb{R}$, where $\tau$ is the projection on the second factor, which plays the role of the extended phase space in \cite{C-DC-I-M-S-2020-01}, and we denote by $\Gamma$ the space of sections of this bundle.
Following what is done in the previous section, we can find a one-to-one correspondence between $\Gamma$ and $\mathscr{F}_{\mathcal{Q}}$. Now, we define the Hamiltonian function
\begin{equation}
\mathcal{H}(\varphi^{a},\beta^{j}_{a};s) = \int_{\Sigma} -\frac{1}{2}\delta^{jk}\left( \frac{\partial \varphi^R}{\partial x^j}\frac{\partial \varphi^R}{\partial x^k} + \frac{\partial \varphi^I}{\partial x^j}\frac{\partial \varphi^I}{\partial x^k} \right) \mathrm{vol}_{\Sigma}=:\int_{\Sigma}\,H\,\mathrm{vol}_{\Sigma}\,,
\end{equation}
the one-form
\begin{equation}\label{eqn: infinite theta2}
\Theta_{\mathcal{H}}=2\int_{\Sigma} \varphi^I \delta\varphi^R \mathrm{vol}_{\Sigma} - \mathcal{H}\wedge {\rm d} t\,,
\end{equation}
and the action functional
\begin{equation}
S[\gamma]=\int_{\mathbb{R}}\,\gamma^{*}\Theta_{\mathcal{H}}.
\end{equation}
Developing the variation of $S$ as done for the Klein-Gordon case, we obtain the constraints
\begin{equation}\label{constraints schroedinger}
\frac{\partial \varphi^R }{\partial x^j} = - \delta_{jk}\beta^k_R\,, \quad \frac{\partial \varphi^I }{\partial x^j} = - \delta_{jk}\beta^k_I
\end{equation}
and the ``evolution equations''
\begin{equation} \label{Cauchy-form schroedinger equation}
\frac{{\rm d} \varphi^R}{{\rm d} s} = - \frac{1}{2} \Delta_{\Sigma}\varphi^I , \quad \frac{{\rm d} \varphi^I}{{\rm d} s} = \frac{1}{2} \Delta_{\Sigma}\varphi^R \,.
\end{equation}
The constraint conditions in Eq.\eqref{constraints schroedinger} determine a submanifold $\mathcal{C}\subset \mathcal{F}_{\Sigma}$ which is a linear subspace.
We define the trivial bundle $\tau_{\mathcal{C}}\colon\mathcal{C}\times\mathbb{R}\rightarrow\mathbb{R}$, where $ \tau_{\mathcal{C}}$ is the projection on the second factor, and we denote by $\Gamma_{\mathcal{C}}$ the space of sections of this bundle.
Moreover, we write, with an evident abuse of notation, $\Theta_{\mathcal{H}}$ for the pullback to $\mathcal{C}\times\mathbb{R}$ of the one-form in equation \eqref{eqn: infinite theta2}.
Then, on $\mathcal{C}\times\mathbb{R}$, it is clear that Eq.\eqref{Cauchy-form schroedinger equation} may be read as defining the integral curves of the (densely defined) vector field
\begin{equation}
X_{H} = \frac{\partial}{\partial t} + \frac{\delta H}{\delta \varphi^{I}}\,\frac{\delta}{\delta \varphi^{R}} - \frac{\delta H}{\delta \varphi^{R}}\,\frac{\delta}{\delta \varphi^{I}}\,
\end{equation}
which is in the kernel of the two-form
\begin{equation}
{\rm d} \Theta_{\mathcal{H}}\, =\, 2\int_{\Sigma}\left( \delta \varphi^{I} \wedge \delta\varphi^{R}\right)\mathrm{vol}_{\Sigma} - {\rm d}\mathcal{H} \wedge {\rm d} t \,.
\end{equation}
This two-form plays the role of the contact two-form on $\mathbf{T}^{*}\mathcal{Q}\times\mathbb{R}$ in the case of non-realtivistic Hamiltonian mechanics, and of the contact two-form on the mass-shell in the case of the relativistic particle considered in \cite{C-DC-I-M-S-2020-01}.
Eventually, we obtained the de Donder-Weyl equations for the free quantum particle in $\mathbb{R}^{3}$ as a Hamiltonian system on an infinite-dimensional manifold, in such a way that a section $\gamma$ satisfying Eq.\eqref{Cauchy-form schroedinger equation} provides a representation in terms of Cauchy data on $\Sigma$ of the points in $\mathcal{EL}_{\mathcal{M}}$.
We denote by $\mathcal{EL}_{\mathcal{C}}$ the space of all $\gamma\in\Gamma_{\mathcal{C}}$ satisfying Eq.\eqref{Cauchy-form schroedinger equation}.
Again, all vector fields $fX_{H}$ with $f$ a non-vanishing function are in the kernel of ${\rm d}\Theta_{\mathcal{H}}$, and the integral curves of $f\,X_{H}$ can be interpreted as reparametrizations of the dynamical trajectories.
Moreover, just as we said for the free Klein-Gordon theory, under suitable regularity properties for $X_H$, the family of its integral curves defines a regular foliation of the manifold $\mathcal{C}\times \mathbb{R}$, and every point in the quotient manifold, say $\mathrm{Q}$, associated with the foliation can be identified with one and only one element in $\mathcal{EL}_{\mathcal{C}}$.
This allows us to look at the space of smooth functions on $\mathrm{Q}\cong\mathcal{EL}_{\mathcal{C}}\cong\mathcal{EL}_{\mathcal{M}}$ as the subalgebra $C^{\infty}_H(\mathcal{C}\times \mathbb{R})\subset C^{\infty} (\mathcal{C}\times \mathbb{R})$ of smooth functions such that $\mathrm{L}_{X_H}f=0$.
Then, as we did for the Klein-Gordon equation, we may look for a bivector $\Lambda$ on $\mathcal{C}\times \mathbb{R}$ satisfying the relations
\begin{equation}
\left[\Lambda,\Lambda\right]_{S}\,=\,2\,\Gamma_{H}\,\wedge\,\Lambda\,,\quad\,\mathrm{L}_{\Gamma_{H}}\Lambda\,=\,0,
\end{equation}
where $[\cdot , \cdot ]_{S}$ denotes the Schouten-Nijenhuis bracket and $\Gamma_H$ is the Reeb vector field satisfying $i_{\Gamma_H}\Theta_{\mathcal{H}}=1$. The associated Jacobi bracket\cite{AsoCiagliaDCosmoIbortMarmo2017-Covariant_Jacobi_brackets,C-C-M-2018} $[\cdot , \cdot ]_{J}$ on $ C^{\infty} (\mathcal{C}\times \mathbb{R})$ is then defined as follows
\begin{equation}
[f,g]_{J}\,:=\,\Lambda(\mathrm{d}f,\mathrm{d}g) + f\,\mathrm{L}_{\Gamma_{H}}g - g\,\mathrm{L}_{\Gamma_{H}}f\,.
\end{equation}
Recalling that $\Gamma_{H}$ annihilates the elements in $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$, it is immediate to check that $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$ becomes a Poisson subalgebra of $C^{\infty}(\mathcal{C}\times \mathbb{R})$ with respect to the Jacobi bracket defined above.
Upon identifying $\mathcal{EL}_{\mathcal{M}}$ with $\mathcal{EL}_{\mathcal{C}}$, and then $\mathcal{EL}_{\mathcal{C}}$ with $\mathrm{Q}$, the Poisson bracket on $C^{\infty}_H(\mathcal{C}\times \mathbb{R})$ is precisely the Poisson bracket associated with the two-form $\Omega$ in Eq. \eqref{eqn: omega on solutions of de Donder Weyl for Schroedinger}.
To explicitely write the Jacobi and Poisson bracket, we can look for an analogue of the generalized Darboux coordinates used in the particle case in the companion letter \cite{C-DC-I-M-S-2020-01}.
At this purpose, we first move from $\mathcal{C}$ to $\mathcal{C}'$ introducing the Fourier transforms
\begin{equation}
\varphi_{R}(x)\,=\,\int_{\overline{\Sigma}}\,\hat{\varphi}_{R}(k)\,\mathrm{e}^{i k\cdot x}\,\,\mathrm{vol}_{\overline{\Sigma}}, \quad \varphi_{I}(x)\,=\,\int_{\overline{\Sigma}}\,\hat{\varphi}_{I}(k)\,\mathrm{e}^{i k\cdot x}\,\,\mathrm{vol}_{\overline{\Sigma}}.
\end{equation}
Note that the fact that $\varphi_{R}$ and $\varphi_{I}$ are real-valued imposes the constraints $\overline{\hat{\varphi}}_{R}(k)=\hat{\varphi}_{R}(-k)$ and $\overline{\hat{\varphi}}_{I}(k)=\hat{\varphi}_{I}(-k)$.
Now, we consider the ``change of coordinates'' in $\mathcal{C}'\times\mathbb{R}$ given by
\begin{equation}
\begin{split}
W\,=\, \frac{1}{2}\,\int_{\overline{\Sigma}}\ \left( |\hat{\Phi}_{R}|^{2} - |\hat{\Phi}_{I}|^{2}\right) &\sin(k^{2}s) + 2 (\hat{\Phi}_{R}\overline{\hat{\Phi}}_{I} + \overline{\hat{\Phi}}_{R}\hat{\Phi}_{I})\sin\left(\frac{ k^{2}}{2} s\right) \,\mathrm{vol}_{\overline{\Sigma}} \\
\hat{\Phi}_{R} &\,=\, \hat{\varphi_{R}}\,\cos\left(\frac{k^{2}}{2}s\right) - \hat{\varphi_{I}}\,\sin\left(\frac{k^{2}}{2}s\right)\\
\hat{\Phi}_{I} &\,=\, \hat{\varphi_{I}}\,\cos\left(\frac{k^{2}}{2}s\right) + \hat{\varphi_{R}}\,\sin\left(\frac{k^{2}}{2}s\right) ,.
\end{split}
\end{equation}
where, again, it is $\overline{\hat{\Phi}}_{R}(k)=\hat{\Phi}_{R}(-k)$ and $\overline{\hat{\Phi}}_{I}(k)=\hat{\Phi}_{I}(-k)$.
In this coordinate system, it is possible to see that the one-form $\Theta_{\mathcal{H}}$ in equation \eqref{eqn: infinite theta2} becomes
\begin{equation}
\Theta_{\mathcal{H}}\,=\,2\int_{\overline{\Sigma}} \,\overline{\hat{\Phi}}_{I} \,\delta\hat{\Phi}_{R} \,\mathrm{vol}_{\overline{\Sigma}} + {\rm d} W\,.
\end{equation}
The Reeb vector field $\Gamma_{H}$ becomes $\Gamma_{H}\,=\,\frac{\partial}{\partial W}$, and, in analogy with what we did in the particle case in the companion letter \cite{C-DC-I-M-S-2020-01}, the bivector field $\Lambda$ in the definition of the Jacobi bracket reads
\begin{equation}
\Lambda\,=\,\frac{1}{2}\int_{\overline{\Sigma}}\,\left(\frac{\delta}{\delta \hat{\Phi}_{R} } - \hat{\Phi}_{I} \,\frac{\partial}{\partial W}\right)\,\wedge\,\frac{\delta}{\delta \overline{\hat{\Phi}}_{I} }\,\mathrm{vol}_{\overline{\Sigma}} \,.
\end{equation}
Clearly, elements in $C^{\infty}_H(\mathcal{C}'\times \mathbb{R})$ are just those functions which do not depend on $W$, and thus the Jacobi bracket among them becomes the Poisson bracket given by
\begin{equation}
\Lambda({\rm d} F,{\rm d} G)\,=\,\frac{1}{2}\int_{\overline{\Sigma}}\, \frac{\delta F}{\delta \hat{\Phi}_{R} } \,\wedge\,\frac{\delta G}{\delta \overline{\hat{\Phi}}_I }\,\mathrm{vol}_{\overline{\Sigma}} \,.
\end{equation}
\section{Conclusions}
In this letter, we continued the analysis of the description of the covariant bracket on the space of functionals on solutions to variational problems in the framework of contact geometry initiated in the companion letter \cite{C-DC-I-M-S-2020-01}.
We analysed in detail the case of free Klein-Gordon theory on Minkowski spacetime, and of the free Schr\"{o}dinger equation for a particle in $\mathbb{R}^{3}$.
Both systems were described by means of two different but equivalent formulations.
On the one hand, we exploited the multisymlectic formalism to give a description in which the fields of the theory are sections of a suitable bundle over the spacetime manifold, but for which it is not possible to appreciate the role of contact geometry.
Indeed, this formulation seems to point at the need to develop a generalization of contact geometry which is analogue to the multisymplectic generalization of symplectic geometry, and that will be pursued in a future work.
On the other hand, we presented a description in terms of a vector field on a suitable infinite-dimensional manifold that leads to the identification of the covariant bracket in terms of a Poisson subalgebra of the algebra of smooth functions on the infinite-dimensional manifold endowed with the Jacobi bracket.
\section*{Acknowledgments}
F.D.C. and A.I. would like to thank partial support provided by the MINECO research project MTM2017-84098-P and QUITEMAD++, S2018/TCS-A4342. A.I. and G.M. acknowledge financial support from the Spanish Ministry of Economy and Competitiveness, through the Severo Ochoa Programme for Centres of Excellence in RD(SEV-2015/0554). G.M. would like to thank the support provided by the Santander/UC3M Excellence Chair Programme 2019/2020, and he is also a member of the Gruppo Nazionale di Fisica Matematica (INDAM), Italy.
| 03e6455a1410ea6df9759c3b2f1886ae9435579b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Modern time series analysis has been transformed by emerging and innovative mathematical methods from machine learning and data science. The ubiquitous availability of measurement data across the sciences, coupled with advances in storage and processing power, has led to a resurgence of interest in the question of how diagnostic and predictive models can be discovered directly from time series data. Using a dynamical systems perspective allows one to take advantage of a broad variety of existing systems analysis methods with applications to forecasting and control. Of particular interest are techniques which are able to map an observed measurement series into a space where it can be accurately reproduced by a set of linear governing equations. This is the central tenet of Koopman theory, first introduced in 1931, which states that nonlinear dynamical systems can be reproduced by linear evolution in a space of observables on the state variables~\cite{koopman1931,koopmanvonneumann1932}. However, this linear representation typically comes at the price of dimensionality: finite-dimensional nonlinear evolution generically requires an infinite-dimensional lifting to be accurately recreated by a linear operator. Linear models are desirable because they allow for the use of many powerful linear algebraic methods for estimation, prediction, and control, all of which can be extended to nonlinear systems if one is able to construct an accurate finite-dimensional coordinate system for the Koopman operator~\cite{Brunton2019book}. As we show, time-delay embeddings provide such a coordinate system, providing a mathematical framework for the extraction of {\em principal component trajectories} (PCT) which allows for the construction of dynamics, actuated or not, via superposition.
Dynamic Mode Decomposition (DMD) is the leading approach to approximating Koopman operators through regression~\cite{Schmid2010jfm,rowleymezicetal2009,Kutz2016book}. Specifically, it offers a means of efficiently discovering a finite-rank linear approximator to nonlinear dynamics directly from observation data, even in very high dimensions. The simplest implementation of the algorithm regresses an operator which acts directly on the measured state variables, but there is no reason, \textit{a priori}, to expect this space to admit a faithful linear representation. Central to the successful implementation of DMD is the problem of identifying a space of Koopman observables into which the dynamics can be lifted to better suit the assumption of linearity \cite{mezic2004,Mezic2005nd,Budivsic2012chaos,Mezic2013arfm}. A variety of approaches have been taken, including constructing libraries of simple monomial functions evaluated on the data and regressed densely \cite{williamskevrekidisrowley2015} or sparsely \cite{Brunton2016pnas,Brunton2016plosone}, kernel methods such as diffusion mapping \cite{Giannakis2019}, and deep learning of coordinate transformations \cite{Takeishi2017nips,Lusch2018,Mardt2018natcomm,Wehmeyer2018jcp}. The method of interest in this work is the use of time-delay embedding of measurement data, which has been used to great effect in the HAVOK~\cite{bruntonproctorkaiserkutz2017} and the subsequent Hankel DMD~\cite{arbabimezic2017} algorithms.
DMD on time-delay coordinates offers advantageous properties with respect to Koopman approximation which are universal to any input data \cite{kamb18}.
Time-delay embedding refers to a coordinate transformation in which a time-localized measurement $x(t)$ is augmented by time-shifted copies of itself $x(t-\tau)$. The use of this technique for data-driven modeling dates back to the seminal Takens embedding theorem, which showed that a chaotic attractor can be reconstructed, up to a diffeomorphism, from the time series of a single measured variable \cite{takens1981}.
The method has since found purchase in a number of widely-used algorithms, including {\em singular spectrum analysis} (SSA) \cite{broomhead1989}, {\em nonlinear Laplacian spectral analysis} (NLSA) \cite{giannakis2012}, and the {\em eigensystem realization algorithm} (ERA) \cite{juangpappa1985}. Particular attention has been devoted to the use of techniques such as singular value decomposition (SVD) to identify dominant modal content of data represented in delay coordinates \cite{hassani07,kamb18}.
For a time-delayed scalar measurement series, the principal components obtained by SVD form a temporal basis, with functions similar to those in a Fourier or wavelet basis.
Indeed, it has been demonstrated that for sufficiently long embedding windows the delay coordinate SVD converges to an estimate of the discrete Fourier decomposition \cite{vautardghil1989}. This property suggests a strong compatibility with the Koopman operator-theoretic approach to systems modeling, in which the eigenvalue spectrum of the desired operator has been shown to relate to the same harmonic averages used to compute the Fourier transform \cite{Mezic2005nd}. This connection has been borne out by recent work on time-delay DMD~\cite{turowleyetal2014,bruntonjohnsonojemannkutz2016}, variations on which form analogous models on their SVD projections \cite{bruntonproctorkaiserkutz2017,arbabimezic2017} or on exact Fourier basis projections \cite{panduraisamy2019}.
The fidelity of a DMD model depends heavily on the Fourier spectral properties of the true dynamics. A finite-dimensional linear operator generates dynamics on a discrete set of frequencies determined by its eigenvalues; if the input data has a continuous spectrum it cannot be fully reproduced by DMD (time-delayed or otherwise). Moreover, the number of discrete spectral peaks that can be captured by a linear model depends on its dimension. DMD represents oscillatory modes by complex conjugate eigenvalue pairs, so a rank-$r$ model will admit at most $r/2$ distinct frequency components. This offers one perspective on the motivation for building models in delay coordinates: a discrete eigenvalue spectrum of arbitrary finite length can be obtained simply by embedding additional shifted copies of the data to increase the dimension of the augmented space. It can be shown that for scalar time series, the minimum number of such embeddings is determined fully by the sparsity of the Fourier spectrum \cite{liebertschuster1988,panduraisamy2019}. The continuous spectrum case, however, presents a much greater challenge for Koopman modeling. While a sufficiently dense point spectrum can approximate a continuous one in the high-dimensional embedding limit \cite{Das2019}, the practical utility of this for numerical methods is limited. It has long been suggested \cite{Mezic2005nd} that a Koopman approach might be taken for the ``almost periodic'' (i.e. spectrally discrete) portion of dynamics and augmented somehow to account for the continuous remainder. Previous attempts at this have used intermittent parametric forcing derived from rank-reduced SVD time series \cite{bruntonproctorkaiserkutz2017} or direct translational manipulation of eigenvalues \cite{Lusch2018} to fill out the spectrum.
Building on the approach of \cite{bruntonproctorkaiserkutz2017}, this paper makes two principal methodological contributions to the problem of modified Koopman representations for such systems. First, we introduce a means of deriving a novel projectional basis for dynamics using time delay methods inspired by SSA. This "PCT" representation offers properties generically favorable to approximation of the Koopman operator \textit{and} a basis which naturally decomposes a signal into its continuous and discrete spectral components. Second, we propose a technique for Koopman modeling of such mixed-spectrum systems using the DMD with Control (DMDc) algorithm~\cite{proctorbruntonkutz2016}, which fits measurement data to a linear control model using a known control signal. We present a fully unsupervised method to learn a control model and a concomitant forcing signal which together fully reproduce the observed dynamics, with discrete-spectrum behavior modeled by endogenous linear evolution and the continuous-spectrum remainder isolated in the exogenous control. This approach is made possible by the projection of dynamics onto the PCT basis, in which components of the dynamics which are not amenable to low-rank Koopman representation are by default relegated to the projectional residual.
The remainder of this paper is structured as follows: Sec. \ref{sec:delay_embedding} summarizes the process by which delay coordinate representations are obtained from data. The approach is modeled on that of SSA, but focuses on the less common case of spatiotemporal state delay embedding. We present a novel interpretation for the coordinate bases that this generates and illustrate the information-richness of the resultant modes relative to the highly generic output of the same approach on scalar data. Sec. \ref{sec:dmd} discusses the regression of linear DMD models in the learned delay-coordinate space, using data with discrete and mixed spectra. A procedure is introduced for the extraction of an external forcing signal to account for observed nonperiodic dynamics, with results validating its success on systems with known exogenous forcing. Sec. \ref{sec:dmdc} joins this method with the existing DMDc algorithm, allowing these results to be integrated into a single unified linear control model. Finally, Sec. \ref{sec:power_grid} offers an example of how this technique might be applied to a real-world data set as an informative diagnostic tool.
\section{Full-State Embedding: Methods and Interpretations}
\label{sec:delay_embedding}
In this section we outline a procedure for time-delay analysis of a measurement time series. The steps taken are similar to those of SSA. But where SSA typically entails delay embedding on a scalar signal, we extend this approach to the case of a multivariate vector signal and briefly discuss the interpretation of the principal components that this produces.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{mode_trajectory_protocol_edit.pdf}
\caption{Protocol for extracting delay-embedded SVD trajectories, or {\em principal component trajectories} (PCT), from time series data.}
\label{fig:mode_trajectory_protocol}
\end{figure*}
\subsection{Time Delay Embedding}
Delay embedding is the process of lifting a time series signal into a higher dimensional space by stacking it with time-shifted copies of itself. For a scalar $y(t)\in\mathbb{R}$, the lifted Hankel matrix $\mathbf{H}$ is obtained as follows:
\begin{equation}
\begin{split}
\mathbf{Y} &= \left[\begin{matrix} y(t_0) & y(t_1) & \cdots & y(t_m)\end{matrix}\right]
\end{split}
\end{equation}
becomes
\begin{equation}
\begin{split}
\mathbf{H} &= \left[\begin{matrix} y(t_0) & y(t_1) & \cdots & y(t_{m-d}) \\
y(t_1) & y(t_2) & \cdots &y(t_{m-d+1})\\
\vdots & \vdots & \ddots & \vdots\\
y(t_d) & y(t_{d+1}) & \cdots &y(t_m)\end{matrix}\right]
\end{split} .
\label{eq:scalar_delay}
\end{equation}
The matrix
${\bf H}$ contains $d$ shifted copies of the original signal ${\bf Y}$, each offset from the last by one time step. If ${\bf Y}$ had dimension $1\times m$, ${\bf H}$ has dimension $d \times (m-d+1)$. Note that there is considerable redundancy in the elements of ${\bf H}$: all elements belonging to diagonal sets $\{H_{ij}\}$ such that $i+j = \rm(const)$ are equal.
The extension of this procedure to a vector signal $\mathbf{y} \in \mathbb{R}^n$ is straightforward. Each column of ${\bf H}$ now consists of $d$ vectors in $\mathbb{R}^n$ stacked on top of each other with the same time shifting scheme as before:
\begin{equation}
\begin{split}
\mathbf{Y} &= \left[\begin{matrix} | & | & & | \\
\mathbf{y}(t_0) & \mathbf{y}(t_1) & \cdots & \mathbf{y}(t_m) \\
| & | & & | \end{matrix}\right]
\end{split}
\end{equation}
becomes
\begin{equation}
\begin{split}
\mathbf{H} &= \left[\begin{matrix} | & | & & | \\
\mathbf{y}(t_0) & \mathbf{y}(t_1) & \cdots & \mathbf{y}(t_{m-d}) \\
| & | & & |\\
| & | & & | \\
\mathbf{y}(t_1) & \mathbf{y}(t_2) & \cdots & \mathbf{y}(t_{m-d+1}) \\
| & | & & |\\
\vdots & \vdots & \ddots & \vdots \\
| & | & & | \\
\mathbf{y}(t_d) & \mathbf{y}(t_{d+1}) & \cdots & \mathbf{y}(t_m) \\
| & | & & |\end{matrix}\right]
\end{split}.
\label{eq:delay_embed}
\end{equation}
The vectorized Hankel matrix is often used for linear system identification in control applications via ERA~\cite{juangpappa1985}.
This procedure offers two tunable parameters to be independently chosen. Firstly, the number of delay embeddings $d$ performed on the input signal. This determines the dimension of the delay coordinate space into which the columns of ${\bf y}$ are lifted. Previous work on this topic has sought to quantify the lowest value $d$ can take while still admitting a linear dynamical model which faithfully reproduces observed dynamics \cite{broomheadking1986,Gibson1992phD,panduraisamy2019,liebertschuster1988}. For the purposes of this analysis, however, we assume the availability of input data of arbitrary duration, so there is no need for parsimony in the selection of $d$. While a high value of $d$ does lead to a high-dimensional embedding space, our subsequent model-building efforts will be carried out in a rank-reduced basis spanned by $r$ principal component vectors, or the {\em principal componant trajectories} (PCT) of the time-series. Thus a large number of embeddings $d$ does not carry with it any risk of causing the model regression to be underdetermined.
The second parameter to be selected is the embedding period $T^d$. In Eq.~(\ref{eq:delay_embed}) the time lag introduced in each successive embedding is equal to the sampling resolution of ${\bf y}(t)$, so $T^d = t_d-t_0 = d\Delta t$, assuming the columns of ${\bf Y}$ are sampled at evenly spaced intervals of $\Delta t$. However, the time shift per embedding could just as easily be set to any integer multiple of $\Delta t$, leading to a modified embedding period $T^d = qd\Delta t$, where $q\in \mathbb{Z}^+$. The governing principle for making this choice should be to match the time scale of the dynamics of interest: in the low-$T^d$ limit the stacked state vectors which comprise the columns of ${\bf H}$ will be highly redundant, encoding the passage of a time interval too short for any meaningful dynamical evolution of the system. In the high-$T^d$ limit, the delay coordinate sampling resolution is too coarse to resolve the dynamics; unless ${\bf y}(t)$ is perfectly periodic the delay registers of ${\bf H}$ will become completely uncorrelated. If the system in question contains multiscale dynamics with periods separated by orders of magnitude, multiple sets of embedding parameters may be required. A more extensive discussion of this case with respect to the application of dynamic mode decomposition can be found in \cite{dylewsky2019}. For the purposes of this paper, we restrict our analysis to monoscale dynamics and choose $T^d$ such that it spans a few periods of the lowest-frequency peak of the Fourier spectrum of ${\bf y}(t)$. A more detailed discussion of this parameter selection is available in \cite{Gibson1992phD}.
\subsection{SVD and Full-State Embedding}
Of central importance to the signal processing technique of SSA \cite{broomheadking1986,Fraedrich1986,vautardghil1989}, and to more recent dynamical systems work with time delay coordinates \cite{arbabimezic2017,bruntonproctorkaiserkutz2017,kamb18}, is the application of singular value decomposition (SVD) to the delay-embedded Hankel matrix ${\bf H}$ (as defined in Eq.~(\ref{eq:delay_embed})). In this section we present a brief discussion of the motivation for this step, convergence properties of the results, and interpretations of the modes in delay space specifically for the case of full-state vector embedding.
The SVD of a matrix $\mathbf{Y}$ yields $\mathbf{Y} = \mathbf{U} \mathbf{S} \mathbf{V}^T$. If $\mathbf{Y}^{n\times m}$ is a time series with state dimension $n$ and time dimension $m$, then the columns (or \textit{modes}) of $\mathbf{U}$ form an orthonormal basis for the state space, the columns of $\mathbf{V}$ are time series representing normalized projections of the state onto the $\mathbf{U}$ modes, and the elements of the diagonal matrix $\mathbf{S}$ carry the energetic (correlation) weight of each mode's contribution to the full signal $\mathbf{Y}$. Because modal amplitudes are encoded entirely by $\mathbf{S}$, modes can be hierarchically ranked by relative magnitude. In the examples presented in this paper we assume each coordinate of the input data has mean 0, which can be accomplished for arbitrary $\mathbf{Y}$ by applying the transformation $\mathbf{Y}\to\mathbf{Y}-\bar{\mathbf{Y}}$. Under this condition the SVD is equivalent to Principal Component Analysis (PCA), and the singular values can be interpreted as the variance of the data along the axes of the left singular vectors (the columns of $\mathbf{U}$). Truncating the decomposition to contain only the top $r$ modes yields a rank-reduced representation which produces a least-squares optimal reconstruction of $\mathbf{Y}$. [Note that strictly speaking, centering $\mathbf{Y}$ on the origin does not precisely center $\mathbf{H}$; each row of $\mathbf{H}$ contains a truncated subset of the corresponding row in $\mathbf{Y}$. However, we are concerned with the limit of $m>>d$ in which this discrepancy is negligible, so mean subtraction of either matrix yields results which are equivalent in a practical sense.]
The SVD can be applied to a time series lifted into delay coordinates, as in Eq. \ref{eq:delay_embed}, to obtain a decomposition of the same form:
\begin{equation}\label{HankelSVD}
\begin{split}
\mathbf{H} = \mathbf{U} \mathbf{S} \mathbf{V}^T
\end{split}
\end{equation}
Resultant modes have additional degree of interpretability owing to the temporal structure embedded in $\mathbf{H}$ . Columns of $\mathbf{U}$ are vectors in the input state space, which in this case is the high-dimensional delay space. If $\mathbf{H}$ was constructed from a scalar time series $y(t)$, the SVD modes can be treated as $d$-element time series spanning a duration of $T^d$. If $\mathbf{H}$ were instead the full-state embedding of a vector times series $\mathbf{y}(t) \in \mathbb{R}^n$, its modes can be interpreted as trajectories in $\mathbb{R}^n$, also of duration $T^d$. An intuitive parallel for these decompositions can be found in time-frequency analysis methods such as windowed Fourier and wavelet transforms, both of which have been applied to the DMD method~\cite{kutz2016multiresolution,dylewsky2019}. Thus a vector projection of ${\bf H}$ onto the columns of $\mathbf{U}$ in delay coordinates is analogous to a time-localized projection onto some template function, e.g. a wavelet confined to a length of $T^d$.
This procedure is shown in Fig. \ref{fig:mode_trajectory_protocol}. Note that when the input signal is multivariate, vectors in the delay embedding space have elements spanning both spatial coordinates $x,y,z$ and measurement times $t_j, t_{j+1},...,t_{j+d}$. To treat this as a normed vector space (as SVD does) is to make an implicit assumption about the relative weighting of spatial versus temporal coordinate separations. This assumption is carried in the choice of time spacing between delay embeddings, or equivalently in the choice of embedding period $T^d$. A more concrete resolution could be obtained by explicitly defining a space-time inner product in the Hankel vector space as described in \cite{schmidt2019}, which would also offer a more rigorous basis for interpreting singular value weightings as mode energies of the PCTs.
\subsection{Interpreting SVD Modes in Delay Coordinates}
The result of the decomposition process outlined in Fig. \ref{fig:mode_trajectory_protocol} depends heavily on the structure of ${\bf H}$. By construction, elements of ${\bf H}$ are represented redundantly up to $d$ times (specifically, all $H_{ij}$ where $i+n\times j=(\rm{const})$ are equal). This property holds almost regardless of the dynamical content of the signal in question, and leads to certain guarantees on the SVD modes of ${\bf H}$. In particular, it can be shown that for embeddings of a scalar time series ($n=1$), modes converge to simple functions in the limiting cases of $T^d$. For small $T^d$, the columns of ${\bf U}$ resemble Legendre polynomials, while for large $T^d$ they become sinusoidal \cite{broomheadking1986,vautardghil1989}. (An exception to this convergence property would occur in a purely mixing system with continuous spectrum measured with perfect precision; here all but one singular value would approach 0 in the $d\to\infty$ limit. But in the practical case of large-but-finite $d$ this does not pose any problem.) Indeed, in some circumstances exact instantiations of these functions have been used instead of the data-driven approximations obtained with the SVD \cite{Gibson1992phD,panduraisamy2019}. In this paper we focus on the large-$T^d$ limit, in which delay-coordinate SVD can be thought of as reproducing the discrete Fourier transform on a sparsified frequency basis: the near-sinusoidal ${\bf U}$ modes resemble the projection basis used in the DFT, but instead of a large number of basis functions densely populating frequency space, only the (SVD rank) $r$ most prominently represented frequencies are retained.
Shifting attention to the case of multivariate vector embeddings, we observe similar behavior in the long embedding time limit. Some sample modes obtained for the chaotic Lorenz system are plotted in Fig. \ref{fig:mode_trajectory_breakdown}. The trajectories in $\mathbb{R}^3$ traced out by the modes result from sinusoidal oscillation in $x$, $y$, and $z$. The frequencies of oscillation for different state variables within the same mode are typically equal or related by an integer ratio, but the relative phases and amplitudes are free to vary. Consequently, modes which in the scalar high-$T^d$ case could be completely defined by a single frequency value take on much richer geometric structure in the vector high-$T^d$ case. The figures which can be thus formed by parametric sinusoids belong to the class of generalized $N$-dimensional Lissajous figures \cite{albertazzi2013}, in the special case where frequencies differ by integer ratios. Note that the periods of oscillation are independent of the embedding period $T^d$, so these curves do not in general form closed orbits.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{mode_trajectory_breakdown_edit.pdf}
\caption{Time-delayed SVD modes for a chaotic Lorenz system. Full 3D representations of the PCTs (left) are composed of approximately sinusoidal oscillation in the $x,y,z$ coordinates with (often) incommensurate frequencies and amplitudes.}
\label{fig:mode_trajectory_breakdown}
\end{figure}
The SVD reconstruction~\eqref{HankelSVD} can be equivalently written as a sum over individual modes:
\begin{equation}
\begin{split}
{\bf H} &= \sum_j {\bf u}_j s_j {\bf v}_j^*
\end{split}
\end{equation}
Each column of ${\bf H}$, which represents a windowed snapshot of the dynamics over a period of $T^d$, is reproduced by a linear combination of the orbits $\{{\bf u}_j\}$ weighted by the singular values $\{s_j\}$ and the time series projections $\{{\bf v}_j(t)\}$. The delay-coordinate SVD therefore functions as an automated means of decomposing a complex trajectory into simple orbital components (Fig. \ref{fig:spirograph3D}), or the PCT coordinates of interest. The result is reminiscent of the Ptolemaic model of the solar system, in which the apparent retrograde motion of planets in the sky is explained using a hierarchy of superimposed epicycles in geocentric coordinates.
It should be noted that although PCTs often look periodic as a result of the Fourier converge property discussed above, they do not correspond to true closed orbits on the original attractor. They should be interpreted as an optimal basis for low-rank projection in the spirit of Principal Component Analysis; the dynamics of the system are unlikely to resemble any given one of the PCTs, but should be reproducible by their linear superposition. The novelty of PCTs relative to traditional PCA and SSA lies in the dynamical richness of the modes. Multivariate time delay embedding lifts data into a high dimensional, hybrid spatial-temporal representation which admits a modal decomposition which offers a much more compact, low-rank representation for a large class of dynamics than other principal component or Fourier projection methods.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{spirograph3D_edit3.pdf}
\caption{Pointwise linear combinations of simple modal orbits are used to reconstruct a complex trajectory in $\mathbb{R}^3$ (a). Modes are grouped together (e.g. $u_1+u_2$) because delay-coordinate SVD naturally identifies pairs of modes which are identical up to a $90^\circ$ phase shift (b). The trajectory reconstituted from these 8 modes is plotted in (c).}
\label{fig:spirograph3D}
\end{figure}
\subsection{PCT of a Nonlinear Oscillator with Discrete Fourier Spectrum}
A more concrete illustration of the analogy to multivariate Fourier decomposition can be found in the example of the Van der Pol system. This canonical nonlinear oscillator is defined by the coupled first-order equations:
\begin{equation}
\begin{split}
\dot{y}_1 &= y_2\\
\dot{y}_2 &= \mu\left(1-y_1^2\right)y_2 - y_1
\end{split}
\label{eq:vdp_def}
\end{equation}
with $\mu$, which controls the strength of the nonlinearity, taken to be $1$ from here on. This system admits solutions which form a closed periodic orbit, so the Fourier spectrum of the trajectory it generates is discrete (only integer harmonics of the base periodic frequency are present). Applying the procedure from Fig. \ref{fig:mode_trajectory_protocol} with $T^d = 20$ produces the delay-coordinate modes plotted in Fig. \ref{fig:vdp_svd_modes}. The state space of Eq. \ref{eq:vdp_def} is represented by the $x-y$ plane, and the time coordinate on the interval $[0, T^d]$ is extended up the $z$ axis to better illustrate frequency distinctions. In the rightmost column, Fourier spectral content of each (combined) mode pair is overlaid on the spectrum of the full Van der Pol trajectory.
The results for this simple example are illustrative of three key features of vector-embedded PCT. First, the modes naturally adhere to the frequencies present in the power spectrum of the input signal. In other words, the PCT not only converge to a sinusoidal basis, but it discovers a sparse Fourier basis made up of those sinusoids most strongly represented in the data. Second, the ordering of the modes corresponds to the heights of their corresponding spectral peaks. This follows from the fact that in the limit as the columns of ${\bf U}$ converge to sinusoids, the singular values associated with them become directly analogous to Fourier power measurements in that both represent a magnitude of the energetic content of the full dynamics contained in a projection onto a given oscillatory mode. Third, the PCT modes naturally self-organize into pairs which share a common frequency and singular value, which is often observed in the fluid flow past bluff bodies~\cite{Noack2003jfm,Taira2017aiaa}. In each row in Fig. \ref{fig:vdp_svd_modes}, the modes are identical up to a $90^\circ$ rotation. This again suggests a parallel to Fourier analysis: a pair of out-of-phase sinusoids of equal frequency (e.g. $\sin(\omega t)$ and $\cos(\omega t)$) form a complete basis for dynamics of arbitrary phase at that frequency.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{vdp_svd_modes.pdf}
\caption{First 6 SVD modes for the delay-embedded Van der Pol oscillator. Each mode can be thought of as a trajectory in $\mathbb{R}^2$, with the $z$ coordinate here representing the delay time. They organize into pairs $90^{\circ}$ out of phase in order to form a complete orthogonal basis for a given frequency. Power spectra plotted on the right show that the dominant SVD modes naturally correspond to the frequencies most prominently represented in the signal.}
\label{fig:vdp_svd_modes}
\end{figure}
\subsection{PCT and Continuous-Spectrum Dynamics}
As the spectral plots in Fig. \ref{fig:vdp_svd_modes} suggest, PCT is a method suited to systems whose dynamics admit a discrete Fourier representation. In the long-embedding limit in which PCT modes converge to sinusoids, a decomposition of finite rank $r$ can at most reproduce $r/2$ spectral peaks in its reconstruction. For a system with a continuous spectrum, this is clearly insufficient: no finite number of discrete frequencies can densely cover a continuous interval on the frequency space. Thus, while the delay method of Fig. \ref{fig:mode_trajectory_protocol} can be applied to any uniformly-sampled time series data in $n$ dimensions, its utility as a means of sparsely representing the dynamical content of the given signal is limited to the discrete-spectrum case.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{power_spectrum_comparison.pdf}
\caption{Power spectrum of a Van der Pol oscillator system subjected to different types of forcing.}
\label{fig:power_spectrum_comparison}
\end{figure*}
This is a highly restrictive condition, given that many systems of mathematical interest and almost all real-world systems do not exhibit fully discrete Fourier spectra. Even when dynamics are very nearly periodic, the introduction of any measurement noise or stochastic/chaotic forcing will populate the gaps in the spectrum and undermine any attempt to build a linear model for the system in delay coordinates. This is illustrated in Fig. \ref{fig:power_spectrum_comparison}, in which the Van der Pol oscillator from Eq. \ref{eq:vdp_def} is subjected to different types of weak parametric forcing on the $y_2$ variable. Though the magnitude of forcing is sufficiently small that the state-space trajectories (top) look nearly identical, the differences in the power spectra (bottom) are quite obvious. The consequences of this spectral contamination are evident when the plotted time series are subjected to linear model discovery using DMD, which is discussed in the next section. Even when the exogenous forcing is an order of magnitude weaker than the intrinsic dynamics, it yields a model whose reconstruction error is many times greater than that of the unforced system. Because the base periodic oscillation still dominates the observed dynamics, a linear delay-coordinate model should offer a useful approximation to the true behavior. The remainder of this paper is devoted to understanding how such a model can be obtained by isolating the continuous-spectrum forcing from the underlying periodic motion and combining them in the framework of linear control.
\section{Dynamic Mode Decomposition and Time-Delay Embedding}\label{sec:dmd}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{dmd_schematic.png}
\caption{Procedure for building linear models from the delay-coordinate SVD illustrated in Fig. \ref{fig:mode_trajectory_protocol}. DMD performed on the time series $V$ matrix yields a linear model whose stepwise prediction error is interpreted as a learned forcing signal $u^{\rm{disc}}$. The same $V$ matrix can then be fed to the DMD with control (DMDc) algorithm along with $u^{\rm{disc}}$ in order to construct a linear control model of the form shown.}
\label{fig:dmd_schematic}
\end{figure*}
\subsection{Dynamic Mode Decomposition}\label{sec:dmd_background}
Dynamic Mode Decomposition (DMD) is a model regression algorithm which produces a best-fit linear operator to reproduce the dynamics observed in some time series data set \cite{Schmid2010jfm,rowleymezicetal2009,Tu2014jcd,Kutz2016book}. The resulting model offers the benefits of interpretability (via its eigenvalue spectrum and spatial eigenvectors) and future-state forecasting to arbitrary time. Given a set of sequential measurements $\mathbf{X} = \left[\mathbf{x}_1, \mathbf{x}_2, ... \mathbf{x}_m\right] \in \mathbb{R}^{n\times m}$, DMD seeks to solve
\begin{equation}
\begin{split}
\dot{\mathbf{X}}= \mathbf{A}\mathbf{X}
\end{split}
\end{equation}
The simplest implementation of the algorithm, known as {\em exact DMD}, simply seeks a least-squares best fit which minimizes $\|\dot{\mathbf{X}} - \mathbf{A}\mathbf{X}\|_F$, solved by $\mathbf{A} = \dot{\mathbf{X}}\mathbf{X}^\dagger$ ($\dagger$ denotes the Moore-Penrose pseudo-inverse) \cite{turowleyetal2014}. This is sometimes formulated in discrete time, where one seeks an operator which maps the state one time step into the future. The resultant $\mathbf{A}$ defines a first-order linear differential equation in the state space $\mathbb{R}^n$, the closed-form solution to which can be written using the operator's eigendecomposition:
\begin{equation}
\begin{split}
\label{eq:dmd_sol}
\tilde{\mathbf{x}}(t) &= \sum_j b_j\mathbf{w}_j e^{\omega_j t}
\end{split}
\end{equation}
where $\mathbf{w}_j$ and $\omega_j$ are eigenvectors and eigenvalues of the DMD operator and $b_j$ are scalar weighting coefficients. This illustrates the power of linear model regression: the obtained DMD model admits a closed-form analytic solution which can easily be broken down into simple exponential modes and can be evaluated over any time domain. DMD has become a heavily used algorithm in the analysis of spatio-temporal data analysis, with many recent innovations including a sparsity promoting variant~\cite{Jovanovic2014pof}, an optimized framework~\cite{askham2018variable}, extensions to control~\cite{proctorbruntonkutz2016}, a Bayesian formulation~\cite{Takeishi2017JCAI}, a variational formulation~\cite{nuske2014jctc,noe2013variational,azencot2019consistent}, a higher-order formulation (HO-DMD)~\cite{leclainchevega2017a}, a stochastic formulation~\cite{Yeung2017arxiv} and a multi-resolution analysis~\cite{kutz2016multiresolution,dylewsky2019}.
\subsection{Time-Delay DMD}
A resurgence of interest in Koopman theory and consequently in DMD in recent years has led to a variety of extensions of the algorithm focusing on two principal tasks: modifying the optimization objective function to redefine what constitutes a ``best fit'' for ${\bf A}$ and identifying lifting transformations which render the data more amenable to linear representation \cite{williamskevrekidisrowley2015,Giannakis2019,Lusch2018}. The results presented in this work remain agnostic on the former issue and focus on a particular solution to the latter: namely, the use of delay-embedding coordinates, or PCT. The lifting procedure described in Eq.~(\ref{eq:scalar_delay}) has been shown to quite generically improve DMD representations for continuous time-series data \cite{Mezic2005nd,leclainchevega2017a}.
Of particular note is the original Hankel alternative view of Koopman (HAVOK) formulation~\cite{bruntonproctorkaiserkutz2017}, and the subsequent and closely related Hankel DMD~\cite{arbabimezic2017}.
The procedure for constructing a delay-coordinate linear model is diagrammed in the first two rows of Fig. \ref{fig:dmd_schematic}. Starting from the SVD approach illustrated in Fig. \ref{fig:mode_trajectory_protocol}, a linear DMD operator $\mathbf{A}$ is regressed using data from the first $r$ rows of the time-series projection matrix $\mathbf{V}$:
\begin{equation}
\begin{split}
\dot{\mathbf{V}} = \mathbf{A}\mathbf{V}
\end{split}
\end{equation}
The resultant model can reproduce the dynamics in the space of the SVD modes by integrating this differential equation (or using the analytic solution from Eq. \ref{eq:dmd_sol}). It also enables state forecasting simply by integrating further out to some future time. Results can be transformed back into the original delay-coordinate space via matrix multiplication: $\hat{\mathbf{H}} = \mathbf{U}\mathbf{S}\hat{\mathbf{V}}^T$, where $\hat{\mathbf{V}}$ is the DMD reconstruction of the data matrix $\mathbf{V}$.
The Hankel matrix formed by the embedding procedure of Eq.~(\ref{eq:delay_embed}) is guaranteed by construction to have redundant values along matrix diagonals $\{H_{ij}$ where $i+j = \rm(const)\}$. No such property is assured for the DMD reconstruction $\hat{\mathbf{H}}$, so the task of collapsing the resultant signal back down into the original state space is nontrivial. There is an approach used in SSA known as Hankelization, or diagonal averaging, in which a state-space reconstruction $\hat{\mathbf{x}}$ is obtained from $\hat{\mathbf{H}}$ by averaging over the elements which would in a true Hankel matrix be redundant \cite{hassani07}. However, in the case of model forecast residuals discussed in the following section, we find there is often coherent oscillatory behavior along these diagonals which can be obfuscated by averaging. In these cases we therefore simply sample the first $n$ rows of $\hat{\mathbf{H}}$ to compress results back into the state space. We find that the delay registers largely differ from one another only by a global phase factor, so the choice to use the first register is not particularly consequential.
\section{Data-Driven Decomposition of Nonlinear Systems into Forced Linear Models}
DMD generates models with complex eigenvalues, but for real and stationary input data its modes converge to pure-imaginary conjugate pairs. In this case, a rank-$r$ DMD model yields dynamics at at most $r/2$ distinct frequencies. Thus, while Hankel DMD can produce a linear model which reproduces a discrete-spectrum nonlinear oscillator to arbitrary precision, it will unavoidably fall short when it comes to systems with continuous spectra. In this section, we present an extension of Hankel DMD to model nonlinear dynamics as a forced linear system, for which the model and the forcing signal are learned simultaneously. This approach retains the benefits of linearity for the portion of the model which captures the dominant, (quasi-)periodic dynamics, while adding the versatility to simulate a much broader class of nonlinear systems.
A standard delay-coordinate DMD model is first generated as explained in the previous section. Working in the space of PCT modes, this means learning the linear operator ${\bf A}$ which offers a best-fit solution to $\dot{\mathbf{v}}(t) = \mathbf{A}\mathbf{v}(t)$ over the full duration of available data. The resulting model is then used for stepwise forecasting: given some $\mathbf{v}(t_j)$, the succeeding observation $\mathbf{v}(t_{j+1})$ can be approximated as follows:
\begin{equation}
\begin{split}
\hat{\mathbf{v}}(t_{j+1}) &= \mathbf{v}(t_j) + \Delta t \mathbf{A}\mathbf{v}(t_j)
\label{eq:stepwise_forecast}
\end{split}
\end{equation}
(This simple Euler time step could be replaced by any forward numerical integration method). Iterating over the first $(m-1)$ columns of $\mathbf{V}$, this approach can be used to construct a full single-step prediction matrix $\hat{\mathbf{V}}$:
\begin{equation}
\begin{split}
\hat{\mathbf{V}} &= \left[ \begin{matrix}
| & | & | & \\
\hat{\mathbf{v}}(t_1) & \hat{\mathbf{v}}(t_2) & \hat{\mathbf{v}}(t_3) & \cdots\\
| & | & | &
\end{matrix}\right]
\end{split}
\end{equation}
The discrepancy between $\mathbf{V}$ and $\hat{\mathbf{V}}$ can be thought of as a quantitative account of the shortcoming of the DMD representation. We assert that the true dynamics can be written in the form
\begin{equation}
\begin{split}
\dot{\mathbf{v}} &= \mathbf{A}\mathbf{v} + \mathbf{u}(t)
\label{eq:control_model}
\end{split}
\end{equation}
for some parametric exogenous forcing $\mathbf{u}(t)$. In this case a forward Euler step would yield
\begin{equation}
\begin{split}
\mathbf{v}(t_{j+1}) &= \mathbf{v}(t_j) + \Delta t \mathbf{A}\mathbf{v}(t_j) +\Delta t \mathbf{u}(t_j)
\end{split}
\end{equation}
and so
\begin{equation}
\begin{split}
\mathbf{v}(t_{j+1}) - \hat{\mathbf{v}}(t_{j+1}) = \Delta t \mathbf{u}(t_j)
\end{split}
\end{equation}
leading to
\begin{equation}
\begin{split}
\mathbf{u}(t) &= \frac{\mathbf{v}(t+\Delta t) - \hat{\mathbf{v}}(t+\Delta t)}{\Delta t} .
\end{split}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{control_signal_protocol_edit.pdf}
\caption{Extracting stepwise DMD forecast error to construct a forcing time series. Actuated by this signal, the linear DMD model will perfectly reproduce the true nonlinear dynamics over the observation interval. The model error shown below in red has been scaled up for visual clarity.}
\label{fig:control_signal_protocol}
\end{figure}
An inferred forcing signal can thus be extracted simply by computing the stepwise forecast $\hat{\mathbf{V}}$ using Eq.~(\ref{eq:stepwise_forecast}). This process is illustrated schematically in Fig. \ref{fig:control_signal_protocol}. In one sense this process is tautological: the learned $\mathbf{u}(t)$ takes on whatever value is required to match prediction to reality at a given time. The control model posited in Eq.~(\ref{eq:control_model}) is fully accurate by construction. However, the decomposition it achieves is nontrivial. For any system whose dynamics are dominated by near-periodic evolution (or linear combinations thereof), Hankel DMD can generate a model which captures a large fraction of observed behavior. Under these circumstances, and assuming the rank of ${\bf A}$ is sufficient to account for all prominent spectral peaks, the discovered forcing consists solely of the continuous-spectrum mixing dynamics which lie beyond the reach of DMD. In other words, to the extent that Hankel DMD can be considered an optimal means of discovering a set of $r$ observables in whose space dynamics are most nearly linear (known as a Koopman invariant subspace), ${\bf u}(t)$ offers a window into the residual dynamics which unfold in the orthogonal complement to this subspace.
As an example, we consider the forced Van der Pol system defined by:
\begin{equation}
\begin{split}
\dot{y}_1 &= y_2\\
\dot{y}_2 &= \left(1-y_1^2\right)y_2 - y_1 + u(t)
\end{split}
\label{eq:forced_vdp}
\end{equation}
As plotted in Fig. \ref{fig:power_spectrum_comparison}, we consider cases where the forcing $u(t)$ is zero (unforced), where it is sinusoidal, and where it is a chaotic oscillatory signal derived from the $z$ coordinate of a trajectory on the canonical Lorenz attractor. In each case, a Hankel DMD model is generated from data generated over an interval of $1000$ time units sampled at a resolution of $0.01$ time units. The Hankel matrix is constructed using $d = 256$ delay embeddings spaced $q = 4$ time steps apart, for an embedding period of $T^d = qd\Delta t = 10.24$ (which is on the same order as the Van der Pol oscillatory period). The results of the forcing discovery procedure are plotted side by side with the true forcings used to simulate the data in Fig. \ref{fig:discovered_forcings}. While the phases of the corresponding signals clearly do not match, their structures are quite similar and their autocorrelations plotted on the right match very closely.
The DMD models used to generate Fig. \ref{fig:discovered_forcings} were selected from models with ranks varying from $2$ to $48$. The results plotted (models of $r=12$ and $r=6$, respectively) were chosen as particularly clean representations of the desired results. Some of the models of other ranks produced forcing signals which were somewhat obfuscated by additional high-frequency oscillation. This is because the ability of this method to separate periodic and aperiodic contributions to observed dynamics is limited by the capacity of DMD to fully capture the periodic behavior without any additional, spurious oscillatory modes. The strength of a DMD model in this regard can be evaluated by comparison between its eigenvalue spectrum and the Fourier spectrum of the input data: there should be eigenfrequencies corresponding to all prominent Fourier peaks no more than that. In Fig. \ref{fig:power_spectrum_comparison} it is apparent that the sinusoidal forcing introduced additional peaks at half-integer harmonics of the base frequency, so it is not surprising that a higher-rank DMD model was required to isolate a clean forcing signal relative to the chaotically forced system.
It should be noted that in the case of sinusoidal forcing, the separation that has occurred is not between periodic and aperiodic dynamics: the forcing signal itself is, of course, periodic. But ${\bf u}(t)$ oscillates at a frequency orders of magnitude lower than those which dominate the Fourier power spectrum, and DMD is well known to be ill-equipped to capture dynamics at such disparate time scales \cite{dylewsky2019}. The obtained linear model therefore ignores that slow oscillatory behavior and fully relegates it to the exogenous forcing.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{discovered_forcings_v2.pdf}
\caption{Left: discovered forcing signals (red, lower) plotted alongside the true signals (blue, upper) applied to $y_2$ in the simulations. Right: autocorrelation profiles illustrate their similarity up to a global phase factor.}
\label{fig:discovered_forcings}
\end{figure*}
\section{Integration with DMDc: Discovering Linear Control Systems Without Prior Knowledge of the Forcing Signal}
\label{sec:dmdc}
DMD with control (DMDc) extends traditional DMD to fit a model of the form~\cite{proctorbruntonkutz2016}:
\begin{equation}
\begin{split}
\dot{\mathbf{x}} &= \mathbf{A}\mathbf{x} + \mathbf{B}\mathbf{u}(t).
\end{split}
\end{equation}
The algorithm offers a means of simultaneous regression of the intrinsic linear dynamics $\mathbf{A}$ and the control mapping $\mathbf{B}$, but requires a known control signal $\mathbf{u}(t)$ in addition to the state data $\mathbf{x}(t)$. In many circumstances this information is not available, either because a known forcing variable cannot be measured or because the exogenous influence on the system is carried by many environmental variables which cannot even be enumerated, much less measured.
The control architecture can be extended to a Koopman theoretic framework as well~\cite{KordaMezic2018,Peitz2017arxiv,Kaiser2017arxiv,proctor2018generalizing}.
Incorporating our method for unsupervised forcing discovery into the DMDc approach allows for the construction of linear control models from time series measurements of the state alone. This is accomplished by first running the standard DMD algorithm and collecting the stepwise forecast errors as explained in the previous section, and then performing a second decomposition this time using the DMDc formalism to obtain a modified $\mathbf{A}$ approximating the intrinsic linear dynamics and a control mapping $\mathbf{B}$. The latter is not strictly necessary; the method by which the control signal is generated implicitly assumes equally-weighted forcing on all state variables, so working in the space of the truncated SVD on delay coordinates it can be assumed that $\mathbf{B}$ is the identity ${\bf I}_r$. There exists a variation on DMDc which assumes $\mathbf{B}$ is known, but even the general algorithm reliably discovers this to a good approximation, so there is little practical difference.
Figure~\ref{fig:DMDc_comparison} shows results from the application of this method to the chaotically forced Van der Pol system used in the previous section. Two DMDc models are trained: one using the ground truth forcing used in the initial simulation, and one using the discovered forcing obtained from the delay DMD method. In order to apply the original forcing to the model in rank-reduced SVD delay coordinates, it is transformed $\mathbf{u}(t) \in \mathbb{R}^2 \rightarrow \tilde{\mathbf{u}}(t) \in \mathbb{R}^r$ by time-delay embedding and projecting the result onto the top $r$ PCT modes.
Note that while the DMDc models are trained on the first half of the plotted time series and tested on the remainder, these results should not be interpreted as true forecasts. The models are supplied with forcing data for the full duration, derived either from the original simulation (true forcing) or from the first pass of DMD without control (discovered forcing). This effectively grants the models some foreknowledge of impending dynamics which would not be available for future-state prediction tasks. Rather, the plots of Fig. \ref{fig:DMDc_comparison} are intended to exhibit the results of the DMDc method applied to a system for which the dominant contribution to the stepwise forecast residual is a known external actuation. The plots illustrate the method’s ability to reconstruct observed dynamics equally well with or without ground truth knowledge of the underlying actuation.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{DMDc_comparison_v2.pdf}
\caption{$u(t)$ is the control signal (in state space) used to generate the full nonlinear signal (black, underlaid). $\tilde{u}$ is the same control signal lifted into delay embedding space and projected onto the top $r$ SVD modes of the state Hankel matrix ${\bf H}$. $\tilde{u}_{\rm{disc}}$ is the discovered forcing obtained by the procedure diagrammed in Fig. \ref{fig:control_signal_protocol}.}
\label{fig:DMDc_comparison}
\end{figure*}
\section{Application: Discovering forcing on real-world data with inherent periodicity}
\label{sec:power_grid}
As an example of the practical diagnostic utility of this method, we apply it to power grid demand data published as part of the RE-Europe data set \cite{jensen15}. Electrical load is sampled hourly over a period from 2012 through 2014. Data is presented for many geolocated nodes of a reduced network representation of the European grid, which we average over to obtain a single time series for each country surveyed.
This subject was chosen as an example of a real-world data series which contains strong periodic signatures: daily, weekly, and yearly cycles are all quite obviously present, and indeed dominate much of the observed behavior. But the dynamics are of course not entirely periodic; electricity demand is mediated by collective human behaviors, which are in turn influenced by weather patterns, economic trends, and uncountably many other variables. Casting these dynamics as a linear control system in time-delay coordinates effectively separates these factors, with predictable periodicities represented in eigenvalues of the DMD $\mathbf{A}$ matrix and everything else contained in the discovered control signal $\tilde{\mathbf{u}}(t)$. There is no ground truth against which to validate results, as the ``true'' forcing is not a measurable quantity, but we can offer some qualitative observations which corroborate the idea that $\tilde{\mathbf{u}}(t)$ should encode anomalies in human behaviors as they pertain to electricity consumption.
Results are plotted in Fig. \ref{fig:power_grid_plot} for the load data from Germany and Italy. As discussed previously, DMD (like most numerical methods) cannot easily accommodate models with dynamics on highly disparate timescales, such as daily and yearly oscillations. To circumvent this issue we run decompositions on a sliding window of approximately 2 weeks over the full sampling period. This is a sufficiently narrow domain to render seasonal variations negligible over the course of any given sub-series. Forcing signals are then inferred individually for each windowed iteration and then averaged to form a global $\tilde{\mathbf{u}}(t)$ for the full two-year span.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{power_grid_plot_v2.pdf}
\caption{Discovered forcing signal (in MW) for mean grid demand per network node in Germany (top) and Italy (bottom). The most obvious feature of the forcing signal is a large annual spike around Christmas in both countries, and an additional annual spike in late summer in Italy. Many businesses in Italy close in August. For reference, the weeks around Christmas (Dec. 11 - Jan. 8) and the month of August (Aug. 1 - Sept. 1) are shaded.}
\label{fig:power_grid_plot}
\end{figure*}
The most salient features in the results are large annual spikes in which forcing takes on an anomalously high value followed immediately by an anomalous low. Both nations exhibit this pattern once per year in December, and Italy additionally evinces a similar phenomenon (albeit slightly horizontally stretched) in the late summer of every year. This is highly consistent with known behavioral signatures in these countries: both celebrate Christmas with school holidays, business closures, etc., and August is widely observed as a national vacation month in Italy. That these breaks in human routine lead to by far the most prominent spikes in the learned forcing signals (while their influence is overshadowed by standard daily oscillations in the original data) suggests even in the absence of ground truth that this method has to some nontrivial extent successfully disambiguated the quasilinear oscillations which dominate the load data from anomalous activity resulting from more complex environmental variables. This could have far-reaching implications for forecasting of time series such as these, in that it separates the more easily predicted, relatively deterministic dynamics from aperiodic factors which may be better modeled stochastically.
The success of this method in recovering the $u(t)$ signal of the Van der Pol system in Eq. \ref{eq:forced_vdp} was owed in part to the contrived simplicity of this example. For more generic input data, the discovered forcing is not solely determined by true exogeneous actuation of the underlying system. The learned signal also incorporates contributions from any endogenous chaotic (continuous-spectrum) dynamics which cannot be reproduced by the linear Koopman model, as well as any sampling errors that might originate from an imperfect DMD model regression or measurement noise. The Van der Pol example from Fig. \ref{fig:discovered_forcings} minimized these confounding effects because it had a discrete spectrum and the DMD model was regressed on ample, noise-free data. Applying the same procedure to the more complex grid load data demands a slightly more careful interpretation of the results. In a case like this, the difference between external forcing and internal chaotic nonlinearity is not particularly clear. The underlying dynamics governing demand for electricity are complex to the point that there is no obvious distinction between what would qualify as an internal state variable versus an external environmental variable. As a result, it is in many cases philosophically appropriate that the method proposed in this paper makes no attempt to disambiguate between the two. The additional contributions of sampling error and measurement noise also bear acknowledgment, but with enough data to capture all major periodicities of the data the former is likely to be negligible, and the latter would simply pass through and produce a noise floor in the learned forcing (provided that measurement errors are uncorrelated). The spikes in the discovered forcing in the electrical grid load data should therefore be interpreted as affirmative indications of strong disturbances to otherwise nearly-linear dynamics, with the caveat that these disturbances may be endogeneous or exogeneous in origin. It is possible that some further distinction could be made using statistical methods designed to chaotic dynamics from stochastic processes, as in ~\cite{rosso2007} or ~\cite{ravetti2014}, but that is beyond the scope of this paper.
\section{Conclusions}
Since the seminal contributions of Takens, time-delay embeddings of dynamical systems have long been known to contain critical and representative information about the underlying dynamical system measured. In recent years, the advent of high-quality time-series measurements from spatio-temporal systems have afforded the community unprecedented opportunities for constructing data-driven models from such measurement data alone. Not only can time-delay embeddings extract meaningful information from unmeasured latent variables, but such embeddings can be used as a data-driven coordinate system approximating the Koopman operator. The work here advocates the use of these time-delay embeddings for constructing {\em principal component trajectories} (PCT) which provide a time-delay coordinate system which can be used to reconstruct dynamical trajectories via superposition. Indeed, the PCT provide an ideal representation of many dynamical systems and their time-series, especially in systems where an unmeasured latent space is critical to the dynamics. The method is shown to be intimately connected to SSA where various theoretical guarantees are available for infinite time-delay embeddings.
PCT also provide a coordinate system that can be used in conjunction with the DMD algorithm for producing good Koopman operator approximations. It can even be augmented to include external forcing terms in order to model the effects of a continuous spectrum. Such forcing terms are determined and constructed in a completely unsupervised fashion, allowing for a data-driven discovery process that can produce Koopman models which approximate both discrete and continuous spectral dynamics. We offer a demonstration of this method applied to real-world power grid load data to show its utility for diagnostics and interpretation on systems in which somewhat periodic behavior is strongly actuated by unknown and unmeasurable environmental variables. This example illustrates how a completely data-driven method and PCT coordinates can be used in practice.
It will be interesting to apply this approach to analyze other complex systems, such as for fluid flow control~\cite{Brunton2015amr}, which would benefit from the disambiguation of the linear dynamics and external forcing. Extending these approaches to more complex systems is an area of future research.
\begin{acknowledgments}
We acknowledge support from the UW Engineering Data Science Institute, NSF HDR award 1934292.
SLB further acknowledges support from the Army Research Office (ARO W911NF-17-1-0306).
JNK further acknowledges support from the Air Force Office of Scientific Research (AFOSR) grant FA9550-17-1-0329.
\end{acknowledgments}
| 8d997672ec473b54288240baab64c9d6ff7cf9f5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
An interesting direction of combinatorics in recent years is the study of multicolored version of classical extremal results, whose origin can be traced back to a question of Erd\H{o}s and Rothschild~\cite{E} in 1974. They asked which $n$-vertex graph admits the maximum number of $2$-edge-colorings without monochromatic triangles, and conjectured that the complete balanced bipartite graph is the optimal graph. About twenty years later, Yuster~\cite{yuster1996number} confirmed this conjecture for sufficiently large $n$.
\subsection{Erd\H{o}s-Rothschild problems in various settings}
There are many natural generalizations of the Erd\H{o}s-Rothschild problem. The most obvious one may be to ask it for graphs other than the triangles, and one may also increase the number of colors used. A graph G on $n$ vertices is called \textit{$(r, F)$-extremal} if it admits the maximum number of $r$-edge-colorings without any monochromatic copies of $F$ among all $n$-vertex graphs. Alon, Balogh, Keevash and Sudakov~\cite{ABKS} greatly extended Yuster's result and showed that the Tur\'{a}n graph $T_{k}(n)$ is the unique $(r, K_{k+1})$-extremal graph for $k\geq 2$ and $r\in\{2, 3\}$. Interestingly, they also showed that Tur\'{a}n graphs $T_{k}(n)$ are no longer optimal for $r\geq 4$. Indeed, Pikhurko, and Yilma~\cite{pikhurko2012maximum} later proved that $T_{4}(n)$ is the unique $(4, K_3)$-extremal graph, while $T_{9}(n)$ is the unique $(4, K_4)$-extremal graph. Determining the extremal configurations in general for $k\geq 2$ and $r\geq 4$ turned out to be a difficult problem. For further results along this line of research (when $F$ is a non-complete graph or a hypergraph), we refer to~\cite{hoppen2012edge, hoppen2014edge, hoppen2015edge, lefmann2013exact, lefmann2009colourings, lefmann2010structural}.
Another variant of this problem is to study edge-colorings of a graph avoiding a copy of $F$ with a prescribed color pattern. For an $r$-colored graph $\hat{F}$, a graph $G$ on $n$ vertices is called $(r, \hat{F})$-extremal if it admits the maximum number of $r$-colorings which contain no subgraph whose color pattern is isomorphic to $\hat{F}$.
This line of work was initiated by Balogh~\cite{balogh2006remark}, who showed that the Tur\'{a}n graph $T_{k}(n)$ once again yields the maximum number of 2-colorings avoiding $H_{k+1}$, where $H_{k+1}$ is any 2-coloring of $K_{k+1}$ that uses both colors. For $r\geq 3$, the behavior of $(r, H_{k+1})$-extremal graphs was studied by Benevides, Hoppen, Sampaio, Lefmann, and Odermann, see~\cite{BHS, hoppen2015color, hoppen2015rainbow, hoppen2017graphs, hoppen2017rainbow}. In particular, the case when $\hat{F}=\hat{K}_3$ is a triangle with rainbow pattern has recently received a lot of attention (for its relation to Gallai colorings).
Hoppen, Lefmann and Odermann~\cite{hoppen2017graphs} first proved that the Tur\'{a}n graph $T_2(n)$ is the unique $(r, \hat{K}_3)$-extremal graph for $r\geq 5$. Very recently, Balogh and Li~\cite{balogh2018typical}, confirming conjectures of~\cite{BHS} and~\cite{hoppen2017graphs}, showed that the complete graph $K_{n}$ is the unique $(3, \hat{K}_3)$-extremal graph, while the Tur\'{a}n graph $T_2(n)$ becomes optimal as $r\geq 4$.
The Erd\H{o}s-Rothschild problem can also be extended to other discrete structures. In the domain of extremal set theory, Hoppen, Kohayakawa and Lefmann~\cite{hoppen2012hypergraphs} solved the Erd\H{o}s-Rothschild extension of the famous Erd\H{o}s-Ko-Rado Theorem. They, for instance, showed that the optimal $\ell$-intersecting families (each set is of size $k$) yields the maximum number of $r$-colorings in which every color class is $\ell$-intersecting for $r\in\{2, 3\}$, and also provided a fairly complete characterization of the corresponding extremal family for $r\geq 4$. Hoppen, Lefmann and Odermann~\cite{hoppen2016coloring}, and Clemens, Das and Tran~\cite{clemens2018colourings} later studied the Erd\H{o}s-Rothschild extension of the Erd\H{o}s-Ko-Rado Theorem for vector spaces. Moving the problem to the context of power set lattice, recently, Das, Glebov, Sudakov and Tran~\cite{das2019colouring} investigated the the Erd\H{o}s-Rothschild extension of the Sperner's Theorem, and proved that the largest antichain yields the maximum number of $r$-colorings, in which each color class is an antichain, for $r\in\{2, 3\}$. As for many of the previous results, they demonstrated that as $r$ grows, the largest antichain is no longer optimal. They also determined that the extremal configurations for 2-colorings without monochromatic $k$-chains are the largest $k$-chain-free family. The extremal configurations for $r\geq 3$ and $k\geq 2$ are widely unknown.
\subsection{Erd\H{o}s-Rothschild problems for sum-free sets}
Given integers $n\geq m \geq 1$, write $[m,n]:=\{m,\ldots,n\}$ and $[n]:=\{1,...,n\}$.
\begin{defi}[Schur triple \& Sum-free set]
A \textit{Schur triple} or a \textit{sum} in an abelian group $G$ (or in $[n]$) is a triple $\{a, b, c\}$ with $a+b=c$. A set $A\subseteq G$ (or $A\subseteq [n]$) is \textit{sum-free} if $A$ contains no such triple.
\end{defi}
Given a set A of numbers, an \textit{r-coloring} of $A$ is a mapping $f: A \rightarrow [r]$, which assigns one color to each element of $A$. An $r$-coloring of A is called a \textit{sum-free $r$-coloring} if each of the color classes is a sum-free set. Sum-free colorings are among the classical objects studied in extremal combinatorics and can be traced back to Schur's theorem, one of the seminal results in Ramsey theory.
The Erd\H{o}s-Rothschild extension for sum-free sets has been pursued by Liu, Sharifzadeh and Staden~\cite{liu2017maximum} for subsets of the integers, and H\`{a}n and Jim{\'e}nez~\cite{han2018maximum} for finite abelian groups. More specifically, they investigated the extremal configurations which maximize the number of sum-free $r$-colorings.
In the setting of integers, it is well known that the largest sum-free set in $[n]$ has size $\lceil n/2\rceil$. Liu, Sharifzadeh and Staden~\cite{liu2017maximum} determined the extremal configurations for $r=2$.
\begin{thm}\cite{liu2017maximum}
There exists $n_0 > 0$ such that for all integers $n\geq n_0$ , the number of sum-free $2$-colorings of a subset $A\subseteq [n]$ is at most $2^{\lceil n/2\rceil}$. Moreover, the extremal subsets are $\{1, 3, 5, \cdots, 2\lceil n/2\rceil-1\}$, and $[\lfloor n/2\rfloor+1, n]$; and if $n$ is even, we additionally have $[n/2, n-1]$, and $[n/2, n]$.
\end{thm}
Unlike the graph case, in the sum-free setting, there are extremal configurations which are not sum-free even for 2 colors. Therefore, one would expect a more sophisticated extremal behavior as $r$ grows. Although some asymptotic bounds were obtained in~\cite{liu2017maximum}, the characterization of extremal sets for $r\geq 3$ remains widely open.
Such problem was also studied for finite abelian groups. Let $G$ denote a finite abelian group. Over fifty years ago, Diananda and Yap~\cite{diananda1969maximal} determined the maximum density $\mu(G)$ of a sum-free set in $G$ whenever $|G|$ has a prime factor $q \not\equiv 1 \mod 3$, but it was not until 2005 that Green and Ruzsa~\cite{green2005sum} completely solved this extremal question for all finite abelian group. H\`{a}n and Jim{\'e}nez~\cite{han2018maximum} investigated the Erd\H{o}s-Rothschild extension for sum-free sets on some special abelian groups.
\begin{thm}\cite{han2018maximum}
Let $r\in \{2,3\}$, $q\in N$ and let $G$ be a abelian group of sufficiently large order, which has a prime divisor $q$ such that $q\equiv 2 \mod 3$. Then the number of sum-free $r$-coloring of a set $A\subseteq G$ is at most $r^{\mu(G)}$. Moreover, the maximum is only achieved by the largest sum-free set.
\end{thm}
For more than three colors this phenomenon does not persist in general and the problem becomes considerably more complicated. For more details, we refer the readers to~\cite{han2018maximum}. For other abelian groups, despite some asymptotic bounds presented in~\cite{han2018maximum}, the exact extremal phenomena is unknown even for 2 colors.
\subsection{Our results}
In this paper, we consider a rainbow variant of the Erd\H{o}s-Rothschild problem for sum-free sets in $[n]$.
A Schur triple or a sum $\{x,y,z\}$ is a \textit{rainbow sum} if $x,y,z$ are colored with different colors. Note that a rainbow sum must have three distinct elements. For convenience, sometimes we would use the following definitions, which are slightly different with the classical notations on sum-free sets.
\begin{defi}[Restricted Schur triple \& Restricted sum-free set]
A \textit{restricted Schur triple} or a \textit{restricted sum} in $[n]$ is an ordered triple $(a, b, c)$ with $a<b<c$ and $a+b=c$. A set $A\subseteq [n]$ is \textit{restricted sum-free} if $A$ contains no such triple.
\end{defi}
For any integer $n\geq7$, it is not hard to show that the largest restricted sum-free sets in $[n]$ have size $\lfloor n/2 \rfloor+1$. If $n$ is even, then the only subset attaining this bound is $\left[\frac{n}{2}, n\right]$; if $n$ is odd, then the maximum restricted sum-free sets are attained by the following four sets: $\left\{\frac{n-1}{2}, \frac{n-1}{2}+1, \ldots, n-1\right\}$, $\left\{\frac{n-1}{2}, \frac{n-1}{2}+2, \ldots, n\right\}$, $\left[\frac{n+1}{2}, n\right]$, and $\{1, 3, 5, \ldots, n\}$.
Given a set of positive integers $A \subseteq [n]$, an $r$-coloring of $A$ is \textit{rainbow sum-free} if it contains no rainbow sum.
For a positive integer $r$ and a set $A\subseteq [n]$, we write $g(A, r)$ for the number of rainbow sum-free $r$-colorings of $A$ and define
\[
g(n, r):=\max_{A\subseteq [n]}g(A,r).
\]
A set $A\subseteq [n]$ is \textit{rainbow $r$-extremal} if $g(A, r)=g(n, r)$.
When $r\in\{1,2\}$, it is trivial to see that $g(n, r)=r^n$ for all positive integers $n$, and the only extremal set is the interval $[n]$, since for every subset $A\subseteq [n]$, all $r$-colorings of $A$ are rainbow sum-free. For $r \geq 3$, the characterization of the extremal sets requires substantially more work.
Our first main result is an upper bound on the number of rainbow sum-free $r$-colorings of dense sets
\begin{thm}\label{thm: hrange}
For every integer $r\geq 3$, there exists $n_0$ such that for all $n>n_0$ the following holds. For a set $A\subseteq [n]$ with $|A|\geq (1-{r}^{-3})n$, the number of rainbow sum-free $r$-colorings $g(A,r)$ satisfies
$$g(A, r)\leq \binom{r}{2}\cdot 2^{|A|}+2^{-\frac{n}{26\log n}}2^n.$$
\end{thm}
By choosing two of the $r$ colors and coloring the elements of $[n]$ arbitrarily with these two colors, one can easily obtain that
\begin{equation}\label{eq: int}
g([n], r)\geq \binom{r}{2}(2^n - 2) + r =\binom{r}{2}2^n - (r^2 - 2r).
\end{equation}
Therefore, Theorem~\ref{thm: hrange} is asymptotically sharp for $A=[n]$ and then the typical structure of rainbow sum-free $r$-colorings of $[n]$ immediately follows from~(\ref{eq: int}).
\begin{cor}\label{cor1}
For every integer $r\geq 3$, almost all rainbow sum-free $r$-colorings of $[n]$ are 2-colorings.
\end{cor}
Now we turn to the extremal configurations of rainbow sum-free $r$-colorings. Let us first consider the case $r=3$. Similarly as in the Gallai coloring problem, two natural candidates of the extremal sets are the maximum restricted sum-free sets and the interval $[n]$.
Note that for every restrict sum-free set $A$, we have $g(A, 3)\le 3^{\lfloor n/2 \rfloor+1} \ll g([n], 3)$. Our second theorem shows that for three colors the interval $[n]$ is indeed optimal.
\begin{thm}\label{thm: three}
There exists $n_0$ such that for all $n > n_0$, among all subsets of $[n]$, the interval $[n]$ is the unique rainbow 3-extremal set.
\end{thm}
Just as for the Erd\H{o}s-Rothschild extension for Gallai colorings~\cite{balogh2018typical}, we may not expect that the same phenomena persists for $r\geq 4$.
Define $O:=\{1, 3, 5, \cdots, 2\lceil n/2\rceil-1\}$, and $I_0:=[\lfloor n/2\rfloor+1, n]$.
We prove the following stability theorem.
\begin{thm}\label{thm:sta}
For every positive integer $r\geq 4$, we have
\[
g(n, r)= r^{n/2 + o(n)}.
\]
Moreover, for every $\varepsilon >0$, there exist $\delta, n_0>0$ such that for all integers $n\geq n_0$ the following holds.
Let $A$ be a subset of $[n]$ with $g(A, r)\geq r^{n/2 - \delta n}.$ Then
\smallskip
\begin{compactenum}
\item[\rm (i)] for $r\geq 5$, we have that either $|A \bigtriangleup O|\leq \varepsilon n$, or $|A \bigtriangleup I_0|\leq \varepsilon n$;\smallskip
\item[\rm (ii)] for $r=4$, we have that either $|A \bigtriangleup [n]|\leq \varepsilon n$, or $|A \bigtriangleup O|\leq \varepsilon n$, or $|A\bigtriangleup I_0|\leq \varepsilon n$.
\end{compactenum}
\end{thm}
The behavior of the exact extremal configurations not only depends on the number of colors, but also depend on the parity of $n$.
For even $n$, we define
\[
I_1=\left[\frac{n}{2}-1, n\right], \qquad
I_2=\left[\frac{n}{2}, n\right].
\]
Observe that $I_1$ contains exactly two restricted Schur triples $(n/2-1, n/2, n-1)$, $(n/2-1, n/2+ 1, n)$, and it is not hard to compute that $g(I_1, r)=r^{n/2}\left(3 - 2/r\right)^2$. On the other hand, the set $I_2$ is a restricted sum-free set and therefore $g(I_2, r)=r^{|I_1|}=r^{n/2 + 1}$.
For odd $n$, we define
\[
I_3=\left[\frac{n-1}{2}, n\right].
\]
Again, the set $I_3$ contains exactly one restricted Schur triple $(\frac{n-1}{2}, \frac{n-1}{2}+1, n)$, and one can show that $g(I_3, r)=r^{\lceil n/2\rceil}\left(3 - 2/r\right)$, which is already greater than the number of colorings for any restricted sum-free set.
When a set $A$ is of size at least the size of the maximum restricted sum-free sets and not one of the above three sets, we believe that the restrictions from the triples would more than counteract the extra possibilities offered by the additional vertices. Therefore, we make the following conjecture.
\begin{conj}\label{conj}
Let $n, r$ be positive integers and $r\geq 4$.\smallskip
\begin{compactenum}[\rm (i)]
\item If $n$ is even and $r\leq 7$, then $g(n, r)=r^{n/2}\left(3 - 2/r\right)^2$, and $I_1$ is the unique rainbow $r$-extremal set.\smallskip
\item If $n$ is even and $r\geq 8$, then $g(n, r)=r^{n/2 + 1}$, and $I_2$ is the unique rainbow $r$-extremal set.\smallskip
\item If $n$ is odd and $r=4$, then $g(n, r)=g([n], r)$, and $[n]$ is the unique rainbow $r$-extremal set.\smallskip
\item If $n$ is odd and $r\geq 5$, then $g(n, r)=r^{\lceil n/2\rceil}\left(3 - 2/r\right)$, and $I_3$ is the unique rainbow $r$-extremal set.
\end{compactenum}
\end{conj}
Our forth main result verifies Conjecture~\ref{conj} for $r\geq 8$ and $n$ sufficiently large.
\begin{thm}\label{thm: r8}
For an integer $r\geq 8$, there exists $n_0=n_0(r)$ such that for all $n > n_0$ the following holds. Let $A$ be a subset of $[n]$ with $|A|\geq\lceil n/2 \rceil +1$. \smallskip
\begin{compactenum}[\rm (i)]
\item If $n$ is even, then $g(A, r)\leq r^{\lceil n/2 \rceil +1}$, and the equality holds if and only if $A=I_2$.\smallskip
\item If $n$ is odd, then $g(A, r)\leq r^{\lceil n/2\rceil}\left(3 - 2/r\right)$, and the equality holds if and only if $A=I_3$.
\end{compactenum}
\end{thm}
The paper is organized as follows. In the next section, we list some structural results on sum-free sets, which are essential for the proof, and introduce the multi-color container theorem. In Section~3, we prove Theorem~\ref{thm: hrange}. In Section 4, we prove the stability theorem, Theorem~\ref{thm:sta}, and determine $g(n,3)$ for $n$ sufficiently large. In Section 5, we determine $g(n,r)$ for $r\geq8$, and describe the corresponding extremal configurations.
We close the paper with some concluding remarks in Section 6.
Throughout the paper, all logarithms have base 2.
\section{Notation and preliminaries}
\subsection{Basic properties of restricted sum-free sets}
We use the following result of Staden~\cite{staden2017note} on the minimum number of additive triples
among all sets of a given size.
\begin{thm}\cite{staden2017note}\label{thm:staden}
Let $A$ be a subset of $[n]$ with $|A|> \lceil n/2 \rceil$. Then the number of Schur triples in $A$ is at least
\[
(|A| - \lceil n/2 \rceil)(|A| - \lfloor n/2 \rfloor),
\]
where the unique minimising set
is $[n - |A| + 1, n]$.
\end{thm}
For a set $A \subseteq [n]$, we write $\mathcal{S}(A)$ for the set of all restricted Schur triples in $A$, and let $s(A)=|\mathcal{S}(A)|$. For an integer $t\in A$, denote by $\mathcal{S}(t, A)$ the set of all triples in $\mathcal{S}(A)$ containing $t$, and let $s(t, A)=|\mathcal{S}(t, A)|$. Then from Theorem~\ref{thm:staden}, we immediately obtain the following proposition.
\begin{prop}\label{numberofsum}
Let $A$ be a subset of $[n]$ with $|A|> \lceil n/2 \rceil$. Then
\[
s(A)\geq (|A| - \lceil n/2 \rceil)(|A| - \lfloor n/2 \rfloor) - |A|/2.
\]
In particular, we have
\[
s([n])= \begin{cases}
\frac{n^2-2n}{4}&\text{if } n \text{ is even};\\
\frac{n^2-2n+1}{4}&\text{otherwise}.
\end{cases}
\]
\end{prop}
\subsection{Structural properties of sum-free sets}
We will use standard definitions and notation in additive combinatorics as given in \cite{additive}. Given $A,B\subseteq \mathbb{Z}$, let
\[
A+B:=\{a+b\mid a\in A,b\in B\},\quad\text{ and }\quad A-B:=\{a-b\mid a\in A,b\in B\}.
\]
When $B=\{x\}$, we simply write $A+x$ and $A-x$.
The following lemma is known as Green's removal lemma, which was first proved by Green \cite{green2005re-lemma}, and was later generalized to non-abelian groups by Kr\'al and Vena \cite{KSV2009}.
\begin{lemma}\cite{KSV2009, green2005re-lemma}\label{Stabilitylem2}
For all $\varepsilon>0$, there exists $\delta, n_0>0$ such that the following holds for all integers $n\geq n_0$. Suppose that $A \subseteq [n]$ is a set containing at most $\delta n^2$ Schur triples. Then there exist $B, C\subseteq [n]$ such that $A=B\cup C$ where $B$ is sum-free and $|C|\leq \varepsilon n$.
\end{lemma}
We also require a very strong stability theorem for sum-free sets proved by Deshouillers, Freiman, and Odermann \cite{Deshouillers1999sum-free}.
\begin{lemma}\cite{Deshouillers1999sum-free}\label{Stabilitylem3}
Every sum-free set $S$ in $[n]$ satisfies at least one of the following conditions:\smallskip
\begin{compactenum}[\rm (i)]
\item $|S|\leq 2n/5$;\smallskip
\item $S$ consists of odd numbers;\smallskip
\item $|S| \leq \min(S)$.
\end{compactenum}
\end{lemma}
\subsection{Multi-color container theorem}
An important tool in our proof is the hypergraph container theorem.
We use the following version from~\cite{balogh2017container}. Let $\mathcal{H}$ be a $k$-uniform hypergraph with average degree \textit{d}. The \textit{co-degree} of a set of vertices $X\subseteq V(\mathcal{H})$ is the number of edges containing $X$; that is,$$d(X)=\{e\in E(\mathcal{H})\ |\ X\subseteq e\}.$$
For every integer $2 \leq j\leq k$, the $j$-th maximum co-degree of $\mathcal{H}$ is
$$\Delta_j(\mathcal{H})=\max\{d(X)\ |\ X\subseteq V(\mathcal{H}),\ |X|=j\}.$$
When the underlying hypergraph is clear, we simply write it as $\Delta_j$. For $0<\tau<1$, the \textit{co-degree function} $\Delta(\mathcal{H},\tau)$ is defined as
$$\Delta(\mathcal{H},\tau)=2^{\binom{k}{2}-1}\sum_{j=2}^k2^{-\binom{j-1}{2}}\frac {\Delta_j}{d\tau^{j-1}}.$$
In particular, when $k=3$,
$$\Delta(\mathcal{H},\tau)=\frac {4\Delta_2}{d\tau}+\frac {2\Delta_3}{d\tau^{2}}.$$
\begin{thm}\cite{balogh2017container}\label{baloghcontainer}
Let $\mathcal{H}$ be a $k$-uniform hypergraph on vertex set $[N]$. Let $0<\varepsilon, \tau< 1/2$.
Suppose that $\tau<1/(200k!^2k)$ and $\Delta(\mathcal{H},\tau)\leq \varepsilon/(12k!)$. Then there exists $c=c(k)\leq 1000k!^3k$ and a collection of vertex subsets $\mathcal{C}$ such that\smallskip
\begin{compactenum}[\rm (i)]
\item every independent set in $\mathcal{H}$ is a subset of some of $A\in \mathcal{C}$;\smallskip
\item for every $A\in \mathcal{C}$, $e(\mathcal{H}[A])\leq \varepsilon\cdot e(\mathcal{H})$;\smallskip
\item $\log|\mathcal{C}|\leq cN\tau \log(1/\varepsilon)\log(1/\tau)$.
\end{compactenum}
\end{thm}
A key concept in applying container theory to such coloring problems is the notion of \textit{template}, which was first introduced in \cite{FOSU}, although the concept had already appeared in~\cite{ST} under the name of `2-colored multigraphs' and later in~\cite{balogh2016further}, simply referred as `containers'.
\begin{defi}[Template\ and\ palette]
An \textit{$r$-template} of order $n$ is a function $P: [n] \to 2^{[r]}$, associating to each element $x \in [n] $ a list of colors $P(x) \subseteq [r]$. We refer to this set $P(x)$ as the \textit{palette} available at $x$.
For a set $A \subseteq [n]$, any $r$-coloring of $A$ can be considered as an $r$-template of order $n$, with only one color allowed at each element in $A$, and no color allowed for elements not belonging to $A$.
\end{defi}
\begin{defi}[Subtemplate]
Let $P_1$, $P_2$ be two $r$-templates of order $n$. We say that $P_1$ is a \textit{subtemplate} of $P_2$ (written as $P_1 \subseteq P_2$) if $P_1(x) \subseteq P_2(x)$ for each element $x \in [n]$.
\end{defi}
For an $r$-template $P$ of order $n$, write $RS(P)$ for the number of subtemplate of $P$ that are rainbow restricted sums. We say that $P$ is a \textit{rainbow restricted sum-free} $r$-template if $RS(P)=0$.
\smallskip
Using Theorem \ref{baloghcontainer}, we obtain the following.
\begin{thm} \label{container}
For every integer $r\geq 3$, there exists a constant $c=c(r)$ and a collection $\mathcal{C}$ of $r$-templates of order $n$ such that\smallskip
\begin{compactenum}[\rm (i)]
\item every rainbow restricted sum-free $r$-template of order $n$ is a subtemplate of some $P\in \mathcal{C}$;\smallskip
\item for every $P\in \mathcal{C}$, $RS(P)\leq n^{-1/3}s([n])$;\smallskip
\item $|\mathcal{C}|\leq 2^{cn^{2/3} \log^2n}$.
\end{compactenum}
\end{thm}
\begin{proof}
Let $\mathcal{H}$ be a $3$-uniform hypergraph with vertex set $[n]\times \{1,2,...,r\}$, whose edges are all triples $\{(x_1, c_1), (x_2, c_2), (x_3, c_3)\}$ such that $(x_1, x_2, x_3)$ forms a restricted Schur triple in $[n]$ and $c_1, c_2, c_3$ are all different. In other words, every hyperedge in $\mathcal{H}$ corresponds to a rainbow restricted Schur triple. Note that there are exactly $r(r-1)(r-2)$ ways to rainbow color a restricted Schur triple with $r$ colors. Hence, the average degree $d$ of $\mathcal{H}$ is equal to
$$d=\frac{3e(\mathcal{H})}{v(\mathcal{H})}
=\frac{3r(r-1)(r-2)s([n])}{nr}
\geq \frac{3(r-1)(r-2)n}{8}.$$
Now we apply Theorem \ref{baloghcontainer} on $\mathcal{H}$. Let $\varepsilon=n^{-1/3}/r(r-1)(r-2)$ ands
$\tau=\sqrt{96\cdot 3! \cdot r}n^{-\frac{1}{3}}$.
Observe that $\Delta_2(\mathcal{H})=2(r-2), \Delta_3(\mathcal{H})=1.$
For $n$ sufficiently large, we can get $\tau<1/(200\cdot3!^2\cdot3)$ and $$\Delta(\mathcal{H},\tau)=\frac{4\Delta_2}{d\tau}+\frac{2\Delta_3}{d\tau^2}=\frac{8(r-2)}{d\tau}+\frac{2}{d\tau^2}\leq \frac{3}{d\tau^2}\leq \frac{\varepsilon}{12\cdot 3!}.$$
Hence, there is a collection of vertex subsets $\mathcal{C}$ satisfying properties (i)-(iii) of Theorem \ref{baloghcontainer}. Observe
that every vertex set of $\mathcal{H}$ corresponds to an $r$-template of order $n$; every rainbow restricted sum-free $r$-template of order $n$ corresponds to an independent set in $\mathcal{H}$. Therefore, $\mathcal{C}$ is a desired collection of $r$-templates.
\end{proof}
\begin{defi}[Good $r$-template]
For $A\subseteq [n]$, an $r$-template $P$ of order $n$ is a \textit{good $r$-template} of $A$ if it satisfies the following properties:\smallskip
\begin{compactenum}[(i)]
\item For each element $i \in A$, $|P(i)|\geq 1$;\smallskip
\item $RS(P)\leq n^{-1/3}s([n])$.
\end{compactenum}
\end{defi}
For a set $A\subseteq [n]$ and a collection of templates $\mathcal{P}$, denote by $G(\mathcal{P},A)$ the set of rainbow sum-free $r$-colorings of $A$, which is a subtemplate of some $P\in \mathcal{P}$. Let $g(\mathcal{P},A)=|G(\mathcal{P},A)|$.
If $\mathcal{P}$ consists of a single $r$-template $P$, then we simply write $G(P,A)$ and $g(P,A)$.
\section{Proof of Theorem \ref{thm: hrange}}
Throughout this section, we fix an integer $r\geq 3$, a sufficiently large integer $n$ and an arbitrary set $A\subseteq [n]$ with $|A|=(1-\xi)n$, where
\[
0\leq \xi \leq r^{-3}.
\]
Let $\mathcal{C}$ be the collection of containers given by Theorem \ref{container}, and $\delta =1/(24\log n)$.
We divide $\mathcal{C}$ into two classes
\begin{equation}\label{def:coll}
\mathcal{C}_1=\{P\in \mathcal{C} : g(P,A)\leq 2^{(1-\delta)n}\},
\quad
\mathcal{C}_2=\{P\in \mathcal{C} : g(P,A)> 2^{(1-\delta)n}\}.
\end{equation}
Note that every $P\in\mathcal{C}_2$ is a good $r$-template of $A$.
The crucial part of the proof is to estimate $g(\mathcal{C}_2,A)$, which replies on the following four lemmas.
\begin{lemma}\label{(x,y)sumfree}
Let $F$ be the collection of ordered pairs $(a,b)\in A^2$ with $a<b$ such that $\{a,b\} \nsubseteq S$ for all $S\in \mathcal{S}(A)$. Then we have $|F|\leq \xi n^2+n/6$.
\end{lemma}
\begin{proof}
Let
\begin{gather*}
F_1= \{(a,b)\in A^2\mid a+b\in [n]\backslash A, \ b=2a\},
\quad
F_2= \{(a,b)\in A^2\mid a+b>n, \ b=2a\},\\
F_3= \{(a,b)\in A^2\mid a+b\in [n]\backslash A, \ b-a\in [n]\backslash A\},
\quad
F_4= \{(a,b)\in A^2\mid a+b>n, \ b-a\in [n]\backslash A\}.
\end{gather*}
Clearly, $|F|=\sum_{i=1}^{4}|F_i|$ and $|F_1|\leq |[n]\backslash A|=\xi n$.
Since every $(a,b)\in F_2$ satisfies $b=2a\leq n$ and $a+b=3a>n$, we have $|F_2|\leq n/6$. Moreover, we obtain that $|F_3|\leq \binom{\xi n}{2}\leq {{\xi}^2n^2/2}$, as $a+b\in[n]\backslash A$ and $b-a\in[n]\backslash A$.
Similarly, we have $|F_4|\leq \xi n^2/2$, as $b>n/2$ and $b-a\in[n]\backslash A$.
Finally, we conclude that $|F|\leq \xi n+n/6+{\xi}^2n^2/2+\xi n^2/2
\leq \xi n^2+n/6$.
\end{proof}
For a template $P$ of $A$, let
\[
X_1=\{x\in A\mid |P(x)|=1\},\quad X_2=\{x\in A \mid |P(x)|=2\},\quad X_3=\{x\in A \mid |P(x)|\geq3\},
\] and $x_i=|X_i|$ for $i\in[3]$.
\begin{lemma} \label{x_3}
Suppose that $P$ is a template of $A$ in $\mathcal{C}_2$. Then we have \[
\max\left\{\frac {(\xi- \delta)n+x_1}{\log{r}-1},\ 0\right\}<
x_3 \leq 2n^{-1/3}n.
\]
In particular, if $\xi\geq 2(\log r -1)n^{-1/3} + \delta$, then $\mathcal{C}_2$ is empty.
\end{lemma}
\begin{proof}
By the definitions of $G(P,A)$ and $\mathcal{C}_2$, we have
\begin{equation}\label{ineq:courainc}
2^{x_2}r^{x_3} \geq g(P,A)>2^{(1-\delta)n}.
\end{equation}
Since $x_2=|A|-x_1-x_3$ and $|A|=(1-\xi)n$, we obtain that
\begin{equation}\label{eq:x3lb}
x_3>\frac {(\xi- \delta)n+x_1}{\log{r}-1}.
\end{equation}
We first claim that $x_2\geq (1-\xi)n/3$. Otherwise, we immediately have $x_1+x_3 > 2(1-\xi)n/3$.
Together with~(\ref{eq:x3lb}), we obtain that
$x_3 > \frac{(2+\xi-3\delta)n}{3\log r} > \frac{n}{2\log r}$.
By Lemma \ref{(x,y)sumfree} and $0\leq \xi\leq r^{-3}$, there are at least
\[\binom{x_3}{2}-({\xi}n^2 + n/6)\geq \frac{n^2}{16\log^2 r}\]
pairs in $X_3$, which are contained in some restricted Schur triples in $A$.
This contradicts the definition of good $r$-templates, as
$RS(P)\geq n^2/(3\cdot 16\log^2 r)> n^{-1/3}s([n])$.
For each $a\in X_3$, let $B_a:=\{b\in X_2\mid \{a,b\}\subseteq S, \text{ for some} \ S\in \mathcal{S}(A)\}$.
Note that every $b\in X_2 \setminus B_a$ satisfies either $|b -a|\in [n]\setminus A$ or $|b - a| = \min\{a, b\}$. Then we have
$|B_a|\geq (1-\xi)n/3-2\xi n-1>n/4$.
Since $P$ is a good $r$-template of $A$, we obtain that
$$n^{-1/3}s([n])
\geq RS(P)
\geq \frac12\sum_{a\in X_3}|B_a|\geq \frac{x_3\cdot n}{8},$$
which indicates $x_3\leq 2n^{-1/3}n.$
\end{proof}
Next, we will prove a stability result on good templates with many rainbow sum-free colorings.
\begin{lemma} \label{(i,j)palette number}
Let $0\leq \xi< 2(\log r -1)n^{-1/3} + \delta $. Then for every $P\in \mathcal{C}_2$, there exist two colors $\{i,j\}\in [r]$ such that the number of elements in $A$ with palette $\{i,j\}$ is at least $(1-2\delta)n$.
\end{lemma}
\begin{proof}
By Lemma~\ref{x_3} and~(\ref{ineq:courainc}), we have \[x_2\geq(1-\delta-3\log r \cdot n^{-1/3})n.\]
For $1\leq i<j\leq r$, define $Y_{i,j}:=\{x\in X_2 \mid P(x)=\{i,j\}\}$.
Without loss of generality, we can assume that $|Y_{1,2}|\geq x_2/\binom{r}{2}$.
Let $Y'=X_2\setminus Y_{1,2}$.
For each $a\in Y'$, let $B_a=\{b\in Y_{1,2}\mid \{a,b\}\subseteq S, \text{ for some } S\in \mathcal{S}(A)\}$. Similarly as in Lemma~\ref{x_3}, we obtain that $|B_a|\geq x_2/\binom{r}{2}-2\xi n-1$, and then
$$n^{-1/3}s([n])\geq RS(P)\geq \frac{1}{2}\sum_{x\in Y'}|B_a|\geq \frac{|Y'|\cdot n}{2r(r-1)}
.$$
Since $\delta \gg n^{-1/3}$, we have $|Y_{1,2}|=x_2-|Y'|\geq (1-2\delta)n$, which completes the proof.
\end{proof}
\begin{lemma} \label{P_{i,j}}
For two colors $i,j \in [r]$, denote by $\mathcal{P}=\mathcal{P}(i, j)$ the set of good $r$-template of $A$, in which there are at least $(1-2\delta)n$ elements in $A$ with palette $\{i,j\}$. Then
$$g(\mathcal{P}, A)\leq 2^{|A|}(1 + 2^{-n/12}).$$
\end{lemma}
\begin{proof}
For an $r$-coloring $g\in G(\mathcal{P}, A)$, let $S(g)$ the set of elements in $A$, which are not colored by $i$ or $j$.
By the definition of $\mathcal{P}$, we have $|S(g)|\leq 2\delta n$.
Define
\[\mathcal{G}_0=\{g\in G(\mathcal{P}, A) \mid S(g)=\varnothing\}, \quad\text{and}\quad\mathcal{G}_1=\{g\in G(\mathcal{P}, A) \mid |S(g)|\geq 1\}.\] Clearly, we have $g(\mathcal{P}, A)=|\mathcal{G}_0| + |\mathcal{G}_1|$ and $|\mathcal{G}_0|\leq 2^{|A|}$. It remains to show that $|\mathcal{G}_1|\leq 2^{|A| - n/12}.$
Let us consider the ways to color $A$ so that the resulting colorings are in $\mathcal{G}_1$. We first choose a set $A_0\subseteq A$ of size at most $2\delta n$, which will be colored by the colors in $[r]\setminus \{i, j\}$. The number of options is at most $\sum_{1\leq k\leq 2\delta n}\binom{n}{k}$, and the number of colorings is at most $r^{2\delta n}$.
Once we fix $A_0$ and its color, take an arbitrary vertex $t\in A_0$.
\begin{claim}\label{claim:dispair}
Let $\mathcal{D}(t)$ be the collection of disjoint pairs $\{a, b\}$ in $A\setminus A_0$ such that $\{a, b, t\}$ forms a restricted Schur triple.
Then $|\mathcal{D}(t)|\geq n/6$.
\end{claim}
\begin{proof}
Define
\[
\mathcal{S}_1= \{(a,b)\in [n]^2\mid a + b=t,\ a<b\},\ and\
\mathcal{S}_2= \{(a, b)\in [n]^2\mid t+a=b,\ a<b\}.
\]
We first observe that $|\mathcal{S}_1|= \lfloor(t-1)/2\rfloor$ for every $t\in [n]$. Note that all pairs in $\mathcal{S}_1$ are disjoint. Therefore, if $t> 2n/5$, we have
$|\mathcal{D}(t)|\geq |\mathcal{S}_1| - \xi n - |A_0|\geq |\mathcal{S}_1| - (2\delta + \xi) n\geq n/6$.
If $t\leq 2n/5$, observe that $|\mathcal{S}_2|= n -2t$ and all pairs in $\mathcal{S}_2$ are disjoint. Therefore, we obtain that $|\mathcal{D}(t)|\geq |\mathcal{S}_2| - \xi n - |A_0|\geq |\mathcal{S}_2| - (2\delta + \xi) n\geq n/6$.
\end{proof}
For every pair $(a, b)\in \mathcal{D}(t)$, since $t$ is colored by some color in $[r]\setminus\{i, j\}$, and $a, b$ can only be colored by $i$ or $j$, the elements $a$ and $b$ must receive the same color in order to avoid the rainbow Schur triple. Therefore, together with Claim~\ref{claim:dispair}, the number of ways to finish the colorings is at most
\[
2^{|A| - |A_0| - |\mathcal{D}(t)|}\leq 2^{|A| - n/6}.
\]
Hence, we obtain that
\[
|\mathcal{G}_2|\leq
\sum_{1\leq k\leq 2\delta n}\binom{n}{k}r^{2\delta n}2^{|A| - n/6}
\leq 2^{|A| + 2\delta n(\log n + \log r) - n/6}\leq 2^{|A| - n/12},
\]
where the last inequality follows from $\delta=1/(24\log n)$.
\end{proof}
Now we have all ingredients to prove Theorem \ref{thm: hrange}.
\begin{proof}[{\bf Proof of Theorem \ref{thm: hrange}}]
First, by property (i) of Theorem~\ref{container} , every rainbow sum-free $r$-coloring of $A$ is a subtemplate of some
$P \in \mathcal{C}$.
By Property (iii) of Theorem \ref{container} and the definition of $\mathcal{C}_1$ (see~(\ref{def:coll})), we have
$$g(\mathcal{C}_1, A) \leq |\mathcal{C}_1|\cdot 2^{(1-\delta)n} \leq | \mathcal{C}|\cdot 2^{(1-\delta)n} <2^n\cdot 2^{-n/(25\log n)}.$$
If $\xi\geq 2(\log r -1)n^{-1/3} + \delta$, using Lemma~\ref{x_3}, we are done by $g(\mathcal{C}, A)=g(\mathcal{C}_1, A)$.
Otherwise, by Lemma \ref{(i,j)palette number} and Lemma \ref{P_{i,j}}, we obtain that
$$g(\mathcal{C}_{2}, A)= \sum_{1\leq i<j\leq r}g(\mathcal{P}(i,j), A)\leq \binom{r}{2}2^{|A|}(1+2^{-n/12}).$$
Hence, we have
$$g(\mathcal{C}, A)=g(\mathcal{C}_{1}, A)+g(\mathcal{C}_{2}, A)\leq 2^n\cdot 2^{-\frac{n}{25\log n}}+\binom{r}{2}2^{|A|}(1+2^{-n/12})\leq \binom{r}{2}\cdot 2^{|A|}+2^{-\frac{n}{26\log n}}2^n,$$
which gives the desired upper bound on the number of rainbow sum-free $r$-colorings of $A$.
\end{proof}
\iffals
\section{Proof of Theorem~\ref{thm: three}}
In this section, let $r=3$ and $\mathcal{C}$ be the collection of $3$-templates obtained from Theorem~\ref{container}.
\begin{lemma}\label{le: 3sparse}
For any $A\subseteq [n]$ with $|A|=(1 - \xi)n$ and $3^{-3}<\xi\leq 1$, we have
$g(A,r)\leq 2^n$.
\end{lemma}
\begin{proof}
First, we can assume $\xi< {\frac 25}$, otherwise it is trivial as $g(A, 3)\leq 3^{|A|}\leq 2^{0.955n}$.
Let $\mathcal{C}$ be the collection of containers given by Theorem~\ref{container}, and $g_{\max}(P,A)=\max_{P\in \mathcal{C}}g(P, A)$.
For a template $P\in\mathcal{C}$, which is not good of $A$, there must be an element $i\in A$ with $|P(i)|=0$, which immediately gives $g(P, A)=0$. Therefore, $g_{\max}(P,A)$ is always achieved by a good template.
Let $P$ be a \emph{good} template of $A$.
Since $RS(P)=o(n^2)$, by Green's removal lemma,
there is a set $E\subseteq [n]$ and a template $P':[n]\setminus E\to 2^{[r]}$, such that $P\mid_{[n]\setminus E}=P'$, $|E|=o(n)$, and $P'$ has no rainbow Schur triples.
Define
\[
X_1=\{a\in [n]\setminus E\mid |P'(a)|=1\},
\quad X_2=\{a\in [n]\setminus E\mid |P'(a)|=2\},
\]and
\[
X_3=\{a\in [n]\setminus E\mid |P'(a)|\geq 3\}.
\]
Let $T=X_2\cup X_3$ and $x_i=|X_i|$ for $i\in [3]$. Therefore, we have
\begin{equation}\label{eq:TT}
(X_3+X_3)\cap X_3=\varnothing,\quad (T+T)\cap X_3=\varnothing,\quad (X_3+T)\cap T=\varnothing.
\end{equation}
We first assume that $x_2\leq(1-\frac52 \xi)n$. Let $m$ be the largest element in $X_3$. By (\ref{eq:TT}), for every $i< m$, at least one of $\{i,m-i\}$ is not in $T$. Hence, we have
\begin{equation}
|T|\leq n-\Big\lceil\frac{m-1}{2}\Big\rceil,\quad x_3\leq m-\Big\lceil\frac{m-1}{2}\Big\rceil.
\end{equation}
Therefore,
\begin{align}
\log g(\mathcal{C},A)&\leq
\log (|\mathcal{C}|\cdot g_{\max}(P,A))
\leq cn^{\frac{2}{3}}\log^2n+|E|\log 3+x_2+x_3\log 3\nonumber\\
&=o(n)+x_2+x_3\log
= \left(o(n)+\frac{1}{2}\left(|T|+x_3-\left(1-\frac{2}{\log 3}\right)x_2\right)\right)\log 3\nonumber\\
&\leq \left(\frac{n}{2}+o(n)+\frac{1}{2}\left(\frac{2}{\log 3}-1\right)\frac{2-5\xi}{2}n\right)\log 3< n,
\end{align}
since $\xi>3^{-3}$, and we take $o(n)<0.01n$.
Finally, we may assume that $x_2> (1-\frac52 \xi)n$. Then $x_3\leq |A|-x_2<\frac32 \xi n$. Thus we obtain
\[
\log g(\mathcal{C},A)\leq o(n)+x_2+x_3\log 3
\leq o(n)+|A|+(\log 3-1)x_3\leq n-(2-\log 3)\xi n+o(n)
<n,
\]
where we take $o(n)<0.01n$.
\end{proof}
\begin{proof}[{\bf Proof of Theorem~\ref{thm: three}}]
Since $g([n], 3)\geq 3\cdot 2^n -3$, it is equivalent to prove that for all sets $A\subsetneq [n]$, we have $g(A, 3)\leq (1.5 + o(1))2^n$, which directly follows from Theorem~\ref{thm: hrange} and Lemma~\ref{le: 3sparse}.
\end{proof}
\newpage\fi
\section{Proof of Theorems~\ref{thm: three} and \ref{thm:sta}}
The following lemma gives us a structural description of large sum-free sets.
\begin{lemma}\label{lem:str}
Let $\varepsilon,c>0$, $c>10\varepsilon$, and $\varepsilon<1/10$. Let $A,B\subseteq[n]$ such that $A\cap B=\varnothing$, $B$ is sum-free, and $|A|=cn$. If $|B|\geq(1/2-\varepsilon) n$, then $(A+B)\cap B\neq \varnothing.$
\end{lemma}
\begin{proof}
Suppose, to the contrary, that $(A+B)\cap B=\varnothing$. Since $|B|>2n/5$, by Lemma~\ref{Stabilitylem3}, either $B$ only contains odd numbers, or the minimum element of $B$ is at least $|B|\geq(1/2-\varepsilon) n$. If $B$ only contains odd numbers, then there is $d\geq c-\varepsilon$, such that $|A\cap E|=dn$, where $E\subseteq [n]$ is the collection of all even numbers.
Thus, there exists an $a\in A\cap E$ such that $dn\leq a\leq (1-d)n$.
Let $P$ be the collection of all pairs $(i,\ i+a)$, where $i$ is odd, and $1\leq i\leq dn$. Observe that all pairs in $P$ are pairwise disjoint and there are at least $dn/2$ of them.
Since $(A+B)\cap B=\varnothing$, for each pair $(i,j)$ in $P$, at least one of $\{i,j\}$ is not in $B$. This implies
\[
|B|\leq \frac n2- |P|\leq \frac n2 - \frac {dn}{2}
\leq \frac n2 - \frac{(c - \varepsilon)n}{2}\leq
\left(\frac{1}{2}-2\varepsilon\right)n,\] which contradicts the assumption of $B$.
If the minimum element of $B$ is at least $|B|\geq(1/2-\varepsilon) n$, let $b$ be the smallest element in $B$, then there is $d\geq c-2\varepsilon$ such that $|A\cap[b-1]|=dn$. This implies that there exists $a\in A$ with $dn/2\leq a\leq b-dn/2$. We define $P$ to be the collection of all pairs $(i,j)$, where $b\leq i< (1/2+3\varepsilon)n$ and $j=i+a$. Then the number of pairs in $P$ is at least $2\varepsilon n$, as $b\leq (1/2+\varepsilon) n$. Moreover, for every $(i,j)$ in $P$, we have $j<n$ since $b-dn/2\leq (1/2-3\varepsilon)n$ and since $b+a>(1/2+3\varepsilon)n$, $P$ is a set of disjoint pairs. Since $(A+B)\cap B=\varnothing$, for every $(i,j)$ in $P$, at least one of $\{i,j\}$ is not in $B$. Similarly we obtain that $|B|\leq (1/2-2\varepsilon)n$, a contradiction.
\end{proof}
Our next lemma says that when the number of colorings $r=3$ and the size of $A$ is significantly smaller than $n$, the number of rainbow sum-free $r$-colorings will be much less than $2^n$. And when $r\geq 5$ and the size of $A$ is significantly larger than $n/2$, much less than $r^{n/2}$ when $r$, the number of rainbow sum-free $r$-colorings will be much less than $r^{n/2}$.
\begin{lemma}\label{lem:mid}
Let $\varepsilon>0$, $r$ be a positive integer, and let $A$ be a subset of $[n]$. Then the followings hold.\smallskip
\begin{compactenum}[\rm (i)]
\item If $r=3$, and $|A|\leq (1-\varepsilon)n$, then here is a constant $\delta_1=\delta_1(\varepsilon)>0$, such that $$g(A,3)\leq 2^{(1-\delta_1)n}.$$
\item If $r\geq5$, and $|A|\geq \big(1/2+\varepsilon\big)n$,then there is a constant $\delta_1=\delta_1(\varepsilon,r)>0$ such that $$g(A,r)\leq r^{(1/2-\delta_1)n}.$$
\end{compactenum}
\end{lemma}
\begin{proof}
Let $\mathcal{C}$ be the collection of containers given by Theorem~\ref{container}, and let $$g_{\max}(P,A)=\max_{P\in \mathcal{C}}g(P, A).$$
For a template $P\in\mathcal{C}$, suppose $P$ is not a good template. Then there must be an element $i\in A$ with $|P(i)|=0$, which immediately gives $g(P, A)=0$. Therefore, $g_{\max}(P,A)$ is always achieved by a good template.
Let $P$ be a good template of $A$.
Since $RS(P)=o(n^2)$, by Green's arithmetic removal lemma,
there is a set $E\subseteq [n]$ and a template $P':[n]\setminus E\to 2^{[r]}$, such that $P\mid_{[n]\setminus E}=P'$, $|E|=o(n)$, and $P'$ has no rainbow Schur triples.
Define
\[
X_1=\{a\in [n]\setminus E: |P'(a)|=1\},
\quad X_2=\{a\in [n]\setminus E: |P'(a)|=2\},
\]and
\[
X_3=\{a\in [n]\setminus E: |P'(a)|\geq 3\}.
\]
Let $T=X_2\cup X_3$ and let $x_i=|X_i|$ for $i=1,2,3$. Therefore, we have
\begin{equation}\label{eq:T}
(X_3+X_3)\cap X_3=\varnothing,\quad (T+T)\cap X_3=\varnothing,\quad (X_3+T)\cap T=\varnothing.
\end{equation}
As $X_3$ is sum-free, we have $x_3\leq \lfloor n/2 \rfloor+1$.
Let $m$ be the largest element in $X_3$. By (\ref{eq:T}), for every $i< m$, at least one of $\{i,\ m-i\}$ is not in $T$ which is also the same for $X_3$. Hence, we have
\begin{equation}\label{eq:P3}
|T|\leq n-\Big\lceil\frac{m-1}{2}\Big\rceil,\quad x_3\leq m-\Big\lceil\frac{m-1}{2}\Big\rceil.
\end{equation}
\noindent\textbf{Case 1: $r=3$.}
Observe that we may assume $\varepsilon< 2/5$, as otherwise we will get $g(A, 3)\leq 3^{|A|}\leq 2^{0.955n}$, which completes the proof with $\delta_1=0.045$. We first consider the case when $x_2\leq(1-5\varepsilon/2)n$.
Then we have
\begin{align*}
\log g(\mathcal{C},A)&\leq
\log (|\mathcal{C}|\cdot g_{\max}(P,A))
\leq cn^{2/3}\log^2n+|E|\log 3+x_2+x_3\log 3\nonumber\\
&=o(n)+x_2+x_3\log
= \left(o(n)+\frac{1}{2}\left(|T|+x_3-\left(1-\frac{2}{\log 3}\right)x_2\right)\right)\log 3\nonumber\\
&\leq n+o(n)-\frac{5\varepsilon}{2}\left(1-\frac{\log3}{2}\right)n< (1-\delta_1)n,
\end{align*}
where we take $\delta_1=\frac{5\varepsilon}{4}(1-\frac{\log3}{2})$.
Now, we may assume that $x_2> (1-5\varepsilon/2)n$. Then $x_3\leq |A|-x_2<3\varepsilon n/2$. Thus we obtain
\begin{align*}
\log g(\mathcal{C},A)&\leq o(n)+x_2+x_3\log 3
\leq o(n)+|A|+(\log 3-1)x_3\\
&\leq n+o(n)-\frac{\varepsilon}{2}\left(5-3\log3\right)n
<(1-\delta_1)n,
\end{align*}
and we take $\delta_1=\frac{1}{4}\left(5-3\log3\right)\varepsilon$.\smallskip
\newline
\noindent\textbf{Case 2: $r\geq5$.}
Since $|A|\geq (1/2+\varepsilon)n$, and $P$ is good, we have that $x_1+x_2\geq |A|-x_3-|E|\geq \varepsilon n/2$ for large enough $n$. We first assume that $x_2\geq\frac{\varepsilon n}{100}$.
Similarly, we get
\begin{align}
\log g(\mathcal{C},A)&\leq
\log (|\mathcal{C}|\cdot g_{\max}(P,A))
\leq cn^{2/3}\log^2n+|E|\log r+x_2+x_3\log r\nonumber\\
&=o(n)+x_2+x_3\log
= \left(o(n)+\frac{1}{2}\left(|T|+x_3-\left(1-\frac{2}{\log r}\right)x_2\right)\right)\log r\nonumber\\
&\leq \left(\frac{n}{2}+o(n)-\frac{1}{2}\left(1-\frac{2}{\log r}\right)\frac{\varepsilon}{100}n\right)\log r< \left(\frac{n}{2}-\delta_1 n\right)\log r,\label{eq:mid}
\end{align}
where we take $\delta_1=\frac{1}{300}\left(1-\frac{2}{\log r}\right)\varepsilon$. Note that $\delta_1>0$ as $r\geq5$.
Finally, we may assume that $x_2\leq \frac{\varepsilon n}{100}$, and then $x_1\geq \frac{\varepsilon n}{2} - x_2\geq \frac{\varepsilon n}{3}$.
We claim that $x_3\leq(1/2-\varepsilon/40)n$.
Otherwise, by the way we construct $P'$, we also have
$(X_1+X_3)\cap X_3=\varnothing$
and this contradicts Lemma~\ref{lem:str}.
Similarly as before, we can conclude that
\[
\log g(\mathcal{C},A)\leq o(n)+x_2+x_3\log r
\leq \left(\frac{n}{2}+o(n)+\frac{\varepsilon}{100\log r}n-\frac{\varepsilon}{80}n\right)\log r\leq \left(\frac{n}{2}-\delta_1 n\right)\log r,
\]
where we take $\delta_1=\frac{\varepsilon}{400}$.
\end{proof}
The case when $r=4$ is more involved, and we will discuss it later in this section. But the result in Lemma~\ref{lem:mid} (i) is enough to imply Theorem~\ref{thm: three}.
\begin{proof}[{\bf Proof of Theorem~\ref{thm: three}}]
Observe that $g([n], 3)\geq 3\cdot 2^n -3$. Suppose $A\subseteq[n]$ and $A\neq[n]$. When $|A|\leq (1-3^{-3})n$, by Lemma~\ref{lem:mid} (i), there is $\delta_1>0$ such that $g(A,3)\leq 2^{(1-\delta_1)n}<g([n],3).$ Now, we have $(1-3^{-3})n<|A|\leq n-1$. By Theorem~\ref{thm: hrange}, $g(A,3)\leq (1.5+o(1))2^n<g([n],3)$.
\end{proof}
\iffals
\newpage
\section{Proof of Theorem~\ref{thm:sta}}
In this section, we prove the stability theorem. For this, we will need five lemmas. The first lemma gives us a structural description of large sum-free sets.
\begin{lemma}\label{lem:str}
Let $\varepsilon,c>0$, $c>10\varepsilon$, and $\varepsilon<\frac{1}{10}$. Let $A,B\subseteq[n]$ such that $A\cap B=\varnothing$, $B$ is sum-free, and $|A|=cn$. If $|B|\geq(1/2-\varepsilon) n$, then $(A+B)\cap B\neq \varnothing.$
\end{lemma}
\begin{proof}
We assume that $(A+B)\cap B=\varnothing$. Since $|B|>2n/5$, by Lemma~\ref{Stabilitylem3}, either $B$ only contains odd numbers, or the minimum element of $B$ is at least $|B|\geq(1/2-\varepsilon) n$. For the first case, there is $d\geq c-\varepsilon$, such that $|A\cap E|=dn$, where $E\subseteq [n]$ is the collection of all even numbers. Thus, there exists an $a\in A\cap E$ such that $dn/2\leq a\leq (1-d/2)n$. Let $P$ be the collection of all pairs $(i,j)$, where both $i$ and $j$ are odd, $1\leq i\leq a-1$ and $j-i=a$.
Observe the all pairs in $P$ are disjoint and there are at least $a/2$ of them. For each pair $(i,j)$ in $P$, at least one of $\{i,j\}$ is not in $B$ since $(A+B)\cap B=\varnothing$. This implies
\[
|B|\leq \frac n2- |P|\leq \frac n2 - \frac a2
\leq \frac n2 - \frac{(c - \varepsilon)n}{4}\leq
\left(\frac{1}{2}-2\varepsilon\right)n,\] which contradicts the assumption of $B$.
Let $b$ be the smallest element in $B$, and now we may assume that $b\geq (1/2-\varepsilon) n$. Then there is $d\geq c-2\varepsilon$ such that $|A\cap[b-1]|=dn$. This implies that there exists $a\in A$ with $dn/2\leq a\leq y-dn/2$. We define $P$ to be the collection of all pairs $(i,j)$, where $b\leq i< (1/2+3\varepsilon)n$ and $j=i+a$. Then the number of pairs in $P$ is at least $2\varepsilon n$, as $b\leq (1/2+\varepsilon) n$. Moreover, for every $(i,j)$ in $P$, we have $j<n$ since $b-\frac{dn}{2}\leq (\frac{1}{2}-3\varepsilon)n$. Since $(A+B)\cap B=\varnothing$, for every $(i,j)$ in $P$, at least one of $\{i,j\}$ is not in $B$. Similarly we obtain that $|B|\leq (\frac{1}{2}-2\varepsilon)n$, a contradiction.
\end{proof}
Our next lemma says that if the size of $A$ is significantly larger than $n/2$, then the number of rainbow sum-free $r$-colorings will be much less than $r^{n/2}$ when $r$ is at least $5$.
\begin{lemma}\label{lem:mid}
For a constant $\varepsilon>0$ and an integer $r\geq5$, let $A$ be a subset of $[n]$ with $|A|\geq \big(1/2+\varepsilon\big)n$. Then there is a constant $\delta_1=\delta_1(\varepsilon,r)>0$ such that
\[
g(A,r)\leq r^{(1/2-\delta_1)n}.
\]
\end{lemma}
\begin{proof}
Let $\mathcal{C}$ be the collection of containers given by Theorem~\ref{container}, and $g_{\max}(P,A)=\max_{P\in \mathcal{C}}g(P, A)$.
For a template $P\in\mathcal{C}$, which is not good of $A$, there must be an element $i\in A$ with $|P(i)|=0$, which immediately gives $g(P, A)=0$. Therefore, $g_{\max}(P,A)$ is always achieved by a good template.
Let $P$ be a \emph{good} template of $A$.
Since $RS(P)=o(n^2)$, by Green's removal lemma,
there is a set $E\subseteq [n]$ and a template $P':[n]\setminus E\to 2^{[r]}$, such that $P\mid_{[n]\setminus E}=P'$, $|E|=o(n)$, and $P'$ has no rainbow Schur triples.
Define
\[
X_1=\{a\in [n]\setminus E\mid |P'(a)|=1\},
\quad X_2=\{a\in [n]\setminus E\mid |P'(a)|=2\},
\]and
\[
X_3=\{a\in [n]\setminus E\mid |P'(a)|\geq 3\}.
\]
Let $T=X_2\cup X_3$ and $x_i=|X_i|$ for $i\in [3]$. Therefore, we have
\begin{equation}\label{eq:T}
(X_3+X_3)\cap X_3=\varnothing,\quad (T+T)\cap X_3=\varnothing,\quad (X_3+T)\cap T=\varnothing.
\end{equation}
As $X_3$ is sum-free, we have $x_3\leq \lfloor n/2 \rfloor+1$. Since $|A|\geq (1/2+\varepsilon)n$, and $P$ is good, we have that
$x_1+x_2\geq |A|-x_3-|E|\geq \varepsilon n/2$ for large $n$.
We first assume that $x_2\geq\frac{\varepsilon n}{100}$. Let $m$ be the largest element in $X_3$. By (\ref{eq:T}), for every $i< m$, at least one of $\{i,m-i\}$ is not in $T$. Hence, we have
\begin{equation}\label{eq:P3}
|T|\leq n-\Big\lceil\frac{m-1}{2}\Big\rceil,\quad x_3\leq m-\Big\lceil\frac{m-1}{2}\Big\rceil.
\end{equation}
Therefore,
\begin{align}
\log g(\mathcal{C},A)&\leq
\log (|\mathcal{C}|\cdot g_{\max}(P,A))
\leq cn^{\frac{2}{3}}\log^2n+|E|\log r+x_2+x_3\log r\nonumber\\
&=o(n)+x_2+x_3\log
= \left(o(n)+\frac{1}{2}\left(|T|+x_3-\left(1-\frac{2}{\log r}\right)x_2\right)\right)\log r\nonumber\\
&\leq \left(\frac{n}{2}+o(n)-\frac{1}{2}\left(1-\frac{2}{\log r}\right)\frac{\varepsilon}{100}n\right)\log r< \left(\frac{n}{2}-\delta_1 n\right)\log r,\label{eq:mid}
\end{align}
where we take $\delta_1=\frac{1}{300}\left(1-\frac{2}{\log r}\right)\varepsilon$. Note that $\delta_1>0$ as $r\geq5$.
Finally, we may assume that $x_2\leq \frac{\varepsilon n}{100}$, and then $x_1\geq \frac{\varepsilon n}{2} - x_2\geq \frac{\varepsilon n}{3}$.
We claim that $x_3\leq(\frac12-\frac{\varepsilon}{40})n$.
Otherwise, by the way we construct $P'$, we also have
$(X_1+X_3)\cap X_3=\varnothing$
and this contradicts Lemma~\ref{lem:str}.
Similarly as before, we can conclude that
\[
\log g(\mathcal{C},A)\leq o(n)+x_2+x_3\log r
\leq \left(\frac{n}{2}+o(n)+\frac{\varepsilon}{100\log r}n-\frac{\varepsilon}{80}n\right)\log r\leq \left(\frac{n}{2}-\delta_1 n\right)\log r,
\]
where we take $\delta_1=\frac{\varepsilon}{400}$.
\end{proof}
\f
The next lemma records an easy fact about intervals for convenience
in the proof of the analogue result of Lemma~\ref{lem:mid} when we have only 4 colors.
\begin{lemma}\label{lem:inter}
Let $\varepsilon>0$, and let $a,b$ be integers such that $0<a<b<n$, $3\varepsilon n<a<n/2-2\varepsilon n$, and $a+2+3\varepsilon n<b\leq 2a$. Suppose $A\subseteq [a+1,b]$, $B\subseteq [b+1,n]$, and $|A|>b-a-\varepsilon n$, $|B|>n-b-\varepsilon n$. Then $(A+A)\cap B\neq\varnothing$.
\end{lemma}
\begin{proof}
Let $\alpha$ be the smallest element in $A$, then $\alpha\leq a+\varepsilon n+1$. Let $J=[\alpha+1,\alpha+1+\lceil 2\varepsilon n\rceil]\subseteq [a+1,b]$. Observe that $\alpha+J\subseteq [b+1,n]$. Since $|J|=\lceil2\varepsilon n\rceil$, and $|[a+1,b]\setminus A|<\varepsilon n$, $|[b+1,n]\setminus B|<\varepsilon n$, this implies there is $\beta\in A\cap J$ such that $\alpha+\beta\in B$.
\end{proof}
The next lemma, Lemma~\ref{lem:mid4}, is similar to Lemma~\ref{lem:mid}, but here we consider the case when the number of colors is $4$. In order to obtain the same conclusion in Lemma~\ref{lem:mid}, we further require that the size of $A$ is significantly smaller than $n$, since if $A$ is close to $[n]$, when we color all the elements in $A$ by two colors, the number of colorings we obtained is also close to the extremal case. Note that if we use the same proof as in Lemma~\ref{lem:mid} for $r=4$, equation (\ref{eq:mid}) does not give us the conclusion we want. Hence the proof of Lemma~\ref{lem:mid4} requires a more careful and complicated analysis of the structures of the containers.
\begin{lemma}\label{lem:mid4}
Let $\varepsilon>0$ such that $ \big(1/2+\varepsilon\big)n\leq |A|\leq (1-\varepsilon)n$. Then there is $\delta_2=\delta_2(\varepsilon)>0$ such that
\[
g(A,4)\leq 4^{n/2-\delta_2n}.
\]
\end{lemma}
\begin{proof}
We apply Theorem~\ref{container} on $A$. Let $\mathcal{C}$ be the collection of containers, and let $P\in\mathcal{C}$ be a good template of order $n$. As what we did in the proof of Lemma~\ref{lem:mid}, we similarly apply Green's removal lemma on $P$, and obtain a template $P':[n]\setminus E\to 2^{[r]}$, such that $P\mid_{[n]\setminus E}=P'$, $|E|=o(n)$, and $P'$ is sum-free. Let $X_1,X_2,X_3\subseteq A\setminus E$ such that $X_1=\{a\in A\mid |P'(a)|=1\}$, $X_2=\{a\in A\mid |P'(a)|=2\}$, and $X_3=\{a\in A \mid |P'(a)|\geq 3\}$. Let $T=X_2\cup X_3$ and $x_i=|X_i|$ for $i\in [3]$.
Therefore, we have equations (\ref{eq:T}) still hold, and in particular, $X_3$ is sum-free. Thus $x_3\leq(n+1)/2$. Since $|A|\geq(1/2+\varepsilon)n$, and $P$ is good, we also obtain that $x_1+x_2\geq \varepsilon n/2$. Let $m=n-\alpha$ be the maximum element in $X_3$, by using the same argument, equations (\ref{eq:P3}) still hold.
Suppose we have either
\[
|T|\leq n-\Big\lceil\frac{m-1}{2}\Big\rceil-\frac{\varepsilon n}{1000},\quad\text{ or }\quad x_3\leq m-\Big\lceil\frac{m-1}{2}\Big\rceil-\frac{\varepsilon n}{1000}.
\]
Thus
\begin{align*}
g(\mathcal{C},A)\leq 2^{cn^{2/3}\log^2n}r^{|E|}2^{x_2}r^{x_3}
= r^{o(n)+\frac{1}{2}(|T|+x_3)}\leq r^{\frac{n}{2}+o(n)-\frac{\varepsilon n}{2000}}< r^{\frac{n}{2}-\delta_2 n},
\end{align*}
and we take $\delta_2=\frac{\varepsilon}{3000}.$ Therefore, we may assume that
\begin{equation}\label{eq:P3r=4}
n-\Big\lceil\frac{m-1}{2}\Big\rceil-\frac{\varepsilon n}{1000}\leq |T|\leq n-\Big\lceil\frac{m-1}{2}\Big\rceil,\text{ and }\ m-\Big\lceil\frac{m-1}{2}\Big\rceil-\frac{\varepsilon n}{1000}\leq x_3\leq m-\Big\lceil\frac{m-1}{2}\Big\rceil.
\end{equation}
In the rest of the proof, we are going to show that this is impossible.
Suppose that $\alpha\leq\frac{\varepsilon n}{40}$. Note that $\max\{x_1,x_2\}\geq\frac{\varepsilon n}{4}$, and
\[
(X_1+X_3)\cap X_3=\varnothing, \quad (X_2+X_3)\cap X_3=\varnothing,
\]
this contradicts Lemma~\ref{lem:str}. Thus we have $m\leq (1-\frac{\varepsilon}{40})n$. Since $|A|\leq (1-\varepsilon)n$, thus by the lower bound on $|T|$ in (\ref{eq:P3r=4}), $m\geq 3\varepsilon n/2$.
We now partition $[n]$ into three parts $J_1,J_2,J_3$, such that $J_1=[n-\alpha+1,n]$, $J_2=[1,\alpha]$, and $J_3=[\alpha+1,n-\alpha]$. By (\ref{eq:P3r=4}), we obtain that
\begin{equation}\label{eq:P2r=4}
|J_1\cap X_2|\geq \alpha-\frac{\varepsilon n}{1000}.
\end{equation}
Take $d=\frac{\varepsilon}{400}$. Suppose $|J_2\cap X_3|\geq dn$, then we can find $\beta\in X_3$ such that $\frac{dn}{2}\leq \beta\leq \alpha-\frac{dn}{2}$. Let $J_1'=[n-\alpha+1,n-\alpha+\frac{dn}{2}]\subseteq J_1$. Note that $(J_1'+\beta)\cap J_1'=\varnothing$, and $J_1'+\beta\subseteq J_1$, since $(X_2+X_3)\cap X_2=\varnothing$, and for every $i\in J_1'$, there is at least one element in the pair $\{i, i+\beta\}$ that is not contained in $X_2$, we have that $|J_1\cap X_2|\leq \alpha-\frac{dn}{2}=\alpha-\frac{\varepsilon n}{800}$, contradicts (\ref{eq:P2r=4}). Hence we may assume $|J_2\cap X_3|\leq dn$. Therefore, by (\ref{eq:P3r=4}), $|J_3\cap X_3|\geq \frac{n-\alpha}{2}-\frac{\varepsilon n}{1000}-dn$. This gives us an upper bound on $\alpha$ since $|J_3|\geq|J_3\cap X_3|$, that $\alpha\leq \frac{n}{3}+\frac{\varepsilon n}{1500}+\frac{2dn}{3}$. Next, we are going to show that actually we have $\alpha\leq \frac{n}{3}$. Suppose $\alpha>\frac{n}{3}$. Note that $|J_3\cap X_3|\geq \frac{n-\alpha}{2}-dn-\frac{\varepsilon n}{1000}$ implies
$$
|J_3\setminus X_3|\leq\frac{n-3\alpha}{2}+dn+\frac{\varepsilon n}{1000} \leq dn+\frac{\varepsilon n}{1000}<\frac{\varepsilon n}{250}.$$
By (\ref{eq:P2r=4}) and Lemma~\ref{lem:inter}, we get $(X_3+X_3)\cap X_2\neq\varnothing$, this contradicts (\ref{eq:T}).
Let $J_3'=[n-2\alpha,n-\alpha]$. We claim that
\begin{equation}\label{eq:J3r=4}
|J_3'\setminus X_3|\leq dn+\frac{\varepsilon n}{800}=\frac{3\varepsilon n}{800}.
\end{equation}
Otherwise, observe that at least one of $\{i,m-i\}$ is not in $X_3$, then $|X_3\cap (J_3\setminus J_3')|\leq\frac{n-3\alpha}{2}$ since $\alpha\leq n/3$. Hence
\[
x_3\leq \frac{m}{2}-\alpha+dn+\alpha-\Big(dn+\frac{\varepsilon n}{800}\Big)\leq \frac{m}{2}-\frac{\varepsilon n}{800},
\]
contradicts (\ref{eq:P3r=4}).
Next, let $d'=\frac{\varepsilon}{60}$, and suppose that $|(J_2+\alpha)\cap X_3|\geq d'n$. Thus there is $\gamma\in X_3$ such that $\alpha+\frac{d'n}{2}\leq \gamma\leq 2\alpha-\frac{d'n}{2}$. Let $J_1''=[n-2\alpha,n-2\alpha+\frac{d'n}{2}]$. Observe that $(\gamma+J_1'')\cap J_1''=\varnothing$, and $\gamma+J_1''\subseteq J_1$. Since $(X_2+X_3)\cap X_2=\varnothing$, we have either $|X_3\cap J_3'|\leq \alpha-\frac{d'n}{4}<\alpha-\frac{3\varepsilon n}{800}$, or $|X_2\cap J_1|\leq \alpha-\frac{d'n}{4}<\alpha-\frac{\varepsilon n}{1000}$, and in either case we get a contradiction with (\ref{eq:P2r=4}) or (\ref{eq:J3r=4}). Thus, we have $|(J_2+\alpha)\cap X_3|\leq d'n$. Note that $J_3'=\{m\}\cup (m-J_2)$, clearly, $|X_3\cap (J_2\cup J_3')|\leq \alpha+1$. Then by (\ref{eq:P3r=4}),
\[
\frac{n-3\alpha}{2}-\frac{\varepsilon n}{1000}\leq \big|X_{3}\cap [\alpha,n-2\alpha]\big|\leq n-4\alpha+d'n,
\]
hence $\alpha\leq\frac{n}{5}+\frac{2d'n}{5}+\frac{\varepsilon n}{2500}$. Suppose now $\alpha\geq \frac{n}{5}$. We have that $|X_3\cap [2\alpha,n-\alpha]|\geq \frac{n-\alpha}{2}-dn-d'n-\frac{\varepsilon n}{1000}$, which implies $|[2\alpha,n-\alpha]\setminus X_3|\leq \frac{\varepsilon n}{1000}+dn+d'n$. By (\ref{eq:P2r=4}) and Lemma~\ref{lem:inter}, we obtain that $(X_3+X_3)\cap X_2\neq\varnothing$, and this contradicts (\ref{eq:T}).
Finally, we get $\frac{\varepsilon n}{40}\leq \alpha<\frac{n}{5}$. By (\ref{eq:P3r=4}), we have that $x_3\leq\frac{2n}{5}$. By Lemma~\ref{Stabilitylem3}, either $X_3$ consists of odd integers, or the minimum element in $X_3$ is at least $\frac{n-\alpha}{2}-\frac{\varepsilon n}{1000}$. The first case is impossible. Otherwise, since $|J_2|=\alpha>2dn+\frac{\varepsilon n}{1000}$, we have
\[
x_3=|X_3\cap J_2|+ |X_3\cap J_3|\leq dn+\frac{n-2\alpha}{2}\leq\frac{n-\alpha}{2}-\frac{9\varepsilon n}{800},
\]and this contradicts the lower bound on $x_3$ in (\ref{eq:P3r=4}). Now, we assume $a\in X_3$ is the minimum element, and $a\geq \frac{n-\alpha}{2}-\frac{\varepsilon n}{1000}$. Observe that $|[\frac{n-\alpha}{2}+1,n-\alpha]\setminus X_3|\leq\frac{\varepsilon n}{500}$, $\frac{n-\alpha}{2}\leq\frac{n}{2}-\frac{\varepsilon n}{80}$, and by (\ref{eq:P2r=4}) and Lemma~\ref{lem:inter}, we have $(X_3+X_3)\cap X_2\neq\varnothing$, this contradicts (\ref{eq:T}).
\end{proof}
The final lemma consider the case when $A$ contains many Schur triples.
\begin{lemma}\label{lem:manyschur}
Let $r\geq4$ be an integer. Suppose there is $\mu>0$, such that $s(A)\geq\mu n^2$. Then
\[g(A,r)\leq r^{|A|-\frac{3(2\log r-\log (3r-2))}{2\log r}\mu n}.
\]
\end{lemma}
\begin{proof}
Since $s(A)\geq\mu n^2$, by Pigeonhole Principle, there is $t\in A$, such that
\[
s(t,A)\geq \frac{3\mu n^2}{|A|}\geq 3\mu n.
\]
Let the \textit{link graph} $L_t(A)$ to be the simple graph defined on the vertex set $A\setminus\{t\}$, such that $xy\in E(L_t(A))$ if and only if $\{t, x, y\}\in \mathcal{S}(t, A)$.
Let $k$ be the size of the maximum matching in $L_t(A)$.
Observe that $\Delta(L_t(A))\leq2$, and $|E(L_t(A))|=s(t,A)\geq 3\mu n$. Then we have $k\geq3\mu n/2$.
Now we consider the possible number of rainbow sum-free colorings of $A$. We first fix a maximum matching $M$ of $L_t(A)$. For the elements in $A\setminus V(M)$, we color them arbitrarily.
For each edge $ab\in E(M)$, in order to avoid a rainbow Schur triple, we either let $a, b$ share the same color, or color one of $a, b$ by the color of $t$, and color another vertex by a different color.
In this way, $a, b$ have exactly $r+2(r-1)$ effective colorings. Hence, we have
\begin{align*}
g(A,r)&\leq r^{|A|-2k-1}r(3r-2)^k
\leq r^{|A|}\Big(\frac{3r-2}{r^2}\Big)^{\frac{3\mu n}{2}}=r^{|A|-\frac{3(2\log r-\log (3r-2))}{2\log r}\mu n},
\end{align*}
as desired.
\end{proof}
Now we can prove the stability theorem.
\begin{proof}[{\bf Proof of Theorem \ref{thm:sta}}]
The first part of the statement, that $g(n,r)\leq r^{n/2+o(n)}$, follows easily from the fact $g(A,r)\leq r^{|A|}$ when $|A|\leq n/2+o(n)$. If $|A|\geq n/2+\eta n$ for some constant $\eta$, the result follows from Lemma~\ref{lem:mid} (ii) when $r\geq5$. For the case $r=4$, after applying Lemma~\ref{lem:mid4} we still have one extra case that $|A|\geq (1-\eta)n$, and this follows from Theorem~\ref{thm: hrange}.
For the second part of the statement, we will prove it by contrapositive. Let $c=\frac{3(2\log r-\log(3r-2))}{2\log r}$, clearly $c>0$ when $r\geq4$. Let $\mu$ be the value of $\delta(\frac{\varepsilon}{20})$ given in Lemma~\ref{Stabilitylem2}, and let $\varepsilon'=\min\{\frac{c\mu}{2},\varepsilon\}$. We first consider $r\geq5$, and suppose that we have both
\begin{equation}\label{eq:r=5}
|A\triangle O|>\varepsilon n, \quad\text{ and }\quad |A\triangle I_0|>\varepsilon n.
\end{equation}
In this case we take $\delta=\min\{\delta_1(\varepsilon'),\varepsilon',\frac{\varepsilon}{20}\}$, where $\delta_1(\varepsilon')$ is given in Lemma~\ref{lem:mid} (ii).
If $|A|\geq \frac{n}{2}+\varepsilon' n$, we apply Lemma~\ref{lem:mid} (ii) with parameter $\varepsilon'$, then we obtain that $g(A,r)\leq r^{n/2-\delta n}$. Thus we may assume that $|A|\leq (1/2+\varepsilon')n$. If $s(A)\geq\mu n^2$, applying Lemma~\ref{lem:manyschur}, then we have
\[
g(A,r)\leq r^{n/2+\varepsilon'n-c\mu n}\leq r^{n/2-\varepsilon' n}\leq r^{n/2-\delta n}.
\]
Finally, we have $s(A)<\mu n^2$. By Lemma~\ref{Stabilitylem2}, we get the partition $A=B\cup C$, where $B$ is sum-free and $|C|<\frac{\varepsilon}{20} n$. Note that we may assume $|A|\geq\frac{n}{2}-\frac{\varepsilon}{20} n$, otherwise $g(A,r)\leq r^{|A|}\leq r^{n/2-\delta n}$. Now we have $$|B|\geq |A|-|C|\geq \frac{n}{2}-\frac{\varepsilon n}{10}\geq\frac{2n}{5}$$ since $\varepsilon\leq1$. We apply Lemma~\ref{Stabilitylem3} on $B$. Hence either $B$ contains only odd integers, or the minimum element of $B$ is at least $|B|$. Suppose $B$ consists of odd integers. Thus
\[
|A\triangle O|\leq |C|+|O\setminus B|\leq \frac{\varepsilon}{20} n+\frac{\varepsilon}{10} n<\varepsilon n,
\]
contradicts (\ref{eq:r=5}). Thus, let $a$ be the minimum element in $B$, then $a\geq \frac{n}{2}-\frac{\varepsilon n}{10}$. Therefore,
\[
|A\triangle I_0|\leq |C|+|B\triangle I_0|\leq\frac{\varepsilon}{20}n+\frac{\varepsilon}{10}n+\frac{\varepsilon}{5}n<\varepsilon n,
\]
which also contradicts (\ref{eq:r=5}).
Next, let us consider the case when $r=4$. Besides (\ref{eq:r=5}), we further require\begin{equation}\label{eq:r=4}
|A\triangle [n]|>\varepsilon n.
\end{equation}
We now take $\delta=\min\{\delta_2(\varepsilon'),\varepsilon', \frac{\varepsilon}{20}\}$, where $\delta_2(\varepsilon')$ is given in Lemma~\ref{lem:mid4}. The case when $|A|\leq \frac{n}{2}+\varepsilon'n$ is same as when $r\geq5$. When $\frac{n}{2}+\varepsilon'n\leq |A|\leq n-\varepsilon'n$, by applying Lemma~\ref{lem:mid4} we get $g(A,4)\leq4^{n/2-\delta n}$. When $|A|\geq n-\varepsilon' n$, we get $|A\triangle [n]|\leq \varepsilon' n\leq\varepsilon n$, which contradicts (\ref{eq:r=4}).
\end{proof}
\section{Proof of Theorem~\ref{thm: r8}}
For a subset $A \subseteq [n]$ and an integer $t$, recall that $L_t(A)$ is the simple graph defined on $A\setminus\{t\}$, in which $xy\in E(L_t(A))$ if and only if $\{t, x, y\}\in \mathcal{S}(t, A)$.
Let $k(t, A)$ be the size of the maximum matching of $L_t(A)$. Note that $\Delta(L_t(A))\leq 2$.
Therefore we have
\begin{equation}\label{ineq:mat}
k(t, A)\geq |E(L_t(A))|/2=s(t, A)/2.
\end{equation}
\begin{prop}\label{prop:larmat}
Let $n, r, c \in N$ with $r\geq 8$ and $c>1$. Suppose that $A$ is a subset of $[n]$ of size $\lceil n/2\rceil+c$.\smallskip
\begin{compactenum}[\rm (i)]
\item If there exists an element $t\in A$ such that $k(t, A)\geq 2(c-1)$, then we have $g(A,r)< r^{\lceil n/2\rceil+1}$.\smallskip
\item If there exists an element $t\in A$ such that $k(t, A)\geq 2(c-1)+1$, then we have $g(A,r)< r^{\lceil n/2\rceil}\left(3-2/r\right)$.
\end{compactenum}
\end{prop}
\begin{proof}
First note that for $r\geq 8$ we have
\begin{equation}\label{eq:estr}
\frac{(3r-2)^2}{r^3}<1.
\end{equation}
Let $k=k(t, A)$. Similarly as in the proof of Lemma~\ref{lem:manyschur}, we obtain that
$$g(A,r)\leq r^{\lceil n/2\rceil+c-2k}(3r-2)^k
=r^{\lceil n/2\rceil+c}\left(\frac{3r-2}{r^2}\right)^{k}.$$
For $k\geq 2(c-1)$, we have
$$
g(A,r)\leq r^{\lceil n/2\rceil+c}\left(\frac{3r-2}{r^2}\right)^{2(c-1)}
=r^{\lceil n/2\rceil+1}\left(\frac{(3r-2)^2}{r^3}\right)^{c-1}
< r^{\lceil n/2\rceil+1},
$$
where the last inequality follows from (\ref{eq:estr}) and $c>1$.
Similarly, for $k\geq 2(c-1)+1$, we have
\[
\begin{split}
g(A,r)&\leq r^{\lceil n/2\rceil} \left(3-\frac{2}{r}\right) r^{c-1}\left(\frac{3r-2}{r^2} \right)^{k-1}
\leq r^{\lceil n/2\rceil} \left(3-\frac{2}{r}\right)r^{c-1}\left(\frac{3r-2}{r^2} \right)^{2(c-1)}\\
&= r^{\lceil n/2\rceil} \left(3-\frac{2}{r}\right)\left(\frac{(3r-2)^2}{r^3}\right)^{c-1}
< r^{\lceil n/2\rceil} \left(3-\frac{2}{r}\right).
\end{split}
\]
Together with the previous inequality, this completes the proof.
\end{proof}
\begin{lemma}\label{lem:staodd}
Let $r\geq 8$, $0<\varepsilon \leq 1/36$, and $A$ be a subset of $[n]$ of size $\lceil n/2\rceil + c$, where $1 < c \leq \varepsilon n$. Suppose that there exists a partition $A=B\cup C$ such that $B$ consists of odd numbers and $|C|\leq \varepsilon n$. Then we have $g(A,r)< r^{\lceil n/2\rceil}\left(3-2/r\right)$.
\end{lemma}
\begin{proof}
From the assumption of $A$, there must be an even number $t\in A$.
By Proposition~\ref{prop:larmat}(ii) and $\varepsilon \leq 1/36$, it is sufficient to show that $k(t, A)\geq (1/12-\varepsilon)n -1$.
Recall that $O$ is the set of all odd numbers in $[n]$. Since $|A|\geq \lceil n/2\rceil + 1$ and $|C|\leq \varepsilon n$, we have $|O\setminus B|\leq \varepsilon n$.
Then,
\[
k(t, A)\geq k(t, B)\geq k(t, O)-\varepsilon n,
\]
and thus it is equivalent to show that $k(t, O)\geq n/12 - 1$.
If $t\geq n/3$, then we immediately have
\[
k(t, O)\geq \left|\left\{(i, t-i, t), i\in O\cap\left[t/2-1\right]\right\}\right|\geq t/4-1\geq n/12-1.
\]
If $t < n/3$, then by~(\ref{ineq:mat}) we obtain that
\[
k(t, O)\geq s(t, O)/2\geq \left|\left\{(t, i, t+i), i\in O\cap\left[t+1,\ n-t\right]\right\}\right|/2
\geq (n-2t)/4\geq n/12.
\]
This completes the proof.
\end{proof}
\begin{lemma}\label{lem:staint}
Let $r\geq 8$, $0<\varepsilon \ll 1$, and $A$ be a subset of $[n]$ of size $\lceil n/2\rceil + c$, where $1 < c \leq \varepsilon n$. Suppose that there exists a partition $A=B\cup C$ such that $B\subseteq I_0$ and $|C|\leq \varepsilon n$. Then the following holds.\smallskip
\begin{compactenum}[\rm (i)]
\item If $n$ is even, then $g(A,r)< r^{\lceil n/2\rceil+1}$.\smallskip
\item If $n$ is odd, then $g(A,r)< r^{\lceil n/2\rceil}\left(3-2/r\right)$.
\end{compactenum}
\end{lemma}
\begin{proof}
Let $m$ be the minimum element of $A$, and clearly $m \leq\lfloor n/2\rfloor -(c-1)$.
Recall that $I_0=[\lfloor n/2\rfloor +1, n]$.
Let $d=|I_0\setminus A|$. From the assumption of $A$, we have $d\leq \varepsilon n$. We divide the proof into four cases.\smallskip
\noindent\textbf{Case 1: $m\leq d+3(c-1)$.} In this case, we have $m\leq 4\varepsilon n$.
Similar to the proof of Lemma~\ref{lem:staodd}, we have
\[
k(m, A)\geq k(m, B)\geq k(m, I_0) - d
\geq s(m, I_0)/2 - \varepsilon n
\geq (n/2 - m)/2 - \varepsilon n
\geq n/4 - 3\varepsilon n,
\]
which, together with Proposition~\ref{prop:larmat}(ii) and $\varepsilon\ll 1$, completes the proof.\smallskip
\noindent\textbf{Case 2: $d+3(c-1)< m \leq \lceil n/2 \rceil-d-3(c-1)$.}
Since $m\leq n/2$, each nontrivial component of $L_m(I_0)$ is a path, and there are $\min\left\{m, \lceil n/2\rceil - m\right\}\geq d+3(c-1)$ of them.
Therefore we have
\[
k(m, A)\geq k(m, I_0) - d
\geq d+3(c-1) - d=3(c-1)
\]
which, together with Proposition~\ref{prop:larmat}(ii) and $c>1$, completes the proof.\smallskip
\noindent\textbf{Case 3: $\lceil n/2\rceil-d-3(c-1)< m \leq \lceil n/2\rceil-2(c-1)$.}
By the choice of $m$, each nontrivial component of $L_m(A)$ is a path of length 1, and the number of them is exactly
\[
s(m, [m+1, n])- |[m+1, n]\setminus A|
= n-2m - (n -|A| -(m-1))
= \lceil n/2\rceil +(c-1)-m.
\]
Therefore, we obtain that
\[
k(m, A)=\lceil n/2\rceil +(c-1)-m\geq 3(c-1),
\]
which completes the proof together with Proposition~\ref{prop:larmat}(ii) and $c>1$.\smallskip
\noindent\textbf{Case 4: $\lceil n/2\rceil-2(c-1)<m \leq \lfloor n/2\rfloor-(c-1)$.}
Similarly as in Case 3, we obtain that $k(m, A)=\lceil n/2\rceil +(c-1)-m$. By the choice of $m$, for even $n$, we have $k(m, A)\geq 2(c-1)$, while for odd $n$, $k(m, A)\geq 2(c-1)+1$.
By Proposition~\ref{prop:larmat}, this gives the desired upper bounds.
\end{proof}
\begin{proof}[{\bf Proof of Theorem~\ref{thm: r8}.}]
Here we only prove (i) as the proof of (ii) is similar.
If $|A|=\lceil n/2 \rceil +1$ and $A\neq I_2$, then $A$ must have at least one restricted Schur triple, and therefore $g(A, r)<g(I_2, r)=r^{\lceil n/2 \rceil +1}$.
When $|A|>\lceil n/2 \rceil +1$, choose a constant $\varepsilon \ll 1$, which satisfies the assumptions of Lemmas~\ref{lem:staodd} and~\ref{lem:staint}. Then by Theorem~\ref{thm:sta}, we can further assume that $|A \bigtriangleup O|\leq \varepsilon n$, or $|A\bigtriangleup I_0|\leq \varepsilon n$. Applying Lemmas~\ref{lem:staodd} and~\ref{lem:staint}(i) on $A$, for both cases, we obtain $g(A, r)<r^{\lceil n/2 \rceil +1}$.
\end{proof}
\section{Concluding Remarks}
Our investigation raises many open problems.
In this paper, we determine the rainbow $r$-extremal sets, that is, the subsets of $[n]$ which maximize the number of rainbow sum-free $r$-colorings, for $r\leq3$ and $r\geq8$. However, for $r\in\{4,5,6,7\}$, although Theorem~\ref{thm:sta} says the rainbow $r$-extremal sets should be close to what we expect, our proofs cannot give the exact structure of the extremal sets.
Therefore, the most interesting question is to determine the unsolved cases of Conjecture~\ref{conj}.
Recall that $I_1=[\frac{n}{2}-1,n]$ and $I_3=[\frac{n-1}{2},n]$.
\begin{conj}\label{conj:concluding}
Let $n, r$ be positive integers and $4\leq r\leq 7$.\smallskip
\begin{compactenum}[\rm (i)]
\item If $n$ is even, then $g(n, r)=r^{n/2}\left(3 - 2/r\right)^2$, and $I_1$ is the unique rainbow $r$-extremal set.\smallskip
\item If $n$ is odd and $r=4$, then $g(n, r)=g([n], r)$, and $[n]$ is the unique rainbow $r$-extremal set.\smallskip
\item If $n$ is odd and $5\leq r\leq 7$, then $g(n, r)=r^{\lceil n/2\rceil}\left(3 - 2/r\right)$, and $I_3$ is the unique rainbow $r$-extremal set.
\end{compactenum}
\end{conj}
Another direction is that one can consider various generalization of this problem.
Recall that a sum-free set is a set forbidding the solutions of the linear equation $x_1+x_2=y$.
It is natural to extend the Erd\H{o}s-Rothschild problems on sets forbidding solutions of other linear equations, for example, the $(k, \ell)$-free sets, that is, the sets without nontrivial tuples $\{x_1, \ldots, x_{k}, y_1,\ldots, y_{\ell}\}$ satisfying $\sum_{i=1}^kx_i = \sum_{j=1}^\ell y_j$.
It is possible that the method used to prove Theorem~\ref{thm: hrange} can prove the analogous results for some other $(k, \ell)$-free sets. However, the stability analysis on other parts would be very involved.
One could also broaden the study of rainbow Erd\H{o}s-Rothschild problems to various other extremal problems in this fashion.
In the rainbow Erd\H{o}s-Rothchild problems studied to date, that is, the Gallai colorings and the rainbow sum-free colorings, for $r=3$ the configurations maximizing the number of such colorings are complete graphs or the whole intervals, while for sufficiently large $r$ the optimal configurations are those solving the original extremal problems.
It would be very interesting to determine the threshold of $r$ to ensure that the extremal configurations for the uncolored problems are optimal for rainbow Erd\H{o}s-Rothschild problems.
\section*{Acknowledgements}
This work was supported by the Natural Science Foundation of China (11871311, 11631014) and Shandong University multidisciplinary research and innovation team of young scholars.
Part of the work was done while the second and the third authors were visiting the School of Mathematics at Shandong University. They would like to thank the school for the hospitality they received.
\bibliographystyle{abbrv}
| 35b276d518d25fab4a018337312fed6d1b8c86e1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The thermodynamic properties of hydrogen and helium gases enter as a basic and key ingredient in the study of the structure and evolution of dense astrophysical bodies like stars, brown dwarfs and giant planets, for these objects are mostly made of a mixture of H and He. Many works have been performed to determine the equation of state of H-He mixtures and also to address the related question of the helium solubility in hydrogen, which is important for a correct description of planetary interiors~\cite{Fantoni2015,Morales2013,Lorenzen2009}. The hydrogen-helium mixture is fundamentally a quantum plasma made of electrons and nuclei interacting via the Coulomb potential. Numerical simulation techniques, like density-functional theory molecular dynamics~\cite{Militzer2013,Lorenzen2009,Vorberger2007,Lenosky2000,Perrot1995}, Path Integral Monte Carlo~\cite{Militzer2001,Militzer2000b,Pierleoni1994} and Quantum Monte Carlo~\cite{Mazzola2018}, have been used to calculate, with good precision, some thermodynamical properties of H-He mixtures in strongly interacting regimes. Besides simulations, analytical calculations, in particular asymptotic expansions, are useful to provide theoretical insights and to complement the simulation data with reliable results in asymptotic regimes, like at low density or at high density or high temperature~\cite{Alastuey2012,Chabrier2019}. Such expansions have been derived using various analytical tools: the effective potential method~\cite{Morita1959,Ebeling1967,Deutsch1977}, many-body perturbation theory~\cite{deWitt1995,Kremp2005,Alastuey1994} and two different path-integral formalisms: Mayer diagrammatical expansions in the ring-polymer representation~\cite{Alastuey1994,Alastuey1996,Brydges1999,Ballenegger2002} and an effective field theory~\cite{Brown2001}. It has been checked explicitly that these quite different theoretical frameworks all lead to the same expansion at low densities~\cite{Alastuey2015}. Beyond asymptotic expansions at low or high densities, analytical theories can also provide insights at intermediate densities via the introduction of suitable approximations (see e.g. \cite{Ebeling2012,Alastuey2012a,Ramazanov2015,Bunker1997}).
\bigskip
In this article, we derive two general results that enable easier (full or partially) analytical calculations of the equation of state of H-He mixtures, and also other plasmas, in the low and moderate density regimes. First, we obtain a new exact representation for all terms in the activity expansion of the grand-potential $\Omega = -P \Lambda$ ($\Lambda$ denotes the volume) of a quantum multi-component plasma. This so-called screened Mayer activity-series for $\Omega$, or equivalently for the pressure $P$, complements the activity-series for distribution functions derived in Ref.~\cite{Ballenegger2002} and provides a much more direct route for computing the equation of state: less diagrams need to be computed and no term-by-term integration of contributions to distribution functions needs to be performed. Since $\Omega$ is a thermodynamic potential, all thermodynamic properties can furthermore be deduced from it via standard thermodynamic relations. The screened Mayer series are not perturbative with respect to the strength of the interaction, in contrast to the expressions of standard many-body perturbation theory (thermodynamic Green function formalism~\cite{Kremp2005}), and allow therefore calculations not only in the fully ionized regimes, at low or high densities, but also in moderately dense regimes where the particles are bound into atoms and/or molecules.
\bigskip
In the grand-canonical ensemble, the particle densities are deduced from the pressure thanks to the standard thermodynamical relation
$\rho_\alpha = {\partial P}/{\partial \mu_\alpha}$.
If charge neutrality is automatically satisfied for the exact expression of $P$, it is not necessarily the case for an approximate expression $P_{\rm A}$ of $P$. We introduce therefore, and this constitutes our second main result, general procedures for making any approximation $P_{\rm A}$ automatically consistent with electroneutrality. In these procedures, either dressed or neutral-group activities, are introduced based on general properties of quantum plasmas at equilibrium.
We then deduce directly from $P_{\rm A}$, without any equation to solve, an associated thermodynamical potential that is compatible with electroneutrality.
We show that the various neutralization prescriptions do not lead in general to identical results for the equation of state.\footnote{unless the original approximation $P_{\rm A}$ is already compatible with electroneutrality, as for example in an asymptotic expansion, in which case these procedures are without effect.} The choice of a particular prescription is hence worthy of attention since it is not inconsequential.
\bigskip
The paper is organized as follows. In Section~\ref{sec:S2}, we define the model and recall that electroneutrality always holds in the bulk of a plasma, whatever the activities $\{z_\alpha \}$ are. This implies that one can impose the pseudo-neutrality condition $\sum_{\alpha=1}^{\mathcal{S}} \mathrm{e}_\alpha z_\alpha = 0$ with $\mathcal{S}$ the number of species, without loss of generality and that the average bulk properties necessarily depend only on
the temperature and on $(\mathcal{S}-1)$ independent variables $\{y_i \}$, called neutral-group activities. An approximate expression~$P_{\rm A}(T;\{ z_{\alpha} \})$ of the pressure, which is not necessarily compatible with electroneutrality, can then be made compatible by adjusting it so that it depends on the activities only through
$(\mathcal{S}-1)$ neutral-group activities.
This defines the neutral-group neutralization prescription, which is new to our knowledge. This prescription is not unique for a plasma with three or more components because there are several ways, when $\mathcal{S} \geq 3$, of grouping particles together such that each group is charge neutral.
\bigskip
The screened activity series for $P(T;\{ z_{\alpha} \})$ is derived in Section~\ref{sec:S3}. This series is obtained within the path-integral representation
of the quantum system in terms of an equivalent classical gas of loops. This allows one to apply two standard classical techniques, namely Mayer diagrammatical expansions and Abe-Meeron summations~\cite{Mayer1940,Abe1959,Meeron1958}.
The known screened diagrammatic series for the particle densities~\cite{Alastuey2003} are recovered, as it should,
by differentiating the present series for the pressure. The
$z$-series for the particle densities do satisfy the local charge neutrality order by order, thanks to
the combination of the pseudo-neutrality condition with a Debye-dressing mechanism. We
show how the well-known virial expansion of the EOS up to order $\rho^2$ can be recovered by
keeping a few simple diagrams.
\bigskip
In Section~\ref{DDR}, we introduce a Debye-Dressing prescription which automatically ensures that the particle densities
inferred from any approximate function~$P_{\rm A}(T;\{ z_{\alpha} \})$ do satisfy local charge neutrality. This prescription is directly inspired by
the Debye-dressing mechanism at work in the screened Mayer diagrammatic series. This method is compared to other
procedures for ensuring electroneutrality, like the Neutral-Group procedure (Section~\ref{sec:S2}) and
the Enforced-Neutrality method~\cite{Starostin2005}.
\bigskip
In Section~\ref{Sec:S5}, we show how approximate equations of state at moderate densities
can be constructed by using the diagrammatic series
for $P(T;\{ z_{\alpha} \})$, together with densities deduced via simple derivatives in which either the neutral-group activities (Section~II) or the Debye-Dressed activities (Section~IV) are used to ensure that electroneutrality is satisfied.
We point out that important physical mechanisms, like the recombination of nuclei and electrons into chemical species and atom-charge interactions, can be taken into account by retaining a few selected diagrams, in the spirit of the ACTEX method introduced by Rogers~\cite{Rogers1974,Rogers1981,Rogers1997,Rogers2002b} which underlie the OPAL thermodynamic tables~\cite{OPAL1996,Rogers2002}.
Some conclusions and perspectives are eventually given in Section~\ref{sec:S6}.
\section{Neutrality in the grand-canonical ensemble}\label{sec:S2}
\subsection{Quantum multi-component Coulomb system and the thermodynamic limit}
We consider a quantum multi-component plasma made of $\mathcal{S}$ species of charged point particles enclosed
in a box with volume $\Lambda$.
The species index is denoted by $\alpha \in \{1,...,\mathcal{S}\}$. Each particle of species $\alpha$
has a mass $m_\alpha$, while it carries a charge $e_\alpha$ and a spin $s_\alpha$. Each of them
obeys to either Bose or Fermi statistics, according to the integer or half-integer value of $s_\alpha$ respectively.
In order to ensure thermodynamic stability, at least one species needs to be fermions~\cite{Lieb1972} and there must be both positively and negatively
charged species. The species
$\alpha$ and the position $\mathbf{x}$ of a given particle is denoted by the single notation
$\textsf{\textbf{x}}=(\alpha,\mathbf{x})$. The total interaction potential $U(\textsf{\textbf{x}}_1,...,\textsf{\textbf{x}}_N)$ of $N$ particles
is the sum of pairwise pure Coulomb interactions,
\begin{equation}
\label{IX.QMG1}
U(\textsf{\textbf{x}}_1,...,\textsf{\textbf{x}}_N)= \sum_{i <j} V_{\rm C}(\textsf{\textbf{x}}_i,\textsf{\textbf{x}}_j)
\end{equation}
with
\begin{equation}
\label{IX.QMG2}
V_{\rm C}(\textsf{\textbf{x}}_i,\textsf{\textbf{x}}_j)=e_{\alpha_i} e_{\alpha_j} v_{\rm C}(|\mathbf{x}_i - \mathbf{x}_j|)
\end{equation}
and $v_{\rm C}(r)=1/r$. The corresponding non-relativistic Coulomb Hamiltonian reads
\begin{equation}
H_{N}=-\sum_{i=1}^{N}\frac{\hbar^2}{2m_{\alpha_i}} \Delta_i + U(\textsf{\textbf{x}}_1,...,\textsf{\textbf{x}}_N)
\label{IX.QMG3}
\end{equation}
where $\Delta_i$ is the Laplacian with respect to position $\mathbf{x}_i$. The nucleo-electronic plasma
is an example of such multi-component system, where the negative point charge are electrons,
(species $\alpha=\mathcal{S}$) while all positive point charges are nuclei (species $\alpha=1,...,\mathcal{S}-1$).
\bigskip
As proved by Lieb and Lebowitz~\cite{Lieb1972}, the present quantum multi-component plasma has a well-behaved thermodynamic
limit (TL), and all statistical ensembles become equivalent in this limit. In the grand-canonical
ensemble the TL is defined by fixing the chemical potentials $\mu_\alpha$ of each species as well as the
inverse temperature $\beta=1/(k_{\rm B}T)$, and letting the volume $\Lambda \to \infty$.
The grand-partition function $\Xi_{\Lambda}$ of the finite system reads
\begin{equation}
\Xi_{\Lambda}= \mathrm{Tr} \exp[-\beta(H-\sum_{\alpha=1}^{\mathcal{S}} \mu_{\alpha}N_{\alpha})] \; ,
\label{VI.48}
\end{equation}
where the trace runs on all particle numbers, not only on neutral configurations.
The grand canonical pressure
\begin{align}
P_\Lambda(T; \{ \mu_{\alpha} \})=\frac{k_{\rm B}T \ln \Xi_{\Lambda}}{\Lambda}
\label{VI.48bis}
\end{align}
has a well-defined thermodynamic limit
\begin{align}
P(T; \{ \mu_{\alpha} \})= k_{\rm B}T \lim_{\rm \Lambda \to \infty}\frac{ \ln \Xi_{\Lambda}}{\Lambda} \; .
\label{PressureTL}
\end{align}
\bigskip
As a consequence of elementary electrostatics, for non-neutral configurations
associated with $\sum_{\alpha=1}^{\mathcal{S}} e_{\alpha}N_{\alpha} = Q \neq 0$, the excess charges are expelled to the surface~\cite{Lieb1972},
so the system maintains charge neutrality in the bulk. Moreover, the Coulomb energy associated with these excess charges
is of order $Q^2/(2R)$ for a spherical box of radius $R$. Non-neutral configurations with a macroscopic charge proportional to the volume
$\Lambda$ do not contribute to $\Xi$, since their weights involve the factor $\exp(-\beta Q^2/R)$ which
vanishes faster than $\exp(- C \Lambda^{1+ \epsilon})$ with $C,\epsilon >0$. In fact, $\langle Q \rangle_{\Lambda,\rm GC}$ remains
of order $R$ in the TL whatever the chemical potentials are, so the average charge density vanishes in the TL,
\begin{equation}
\label{AverageCharge}
\lim_{\rm TL} \frac{\langle Q \rangle_{\Lambda,{\rm GC}}}{ \Lambda} = 0,
\end{equation}
while the total surface charge-density carried by the walls of the box
also vanishes in the TL~\cite{Lieb1972}.
Hence the average densities defined by
\begin{equation}
\label{AverageDensity}
\rho_{\alpha}= \lim_{\rm TL} \frac{\langle N_\alpha \rangle_{\Lambda,\rm GC}}{\Lambda}
\end{equation}
satisfy the overall charge neutrality
\begin{equation}
\sum_{\alpha=1}^{\mathcal{S}}e_{\alpha}\rho_{\alpha}=0 \; .
\label{VI.49}
\end{equation}
Moreover, they coincide with the local bulk densities in the TL in a fluid phase.
Thanks to the thermodynamic identity
\begin{equation}
\label{DensityChemicalPotential}
\rho_{\alpha}= \frac{\partial P}{\partial \mu_{\alpha}}(T;\{ \mu_{\gamma} \}) \; ,
\end{equation}
charge neutrality~(\ref{VI.49}) can be recast as
\begin{equation}
\sum_{\alpha=1}^{\mathcal{S}}e_{\alpha} \frac{\partial P}{\partial \mu_{\alpha}} (T;\{ \mu_{\gamma} \})=0 \; .
\label{VI.50}
\end{equation}
The identity~(\ref{VI.50}) is valid for any set $\{ \mu_{\gamma} \}$. Because of this identity,
the bulk properties do not depend independently on all $\mathcal{S}$ chemical potentials: there is necessarily
a combination of these chemical potentials that is irrelevant. This remarkable property allows one to
introduce $(\mathcal{S} -1)$ independent neutral-group chemical potentials, as well as the pseudo-neutrality condition~\eqref{PseudoNeutrality},
as detailed in the next sections~\ref{sec:II.B} and \ref{sec:II.Bbis}.
\subsection{Introduction of $(\mathcal{S} - 1)$ independent Neutral-Group chemical potentials}
\label{sec:II.B}
We start with the simplest case of a two-component system ($\mathcal{S}=2$) made with nuclei ($\alpha = 1={\rm n}$) carrying a charge $e_1= Z e$ and electrons ($\alpha=2={\rm e}$) carrying a charge $e_2=-e$.
Due to the identity~\eqref{VI.50}, there is one relevant combination of chemical potentials which entirely determines the equilibrium state in the TL.
As discussed previously the leading configurations which contribute to the grand-canonical trace~\eqref{VI.48} are almost neutral.
i.e. the numbers $(N_{\rm n}, N_{\rm e})$ of nuclei and electrons are such that $N_{\rm e} \simeq N_{\rm n} Z$. Accordingly,
$(\mu_{\rm n} N_{\rm n} +\mu_{\rm e} N_{\rm n})$ is close to $\mu N_{\rm n}$ with
\begin{equation}
\label{RelevantCombination}
\mu= \mu_{\rm n} + Z \mu_{\rm e} \; ,
\end{equation}
which can be viewed as the chemical potential of an elementary neutral-group made with a single nuclei and $Z$ electrons.
Such a linear combination, together with $T$, entirely determines the pressure,\textsl{ i.e.}
$P(T; \mu_{\rm n},\mu_{\rm e})=P(T;\mu)$, { in agreement with the Lieb-Lebowitz theorem~\cite{Lieb1972,Brydges1999}}. The particle densities~(\ref{DensityChemicalPotential}) can then be recast as
\begin{align}
\label{DensityChemicalBis}
&\rho_{\rm n}= \frac{\partial P}{\partial \mu}(T;\mu)\frac{\partial \mu}{\partial \mu_{\rm n}}
=\frac{\partial P}{\partial \mu}(T;\mu) \nonumber \\
&\rho_{\rm e}= \frac{\partial P}{\partial \mu}(T;\mu)\frac{\partial \mu}{\partial \mu_{\rm e}} \; ,
=Z \frac{\partial P}{\partial \mu}(T;\mu) \; .
\end{align}
and they obviously satisfy local charge neutrality.
\bigskip
For multi-component systems with three or more components, we can determine in a similar way $(\mathcal{S}-1)$ relevant combinations of the chemical potentials.
Let us consider that species $(\alpha=1,...,\mathcal{S}-1)$ are nuclei with charges $Z_\alpha e$, while species $\alpha=\mathcal{S}={\rm e}$ are electrons with charges
$-e$. Elementary neutral-groups can be constructed by associating $Z_\alpha$ electrons to a single given nuclei with species $\alpha$.
The associated neutral-group (NG) chemical potentials are the $(\mathcal{S}-1)$ combinations
\begin{equation}
\label{RelevantCombinationGeneral}
\mu_\alpha^{\rm NG} = \mu_\alpha + Z_\alpha \mu_{\rm e} \, , \quad \alpha=1,...,\mathcal{S}-1 \; ,
\end{equation}
which, together with the temperature, entirely determine the equilibrium state. Of course, when $\mathcal{S} \geq 3$, there are several ways to constitute $(\mathcal{S}-1)$ elementary neutral groups \footnote{For instance, one could define another set of neutral groups from $(\mathcal{S}-1)$ vectors orthogonal to the charge vector $\pmb{e}=(Z_1, ..., Z_{\mathcal{S}-1}, -1)e$ by setting the abundance of particles of species $\alpha$ in the neutral group associated to a vector $\vec{v}$ orthogonal to~$\pmb{e}$ to be proportional to the component $v_\alpha$ of that vector.
The particular choice~\eqref{RelevantCombinationGeneral} for neutral-groups is associated with a particularly simple basis for the subspace orthogonal to $\pmb{e}$.}. This freedom of choice for the set of independent relevant variables $\{\mu_\alpha^{\rm NG}\}$ is however inconsequential in an exact calculation and it would not affect any physical prediction. This arbitrariness is due to the fact that there are several ways of grouping particles together such that each group is charge-neutral.
\subsection{Neutral-Group activities} \label{sec:II.Bbis}
It is useful to translate the previous considerations in terms of the particle activities
\begin{equation}
z_\alpha=(2s_\alpha+1)\frac{\mathrm{e}^{\beta\mu_{\alpha}}}{(2\pi \lambda_\alpha^{2})^{3/2}}\; ,
\label{ParticleActivity}
\end{equation}
where $\lambda_\alpha=(\beta \hbar^2/m_\alpha)^{1/2}$ is the de Broglie thermal wavelength of the
particles of species $\alpha$. Let us consider the neutral-group chemical potentials~(\ref{RelevantCombinationGeneral}).
They provide $(\mathcal{S}-1)$ neutral-group activities
\begin{equation}
\label{RelevantActivity}
y_i =
\left[z_i z_{\rm e}^{{{Z}}_i} \right]^{1/(1+{Z}_i)}
\end{equation}
where the exponent $1/(1+Z_i)$ has been introduced in the definition of $y_i$ so that it has the dimension of an activity, \textit{i.e.} a density.
The pressure depends solely on the $(\mathcal{S}-1)$ neutral-group activities $y_i$ and on the temperature, i.e.
$P(T;\{ \mu_{\alpha} \})=P(T;\{ \mu_{\alpha}^{\rm NG}\})=P(T;\{ y_i \})$. The thermodynamical identity~(\ref{DensityChemicalPotential})
which provides the particle densities is then rewritten as
\begin{equation}
\label{DensityNewActivity}
\rho_{\alpha}= z_\alpha \sum_{i=1}^{\mathcal{S}-1} \frac{\partial \beta P}{\partial y_i}(T;\{ y_j \})
\frac{\partial y_i}{\partial z_{\alpha}}(\{ z_{\gamma} \})\; .
\end{equation}
The total local charge density reads
\begin{equation}
\label{ChargeDensityNewActivity}
\sum_\alpha e_\alpha \rho_{\alpha}= \sum_{i=1}^{\mathcal{S}-1} \frac{\partial \beta P}{\partial y_i}(T;\{ y_j \})
\sum_\alpha e_\alpha z_\alpha \frac{\partial y_i}{\partial z_{\alpha}}(\{ z_{\gamma} \})\; ,
\end{equation}
and it indeed always vanishes since
\begin{equation}
\label{OrthogonalityProperty}
\sum_\alpha e_\alpha z_\alpha \frac{\partial y_i}{\partial z_{\alpha}}(\{ z_{\gamma} \})
= Z_i e z_i\frac{\partial y_i}{\partial z_i} -e z_{\rm e}\frac{\partial y_i}{\partial z_{\rm e}} = e y_i \left(\frac{Z_i}{Z_i+1}-\frac{Z_i}{Z_i+1} \right) = 0\; .
\end{equation}
\bigskip
Since the particle densities are determined solely by the $(\mathcal{S}-1)$
neutral-group activities $y_i$ and by the temperature, different sets of activities can lead to the same set of densities $\{\rho_\alpha\}$. It is common to break this redundancy by imposing, without any loss of generality
as far as bulk properties are concerned, the so-called pseudo-neutrality (or bare-neutrality) condition~\cite{Brydges1999,Brown2001}
\begin{equation}
\label{PseudoNeutrality}
\sum_\alpha e_\alpha z_\alpha=0 \; .
\end{equation}
{Notice that fixing the electrons' activity in terms of the $(\mathcal{S}-1)$ nuclei's activities via this relation does not affect the range of variations $[0, \infty]$ of each $y_i$ variable.}
The choice~(\ref{PseudoNeutrality}) is particularly useful for various purposes, in particular it simplifies the derivation of the
low-density expansion of the EOS as explained in Section~\ref{sec:S3}.
\subsection{Neutral-Group neutralization scheme}
\label{sec:S23}
\subsubsection{NG neutralization prescription}
A given approximate theory, that is a given function for the pressure $P_{\rm A}(T; \{z_\alpha\})$ in the TL, is not necessarily
compatible with neutrality, {\it i.e.} the particle densities inferred \textsl{via } the standard identities
\begin{equation}
\label{DensityPressureA}
\rho_{\alpha}= z_{\alpha} \frac{\partial P_{\rm A}}{\partial z_{\alpha}}(T;\{ z_{\gamma} \}) \; ,
\end{equation}
do not satisfy the local charge neutrality~(\ref{VI.49}) in general. In other words,
it is not possible to express $P_{\rm A}(T; \{z_\alpha\})$ solely in terms of the neutral-group activities $\{y_i\}$ and the temperature,
as it can be done for the exact pressure. However, one can modify the approximate theory via the following general procedure
to make it compatible with neutrality.
\bigskip
Let us introduce the associated approximation
\begin{equation}
\label{NGpressure}
P_{\rm A}^{\rm NG}(\beta;\{ z_\alpha \})=P_{\rm A}(\beta;\{ z_\alpha^{\rm NG}(y_1(\{z_\alpha\}),...,y_{\mathcal{S}-1}(\{z_\alpha\})) \})
\end{equation}
where each $z_\alpha$ in $P_{\rm A}(\beta;\{ z_\alpha\})$ is replaced by a Neutral-Group function
$z_\alpha^{\rm NG}(y_1(\{z_\alpha\}),$ $..., y_{\mathcal{S}-1}(\{z_\alpha\}))$ which depends on the genuine activities $\{z_\alpha\}$
through the neutral-group activities $\{y_i\}$ [Eq.~\eqref{RelevantActivity}].
The dependence of the NG functions $\{z_\alpha^{\rm NG}\}_{\alpha=1, ..., \mathcal{S}}$ on the variables $\{y_i\}_{i, ..., \mathcal{S}-1}$
is obtained by inverting the system of equations
\begin{subequations}
\begin{numcases}{}
%
y_i = \left[z_i^{\rm NG} (z_{\rm e}^{\rm NG})^{{{Z}}_i} \right]^{1/(1+{Z}_i)}, \qquad i = 1, ..., \mathcal{S}-1 \label{29a} \\
%
\sum_{\alpha=1}^\mathcal{S} e_\alpha z_\alpha^{\rm NG} = 0 \hspace{5.2cm}\mbox{}
\label{29b}
\end{numcases}
\end{subequations}
which combines the definitions~\eqref{RelevantActivity} of the neutral-group variables [where $z_i$ is replaced by $z_i^{\rm NG}$] with pseudo-neutrality.
Hence, for the specific set of genuine activities $\{z_\alpha\}$ which satisfy pseudo-neutrality,
each function $\{ z_\alpha^{\rm NG}(\{y_i\}) \}$ takes the value $\{z_\alpha\}$. Notice that the variations of the functions
$\{z_\alpha^{\rm NG}\}$, with the $\{z_\alpha\}$'s treated as independent variables, are entirely defined by the choice of
neutral groups. The Neutral-Group functions do not depend in particular on the considered approximate theory.
The present ``back-and-forth'' conversion, from the $\mathcal{S}$ genuine activities to $(\mathcal{S}-1)$ neutral-group activities $\{y_i\}$
to $\mathcal{S}$ activity-functions $\{z_\alpha^{\rm NG}(z_{\gamma})\}$, ensures that the associated approximation $P_{\rm A}^{\rm NG}$
depends on the activities only via the neutral-group activities, and therefore that $\Omega^{\rm NG} = -P_{\rm A}^{\rm NG}(T,\{z_\alpha\})\Lambda$ is a thermodynamic potential compatible with electroneutrality.
\bigskip
By construction, the associated pressure $P_{\rm A}^{\rm NG}(\beta;\{ z_\alpha \})$ only depends on the activities $\{z_\alpha\}$ via the relevant neutral-group activities, so it leads to particle densities that satisfy local charge neutrality.
The particle densities can be computed by applying
the general rules for partial derivatives of composite functions, which provides
\begin{align}
\label{DensityPressureANG}
\rho_{\alpha} &= z_{\alpha} \frac{\partial P_{\rm A}^{\rm NG}}{\partial z_{\alpha}}(T;\{ z_{\gamma} \})
= z_{\alpha} \frac{\partial }{\partial z_{\alpha}} P_{\rm A}(T;\{ z_\gamma^{\rm NG}(y_1(\{z_\delta\}),...,y_{\mathcal{S}-1}(\{z_\delta\})) \})
\nonumber \\
&= z_{\alpha} \sum_{\delta=1}^{\mathcal{S}} \sum_{i=1}^{\mathcal{S}-1} \frac{\partial P_{\rm A}}{\partial z_{\delta}} (T;\{z_\gamma^{\rm NG} \})
\frac{\partial z_{\delta}^{\rm NG}}{\partial y_i}
\frac{\partial y_i}{\partial z_{\alpha}}\; ,
\end{align}
The partial derivatives $\partial P_{\rm A}/\partial z_{\gamma}$, with $z_\gamma$ treated as an independent variable, have to be evaluated
at the end for the set $\{ z_\gamma^{\rm NG}(y_1,...,y_{\mathcal{S}-1}) \}$
that satisfies the pseudo-neutrality condition~(\ref{PseudoNeutrality}).
It is convenient to consider that the set of genuine activities $\{ z_\gamma \}$ already
satisfy this condition since each function $z_\gamma^{\rm NG}(y_1,...,y_{\mathcal{S}-1})$ then exactly coincides with $z_\gamma$ at the end.
\bigskip
With this neutralization prescription, the pressure is left unchanged for a pseudo-neutral set $\{z_\alpha\}$,
\textit{i.e.} $P_{\rm A}^{\rm NG} = P_{\rm A}$ for such set, and the theory is internally consistent since the densities are deduced from the pressure via the standard thermodynamic relation $\rho_\alpha = z_\alpha \frac{\partial}{\partial z_\alpha} P_{\rm A}^{\rm NG}(T,\{ z_\gamma\})$.
\subsubsection{Explicit neutralization formulae}
Using the definition~(\ref{RelevantActivity}) of the Neutral-Group activities, the densities~(\ref{DensityPressureANG})
can be recast as
\begin{align}
\label{DensityPressureANGbis}
&\rho_i = \sum_{\theta=1}^{\mathcal{S}} \frac{\partial P_{\rm A}}{\partial z_{\theta}} (T;\{z_\gamma \})
\frac{y_i}{Z_i+1} \frac{\partial z_{\theta}^{\rm NG}}{\partial y_i}
\quad \text{for} \quad i=1,...,\mathcal{S}-1 \nonumber \\
&\rho_{\rm e}= \sum_{\theta=1}^{\mathcal{S}} \frac{\partial P_{\rm A}}{\partial z_{\theta}} (T;\{z_\gamma \})
\sum_{j=1}^{\mathcal{S}-1} \frac{Z_j y_j}{Z_j+1} \frac{\partial z_{\theta}^{\rm NG}}{\partial y_j}
\; .
\end{align}
Taking the logarithm of the definitions~(\ref{RelevantActivity}), we obtain
\begin{equation}
\label{RelevantActivityLogarithm}
(Z_i+1) \ln y_i = \ln z_i^{\rm NG} + Z_i \ln z_{\rm e}^{\rm NG} \quad i=1,...,\mathcal{S}-1 \; .
\end{equation}
The partial derivatives ${\partial z_{\theta}^{\rm NG}}/{\partial y_i}$ can be calculated by differentiating each side of Eq.~(\ref{RelevantActivityLogarithm})
with respect to $y_i$ in a first step, and then with respect to $y_j$ for $j \neq i$ in a second step. Inserting the resulting
expressions for ${\partial z_{\theta}^{\rm NG}}/{\partial y_i}$ into the
formula~(\ref{DensityPressureANGbis}), we eventually find
\begin{align}
\label{DensityPressureANGter}
&\rho_i = \sum_{\theta=1}^{\mathcal{S}} C_{i,\theta} \; z_\theta \; \frac{\partial P_{\rm A}}{\partial z_{\theta}} (T;\{z_\gamma \})
\quad \text{for} \quad i=1,...,\mathcal{S}-1 \nonumber \\
&\rho_{\rm e}= \sum_{j=1}^{\mathcal{S}-1} Z_j \rho_j \; ,
\end{align}
with coefficients
\begin{equation}
\label{CoefficientsNGDensityA}
C_{i,\theta}=
\begin{cases}
\displaystyle
- \frac{Z_i Z_j z_i}{2 z_{\rm e} + Z_i(Z_i-1)z_i} & \text{if $\theta \neq i$ and $\theta \neq \mathcal{S}$} \\
\displaystyle\rule{0mm}{8mm}
1-\frac{Z_i^2 z_i}{2 z_{\rm e} + Z_i(Z_i-1)z_i} & \text{if $\theta = i$} \\
\displaystyle\rule{0mm}{8mm}
\frac{Z_i z_i}{2 z_{\rm e} + Z_i(Z_i-1)z_i} & \text{if $\theta = S$ [\textsl{i.e.} $\theta$ = e]}
\end{cases}
\end{equation}
\bigskip
By construction, the particle densities generated by the Neutral-Group prescription \eqref{DensityPressureANGter}-\eqref{CoefficientsNGDensityA}
do satisfy the local charge
neutrality~(\ref{VI.49}), whatever the partial derivatives $\partial P_{\rm A}/\partial z_{\gamma}$ are. We recall that
such derivatives must be calculated for a set $\{ z_\gamma \}$ which fulfills the pseudo-neutrality condition~(\ref{PseudoNeutrality}).
\bigskip
As a first check, we verify that if the approximate pressure $P_{\rm A}$ is consistent with local charge neutrality, namely if
\begin{equation}
\label{APconsistencyNeutrality}
z_{\rm e} \frac{\partial P_{\rm A}}{\partial z_{\rm e}} = \sum_{j=1}^{\mathcal{S}-1} Z_j z_j \frac{\partial P_{\rm A}}{\partial z_j} \; ,
\end{equation}
then formula~(\ref{DensityPressureANGter}) for each density $\rho_i$ does reduce to $z_i\partial P_{\rm A}/\partial z_{i} $ by virtue of the
identities
\begin{align}
\label{CoefficientsIdentity}
& Z_j C_{i,{\rm e}} + C_{i,j}= 0 \quad \text{for} \quad j=1,...,\mathcal{S}-1 \; , \; j \neq i \nonumber \\
& Z_i C_{i,{\rm e}} + C_{i,i}= 1 \; .
\end{align}
Moreover, if the approximate pressure is the ideal Maxwell-Boltzmann expression, i.e. if
\begin{equation}
\label{MBPressure}
\beta P_{\rm A}= \beta P_{\rm MB}= z_{\rm e}+\sum_{j=1}^{\mathcal{S}-1} z_j
\end{equation}
each density $\rho_i$ given by formula~(\ref{DensityPressureANGter}) does reduce to $z_i$ as a consequence of
$
\sum_{j=1}^{\mathcal{S}-1} C_{i,j} z_j + C_{i,{\rm e}} z_{\rm e}= z_i
$.
\subsubsection{Explicit formulae for two- and three-component plasmas}
Let us consider a two-component plasma, $\mathcal{S}=2$ with $(1={\rm n},Z_1=Z$).
The nuclei and electron densities~(\ref{DensityPressureANGter}) then become
\begin{align}
\label{DensityPressureANGTCP}
&\rho_{\rm n} = \frac{z}{(Z+1)}
\left[ \frac{\partial P_{\rm A}}{\partial z_{\rm n}} + Z \frac{\partial P_{\rm A}}{\partial z_{\rm e}} \right]
\nonumber \\
&\rho_{\rm e}= \frac{Z z}{(Z+1)}
\left[ \frac{\partial P_{\rm A}}{\partial z_{\rm n}} + Z \frac{\partial P_{\rm A}}{\partial z_{\rm e}} \right]
\; ,
\end{align}
where $\partial P_{\rm A}/\partial z_{\rm n}$ and $\partial P_{\rm A}/\partial z_{\rm e}$ are calculated for
the set $(z_{\rm n}=z,z_{\rm e}=Z z )$. Note that for the hydrogen plasma, the nuclei are protons with $Z=1$, while for
the helium plasma the nuclei are alpha-particles with $Z=2$.
\bigskip
For three-component systems, $\mathcal{S}=3$, the nuclei densities~(\ref{DensityPressureANGter}) read
\begin{multline}
\label{DensityPressureANG3CP1}
\rho_1 = \frac{(Z_1z_1 +2Z_2 z_2) z_1 }{Z_1(Z_1+1) z_1 + 2Z_2 z_2} \; \frac{\partial P_{\rm A}}{\partial z_1} (T;\{z_\gamma \})
-\frac{Z_1 Z_2 z_1 z_2}{Z_1(Z_1+1) z_1 + 2Z_2 z_2} \frac{\partial P_{\rm A}}{\partial z_2} (T;\{z_\gamma \}) \\
+ \frac{Z_1 z_1 (Z_1z_1+Z_2z_2)}{Z_1(Z_1+1) z_1 + 2Z_2 z_2} \frac{\partial P_{\rm A}}{\partial z_{\rm e}} (T;\{z_\gamma \}) \; ,
\end{multline}
and
\begin{multline}
\label{DensityPressureANG3CP2}
\rho_2 = \frac{(2Z_1z_1 + Z_2 z_2)z_2}{2 Z_1 z_1 + Z_2(Z_2+1)z_2} \; \frac{\partial P_{\rm A}}{\partial z_2} (T;\{z_\gamma \})
-\frac{Z_1 Z_2 z_1 z_2}{2 Z_1 z_1 + Z_2(Z_2+1)z_2} \; \frac{\partial P_{\rm A}}{\partial z_1} (T;\{z_\gamma \}) \\
+ \frac{Z_2 z_2(Z_1z_1+Z_2z_2)}{2 Z_1 z_1 + Z_2(Z_2+1)z_2} \; \frac{\partial P_{\rm A}}{\partial z_{\rm e}} (T;\{z_\gamma \})\; ,
\end{multline}
with the electron density $\rho_{\rm e}=Z_1 \rho_1 + Z_2 \rho_2$. These formulae can be applied to the case of
the hydrogen-helium mixture made with protons $(1={\rm p},Z_1=1$) and alpha-nuclei ($2={\rm alpha}, Z_2=2$). The proton and alpha-particle activities
$z_{\rm p}$ and $z_{\rm alpha}$ can take arbitrary values, while the electron activity is set to $z_{\rm e}=z_{\rm p} + 2 z_{\rm alpha}$.
The nuclei-densities depend on the two independent
activities $z_{\rm p}$ and $z_{\rm alpha}$. Note that the relative concentrations of hydrogen and helium, determined
by the ratio $\rho_{\rm p}/\rho_{\rm alpha}$, do not depend only on the ratio $z_{\rm p}/z_{\rm alpha}$ in general.
\subsubsection{Comments}
The present neutral-group neutralization prescription if quite appealing because:
(i) It is general and straightforward to implement since no equation needs to be solved;
(ii) It is based on exact properties of the system, namely that it maintains neutrality in the bulk and that the dependence
of the pressure on the activities occurs only via $(\mathcal{S}-1)$ neutral-group activities $\{y_i\}$ which have a clear physical interpretation;
(iii) The associated electroneutrality-compatible theory $P_{\rm A}^{\rm NG}(T,\{z_\alpha\})$
is internally consistent since the densities are deduced from this function via the standard thermodynamic relation;
(iv) The original value of the pressure is left unchanged after neutralization $P_{\rm A}^{NG} = P_{\rm A}$ if the genuine $z_\alpha$'s satisfy the pseudo-neutrality condition.
\bigskip
The neutral-group neutralization prescription is not unique in a plasma with 3 or more components. Indeed, when $\mathcal{S} \geq 3$,
other choices of the neutral groups would lead to expressions of the particle densities
similar to Eqns.~(\ref{DensityPressureANGter}) but with coefficients different from Eqns.~(\ref{CoefficientsNGDensityA}).
However, the formulae for these coefficients in terms of the particle activities are expected to be much more complicated than the
rational fractions~(\ref{CoefficientsNGDensityA}). In fact the choice~(\ref{RelevantActivity}) ensures that
$\partial y_i /\partial z_j = 0$ for $j \neq i$, which greatly simplifies the calculations of the $C_{i,\delta}$'s.
\section{Activity expansion of the pressure} \label{sec:S3}
Mayer diagrams have been introduced while ago~\cite{Mayer1940} in order to derive
low-density expansions of equilibrium quantities for classical
systems with short-range pair interactions. For charged fluids, every Mayer diagram
diverges because of the long-range of Coulomb interactions.
Abe~\cite{Abe1959} and Meeron~\cite{Meeron1958} showed that such divergences can be removed \textit{via}
systematic summations of convolution chains built with the Coulomb interaction. The whole Mayer series is then
exactly transformed into a series of so-called prototype graphs, with the same topological structure as the Mayer diagrams, but
with effective bonds built with the familiar Debye potential in place of the bare Coulomb interaction. The contribution of each
prototype graph is finite thanks to the screening collective effects embedded in the Debye potential.
\subsection{The equivalent classical gas of loops}
The trace~(\ref{VI.48}) defining
$\Xi_{\Lambda}$ can be expressed in position and spin space, where the corresponding states have to be symmetrized
according to Bose or Fermi statistics. The corresponding sum involves both diagonal and off-diagonal
matrix elements of $\exp(-\beta H_N)$ in position space. Diagonal matrix elements account for Maxwell-Boltzmann
statistics, while off-diagonal matrix elements describe exchange contributions.
Within the Feynman-Kac representation,
all the matrix elements of $\exp(-\beta H_N)$ in position space can be rewritten as
functional integrals over paths followed by the particles.
The off-diagonal matrix elements generate open paths. However all the open paths followed by
the particles exchanged in a given cyclic permutation, can be collected into a closed filamentous object,
called a loop $\mathcal{L}$, or sometimes a ring-polymer, in the literature.
Each contribution of a given spatial matrix element of $\exp(-\beta H_N)$ for a given set of particles can be
related to that of a classical Boltzmann factor for a set of loops. In a last non-trivial step, the sum
of all these contributions, namely $\Xi_{\Lambda}$, is recast as the grand-partition function of
a classical gas of loops~\cite{Ginibre1971,Cornu1996,Martin2003}
\begin{equation}
\Xi_{\Lambda}=\Xi_{\Lambda,{\rm Loop}}=\sum_{N=0}^{\infty}\frac{1}{N\;!}\[ \prod_{i=1}^{N} \int_{\Lambda}D(\mathcal{L}_{i})z(\mathcal{L}_{i})\]
\mathrm{e}^{-\beta U(\mathcal{L}_{1},\mathcal{L}_{2},\ldots,\mathcal{L}_{N})} \;,
\label{IX.QMG6}
\end{equation}
thereby establishing a mapping, at equilibrium, between the quantum gas and a classical gas of loops.
The loop phase-space measure $D(\mathcal{L})$, loop fugacity $z(\mathcal{L})$, and total interaction potential
$U(\mathcal{L}_{1},\mathcal{L}_{2},\ldots,\mathcal{L}_{N})$ are defined as follows.
\bigskip
\begin{figure}[h]
\includegraphics[scale=0.8]{loop_k2opt.pdf}
\caption{A loop made with the paths of 3 particles exchanged in a permutation cycle ($1 \to 2 \to 3 \to 1$).}
\label{Fig1}
\end{figure}
A loop $\mathcal{L}$ located at $\mathbf{x}$ containing $q$ particles of species $\alpha$,
is a closed path $\mathbf{X}(s)=\mathbf{x}+\lambda_{\alpha}\pmb{\mathcal{X}}(s)$, parametrized by an
imaginary time $s$ running from $0$ to $q$ where $\pmb{\mathcal{X}}(s)$, the shape of the loop, is a Brownian bridge
subjected to the constraints $\pmb{\mathcal{X}}(0)=\pmb{\mathcal{X}}(q)=\mathbf{0}$ {{(Fig.~\ref{Fig1})}}.
The state of a loop, collectively denoted by $\mathcal{L}=\{\mathbf{x}, \chi\}$, is defined by its position $\mathbf{x}$ together with an
internal degree of freedom $\chi=\{\alpha, q,\pmb{\mathcal{X}}\}$, which includes its shape $\pmb{\mathcal{X}}$ as well as
the number $q$ of exchanged particles of species $\alpha$. The loop phase-space measure $D(\mathcal{L})$ means summation over all these
degrees of freedom,
\begin{equation}
\int_{\Lambda} D(\mathcal{L})\cdots=\sum_{\alpha=1}^{\mathcal{S}}\sum_{q=1}^{\infty}\int_{\Lambda}\d \mathbf{x}\int D_{q}(\pmb{\mathcal{X}})\cdots \; .
\label{L.27}
\end{equation}
The functional integration over the loop shape $D_{q}(\pmb{\mathcal{X}})$ is the normalized Gaussian measure for
the Brownian bridge $\pmb{\mathcal{X}}(s)$ entirely defined by its covariance
\begin{equation}
\int D_{q}(\pmb{\mathcal{X}})\pmb{\mathcal{X}}_{\mu}(s_{1})\pmb{\mathcal{X}}_{\nu}(s_{2})=q\delta_{\mu\nu}\[\min\left(\frac{s_{1}}{q},\frac{s_{2}}{q}\right)-\frac{s_{1}}{q}\frac{s_{2}}{q}\] \; .
\label{IX.QMG7bis}
\end{equation}
The loop activity reads
\begin{equation}
z(\mathcal{L})=(2s_\alpha+1)\frac{\eta_\alpha^{q-1}}{q}\frac{\mathrm{e}^{\beta\mu_{\alpha}q}}{(2\pi q\lambda_\alpha^{2})^{3/2}}\mathrm{e}^{-\beta U_{\rm self}(\mathcal{L})} \; ,
\label{IX.QMG8}
\end{equation}
where the factor $\eta_\alpha=1$ for bosons and $\eta_\alpha=-1$ for fermions.
Moreover, $U_{\rm self}(\mathcal{L})$
is the self-energy of the loop which is generated by the interactions between the exchanged particles,
\begin{align}
U_{\rm self}(\mathcal{L})=\frac{e_\alpha^2}{2}\int_{0}^{q}\d s\int_{0}^{q}\d s'(1-\delta_{[s][s']})\tilde{\delta}(s-s')
v_{\rm C}(\lambda_\alpha \pmb{\mathcal{X}}(s)-\lambda_\alpha \pmb{\mathcal{X}}(s')) \;,
\label{IX.QMG9}
\end{align}
with the Dirac comb
\begin{equation}
\tilde{\delta}(s-s')=\sum_{n=-\infty}^{\infty}\delta(s-s'-n)=\sum_{n=-\infty}^{{\infty}}\mathrm{e}^{2i\pi n(s-s')} \; .
\label{IX.QMG10}
\end{equation}
The Dirac comb ensures that particles only interact at equal times $s$ along their paths, as required by the Feynman-Kac formula,
while the term $(1-\delta_{[s][s']})$ removes the contributions of self-interactions ($[s]$ denote the integer part of~$s$).
Eventually, the total interaction potential $U(\mathcal{L}_{1},\mathcal{L}_{2},\ldots,\mathcal{L}_{N})$ is a sum of pairwise interactions,
\begin{equation}
\label{IX.QMG11}
U(\mathcal{L}_{1},\mathcal{L}_{2},\ldots,\mathcal{L}_{N})=\frac{1}{2} \sum_{i \neq j} V(\mathcal{L}_i,\mathcal{L}_j)
\end{equation}
with
\begin{align}
&V(\mathcal{L}_i,\mathcal{L}_j)=e_{\alpha_i} e_{\alpha_j} \int_{0}^{q_i}\d s_i\int_{0}^{q_j}\d s_j \tilde{\delta}(s_i-s_j)
v_{\rm C}(\mathbf{x}_i+\lambda_{\alpha_i}\pmb{\mathcal{X}}_i(s_i)-\mathbf{x}_j-\lambda_{\alpha_j}\pmb{\mathcal{X}}_j(s_j)) \; .
\label{IX.QMG12}
\end{align}
The loop-loop interaction $V(\mathcal{L}_i,\mathcal{L}_j)$ is generated by the interactions between any particle inside $\mathcal{L}_i$ and any particle inside $\mathcal{L}_j$.
Like in formula~(\ref{IX.QMG9}), the Dirac comb~(\ref{IX.QMG10}) guarantees that interactions are taken at equal times along particle paths.
\bigskip
The introduction of the gas of loops is particularly useful at low densities, because the standard Mayer diagrammatic expansions, valid
for classical systems with pairwise interactions,
can be straightforwardly applied by merely replacing points by loops. However, as in the case of classical Coulomb systems, the Mayer diagrams for the loop gas are plagued with divergences arising from the large-distance behavior
\begin{equation}
\label{IX.QMG28}
V(\mathcal{L}_{i},\mathcal{L}_{j}) \sim\frac{q_{i} e_{\alpha_{i}} q_{j} e_{\alpha_{j}}}{|\mathbf{x}_{i}-\mathbf{x}_{j}|} \;\quad \text{when} \; |\mathbf{x}_{i}-\mathbf{x}_{j}| \to \infty \; .
\end{equation}
Note that such behavior is nothing but the Coulomb interaction between point charges, because the finite spatial extents of loops $\mathcal{L}_i$
and $\mathcal{L}_j$ can be neglected with respect to their large relative distance $|\mathbf{x}_{i}-\mathbf{x}_{j}|$. It has been shown that all these long-range
divergences can be removed within a suitable extension of the Abe-Meeron summation process
introduced long ago for classical Coulomb fluids. The method has been applied for both the one- and two-body distribution functions~\cite{Cornu1996,Alastuey2003}.
In the next Section, we derive the corresponding Abe-Meeron series for the pressure.
\subsection{Abe-Meeron like summations for the pressure}
The Mayer diagrammatical expansion of the pressure,
\begin{equation}
\label{VBMayerP}
\beta P =
\sum_{{\cal G}} \frac{1}{S({\cal G})}
\int \left[ \prod D(\mathcal{L}) z(\mathcal{L}) \right]
\left[ \prod b_{\rm M}\right]_{{\cal G}} \;,
\end{equation}
involves simply connected diagrams ${\cal G}$ made with $N=1, 2, ...$ field (black) points, representing loops with statistical weight $z(\mathcal{L})$, and Mayer bonds $b_{\rm M}$ defined by
\begin{equation}
\label{MayerBond}
b_{\rm M}(\mathcal{L}_i,\mathcal{L}_j)= \exp(-\beta V(\mathcal{L}_i,\mathcal{L}_j)) -1 \; .
\end{equation}
The contribution of a given ${\cal G}$ is
calculated by labeling arbitrarily the $N$ field points (loops). $S({\cal G})$ denotes the symmetry factor, which is the number of permutations of those labeled field loops that leave the product of bonds
and weights unchanged.
An integration \eqref{L.27} is performed over the degrees of freedom of each field loop.
Thanks to translation invariance, once the
integration over $(N-1)$ black loops have been performed in ${\cal P}$, the result no longer depends on the position of the remaining black loop. The
$1/\Lambda$ factor in the definition~(\ref{VI.48bis}) of the pressure of the finite system can then be absorbed in the thermodynamic limit
by keeping the position of one loop fixed, i.e. by integrating only over $(N-1)$ loops and on the internal degrees of freedom of the fixed loop.
\bigskip
Due to the large-distance behavior~\eqref{IX.QMG28}, any Mayer diagram ${\cal G}$ involving more than one loop is divergent in the thermodynamic limit. Let us eliminate these divergences systematically by summing diagrams in classes, as in the classical case~\cite{Abe1959,Meeron1958}. Since exactly the same counting and combinatorics formulae intervene in these summations as in the classical case, we won't detail them. Note that
simplified presentations of the summation process for the one-body loop density are given in Refs.~\cite{Cornu1996} and~\cite{Alastuey2003}.
The key staring point is the decomposition of the Mayer bond \eqref{MayerBond}
into
\begin{equation}
\label{DecompositionBond}
b_{\rm M}(\mathcal{L}_i,\mathcal{L}_j)=b_{\rm T}(\mathcal{L}_i,\mathcal{L}_j) + b_{\rm I}(\mathcal{L}_i,\mathcal{L}_j) \; ,
\end{equation}
with the interaction bond
\begin{equation}
\label{InteractionBond}
b_{\rm I}(\mathcal{L}_i,\mathcal{L}_j)=-\beta V(\mathcal{L}_i,\mathcal{L}_j) \;
\end{equation}
and the truncated bond
\begin{equation}
\label{TruncatedBond}
b_{\rm T}(\mathcal{L}_i,\mathcal{L}_j)= \exp(-\beta V(\mathcal{L}_i,\mathcal{L}_j)) -1 + \beta V(\mathcal{L}_i,\mathcal{L}_j) \; .
\end{equation}
\begin{figure}[h]
\includegraphics[scale=1.15]{BareBonds.pdf}
\caption{The bonds before summations. The last four bonds are generated by decomposing the original Mayer bond.}
\label{FigBareBonds}
\label{Fig2}
\end{figure}
\begin{figure}[h]
\includegraphics[scale=1.15]{EndPoint.pdf}
\caption{Diagram with an ending loop (point 4).}
\label{Fig3}
\end{figure}
Graphical representations for these bonds are given in Fig.~\ref{FigBareBonds}.
A loop $\mathcal{L}_i$ which is singly connected to a loop $\mathcal{L}_j$ is called an ending loop (Fig.~\ref{Fig3}).
A Mayer bond $b_{\rm M}(\mathcal{L}_i,\mathcal{L}_j)$ connected to such a loop is decomposed as
\begin{equation}
\label{DecompositionBondEnding}
b_{\rm M}(\mathcal{L}_i,\mathcal{L}_j)=b_{\rm TE}(\mathcal{L}_i,\mathcal{L}_j) + b_{\rm I}(\mathcal{L}_i,\mathcal{L}_j) + [b_{\rm I}(\mathcal{L}_i,\mathcal{L}_j)]^2/2\; ,
\end{equation}
with the truncated ending bond
\begin{equation}
\label{TruncatedEndingBond}
b_{\rm TE}(\mathcal{L}_i,\mathcal{L}_j)=\exp (-\beta V(\mathcal{L}_i,\mathcal{L}_j)) - 1 + \beta V(\mathcal{L}_i,\mathcal{L}_j)-(\beta V(\mathcal{L}_i,\mathcal{L}_j))^2/2 \; .
\end{equation}
These two decompositions can be represented graphically
\begin{align}
&\includegraphics{DecompositionT.pdf}\\
&\includegraphics{DecompositionTE.pdf}\;.
\end{align}
After inserting these decompositions into every diagram ${\cal G}$, a pair of loops $\mathcal{L}_k$ and $\mathcal{L}_l$ can be connected either by $b_{\rm I}$ or $b_{\rm T}$ (if none of the two loops is an ending loop) or by $b_{\rm I}$, $\frac{1}{2} b^2_{\rm I}$ or $b_{\rm TE}$ (if at least one of the two loops is an ending loop). We proceed then to systematic summations of all chain convolutions $b_{\rm I} \ast b_{\rm I} \ast...b_{\rm I} \ast b_{\rm I} $ made with arbitrary numbers $p$
of interaction bonds $b_{\rm I}$. Such a convolution chain can link a loop $\mathcal{L}_i$ to another loop $\mathcal{L}_j$ or to itself ($\mathcal{L}_j = \mathcal{L}_i$), in which case we call this convolution chain a ring.
The sum of $p=1,2,... \infty$ single convolution chains between two fixed loops $\mathcal{L}_i$ and~$\mathcal{L}_j$,
\begin{equation} \label{defPhi}
\begin{graph}
\node[label=below:{$i$},blanc] (1) {};
\node[label=below:{$j$},blanc] [right of=1] (2) {};
\drawLinkD{1}{2} ;
\end{graph}
\quad = \quad
\begin{graph}
\node[blanc] (1) {};
\node[blanc] [right of=1] (2) {};
\drawLinkFC{1}{2} ;
\end{graph}
\;\, + \;\,
\begin{graph}
\node[blanc] (1) {};
\node[noir] [right of=1] (2) {};
\node[blanc] [right of=2] (3) {};
\drawLinkFC{1}{2};
\drawLinkFC{2}{3} ;
\end{graph}
\;\, + \;\,
\begin{graph}
\node[blanc] (1) {};
\node[noir] [right of=1] (2) {};
\node[noir] [right of=2] (3) {};
\node[blanc] [right of=3] (4) {};
\drawLinkFC{1}{2} ;
\drawLinkFC{2}{3} ;
\drawLinkFC{3}{4} ;
\end{graph}
\;\, + \;\,
\cdots
\end{equation}
generates the Debye bond
\begin{equation}
\label{DebyeBond}
b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)=-\beta e_{\alpha_i} e_{\alpha_j} \phi(\mathcal{L}_i,\mathcal{L}_j) \; ,
\end{equation}
where $\phi(\mathcal{L}_i,\mathcal{L}_j)$ is the quantum analogue of the Debye potential, which reads~\cite{Ballenegger2002}
\begin{equation}
\label{IX.QMG47}
\phi(\mathcal{L}_{i},\mathcal{L}_{j})= \int_{0}^{q_i}\d s_i\int_{0}^{q_j}\d s_j
\; \psi_{\rm loop}(\mathbf{x}_j+\lambda_{\alpha_{j}}\pmb{\mathcal{X}}_j(s_j)-\mathbf{x}_i-\lambda_{\alpha_i}\pmb{\mathcal{X}}_i(s_i), s_i-s_j) \; ,
\end{equation}
with
\begin{equation}
\label{IX.QMG48}
\psi_{\rm loop}(\mathbf{r},s)= \sum_{n=-\infty}^\infty \exp(2 i \pi n s) \tilde{\psi}_{\rm loop}(\mathbf{r},n) \; \;
\end{equation}
and
\begin{equation}
\label{IX.QMG49}
\tilde{\psi}_{\rm loop}(\mathbf{r},n)= \int \frac{ \d^3 \mathbf{k}}{(2\pi)^3} \exp(i \mathbf{k} \cdot \mathbf{r}) \frac{4 \pi }{k^2 + \kappa^2(k,n)} \; .
\end{equation}
Note that $\tilde{\psi}_{\rm loop}(\mathbf{r},n)$ has a structure analogous to the classical Debye form,
except that an infinite number of frequency-dependent screening factors $\kappa^{2}(k,n)$ occur,
\begin{equation}
\label{IX.QMG45a}
\kappa^2(k,n)=4\pi \beta \sum_\alpha \sum_{q=1}^\infty q e_{\alpha}^2 \int_0^q \d s \exp(2 i \pi n s) \int D_{q}(\pmb{\mathcal{X}})
\exp(i \mathbf{k} \cdot \lambda_{\alpha}\pmb{\mathcal{X}}(s)) z(\chi) \; .
\end{equation}
The collective effects are embedded in these screening factors $\kappa^2(k,n)$, while the frequencies $2 \pi n$ are the analogues
of the familiar Matsubara frequencies in the standard many-body perturbative series.
\bigskip
Similarly to the case of the Mayer diagrams for the one-body loop density, the summation of all convolution chains in the Mayer
diagrams for the pressure can be expressed in terms of $\phi$, except in the single ring diagrams built with arbitrary numbers $p \geq 2$ of interaction bonds $b_{\rm I}$,
\begin{equation} \label{eq:rings}
\beta P_{\rm R} = \quad
\begin{graph}
\twoNodesUnNamed;
\drawLinkFCC{1}{2};
\end{graph}
\,+\,
\begin{graph}
\ringsAttached{3}{0.6}{1};
\end{graph}
\,+\,
\begin{graph}
\ringsAttached{4}{0.6}{1};
\end{graph}
\,+\,
\begin{graph}
\ringsAttached{5}{0.6}{1};
\end{graph}
\,+\,\ldots
\quad\equiv \begin{graph}
\node[noir, thick] (1) {};
\drawLoopNsnakeAngle{1}{90};
\end{graph}
\end{equation}
which provide the contribution $\beta P_{\rm R}$.
In such diagrams made with $p$ black points, the symmetry factor is $1/(2p)$, in contrast to the single chain diagrams in Eq.~\eqref{defPhi} where the symmetry factor is $1$ for any $p$.
After expressing each bare interaction $V$ in Fourier space, we find that the contribution to the pressure of a
single ring made with $p$ black loops and $p$ bonds $b_{\rm I}=-\beta V$ reduces to
\begin{equation}
\label{RingPp}
\frac{1}{2}\sum_{n=-\infty}^\infty \int \frac{\d^3 \mathbf{k}}{(2\pi)^3}\frac{1}{p} \left[\frac{-\kappa^2(k ,n)}{k^2}\right]^p \; .
\end{equation}
The calculation is similar to that involved in the convolution chain and gives again rise to the screening factors $\kappa^{2}(k,n)$.
Now, the summation over $p$ of all ring contributions leads to a logarithmic function instead of the rational fraction $1/[k^2 + \kappa^{2}(k,n) ]$ for
the chain contributions, namely
\begin{equation}
\label{RingSum}
\beta P_{\rm R} = \frac{1}{2}\sum_{n=-\infty}^\infty \int \frac{\d^3 \mathbf{k}}{(2\pi)^3} \left[
\frac{\kappa^2(k ,n)}{k^2} - \ln\left(1+ \frac{\kappa^2( k,n)}{k^2}\right) \right] \; .
\end{equation}
\bigskip
The summations for all the remaining diagrams are carried out as for the one-body density~\cite{Ballenegger2002}. They generate the same screened bonds and dressed activities (see Fig.~\ref{FigScreenedDiagrammatics}).
Besides the Debye bond \eqref{DebyeBond}, the so-called Abe-Meeron bond
\begin{equation}
\label{AMBond}
b_{\rm AM}(\mathcal{L}_i,\mathcal{L}_j) = \exp \( b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j) \) - 1 - b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)
\end{equation}
is generated by summing more complex structures connecting
the fixed pair $\mathcal{L}_i$ and $\mathcal{L}_j$ (see Fig.~\ref{figBAM}).
If $\mathcal{L}_i$ is an ending loop, a similar summation provides the Abe-Meeron ending bond
\begin{equation}
\label{AMEBond}
b_{\rm AME}(\mathcal{L}_i,\mathcal{L}_j) = \exp \( b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j) \) - 1 - b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j) - [b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)]^2/2 \; .
\end{equation}
\begin{figure}
\vskip 8mm
\includegraphics[scale=1.15]{SumB_AM_k2opt.pdf}
\caption{Examples of diagrams contributing to the bond $b_{\rm AM}(\mathcal{L}_i,\mathcal{L}_j)$. Since the truncated Mayer bond $b_{\rm T}$ can be interpreted as the sum of $n=2,3, ...,\infty$ direct interaction bonds $b_{\rm I}$ in parallel, the summed diagrams involve arbitrary number of links, either direct or via convolution chains of $b_{\rm I}$ bonds, between the two fixed loops $\mathcal{L}_i$ and $\mathcal{L}_j$.}
\label{figBAM}
\end{figure}
\begin{figure}
\includegraphics[scale=1.08]{ScreenedDiagrammatics2.pdf}
\caption{Bonds and weights in the screened Mayer expansions for the pressure and for loop distribution functions.
In the expansion of $\beta P$, the diagrams made with only one or two loops need a special treatment, see Eq.~\eqref{eq:ZZZ} and the comment after Eq.~\eqref{ruleVBsimplified}.
\label{FigScreenedDiagrammatics}
}
\end{figure}
\bigskip
The sum of all convolution rings involving $p \geq 1$ loops and $(p+1)$ bonds $b_{\rm I}$ attached to a loop $\mathcal{L}_i$ that is connected by more than two bonds $b_{\rm I}$ (or that is a root loop),
\begin{equation} \label{eq:27}
I_{\rm R}(\mathcal{L}_i)
=\;\,
\begin{graph}
\twoNodesUnNamed;
\colorNode{blanc}{1};
\drawLinkFCC{1}{2};
\end{graph}
\;\, + \;\,
\begin{graph}
\ringsAttachedMirror{3}{0.6}{1};
\colorNode{blanc}{3};
\end{graph}
\;\, + \;\,
\begin{graph}
\ringsAttached{4}{0.6}{1};
\colorNode{blanc}{2};
\end{graph}
\;\, + \;\,
\begin{graph}
\ringsAttachedMirror{5}{0.6}{1};
\colorNode{blanc}{5};
\end{graph}
\;\, + \;\,
\cdots
\end{equation}
provides the ring sum
\begin{equation}
\label{XIIIRingSum}
I_{\rm R}(\mathcal{L}_i) = \frac{1}{2}\left[ b_{\rm D}(\mathcal{L}_i,\mathcal{L}_i)-b_{\rm I}(\mathcal{L}_i,\mathcal{L}_i) \right] \;.
\end{equation}
Notice that the symmetry factor of each of these rings is $1/2$ because of the particular role of the attaching loop $\mathcal{L}_i$. The sum of $n\geq 1$ such rings attached to loop $\mathcal{L}_i$ generates the ring dressing factor $\left( \exp\left( I_{\rm R}(\mathcal{L}_i) \right)-1 \right)$ in the definition of the dressed activity
\begin{equation}
\label{XIIIRingActivityWeak}
z_{\rm D}(\mathcal{L}_i)=z(\mathcal{L}_i)\left( \exp\left( I_{\rm R}(\mathcal{L}_i) \right)-1 \right)\;.
\end{equation}
The ring dressing factor $\exp(I_{\rm R}(\mathcal{L}_i))$ accounts for the interaction energy of loop $\mathcal{L}_i$ with the surrounding polarization cloud of loops within a (non-linear) mean-field description.
\bigskip
The final screened Mayer series of the pressure reads
\begin{multline}
\label{ScreenedMayerP}
\beta P= \int D(\chi) z(\mathcal{L}) + \beta P_{\rm R} + \int D(\chi) z(\mathcal{L}) [e^{I_{\rm R}(\mathcal{L})}-1- I_{\rm R}(\mathcal{L})] \\
+ \sum_{{\cal P}} \frac{1}{S({\cal P})}
\int \left[ \prod D(\mathcal{L}) z^\ast(\mathcal{L}) \right]
\left[ \prod b^\ast \right]_{{\cal P}} \;
\end{multline}
where $b^*$ and $z^*$ are generic notations for the bonds and weights listed in Fig.~\ref{FigScreenedDiagrammatics} ($\chi=\{\alpha, q,\pmb{\mathcal{X}}\}$).
The diagram made with a single field point, which is treated separately, provides the three first terms in this formula: the ideal term, the ring pressure \eqref{RingSum} and the contribution of a single black loop to which are attached at least two rings, represented graphically by
\begin{equation} \label{eq:ZZZ}
\begin{graph}
\node[noir, thick] (1) {};
\end{graph}
\;\quad + \;\,
\begin{graph}
\node[noir, thick] (1) {};
\drawLoopNsnakeAngle{1}{90};
\end{graph}
\;\, + \;\,
\begin{graph}
\node[noir, thick] (1) {};
\drawGrayLoopSnakeAngle{1}{140}
\drawGrayLoopSnakeAngle{1}{40}
\end{graph}
\!\!\! + \,\ldots
\quad \equiv \quad
\begin{graph}
\node[noir, thick] (1) {};
\dressNodeSnakeGrayCircle{1};
\end{graph}
\end{equation}
The ring term $\beta P_{\rm R}$ reduces in the classical limit to the familiar Debye mean-field correction $\kappa^3/(12\pi)$,
while the diagrams with two rings or more in Eq.~\eqref{eq:ZZZ} accounts for corrections
beyond mean-field to the interaction energy of a loop with its surrounding polarization cloud~\cite{Ballenegger2017}.
We recall that the position of an arbitrarily chosen loop in each diagram ${\cal P}$ is kept fixed thanks to translational invariance.
For instance, in the contribution of the
graphs \eqref{eq:ZZZ}, it is understood that the integration $D(\mathcal{L})$ is carried out over all the internal degrees of
freedom of loop $\mathcal{L}$ except its position.
The sum in the second line is carried over all unlabelled topologically different prototype graphs ${\cal P}$ made with $N \geq 2$ black points.
These diagrams have the same topological structure
as the genuine Mayer diagrams. They are simply
connected and may contain articulation points.
Each point carries a statistical weight $z^\ast(\mathcal{L})$ which is either
\begin{equation} \label{2weights}
z(\mathcal{L})\text{\quad(bare loop)}\qquad \text{or}\qquad z_{\rm D}(\mathcal{L})
\text{\quad(dressed loop)}\;.
\end{equation}
When both weights are allowed for a point, the two possibilities can be added together to form the weight
\begin{equation}
\label{RingActivity}
z_{\rm R}(\mathcal{L}) = z(\mathcal{L}) + z_{\rm D}(\mathcal{L}) = z(\mathcal{L}) \exp(I_{\rm R}(\mathcal{L}))\;.
\end{equation}
There exists three possible bonds $b^* = b_{\rm D}$, $b_{\rm AM}$ and $b_{\rm AME}$. The bond $b_{\rm AME}$ can only be used to connect an ending bare field loop to the rest of the diagram (this rest can consist in a single bare or dressed loop in the particular case of a diagram made with two loops).
In general, the two weights~\eqref{2weights} are possible, except in the following cases~:
\begin{itemize}
\item If $\mathcal{L}$ is an ending field loop connected by a bond $b_{\rm AM}$ or $b_{\rm AME}$, its weight is
\begin{equation}
\label{ruleVBsimplified}
z^*(\mathcal{L}) =
\begin{cases}
z(\mathcal{L})& \text{if $b^* = b_{\rm AME}$} \\
z_{\rm D}(\mathcal{L}) & \text{if $b^* = b_{\rm AM}$} \\
\end{cases}
\end{equation}
\item If $\mathcal{L}$ is an intermediate field loop in a convolution $b_{\rm D}*b_{\rm D}$ of two Debye bonds, its weight is $z_{\rm D}(\mathcal{L})$.
\end{itemize}
Moreover, the case of a diagram made with only two loops is special, because both loops are then ending. The case $b^*=b_{\rm AME}$ in rule \eqref{ruleVBsimplified} is then modified to allow not only the case where both field loops are bare, but also the case where one loop is bare and one loop is dressed (see Fig.~6).
\bigskip
These diagrammatic ingredients and rules are summarized in Fig.~\ref{FigScreenedDiagrammatics}. These rules are valid not only for diagrams with $N \geq 2$ points in the screened Mayer series of the pressure, but also for all diagrams in the screened Mayer series of any loop distribution function $\rho^{(n)}(\mathcal{L}_1, ..., \mathcal{L}_n)$. In the latter case, each diagram contains $n$ root points and an arbitrary number of field points. Fig.~\ref{figDiagrams123} shows all diagrams in the screened Mayer series of the pressure made with 1, 2 or 3 loops.
\begin{figure}[t]
\includegraphics[scale=1.0]{Diagrams123.pdf}
\caption{All diagrams involving one, two or three loops in the screened Mayer series of the pressure.}
\label{figDiagrams123}
\end{figure}
\bigskip
The central quantity is the Debye bond
$b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)=-\beta e_{\alpha_i} e_{\alpha_j} \phi(\mathcal{L}_i,\mathcal{L}_j)$. As shown in Ref.~\cite{Ballenegger2002},
$\phi$ decays as $1/r^3$ at large distances $r$
between two loops. Thus bonds $b_{\rm AM}$ and $b_{\rm AME}$ decay respectively as $1/r^6$ and $1/r^9$, and they are integrable.
The bond $b_{\rm D}$ decays as $\phi$ itself, \textit{i.e.} as $1/r^3$, which is at the border line for integrability.
Accordingly, the graphs with ending loops connected to the rest of the diagram by bonds $b_{\rm D}$ have to be dealt with some care.
In fact, since the corresponding weight of the ending loop, $z_{\rm R}(\mathcal{L})$, is an even function of the loop shape $\pmb{\mathcal{X}}(s)$, if we proceed first to
functional integrations over the shape, then the $1/r^3$-algebraic tails vanish, because their amplitudes are odd functions of $\pmb{\mathcal{X}}(s)$ and
every prototype graph provides a finite contribution~\cite{Ballenegger2002}.
\subsection{Link with the activity-series for the particle densities}
\label{SectionLink}
The screened activity expansion of the loop density can be readily inferred from
the expansion~(\ref{ScreenedMayerP}) of the pressure by using
\begin{equation}
\label{FunctionalDerivativeP}
\rho(\mathcal{L}_{\rm a})=z(\mathcal{L}_{\rm a}) \frac{ \delta \beta P}{\delta z(\mathcal{L}_{\rm a})} \;.
\end{equation}
The activity-expansion of the particle densities follows then from
\begin{equation}
\label{ParticleDensityLoop}
\rho_{\alpha_{\rm a}} = \sum_{q_{\rm a}=1}^{\infty}q_{\rm a} \;\int D_{q}(\pmb{\mathcal{X}}_{\rm a}) \; \rho(\mathcal{L}_{\rm a}) \; .
\end{equation}
Notice that $P=k_{\rm B}T \ln{\Xi}/\Lambda$ is viewed in Eq.~\eqref{FunctionalDerivativeP} as a functional of the loop activity $z(\mathcal{L})$, which is present in Eq.~\eqref{IX.QMG6} and also in bonds and weights of the resummed diagrammatics. We consider here that the function $z(\mathcal{L})$ can also vary with the root position $\pmb{x}$ of the loop $\mathcal{L}$, as it does in an inhomogeneous system. The rules of functional derivatives generate then straightforwardly the expressions for the potentially space-dependent loop density $\rho(\mathcal{L})$.
When computing the particle density $\rho_\alpha$ in a homogeneous plasma, as in sections \ref{LDExpansionEOS} and \ref{DDR}, the functional derivative can be replaced by an ordinary partial derivative with respect to the one-dimensional variable $z_\alpha$.
\bigskip
The functional derivative of each prototype diagram ${\cal P}$ is calculated by either whitening a black loop $\mathcal{L}$
with weight $z(\mathcal{L})$ into the root loop $\mathcal{L}_{\rm a}$ with weight $z(\mathcal{L}_{\rm a})$ or by taking the functional derivative with
respect to $z(\mathcal{L}_{\rm a})$ of $\beta P_{\rm R}$ and $b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)$, namely
\begin{equation}
\label{FDRingP}
z(\mathcal{L}_{\rm a}) \frac{ \delta \beta P_{\rm R}}{\delta z(\mathcal{L}_{\rm a})}= z(\mathcal{L}_{\rm a}) I_{\rm R}(\mathcal{L}_{\rm a})
\end{equation}
and
\begin{equation}
\label{FDphi}
z(\mathcal{L}_{\rm a}) \frac{ \delta b_{\rm D}(\mathcal{L}_i,\mathcal{L}_j)}{\delta z(\mathcal{L}_{\rm a})} = z(\mathcal{L}_{\rm a})b_{\rm D}(\mathcal{L}_i,\mathcal{L}_{\rm a})b_{\rm D}(\mathcal{L}_{\rm a},\mathcal{L}_j) \; .
\end{equation}
Note that the functional derivatives of the dressed activities and of the other bonds, which can be all expressed in terms of
$b_{\rm D}$, are then obtained by using Eq.~(\ref{FDphi}).
In particular, the derivative of the ring factor $I_{\rm R}(\mathcal{L}_i)$ generates the bond $\frac{1}{2} b_{\rm D}^2(\mathcal{L}_a, \mathcal{L}_i)$.
This calculation provides
\begin{equation}
\label{IX.QMG88}
\rho(\mathcal{L}_{\rm a}) = \sum_{{\cal P}_{\rm a}} \frac{1}{S({\cal P}_{\rm a})}
\int \left[ \prod D(\mathcal{L}) z^\ast(\mathcal{L}) \right]
\left[ \prod b^\ast \right]_{{\cal P}_{\rm a}} \; .
\end{equation}
which can be also obtained by a direct Abe-Meeron summation of the Mayer diagrammatic series for the loop density~\cite{Ballenegger2002}. In obtaining Eq.~\eqref{IX.QMG88}, we have used that each bond $\frac{1}{2} b_{\rm D}^2(\mathcal{L}_a, \mathcal{L}_i)$ can be added to the bond $b_{\rm AME}(\mathcal{L}_a, \mathcal{L}_i)$ in diagrams with the same topological structure to provide the bond $b_{\rm AM}$.
\bigskip
The prototype diagrams ${\cal P}_{\rm a}$ have
one root (white) point with weight $z_{\rm R}(\mathcal{L}_a)$, $N=0, 1, 2, ...$ field (black) points and obey the diagrammatical rules summarized in Fig.~\ref{FigScreenedDiagrammatics}. The first few diagrams in the series~\eqref{IX.QMG88} are
\begin{equation}
\rho(\mathcal{L}) =
\begin{graph}
\node[blanc] (1) {};
\dressNodeDashedCircle{1};
\end{graph}
\;+\;
\begin{graph}
\twoNodesUnNamed;
\colorNode{blanc}{1};
\dressNodeDashedCircle{1};
\dressNodeDashedCircle{2};
\drawLinkD{1}{2};
\end{graph}
\;+\;
\begin{graph}
\twoNodesUnNamed;
\drawLinkAME{1}{2};
\colorNode{blanc}{1};
\dressNodeDashedCircle{1};
\end{graph}
\;+\;
\begin{graph}
\twoNodesUnNamed;
\drawLinkAM{1}{2};
\colorNode{blanc}{1};
\dressNodeDashedCircle{1};
\dressNodeCircle{2};
\end{graph}
\;+\; \ldots
\end{equation}
It can be checked that there are 16 topologically different diagrams made with 3 loops.
\subsection{Neutrality and low-density expansion of the EOS}
\label{Section3.D}
\subsubsection{Neutrality, pseudo-neutrality and Debye dressing}
\label{ChargeNeutralityDensity}
The collective electrostatic effects in a finite box
which enforce charge neutrality in the grand-canonical ensemble (see Section~\ref{sec:S2}), do not show in each individual term of the activity series,
where only a finite number of particles intervene. This is particularly striking for the ideal contribution
in series~(\ref{IX.QMG88}) for the particle density, whose Maxwell-Boltzmann
(weak-degeneracy) limit involves a single particle. The pseudo-neutrality condition,
$
\sum_\alpha e_\alpha z_\alpha=0
$,
which can be safely imposed as argued in
Section~\ref{sec:S2}, restores the previous collective electrostatic effects at this lowest order in the particle activities
Such effects might otherwise be erased in the series~(\ref{ScreenedMayerP}) and~\eqref{IX.QMG88} since the boundaries have been sent to infinity
without worrying about surface effects.
Importantly, the pseudo-neutrality condition ensures moreover that, at a given order in the small activities~$z$,
the expansions~(\ref{ScreenedMayerP}) and~(\ref{IX.QMG88}) for the pressure and the particle densities can be calculated
by keeping only a finite number of diagrams. We show first this point, and demonstrate then that local charge neutrality is always ensured due to the structure of these series.
\bigskip
Let us consider the series~(\ref{ScreenedMayerP}) for the pressure.
For any given graph ${\cal P}$, there exists a Debye dressed graph ${\cal P}^{\rm D}$ obtained by
adding a black loop $\mathcal{L}$ with weight $z(\mathcal{L})$ connected to ${\cal P}$ \textsl{via} a single bond $b_{\rm D}(\mathcal{L},\mathcal{L}')$
where $\mathcal{L}'$ is a black loop inside ${\cal P}$, that is
\begin{equation}
\label{diagramPD}
{\cal P}^{\rm D} =
\begin{graph}
\draw[black,fill=gray] (0,0cm) circle [radius=0.75 cm];
\node[transparent] (3) at (+0.2,0.45) [label=below:${\cal P}$] {};
\node[noir] (1) at (-0.35,0) [label={[xshift=1mm,yshift=0.75mm]below:${\cal L}'$}] {};
\node[noir] (2) at (-1.75,0) [label={[xshift=1mm,yshift=0.75mm]below:${\cal L}$}] {};
\drawLinkD{1}{2};
\end{graph}
\end{equation}
In the low-activity limit, the
potential $\phi$ reduces to its classical Debye counterpart~\cite{Ballenegger2002}, so
\begin{equation}
\label{DebyeBondLowA}
b_{\rm D}(\mathcal{L},\mathcal{L}') \sim -\beta q_\alpha e_\alpha q_{\alpha'}e_{\alpha'} \frac{\exp(-\kappa_z |\mathbf{x}-\mathbf{x}'| )}{|\mathbf{x}-\mathbf{x}'|}
\end{equation}
and where we have used
\begin{equation}
\label{KappaLowA}
\kappa^2(0,0) \sim \kappa_z^2= 4 \pi \beta \sum_\gamma e_\gamma^2 z_\gamma \; .
\end{equation}
At leading order in the small activities, the contribution of the graph ${\cal P}^{\rm D}$ is obtained by keeping only the loop
$\mathcal{L}$ made with a single particle, i.e. $q_\alpha=1$, while the bond $b_{\rm D}(\mathcal{L},\mathcal{L}')$ is replaced by its classical Debye
expression~(\ref{DebyeBondLowA}). The leading contribution of ${\cal P}^{\rm D}$ reduces hence to that of graph ${\cal P}$ multiplied
by
\begin{multline}
\label{DebyeBondLowAbis}
\int D(\mathcal{L}) z(\mathcal{L}) \; b_{\rm D}(\mathcal{L},\mathcal{L}') \sim -\beta q_{\alpha'}e_{\alpha'} \sum_\alpha e_\alpha z_\alpha
\int \d \mathbf{x} \frac{\exp(-\kappa_z |\mathbf{x}-\mathbf{x}'| )}{|\mathbf{x}-\mathbf{x}'|} \\
=-\frac{4 \pi \beta q_{\alpha'}e_{\alpha'}}{\kappa_z^2} \sum_\alpha e_\alpha z_\alpha \; .
\end{multline}
At leading order, the contribution of ${\cal P}^{\rm D}$ has obviously the same order as
that of ${\cal P}$ for arbitrary sets $\{ z_\alpha\}$ of particle activities. In other words, in order to compute the pressure at
a given order for such sets, one would have to keep an infinite number of graphs in the series~(\ref{ScreenedMayerP}), since the dressing
of a given ${\cal P}$ can be repeated an arbitrary number of times. This infinite sum might actually not converge, meaning
that the screened Mayer series~\eqref{ScreenedMayerP} and~\eqref{IX.QMG88} in an unbounded volume might make sense only
when the pseudo-neutrality condition is imposed. The pseudo-neutrality condition~(\ref{PseudoNeutrality}) greatly simplifies the calculations at a given order. Indeed, the graph ${\cal P}^{\rm D}$ contributes then at a higher order than graph ${\cal P}$. Only a finite number of graphs ${\cal P}$ in the series~(\ref{ScreenedMayerP}) and \eqref{IX.QMG88} need then to be kept.
\bigskip
The property of a quantum plasma to be locally charge neutral at equilibrium
can be proved by combining Eq.~\eqref{ParticleDensityLoop} for the particle density with the Mayer series \eqref{IX.QMG88} for the density of loops. This proof is based on a simple Debye-dressing mechanism at work in the resulting series for the particle densities.
\begin{figure}[h]
\includegraphics[scale=1.15]{Dressing_density.pdf}
\caption{
A diagram ${\cal P}_{\rm a} \in {\cal C}_{\rm a}$, and its Debye-dressed companion diagram ${\cal P}_{\rm a}^{\rm D}$. In diagram ${\cal P}_{\rm a}$, loop ${\cal L}_{\rm a}$ cannot be a bare loop connected only by a bond $b_{\rm D}$, whereas ${\cal L}_{\rm a}$ is precisely such a loop in diagram ${\cal P}^{\rm D}_{\rm a}$.
}
\label{figDebyeDressingDensity}
\end{figure}
Remembering that each loop in a prototype diagram is either bare or dressed, we classify the diagrams into two groups: the class ${\cal C}^{\rm D}_{\rm a}$ of diagrams where the root loop $\mathcal{L}_a$ is an ending bare loop connected only by a Debye bond $b_{\rm D}$, and the class ${\cal C}_{\rm a}$ containing all other diagrams.
For any diagram ${\cal P}_{\rm a}$ in the class ${\cal C}_{\rm a}$, including the most simple diagram made with only one single bare loop, there exists a unique corresponding diagram ${\cal P}^{\rm D}_{\rm a}$ in the class ${\cal C}^{\rm D}_{\rm a}$ where $\mathcal{L}_{\rm a}$ is a bare loop connected by a (single) bond $b_{\rm D}$ to a subdiagram identical to ${\cal P}_{\rm a}$ but where the root point is replaced by a field point, which we label $\mathcal{L}_1$
(see Fig.~\ref{figDebyeDressingDensity}). This establishes a one-to-one correspondence between the diagrams in the two classes because no convolution
$b_{\rm D}*b_{\rm D}$ with an intermediate bare field loop is allowed in the prototype diagrams. The diagram ${\cal P}^{\rm D}_{{\rm a}}$ is said to be the Debye-dressed companion of diagram ${\cal P}_{\rm a}$.
The contribution of diagram ${\cal P}^{\rm D}_{\rm a}$ to the density $\rho_{\alpha_{\rm a}}$,
\begin{equation}
\label{GraphDDensityBareBis}
\rho_{\alpha_{\rm a}}[{\cal P}^{\rm D}_{\rm a}]=- \frac{4 \pi \beta e_{\alpha_{\rm a}}}{\kappa^2(0,0) }
\sum_{q_{\rm a}=1}^\infty q_{\rm a}^2 \int D_{q_{\rm a}}(\pmb{\mathcal{X}}_{\rm a}(\cdot)) z(\mathcal{L}_{\rm a})
\sum_{\alpha_{1}} e_{\alpha_{1}} \rho_{\alpha_{1}}[{\cal P}_{1}] \; ,
\end{equation}
is easily calculated with the frequency decomposition~(\ref{IX.QMG48})
of $\phi(\mathcal{L}_{\rm a},\mathcal{L}_1)$ and translation invariance.
Notice that this contribution has the same order, when $z\to 0$, as the one of diagram $\rho_{\alpha_{\rm a}}[{\cal P}_{\rm a}]$ because $z(\mathcal{L}) \propto z$ and $\kappa^2(0,0)\propto z$.
Using expression~(\ref{IX.QMG45a}) for $\kappa^2(0,0)$, this contribution to the local charge density reduces to
\begin{equation}
\label{GraphDChargeDensityBare}
\sum_{\alpha_{\rm a}} e_{\alpha_{\rm a}}\rho_{\alpha_{\rm a}}[{\cal P}^{\rm D}_{\rm a}]=
- \sum_{\alpha_{\rm a}} e_{\alpha_{\rm a}}\rho_{\alpha_{\rm a}}[{\cal P}_{\rm a}]\; .
\end{equation}
The contribution to the local charge density of any diagram ${\cal P}_{\rm a}$ is thus exactly compensated by the contribution of its
companion diagram ${\cal P}^{\rm D}_{\rm a}$. The local charge density therefore vanishes. Notice that this proof does not require the pseudo-neutrality condition to be satisfied. If pseudo-neutrality does not hold, there is an infinite number of diagrams contributing at the same order, rendering the proof only formal, whereas there is only a finite number of diagrams contributing at a given order when the pseudo-neutrality condition holds.
\subsubsection{Expansion of the EOS at order $\rho^2$}
\label{LDExpansionEOS}
The low-density expansion of the EOS has been computed up to order $\rho^{5/2}$
by various methods~\cite{Ebeling1967,Kraeft1986,Alastuey1992,deWitt1995,Brown2001},
which all provide eventually identical physical predictions~\cite{Alastuey2015}. Our purpose in this section is to illustrate the efficiency of the method
based on the screened activity expansion~\eqref{ScreenedMayerP} of the pressure by outlining how all terms up to order $\rho^2$ in the EOS can be computed.
\bigskip
\mbox{}
\vskip -6mm
Since at low densities, $z \sim \rho$, in order to obtain the EOS at order $\rho^2$, we need to start with the $z$-expansion of $\beta P$ at the order
$z^2$. We assume that the pseudo-neutrality condition holds. Then, at this order, one only needs to consider the diagrams made with 1 or 2 loops, i.e. the first five diagrams in Fig.~\ref{figDiagrams123}. In the first diagram made with 2 loops in this figure, we can discard the cases where one or both loops are bare because their contributions are $o(z^2)$ thanks to pseudo-neutrality.
Since $z_{\rm D} \propto z^{3/2}$, the next diagram and the last one made with two loops are of order $z^{5/2}$ and can hence also be discarded.
Only three diagrams remain,
\begin{equation}
\label{P2}
\beta P \quad=\quad
\begin{graph}
\node[noir, thick] (1) {};
\dressNodeSnakeGrayCircle{1};
\end{graph}
\quad+\quad
\begin{graph}
\twoNodesUnNamed;
\drawLinkD{1}{2};
\dressNodeCircle{1};
\dressNodeCircle{2};
\end{graph}
\quad+\quad
\begin{graph}
\twoNodesUnNamed;
\drawLinkAME{1}{2};
\end{graph} \quad+\; o(z^2)\;.
\end{equation}
\noindent
Significantly more diagrams contribute to the density at the same order,
\begin{align} \notag
\rho_\alpha \quad=\quad
\begin{graph}
\node[blanc] (1) {};
\dressNodeDashedCircle{1};
\end{graph}
\;&+\;
\begin{graph}
\twoNodesUnNamed;
\colorNode{blanc}{1};
\dressNodeCircle{1};
\dressNodeCircle{2};
\drawLinkD{1}{2};
\end{graph}
\;+\;
\begin{graph}
\twoNodesUnNamed;
\drawLinkAME{1}{2};
\colorNode{blanc}{1};
\end{graph}
\;+\;
\begin{graph}
\twoNodesUnNamed;
\drawLinkAM{1}{2};
\colorNode{blanc}{1};
\dressNodeCircle{2};
\end{graph} \\ \notag
&+\;
\begin{graph}
\node[blanc] (1) {};
\node[noir] (2) at($(1) + (0:1) $){};
\node[noir] (3) at($(1) + (60:1) $){};
\drawLinkD{1}{2};
\drawLinkD{1}{3};
\colorNode{blanc}{1};
\dressNodeCircle{2};
\dressNodeCircle{3};
\end{graph}
\;+\;
\begin{graph}
\node[blanc] (1) {};
\node[noir] (2) at($(1) + (0:1) $){};
\node[noir] (3) at($(1) + (60:1) $){};
\drawLinkAM{1}{2};
\drawLinkD{2}{3};
\colorNode{blanc}{1};
\dressNodeCircle{3};
\end{graph}
\;+\;
\begin{graph}
\node[blanc] (1) {};
\node[noir] (2) at($(1) + (0:1) $){};
\node[noir] (3) at($(1) + (60:1) $){};
\drawLinkD{1}{2};
\drawLinkD{1}{3};
\drawLinkAM{2}{3}
\colorNode{blanc}{1};
\end{graph} \\
&+\,\text{7 Debye-dressed companion diagrams}
\quad+\;o(z^2).
\label{DensityPartialDerivativeP}
\end{align}
Since Eq.~\eqref{DensityPartialDerivativeP} includes all companion DD diagrams, it leads to particle densities that satisfy the charge neutrality
$\sum_\alpha e_\alpha \rho_\alpha = 0$, as shown in the previous section.
In the last four drawn diagrams, the dressed weight $z_{\rm D}$ of the root point has been discarded into the $o(z^2)$ remainder.
The seven DD companion diagrams in Eq.~\eqref{DensityPartialDerivativeP} are
\begin{multline}
\label{classicalDD}
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\dressNodeDashedCircle{2};
\drawLinkDclas{1}{2}
\end{graph}
\;\;+\;\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] (3) at($(1) + (60:1) $){};
\dressNodeCircle{2};
\dressNodeCircle{3};
\drawLinkDclas{1}{2}
\drawLinkD{2}{3};
\end{graph}
\;\;+\;\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] (3) at($(1) + (60:1) $){};
\drawLinkDclas{1}{2}
\drawLinkAME{2}{3};
\end{graph}
\;\;\;\;\\
+\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] (3) at($(1) + (60:1) $){};
\drawLinkDclas{1}{2}
\drawLinkAM{2}{3};
\dressNodeCircle{3};
\dressNodeCircle{2};
\end{graph}
\;\;+\;\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] [above of=1,yshift=-1mm] (3) {};
\node[noir] [above of=2,yshift=-1mm] (4) {};
\drawLinkDclas{1}{2}
\drawLinkD{3}{4};
\drawLinkAM{2}{4};
\dressNodeCircle{3};
\dressNodeCircle{4};
\end{graph}
\;\;+\;\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] [above of=1,yshift=-1mm] (3) {};
\node[noir] [above of=2,yshift=-1mm] (4) {};
\drawLinkDclas{1}{2}
\drawLinkD{2}{3};
\drawLinkD{2}{4};
\dressNodeCircle{3};
\dressNodeCircle{4};
\end{graph}
\;\;+\;\;
\begin{graph}
\node[blanc,scale=0.7,semithick] (1) {} ;
\node[noir] [right of=1] (2) {};
\node[noir] [above of=1,yshift=-1mm] (3) {};
\node[noir] [above of=2,yshift=-1mm] (4) {};
\drawLinkDclas{1}{2}
\drawLinkD{2}{3}
\drawLinkD{2}{4}
\drawLinkD{3}{4};
\drawLinkAM{3}{4};
\end{graph}
\end{multline}
At the considered order $O(z^2)$, the root point is a single particule of species $\alpha$, i.e. a loop with $q=1$, with weight $z_{\alpha}$, and the wavy and dotted bond represents the classical Debye screened bond given by Eq.~\eqref{IX.QMG47} where $\psi_{\rm loop}(\pmb{r},s)$ is replaced by $\exp(-\kappa_z r)/r$ with
$r = |\pmb{r}|$ while $\kappa^2_z$ is defined in Eq.~\eqref{KappaLowA}. The contribution of a diagram ${\cal P}^{\rm D}$ companion of ${\cal P}$
to density $\rho_{\alpha}$, is given by the general
formula~(\ref{GraphDDensityBareBis}) with $\alpha_{\rm a}=\alpha$ , which then reduces to
\begin{equation}
\label{CompanionClassical}
\rho_{\alpha}[{\cal P}^{\rm D}]=- \frac{4 \pi \beta e_{\alpha} z_{\alpha}}{\kappa_z^2 }
\sum_{\gamma} e_{\gamma} \rho_{\gamma}[{\cal P}]
\end{equation}
discarding terms of order $o(z^2)$. Since the contributions $\rho_{\gamma}[{\cal P}]$ of the 7 diagrams~(\ref{DensityPartialDerivativeP}) are obtained by
merely taking the partial derivative $ z_\alpha \partial(\beta P)/\partial z_\alpha$ of the retained pressure diagrams
at the same order $O(z^2)$, we see that the contributions $\rho_{\alpha}[{\cal P}^{\rm D}]$ of their companions at the same order $O(z^2)$
are also fully determined by the pressure diagrams~(\ref{P2}).
\bigskip
After computing the pressure at order $z^2$, denoted $P^{(2)}$ and the density at the same order,
denoted $\rho_{\alpha}^{(2)}(\{z_\gamma \})$, one determines the equation of state $P(T,\{\rho_\gamma\})$.
The ${\cal S}$ pseudo-neutral activities $\{z_\gamma\}$ up to order $\rho^2$ included, denoted $\{z_\gamma^{(2)}\}$, are
first obtained as function of the physical
densities $\{\rho_\alpha\}$ by inverting perturbatively the ${\cal S}$
independent relations
\begin{align}
\label{DensityActivity}
&\rho_{\alpha}^{(2)}(\{z_\gamma^{(2)} \}) = \rho_{\alpha},
\quad \quad \alpha_{\rm a}=1,...,{\cal S}-1 \nonumber \\
& \sum_\gamma e_\gamma z_\gamma^{(2)}= 0 \; .
\end{align}
The required EOS up at order $\rho^2$ included follows from inserting $\{z_\gamma^{(2)}\}$
into Eq.~\eqref{P2}, namely
\begin{equation}
\label{EOS2}
\beta P(\beta;\{\rho_\alpha \})=P^{(2)}(\beta;\{z_\gamma^{(2)}(\{\rho_\alpha \}) \}) + o(\rho^2) \; .
\end{equation}
It can be checked that the known density expansion of the pressure is indeed recovered.
\bigskip
This calculation illustrates the usefulness of the activity series for the pressure which considerably reduces the number of diagrams
which need to be computed. A similar scheme can be repeated at the next orders.
However, even if the number of diagrams which need to be computed is reduced with respect to other methods, it remains a formidable task to
obtain the terms of order $\rho^3$. In particular, one has to take into account
quantum effects embedded in $\phi$ and $I_{\rm R}$, so a classical treatment of the dressing mechanism is no longer sufficient.
\subsubsection{Computation of the densities by differentiation of the pressure with Debye-dressed activities}
\label{sectionDDActivities}
It is instructive to interpret the previous results in terms of Debye-Dressed activities.
Firstly, we note that the $7$ companion-diagrams~(\ref{classicalDD}) in the density series arise from other diagrams in the pressure series
which do not contribute at order $z^2$ by virtue of the pseudo-neutrality condition. In fact, the pseudo-neutrality condition must be applied only after the derivative $ z_\alpha \partial(\beta P)/\partial z_\alpha$ has been taken, whereas it has already been applied in an expression like~\eqref{P2}. Since a contribution to the pressure that vanishes by pseudo-neutrality can have a non-vanishing derivative with respect to $z$, and
hence a non-vanishing contribution to the density $\rho_\alpha$, one needs to consider also such contributions in the pressure series.
These contributions can be seen as decorations of the diagrams that do not increase their order. The Debye-Dressing (DD) of a loop in a diagram is an example of such a contribution (recall Eq.~\eqref{diagramPD}).
Adding DD decorations successively to each points in the three diagrams of Eq.~\eqref{P2} provides a set of diagrams which generate, after differentiation,
the corresponding density diagrams in Eq.~\eqref{DensityPartialDerivativeP} except for the last four companion diagrams in Eq.~\eqref{classicalDD}.
In fact, these four diagrams arise from other diagrams in the pressure series which are nothing but the same diagrams where the root white point is transformed
into a black point. Nevertheless, we will show that all the 14 diagrams can be computed by direct partial differentiation of only the 3 pressure diagrams~(\ref{P2}) when Debye-dressed activities are used.
\bigskip
Let us define the Debye-Dressed activity
\begin{equation}
\label{DDfunction}
z_{\alpha}^{\rm DD} =z_\alpha \left( 1 -\frac{4 \pi \beta e_\alpha }{\kappa_z^2} \sum_\gamma e_\gamma z_{\gamma} \right) \; ,
\end{equation}
where the second term in Eq.~\eqref{DDfunction} is the classical DD factor~\eqref{DebyeBondLowAbis}.
Let us replace, in the 3 pressure diagrams~(\ref{P2}),
the weights $z_\alpha$ of the black points by $z_{\alpha}^{\rm DD}$.
This amounts to decorate the considered diagrams and it provides after differentiation
the first 10 diagrams in Eq.~\eqref{DensityPartialDerivativeP}. Since the remaining four companion diagrams in Eq.~\eqref{classicalDD} are
associated with diagrams obtained by taking the derivative of the ring factor $I_{\rm R}$ and of the bond $b_{\rm D}$, we see that
if we also replace $z_\alpha$ by $z_{\alpha}^{\rm DD}$ in both $I_{\rm R}$ and $b_{\rm D}$ in the 3 pressure diagrams~(\ref{P2}), the standard rules of partial derivatives of composite functions generate the last four companion diagrams in Eq.~\eqref{classicalDD}, thanks
to the expression~(\ref{PartialDerivativeDDbis}) of the partial derivative $\partial z_{\theta}^{\rm DD}/\partial z_{\alpha}$. Again, and
as mentioned above, this partial derivative
has to be calculated for any set of independent activities, while the pseudo-neutral condition is applied afterward.
Hence, if we replace all the activities $z_\alpha$ by the functions $z_{\alpha}^{\rm DD}$ in the weights and bonds of the 3 pressure
diagrams~(\ref{P2}), the corresponding function $P^{\rm DD}$ is such that the derivative $z_\alpha \partial(\beta P^{\rm DD})/\partial z_\alpha$ generates automatically the values of all the 14 diagrams in Eq.~\eqref{DensityPartialDerivativeP} at order $z^2$ included.
\section{Debye-Dressing neutralization prescription}
\label{DDR}
Let us consider a given approximation $P_{\rm A}(\beta;\{ z_i \})$ obtained by selecting specific diagrams in
the series~(\ref{ScreenedMayerP}). As discussed in Section~\ref{sec:S23} for any approximation, the densities inferred from
$P_{\rm A}(\beta;\{ z_i \})$ \textsl{via} the standard identities~(\ref{DensityPressureA}) do not necessarily satisfy the local charge neutrality.
We introduced a general prescription, based on the Neutral-Group activities, which systematically circumvents this drawback. Here,
we propose a different, but closely related, general method inspired by the Debye-Dressing mechanism described and applied to the first terms of
the pressure and density series up to order $z^2$.
\subsection{Debye-Dressing neutralization prescription}
\label{Sec:S41}
The Debye-dressed diagrams (see Fig.~\ref{figDebyeDressingDensity}) in the series for the particle densities are crucial for ensuring the local charge neutrality. If a given diagram contributes to the particle densities in a way that breaks the local charge neutrality, adding the contribution of its DD companion diagram is sufficient to restore electro-neutrality. Inspired by this simple mechanism, one can define the following heuristic Debye-Dressing neutralization prescription
\begin{equation}
\label{DressingRecipe}
\rho_{\alpha} = z_{\alpha} \frac{\partial P_{\rm A}}{\partial z_{\alpha} }
-\frac{4 \pi \beta e_\alpha z_{\alpha}}{\kappa_z^2} \sum_\gamma e_\gamma z_{\gamma} \frac{\partial P_{\rm A} }{\partial z_{\gamma} } \; .
\end{equation}
The two terms in this equation are the analogs of the two diagrams in Fig.~\ref{figDebyeDressingDensity}. The classical expression for the Debye-dressing factor is used in Eq.~\eqref{DressingRecipe}, as in Section~(\ref{Section3.D}) for exact calculations at order $z^2$, because it is sufficient to ensure electroneutrality.
We stress that the partial derivatives in the dressed expression~(\ref{DressingRecipe}) are calculated as usual, namely for independent activities $z_\alpha$.
However, at the end, their values are determined for a set $\{ z_{\alpha} \} $ satisfying the pseudo-neutrality
condition~(\ref{PseudoNeutrality}).
The local charge neutrality is automatically satisfied by the dressed densities~(\ref{DressingRecipe}).
If the undressed densities
\begin{equation}
\label{UndressedD}
z_{\alpha} \frac{\partial P_{\rm A}}{\partial z_{\alpha} }
\end{equation}
carry a non-zero net charge $q_{\rm exc}$, the dressed density $\rho_\alpha$~(\ref{DressingRecipe}) is shifted from its undressed counterpart
by a term proportional to $e_\alpha z_\alpha q_{\rm exc}$. As it should, this shift vanishes if $q_{\rm exc}=0$,
namely if the undressed densities~(\ref{UndressedD})
already satisfy local charge neutrality.
\bigskip
In Section~\ref{sectionDDActivities}, Debye-dressed activities $\{z^{\rm DD}_\alpha\}$ have been introduced to take into account systematically, at any order in the particle activities, the classical Debye screening effect when determining the particle densities associated with some diagrams
in the pressure series~\eqref{ScreenedMayerP}. The DD activities can therefore also be used to determine, from an
approximate expression~$P_{\rm A}(\beta;\{ z_i \})$ for the pressure, particle densities that satisfy electroneutrality, and also a grand-potential
from which these densities derive.
Let us show that this way of ensuring electroneutrality, which is exact at order~$z^2$, leads to the same particle
densities as the prescription~\eqref{DressingRecipe}.
In that approach, to any approximation $P_{\rm A}(\beta;\{ z_i \})$ for the pressure, we introduce the associated approximation
\begin{equation}
\label{DDpressure}
P_{\rm A}^{\rm DD}(\beta;\{ z_\alpha \})=P_{\rm A}(\beta;\{ z_\alpha^{\rm DD} \})
\end{equation}
where each $z_\alpha$ in $P_{\rm A}$ is replaced by the Debye-Dressed function~\eqref{DDfunction} of the activities. The particle densities inferred
from $P_{\rm A}^{\rm DD}$, namely
\begin{equation}
\label{DensityPressureADD}
\rho_{\alpha} = z_{\alpha} \frac{\partial P_{\rm A}^{\rm DD}}{\partial z_{\alpha}}(T;\{ z_{\gamma} \}) \; ,
\end{equation}
can be calculated by applying the rules of composition of partial derivatives,
\begin{equation}
\label{DensityPressureADDbis}
\rho_{\alpha} = z_{\alpha} \frac{\partial }{\partial z_{\alpha}} P_{\rm A}(T;\{ z_\gamma^{\rm DD} \}) \nonumber \\
= \sum_{\theta=1}^{\mathcal{S}} \frac{\partial P_{\rm A}}{\partial z_{\theta}} (T;\{z_\gamma^{\rm DD} \})
z_{\alpha} \frac{\partial z_{\theta}^{\rm DD}}{\partial z_{\alpha}}\; .
\end{equation}
The partial derivatives $\partial z_{\theta}^{\rm DD}/\partial z_{\alpha}$ calculated by using expressions~(\ref{DDfunction}) are
\begin{equation}
\label{PartialDerivativeDD}
z_{\alpha}\frac{\partial z_{\theta}^{\rm DD}}{\partial z_{\alpha}} = z_\alpha^{DD}\delta_{\alpha,\theta}
- \frac{4 \pi \beta e_\alpha e_\theta z_{\alpha} z_\theta}{\kappa_z^2}
-\frac{(4 \pi \beta)^2 e_\alpha^2 e_\theta z_{\alpha} z_\theta}{\kappa_z^4} \sum_\gamma e_\gamma z_{\gamma}
\end{equation}
where $\delta_{\alpha,\theta}$ is the Kronecker symbol.
For any set $\{ z_{\alpha} \} $ satisfying the pseudo-neutrality
condition~(\ref{PseudoNeutrality}), these expressions become
\begin{equation}
\label{PartialDerivativeDDbis}
z_{\alpha}\frac{\partial z_{\theta}^{\rm DD}}{\partial z_{\alpha}} = z_\alpha \delta_{\alpha,\theta} - \frac{4 \pi \beta e_\alpha e_\theta z_{\alpha} z_\theta}{\kappa_z^2} \; ,
\end{equation}
while the Debye-Dressed functions $z_\alpha^{\rm DD}$ reduce to $z_\alpha$. Inserting these results into Eq.~(\ref{DensityPressureADDbis}),
we exactly recover the expressions~(\ref{DressingRecipe}) for the dressed densities.
\subsection{Comparison with other prescriptions ensuring electroneutrality}
\label{Sec:S42}
In order to compare the Debye-Dressing neutralization prescription~\eqref{DressingRecipe} with that of the Neutral Groups, it is useful to rewrite Eqn.~(\ref{DressingRecipe}) in a way similar to expressions~(\ref{DensityPressureANGter}), namely
\begin{align}
\label{DensityPressureADDter}
&\rho_i = \sum_{\delta=1}^{\mathcal{S}} D_{i,\delta} \; z_\delta \; \frac{\partial P_{\rm A}}{\partial z_{\delta}} (T;\{z_\gamma \})
\quad \text{for} \quad i=1,...,\mathcal{S}-1 \nonumber \\
&\rho_{\rm e}= \sum_{j=1}^{\mathcal{S}-1} Z_j \rho_j \;
\end{align}
with coefficients
\begin{equation}
\label{CoefficientsNGDensityB}
D_{i,j}=
\begin{cases}
\displaystyle
1-\frac{Z_i^2 z_i}{z_{\rm e} + \sum_{l=1}^{\mathcal{S}-1}Z_l^2 z_l} &\text{if $j=i$} \\
\displaystyle\rule{0mm}{8mm}
-\frac{Z_i Z_j z_i}{ z_{\rm e} + \sum_{l=1}^{\mathcal{S}-1}Z_l^2 z_l} &\text{for } j=1,...,\mathcal{S}-1 \; , \; j \neq i \\
\displaystyle\rule{0mm}{8mm}
\frac{Z_i z_i}{ z_{\rm e} + \sum_{l=1}^{\mathcal{S}-1}Z_l^2 z_l} &\text{if $j=\mathcal{S}$ [i.e. $j={\rm e}$]}
\end{cases}.
\end{equation}
For two-component systems, like the hydrogen or the helium plasmas for instance, it turns out that
$D_{1,1}=C_{1,1}=1/(Z+1)$ and $D_{1,{\rm e}}=C_{1,{\rm e}}=Z/(Z+1)$, so both recipes are equivalent. For
systems with three or more components, like the hydrogen-helium mixture, these methods are no longer
equivalent, at least for the choice~(\ref{RelevantActivity}) of the Neutral-Group activities. Nevertheless it is worthy to note that both
prescriptions become equivalent if the approximate pressure $P_{\rm A}$ is consistent with local charge neutrality, i.e.
if the undressed densities~(\ref{UndressedD}) do not carry a net charge $q_{\rm exc}=0$. {Of course, they become
exact for an exact expression of the pressure. }
\bigskip
Let us mention that yet another neutralization prescription has been used in the literature~\cite{Starostin2005}, which we call the Enforced-Neutrality prescription. Contrarily to the previous
Neutral-Group or Debye-Dressed procedures, it does not rely on a general transformation valid for any approximate
pressure $P_{\rm A}$. For a given set of nuclei activities $\{ z_i; i=1,...,\mathcal{S}-1 \}$, it consists in choosing the
electron activity $z_{\rm e}$ in such a way that the local charge neutrality for the densities directly calculated within
the standard formulae is indeed observed. The particular value of $z_{\rm e}$ if found by solving a non-linear equation that is specific to considered model. Notice that this prescription disregards the pseudoneutrality condition~\eqref{PseudoNeutrality}.
\bigskip
Eventually, let us illustrate the various neutralization methods for a two-component system in the case of the following simple approximation for the pressure
\begin{equation}
\label{ApproximationPring}
\beta P_{\rm A}(\beta;z_1,z_{\rm e}) = z_1 + z_{\rm e} + \frac{\kappa_z^{3}}{12 \pi}
\end{equation}
with $\kappa_z=[4\pi \beta e^2 (Z^2 z_1 + z_{\rm e} )]^{1/2}$. The first two terms are nothing
but the ideal Maxwell-Botzmann contributions, while the last
term is the classical mean-field (or ring) contribution. The Neutral-Group and Debye-Dressed methods
provide the same densities
\begin{equation}
\label{DensityRingTCP}
\rho_{\rm n} = z \left[ 1 + \beta e^2 Z \kappa_z/2 \right] \quad \text{and} \quad \rho_{\rm e}= Z \rho_{\rm n}
\end{equation}
where the subscript `n' refers to nuclei ($z_1 = z_{\rm n} = z, z_{\rm e}=Zz$).
These expressions also coincide with the exact small-activity expansion of $\rho_{\rm n}$ and $\rho_{\rm e}$
up to order $z^{3/2}$ included, which can be calculated within the diagrammatic series~(\ref{IX.QMG88}). Hence, the
approximate EOS associated with~(\ref{ApproximationPring}) are identical in both methods, namely
$ P_{\rm A}^{\rm NG}(\beta;\rho_{\rm n},\rho_{\rm e})= P_{\rm A}^{\rm DD}(\beta;\rho_{\rm n},\rho_{\rm e})$.
\bigskip
Within the Enforced-Neutrality procedure, since
\begin{equation}
\label{ApproximationPringDerivative}
z_1 \frac{\partial \beta P_{\rm A}}{\partial z_1} = z_1 + \beta e^2 Z^2 \kappa_z/2 \quad \text{and}
\quad z_{\rm e} \frac{\partial \beta P_{\rm A}}{\partial z_{\rm e}} = z_{\rm e} + \beta e^2 \kappa_z/2 \; ,
\end{equation}
if we set $z_1=z$, the electron activity $z_{\rm e}^{\rm EN}$ is such that
\begin{equation}
\label{ENactivity}
Z \left[z + \beta e^2 Z^2 \kappa_z/2 \right] = z_{\rm e}^{\rm EN} + \beta e^2 \kappa_z/2 \quad \text{with}
\quad \kappa_z=\left[4\pi\beta e^2(Z^2 z +z_{\rm e}^{\rm EN}) \right]^{1/2} \; ,
\end{equation}
which can be recast as a cubic polynomial equation for $z_{\rm e}^{\rm EN}$. For $Z \neq 1$,
$z_{\rm e}^{\rm EN}$ is different from the electron activity $Zz$ satisfying the pseudo-neutrality condition,
and the resulting EOS $\beta P_{\rm A}^{\rm EN}(\beta;\rho_{\rm n},\rho_{\rm e})$ is different from the previous EOS
$P_{\rm A}^{\rm NG}(\beta;\rho_{\rm n},\rho_{\rm e})=P_{\rm A}^{\rm DD}(\beta;\rho_{\rm n},\rho_{\rm e})$.
If $Z=1$, \textsl{i.e.} for the hydrogen plasma, $z_{\rm e}^{\rm EN}=z$ so the Enforced-Neutrality procedure is equivalent to the
previous methods for the specific model~\eqref{ApproximationPring}. However, as soon as quantum corrections are added to the model, this equivalence no longer holds, even in the hydrogen plasma, because
quantum effects involve particle masses $m_{\rm e}$ and $m_{\rm p}$ which are not identical.
\section{Derivation of approximate equations of state}
\label{Sec:S5}
The screened activity expansion for the pressure appears to be quite useful for
constructing approximate expressions $P_{\rm A}(\beta;\{ z_i \})$ at moderate densities.
In such regimes, recombination processes into chemical species made with three or more particles
become important. The contributions of the relevant chemical species, including their interactions, are
included in cluster functions. In a first step, the graphs which are expected to provide
the main contributions are selected on the basis of physical arguments. In a second step, their
contributions can be numerically computed by using simplified versions of $\phi$~\cite{Ballenegger2017}, while the
functional integrations over loop shapes require the introduction
of suitable quantum Monte Carlo techniques~\cite{Wendland2014}.
\subsection{Cluster functions associated with a given number of particles}
\label{ClusterFunctions}
The contributions of familiar chemical species can be easily identified in terms of specific diagrams
in the screened activity expansion~(\ref{ScreenedMayerP}) of the pressure, following the method first
introduced for the particle densities~\cite{Alastuey2003}. It consists in rewriting the phase space measure of each loop $\mathcal{L}$, as a sum over
the number $q$ of elementary particles (nuclei or electrons) which are contained in $\mathcal{L}$. Each graph ${\cal P}$ then generates an infinite number
of graphs ${\cal P}[N_1,...,N_{\cal S}]$ with the same topological structure. Each $N_\alpha$ is the total number of
particles of species $\alpha$, obtained by summing the particle numbers in all the loops of species $\alpha$. The
corresponding loop phase-space integration becomes
\begin{equation}
\label{PhaseSpaceLoopFixedN}
\int D(\mathcal{L}) \rightarrow \int \d \mathbf{x} \int D_{q}(\pmb{\mathcal{X}}(\cdot))
\end{equation}
Similarly to what occurs for the screened representation of particle
densities~\cite{Alastuey2003}, ideal-like contributions of familiar chemical species $\mathcal{E}[N_1,...,N_{\cal S}]$
made with $N_\alpha$ particles of species $\alpha$, $\alpha=1,...,{\cal S}$, are contained in the
sum of all the contributions of graphs ${\cal P}[N_1,...,N_{\cal S}]$.
\bigskip
Within the present formalism, the contributions of $\mathcal{E}[N_1,...,N_{\cal S}]$ are dressed
by the collective effects embedded in the screened potential $\phi$ as well as in the ring sum $I_{\rm R}$.
The sum of the contributions of all
graphs ${\cal P}[N_1,...,N_{\cal S}]$ for a given set $(N_1,...,N_{\cal S})$ defines a cluster function
$Z[N_1,...,N_{\cal S}]$. It includes ideal-like contribution for the dressed chemical species $\mathcal{E}[N_1,...,N_{\cal S}]$,
as well as interactions between the chemical species resulting from the dissociation of $\mathcal{E}[N_1,...,N_{\cal S}]$.
\bigskip
Let us consider the case of the hydrogen-helium mixture ${\cal S}=3$, made with protons ($\alpha=1$),
alpha-nuclei ($\alpha=2$) and electrons ($\alpha=3={\rm e}$).
Hydrogen atoms are associated with graphs ${\cal P}[1,0,1]$ made with one proton and one electron, helium atoms with graphs ${\cal P}[0,1,2]$
made with one alpha-particle and two electrons, etc...
For instance, $Z[0,1,2]$ accounts for a dressed atom ${\rm He}$ as well
as interactions between (i) one ion ${\rm He}^+$ and one electron (ii) one alpha-nuclei and one electron.
Also, $Z[2,0,2]$ describes a dressed molecule ${\rm H}_2$, interactions between two dressed atoms ${\rm H}$, etc...
\bigskip
In the zero-density limit, the cluster functions can be related to suitably defined bare partition
functions of the chemical species in the vacuum~\cite{Alastuey2003}. We stress that the systematic prescriptions
defining these cluster functions avoid double counting problems. Moreover, they properly account for the collective
screening effects which ensure the finiteness of the bare partition functions, without introducing ad-hoc regularizations
as in the phenomenological Planck-Larkin partition functions (see \textsl{e.g. }~\cite{Ebeling2017,Ballenegger2012}).
For instance, in the case of the hydrogen plasma made with protons ($\alpha=1$) and electrons ($\alpha=2$), the zero-density limit of $Z[1,1]$ gives rise to the bare partition function
$Z_{{\rm H}}$ of the hydrogen atom in the vacuum, which is close to the virial second-order function $Q$ first introduced
by Ebeling~\cite{Ebeling1967}. Similar partition functions $Z_{{\rm H}_2^+}$, $Z_{{\rm H}^-}$ and $Z_{{\rm H}_{2}}$ for ions and molecules can
be defined. They control the systematic corrections to Saha theory for a partially ionized atomic gas~\cite{Alastuey2008}.
\subsection{Simple scheme using the Debye-Dressing neutralization prescription}
In order to calculate the particle densities associated with a given $P_{\rm A}(\beta;\{ z_i \})$,
use of the DD prescription is particularly attractive. Firstly, it is based on an important physical mechanism related to
Debye screening. Secondly, the dressed densities~(\ref{DressingRecipe}) are given by
a general expression which does not depend on the form of $P_{\rm A}(\beta;\{ z_i \})$.
Eventually, the resulting EOS can be determined within the following scheme
which is simple to implement in practice. For fixing ideas, we illustrate this scheme for a three-component system
like the hydrogen-helium mixture for instance:
\begin{enumerate}
\item Consider various sets $(z_1,z_2,z_{\rm e})$ that satisfy the pseudo-neutrality condition~(\ref{PseudoNeutrality}), i.e. $z_{\rm e}=Z_1z_1 + Z_2z_2$. For each set, compute\\[-1.8em]
\begin{itemize}
\item the pressure $P_{\rm A}(\beta, z_1,z_2, z_{\rm e})$
\item the (DD) particle densities~(\ref{DressingRecipe}) through numerical partial differentiations of $P_{\rm A}$.
\end{itemize}
\item From the pressures and the associated densities computed at the previous step, determine the EOS $\beta P_{\rm A}^{\rm DD}(\beta;\rho_1,\rho_2)$.
\end{enumerate}
This scheme avoids having to invert the relation between the pseudo-neutral sets $(z_1,\, z_2,\, Z_1z_1 + Z_2z_2)$ and
the nuclei densities $(\rho_1,\rho_2)$ for computing $\beta P_{\rm A}^{\rm DD}(\beta;\rho_1,\rho_2)$. Other approximate EOS would be obtained
by using either the Neutral-Group method or the Enforced-Neutrality procedure. However, for approximate functions $P_{\rm A}(\beta;\{ z_i \})$
obtained within the diagrammatic series~(\ref{ScreenedMayerP}), the Debye-Dressing recipe is more directly related to a crucial mechanism at
work than these methods. Hence, it can be reasonably expected to provide better EOS than the Neutral-Group or Enforced-Neutrality procedures.
\section{Conclusions et perspectives}
\label{sec:S6}
We have derived the screened activity series \eqref{ScreenedMayerP} of the pressure of a quantum multicomponent plasma, which provides a convenient route for computing the equation of state of such systems at low and moderate densities. We have demonstrated that this new series simplifies significantly the calculation of the EOS by reducing drastically the number of diagrams to be computed and by being more efficient for a numerical perspective since it avoids integrating term-by-term diagrams contributing to the particle densities. This representation is also quite promising for deriving approximate EOS for moderately dense plasmas. In particular it
accounts, in a non-perturbative way, for the emergence of any chemical species, atoms, molecules, ions, which are formed through
recombination processes of nuclei and electrons.
Use of the screened activity expansion of the pressure offers a wide
flexibility for various approximations, through the selection of relevant graphs associated with
crucial mechanisms at work. Accurate approximations for the screening potential $\phi$, which simplify the task of computing such graphs, are also available~\cite{Ballenegger2017}.
\bigskip
When devising an approximate theory, it is crucial to ensure that it is compatible with the local charge neutrality. We have devised two schemes for enforcing electroneutrality in approximate theories. The first scheme, the Neutral-Group (NG) neutralization prescription, is based on the Lieb-Lebowitz theorem which implies that the exact pressure depends on the activities only via neutral-group activities, which are variables with a clear physical interpretation. This prescription is very general and several implementations of this scheme are possible in plasmas with three or more components. It is straightforward to use since the corresponding densities are given by fully explicit formulae. The second neutralization scheme, the Debye-Dressing (DD) prescription, is also new and uses the Debye screening effect to enforce electroneutrality. More specifically, the appearance of a neutralizing polarization cloud around each particle is accounted for in that scheme at all orders in the particle activities at a mean-field classical (Debye-Hückel) level. The choice of a particular scheme is worthy of attention because it can affect the computed equation of state, as shown on a simple example. Contrary to the Enforced-Neutrality (EN) scheme which has been used previously, the NG and DD schemes do not break the pseudo-neutrality condition, which is often employed in EOS calculations in the grand-canonical ensemble. The latter two schemes being fully explicit, they do not require solving any equation specific to the studied system. The DD prescription is closely related to the NG scheme.
Whether the DD prescription is a special case of a NG prescription for a specific choice of basis for the neutral groups remains on open question. When calculating an approximate equation of state by using the new diagrammatical series~\eqref{ScreenedMayerP} for the pressure, the DD prescription should be preferred because it is based on a physical phenomenon and because double-counting of screening effects can be avoided by a proper selection of the retained diagrams in the pressure series.
\bigskip
Eventually, the methods presented in this paper will be applied to derive accurate approximate equations of state for hydrogen and hydrogen-helium mixtures at moderate densities. The EOS of such plasmas can be studied by computing the screened Mayer diagrams using analytical and numerical techniques.
The cluster functions for fixed number of particles, defined by summing Mayer diagrams with a constraint on the total number of particles, play a central role in such calculations~\cite{Ballenegger2017,Alastuey2012a,Alastuey2008}. A partial account of calculations along the Sun adiabat is given in Refs.~\cite{WendlandThesis,Ballenegger2018}. A more systematic study including denser regimes will be published elsewhere.
\section*{Acknowledgements}
Financial support from the CNRS (contract 081912) and from the Conseil r\'egional de Franche-Comt\'e (contract 362887) are gratefully acknowledged.
| b73e5e714be29d6398abd159a258fdbc676b06e6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The Heisenberg uncertainty relation is perhaps best known in the form
\begin{equation}\label{QP}
\sigma_{Q,\psi}\, \sigma_{P,\psi} \geq \frac{\hbar}{2}
\end{equation}
due to Kennard \cite{Ken27}, where $\sigma_{Q,\psi}$ and $\sigma_{P,\psi}$ are the standard deviations of the probability distributions of position and momentum associated with a given wave function $\psi$ of norm 1. A well-known generalization to arbitrary self-adjoint operators $A,B$ in the place of $Q,P$, established by Robertson \cite{Rob} and Schr\"odinger \cite{Schr}, asserts that
\begin{equation}\label{AB}
\sigma_{A,\psi} \,\sigma_{B,\psi} \geq \tfrac{1}{2} \Bigl| \scp{\psi}{[A,B]|\psi} \Bigr|\,.
\end{equation}
It was expected already in the early days of quantum mechanics that a relation of the form
\begin{equation}\label{TE}
\sigma_{T,\psi}\, \sigma_{E,\psi}\geq \frac{\hbar}{2}
\end{equation}
should hold between time $T$ and energy $E$; however, it is far from obvious which ``time observable'' should be meant here, and various possibilities have been discussed in the literature, see, e.g., \cite{AB61,Busch08,KRSW12,DD15} and the references therein.
In Section~\ref{sec:pfTE}, we prove \eqref{TE} for a particular time observable, the time at which a non-relativistic quantum particle, whose initial wave function $\psi_0$ is concentrated in some region $\Omega\subset \mathbb{R}^3$ of physical space, gets registered by ideal detectors placed along the surface $\partial \R$ of $\Omega$. Mathematically, we take this observable to be defined by the \emph{absorbing boundary rule} \cite{Wer87,detect-rule,detect-several,detect-dirac,detect-imaginary,detect-thm}, which asserts that, in the presence of such detectors, $\psi$ evolves for $t\geq 0$ according to the Schr\"odinger equation
\begin{equation}\label{Schr}
i\hbar\frac{\partial \psi}{\partial t} = -\frac{\hbar^2}{2m} \nabla^2 \psi
\end{equation}
with the absorbing boundary condition
\begin{equation}\label{abc}
\boldsymbol{n}(\boldsymbol{x}) \cdot \nabla \psi(\boldsymbol{x}) = i\kappa\psi(\boldsymbol{x})
\end{equation}
for every $\boldsymbol{x}\in\partial \R$, where $\boldsymbol{n}(\boldsymbol{x})$ is the outward unit normal vector to $\partial \R$ at $\boldsymbol{x}$ and $\kappa>0$ is a constant (the wave number of sensitivity of the detectors). The rule asserts further that the probability that the time $T$ and location ${\boldsymbol{X}}$ of detection lie in the interval $[t,t+dt]$ and the surface element $d^2\boldsymbol{x}$ in $\partial \R$ around $\boldsymbol{x}$, respectively, is given by
\begin{align}
\mathrm{Prob}_{\psi_0}\bigl(T\in dt, {\boldsymbol{X}}\in d^2\boldsymbol{x}\bigr)
&= \boldsymbol{n}(\boldsymbol{x}) \cdot {\boldsymbol{j}}^{\psi_t}(\boldsymbol{x}) \,dt\, d^2\boldsymbol{x} \label{prob1}\\
&= \tfrac{\hbar\kappa}{m} |\psi_t(\boldsymbol{x})|^2 \, dt\,d^2\boldsymbol{x}\,,\label{prob2}
\end{align}
where it is assumed that $\|\psi_0\|=1$, ${\boldsymbol{j}}^{\psi}$ is the usual probability current associated with $\psi$,
\begin{equation}\label{jdef}
{\boldsymbol{j}}^{\psi} = \tfrac{\hbar}{m} \, \Im[\psi^* \nabla \psi]\,,
\end{equation}
and \eqref{prob1} and \eqref{prob2} are equivalent by virtue of the boundary condition \eqref{abc}.
The absorbing boundary rule represents an ideal ``hard'' detector, i.e., one that detects the particle as soon as it reaches $\partial \R$, as opposed to a ``soft'' detector that may take a while to notice a particle moving through the detector volume. Soft detectors can be conveniently modeled through imaginary potentials \cite{All69b,detect-imaginary}; an energy--time uncertainty relation concerning the time at which a soft detector clicks was established by Kiukas et al.~\cite{KRSW12}. That result, together with the facts that the relation does not depend on the parameters of the soft detector and that the absorbing boundary rule can be obtained as a limit of soft detectors \cite{detect-imaginary}, could also provide a strategy for proving \eqref{TE} for hard detectors; however, this strategy is not straightforward because the relevant limit involves changing the detector volume and thus the Hilbert space. Be that as it may, we will give a different, direct proof of \eqref{TE} for the absorbing boundary rule.
In the event that the particle is never detected, we write $T=\infty$. The probability
\begin{equation}
p:= \mathrm{Prob}_{\psi_0}(T<\infty)
\end{equation}
is 1 if $\Omega$ is a bounded set, but may otherwise be less than 1, for example when $\Omega$ is a half space. If $0<p<1$ then $T$ necessarily has infinite mean and variance, so \eqref{TE} is trivially true because $\sigma_{T,\psi}=\infty$. In such a scenario it may be of interest (as suggested in \cite{KRSW12}) to consider the conditional distribution of $T$, given that $T<\infty$, and define $\sigma_{T,\psi}$ as the standard deviation of \emph{that} distribution; for this case we will show in Section~\ref{sec:pfTEp} that
\begin{equation}\label{TEp}
\sigma_{T,\psi} \, \sigma_{E,\psi} \geq \sqrt{p}\frac{\hbar}{2}
\end{equation}
instead of \eqref{TE}.
It is noteworthy that the constant in \eqref{TE} does not depend on the detector parameter $\kappa$, and that the constant in \eqref{TEp} depends on it only through $p$. The question thus arises whether the constants in \eqref{TE} and \eqref{TEp} are sharp; at present, I cannot answer this question. (I give some discussion at the end of Section~\ref{sec:pfTE}.)
It would also be of interest to study whether the relation
\begin{equation}
\mathbb{E}_\psi\bigl[ T |T<\infty\bigr] \: \sigma_{E,\psi} \geq C \, \sqrt{p} \, \hbar
\end{equation}
with $\mathbb{E}_\psi[ T |T<\infty]$ the expectation of $T$ conditional on $T<\infty$, proved in \cite{KRSW12} for soft detectors, is also true for the absorbing boundary rule, and if so, what the optimal value of the numerical constant $C$ is. Likewise, relations between the means of $T$ and $E$, i.e., $\mathbb{E}_\psi[ T |T<\infty]$ and $\mathbb{E}_\psi E$, would be of interest. These questions will not be addressed here.
\section{Definition of Energy}
We need to say more about the exact statement involving \eqref{TE}, in particular what exactly the quantities in \eqref{TE} mean for us. As mentioned, the quantity $\sigma_{T,\psi}$ in \eqref{TE} is the standard deviation of the probability distribution of $T$,
\begin{equation}\label{sigmaT}
\sigma_{T,\psi}^2 = \mathrm{Var}(T)=\mathbb{E}_\psi \bigl[ (T-\mathbb{E}_\psi T)^2 \bigr]\,.
\end{equation}
We assume that $\psi=\psi_0$ is smooth with compact support that does not touch $\partial \R$; for such $\psi$ we write $\psi\in C_0^\infty(\Omega\setminus\partial \R)$. For the proof to go through, however, it suffices to make the weaker assumption that $\psi$ lies in the domain $D(H^2)$ of $H^2$ and is such that both $\psi$ and $H\psi$ vanish on the boundary.
It follows \cite{detect-rule,detect-thm} from \eqref{prob2} that the probability distribution of $T$ is given by a POVM (positive-operator-valued measure \cite{Nai40,Lud85,povm}) $F(\cdot)$ in the sense that
\begin{equation}
\mathrm{Prob}_{\psi_0}(T\in\Delta) = \scp{\psi_0}{F(\Delta)|\psi_0}
\end{equation}
for every (measurable) set $\Delta \subseteq [0,\infty]$. Since the distribution of $T$ is given by a POVM, not a self-adjoint operator, the Robertson--Schr\"odinger inequality \eqref{AB} does not directly apply to yield the desired relation \eqref{TE}. The POVM can be expressed as
\begin{align}
F \bigl( dt \bigr)
&= \tfrac{\hbar\kappa}{m} \, W_t^* \biggl(\int_{\partial \R} d^2\boldsymbol{x} \,
|\boldsymbol{x}\rangle\langle\boldsymbol{x}| \biggr) W_t\, dt \\
&= -\tfrac{i}{\hbar} W_t^*(H^*-H)W_t \, dt\\[2mm]
F (\{\infty\})
&=\lim_{t\to\infty} W_t^* W_t
\end{align}
with ${}^*$ denoting the adjoint operator and $W_t$ the (non-unitary) linear operator that maps $\psi_0$ to $\psi_t$ solving \eqref{Schr} and \eqref{abc}. The operators $W_t$ are of the form $W_t=e^{-iHt/\hbar}$ for $t\geq 0$ but the Hamiltonian $H$ is \emph{not} self-adjoint. The $W_t$ have the properties $W_0=I$, $W_t W_s=W_{t+s}$, and $\|W_t\psi\|\leq \|\psi\|$; that is, they form a \emph{contraction semigroup}.
This brings us to the question of what exactly is meant by $\sigma_{E,\psi}$. The question arises because the Hamiltonian $H$ is not self-adjoint and, in fact, not unitarily diagonalizable (it is not a ``normal'' operator, its self-adjoint and its skew-adjoint part do not commute and therefore cannot be simultaneously diagonalized \cite{detect-thm}). As a consequence, there is no PVM (projection-valued measure) or POVM associated with $H$, so it is far from clear whether and how a probability distribution (over real or complex ``energies'') can be associated with $H$ and $\psi=\psi_0$. We will use instead the free Hamiltonian
\begin{equation}
H_{\mathrm{free}}=-\frac{\hbar^2}{2m}\nabla^2
\end{equation}
on (the appropriate domain in) $L^2(\mathbb{R}^3)$ (i.e., the second Sobolev space). This is anyhow what one would naturally regard as the ``energy distribution'' of $\psi\in C_0^\infty(\Omega\setminus\partial \R)$. The free Hamiltonian is self-adjoint and thus associated with a PVM that defines, for every $\psi\in L^2(\mathbb{R}^3)$, a probability distribution on the energy axis. In fact, this distribution is concentrated on the positive half axis and there has density
\begin{equation}
\rho(E)=
\frac{\sqrt{2m^3E}}{\hbar^3}\int_{\mathbb{S}^2}d^2{\boldsymbol{\omega}} \,\biggl|\widehat{\psi}\biggl(\frac{\sqrt{2mE}}{\hbar}{\boldsymbol{\omega}}\biggr)\biggr|^2 \,,
\end{equation}
where $\mathbb{S}^2=\{{\boldsymbol{\omega}}\in\mathbb{R}^3:|{\boldsymbol{\omega}}|=1\}$ is the unit sphere, $d^2{\boldsymbol{\omega}} $ the surface element ($\sin \theta \, d\theta \, d\varphi$ in spherical coordinates), and $\widehat{\psi}$ the Fourier transform of $\psi$. The standard deviation of this distribution is $\sigma_{E,\psi}$. Equivalently,
\begin{equation}\label{sigmaEHfree}
\sigma_{E,\psi}^2 = \scp{\psi}{H_{\mathrm{free}}^2|\psi}-\scp{\psi}{H_{\mathrm{free}}|\psi}^2\,.
\end{equation}
The following observation about $\psi\in C_0^\infty(\Omega\setminus\partial \R)$ is relevant here: Consider the three different viewpoints of taking the Hamiltonian to be $H$ (and regarding $\psi$ as an element of $L^2(\Omega)$), or of taking the Hamiltonian to be $H_{\mathrm{free}}$ (and regarding $\psi$ as an element of $L^2(\mathbb{R}^3)$), or of taking the Hamiltonian to be $H_{\mathrm{Dir}}=-(\hbar^2/2m)\nabla^2$ with Dirichlet boundary conditions on $\partial \R$ (and regarding $\psi$ as an element of $L^2(\Omega)$ again). The latter two viewpoints lead to completely different probability distributions over the energy axis (one is continuous, the other discrete for bounded $\Omega$), but both have the same mean and variance, as the mean is given by
\begin{equation}
\scp{\psi}{H_{\mathrm{free}}|\psi}=\scp{\psi}{H_{\mathrm{Dir}}|\psi}
\end{equation}
(because for a function not touching the boundary, both $H_{\mathrm{free}}$ and $H_{\mathrm{Dir}}$ are given by $-(\hbar^2/2m)\nabla^2$), and the variance by
\begin{equation}
\scp{\psi}{H_{\mathrm{free}}^2|\psi}-\scp{\psi}{H_{\mathrm{free}}|\psi}^2=\scp{\psi}{H_{\mathrm{Dir}}^2|\psi}-\scp{\psi}{H_{\mathrm{Dir}}|\psi}^2
\end{equation}
(because both $H_{\mathrm{free}}^2$ and $H_{\mathrm{Dir}}^2$ are $(\hbar^4/4m^2)\nabla^4$). Furthermore,
the standard formulas for the mean and variance, $\scp{\psi}{H|\psi}$ and $\scp{\psi}{H^2|\psi}-\scp{\psi}{H|\psi}^2$, applied to the non-self-adjoint $H$, still yield the same values, for $\psi\in C_0^\infty (\Omega\setminus\partial \R)$, as $H_{\mathrm{free}}$ and $H_{\mathrm{Dir}}$. That is why $\sigma_{E,\psi}$ can still be expressed in terms of $H$ as
\begin{equation}\label{sigmaH}
\sigma_{E,\psi}^2 = \scp{\psi}{H^2|\psi}-\scp{\psi}{H|\psi}^2\,.
\end{equation}
As another remark, I mention that our proofs of \eqref{TE} and \eqref{TEp} remain valid if a bounded, smooth potential $V:\mathbb{R}^3\to\mathbb{R}$ with bounded derivatives is added to the Schr\"odinger equation \eqref{Schr}, and $\sigma_{E,\psi}$ is understood as the standard deviation for $-(\hbar^2/2m)\nabla^2+V$, in agreement with \eqref{sigmaH}. While for the uncertainty relation \eqref{QP} between position and momentum, the choice of potential is irrelevant, the relation \eqref{TE} between time and energy is affected by the choice of $V$ in two ways: first, because $V$ is part of the meaning of the energy $E$, and second, because the choice of $V$ affects the time evolution and thus the probability distribution of $T$.
\section{Proof of \eqref{TE}}
\label{sec:pfTE}
We first focus on the case that $p=\mathrm{Prob}(T<\infty)=1$ for all $\psi$. We formulate our result for $V=0$. For definiteness, we take $\Omega$ to be open, $\Omega=\Omega\setminus \partial \R$.
\begin{thm}
Suppose that $\Omega\subset \mathbb{R}^3$ is open with a boundary $\partial \R$ that is locally Lipschitz and piecewise $C^1$, and such that
\begin{equation}\label{WtWt0}
\lim_{t\to\infty} W_t^*W_t = 0
\end{equation}
(or, equivalently, $p=1$ for every $\psi\in L^2(\Omega)$).
For every $\psi\in D(H^2)$ with $\|\psi\|=1$ and $\psi$ and $H\psi$ vanishing on $\partial \R$, in particular for $\psi\in C_0^\infty(\Omega)$ with $\|\psi\|=1$, \eqref{TE} is true with \eqref{sigmaT} and \eqref{sigmaEHfree}.
\end{thm}
\begin{proof}
Theorem 1 of \cite{detect-thm} guarantees that the operators $W_t$ exist and the distribution of $T$ is well defined.
Due to \eqref{WtWt0},
\begin{subequations}\label{Jdef}
\begin{equation}
J: L^2(\Omega,d^3\boldsymbol{x}) \to L^2([0,\infty)\times \partial \R, dt\, d^2\boldsymbol{x})\,,
\end{equation}
defined by
\begin{equation}
J\psi(t,\boldsymbol{x}) = \sqrt{\hbar\kappa/m} \,(W_t\, \psi)(\boldsymbol{x})
\end{equation}
\end{subequations}
on $\psi$ from the domain of $H$ and by continuation on other $\psi$s from $L^2(\Omega)$ \cite[Sec.~5]{detect-thm}, is a unitary isomorphism between $\mathscr{H}=L^2(\Omega)$ and a subspace of $\mathscr{H}_{>}=L^2([0,\infty)\times\partial \R)$; the natural PVM on $\mathscr{H}_{>}$ is the Naimark dilation \cite{Nai40} of the joint POVM for detection time $T$ and detection location ${\boldsymbol{X}}$; $T$ corresponds to a self-adjoint operator $\hat T = t$ on $\mathscr{H}_{>}$, where $t$ means multiplication by the variable $t$, and the probability distribution of $T$ for $\psi\in\mathscr{H}$ with $\|\psi\|=1$ is exactly the distribution associated with $\hat T$ and $J\psi$. The time evolution $W_t$ gets mapped by $J$ to the shift $S_t$ on $\mathscr{H}_{>}$, $S_t \phi(s,\boldsymbol{x})=\phi(s+t,\boldsymbol{x})$ (by virtue of the semigroup property):
\begin{equation}\label{JW}
JW_t=S_tJ\,.
\end{equation}
Correspondingly, $H$ gets mapped by $J$ to the generator $H_{>}$ of the semigroup $(S_t)_{t\geq 0}$, i.e., $H_{>}= i\hbar\partial/\partial t$, which is no longer self-adjoint because of the boundary at $t=0$.
We will also regard $\mathscr{H}_{>}$ as a subspace of $\widetilde\mathscr{H}= L^2(\mathbb{R}\times \partial \R)$ (by setting $\phi(t,\boldsymbol{x})=0$ for $t<0$) with $\widetilde S_t$ the shift defined for positive or negative $t$. Note that $\widetilde S_t$ does not agree with $S_t$ even for positive $t$; $\mathscr{H}_{>}$ is not invariant under $\widetilde S_t$ for $t>0$, and $S_t f \neq \widetilde S_t f$ for $f\in\mathscr{H}_{>}\subset \widetilde \mathscr{H}$. Rather, for such $f$, $S_t f=P_{>}\widetilde{S}_t f$ with $P_{>}$ the projection $\widetilde{\mathscr{H}}\to \mathscr{H}_{>}$. The group $S_t$ is generated by the operator $\widetilde H = i\hbar\partial/\partial t$, whose formula looks the same as that of $H_{>}$, but the domain of $\widetilde H$ is different from that of $H_{>}$, and $\widetilde H$ is self-adjoint whereas $H_{>}$ is not. The domain of $\widetilde H$ is $H^1(\mathbb{R},L^2(\partial \R))$, that of $H_{>}$ is $P_{>}H^1(\mathbb{R},L^2(\partial \R))$.
The multiplication by $t$ also defines a self-adjoint operator $\widetilde T$ on $\widetilde\mathscr{H}$, and for $f\in \mathscr{H}_{>}\subset \widetilde\mathscr{H}$, $\hat T f = \widetilde T f$. Now on $\widetilde\mathscr{H}$, there is an uncertainty relation between $\widetilde T$ and $\widetilde H$: For any $\phi\in\widetilde \mathscr{H}$ with $\|\phi\|=1$,
\begin{equation}\label{uncertaintywidetilde}
\sigma_{\widetilde T,\phi} \, \sigma_{\widetilde H,\phi} \geq \frac{\hbar}{2}\,.
\end{equation}
This is simply the uncertainty relation \eqref{QP} for position and momentum because $t$ is now one of the variables in $\psi$, analogous to position, and $\widetilde H$ is the derivative relative to $t$, analogous to momentum.
Here, we need to comment on the mathematical conditions of the validity of \eqref{QP}. The question is whether, for \eqref{QP} to be valid, $\psi$ needs to lie in the domain of $Q$ and that of $P$, or perhaps even in the domain of $QP$ and that of $PQ$. The answer is that \eqref{QP}, when understood appropriately, is valid for every $\psi\in L^2(\mathbb{R})$ with $\|\psi\|=1$ (whereas this is not true in general for \eqref{AB} \cite[Chap.~12]{Hall}). Let us explain. Let $A$ be a self-adjoint operator in the Hilbert space $\mathscr{H}$ and $\psi\in\mathscr{H}$ with $\|\psi\|=1$. Then $\psi$ defines a probability distribution on the spectrum of $A$, the Born distribution $\scp{\psi}{P(\cdot)|\psi}$ with $P$ the spectral PVM of $A$, $A=\int_{\mathbb{R}}P(d\alpha) \, \alpha$. The Born distribution has finite expectation and variance if and only if $\psi$ lies in the domain of $A$. In that case, the variance is given by
\begin{equation}\label{sigmaA}
\sigma_{A,\psi}^2=\|A\psi\|^2 - \scp{\psi}{A\psi}^2\,.
\end{equation}
Thus, for every $\psi\in L^2(\mathbb{R})$ with $\|\psi\|=1$ that lies in both the domain of $Q$ and that of $P$, $\sigma_{Q,\psi}$ and $\sigma_{P,\psi}$ are well defined and finite. For those $\psi$, \eqref{QP} can be proved \cite[Thm.~12.7]{Hall}. However, since for $\psi$ not in the domain of $Q$, $\sigma_{Q,\psi}=\infty$, and since there is no unit vector in $L^2(\mathbb{R})$ for which either $\sigma_Q$ or $\sigma_P$ vanishes, the left-hand side of \eqref{QP} is infinite (and the relation trivially true) for every $\psi$ not in both the domain of $Q$ and that of $P$. Thus, \eqref{QP} is plainly and cleanly valid for all wave functions.
The remainder of the reasoning is about the implications of the relation \eqref{uncertaintywidetilde} in $\widetilde\mathscr{H}$ for $\psi$ that vanishes along with $H\psi$ on $\partial \R$, such as $\psi\in C_0^\infty(\Omega)$. Since $\psi$ lies in the domain of $H$, $J\psi$ lies in the domain of $H_{>}$, and $H_{>}J\psi=JH\psi$. Moreover, $J\psi(t=0,\boldsymbol{x})=0$ (because $\psi$ vanishes on $\partial \R$), and $(\partial/\partial t)J\psi(t=0,\boldsymbol{x})=0$ in $L^2(\partial \R)$ (because $H\psi$ still lies in the domain of $H$ and thus has a trace on $\partial \R$, which in fact is zero). Therefore, $J\psi$ also lies in the domain of $\widetilde{H}$, and
\begin{equation}
\widetilde{H}J\psi = H_{>}J\psi = JH\psi\,.
\end{equation}
Thus, by \eqref{sigmaA},
\begin{subequations}
\begin{align}
\sigma_{\widetilde{H},J\psi}^2
&= \|\widetilde{H}J\psi\|^2-\scp{J\psi}{\widetilde{H}J\psi}^2\\
&= \|JH\psi\|^2-\scp{J\psi}{JH\psi}^2\\[1.5mm]
&= \|H\psi\|^2-\scp{\psi}{H\psi}^2\\[1mm]
&=\sigma_{E,\psi}^2\,,
\end{align}
\end{subequations}
using that $J$ is a unitary isomorphism and \eqref{sigmaH}.
On the other hand,
\begin{align}
\sigma_{\widetilde{T},J\psi}^2
&=\scp{J\psi}{\widetilde{T}^2|J\psi}-\scp{J\psi}{\widetilde{T}|J\psi}^2\\
&=\scp{J\psi}{\hat{T}^2|J\psi}-\scp{J\psi}{\hat{T}|J\psi}^2\\[2mm]
&=\sigma_{T,\psi}^2\,,
\end{align}
which completes the proof of \eqref{TE}.
\end{proof}
Let us turn again to the question of the sharp constant in \eqref{TE}. It is well known that the constant in \eqref{QP} is sharp and equality holds for suitable Gaussian packets. As a consequence, $\sigma_{\widetilde T} \, \sigma_{\widetilde H}=\hbar/2$ for some states in $\widetilde\mathscr{H}$, and can presumably come arbitrarily close to $\hbar/2$ for suitable states in $\mathscr{H}_{>}$. The difficulty with making corresponding statements about $\sigma_T \, \sigma_E$ for states in $L^2(\Omega)$ is to characterize the functions in the range of $J$ and to control how close they can come to Gaussian packets in $\widetilde{\mathscr{H}}$.
\section{Proof of \eqref{TEp}}
\label{sec:pfTEp}
In \eqref{TEp}, we mean by $\sigma_{T,\psi}$ the standard deviation of $T$, conditional on $T<\infty$:
\begin{equation}\label{sigmaTp}
\sigma_{T,\psi}^2 = \mathrm{Var}(T|T<\infty)\,.
\end{equation}
\begin{thm}
Suppose that $\Omega\subset \mathbb{R}^3$ is open with a boundary $\partial \R$ that is locally Lipschitz and piecewise $C^1$. For every $\psi\in D(H^2)$ with $\|\psi\|=1$ and $\psi$ and $H\psi$ vanishing on $\partial \R$, in particular for $\psi\in C_0^\infty(\Omega)$ with $\|\psi\|=1$, \eqref{TEp} is true with \eqref{sigmaTp} and \eqref{sigmaEHfree}.
\end{thm}
\begin{proof}
The mapping $J$ defined by \eqref{Jdef} is now not unitary but still a contraction,
\begin{equation}\label{Jcontract}
\|J\phi\| \leq \|\phi\| \quad \forall \phi\in L^2(\Omega)\,.
\end{equation}
Since $\|J\psi\|=\sqrt{p}$, $J\psi/\sqrt{p}$ is a unit vector.
As before,
\begin{equation}
S_t J=J W_t\,,~~\text{so}~~
H_{>} J = J H\,,
\end{equation}
and the assumption on $\psi$ implies again that $J\psi$ lies in the domain of $\widetilde H$ and
\begin{equation}
\widetilde H J\psi=JH\psi\,.
\end{equation}
By the Kennard relation for $t$ and $i\partial/\partial t$,
\begin{subequations}\label{Hp}
\begin{align}
\frac{\hbar^2}{4\sigma^2_{T,J\psi/\sqrt{p}}}
&\leq \sigma^2_{\widetilde{H},J\psi/\sqrt{p}}\\
&= \frac{1}{p} \|\widetilde{H}J\psi\|^2 - \frac{1}{p^2} \scp{J\psi}{\widetilde{H}J\psi}^2\\
&\leq \frac{1}{p} \bigl\|\widetilde{H}J\psi\bigr\|^2\\
&= \frac{1}{p} \bigl\| J H \psi\bigr\|^2\\
&\stackrel{\eqref{Jcontract}}{\leq} \frac{1}{p} \bigl\| H \psi\bigr\|^2\\
&= \frac{1}{p} \bigl\| H_\mathrm{free} \psi\bigr\|^2\,.
\end{align}
\end{subequations}
Since for any self-adjoint operator $A$, the variance $\sigma^2_A$ does not change when adding a constant $c\in\mathbb{R}$ to $A$, $\sigma^2_{A+cI}=\sigma^2_A$, we can replace $\widetilde{H}$ in \eqref{Hp} by $\widetilde{H}+E_0I$; thus, for any $E_0\in \mathbb{R}$,
\begin{equation}\label{Hfreep}
\frac{\hbar^2}{4\sigma^2_{T,J\psi/\sqrt{p}}}
\leq \frac{1}{p} \bigl\| (H_\mathrm{free}+E_0I) \psi\bigr\|^2\,.
\end{equation}
Since for any real random variable $X$, its variance can be characterized as
\begin{equation}
\text{Var} \, X = \inf_{c\in\mathbb{R}} \mathbb{E}\bigl[ (X-c)^2 \bigr]\,,
\end{equation}
we have that for any observable $A$,
\begin{equation}
\sigma_{A,\psi}^2 = \inf_{c\in\mathbb{R}} \scp{\psi}{(A-cI)^2|\psi}\,,
\end{equation}
in particular
\begin{equation}
\sigma_{H_\mathrm{free},\psi}^2= \inf_{E_0\in\mathbb{R}} \bigl\| (H_\mathrm{free}+E_0I) \psi\bigr\|^2\,.
\end{equation}
Thus, with \eqref{Hfreep},
\begin{equation}
\frac{\hbar^2}{4\sigma^2_{T,J\psi/\sqrt{p}}}
\leq \frac{1}{p} \sigma_{H_\mathrm{free},\psi}^2\,,
\end{equation}
which is just a different notation for the desired relation \eqref{TEp}, so the proof is complete.
\end{proof}
\bigskip
\noindent{\it Acknowledgment.} I thank Stefan Teufel for helpful discussion.
| f6f614f68efb481c256fef3f8e182ff2ba86b715 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:introduction}
The combination of ubiquitous multimedia and high performance computing resources has inspired numerous efforts to \emph{manipulate} audio-visual content for both benign and sinister purposes. Recently, there has been a lot of interest in the creation and detection of high-quality videos containing facial and auditory manipulations, popularly known as \emph{deepfakes}~\cite{dolhansky2019deepfake,DBLP:DF-TIMIT}. Since fake videos are often indistinguishable from genuine counterparts, detection of deepfakes is challenging but topical given their potential for denigration and defamation, especially against women and in propagating misinformation~\cite{dystopia,AI2018}.
Part of the challenge in detecting deepfakes via artifical intelligence (AI) approaches is that deepfakes are themselves created via AI techniques. Neural network-based architectures like Generative Adversarial Networks (GANs)~\cite{NIPS2014_5423} and Autoencoders~\cite{Vincent2008} are used for generating fake media content, and due to their `learnable' nature, output deepfakes become more naturalistic and adept at cheating fake detection methods over time. Improved deepfakes necessitate novel fake detection (FD) solutions; FD methods have primarily looked at frame-level visual features~\cite{Capsule_Forensics} for statistical inconsistencies, but temporal characteristics~\cite{using_RNNs} have also been examined of late. Recently, researchers have induced audio-based manipulations to generate fake content, and therefore, corruptions can occur in both the visual and audio channels.
Deepfakes tend to be characterized by visual inconsistencies such as a lack of lip-sync, unnatural facial and lip appearance/movements or assymmetry between facial regions such as the left and right eyes (see~Fig.~\ref{fig:localisation} for an example). Such artifacts tend to capture user attention. Authors of~\cite{Grimes1991MildAD} performed a psycho-physical experiment to study the effect of \emph{dissonance}, \textit{i.e.}, lack of sync between the audio and visual channels on user attention and memory. Three different versions of four TV news stories were shown to users, one having perfect audio-visual sync, a second with some asynchrony, and a third with no sync. The study concluded that out-of-sync or \emph{dissonant} audio-visual channels induced a high user cognitive load, while in-sync audio and video (no dissonance condition) were perceived as a \emph{single stimulus} as they `belonged together'.
We adopt the \emph{dissonance} rationale for deepfake detection, and since a fake video would contain either an altered audio or visual channel, one can expect higher audio-visual dissonance for fake videos than for real ones. We measure the audio-visual dissonance in a video via the \emph{\textbf{Modality Dissonance Score}} (MDS), and use this metric to label it as \emph{real}/\emph{fake}. Specifically, audio-visual dissimilarity is computed over 1-second video chunks to perform a fine-grained analysis, and then aggregated over all chunks for deriving MDS, employed as figure-of-merit for video labeling.
MDS is modeled based on the \emph{contrastive} loss, which has traditionally been employed for discovering lip-sync issues in video~\cite{outoftime}. Contrastive loss enforces the video and audio features to be \emph{closer} for real videos, and \emph{farther} for fake ones. Our method also works with videos involving only the audio/visual channel as our neural network architecture includes the video and audio-specific sub-networks, which seek to independently learn discriminative real/fake features via the imposition of a \textit{cross-entropy} loss. Experimental results confirm that these unimodal loss functions facilitate better real-fake discriminability over modeling only the contrastive loss. Experiments on the DFDC~\cite{dolhansky2019deepfake} and DF-TIMIT~\cite{VID-TIMIT} datasets show that our technique outperforms the state-of-the-art by upto 7\%, which we attribute to the following factors: (1) Modeling unimodal losses in addition to the contrastive loss which measures modality dissonance, and (2) Learning discriminative features by comparing 1-second audio-visual \emph{chunks}~to compute MDS, as against directly learning video-level features. Chunk-wise learning also enables \textit{localization} of transient video forgeries (\textit{i.e.}, where only some frames in a sequence are corrupted), while past works have only focused on the \emph{real}-\emph{fake} labeling problem.
The {key contributions} of our work are: (a) We propose a novel multimodal framework for deepfake video detection based on modality dissonance computed over small temporal segments; (b) To our knowledge, our method is the first to achieve temporal forgery localization, and (c) Our method achieves state-of-art results on the DFDC dataset, improving AUC score by up to 7\%.
\section{Literature Review}
Recently, considerable research efforts have been devoted to detecting fake multimedia content automatically and efficiently. Most video faking methods focus on manipulating the video modality; audio-based manipulation is relatively rare. Techniques typically employed for generating fake visual content include 3D face modeling, computer graphics-based image rendering, GAN-based face image synthesis, image warping, \textit{etc}. Artifacts are synthesized either via face swapping while keeping the expressions intact (\textit{e.g.}, DeepFakes\footnote{https://github.com/deepfakes/faceswap}, FaceSwap\footnote{https://github.com/MarekKowalski/FaceSwap/}) or via facial expression transfer, \textit{i.e.}, facial reenactment (\textit{e.g.}, Face2Face \cite{thies2016face}). Approaches for fake video detection can be broadly divided into three categories as follows:
\subsection{Image-based} These methods employ image/frame-based statistical inconsistencies for \emph{real}/\emph{fake} classification. For example,~\cite{visual_artifacts} uses visual artifacts such as missing reflections and incomplete details in the eye and teeth regions, inconsistent illumination and heterogeneous eye colours as cues for fake detection. Similarly,~\cite{inconsistent_headposes} relies on the hypothesis that the 3-D head pose generated using the entire face's landmark points will be very different from the one computed using only the central facial landmarks, in case the face is synthetically generated. Authors of \cite{Li2018ExposingDVFWA}~hypothesize that fake videos contain artifacts due to resolution inconsistency between the warped face region (which is usually blurred) and surrounding context, and models the same via the VGG and ResNet deep network architectures. A capsule network architecture is proposed in~\cite{Capsule_Forensics} for detecting various kinds of spoofs, such as video replays with embedded images and computer-generated videos.
\begin{sloppypar}
A multi-task learning framework for fake detection-cum-segmentation of manipulated image (or video frame) regions is presented in~\cite{multi-task}. It is based on a convolutional neural network, comprising an encoder and a Y-shaped decoder, where information gained by one of the detection/segmentation tasks is shared with the other so as to benefit both tasks. A two-stream network is proposed in~\cite{Two-stream}, which leverages information from local noise residuals and camera characteristics. It employs a GoogLeNet-based architecture for one stream, and a patch based triplet network as second stream. Authors of~\cite{afchar2018mesonet} train a CNN, named MesoNet, to classify real and fake faces generated by the DeepFake and Face2face techniques. Given the similar nature of falsifications achieved by these methods, identical network structures are trained for both problems by focusing on mesoscopic properties (intermediate-level analysis) of images.
\end{sloppypar}
\subsection{Video-based}
Video-based fake detection methods also use temporal features for classification, since many a time, deepfake videos contain realistic frames but the warping used for manipulation is temporally inconsistent. For instance, variations in eye-blinking patterns are utilized in~\cite{in_ictu_oculi} to determine real and fake videos. Correlations between different pairs of facial action units across frames are employed for forgery detection in~\cite{protecting_leaders}. Authors of~\cite{using_RNNs} extract frame-level CNN features, and use them to train a recurrent neural network for manipulation detection.
\subsection{Audio-visual features based}
Aforementioned approaches exploit only the visual modality for identifying deepfake videos. However, examining other modalities such as audio signal in addition to the video can also be helpful. As an example, authors of~\cite{mittal2020emotions} propose a Siamese network-based approach, which compares the multimedia as well as emotion-based differences in facial movements and speech patterns to learn differences between real and fake videos. Lip-sync detection in unconstrained settings is achieved by learning the \emph{closeness} between the audio and visual channels for in-sync vs out-of-sync videos via contrastive loss modeling in~\cite{outoftime}. While this technique is not meant to address deep-fake detection \emph{per se}, lip-sync issues can also be noted from a careful examination of deepfake videos, and the contrastive loss is useful for \emph{tying up} genuine audio-visual pairs.
\subsection{Bimodal Approaches}
While it may be natural to see audio and video as the two main information modes that indicate the genuineness/fakeness of a video, one can also employ multiple cues from the visual modality for identifying fake videos. Multimodal cues are especially useful while tackling sophisticated visual manipulations. In~\cite{using_RNNs}, both intra-frame and inter-frame visual consistencies are modeled by feeding in frame-level CNN features to train a recurrent neural network. Statistical differences between the warped face area and the surrounding regions are learned via the VGG and ResNet architectures in~\cite{Li2018ExposingDVFWA}. Hierarchical relations in image (video frame) content are learned via a Capsule Network architecture in~\cite{Capsule_Forensics}.
\subsection{Analysis of Related Work}
Upon examining related literature, we make the following remarks to situate our work with respect to existing works.
\begin{figure*}[!tbph]
\centering
\includegraphics[width=\textwidth]{BTP-pipeline.pdf}
\caption{MDS-based fake video detection: Features extracted from 1-second audio-visual segments are input to the MDS network. The MDS network comprises the audio and visual sub-networks, whose description is provided in Table~\ref{tab:stream_arch}. Descriptors learned by the video and audio sub-networks are tuned via the cross-entropy loss, while the contrastive loss is employed to enforce higher dissimilarity between audio-visual chunks arising from \emph{fake} videos. MDS is computed as the aggregate audio-visual dissonance over the video length, and employed as a figure of merit for labeling a video as \emph{real}/\emph{fake}.}
\label{fig:pipeline}
\end{figure*}
\begin{itemize}
\item[1.] While frame-based methods that learn spatial inconsistencies have been proposed to detect deepfakes, temporal-based approaches are conceivably more powerful to this end, as even if the manipulation looks natural in a static image, achieving temporally consistent warping even over a few seconds of video is difficult. Our approach models temporal characteristics, as we consider 1-second long audio-visual segments to distinguish between real and fake videos. Learning over tiny video chunks allows for a fine-grained examination of temporal differences, and also enables our method to temporally \emph{localize} manipulation in cases where the forgery targets only a small portion of the video. To our knowledge, prior works have only focused on assigning a \emph{real}/\emph{fake} video label.
\item[2.] Very few approaches have addressed deepfake detection where the audio/video stream may be corrupted. In this regard, two works very related to ours are~\cite{outoftime} and~\cite{mittal2020emotions}. In~\cite{outoftime}, the contrastive loss function is utilized to enforce a smaller distance between lip-synced audio-visual counterparts; our work represents a novel application of the contrastive loss, employed for lip-sync detection in~\cite{outoftime}, to deepfake detection. Additionally, we show that learning audio-visual dissonance over small chunks and aggregating these measures over the video duration is more beneficial than directly learning video-level features.
\item[3.] Both multimedia and emotional audio-visual features are learned for FD in~\cite{mittal2020emotions}. We differ from~\cite{mittal2020emotions} in three respects: (a) While we do not explicitly learn emotional audio-visual coherence, forged videos need not be emotional in nature; audio-visual consistency is enforced via the contrastive loss in our approach. (b) The training framework in~\cite{mittal2020emotions} requires a \emph{real}--\emph{fake} video pair. Our approach does not constrain the training process to involve video-pairs, and adopts the traditional training protocol. (c) Whilst~\cite{mittal2020emotions} perform a video-level comparison of audio-visual features to model dissonance, we compare smaller chunks and aggregate chunk-wise measures to obtain the MDS. This enables our approach to localize transient forgeries.
\item[4.] Given that existing datasets primarily involve visual manipulations (number of datasets do not have an audio component), our architecture also includes audio and visual sub-networks which are designed to learn discriminative unimodal features via the cross-entropy loss. Our experiments show that additionally including the cross-entropy loss is more beneficial than employing only the contrastive loss. Enforcing the two losses in conjunction enables our approach to achieve state-of-the-art performance on the DFDC dataset.
\end{itemize}
\section{MDS-based fake video detection}
As discussed in Section \ref{sec:introduction}, our FD technique is based on the hypothesis that deepfake videos have \emph{higher dissonance} between the audio and visual streams as compared to real videos. The dissimilarity between the audio and visual channels for a \emph{real}/\emph{fake} video is obtained via the Modality Dissonance Score (MDS), which is obtained as the aggregate dissimilarity computed over 1-second visual-audio chunks. In addition, our network enforces learning of discriminative visual and auditory features even if the contrastive loss is not part of the learning objective; this enables FD even if the input video is missing the audio/visual modality, in which case the contrastive loss is not computable. A description of our network architecture for MDS-based deepfake detection follows.
\subsection{Overview}
Given an input video, our aim is to classify it as \emph{real} or \emph{fake}. We begin with a training dataset $D = \{(v^1,y^1), (v^2,y^2), ... , (v^N,y^N)\}$ consisting of $N$ videos. Here, $v^i$ denotes the input video and the label $y^i \ \epsilon \ \{0,1\}$ indicates whether the video is \emph{real} ($y^i=0$) or \emph{fake} ($y^i=1$). The training procedure is depicted in Fig. \ref{fig:pipeline}. We extract the audio signal from input video $v^i$ using the \emph{ffmpeg} library, and split it into $D$-second segments. Likewise for the visual modality, we divide the input video into $D$-second long segments, and perform face tracking on these video segments using the S3FD face detector~\cite{zhang2017s3fd} to extract the face crops. This pre-processing gives us segments of visual frames $\{s^i_1, s^i_2, ... , s^i_n\}$ along with corresponding audio segments $\{a^i_1, a^i_2, ... , a^i_n\}$, where $n$ denotes segment count for an input video $v^i$.
We employ a bi-stream network, comprising the audio and video streams, for deepfake detection. Each video segment $s^i_t, t= 1 \ldots n$ is passed through a visual stream $S_v$, and the corresponding audio segment $a^i_t$ is passed through the audio stream $S_a$. These streams are described in Sections \ref{sec:visual} and \ref{sec:audio}. The network is trained using a combination of the \emph{contrastive loss} and the \emph{cross-entropy loss}.
The contrastive loss is meant to \emph{tie up} the audio and visual streams; it ensures that the video and audio streams are \emph{closer} for real videos, and \emph{farther} for fake videos. Consequently, one can expect a low MDS for real, and high MDS for fake videos. If either the audio or visual stream is missing in the input video, in which case the contrastive loss is not computable, the video and audio streams will still learn discriminative features as constrained by the unimodal cross-entropy losses. These loss functions are described in Sec.~\ref{sec:loss}.
\subsection{Visual Stream}
\label{sec:visual}
The input to the visual stream is $s^i_t$, a video sequence of size ($3 \times h \times w \times D*f$), where 3 refers to the RGB color channels of each frame, $h, w$ are the frame height and width, $D$ is the segment length in seconds, and $f$ is the video frame rate. Table \ref{tab:stream_arch} depicts the architecture of the video and audio sub-networks. The visual stream ($S_v$) architecture is inspired by the 3D-ResNet similar to \cite{DBLP:journals/corr/abs-1711-09577}. 3D-CNNs have been widely used for action recognition in videos, and ResNet is one of the most popular architectures for image classification. The feature representation learnt by the visual stream, in particular the fc8 output, denoted by $f_v$ is used to compute the contrastive loss. We also add a 2-neuron classification layer at the end of this stream, which outputs the visual-based prediction label. So the output of this stream, labeled as $\hat{y}_v^i$, constitutes the unimodal cross-entropy loss.
\definecolor{cadetblue}{rgb}{0.37, 0.62, 0.63}
\definecolor{pistachio}{rgb}{0.58, 0.77, 0.45}
\definecolor{orange}{rgb}{1.0, 0.5, 0.0}
\definecolor{palecopper}{rgb}{0.85, 0.54, 0.4}
\definecolor{mountainmeadow}{rgb}{0.19, 0.73, 0.56}
\definecolor{pastelorange}{rgb}{1.0, 0.7, 0.28}
\definecolor{yellow}{rgb}{1.0, 1.0, 0.0}
\definecolor{violet}{rgb}{0.56, 0.0, 1.0}
\begin{table}[t]
\caption{Structure of the audio and visual streams (initial layers of the visual stream are the same as in the 3D ResNet architecture \cite{DBLP:journals/corr/abs-1711-09577}).}
\label{tab:stream_arch}
\scalebox{0.9}{
\begin{tabular}{c c}
\toprule
\textbf{Visual Stream} & \textbf{Audio Stream}\\
\midrule
\multirow{3}{*}{\colorbox{red!30}{conv1}} & \colorbox{red!30}{conv\_1, 3$\times$3, 1, 64}\\
& \colorbox{white}{batch\_norm\_1, 64}\\
& \colorbox{blue!30}{pool\_1, 1$\times$1, MaxPool}\\
\multirow{3}{*}{\colorbox{red!30}{conv2\_x}} & \colorbox{red!30}{conv\_2, 3$\times$3, 64, 192}\\
& \colorbox{white}{batch\_norm\_2, 192}\\
& \colorbox{blue!30}{pool\_2, 3$\times$3, MaxPool}\\
\multirow{2}{*}{\colorbox{red!30}{conv3\_x}} & \colorbox{red!30}{conv\_3, 3$\times$3, 192, 384}\\
& \colorbox{white}{batch\_norm\_3, 384}\\
\multirow{2}{*}{\colorbox{red!30}{conv4\_x}} & \colorbox{red!30}{conv\_4, 3$\times$3, 384, 256}\\
& \colorbox{white}{batch\_norm\_4, 256}\\
\multirow{3}{*}{\colorbox{red!30}{conv5\_x}} & \colorbox{red!30}{conv\_5, 3$\times$3, 256, 256}\\
& \colorbox{white}{batch\_norm\_5, 256}\\
& \colorbox{blue!30}{pool\_5, 3$\times$3, MaxPool}\\
\multirow{2}{*}{\colorbox{orange!80}{average pool}} & \colorbox{red!30}{conv\_6, 5$\times$4, 256, 512}\\
& \colorbox{white}{batch\_norm\_6, 512}\\
\colorbox{mountainmeadow!80}{fc7, 256$\times$7$\times$7, 4096} & \colorbox{mountainmeadow!80}{fc7, 512$\times$21, 4096} \\
\colorbox{white}{batch\_norm\_7, 4096} & \colorbox{white}{batch\_norm\_7, 4096} \\
\colorbox{mountainmeadow!80}{fc8, 4096, 1024} & \colorbox{mountainmeadow!80}{fc8, 4096, 1024} \\
\colorbox{white}{batch\_norm\_8, 1024} & \colorbox{white}{batch\_norm\_8, 1024} \\
\colorbox{cadetblue!80}{dropout, $p=0.5$} & \colorbox{cadetblue!80}{dropout, $p=0.5$} \\
\colorbox{mountainmeadow!80}{fc10, 1024, 2} & \colorbox{mountainmeadow!80}{fc10, 1024, 2} \\
\bottomrule
\end{tabular}}
\end{table}
\subsection{Audio Stream}
\label{sec:audio}
Mel-frequency cepstral coefficients (MFCC) features are input to the audio stream. MFCC features~\cite{Mogran2004} are widely used for speaker and speech recognition \cite{martinez2012speaker}, and have been the state-of-the-art for over three decades. For each audio segment of $D$ second duration $a^i_t$, MFCC values are computed and passed through the audio stream $S_a$. 13 mel frequency bands are used at each time step. Overall, audio is encoded as a heat-map image representing MFCC values for each time step and each mel frequency band. We base the audio stream architecture on convolutional neural networks designed for image recognition. Contrastive loss $L_1$ uses the output of fc8, denoted by $f_a$. Similar to the visual stream, we add a classification layer at the end of the audio stream, and the output $\hat{y}_a^i$ is incorporated in the cross-entropy loss for the audio modality.
\begin{figure}[!b]
\centering \vspace{-4mm}
\subfloat{\includegraphics[width = 70mm]{Contrastive_Alone_Test_1109}}
\newline
\subfloat{\includegraphics[width = 70mm]{Combined_Loss_Test_1109}}
\vspace{-1mm}
\caption{Effect of loss functions: The graphs show the effect of audio and visual cross-entropy loss functions (Section \ref{sec:AblationStudies}). The top and bottom graphs show the MDS distribution for test samples, when the contrastive loss $ L_1$ alone and combined losses $L$ are used in the training of the MDS network, respectively.}\vspace{-1mm}
\label{fig:Graphs}
\end{figure}
\subsection{Loss functions}
\label{sec:loss}
Inspired by~\cite{outoftime}, we use \emph{contrastive loss} as the key component of the objective function. Contrastive loss enables maximization of the dissimilarity score for manipulated video sequences, while minimizing the MDS for real videos. This consequently ensures separability of the real and fake videos based on MDS (see Fig.~\ref{fig:Graphs}). The loss function is represented by Equation \ref{eq:1}. Here, $y^i$ is the label for video $v^i$ and \emph{margin} is a hyper-parameter. Dissimilarity score $d_t^i$, is the Euclidean distance between the (segment-based) feature representations $f_v$ and $f_a$ of the visual and audio streams respectively.
In addition, we employ the cross-entropy loss for the visual and audio streams to learn feature representations in a robust manner. These loss functions are defined in Equations \ref{eq:3} (visual) and \ref{eq:4} (audio). The overall loss $L$ is a weighted sum of these three losses, $L_1, L_2$ and $L_3$ as in Eq.~\ref{eq:9}.
\begin{equation}
L_1 = \frac{1}{N} \sum_{i=1}^{N} (y^i) \ (d_t^i)^2 + (1-y^i) \ max(margin-d_t^i,0)^2 \label{eq:1} \\
\end{equation}
\begin{equation}
d_t^i = \|f_v - f_a\|_2 \label{eq:2}
\end{equation}
\begin{equation}
L_2 = - \ \frac{1}{N} \sum_{i=1}^{N} y^i \ \log{\hat{y}_v^i} \ + (1-y^i) \ \log{(1-\hat{y}_v^i)} \label{eq:3} \\
\end{equation}
\begin{equation}
L_3 = - \ \frac{1}{N} \sum_{i=1}^{N} y^i \ \log{\hat{y}_a^i} \ + (1-y^i) \ \log{(1-\hat{y}_a^i)} \label{eq:4}
\end{equation}
\begin{align}
L = \ \lambda_1L_1 + \lambda_2L_2 + \lambda_3L_3 \label{eq:9}
\end{align}
\noindent where $\lambda_1,\lambda_2, \lambda_3 =1$ in our design.
\subsection{Test Inference}
\label{sec:test}
During test inference, the visual segments $\{s^i_1, s^i_2, ... , s^i_n\}$ and corresponding audio segments $\{a^i_1, a^i_2, ... , a^i_n\}$ of a video are passed through $S_v$ and $S_a$, respectively. For each such segment, dissimilarity score $d_t^i$ (Equation \ref{eq:2}) is accumulated to compute the MDS as below:
\begin{equation}
MDS_i = \frac{1}{n} \sum_{t=1}^{n} d^i_t \label{eq:5}
\end{equation}
\noindent To label the test video, we compare $MDS_i$ with a threshold $\tau$ using $1\{MDS_i < \tau\}$ where $1\{.\}$ denotes the logical indicator function. $\tau$ is determined on the training set. We compute MDS for both real and fake videos of the train set, and the midpoint between the average values for the real and fake videos is used as a representative value for $\tau$.
\section{Experiments}
\subsection{Dataset Description}
As our method is multimodal, we use two public audio-visual deepfake datasets in our experiments. Their description is as follows:\\ \\
\textbf{Deepfake-TIMIT \cite{DBLP:DF-TIMIT}:} This dataset contains videos of 16 similar looking pairs of people, which are manually selected from the publicly available VIDTIMIT \cite{VID-TIMIT} database and manipulated using an open-source GAN-based\footnote{https://github.com/shaoanlu/faceswap-GAN} approach. There are two different models for generating fake videos, one Low Quality (LQ), with $64\times64$ input/output size, and the other High Quality (HQ), with $128\times128$ input/output size. Each of the 32 subjects has 10 videos, resulting in a total of 640 face swapped videos in the dataset; each video is of $512\times384$ resolution with 25 fps frame rate, and of $\approx 4s$ duration. However, the audio channel is not manipulated in any of the videos.\\
\noindent \textbf{DFDC dataset \cite{dolhansky2019deepfake}:} The preview dataset, comprising of 5214 videos was released in Oct 2019 and the complete one with 119,146 videos in Dec 2019. Details of the manipulations have not been disclosed, in order to represent the real adversarial space of facial manipulation. The manipulations can be present in either audio or visual or both of the channels.
In order to bring out a fair comparison, we used 18,000 videos\footnote{Same as the videos used in \cite{mittal2020emotions}.} in our experiments. The videos are of $\approx10s$ duration each with an fps of 30, so there are $\approx300$ frames per video.
\subsection{Training Parameters}
For both the datasets, we used $D=1$ second segment duration and the margin hyper-parameter described in Equation \ref{eq:1} was set to 0.99. This resulted in (3 x 30 x 224 x 224) dimensional input for the visual stream in case of DFDC dataset and for the other dataset, Deepfake-TIMIT, the input dimension to the visual stream was (3 x 25 x 224 x 224). On DFDC we trained our model for 100 epochs with a batch size of 8 whereas for Deepfake-TIMIT the model only required 50 epochs with 16 batch size for convergence as the dataset size was small. We used Adam optimizer with a learning rate of 0.001 and all the results were generated on Nvidia Titan RTX GPU with 32 GB system RAM. For the evaluation, we use video-wise Area Under the Curve (AUC) metric.
\begin{table*}[!tbph]
\caption{Comparison of our method with other techniques on DFDC and DFTIMIT datasets using the AUC metric. For our method frame-wise results are reported inside the brackets. (*We found a discrepancy in the train-test split of the 18k samples subset of \cite{mittal2020emotions}. Hence, the results have been updated after removing the discrepancy from the test set (20$^{th}$ March, 2021).)}
\label{tab:scores}
\centering
\scalebox{0.96}{
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\cline{1-10}
\multirow{2}{*}{}&
\multicolumn{2}{ |c| }{\multirow{2}{*}{} } &
\multicolumn{7}{ |c| }{\textbf{Methods}} \\ \cline{4-10}
& \multicolumn{1}{c}{}& &
Capsule & Multi-task & HeadPose & Two-stream & VA-MLP & VA-LogReg & Meso4 \\
& \multicolumn{1}{c}{}& &
\cite{Capsule_Forensics} & \cite{multi-task} & \cite{inconsistent_headposes} & \cite{Two-stream} & \cite{visual_artifacts} & & \cite{afchar2018mesonet} \\
\cline{1-10}
\multirow{3}{*}{\textbf{Datasets}} & \multicolumn{2}{|c|}{\textbf{DFDC}} &
53.3 & 53.6 & 55.9 & 61.4 & 61.9 & 66.2 & 75.3 \\ \cline{2-10}
& \multirow{2}{*}{\textbf{DFTIMIT}} & \textbf{LQ} &
78.4 & 62.2 & 55.1 & 83.5 & 61.4 & 77.0 & 87.8 \\ \cline{3-10}
& & \textbf{HQ} &
74.4 & 55.3 & 53.2 & 73.5 & 62.1 & 77.3 & 68.4 \\
\cline{1-10}
\textbf{Modality} & \multicolumn{2}{|c|}{} &
V & V & V & V & V & V & V \\ \cline{1-10}
\thickhline
\multirow{2}{*}{}&
\multicolumn{2}{ |c| }{\multirow{2}{*}{} } &
\multicolumn{7}{ |c| }{\textbf{Methods}} \\ \cline{4-10}
& \multicolumn{1}{c}{}& &
Xception-raw & Xception-c40 & Xception-c23 & FWA & DSP-FWA & Siamese-based & \textbf{Our Method} \\
& \multicolumn{1}{c}{}& &
\cite{rossler2019faceforensics++} & & & \cite{Li2018ExposingDVFWA} & & \cite{mittal2020emotions} & \\
\cline{1-10}
\multirow{3}{*}{\textbf{Datasets}} & \multicolumn{2}{|c|}{\textbf{DFDC}} &
49.9 & 69.7 & 72.2 & 72.7 & 75.5 & \colorbox{red!30}{84.4} & \colorbox{blue!30}{\textbf{90.55 (90.66)*}} \\ \cline{2-10}
& \multirow{2}{*}{\textbf{DFTIMIT}} & \textbf{LQ} &
56.7 & 75.8 & 95.9 & \colorbox{blue!30}{99.9} & \colorbox{blue!30}{99.9} & 96.3 & \colorbox{red!30}{97.9 (98.3)} \\ \cline{3-10}
& & \textbf{HQ} &
54.0 & 70.5 & 94.4 & 93.2 & \colorbox{blue!30}{99.7} & 94.9 & \colorbox{red!30}{96.8 (94.7)} \\
\cline{1-10}
\textbf{Modality} & \multicolumn{2}{|c|}{} &
V & V & V & V & V & AV & AV \\ \cline{1-10}
\end{tabular}}
\end{table*}
\begin{table}[b]
\caption{The AUC based comparison of audio stream, visual stream, contrastive loss only and combined loss (Details in Section \ref{sec:AblationStudies}).} \label{tab:lambdas}
\scalebox{0.86}{
\begin{tabular}{|c|c|c|c|}
\hline
\begin{tabular}[c]{@{}c@{}}$\lambda_1$=0,$\lambda_2$=0,$\lambda_3$=1\\ Audio Stream\end{tabular} & \begin{tabular}[c]{@{}c@{}}$\lambda_1$=1,$\lambda_2$=0,$\lambda_3$=0\\ Contrastive Only\end{tabular} & \begin{tabular}[c]{@{}c@{}}$\lambda_1$=0,$\lambda_2$=1,$\lambda_3$=0\\ Visual Stream\end{tabular} & \begin{tabular}[c]{@{}c@{}}$\lambda_1$=1,$\lambda_2$=1,$\lambda_3$=1\\ Combined Loss\end{tabular} \\ \hline
50.0 & 86.1 & 89.7 & 91.7 \\ \hline
\end{tabular}}
\end{table}
\subsection{Ablation Studies} \label{sec:AblationStudies}
To decide the structure of the network and effect of different components, we chose 5800 videos (4000 real and 1800 fake) from the DFDC dataset, divided them into an 85:15 train-test split, following video-wise AUC and conducted the following experiments:\\
\textbf{Effect of Loss Functions:} We evaluated the contribution of the audio and visual streams based cross-entropy loss functions ($L_2$ and $L_3$, respectively). The hypothesis behind adding these two loss functions to the network is that the feature representations learnt across the audio and visual streams, respectively will be more discriminative towards the task of deep fake detection. This should further assist in the task of computing a segment dissimilarity score $d_t^i$, which disambiguates between fake and real videos. This hypothesis was tested by training the network on the DFDC dataset in two different settings. The first setting is based on training the MDS network with contrastive loss only. The second setting is combination of the three loss functions for training of the MDS network. Figure \ref{fig:Graphs} shows two graphs generated using the MDS scores as predicted by the networks from the two settings above. It is observed that the distributions of real and fake videos have lesser intersection in the case of the network trained with all three losses. Overall, combined loss function and contrastive loss only based networks achieved 92.7\% and 86.1\% AUC scores.
The difference attributes to the observation that the gap between average MDS for real and fake videos widens when cross-entropy loss functions are also used. Hence, there is more clear distinction between the dissonance scores for real and fake.
\textbf{Audio and Visual Stream Performance:} We analysed the individual discriminative ability to identify fake and real videos for the audio and the visual streams. In this case the cross-entropy loss alone was used for training of the streams. It was observed that the audio stream only and visual stream only based deepfake classifiers achieved 50.0 and 89.7\%, respectively. Note that audio stream achieves less performance as in the DFDC dataset, minimal manipulation is performed on the audio signal of the videos.
\begin{figure}[t]
\centering
\includegraphics[width=4cm,height=3.5cm]{exp_aud_loss.png}\hspace{2mm
\includegraphics[width=4.1cm,height=3.5cm]{exp_vid_loss.png}\vspace{-1mm}
\caption{Discriminative-ability of individual streams: The graphs show the distribution of the mean L2 distance, when a fake segment and corresponding real segment are passed through only the audio and visual streams.}
\label{fig:vid_aud_plot}
\end{figure}
\begin{figure}[b]
\centering
\vspace{-2mm}
\includegraphics[width=8.6cm]{Activation.png
\caption{Grad-CAM results: Three frames are overlayed with the important attention regions highlighted using Grad-CAM. Note that forehead, face region and neck regions are highlighted in the 1st, 2nd and 3rd frames respectively.} \vspace{-4mm}
\label{fig:GRADCAM}
\end{figure}
In Equation \ref{eq:5}, for the four configurations: audio stream only, contrastive loss only, visual stream only and combined loss, we set the parameters as follows: ($\lambda_1=0,\lambda_2=0,\lambda_3=1$), ($\lambda_1=1,\lambda_2=0,\lambda_3=0$), ($\lambda_1=0,\lambda_2=1,\lambda_3=0$) and ($\lambda_1=1,\lambda_2=1,\lambda_3=1$). In the case of audio stream and visual stream based classification, the cross-entropy generated real and fake probabilities are generated segment-wise. We compute a maximum over the probabilities to compute the final fake and real for these two streams, individually.
For further assessing the audio and visual stream's individual performances, we generated the plots shown in Figure \ref{fig:vid_aud_plot}. The process is as follows: a fake video segment and corresponding real video segment is passed through the streams and L2 distance is computed between the output of fc8 layer of the visual/audio sub-network. Then the average of these L2 distances for a video pair is plotted. It is observed in the audio stream plot, that most of the videos are centered close to 0 inter-pair L2 distance. This is due to the fact that audio has been modified in few cases in DFDC dataset. On the contrary, in the plot for the visual stream, we observed that the inter-pair L2 scores is spread across the dataset. This supports the argument that the visual stream is more discrimiantive due to the added cross entropy loss.
In Table \ref{tab:lambdas}, we show the AUC of the four configurations mentioned above. Note that these numbers are on a smaller set of DFDC, which is used for ablation studies only. The contrastive loss only based configuration, which uses both the streams, achieves 86.1\%. The fusion of the cross-entropy loss into the final MDS network ($\lambda_1=1,\lambda_2=1,\lambda_3=1$ for Equation \ref{eq:5}), achieves 92.7\%. This backs up the argument that comparing (using contrastive loss) features learned through supervised channels enhances the performance of the MDS network.
\textbf{Segment Duration:} For deciding the optimal length of the temporal segment $D$, we conducted empirical analysis with temporal sizes of $D = [1,2,4]$ seconds. From the evaluation on the DFDC dataset, we observed that the temporal duration $D=1$ second is most optimum. This can attributed to the larger number of segments representing each video, thereby allowing fine-grained comparison of the audio and the visual data of a video.
\subsection{Evaluation on DFDC Dataset}
We compare the performance of our method on the DFDC dataset with other state-of-the-art works \cite{mittal2020emotions,Capsule_Forensics,afchar2018mesonet,rossler2019faceforensics++,visual_artifacts,in_ictu_oculi,inconsistent_headposes,Two-stream}. A total of 18,000 videos are used in this experiment\footnote{Please note that some earlier methods in Table \ref{tab:scores} are trained on DFDC preview (5000 videos), which is no longer available. }. In Table \ref{tab:scores}, it is observed that the MDS network approach outperforms the other visual-only and audio-visual based approaches by achieving 91.54\%. The performance is \textasciitilde8\% more relatively than the other audio-visual based approach \cite{mittal2020emotions}. Please note that the result 91.54\% is achieved with the test set containing 3281 videos out of which 281 videos have two subjects each. We chose the larger face and passed it through the visual stream. The performance of the network without these multiple subject videos is 93.50\%. We also report the frame-wise AUC (91.60\%) as mentioned in brackets in Table \ref{tab:scores}. This is computed by assigning each frame of a video the same label as predicted for the video by our method.
We argue that the gain in performance here is due to: (a) The temporal segmentation of the video into segment helps in fine-grained audio-visual feature comparison; (b) We extract task-tuned features from the audio and visual segments. Here task-tuned means that the features are learnt to discriminate between real and fake with the $L_2$ and $L_3$ loss functions, and (c) The visual stream's input is the face and an extra margin (see Figure \ref{fig:pipeline}) around it, which accounts for some rigid (head) movement along with the non-rigid (facial) movements. We visualise the important regions using the Gradient-weighted Class Activation Mapping (Grad-CAM) method \cite{selvaraju2017grad}. Figure \ref{fig:GRADCAM} shows the important regions localised by Grad-CAM on few frames of a video. Note that the region around the face is highlighted in the middle frame. Also, the forehead and the neck regions are highlighted in the first and third frames, respectively. This supports our argument that the disharmony between the non-rigid and rigid movement is also discriminative for the visual stream to classify between real and fake videos.
\subsection{Evaluation on DFTIMIT Dataset}
The DeepFake-TIMIT (DFTIMIT) dataset is smaller as compared to the DFDC dataset. We trained the MDS network in two resolution settings: LQ and HQ. Table \ref{tab:scores} shows the comparison of our method with the other state-of-the-art methods. It is observed that our method achieves comparable results (LQ: 97.92\% and HQ: 96.87\%) with the top achieving method \cite{Li2018ExposingDVFWA}. In the DFTIMIT test set there are 96 videos in total. This applies that the penalty for mis-classification towards the overall AUC is high in DFTIMIT's case. It is interesting to note that our method mis-classified just 3 video samples in the HQ experiments and 2 videos in the LQ experiments. \cite{Li2018ExposingDVFWA} achieve state-of-the-art results (LQ: 99.9\% and HQ: 99.7\%) on DFTIMIT, however, achieve relatively lower AUC results (72.7\%) on the larger dataset DFDC. In comparison our method achieved 18\% more than \cite{Li2018ExposingDVFWA} on DFDC dataset. This could also mean that the DFTIMIT dataset is now saturated due to smaller size similar to the popular ImageNet dataset \cite{deng2009imagenet}. We also report the frame-wise AUC (LQ: 98.3\% and HQ: 94.7\%) for DFTIMIT as mentioned in brackets in Table \ref{tab:scores}.
\begin{figure}[t]
\centering
\includegraphics[width=8.6cm]{localisation.png} \vspace{6mm}
\includegraphics[width=8.6cm]{localisation_1.png}\vspace{-2mm}
\caption{Forgery Localization results: In the two examples above, the red part of the curve is detected as fake and blue as real by our method. The segment of the video with segment wise dissimilarity above the threshold is labelled as fake and below the threshold is labelled as real. On the x-axis, we have the original segment labels and on the y-axis the segment-level dissimilarity score (Section \ref{sec:localisation}).}
\label{fig:localisation}
\end{figure}
\subsection{Temporal Forgery Localization} \label{sec:localisation}
With the advent of sophisticated forgery techniques, it is possible that an entire video or smaller portions of the video are manipulated to deceive the audience. If in case parts of the original video are corrupted, it would be useful from the FD perspective to be able to locate the timestamps corresponding to the corrupted segments. In an interesting work, Wu \textit{et al.}~\cite{wu2018busternet} proposed a CNN which detects forgery along with the forged locations in images. However, their method is only applicable to copy-paste image forgery. As we process the vidseo by dividing it into temporal segments, a fine-grained analysis of the input video is possible, thereby enabling \textit{forgery localization}. In contrast to the MDS network, earlier techniques \cite{mittal2020emotions,Li2018ExposingDVFWA} computed features over the entire video. We argue that if a forgery has been performed on a small segment of a video, the forgery signatures in that segment may get diluted due to pooling across the whole video. A similar phenomenon is also observed in prior work relating to pain detection and localization \cite{sikka2014classification}. As the pain event could be short and its location is not labeled, the authors divide the video into temporal segments for better pain detection.
Most of the datasets, including the ones used in this paper have data manipulation performed on practically the entire video. To demonstrate the effectiveness of our method for forgery localization in fake videos, we joined segments from real and corresponding fake videos of the same subject at random locations. In Figure \ref{fig:localisation}, we show the outputs on two videos created by mixing video segments of the same subject from the DFDC dataset. Here, the segment-wise score is shown on the $y$-axis. The segments for which the score is above a threshold are assigned as being fake (red on the curve) and below the threshold (blue color on the curve) are assigned are real. In addition to Figs.~\ref{fig:GRADCAM} and~\ref{fig:Graphs}, forgery localization makes the working of the MDS-based fake detection framework more \emph{explainable} and \emph{interpretable}.
\section{Conclusion and Future Work}
We propose a novel bimodal deepfake detection approach based on the modality dissonance score (MDS), which captures the similarity between audio and visual streams for real and fake videos thereby facilitating separability. The MDS is modeled via the contrastive loss computed over segment-level audiovisual features, which constrains genuine audio-visual streams to be \textit{closer} than fake counterparts. Furthermore, cross-entropy loss is enforced on the unimodal streams to ensure that they independently learn discriminative features. Experiments show that (a) the MDS-based FD framework can achieve state-of-the-art performance on the DFDC dataset, and (b) the unimodal cross-entropy losses provide extra benefit on top of the contrastive loss to enhance FD performance. Explainability and interpretability of the proposed approach are demonstrated via audio-visual distance distributions obtained for \emph{real} and \emph{fake} videos, Grad-CAM outputs denoting attention regions of the MDS network, and forgery localization results.
Future work would focus on (a) incorporating human assessments (acquired via EEG and eye-gaze sensing) in addition to content analysis adopted in this work; (b) exploring algorithms such as multiple instance learning for transient forgery detection, and (c) achieving real-time forgery detection (accomplished by online intrusions) given the promise of processing audio-visual information over 1-second segments.
\section{Acknowledgement}
We are grateful to all the brave frontline workers who are working hard during this difficult COVID19 situation.
\bibliographystyle{ACM-Reference-Format}
| c74e984f2b72c9e0cddb9176cda9589cc8164020 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{Deriving repair rules from short-cut rules}
Having defined short-cut rules, they can be operationalized to get edit rules for source graphs and forward rules that repair these edits. As such edits may delete source elements, correspondence elements may be left without corresponding source elements. Hence, the resulting triple graphs show a form of partiality.
They are called \emph{partial triple graphs}.
Given a model, formally considered as triple graph $G_S \xleftarrow{\sigma_G} G_C \xrightarrow{\tau_G} G_T$, a user edit on $G_S$ may consist of the deletion and/or creation of graph elements, resulting in a graph $G_S^{\prime}$.
In general, the \enquote{old} correspondence morphism $\sigma_G: G_C \to G_S$ does not extend to a correspondence morphism from $G_C$ to $G_S^{\prime}$:
The user might have deleted elements in the image of $\sigma_G$.
However, there is
a \emph{partial morphism} $\sigma_G^{\prime}: G_C \dashrightarrow G_S^{\prime}$ that is defined for all elements whose image under $\sigma_G$ still exists.
\begin{definition}[Partial triple graph]
A \emph{partial graph morphism} $f: A \dashrightarrow B$ is a graph morphism $f: A^{\prime} \to B$ where $A^{\prime}$ is a subgraph of $A$; $A^{\prime}$ is called the \emph{domain} of $f$.
A \emph{partial triple graph} $G^{\prime} = G_S^{\prime} \xdashleftarrow{\sigma_G^{\prime}} G_C^{\prime} \xdashrightarrow{\tau_G^{\prime}} G_T^{\prime}$ consists of three graphs $G_S^{\prime},G_C^{\prime},G_T^{\prime}$ and two partial graph morphisms $\sigma_G^{\prime}: G_C^{\prime} \dasharrow G_S^{\prime}$ and $\tau_G^{\prime}: G_C^{\prime} \dasharrow G_T^{\prime}$.
Given a triple graph $G = (G_S \xleftarrow{\sigma_G} G_C \xrightarrow{\tau_G} G_T)$ and a user edit of $G_S$ that results in a graph $G_S^{\prime}$, the \emph{partial triple graph induced by the edit} is $G_S^{\prime} \xdashleftarrow{\sigma_G^{\prime}} G_C \xrightarrow{\tau_G} G_T$ where $\sigma_G^{\prime}$ is obtained by restricting $\sigma_G$ to those elements $x$ of $G_C$ (node or edge) for which $\sigma_G(x) \in G_S$ is still an element of $G_S^{\prime}$.
\end{definition}
According to the above definition, triple graphs are special partial triple graphs, namely those, where the domain of both partial correspondence morphisms is the whole correspondence graph $G_C$.
When operationalizing short-cut rules, i.e., splitting them into a \emph{source} and a \emph{forward rule}, we also have to deal with this kind of partiality:
In contrast to the rules of a given TGG, a short-cut rule might delete an element.
Hence, its forward rule might need to contain a correspondence element for which the corresponding source element is missing; it is referenced in the short-cut rule.
This element is deleted by the corresponding source rule.
\begin{definition}[Source and forward rule of short-cut rule. Repair rule]
Given a pair $(r_1,r_2)$ of plain, monotonic triple rules with short-cut rule $r_{\mathit{sc}} = (L_{\mathit{sc}} \xleftarrow{l_{\mathit{sc}}} K_{\mathit{sc}} \xrightarrow{r_{\mathit{sc}}} R_{\mathit{sc}})$, the \emph{source} and \emph{forward rule} of $r_{\mathit{sc}}$ are defined as
\begin{equation*}
r_{\mathit{sc}}^S \coloneqq (L_{\mathit{sc}}^S \xleftarrow{(l_{\mathit{sc},S}, \mathit{id}_{\emptyset}, \mathit{id}_{\emptyset})} K_{\mathit{sc}}^S \xrightarrow{(r_{\mathit{sc},S}, \mathit{id}_{\emptyset}, \mathit{id}_{\emptyset})} R_{\mathit{sc}}^S)
\end{equation*}
and
\begin{equation*}
r_{\mathit{sc}}^F \coloneqq (L_{\mathit{sc}}^F \xleftarrow{(id_{R_{\mathit{sc},S}}, l_{\mathit{sc},C}, l_{\mathit{sc},T})} K_{\mathit{sc}}^F \xrightarrow{(id_{R_{\mathit{sc},S}}, r_{\mathit{sc},C}, r_{\mathit{sc},T})} R_{\mathit{sc}}^F)
\end{equation*}
where
\begin{align*}
L_{\mathit{sc}}^S & \coloneqq (L_{\mathit{sc},S} \leftarrow \emptyset \rightarrow \emptyset), \\
K_{\mathit{sc}}^S & \coloneqq (K_{\mathit{sc},S} \leftarrow \emptyset \rightarrow \emptyset), \\
R_{\mathit{sc}}^S & \coloneqq (R_{\mathit{sc},S} \leftarrow \emptyset \rightarrow \emptyset), \\
L_{\mathit{sc}}^F & \coloneqq (R_{\mathit{sc},S} \dashleftarrow L_{\mathit{sc},C} \rightarrow L_{\mathit{sc},T}), \\
K_{\mathit{sc}}^F & \coloneqq (R_{\mathit{sc},S} \leftarrow K_{\mathit{sc},C} \rightarrow K_{\mathit{sc},T}), \text{ and } \\
R_{\mathit{sc}}^F & \coloneqq (R_{\mathit{sc},S} \leftarrow R_{\mathit{sc},C} \rightarrow R_{\mathit{sc},T}) \enspace .
\end{align*}
Given a TGG $\mathit{GG}$, a \emph{repair rule for $\mathit{GG}$} is the forward rule $r_{\mathit{sc}}^F$ of a short-cut rule $r_{\mathit{sc}}$ where $r_{\mathit{sc}}$ has been constructed from a pair of rules of $\mathit{GG}$.
\end{definition}
For more details (in particular, the definition of morphisms between partial triple graphs), we refer the interested reader to the literature~\cite{FKST19,KFST19}.
In this paper, we are more interested in conveying the intuition behind these rules by presenting examples.
We next recall the most important property of this operationalization, namely that, as in the monotonic case, an application of a short-cut rule corresponds to the application of its source rule, followed by an application of the forward rule if consistently matched.
\begin{theorem}[{\cite[Theorem~7]{FKST19} and \cite[Theorem~23]{KFST19}}]\label{thm:equivalence}
Given a short-cut rule $r_{\mathit{sc}}$, there is a transformation
\begin{equation*}
(G_S \leftarrow G_C \rightarrow G_T) \Rightarrow_{r_{\mathit{sc}},m_{\mathit{sc}}} (H_S \leftarrow H_C \rightarrow H_T)
\end{equation*}
via this short-cut rule if and only if there is a transformation
\begin{align*}
(G_S \leftarrow G_C \rightarrow G_T) & \Rightarrow_{r_{\mathit{sc}}^S,m_{\mathit{sc}}^S} (H_S \dashleftarrow G_C \rightarrow G_T) \\
& \Rightarrow_{r_{\mathit{sc}}^F,m_{\mathit{sc}}^F} (H_S \leftarrow H_C \rightarrow H_T)
\end{align*}
applying source rule $r_{\mathit{sc}}^S$ with match $m_{\mathit{sc}}^S = (m_{\mathit{sc},S},\emptyset,\emptyset)$ and forward rule $r_{\mathit{sc}}^F$ at match $m_{\mathit{sc}}^F = (n_{\mathit{sc},S},m_{\mathit{sc},C},m_{\mathit{sc},T})$.
\end{theorem}
For practical applications, repair rules should also be equipped with filter NACs.
Let the repair rule $r_{\mathit{sc}}^F$ be obtained from a short-cut rule $r_{\mathit{sc}}$ that has been computed from rule pair $(r_1, r_2)$, both coming from a given TGG.
As the application of $r_{\mathit{sc}}^F$ replaces an application of $r_1^F$ by one of $r_2^F$, $r_{\mathit{sc}}^F$ should be equipped with the filter NAC of $r_2^F$.
However, just copying that filter NAC would not preserve its semantics; a more refined procedure is needed.
The LHS of $r_2^F$ is a subgraph of the one of $r_{\mathit{sc}}^F$ by construction.
There is a known procedure, called \emph{shift along a morphism}, that \enquote{moves} an application condition from a subgraph to the supergraph
preserving its semantics~\cite[Lemma~3.11 and Construction~3.12]{EGHLO14}.
We use this construction to compute the filter NACs of repair rules.
By using this known construction, the filter NACs we construct for our repair rules have the following property:
\begin{lemma}[{\cite[Lemma~3.11 and Construction~3.12]{EGHLO14}.}]\label{lem:property-filter-NACs}
Let $r_{\mathit{sc}}$ be a plain short-cut rule obtained from the pair of monotonic rules $(r_1,r_2)$ where the forward rule $r_2^F$ is equipped with a set $\mathit{NAC}_2^F$ of filter NACs.
Let $\mathit{NAC}_{\mathit{sc}}^F$ be the set of NACs computed by applying the shift construction to $\mathit{NAC}_2^F$ along the inclusion morphism $\iota: L_2^F \hookrightarrow L_{\mathit{sc}}^F$ of the LHS of $r_2^F$ into the LHS of $r_{\mathit{sc}}$ (which exists by construction).
Then, an injective match $m_{\mathit{sc}}^F$ for $r_{\mathit{sc}}^F$ (into any partial triple graph $G$) satisfies the set of NACs $\mathit{NAC}_{\mathit{sc}}^F$ if and only if the induced injective match $m_{\mathit{sc}}^F \circ \iota$ for $r_2^F$ satisfies $\mathit{NAC}_2^F$.
\end{lemma}
\begin{example}
The forward rules of the short-cut rules in Fig.~\ref{fig:scRules} are depicted in Fig.~\ref{fig:fwdSCRules}.
\secondRRule{} is derived to replace an application of \secondTGGForwardRule{} by one of \firstTGGForwardRule{}.
This forward rule is equipped with a filter NAC which ensures that the rule is used only to translate \packages{} at the top of a hierarchy.
Just copying this NAC to the \package{} \p{} in \secondRRule{} would not preserve this behaviour:
The rule would be applicable in situations where the \package{} to which \ssp{} is matched contains a \package{} to which \p{} is matched.
Shifting the NAC from \firstTGGForwardRule{} to \secondRRule{} instead, the forbidden edge between the two \packages{} is introduced in addition.
It ensures that \p{} can be matched to \packages{} at the top of a hierarchy, only.
\fourthRRule{} (see Figure~\ref{fig:fourth-repair-rule}) assumes two connected \packages{} and deletes a \folder{} between their corresponding \folders{} as well as the \doc{} contained in the deleted \folder{} and the correspondence node referencing it.
The LHS of this rule is a proper partial triple graph as there is a correspondence node which is not mapped to any element of the source part.
\begin{figure}%
\includegraphics[width=\columnwidth]{Diagram/OSCRules_4.pdf}%
\caption{Repair rule derived from \fourthSCRule{}}%
\label{fig:fourth-repair-rule}%
\end{figure}
\end{example}
\subsection{Conditions for valid repair rule applications}
Now, we transfer the results obtained so far to the case of repair rules.
To do so, we first define \emph{valid matches} for repair rules (in a restricted kind of transformation sequences).
\begin{definition}[Valid match for repair rule]\label{def:valid-repair-match}
Let a TGG $\mathit{GG}$ and a consistently-marking transformation sequence
\begin{equation}\label{eq:orig-trafo}
G_0 \Rightarrow_{r_1^{\mathit{FN}},m_1^F} G_1 \Rightarrow_{r_2^{\mathit{FN}},m_2^F} \dots \Rightarrow_{r_t^{\mathit{FN}},m_t^F} G_t
\end{equation}
via forward rules $r_i^{\mathit{FN}},\ 1 \leq i \leq t$, (possibly with filter NACs) of $\mathit{GG}$ be given.
Let
\begin{equation*}
G_i = (G_{0,S} \leftarrow G_{i,C} \rightarrow G_{i,T}) \enspace .
\end{equation*}
Let there be some source edit step
\begin{equation*}
G_t \Rightarrow_{r_{\mathit{sc}}^S,m_{\mathit{sc}}^S} G^\prime
\end{equation*}
where $G^\prime = (H_S \dashleftarrow G_{t,C} \rightarrow G_{t,T})$, $r_{\mathit{sc}}$ is a source rule of a short-cut rule derived from a rule pair $(r_j,r)$ where $1 \leq j \leq t$ and $r$ stems from $\mathit{GG}$, and $m_{\mathit{sc}}^S|_{R_{j,S}} = n_{j,S}$, i.e., when restricted to the source part of the RHS $R_j$ of $r_j$ match $m_{\mathit{sc}}^S$ coincides with the source part of the comatch $n_j$.
Moreover, the application of this source edit shall not introduce a violation of any of the filter NACs of $r_1^{\mathit{FN}}, \dots, r_{j-1}^{\mathit{FN}}$.
Then, a match $m_{\mathit{sc}}^F$ for the corresponding forward rule $r_{\mathit{sc}}^F$ in $G^\prime$ is \emph{valid} if the following properties hold.
\begin{enumerate}
\item \emph{Reversing match:} Given comatch $(n_{\mathit{sc},S}^S,\emptyset,\emptyset)$ of the application of the source rule $r_{\mathit{sc}}^S$, its match is
\begin{equation*}
m_{\mathit{sc}}^F = (n_{\mathit{sc},S}^S,m_{\mathit{sc,C}}^F,m_{\mathit{sc,T}}^F)
\end{equation*}
and also $m_{\mathit{sc,C}}^F$ and $m_{\mathit{sc,T}}^F$ coincide with $n_{j,C}$ and $n_{j,T}$ when restricted to $R_{j,C}$ and $R_{j,T}$, respectively.
\item \emph{Sequential independence:}
\begin{enumerate}
\item \emph{Non-disabling match:} The application of $r_{\mathit{sc}}^F$ does not delete elements used in the applications of $r_{j+1}^{\mathit{FN}}, \dots, r_{t}^{\mathit{FN}}$ nor does it create elements forbidden by one of the filter NACs of those forward rules.
\item \emph{Context-preserving match:} The presumed elements of the repair rule $r_{\mathit{sc}}^F$ (which accord to the presumed elements of the short-cut rule $r_{\mathit{sc}}$) are matched to elements of $H_S$ which are marked as translated in $G_{j-1,S}$ and in $G_{t,C}$ and $G_{t,T}$ to elements which are already created in $G_{j-1,C}$ and $G_{j-1,T}$.
This means, elements stemming from the LHS $L$ of $r$ which have not been identified with elements from $L_j$ in the short-cut rule $r_{\mathit{sc}}$, are matched to elements already translated/existing in $G_{j-1}$.
\end{enumerate}
\item \emph{Creation-preserving match:}
All source elements that are newly created by short-cut rule $r_{\mathit{sc}}$, i.e., the source elements of $R_S \setminus L_S$ that have not been merged with an element of $R_{j,S} \setminus L_{j,S}$ during the short-cut rule construction, are matched to elements which are yet untranslated in $G_{t,S}$.
\end{enumerate}
\end{definition}
Note that together, conditions 2. (a) and (b) above constitute sequential independence between the applications of $r_{j+1}^{\mathit{FN}}, \dots, r_{t}^{\mathit{FN}}$ and the one of $r_{\mathit{sc}}^{\mathit{F}}$.
Moreover, the additional requirement on the match to be creation preserving (compared to Theorem~\ref{thm:valid-applications} for short-cut rules) originates from the fact that forward rules do not create but mark source elements.
The following corollary uses Theorem~\ref{thm:equivalence} to transfer the statement of Theorem~\ref{thm:valid-applications} to repair rules.
\begin{corollary}\label{cor:valid-repair-rule-applications}
Let a TGG $\mathit{GG}$ and a consistently marking transformation sequence as in~(\ref{eq:orig-trafo}), followed by an edit step exactly as in Definition~\ref{def:valid-repair-match} above be given.
Then, applying $r_{\mathit{sc}}^F$ at a \emph{valid match} $m_{\mathit{sc}}^F$ in $G^{\prime}$ induces a \emph{consistently marking transformation sequence}
\begin{equation}\label{eq:induced-trafo}
\begin{split}
G_0^{\prime} \Rightarrow_{r_1^{\mathit{FN}},m_1^F} G_1^{\prime} \Rightarrow_{r_2^{\mathit{FN}},m_2^F} \dots & \Rightarrow_{r_{j-1}^{\mathit{FN}},m_{j-1}^F} G_{j-1}^{\prime} \\
& \Rightarrow_{r^{\mathit{FN}},m^F} X
\end{split}
\end{equation}
with $G_i^{\prime} = (H_S \dashleftarrow G_{i,C} \rightarrow G_{i,T})$ for $0 \leq i \leq j-1$.
\end{corollary}
\begin{proof}
For a valid match $m_{\mathit{sc}}^F$ of $r_{\mathit{sc}}^F$, by its \emph{reversing property}, the conditions of Theorem~\ref{thm:equivalence} are met.
Hence, we obtain a sequence
\begin{equation*}
G_0 \Rightarrow_{r_1^{\mathit{FN}},m_1^F} \dots \Rightarrow_{r_t^{\mathit{FN}},m_t^F} G_t \Rightarrow_{r_{\mathit{sc}},m_{\mathit{sc}}} X^\prime \enspace .
\end{equation*}
As a consistently marking sequence of forward rules corresponds to a sequence of TGG rule applications, and the preconditions of Theorem~\ref{thm:valid-applications} are met (\enquote{exists} is exchanged by \enquote{marked} on the source component), this sequence induces a sequence
\begin{equation*}
G_0 \Rightarrow_{r_1^{\mathit{FN}},m_1^F} \dots \Rightarrow_{r_{j-1}^{\mathit{FN}},m_{j-1}^F} G_{j-1} \Rightarrow_{r,m} X
\end{equation*}
(where we do not care for the further applications of forward rules).
Now, we can split $r$ into its source and forward rule.
Its source rule is sequentially independent from the other forward rule applications:
$r_{\mathit{sc}}^S$ does not delete anything, the rules $r_1^{\mathit{FN}}, \dots, r_{j-1}^{\mathit{FN}}$ match, and does not create a filter NAC violation by assumption and, as a consequence, $r^S$ does not.
Hence, by the local Church-Rosser Theorem, we might equivalently switch the application of $r^S$ to the beginning of the sequence and obtain sequence~(\ref{eq:induced-trafo}), as desired.
Moreover, by Lemma~\ref{lem:property-filter-NACs}, the filter NAC of $r^F$ holds whenever $m_{\mathit{sc}}^F$ satisfies the filter NAC of $r_{\mathit{sc}}^F$.
Finally, as the start of the transformation sequence (up to index $j-1$) is context preserving, and by assumption 2. (b), the match $m_{\mathit{sc}}^F$ matches presumed elements of $r_{\mathit{sc}}^F$ to already translated ones (in $H_S$) or already created ones (in $G_{j-1,C}$ and $G_{j-1,T}$), this sequence is context preserving.
Analogously, assumption 3. ensures that it is creation-preserving:
No element which is already marked as translated in $G_{t,S}$ is marked a second time.
Hence, the whole sequence is consistently marking. \qed
\end{proof}
\subsection{Performance Evaluation}
To get realistic models, we extracted five models from Java projects hosted on Github using the reverse engineering tool MoDisco~\cite{HB14} and translated them into our own documentation structure.
In addition, we generated five synthetic models consisting of $n$-level \package{} hierarchies with each non-leaf \package{} containing five sub-\packages{} and each leaf \package{} containing five \classes{}.
While the realistic models shall show that our approach scales to real world cases, the synthetic models are chosen to show scalability in a more controlled way by increasing hierarchies gradually.
To evaluate our synchronization process, we performed several model changes.
We refactored each of the models in four different scenarios; two example refactorings are the moving of a \class{} from one \package{} to another or the complete relocation of a \package{}.
Then we used eMoflon to synchronize these changes in order to restore consistency to the documentation model using two synchronization processes, namely with and without \repairRules{}.
The legacy synchronization process of eMoflon is presented in~\cite{LAFVS17,Leblebici18}; the new synchronization process applying additional repair rules takes place according to the algorithm presented in Sect.~\ref{sec:syncProcess} with the extensions mentioned in Sect.~\ref{sec:prospect}.
These synchronization steps are subject to our evaluation and we pose the following research questions:
\textbf{(RQ1)} {\em For different kinds of model changes, how many elements can be preserved that would be deleted and recreated otherwise?}
\textbf{(RQ2)} {\em How does our new synchronization process affect the runtime performance?}
\textbf{(RQ3)} {\em Are there specific scenarios in which our new synchronization process performs especially good or bad?}
In the following, we evaluate our new {\em synchronization process by repair rules} against the {\em legacy synchronization process} in eMoflon. While the legacy one revokes forward rule applications and re-propagates the source model using forward rules, our new one prefers to apply short-cut repair rules as far as possible and falls back to revoking and re-propagation if there is no possible repair rule application.
To evaluate the performance of the legacy and the new model synchronization processes, we consider the following synchronization scenarios:
Altering a root \package{} by creating a new \package{} as root would imply that many rule applications have to be reverted to synchronize the changes correctly with the legacy synchronization process (Scenario 1).
In contrast, our new approach might perform poorly when a model change does not inflict a large cascade of invalid rule applications.
Hence, we move \classes{} between \packages{} (Scenario 3) and \methods{} between \classes{} (Scenario 4) to measure if the effort of applying \repairRules{} does infer a performance loss when both, the new and old algorithm, do not have to repair many broken rule applications.
Note that Scenario 4 extends our evaluation presented in \cite{FKST19} as it provides a more fine-granular scenario.
Finally, we simulate a scenario which is somewhat between the first three by relocating leaf \packages{} (Scenario 2) which, using the legacy model synchronization, would lead to a re-translation of all underlying elements.
\begin{table*}
\centering
\caption{Legacy synchronizer -- Time in sec. and number of created elements}
\label{tbl:measurements_legacy}
\begin{tabular}{@{}l@{\hskip 7.5pt}*{15}{l}@{}}
\toprule
& \multicolumn{2}{c}{Both} & \phantom{a} & \multicolumn{11}{c}{Legacy Synchronization} \\
\cmidrule{2-3} \cmidrule{5-15}
& \multicolumn{2}{c}{Trans.} & \phantom{i} & \multicolumn{2}{c}{Scen. 1} & \phantom{i} & \multicolumn{2}{c}{Scen. 2} & \phantom{i} & \multicolumn{2}{c}{Scen. 3} & \phantom{i} & \multicolumn{2}{c}{Scen. 4}\\
\cmidrule{2-3} \cmidrule{5-6} \cmidrule{8-9} \cmidrule{11-12} \cmidrule{14-15}
Models & Sec & Elts & & Sec & Elts & & Sec & Elts & & Sec & Elts & & Sec & Elts \\
\midrule
lang.List & 0.3 & 25 & & 0.2 & 20 & & -- & -- & & 0.06 & 5 & & 0.04 & 3 \\
tgg.core & 6.4 & 1.6k & & 39 & 1.6k & & 3.8 & 99 & & 0.64 & 17 & & 0.2 & 3 \\
modisco.java & 9.9 & 3.2k & & 228 & 3.3k & & 18.6 & 192 & & 3.6 & 33 & & 0.4 & 4 \\
eclipse.compare & 10.74 & 3.8k & & 83 & 3.7k & & 3.1 & 76 & & 2.36 & 47 & & 0.1 & 1 \\
eclipse.graphiti & 20.7 & 6.5k & & 704 & 6.5k & & 63.9 & 490 & & 5.65 & 25 & & 0.9 & 3 \\
\midrule
synthetic $n=1$ & 0.6 & 89 & & 0.5 & 84 & & 0.2 & 21 & & 0.07 & 5 & & 0.03 & 1 \\
synthetic $n=2$ & 1.4 & 345 & & 1.7 & 340 & & 0.2 & 21 & & 0.11 & 5 & & 0.04 & 1 \\
synthetic $n=3$ & 3.5 & 1369 & & 13.2 & 1364 & & 0.3 & 21 & & 0.11 & 5 & & 0.07 & 1 \\
synthetic $n=4$ & 14.5 & 5.5k & & 141.5 & 5.5k & & 1 & 21 & & 0.32 & 5 & & 0.09 & 1 \\
synthetic $n=5$ & 58.5 & 22k & & 2863 & 22k & & 10.7 & 21 & & 1.07 & 5 & & 0.23 & 1 \\
\bottomrule
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{New synchronizer -- Time in sec. and number of created elements}
\label{tbl:measurements_sc}
\begin{tabular}{@{}lllllllllllll@{}}
\toprule
& \multicolumn{11}{c}{Synchronization by Repair Rules} \\
\cmidrule{2-12}
& \multicolumn{2}{c}{Scen. 1} & \phantom{i} & \multicolumn{2}{c}{Scen. 2} & \phantom{i} & \multicolumn{2}{c}{Scen. 3} & \phantom{i} & \multicolumn{2}{c}{Scen. 4}\\
\cmidrule{2-3} \cmidrule{5-6} \cmidrule{8-9} \cmidrule{11-12}
Models & Sec & Elts & & Sec & Elts & & Sec & Elts & & Sec & Elts \\
\midrule
lang.List & 0.2 & 0 & & -- & -- & & 0.03 & 0 & & 0.02 & 0 \\
tgg.core & 0.8 & 0 & & 0.11 & 0 & & 0.05 & 0 & & 0.04 & 0 \\
modisco.java & 2.5 & 0 & & 0.2 & 0 & & 0.09 & 0 & & 0.1 & 0 \\
eclipse.compare & 0.7 & 0 & & 0.08 & 0 & & 0.04 & 0 & & 0.03 & 0 \\
eclipse.graphiti & 6.1 & 0 & & 0.21 & 0 & & 0.09 & 0 & & 0.1 & 0 \\
\midrule
synthetic $n=1$ & 0.1 & 0 & & 0.05 & 0 & & 0.03 & 0 & & 0.05 & 0 \\
synthetic $n=2$ & 0.1 & 0 & & 0.05 & 0 & & 0.02 & 0 & & 0.04 & 0 \\
synthetic $n=3$ & 0.1 & 0 & & 0.07 & 0 & & 0.02 & 0 & & 0.04 & 0 \\
synthetic $n=4$ & 0.4 & 0 & & 0.14 & 0 & & 0.04 & 0 & & 0.04 & 0 \\
synthetic $n=5$ & 1.5 & 0 & & 0.37 & 0 & & 0.09 & 0 & & 0.06 & 0 \\
\bottomrule
\end{tabular}
\end{table*}
Tables~\ref{tbl:measurements_legacy} and \ref{tbl:measurements_sc} depict the measured time in seconds (Sec) and the number of re-/created elements (Elts) in each scenario (1)--(4).
The first table additionally shows measurements for the initial translation (Trans.) of the Java AST model into the documentation structure.
For each scenario, Table~\ref{tbl:measurements_legacy} shows the numbers of synchronization steps using the legacy synchronizer without \repairRules{} while Table~\ref{tbl:measurements_sc} reflects the numbers of our new synchronizer with \repairRules{}.
W.r.t. our research questions stated above, we interpret these tables as follows:
The Elts columns of Table~\ref{tbl:measurements_sc} show clearly that using repair rules preserves all those elements in our scenarios that are deleted and recreated by the legacy algorithm otherwise as shown in Table~\ref{tbl:measurements_legacy} \textbf{(RQ1)}.
The runtime shows a significant performance gain for Scenario~1 including a worst-case model change in which the legacy algorithm has to re-translate all elements \textbf{(RQ2)}.
{\em Repair rules} do not introduce an overhead compared to the legacy algorithm as can be seen for the synthetic time measurements in Scenario 4 where only one rule application has to be repaired or reapplied \textbf{(RQ2)}.
Our new approach excels when the cascade of invalidated rule applications is long. Even if this is not the case, it does not introduce any measurable overhead compared to the legacy algorithm as shown in Scenarios 2, 3, and 4 (\textbf{RQ3}).
\paragraph{Threats to validity.}
Our evaluation is based on five real world and five synthetic models.
Of course, there exists a wide range of Java projects that differ significantly from each other w.r.t. their size, purpose, and developer style.
Thus, the results may not be transferable to other projects.
Nonetheless, we argue that the four larger models extracted from Github projects are representative since they are deduced from established tools of the Eclipse ecosystem.
The synthetic models are also representative as they show the scalability of our approach in a more controlled environment with an increasing scaling factor.
Together, realistic and synthetic models show that our approach does not only increase the performance of eMoflons synchronization process but also reduce the amount of re-created elements.
Since each re-created element may contain information that would be lost during the process, we preserve this information and increase the overall quality of eMoflons synchronization results.
In this evaluation, we selected four edit operations that are representative w.r.t. their dependency on other edit operations.
They may not be representative w.r.t. other aspects such as size or kind of change. We consider those aspects to be of minor importance in this context as dependency is the cause for deleting and recreating elements in the legacy synchronization process.
Finally, we limited our evaluation to one TGG rule set only as we experienced similar results for a broader range of TGGs from the eMoflon test zoo\footnote{Accessible via \url{https://github.com/eMoflon/emoflon-ibex-tests}}.
\subsection{Refactorings}
\label{sec:refactorings}
As explained in Sect.~\ref{sec:implementation}, we currently employ two different strategies to overlap two rules and to create a short-cut rule.
We pose the following research question: \textbf{(RQ4)} {\em Are the generated \shortcutRules{} applicable to realistic scenarios? Are further \shortcutRules{} necessary?}
Since our example addresses code changes that are incorporated by the Java AST model primarily, we relate our approach to available code refactorings.
In the following, we refer to the book on code refactorings written by Martin Fowler~\cite{Fowler2018} which presents 66 refactorings.
Our example TGG, depicted in Fig.~\ref{fig:eval_rule_set_full}, defines consistency on a structural level solely, without incorporating behaviour, i.e., the bodies of methods and constructors.
Hence, we selected those refactorings that describe changes on \packages{}, \classes{} and \interfaces{}, {\em MethodDeclarations} and \parameters{}, and \fields{}.
The result is a set of 16 refactorings for which we evaluated if \shortcutRules{} help to directly propagate the corresponding change of the AST model or deletion and recreation has to take place.
\begin{figure*}
\centering
\includegraphics[width=.9\textwidth]{Diagram/Refactorings.pdf}
\caption{Refactorings}
\label{fig:refactoring}
\end{figure*}
Fig.~\ref{fig:refactoring} lists these refactorings together with information on the TGG rules and/or \shortcutRules{} that are applicable in these scenarios.
For some of the refactorings as e.g., {\em Extract Class} and \emph{Push-Down Field}, we identified situations where not only \shortcutRules{} are necessary to propagate the changes.
In these cases, new elements may be created which can be propagated using operationalized TGG rules.
The deletion of elements can be propagated by revoking the corresponding prior propagation step.
However, many refactorings benefit from using \shortcutRules{}, for example, those that move methods and fields.
If recreation of documentation on the target part is necessary, it can lead to information loss as there may not be all the necessary information in the Java AST model.
\textit{Example:} \emph{Push-Up Field} moves and merges a similar field from various subclasses into a common superclass.
If one of the subclass fields is moved to the superclass, we can propagate this change using \emph{Move-Field-Repair-Rule}, which is depicted in Fig.~\ref{fig:move_field_rule}.
\begin{figure}
\includegraphics[width=\columnwidth]{Diagram/Move-Field.pdf}
\caption{Move-Field-Repair-Rule}
\label{fig:move_field_rule}
\end{figure}
In summary, we are able to solve all 16 refactorings using a combination of (inverse) TGG rules and our generated \shortcutRules{} \textbf{(RQ4)}.
\paragraph{Threats to validity.}
Note that \shortcutRules{} are especially useful when elements are moved instead of deleting and recreating them in some other location.
Those changes are hard to detect and are not covered here.
Refactorings such as {\em Push-Up Method}, which moves a method that occurs in several subclasses to their common superclass, can be done in two different ways.
First, one of the methods is moved to the superclass while the methods in the other subclasses are deleted.
This employs the use of \shortcutRules{} for the moved method followed by revocation steps for the deleted methods to delete the corresponding documentation elements.
Second, all methods may be deleted and a new similar method is created in the superclass.
In that case, there is no \shortcutRule{} that helps to preserve information and all propagated documentation elements for the method will be blank.
Hence, our approach depends on the kind of change. In particular, it helps when user edits also try to preserve information instead of recreating them.
In addition, we have not incorporated behaviour in our example; such an extension of our TGG may be considered in future work.
However, we can argue that most of those refactorings can be reduced to the movement of elements, the deletion of superfluous elements and the creation of new elements.
These changes are manageable in general using a sequence of \shortcutRule{} and (inverse) operationalized TGG rule applications.
Finally, we evaluated these cases by hand based on the generated \shortcutRules{} from our implementation.
Test cases implementing the identified refactorings and combinations of them will be accessible via eMoflons test zoo.
\subsection{Tool architecture}
Figure~\ref{fig:architecture} depicts a UML component diagram to show the main components of eMoflon's bidirectional transformation engine.
The architecture has two main components:
\textit{TGG Core} contains the core components of eMoflon and \textit{Repair Framework} adds (short-cut) repair rules to eMoflon\rq{}s functionality.
The \emph{TGG engine} manages the synchronization process and alters source, target, and correspondence model in order to restore consistency.
For this purpose, it applies for\-ward/ back\-ward \textit{operationalized TGG rules} to translate elements or revokes broken rule applications.
Finding matches in an incremental way is an important requirement for efficient model synchronization since minor model changes should be detectable without re-evaluating the whole model.
For this reason, eMoflon relies on \textit{incremental pattern matching} to detect the appearance of new matches as well as the disappearance of formerly detected ones.
It uses different incremental pattern matchers such as Democles~\cite{VD13} and HiPE~\cite{hipe} and allows to switch freely between them for optimizing the performance for each transformation scenario.
Furthermore, eMoflon employs the use of various \textit{integer linear programming} (ILP) solvers such as Gurobi~\cite{Gurobi16} and CPLEX~\cite{CPLEX}, e.g., in order to find correspondence links (mappings) between source and target models, which is referred to as consistency check~\cite{leblebici2017inter}.
We have extended this basic setup by introducing the \emph{Repair Framework}, which consists of the \emph{Repair Strategy} and the \emph{Shortcut Rule Creator}.
The \emph{Repair Strategy} is attached to the \textit{TGG Engine} from which it is called with a set of broken rule matches.
It attempts to repair the corresponding rule applications by using repair rules created by the \emph{Shortcut Rule Creator}, which uses the ILP interface provided by the \emph{TGG Core} in order to find overlaps between TGG rules and finally, to create short-cut repair rules.
For invoking the repair rules, however, we have to find matches of repair rules.
This is done by a \textit{Batch (local-search) Pattern Matcher} which, in contrast to the incremental pattern matcher, does not perform any book-keeping.
As a repair of a rule application is always done locally, the checking of matches throughout the whole model is considered to be too expensive and thus, a \emph{Batch Pattern Matcher} can perform this task more efficiently.
\subsection{ILP-based short-cut rule creation}
\label{sec:sc-rule-creation}
In order to create an overlap between two rules, a morphism between the graphs of both rules has to be found:
Each element may only be mapped once; a context element may only be mapped to another context element. Created elements are mapped to each other, respectively.
Furthermore, a node can only be mapped to a node of the same type as we do not incorporate inheritance between types yet.
Edges are allowed to be mapped to each other only if their corresponding source and target elements are also mapped to each other, respectively.
We use integer linear programming (ILP) to encode the search space of all possible mappings and search for a maximal mapping.
Each possible mapping $m$ is considered to be a variable of our ILP problem such that calculating
\begin{equation*}
max (\sum_{m \in M} m)
\end{equation*}
yields the maximal overlap, with $ M $ being the set of all mappings and $ m \in \{0, 1\} $.
To ensure that each element $e$ is mapped only once, we define a constraint to exclude non-used mappings: $ (\sum_{m \in A_e} m) \leqslant 1 $ with $ A_e $ being the set of all alternative mappings for element $e$.
To ensure that edges are mapped only if their adjacent nodes are mapped as well, we define the following constraint: $ m_e \implies m_v $ which translates to $ m_e \leq m_v $ with $ m_e $ being the edge mapping and $ m_v $ being one of the mappings of node $src(e)$ or $trg(e)$.
Maximizing the number of activated variables yields the common kernel of both input rules, i.e., a maximal overlap between them.
If the overlap between the created elements of both rules is empty, we drop this overlap as the resulting short-cut rule would not preserve any elements.
Given a common kernel of two rules, we glue them along this kernel and yield a short-cut rule.
For all elements of the resulting short-cut rule, which are not in the common kernel, we do the following: (1) Preserved elements remain preserved in the short-cut rule. (2) Created elements of the first rule become deleted ones as the first rule is inverted. (3) Created elements of the second rule remain created ones.
We calculate two kinds of overlap for each pair of rules and hence, two short-cut rules: a maximal and a minimal overlap.
The maximal overlap is calculated by allowing mappings between all created and context elements, respectively.
On the other hand, the minimal overlap is created by allowing mappings between created elements only.
Considering the corresponding ILP problem, this means that all other mapping candidates are dropped.
Finally, the derived short-cut rules are operationalized to obtain the repair rules employed in our synchronization algorithm.
\subsection{Attribute Constraints}
Although attribute constraints have not been incorporated formally in our approach, eMoflon is able to define and solve those within the former legacy translation and synchronization process.
As can be seen in Fig.~\ref{fig:eval_rule_set_full}, many rules have an equality constraint defined between the name attributes of created elements on both, source and target parts.
For TGG rules, this means that the attribute values may be chosen arbitrarily since both nodes would be created from scratch.
In forward rules, source elements are already present which means that an attribute constraint can be interpreted as to propagate or copy the already present value to a newly created element.
We reuse this functionality for our new synchronization process in the following a way:
After applying a repair rule, we ensure that the constraints of the replacing rule are fulfilled.
The definition of attribute constraints and their treatment is due to Anjorin et al.~\cite{AVS12}.\footnote{This approach allows to specify constraints on attributes that involve also operations which are not only equality checks such as the concatenation of values of type \emph{String}.}
\section{Introduction}
\label{sec:intro}
\input{introduction-long}
\section{Informal Introduction to TGG-Based Model Synchronization}
\label{sec:example}
\input{example}
\section{Preliminaries: Triple Graphs, Triple Graph Grammars and their Operationalizations}
\label{sec:preliminaries}
\input{preliminaries}
\section{Short-cut Rules}
\label{sec:sc-rules}
\input{sc-rules}
\section{Constructing Language-Preserving Repair Rules}
\label{sec:constructing-repair-rules}
\input{constructing-repair-rules}
\section{Synchronization Algorithm}
\label{sec:syncProcess}
\input{syncProcess}
\section{Implementation}
\label{sec:implementation}
\input{implementation}
\section{Evaluation}
\label{sec:implAndEvaluation}
\input{evaluation-long}
\section{Related Work}
\label{sec:related-work}
\input{related-work}
\section{Conclusion}
\label{sec:conclusion}
\input{conclusion}
\subsubsection*{Acknowledgments}
This work was partially funded by the German Research Foundation (DFG) project \enquote{Triple Graph Grammars (TGG)~2.0}. \\
We would also like to thank the anonymous reviewers for their thoughtful comments and efforts.
\bibliographystyle{spmpsci}
\subsection{Graphs, triple graphs, and their transformations}
\emph{Graphs} and their (rule-based) \emph{transformations} are suitable to formalize various kinds of models and their evolution, in particular of EMF models~\cite{BET12}.\footnote{Therefore in this paper, we use the terms \emph{graph} and \emph{model} interchangeably. In the formal parts, we will consequently speak of graphs following the formal literature.}
In the context of this work, a \emph{graph} consists of a set of nodes and a set of directed edges which connect nodes.
Graphs may be related by \emph{graph morphisms}, and a \emph{triple graph} consists of three graphs connected by two graph morphisms.
\begin{definition}[Graph, graph morphism, triple graph, and triple graph morphism]\label{def:graphs}
A \emph{graph} $G = (V,E,s,t)$ consists of a set $V$ of vertices, a set $E$ of edges, and source and target functions $s,t: E \rightarrow V$.
An {\em element} $x$ of $G$ is a node or an edge, i.e., $x \in V$ or $x \in E$.
A \emph{graph morphism} $f:G \rightarrow H$ between graphs $G = (V_G, E_G, s_G, t_G)$ and $H = (V_H, E_H, s_H, t_H)$ consists of two functions $f_V: V_G \rightarrow V_H$ and $f_E: E_G \rightarrow E_H$ that are compatible with the assignment of source and target to edges, i.e., $f_V \circ s_G = s_H \circ f_E$ and $f_V \circ t_G = t_H \circ f_E$.
Given a fixed graph $\mathit{TG}$, a \emph{graph typed over $\mathit{TG}$} is a graph $G$ together with a graph morphism $\mathit{type}_G: G \to \mathit{TG}$.
A \emph{typed graph morphism} $f: (G, \mathit{type}_G) \to (H, \mathit{type}_H)$ between typed graphs is a graph morphism $f: G \to H$ that respects the typing, i.e., $\mathit{type}_G = \mathit{type}_H \circ f$ (componentwise).
A (typed) graph morphism $f = (f_V,f_E)$ is \emph{injective} if both $f_V$ and $f_E$ are.
A \emph{triple graph} $G = (G_S \xleftarrow{\sigma_G} G_C \xrightarrow{\tau_G} G_T)$ consists of three graphs $G_S, G_C, G_T$, called \emph{source, correspondence}, and \emph{target graph}, and two graph morphisms $\sigma_G: G_C \rightarrow G_S$ and $\tau_G: G_C \rightarrow G_T$, called \emph{source} and \emph{target correspondence morphism}.
A \emph{triple graph morphism} $f: G \rightarrow H$ between two triple graphs $G$ and $H$ consists of three graph morphisms $f_S: G_S \rightarrow H_S, f_C: G_C \rightarrow H_C$ and $f_T: G_T \rightarrow H_T$ such that $\sigma_H \circ f_C = f_S \circ \sigma_G$ and $\tau_H \circ f_C = f_T \circ \tau_G$.
Given a fixed triple graph $\mathit{TG}$, a \emph{triple graph typed over $\mathit{TG}$} is a triple graph $G$ together with a triple graph morphism $\mathit{type}_G: G \to \mathit{TG}$.
Again, \emph{typed triple graph morphisms} are triple graph morphisms that respect the typing.
A (typed) triple graph morphism $f = (f_S,f_C,f_T)$ is \emph{injective} if $f_S, f_C$, and $f_T$ all are.
\end{definition}
\begin{example}
Figure~\ref{fig:translationExample} depicts three triple graphs; their common type graph is depicted in Fig.~\ref{fig:typegraph}.
The typing morphism is indicated by annotating the elements of the triple graphs with the types to which they are mapped in the type graph.
The nodes in the triple graphs are of types \package{}, \folder{}, \class{}, and \doc{}.
In each case, the source graph is depicted to the left and the target graph to the right.
The hexagons in the middle constitute the correspondence graphs.
Formally, the edges from the correspondence graphs to source and target graphs are morphisms:
The edges encode how an individual correspondence node is mapped by the correspondence morphisms.
For example, the nodes \rootP{} and \rootF{} of types \package{} and \folder{} correspond to each other as they share the same correspondence node as preimage under the correspondence morphisms.
\end{example}
Rules offer a declarative means to specify transformations of (triple) graphs.
While classically rewriting of triple graphs has been performed using non-deleting rules only, we define a less restricted notion of rules\footnote{As used in double pushout rewriting of graphs or objects of other adhesive categories more generally~\cite{EEPT06,LS05}.} right away since short-cut rules and repair rules derived from them are both potentially deleting.
A rule $p$ consists of three triple graphs, namely a \emph{left-hand side} (LHS) $L$ and a \emph{right-hand side} (RHS) $R$ and an \emph{interface} $K$ between them.
Applying such a rule to a triple graph $G$ means to choose an injective morphism $m$ from $L$ to $G$.
The elements from $m(L\setminus l(K))$ are to be deleted; if this results in a triple graph again, the morphism $m$ is called a match and $p$ is applicable at that match.
After this deletion, the elements from $R \setminus r(K)$ are added; the whole process of applying a rule is also called a \emph{transformation (step)}.
\begin{definition}[Rule, transformation (step)]
A \emph{rule} $p = (L \xleftarrow{l} K \xrightarrow{r} R)$ consists of three triple graphs, $L$, $R$, and $K$, called the \emph{left-hand side}, \emph{right-hand side}, and \emph{interface}, respectively, and two injective triple graph morphisms $l: K \to L$ and $r: K \to R$.
A rule is called \emph{monotonic}, or \emph{non-deleting}, if $l$ is an isomorphism.
In this case we denote the rule as $r: L \to R$.
The \emph{inverse rule} of a rule $p$ is the rule $p^{-1} = (R \xleftarrow{r} K \xrightarrow{l} L)$.
Given a triple graph $G$, a rule $p = (L \xleftarrow{l} K \xrightarrow{r} R)$, and an injective triple graph morphism $m: L \to G$, the rule $p$ is \emph{applicable at $m$} if
\begin{equation*}
D \coloneqq G \setminus (m(L \setminus l(K))) \enspace ,
\end{equation*}
is a triple graph again. Operator $\setminus$ is understood as node- and edge-wise set-theoretic difference. The source and target functions of $D$ are restricted accordingly.
If $D$ is a triple graph,
\begin{equation*}
H \coloneqq D \cup n(R \setminus r(K)) \enspace ,
\end{equation*}
is computed. Operator $\cup$ is understood as node- and edge-wise set-theoretic union.
$n(R \setminus r(K))$ is a new copy of newly created elements. $n$ can be extended to $R$ by $n(r(K)) = m(l(K))$.
The values of the source and target functions for edges from $n(R \setminus r(K))$ with source or target node in $K$ are determined by $m \circ l$, i.e.,
\begin{align*}
s_H(e) & \coloneqq m(l(r^{-1}(s_R(e)))) \\
t_H(e) & \coloneqq m(l(r^{-1}(t_R(e))))
\end{align*}
for such edges $e \in n(E_R)$ with $s_R(e) \in r_V(V_K)$ or $t_R(e) \in r_V(V_K)$.
The whole computation is called a \emph{transformation (step)}, denoted as $G \Rightarrow_{p,m} H$ or just $G \Rightarrow H$, $m$ is called a \emph{match}, $n$ is called a \emph{comatch} and $D$ is the \emph{context triple graph} of the transformation.
\end{definition}
An equivalent definition based on computing two \emph{pushouts}, a notion from category theory generalizing the union of sets along a common subset, serves as basis when developing a formal theory~\cite{EEPT06}.
In the following and in our examples, we always assume $K$ to be a common subgraph of $L$ and $R$ and the injective morphisms $l$ and $r$ to be the corresponding inclusions; this significantly eases the used notation.
When we talk about the union of two graphs $G_1$ and $G_2$ along a common subgraph $S$, we assume that $G_1 \cap G_2 = S$.
To enhance expressiveness, a rule may contain \emph{negative application conditions} (NACs)~\cite{EEPT06}.
A NAC extends the LHS of a rule with a forbidden pattern:
A rule is allowed to be applied only at matches which cannot be extended to any pattern forbidden by one of its NACs.
If we want to stress that a rule is not equipped with NACs, we call it a \emph{plain rule}.
\begin{definition}[Negative application conditions]
Given a rule $p = (L \leftarrow K \rightarrow R)$, a set of \emph{negative application conditions} (NACs) for $p$ is a finite set of graphs $\mathit{NAC} = \{N_1, \dots, N_k\}$ such that $L$ is a subgraph of every one of them, i.e., $L \subset N_i$ for $1 \leq i \leq k$.
A rule $(p = (L \leftarrow K \rightarrow R),\mathit{NAC})$ with NACs is applicable at a match $m: L \rightarrow G$ if the plain rule $p$ is and, moreover, for none of the NACs $N_i$ there exists an injective morphism $x_i: N_i \to G$ such that $x_i \circ \iota_i = m$ where $\iota_i: L \hookrightarrow N_i$ is the inclusion of $L$ into $N_i$.
\end{definition}
\begin{example}
Different sets of triple rules are depicted in Figs.~\ref{fig:tggRules}, \ref{fig:tggFwdRules}, \ref{fig:scRules}, and \ref{fig:fwdSCRules}.
All rules in these figures are presented in an \emph{integrated} form:
Instead of displaying LHS, RHS, and the interface as three separate graphs, just one graph is presented where the different roles of the elements are displayed using markings (and color).
The unmarked (black) elements constitute the interface of the rule, i.e., the context that has to be present to apply a rule.
Unmarked elements and elements marked with $(--)$ (black and red elements) form the LHS while unmarked elements and elements marked with $(++)$ (black and green elements) constitute the RHS.
Elements marked with (nac) (blue elements) extend the LHS to a NAC; different NACs for the same rule are distinguished using names.
As triple rules are depicted, their LHSs and RHSs are triple graphs themselves.
For example, the LHS $L$ of \secondTGGRule{} (Fig.~\ref{fig:tggRules}) consists of the nodes \ssp{} and \ssf{} of types \package{} and \folder{} and the correspondence node in between.
While, e.g., all rules in Fig.~\ref{fig:tggRules} are monotonic, \secondSCRule{} is not as it deletes edges and a \doc{}.
Applying \secondSCRule{} to the triple graph (a) in Fig.~\ref{fig:translationExample} leads to the triple graph (c), when \package{}-nodes \ssp{} and \p{} (of the rule) are matched to \rootP{} and \subP{} (in the graph), respectively. (The \folders{} on the target part are mapped accordingly.)
The rules \firstSCRule{} and \secondSCRule{} are inverse to each other.
Finally, \firstTGGForwardRule{} (Fig.~\ref{fig:tggFwdRules}) depicts a rule that is equipped with a NAC:
It is applicable only at \packages{} that are not referenced by other \packages{}.
This means that it is applicable at node \subP{} in the triple graph (b) depicted in Fig.~\ref{fig:translationExample}, but not at node \leafP{}.
\end{example}
\subsection{Triple graph grammars and their operationalization}
Sets of triple graph rules can be used to define languages.
\begin{definition}[Triple graph grammar]
A \emph{triple graph grammar} (TGG) $\mathit{GG} = (\mathcal{R},S)$ consists of a set of plain, monotonic triple rules $\mathcal{R}$ and a start triple graph $S$.
In case of typing, all rules of $\mathcal{R}$ and $S$ are typed over the same triple graph.
The language of a TGG $\mathit{GG}$, denoted as $\mathcal{L}(\mathit{GG})$, is the reflexive and transitive closure of the relation induced by transformation steps via rules from $\mathcal{R}$, i.e.,
\begin{equation*}
\mathcal{L}(\mathit{GG}) \coloneqq \{H \, | \, S \Rightarrow_{\mathcal{R}}^* H\}
\end{equation*}
where $\Rightarrow_{\mathcal{R}}^*$ denotes a finite sequence of transformation steps where each rule stems from $\mathcal{R}$.
The \emph{projection of the language of a TGG to its source part} is the set
\begin{equation*}
\mathcal{L}_S(\mathit{GG}) \coloneqq \{ G_S \, | \, G = (G_S \leftarrow G_C \rightarrow G_T) \in \mathcal{L}(\mathit{GG})\} \enspace ,
\end{equation*}
i.e., it consists of the source graphs of the triple graphs of $\mathcal{L}(\mathit{GG})$.
\end{definition}
In applications, quite frequently, the start triple graph of a TGG is just the empty triple graph.
We use $\emptyset$ to denote the empty graph, the empty triple graph, and morphisms starting from the empty (triple) graph; it will always be clear from the context what is meant.
To enhance expressiveness of TGGs, their rules can be extended with NACs or with some attribution concept for the elements of generated triple graphs.
A recent overview of such concepts and their expressiveness can be found in~\cite{WOR19}.
In the following, we first restrict ourselves to TGGs that contain plain rules only and discuss extensions of our approach subsequently.
\begin{example}
The rule set depicted in Fig.~\ref{fig:tggRules}, together with the empty triple graph as start graph, constitutes a TGG.
The triple graphs (a) and (c) in Fig.~\ref{fig:translationExample} are elements of the language defined by that grammar while the triple graph (b) is not.
\end{example}
The operationalization of triple graph rules into \emph{source} and \emph{forward} (or, analogously, into \emph{target} and \emph{backward}) \emph{rules} is central to working with TGGs.
Given a rule, its source rule performs the rule\rq{}s actions on the source graph only while its forward rule propagates these to correspondence and target graph.
This means that, for example, source rules can be used to generate the source graph of a triple graph while forward rules are then used to translate the source graph to correspondence and target side such that the result is a triple graph in the language of the TGG.
Classically, this operationalization is defined for monotonic rules only~\cite{Schuerr95}. We will later explain how to extend it to arbitrary triple rules.
We also recall the notion of marking~\cite{Leblebici18} and \emph{consistency patterns} which can be used to check if a triple graph belongs to a given TGG.
\begin{definition}[Source and forward rule. Consistency pattern]
Given a plain, monotonic triple rule $r = L \rightarrow R$ with $r = (r_S, r_C, r_T)$, $L = (L_S \xleftarrow{\sigma_L} L_C \xrightarrow{\tau_L} L_T)$ and $R = (R_S \xleftarrow{\sigma_R} R_C \xrightarrow{\tau_R} R_T)$, its \emph{source rule} is defined as
\begin{equation*}
r^S \coloneqq (L_S \leftarrow \emptyset \rightarrow \emptyset) \xrightarrow{(r_S, \mathit{id}_{\emptyset}, \mathit{id}_{\emptyset})} (R_S \leftarrow \emptyset \rightarrow \emptyset) \enspace .
\end{equation*}
Its \emph{forward rule} is defined as
\begin{equation*}
r^F \coloneqq (R_S \xleftarrow{\sigma_R \circ r_C} L_C \xrightarrow{\tau_L} L_T) \xrightarrow{(\mathit{id}_{R_S}, r_C, r_T)} (R_S \xleftarrow{\sigma_R} R_C \xrightarrow{\tau_R} R_T) \enspace .
\end{equation*}
We denote the left- and right-hand sides of source and forward rules of a rule $r$ by $L^S,L^F,R^S$, and $R^F$, respectively.
The \emph{consistency pattern} derived from $r$ is the rule
\begin{equation*}
r^C \coloneqq (R_S \xleftarrow{\sigma_R} R_C \xrightarrow{\tau_R} R_T) \xrightarrow{(\mathit{id}_{R_S}, \mathit{id}_{R_C}, \mathit{id}_{R_T})} (R_S \xleftarrow{\sigma_R} R_C \xrightarrow{\tau_R} R_T)
\end{equation*}
that, upon application, just checks for the existence of the RHS of the rule without changing the instance it is applied to.
Given a rule $r$, each element $x \in R_S \setminus L_S$ is called a \emph{source marking element} of the forward rule $r^F$; each element of $L_S$ is called \emph{required}.
Given an application $G \Rightarrow_{r^F,\mathit{m^F}} H$ of a forward rule $r^F$, the elements of $G_S$ that have been matched by source marking elements of $r^F$, i.e., the elements of the set $\mathit{m}^F(R_S \setminus L_S)$ are called \emph{marked elements}.
A transformation sequence
\begin{equation}\label{eq:forward-sequence}
G_0 \Rightarrow_{m_1^F,r_1^F} G_1 \Rightarrow_{m_2^F,r_2^F} \dots \Rightarrow_{m_t^F,r_t^F} G_t
\end{equation}
is called \emph{creation preserving} if no two rule applications in sequence (\ref{eq:forward-sequence}) mark the same element.
It is called \emph{context preserving} if, for each rule application in sequence (\ref{eq:forward-sequence}), the required elements have been marked by a previous rule application in sequence (\ref{eq:forward-sequence}).
If these two properties hold for sequence~(\ref{eq:forward-sequence}), it is called {\em consistently marking}.
It is called \emph{entirely marking} if every element of the common source graph $G_S$ of the triple graphs of this sequence is marked by a rule application in sequence (\ref{eq:forward-sequence}).
\end{definition}
The most important formal property of this operationalization is that applying a (sequence of) source rule(s) followed by applying the (sequence of) corresponding forward rule(s) yields the same result as applying the (sequence of) original TGG rule(s) assuming consistent matches~\cite{Schuerr95,EEEHT07}.
Moreover, there is a correspondence between triple graphs belonging to the language of a given TGG and consistently and entirely marking transformation sequences via its forward rules.
We formally state this correspondence as it is an ingredient for the proof of correctness of our synchronization algorithm.
\begin{lemma}[{see \cite[Fact~1]{LAFVS17} or ~\cite[Lemma~4]{Leblebici18}}]\label{lem:emccp-series}
Let a TGG $\mathit{GG}$ be given.
There exists a triple graph $G = (G_S \leftarrow G_C \rightarrow G_T) \in \mathcal{L}(\mathit{GG})$ if and only if there exists a transformation sequence like the one depicted in~(\ref{eq:forward-sequence}) via forward rules from $\mathit{GG}$ such that $G_0 = (G_S \leftarrow \emptyset \rightarrow \emptyset),\ G_t = (G_S \leftarrow G_C \rightarrow G_T)$, and the transformation sequence is consistently and entirely marking.
\end{lemma}
For practical purposes, forward rules and consistency patterns may be equipped with so-called \emph{filter NACs} which can be automatically derived from the set of rules of the given TGG.
The simplest examples of such filter NACs arise through the following analysis:
For each rule that translate a node without translating adjacent edges it is first checked if other rules translate the same type of node but also translate an adjacent edge of some type.
If this is the case, it is checked if there are further rules which only translate the detected kind of adjacent edge.
If none is found, the original rule is equipped with a NAC forbidding the respective kind of edges.
This avoids a dead-end in translation processes:
In the presence of such a node with its adjacent edge, using the original rule to only translate the node leaves an untranslatable edge behind.
The filter NAC of \firstTGGForwardRule{} is derived in exactly this way.
For the exact and more sophisticated derivation processes of filter NACs, we refer to the literature~\cite{Hermann2010,KLKS10}.
For our purposes it suffices to recall their distinguishing property:
Filter NACs do not prevent \enquote{valid} transformation sequences of forward rules.
We state this property in the terminology of our paper.
\begin{fact}[{\cite[Fact~4]{Hermann2010}}]
Given a TGG $\mathit{GG} = (\mathcal{R},S)$, for each $r \in \mathcal{R}$, let $r^{\mathit{FN}}$ denote the corresponding forward rule that is additionally equipped with a set of derived filter NACs. (This set might be empty).
For $G_0 = (G_S \leftarrow \emptyset \rightarrow \emptyset)$,
there exists a consistently and entirely marking
transformation sequence
\begin{equation*}
G_0 \Rightarrow_{r_1^F,m_1^F} G_1 \Rightarrow_{r_2^F,m_2^F} \dots \Rightarrow_{r_t^F,m_t^F} G_t
\end{equation*}
via the forward rules (without filter NACs) derived from $\mathcal{R}$ if and only if the sequence
\begin{equation*}
G_0 \Rightarrow_{r_1^{\mathit{FN}},m_1^F} G_1 \Rightarrow_{r_2^{\mathit{FN}},m_2^F} \dots \Rightarrow_{r_t^{\mathit{FN}},m_t^F} G_t
\end{equation*}
exists, i.e, if none of the filter NACs blocks one of the above rule applications.
\end{fact}
\begin{example}
The source rules of the triple rules depicted in Fig.~\ref{fig:tggRules} are depicted in Fig.~\ref{fig:tggSourceRules}.
They allow to create \packages{} and \classes{} on the source side without changing correspondence and target graphs.
The formally existing empty graphs at correspondence and target sides are not depicted.
The corresponding forward rules are given in Fig.~\ref{fig:tggFwdRules}.
Their required elements are annotated with \marked{} and their source marking elements with \marking{}.
The rule \firstTGGForwardRule{} is equipped with a filter NAC:
The given grammar does not allow to create a \package{} that is contained in another one with its original rule \firstTGGRule{}.
Hence, the derived forward rule should not be used to translate a \package{}, which is contained in another one, to a \folder{}.
As evident in the examples, the application of a source rule followed by the application of the corresponding forward rule amounts to the application of the original triple rule if matched consistently.
The consistency patterns that are derived from the TGG rules of our example are depicted in Fig.~\ref{fig:consistency}.
They just check for existence of the pattern that occurs after applying the original TGG rule.
A consistency pattern is equipped with the filter NACs of both its corresponding forward and backward rule.
In our example, only \emph{Root-Consistency-Pattern} receives such NACs; one from \firstTGGForwardRule{} and the second one from the analogous backward rule.
An occurrence of a consistency pattern in our example model indicates that a specific location corresponds to a concrete TGG rule application.
Hence, a disappearance of such a match indicates that a former intact rule application has been broken and needs some fixing.
We call this a \emph{broken match for a consistency pattern} or, short, a \emph{broken consistency match}.
Practically, we will exploit an incremental pattern matcher to notify us about such disappearances.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\columnwidth]{Diagram/Consistency_attr.pdf}
\caption{Example: Consistency Patterns}
\label{fig:consistency}
\end{figure}
\end{example}
\subsection{Sequential independence}
The proof of correctness of our synchronization approach relies on the notion of \emph{sequential independence}.
Transformations that are sequentially independent can be performed in arbitrary order.
\begin{definition}[Sequential independence]
Given two transformation steps $G \Rightarrow_{r_1,m_1} H_1 \Rightarrow_{r_2,m_2} X$, via plain rules $r_1,r_2$ these are \emph{sequentially independent} if
\begin{equation}\label{eq:seq-independence}
n_1(R_1) \cap m_2(L_2) \subseteq n_1(K_1) \cap m_2(K_2)
\end{equation}
where $n_1$ is the comatch of the first transformation.
\end{definition}
By the Local Church-Rosser Theorem the order of sequentially independent transformation can be switched. This means that, given a sequentially independent transformation sequence $G \Rightarrow_{r_1,m_1} H_1 \Rightarrow_{r_2,m_2} X$, there exists a sequentially independent transformation sequence $G \Rightarrow_{r_2,m_2^\prime} H_2 \Rightarrow_{r_1,m_1^\prime} X$~\cite[Theorem~3.20]{EEPT06}.
If $r_1$ and $r_2$ are equipped with NACs $\mathit{NAC}_1$ and $\mathit{NAC}_2$, respectively, transformation steps as above are \emph{sequentially independent} if condition~(\ref{eq:seq-independence}) holds and moreover, the thereby induced matches $m_2^\prime: L_2 \to G$ and $m_1^\prime: L_1 \to H_2$ both satisfy the respective sets of NACs. In particular, the Local Church-Rosser Theorem still holds.
In our setting of graph transformation, it is easy to check the sequential independence of transformations~\cite{EEPT06,EGHLO14}.
A sequence $t_1;t_2$ of two transformation steps is sequentially independent if and only if the following holds.
\begin{itemize}
\item $t_2$ does not match an element that $t_1$ created.
\item$t_2$ does not delete an element that $t_1$ matches.
\item $t_2$ does not create an element that $t_1$ forbids.
\item $t_1$ does not delete an element that $t_2$ forbids.
\end{itemize}
\subsection{Construction of short-cut rules}
We recall the construction of short-cut rules in a semiformal way and reuse an example of~\cite{FKST18} for illustration; a formal treatment (in a category-theoretical setting) can be found in that paper.
Given an inverse monotonic rule (i.e., a rule that purely deletes) and a monotonic rule, a \shortcutRule{} combines their respective actions into a single rule.
Its construction allows to identify elements that are deleted by the first rule as recreated by the second one.
To motivate the construction, assume two monotonic rules $r_1: L_1 \rightarrow R_1$ and $r_2: L_2 \rightarrow R_2$
be given.
Applying the inverse rule of $r_1$ to a triple graph $G$, provides an image of $L_1$ in the resulting triple graph $H$.
When applying $r_2$ thereafter, the chosen match for $L_2$ in $H$ may intersect with the image of $L_1$ yielding a triple graph $L_{\cap}$.
This intersection can also be understood as saying that $L_{\cap}$ provides a partial match for $L_2$.
The inverse application of the first rule deletes elements which may be recreated again.
In this case, it is possible to extend the sub-triple graph $L_{\cap}$ of $H$
to a sub-triple graph $R_{\cap}$ of $H$ with these elements.
In particular, $R_{\cap}$ is a sub-triple graph of $R_1$ and $R_2$ as it includes elements only that have been deleted by the first rule and created by the second.
Based on this observation, the construction of short-cut rules is defined as follows (slightly simplified and directly merged with an example):
\begin{construction}[Short-cut rule]
Let two plain, monotonic rules $r_1 = L_1 \to R_1$ and $r_2= L_2 \to R_2$ be given.
A short-cut rule $r_{\mathit{sc}}$ for the rule pair $(r_1,r_2)$, where $r_1$ is considered to be applied inversely, is constructed in the following way:
\begin{figure*}
\centering
\includegraphics[width=.8\textwidth,trim={1.5mm 1.5mm 1.5mm 1.5mm},clip]{Diagram/example-common-kernel.pdf}
\caption{A common kernel rule pair (\firstTGGRule{},\secondTGGRule{}). The names of the nodes indicate their mappings and the rules are depicted top-down.}
\label{fig:example-common-kernel}
\end{figure*}
\begin{enumerate}
\item \emph{Choice of common kernel:}
A (potentially empty) sub-triple graph $L_{\cap}$ of $L_1$ and $L_2$ and a sub-triple graph $R_{\cap}$ of $R_1$ and $R_2$ with $L_{\cap} \subseteq R_{\cap}$ are chosen.
We call $\Lcap \subseteq \Rcap$ a \emph{common kernel} of both rules.
In Fig.~\ref{fig:example-common-kernel}, an example of such a common kernel is given.
It is a common kernel for rule pair (\firstTGGRule{}, \secondTGGRule{}). The common kernel is depicted in the center of Fig.~\ref{fig:example-common-kernel}.
This choice of a common kernel will lead to \firstSCRule{} as resulting short-cut rule.
In this example, $L_{\cap}$ is empty and $R_{\cap}$ extends $L_{\cap}$ by identifying the \packages{} \p, \folders{} \f, and the correspondence node in between.
The elements of $R_{\cap} \setminus L_{\cap}$, called \emph{recovered elements}, are to become the elements that are preserved by an application of the short-cut rule compared to reversely applying the first rule followed by applying the second one (provided that these applications overlap in $L_{\cap}$).
In the example case, the whole graph $R_{\cap}$ is recovered as $L_{\cap}$ is empty.
\item \emph{Construction of LHS and RHS:}
One first computes the union $L_{\cup}$ of $L_1$ and $L_2$ along $L_{\cap}$.
The result is then united with $R_1$ along $L_1$ and $R_2$ along $L_2$, respectively, to compute the LHS and the RHS of the short-cut rule.
Figure~\ref{fig:example-construction-LHS-RHS} displays this.
\item \emph{Interface construction:}
The interface $K$ of the short-cut rule is computed by taking the union of $L_{\cup}$ and $R_{\cap}$ along $L_{\cap}$.
For our example, this construction is depicted in Figure~\ref{fig:example-construction-interface}.
The elements of $L_2 \setminus L_{\cap}$ are called \emph{presumed elements} since, given a match for the inverse first rule, i.e., for $R_1$, these are exactly the elements needed to extend this match to a match of the short-cut rule.
In our example, these are the \package{} \ssp{}, the \folder{} \ssf{}, and the correspondence node in between.
\end{enumerate}
\begin{figure*}
\centering
\includegraphics[width=\textwidth,trim={1.5mm 1.5mm 1.5mm 1.5mm},clip]{Diagram/example-construction-LHS-RHS.pdf}
\caption{Constructing the LHS and the RHS of the short-cut rule \firstSCRule{}}
\label{fig:example-construction-LHS-RHS}
\end{figure*}
\end{construction}
\begin{figure}
\centering
\includegraphics[width=\columnwidth,trim={1.5mm 1.5mm 1.5mm 1.5mm},clip]{Diagram/example-construction-interface.pdf}
\caption{Constructing the interface of the short-cut \firstSCRule{}; the interface is the resulting graph in the bottom right corner.}
\label{fig:example-construction-interface}
\end{figure}
\begin{example}
More examples of short-cut rules are depicted in Fig.~\ref{fig:scRules}.
Both, \firstSCRule{} and \secondSCRule{}, are constructed for the rules \firstTGGRule{} and \secondTGGRule{}. Switching the role of of the inverse rule, two short-cut rules can be constructed having equal common kernels.
In both cases, the \package{} \p{}, the \folder{} \f{} and the correspondence node between them are recovered elements, as these elements would have been deleted and re-created otherwise.
While in \firstSCRule{}, the presumed elements are the \package{} \ssp{} and the \folder{} \ssf{} with a correspondence node in between, the set of presumed elements of \secondSCRule{} is empty.
Another possible common kernel for \firstTGGRule{} and \secondTGGRule{} is one where $R_{\cap}$ is an empty triple graph as well.
As the resulting short-cut rule just copies both rules (one of them inversely) next to each other, this rule is not interesting for our desired application.
\end{example}
\subsection{Expressivity of short-cut rules}
Given a set of rules, there are {\em two degrees of freedom} when deciding which short-cut rules to derive from them:
First, one has to choose for which \emph{pairs of rules} short-cut rules shall be derived.
Secondly, given a pair of rules, there is typically not only one way to construct a short-cut rule for them:
In general, there are different choices for a \emph{common kernel}.
However, when fixing a common kernel, i.e., $\Lcap$ and $\Rcap$, the result of the construction is uniquely determined. If, moreover, the LHSs and RHSs of the rules are finite, the set of possible common kernels is finite as well.
As short-cut rules correspond to possible (complex) edits of a triple graph, the more short-cut rules are derived, the more user edits are available which can directly be propagated by the corresponding repair rules.
But the number of rules that has to be computed (and maintained throughout the synchronization process) in this way, would quickly grow.
And maybe several of the constructed rules would capture edits that are possible in principle but unlikely to ever be performed in a realistic scenario.
Hence, some trade-off between expressivity and maintainability has to be found.
We shortly discuss these effects of choices:
The construction of short-cut rules is defined for \emph{any two} monotonic rules~\cite{FKST18} -- we do not need to restrict to the rules of a given TGG but may also use monotonic rules that have been constructed as so-called \emph{concurrent rules}~\cite{EEPT06} of given TGG rules as input for the short-cut rule construction.
A concurrent rule combines the actions of two (or more) subsequent rule applications into a single rule.
Hence, deriving short-cut rules from concurrent rules that have been built of given TGG rules leads to short-cut rules that capture even more complex edits into a single rule.
The next example presents such a derived short-cut rule.
While our conceptual approach is easily extended to support such rules, we currently stick with short-cut rules directly derived from a pair of rules of the given TGG in our implementation.
\begin{example}
The short-cut rule \fourthSCRule{} depicted in Fig.~\ref{fig:sc-rule-4} is not directly derived of the TGG rules depicted in Fig.~\ref{fig:tggRules}.
Instead, the concurrent rule of two given applications of \secondTGGRule{} is constructed first.
This concurrent rule directly creates a chain of two \packages{} and \folders{} into an existing pair of \package{} and \folder{}.
The rule in Fig.~\ref{fig:sc-rule-4} is a short-cut rule of this concurrent rule and \secondTGGRule{}. It takes back the creation of a chain such that the bottom package is directly included in the top package in Fig.~\ref{fig:sc-rule-4}.
\begin{figure}
\includegraphics[width=\columnwidth]{Diagram/SC_Rule_4.pdf}
\caption{Example for a short-cut rule not directly derived from the rules of our example TGG}
\label{fig:sc-rule-4}
\end{figure}
\end{example}
Concerning the choice of a common kernel, we follow two strategies.
In both strategies, we overlap as many of the newly created elements of the two input rules as possible since these are the elements that we try to preserve.
A \emph{minimal overlap} overlaps created elements only, i.e. no context elements.
An example is \secondTGGRule{}, which overlapped with itself, results in \thirdSCRule{} and which corresponds to a move refactoring step.
A \emph{maximal overlap} overlaps not only created elements of both rules but also context elements.
Creating such an overlap for \secondTGGRule{} with itself would result in the \emph{Sub-Consistency-Pattern}, which has no effect when applied.
However, when overlapping different rules with each other, it is often useful to re-use context elements.
This is the case, for example, for \emph{VariableDec-2-Parameter-Rule} and \emph{TypeAccess-2-ReturnType-Rule} of our evaluation rule set in Fig.~\ref{fig:eval_rule_set_full} below.
A full overlap between both rules would allow to transform a signature parameter to a return parameter of the same method and of the same type and, vice versa.
Both strategies aim to create different kinds of short-cut rules with specific purposes.
Since generating all possible overlaps and thus short-cut rules is expensive, we chose a heuristic approach to generate a useful subset of them.
As we are dealing with triple graphs being composed of source, target and correspondence graphs, the overlap of source graphs should correspond to that of target graphs.
This restricts the kind of \enquote{reuse} of elements the derived short-cut rules enable.
The allowance of any kind of overlap may include unintended ones.
We argue for the usefulness of these strategies in our evaluation in Sect.~\ref{sec:implAndEvaluation}.
\subsection{Language preserving short-cut rule applications}
The central intuition behind the construction of short-cut rules is to replace the application of a monotonic triple rule by another one.
In this sense, a short-cut rule captures a complex edit operation on triple graphs that (in general) cannot be performed directly using the rules of a TGG.
We illustrate this behaviour in the following. Subsequently, we discuss the circumstances under which applications of short-cut rules are \enquote{legal} in the sense that the result still belongs to the language of the respective TGG.
\medskip
Let a TGG $\mathit{GG}$ and a sequence of transformations
\begin{equation}\label{eq:original-sequence}
G_0 \Rightarrow_{r_1,m_1} G_1 \Rightarrow_{r_2,m_2} G_2 \Rightarrow \dots \Rightarrow_{r_t,m_t} G_t
\end{equation}
be given where all the $r_i$, $1 \leq i \leq t$, are rules of $\mathit{GG}$, all the $m_i$ denote the respective matches, and $G_0 \in \mathcal{L}(\mathit{GG})$; in particular $G_t \in \mathcal{L}(\mathit{GG})$ as well.
Fixing some $j \in \{1, \dots, t\}$ and some rule $r$ of $\mathit{GG}$, we construct a short-cut rule $r_{sc}$ for $(r_j, r)$ with some common kernel $\Lcap \subseteq \Rcap$.
Next, we can consider the transformation sequence
\begin{equation*}
G_0 \Rightarrow_{r_1,m_1} G_1 \Rightarrow_{r_2,m_2} G_2 \Rightarrow \dots \Rightarrow_{r_t,m_t} G_t \Rightarrow_{r_{sc},m_{sc}} G_t'
\end{equation*}
that arises by appending an application of $r_{sc}$ to transformation sequence~(\ref{eq:original-sequence}).
Under certain technical cir\-cumstances (which we will state below) this transformation sequence is equivalent\footnote{The formal notion of equivalence used here is called \emph{switch equivalence} and captures the idea that, in case of sequential independence, the order of rule applications might be switched while using basically the same match for each rule application and receiving the same result; compare, e.g.,~\cite{Kreowski86,BCHKS14}.} to the sequence
\begin{equation}\label{eq:final-sequence}
\begin{split}
G_0 \Rightarrow_{r_1,m_1} G_1 \Rightarrow \dots \Rightarrow_{r_{j-1},m_{j-1}} G_{j-1} \Rightarrow_{r,m_{sc}^{\prime}} G_j^{\prime} \\
\Rightarrow_{r_{j+1},m_{j+1}^{\prime}} \dots \Rightarrow_{r_t,m_t^{\prime}} G_t^{\prime\prime}
\end{split}
\end{equation}
where the application of $r_j$ at match $m_j$ is replaced by an application of $r$ at a match $m_{sc}\rq{}$ that is derived from the match $m_{sc}$ of the \shortcutRule{}. The following matches $m_{j+1},\allowbreak \dots,\allowbreak m_{t}$ have been adapted accordingly.
They still match the same elements but formally they do so in other triple graphs.
In particular, $G_t^{\prime\prime}$, the result of the transformation sequence (\ref{eq:final-sequence}), \emph{is isomorphic} to $G_t^{\prime}$ and hence, $G_t^{\prime}$ can be understood as arising by replacing the $j$-th rule application in the transformation sequence (\ref{eq:original-sequence}) by an application of the rule $r$; thus, $G_t^{\prime}$ also belongs to the language of the TGG:
The sequence~(\ref{eq:final-sequence}) starts at a triple graph $G_0 \in \mathcal{L}(\mathit{GG})$ and solely consists of applications of rules from $\mathit{GG}$.
\begin{figure*}
\centering
\includegraphics[width=.75\textwidth]{Diagram/RuleApplications_Solved}%
\caption{Example: Transforming sequences of rule applications by applying short-cut rules}%
\label{fig:ruleApplications_Solved}%
\end{figure*}
\begin{example}
Consider the triple graph depicted in Fig.~\ref{fig:translationExample} (a).
It arises by applying \firstTGGRule{}, followed by two applications of \secondTGGRule{}, and finally an application of \thirdTGGRule{}.
When matched as already described in the introductory example, an additional application of \secondSCRule{} to this triple graph results in the one depicted in Fig.~\ref{fig:translationExample} (c).
Alternatively, this can be derived by two applications of \firstTGGRule{}, followed by an application of \secondTGGRule{} and \thirdTGGRule{} each.
As schematically depicted in Fig.~\ref{fig:ruleApplications_Solved}, the application of the \shortcutRule{} \secondSCRule{} transforms an transformation sequence deriving the first triple graph into a transformation sequence deriving the second one by replacing an application of \secondTGGRule{} by one of \firstTGGRule{}.
\end{example}
In the following, we state when the above described behaviour is the case (in a somewhat less technical language than originally used).
\begin{theorem}[{\cite[Theorem~8]{FKST19}}]\label{thm:valid-applications}
Let the transformation sequence~(\ref{eq:original-sequence}) be given and let $r_{sc}$ be a \shortcutRule{} that is derived from $(r_j, r)$.
If the following three conditions are met, this sequence is equivalent to sequence~(\ref{eq:final-sequence}) where original TGG rules are applied only.
\begin{enumerate}
\item \emph{Reversing match:} The application of $r_{sc}$ at $m_{sc}$ reverses the application of $r_j$, i.e., $n_j(R_j) = m_{\mathit{sc}}|_{R_j}(R_j)$.
\item \emph{Sequential independence:}
\begin{enumerate}
\item \emph{Non-disabling match:} The application of $r_{sc}$ at $m_{sc}\rq{}$ does not delete elements used in the applications of $r_{j+1}, \dots, r_t$.
\item \emph{Context-preserving match:} The match $m_{sc}$ for $r_{sc}$ already exists in $G_{j-1}$.
Since the assumption on the match to be reversing already ensures this for elements of $L_{\mathit{sc}}$ that stem from $R_j$, context-preservation ensures in particular that the \emph{presumed elements} of $r_{sc}$ are matched to elements already existing in $G_{j-1}$.
\end{enumerate}
\end{enumerate}
\end{theorem}
\begin{example}
We illustrate each of the above mentioned conditions:
\begin{enumerate}
\item \emph{Reversing match:} In our example of matching \firstSCRule{} to the triple graph (c) in Fig.~\ref{fig:translationExample} this means that its nodes \p{} and \f{} (and the correspondence node in between) are allowed to be matched to elements only that have been created using \firstTGGRule{}.
In this way, it is avoided to mis\-use the rule to introduce \packages{} (and \folders{}) that are contained by more than one \package{} (or \folder{}).
\item \emph{Non-disabling match:} For example, \fourthSCRule{} from Fig.~\ref{fig:sc-rule-4} is not allowed to delete \packages{} and \folders{} that already contain \classes{} or \docs{}, respectively.
\item \emph{Context preserving match:} Returning to our example of matching \firstSCRule{} to the triple graph (c) in Fig.~\ref{fig:translationExample} this means that as soon as nodes \subP{} and \subF{} in that triple graph have been chosen as matches for the nodes \p{} and \f{} of \firstSCRule{}, the nodes \leafP{} and \leafF{} are not allowed to be chosen as matches for nodes \ssp{} and \ssf{} of \firstSCRule{}.
The creation of \leafP{} and \leafF{} depends on \subP{} and \subF{} being created first.
In this way, the introduction of cyclic dependencies between elements is avoided.
\end{enumerate}
\end{example}
\subsection{The Basic Setup}\label{sec:basic-setup-algorithm}
We assume a TGG $\mathit{GG}$ with \emph{plain, monotonic rules} to be given.
Its language defines consistency.
This means that a triple graph $G = (G_S \leftarrow G_C \rightarrow G_T)$ is consistent if and only if $G \in \mathcal{L}(\mathit{GG})$.
\paragraph{The problem.}
A consistent triple graph $G = (G_S \leftarrow G_C \rightarrow G_T) \in \mathcal{L}(\mathit{GG})$ is given; by Lemma~1 there exists a corresponding consistently and entirely marking sequence $t$ of forward rule applications.
After editing source graph $G_S$ we get $G\rq{} = (H_S \dashleftarrow G_C \rightarrow G_T)$.
Generally, the result $G\rq{}$ is a partial triple graph and does not belong to $\mathcal{L}(\mathit{GG})$.
We assume that all the edits are performed by applying source rules.
They may be derived from the original TGG rules or from short-cut rules.
Our goal is to provide a \emph{model synchronization algorithm} that, given $G = (G_S \leftarrow G_C \rightarrow G_T) \in \mathcal{L}(\mathit{GG})$ and $G\rq{} = (H_S \dashleftarrow G_C \rightarrow G_T)$ as input, computes a triple graph $H = (H_S \leftarrow H_C \rightarrow H_T) \in \mathcal{L}(\mathit{GG})$.
As a side condition, we want to minimize the amount of elements of $G_C$ and $G_T$ that are deleted and recreated during that synchronization.
\paragraph{Ingredients of our algorithm.}
We provide a rule-based model synchronization algorithm leveraging an incremental pattern matcher.
During that algorithm, rules are applied to compute a triple graph $(H_S \leftarrow H_C \rightarrow H_T) \in \mathcal{L}(\mathit{GG})$ from the (partial) triple graph $(H_S \dashleftarrow G_C \rightarrow G_T)$.
We apply two different kinds of rules, namely
\begin{enumerate}
\item forward rules derived from the rules of the TGG $\mathit{GG}$ and
\item repair rules, i.e., operationalized short-cut rules.
\end{enumerate}
Forward rules serve to propagate the addition of elements.
The use of these rules for model synchronization is standard.
However, the use of additional repair rules and the way in which they are employed are conceptually novel.\footnote{Note that consistency is still defined by the (plain, monotonic) rules of the given TGG; the general repair rules are derived only to improve the synchronization process.}
The repair rules allow to directly propagate more complex user edits.
During the synchronization process, the rules are applied reacting to notifications by an incremental pattern matcher.
We require this pattern matcher to provide the following information:
\begin{enumerate}
\item The original triple graph $G = (G_S \leftarrow G_C \rightarrow G_T)$ is \emph{covered with consistency patterns}.
When considering the induced matches for forward rules, every element of $G_S$ is marked exactly once. The dependency relation between elements required by these matches is acyclic.
This means that the induced transformation sequence of forward rules is consistently and entirely marking.
Such a sequence always exists since $G \in \mathcal{L}(\mathit{GG})$; see Lemma~\ref{lem:emccp-series}.
\item \emph{Broken consistency matches} are reported.
A match for a consistency pattern in $G$ is broken in $G\rq{}$ if one of the elements it matches or creates has been deleted or if an element has been created that violates one of the filter NACs of that consistency pattern.
\item The incremental pattern matcher notifies about newly occurring matches for forward rules.
It does so in a \emph{correct} way, i.e., it only notifies about matches that lead to consistently marking transformations.
\item In addition, the incremental pattern matcher informs a \emph{precedence graph}.
This precedence graph contains information about the mutual dependencies of the elements in the partial triple graph.
Here, an element is dependent on another one if the forward rule application marking the former matches the latter element as required.
We consider the transitive closure of this relation.
\end{enumerate}
\subsection{Synchronization Process}
\label{subsec:sync-process}
Our synchronization process is depicted in Algorithm~\ref{alg:sync}.
It applies rules to translate elements and repair rule applications.
In that, it applies a different strategy than suggested in~\cite{LAFVS17,Leblebici18}.
There, invalid rule applications are revoked as long as there exist any.
Subsequently, forward rules are applied as long as possible.
By trying to apply a suitable repair rule instead of revoking an invalid rule application, we are able to avoid deletion and recreation of elements.
Our synchronization algorithm is defined as follows.
Note that we present an algorithm for synchronizing in forward direction (from source to target) while synchronizing backwards is performed analogously.
The function \textit{synchronize} is called on the current partial triple graph that is to be synchronized.
In line~\ref{alg:sync:update}, \textit{updateMatches} is called on this partial triple graph.
It returns the set of consistency matches currently broken, a set of consistency matches being still intact, and a set of forward TGG rule matches.
By calling the function \emph{isFinished} (line~\ref{alg:sync:isFinished-call}), termination criteria for the synchronization algorithm are checked.
If the set of broken consistency matches and the set of forward TGG rule matches are both empty and all elements of the source graph are marked as translated, the synchronization algorithm terminates (line~\ref{alg:sync:terminate}).
Yet, if both sets are empty but there are still untranslated elements in the source graph, an exception is thrown in line~\ref{alg:sync:exception}, signaling that the (partial) triple graph is in an inconsistent state.
Subsequently, function \emph{translate} is used (line~\ref{alg:sync:translate-call}) to propagate the creation of elements:
If the set of forward TGG rule matches is non-empty (line~\ref{alg:sync:apply}), we choose one of these matches, apply the corresponding rule, and continue the synchronization process (line~\ref{alg:sync:sync1}).
This step is done prior to any repair.
The purpose is to create the context which may be needed to make repair rules applicable.
An example for such a context creation is the insertion of a new root \package{} which has to be translated into a root \folder{} before applying \firstRRule{} thereafter (see Fig.~\ref{fig:tggFwdRules}).
\begin{algorithm*}
\begin{algorithmic}[1]
\Function{synchronize}{tripleGraph}
\State $\mathrm{(brokenCMatches, intactCMatches,fwdMatches) \leftarrow updateMatches(tripleGraph)}$ \label{alg:sync:update}
\\
\If{isFinished(tripleGraph, fwdMatches, brokenCMatches))} \label{alg:sync:isFinished-call}
\State \Return
\EndIf
\\
\If{translate(tripleGraph, fwdMatches)} \label{alg:sync:translate-call}
\State \Return
\EndIf
\\
\State $\mathrm{(cMatch, success) \leftarrow repair(tripleGraph, brokenCMatches)}$ \label{alg:sync:repair-call}
\If{!success}
\State \textbf{throw} InconsistentStateException \label{alg:sync:exception2}
\EndIf
\State \Return
\EndFunction
\\
\Function{isFinished}{tripleGraph, fwdMatches, brokenCMatches} \label{alg:sync:isFinished}
\If{isEmpty(brokenCMatches) \&\& isEmpty(fwdMatches)} \label{alg:sync:termination}
\If{allElementsTranslated(tripleGraph.source)}
\State \Return true \label{alg:sync:terminate}
\Else{}
\State \textbf{throw} InconsistentStateException \label{alg:sync:exception}
\EndIf
\EndIf
\State \Return false
\EndFunction
\\
\Function{translate}{tripleGraph, fwdMatches} \label{alg:sync:translate}
\If{!isEmpty(fwdMatches)} \label{alg:sync:apply}
\State fwdMatch = chooseMatch(fwdMatches)
\State $\mathrm{tripleGraph \leftarrow applyRule(tripleGraph, fwdMatch, getFWDRule(fwdMatch))}$
\State synchronize(tripleGraph) \label{alg:sync:sync1}
\State \Return true
\EndIf
\State \Return false
\EndFunction
\\
\Function{repair}{tripleGraph, brokenCMatches}
\State $\mathrm{cMatch \leftarrow chooseMatch(brokenCMatches)}$ \label{alg:sync:chooseMatch}
\State $\mathrm{scRules \leftarrow getSuitableSCRules(cMatch)}$
\State $\mathrm{scMatches \leftarrow findSCMatches(scRules, cMatch)}$
\\
\While{!isEmpty(scMatches)} \label{alg:sync:while}
\State $\mathrm{scMatch \leftarrow chooseMatch(scMatches)}$
\If{isValidMatch(scMatch)} \label{alg:sync:isValid}
\State $\mathrm{tripleGraph \leftarrow applyRule(tripleGraph, cMatch, getSCRule(scMatch))}$ \label{alg:sync:sync4}
\State synchronize(tripleGraph) \label{alg:sync:sync2}
\State \Return (cMatch, true)
\EndIf
\EndWhile
\State \Return (cMatch, false)
\EndFunction
\end{algorithmic}
\caption{eMoflon -- Synchronization Process} \label{alg:sync}
\end{algorithm*}
If the above cases do not apply, there must be at least one broken consistency match and the corresponding rule application has to be repaired (line~\ref{alg:sync:repair-call}):
Hence, we choose one broken consistency match (line~\ref{alg:sync:chooseMatch}) for which a set of suitable repair rules is determined.
A broken consistency match includes information about the rule it corresponds to (e.g., the name of the rule).
Furthermore, it includes which elements are missing or which filter NACs are violated such that the corresponding application does not exist any more.
We calculate the set of matches of \repairRules{} (i.e., forward \shortcutRules{}) that stem from \shortcutRules{} revoking exactly the rule that corresponds to the broken consistency match.
In particular, by knowing which elements of a broken rule application still exist in the current source graph, we can stick to those repair rules that preserve exactly the still existing elements.
While the calculated set of unprocessed \repairRule{} matches is not empty (line~\ref{alg:sync:while}), we choose one of these matches and check whether it is \emph{valid}.
By constructing the partial match of a \repairRule{}, we only need to ensure that none of its \emph{presumed elements} is matched in such a way that a cyclic dependency is introduced.
This means that they must not be matched to elements that are dependent of elements to which the \emph{recovered elements} are matched.
If a match is valid, we apply the corresponding repair rule and continue the synchronization process (line~\ref{alg:sync:sync2}).
If no such rule or valid match is available, an exception is thrown (line~\ref{alg:sync:exception2}).
\subsection{Formal properties of the synchronization process.}
We discuss the termination, correctness, and completeness of our synchronization algorithm.
Our algorithm terminates as long as every forward rule translates at least one element (which is a quite common condition; compare~\cite[Lemma~6.7]{HEOCDXGE15} or \cite[Theorem~3]{Leblebici18}).
\begin{theorem}
Let a TGG $\mathit{GG}$ with plain, monotonic rules be given.
If every derived forward rule of $\mathit{GG}$ has at least one source marking element, our algorithm terminates for any finite input $G\rq{} = (H_S\rq{} \dashleftarrow G_C \rightarrow G_T)$.
\end{theorem}
\begin{proof}
The algorithm terminates -- by either throwing an exception or returning a result -- if at one point both, the set of broken consistency matches and the set of matches for forward rules are empty; compare the function \emph{isFinished} starting in line~\ref{alg:sync:isFinished}.
The algorithm is called recursively, always applying a forward rule if a match is available.
As every forward rule marks at least one element as translated and forward rules are only matched in such a way that source marking elements are matched to yet untranslated ones, the application of forward rules (lines~\ref{alg:sync:apply} et seq.), i.e., the recursive call of function \emph{translate}, halts after finitely many steps.
Moreover, an application of a forward rule never introduces a new broken consistency match:
As it neither creates nor deletes elements in the source graph, it cannot delete elements matched by a consistency pattern nor create elements forbidden by one.
This means that, as soon as the set of broken consistency matches is empty, the whole synchronization algorithm will terminate.
We show that at some point this set of broken consistency matches will be empty or an exception is thrown.
Whenever the algorithm is called with an empty set of matches for forward rules, broken consistency matches are considered by applying a repair rule, i.e., by calling the function \emph{repair}.
New matches for forward rules can result from this; as discussed above, newly appearing matches for forward rules are unproblematic.
However, an application of a repair rule does not introduce a new violation of any consistency match:
As it does not create source elements, it cannot introduce violations of filter NACs.
And by the condition on valid matches to be non-disabling (condition 2. (a) in Definition~\ref{def:valid-repair-match}), no elements needed by other consistency matches are deleted.
Hence, by application of a repair rule, the number of invalid consistency matches is reduced by one and the algorithm terminates as soon as all broken consistency matches are repaired.
If there is a broken consistency match that cannot be repaired -- either because no suitable repair rule or no valid match is available -- an exception is thrown and the algorithm stops. \qed
\end{proof}
\paragraph{Correctness.}
Upon termination without exception, our algorithm is correct.
\begin{theorem}[Correctness of algorithm]\label{thm:correctness-algorithm}
Let a TGG $\mathit{GG}$ with plain, monotonic rules, a triple graph $G = (G_S \leftarrow G_C \rightarrow G_T) \in \mathcal{L}(\mathit{GG})$, and a partial triple graph $G\rq{} = (G_S\rq{} \dashleftarrow G_C \rightarrow G_T)$ that arises by a user edit step on the source graph be given.
If our synchronization algorithm terminates without exception and yields $H = (H_S \leftarrow H_C \rightarrow H_T)$ as output, then $H_S = G_S\rq{}$ and $H \in \mathcal{L}(\mathit{GG})$.
\end{theorem}
\begin{proof}
We see immediately that $H_S = G_S\rq{}$ since none of the applied rules modifies the source graph.
If the synchronization process terminates without exception, all elements are translated, no matches for forward rules are found, and no consistency match is broken any more.
This means that the collected matches of the forward rules form an entirely marking transformation sequence.
By Lemma~\ref{lem:emccp-series}, we have to show that this sequence is also consistently marking.
Then, the matches of the forward rules that correspond to the matches of the consistency patterns that the incremental pattern matcher has collected encode a transformation sequence that allows to translate the triple graph $(H_S \leftarrow \emptyset \rightarrow \emptyset)$ to a triple graph $(H_S \leftarrow H_C \rightarrow H_T) \in \mathcal{L}(\mathit{GG})$.
We assume that the incremental pattern matcher recognizes all broken consistency matches and reports correct matches for forward rules only.
This means, throughout the application of forward rules, the set of all valid consistency matches remains consistently marking.
We have to show that this is also the case for repair rule applications.
If it is, upon termination without exception, there is an entirely and consistently marking sequence of forward rules which corresponds to a triple graph from $\mathit{GG}$ by Lemma~\ref{lem:emccp-series}.
Whenever we apply a repair rule we are (at least locally) in the situation of Corollary~\ref{cor:valid-repair-rule-applications}:
There is a (maybe empty) sequence of consistently marking forward rule applications and a suitable broken consistency pattern indicates, that a user edit step applying the source rule $r_{\mathit{sc}}^S$ of a short-cut rule $r_{\mathit{sc}}$ has taken place.
Applying the repair rule $r_{\mathit{sc}}^F$ at a valid match amounts to replacing the application of rule $r_j^F$, whose consistency pattern was broken, by rule $r^F$ in a consistently marking way. \qed
\end{proof}
We only informally discuss \emph{completeness}.
We understand completeness as follows: for every input $G\rq{} = (H_S \dashleftarrow G_C \rightarrow G_T)$ with $H_S \in \mathcal{L}_S(\mathit{GG})$, we obtain a result $H = (H_S \leftarrow H_C \rightarrow H_T) \in \mathcal{L}(\mathit{GG})$.
In general, the above proposed algorithm is not complete.
We randomly apply forward rules at available matches (without using backtracking) but the choice and order of such applications can affect the result if the final sequence of forward rule applications leads to a dead-end or translates the given source graph.
However, the algorithm is complete whenever the set of forward rules is of such a form that the order of their application does not make a difference (somewhat more formally: they meet some kind of confluence) and the user edit is of the form discussed in Sect.~\ref{sec:basic-setup-algorithm}.
Analogous restrictions on forward rules hold for other synchronization processes that have been formally examined for completeness~\cite{HEOCDXGE15,Leblebici18}.
Adding filter NACs to the forward rules of a TGG is a technique that can result in such a set of confluent forward rules even if the original set of forward rules is not.
Moreover, there are static methods to test TGGs for such a behaviour~\cite{ALST14,HEOCDXGE15}; they check for sufficient but not for necessary criteria.
If it is known that the set of forward rules of a given TGG guarantees completeness and the edit is of a suitable kind, a thrown exception during our synchronization process implies that $H_S \notin \mathcal{L}_S(\mathit{GG})$.
\subsection{A synchronization example}
\begin{figure*}%
\includegraphics[width=\textwidth]{Diagram/Example-Sync-Alg-adapted.pdf}%
\caption{Example of our proposed synchronization algorithm. Grey background indicates broken consistency matches.}%
\label{fig:algorithm-example}%
\end{figure*}
We illustrate our synchronization algorithm with an example illustrated in Fig.~\ref{fig:algorithm-example}.
For simplicity, we neglect the \textsf{content} attribute and concentrate on the structural behaviour.
As a starting point, we assume that a user edits the source graph of the triple graph depicted in Fig.~\ref{fig:algorithm-example}~(a) (in the following, we will refer to the triple graphs occurring throughout the algorithm just by their numbers).
She adds a new root package above \rootP{}, removes the link between \packages{} \rootP{} and \subP{}, and creates a further class \texttt{c2}.
All these changes are specified by either a source rule of the TGG or the source rule of a derived short-cut rule.
The resulting triple graph is depicted in~(b).
The elements in front of the grey background are considered to be inconsistent, due to a broken consistency match.
Furthermore, \texttt{c2} and \texttt{nRootP} are not translated, yet.
In the first two passes of the algorithm, the two available matches for forward rules are applied (in random order):
\thirdTGGForwardRule{} translates the newly added \class{} \texttt{c2} and \firstTGGForwardRule{} translates the \package{} \texttt{nRootP}; this results in the triple graph (c).
Note that the last rule application creates a match for the repair rule \firstRRule{}.
This is the reason why we start our synchronization process with applications of forward rules.
The incremental pattern matcher notifies about two broken consistency matches, which are dealt with in random order.
\rootP{} is no longer a root package (which is detected by a violation of the according filter NAC in the consistency pattern) and \subP{} is now a root package (which is detected by the missing incoming edge).
Both violations are captured by repair rules, namely \firstRRule{} and \secondRRule{}, whose applications lead to~(d) and (e).
The algorithm terminates with a triple graph that belongs to the TGG.
\subsection{Prospect: Support of further kinds of editing and advanced TGG features}
\label{sec:prospect}
We shortly describe the support of further kinds of editing and more advanced features of TGGs by our approach to synchronization, namely attributed TGGs, rules with NACs, and support for additional attribute constraints.
\paragraph{Further kinds of editing.}
In our implementation (see Sect.~\ref{sec:implementation}), we do not only support the addition of elements and propagation of edits that correspond to source rules of derived edit rules.
Actually, we do not make any assumptions about the kind of editing.
This is achieved by incorporating the application of repair rules into the algorithm suggested by Leblebici et al.~\cite{LAFVS17,Leblebici18}, which has also been proved to be correct and to terminate.
The implemented algorithm first tries to apply a forward or repair rule.
If there is none available with a valid match, the algorithm falls back to revoking of an invalid rule application.
This means that all elements that have been created by this rule application are deleted (and adjacent edges of deleted nodes are implicitly deleted as well).
In line with that revoking of invalid rule applications, it also allows for implicit deletion of adjacent edges in the application of repair rules.
In that way, the application of a repair rule might trigger new appearances of broken consistency matches.
We are convinced that \emph{correctness} is not affected by that more general approach:
Inspecting the proofs of Corollary~\ref{cor:valid-repair-rule-applications} and Theorem~\ref{thm:correctness-algorithm}, the key to correctness is that the sequences of currently valid consistency matches remain consistently marking.
That is achieved via the conditions on matches for repair rules to be \emph{reversing}, \emph{context-preserving}, and \emph{creation-preserving}.
Dropping the condition to be \emph{non-disabling} (by implicitly deleting adjacent edges) does not effect correctness, therefore.
However, proving termination in that more general context is future work.
\paragraph{Advanced features.}
The {\em attribution of graphs} can be formalized by representing data values as special nodes and the attribution of nodes and edges as special edges connecting graph elements with these data nodes~\cite{EEPT06}.
As the rules of a TGG are monotonic, they only set attribute values but never delete or change them. (The deletion or change of an attribute value would include the deletion of the attribution edge pointing to it.)
The formal construction of short-cut rules is based purely on category-theoretic concepts, which can be directly applied to rules on attributed triple graphs as well.
The properties proven for short-cut rules in~\cite{FKST18} are valid also in that case.\footnote{To be precise, in~\cite{FKST18}, all proofs are elaborated for the case of monotonic rules in an \emph{adhesive} category. Attributed triple graphs are \emph{adhesive HLR} which is a weaker notion. However, inspecting the proofs, this does not make any difference as long as the category has so-called \emph{effective pushouts}. This is known to be the case for attributed (triple) graphs; compare, e.g.,~\cite[Remark~5.57]{EEGH15}.}
Hence, we can freely apply the construction of short-cut rules and derivation of repair rules to attributed TGGs.
In fact, our implementation already supports attribution.
For the propagation of attribute changes (made by a user), however, we rely on the inherent support eMoflon offers, which is discussed in Sect.~\ref{sec:implementation}.
Deriving repair rules to propagate such changes is possible in principle but remains future work.
In practical applications, TGGs are often not only attributed but also equipped with \emph{attribute constraints}.
These enable the user to, for example, link the values of attributes of correlated nodes.
eMoflon comes with facilities to detect violations of such constraints and offers support to repair such violations.
In our implementation, we rely on these features of eMoflon to support attribute constraints but do not contribute additional support in our newly proposed synchronization algorithm.
To summarize, while fully formalized for the case of plain TGG rules without attribution, our implementation already supports the synchronization of attributed TGGs with additional attribute constraints.
As these additional features do not affect our construction of short-cut and repair rules, we do not consider them (yet) to improve the propagation of attribute changes (that may lead to violations of attribute constraints).
Instead, we rely on the existing theory and facilities of eMoflon as introduced by Anjorin et al.~\cite{AVS12}.
In contrast, while computing short-cut and repair rules of rules with NACs is straightforward, adapting our synchronization algorithm to that case is future work and no tool support is available yet.
| 4db800b81c417caf6008c1c088ae58383b8d7dce | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subfile{sections/introduction}
\section{Background}
\subfile{sections/background}
\section{Related Work}
\subfile{sections/relatedwork}
\section{Experimental Setup}
\subfile{sections/experimentalsetup}
\section{Baseline Experiment}
\subfile{sections/baselineexperiment}
\section{Reinforcement Learning Experiment}
\subfile{sections/qlearningexperiment}
\section{Conclusion}
\subfile{sections/conclusion}
\bibliographystyle{./bibliography/IEEEtran}
\subsection{Knative Serverless Platform} \label{subsec:Knative}
As an open-source serverless platform, Knative provides a set of Kubernetes-based middleware components to support deploying and serving of serverless applications, including the capability to automatically scale resources on demand \cite{knativedocsautoscaling}.
The auto-scaling function is implemented by different serving components, described by the request flow in \Cref{fig:knative-autoscaling} based on Knative v0.12.
If a service revision is scaled to zero, i.e. the service deployment is reduced to a replica of null operating pods, the ingress gateway forwards incoming requests first to the activator \cite{knativedocsautoscaling}.
The activator then reports the information to the autoscaler, which instructs the revision's deployment to scale-up appropriately.
Further, it buffers the requests until the user pods of the revision become available, which can cause cold-start costs in terms of latency, as the requests are blocked for the corresponding time.
In comparison, if a minimum of one replica is maintained active, the activator is bypassed and the traffic can flow directly to the user pod.
\begin{figure}[t]
\centering
\includegraphics[width=8.5cm]{images/knative_autoscaling3.pdf}
\caption{Simplified request flow in Knative}
\label{fig:knative-autoscaling}
\end{figure}
When the requests reach the pod, they are channeled by the \emph{queue-proxy} container and, subsequently, processed in the \emph{user-container}.
The queue-proxy only allows a certain number of requests to enter the user-container simultaneously, and queues the requests if necessary.
The amount of parallel processed requests is specified by the \textit{concurrency} parameter configured for a particular revision.
By default, the value is set to a concurrency target of 100, defining how many parallel requests are preferred per user-container at a given time.
However, the user can explicitly restrict the number of concurrent requests by specifying a value between 0 and 1000 for the \textit{concurrency limit}.\footnote{A value of 0 allows unlimited concurrent requests and therefore results in no scaling \cite{openshiftknative}.}
Further, each queue-proxy measures the incoming load and reports the average concurrency and requests per second on a separate port.
The metrics of all queue-proxy containers are scraped by the autoscaler component, which then decides how many new pods need to be added or removed to keep the desired concurrency level.
\subsection{Q-learning}
RL refers to a collection of trial-and-error methods in which an agent is trained to make good decisions by interacting with his environment and receiving positive or negative feedback in form of rewards for a respective action.
A popular RL algorithm is the model-free Q-learning.
Q-learning stepwise trains an approximator $Q_{\theta}(s,a)$ of the optimal action-value function $Q^*$.
$Q_{\theta}(s,a)$ specifies the cumulated reward the agent can expect when starting in a state $s$, taking an action $a$, and then acting according to the optimal policy forever after.
By observing the actual reward in each iteration, the optimization of the Q-function is performed incrementally per step $t$:
\[Q(s_t,a_t)\xleftarrow{}(1-\alpha)Q(s_t,a_t) + \alpha[r_t + \gamma \max_{a} Q(s_{t+1},a)]\]
$\alpha$ describes the learning rate, i.e. to what extent newly observed information overrides old information and $\gamma$ a discount factor that serves to balance between the current and future reward.
As RL is a trial-and-error method, during training, the agent has to choose between the exploration of a new action and the exploitation of the current best option \cite{sutton2018rl}.
In research, this is often implemented with an $\epsilon$-greedy strategy, where $\epsilon$ defines the probability of exploration that usually decreases as the learning process advances \cite{rossi2019horizontal, mnih2013playing}.
With a probability of $1-\epsilon$, the agent selects based on the optimal policy and chooses the action that maximizes the expected return from starting in $s$, i.e. the action with the highest Q-value:
\[a^*(s) = \arg \max_{a} Q^*(s,a)\]
In the basic algorithm, the Q-values for each state-action combination are stored in a lookup table, the so-called \textit{Q-table}, indexed by states and actions.
The tabular representation of the agent's knowledge serves as a basis for decision-making during the entire learning episode.
\end{document}
\subsection{Design}
As outlined in the previous section, we use the application parameters \emph{bloat}, \emph{prime} and \emph{sleep} to simulate varying workload characteristics.
Starting with a no-operation workload where no parameters are passed, the memory allocation and CPU load were gradually increased for each new experiment.
The step size of the memory allocating parameter was aligned with the memory buckets commonly used for the standard pricing model of serverless platforms. To simulate compute-intensive and longer-lasting requests, different prime and sleep parameters were chosen correspondingly.
The detailed values are specified in \Cref{tab:concurrencyperformance}.
Per profile, we run performance tests for different concurrency levels according to the process flow described in \Cref{{fig:processflow}}.
Theoretically, the concurrency limit can take all values between 0 and 1000.
To keep the experiments computationally feasible, we proceed in steps of 20, starting at a concurrency limit of 10 and ending at 310.
As stated in related literature, we focus on latency and throughput as key performance measures of serverless applications \cite{li2019understanding}.
Average throughput is defined by requests per second (RPS), mean latency refers to the average time in seconds taken to return a response to a request.
To cover tail latency, we include the 95th percentile of latencies of all requests as an additional metric.
Furthermore, each test is repeated ten times to compensate for outliers or other fluctuations, before the concurrency is updated to the next limit.
\subsection{Results}
We structure the analysis of the baseline experiment results in three parts.
First, we examine the behavior of the individual workload profiles under different concurrency configurations.
Second, we focus on the relation of the target variables throughput and latency during the tests.
Finally, further metrics about resource utilization on container and pod level are analyzed.
As described above, we conducted the experiment for different combinations of the three parameters to simulate possible use cases.
Table \ref{tab:concurrencyperformance} gives an overview of the outcomes with the concurrency limit that lead to the optimal test result in terms of one of the performance measures.
Due to the numerous uncontrollable factors that influence the performance of the cluster, each result forms a snapshot in time.
The respective workload configuration is described by the three columns on the left.
Taking all tests into account, the smallest possible concurrency of 10 is the most common configuration that resulted in the best performance across all three indicators.
Interestingly, this does not correspond to the default setting of the KPA where a target concurrency value of 100 is preferred \cite{knativekpaconfig}.
In particular, workloads that consume memory exclusively perform better with fewer parallel requests per pod instance, e.g tests \#\rom{2}, \#\rom{9}, \#\rom{16} and \#\rom{17}.
Similar observations are made for workloads with additional low CPU usage, i.e. lower \emph{prime} parameter, as in tests \#\rom{3} and \#\rom{10}.
Deviations can be observed when the requests pause for a certain time.
These workloads result in higher throughput and lower mean and tail latency when a higher concurrency is chosen, e.g tests \#\rom{7}, \#\rom{8} and \#\rom{14}.
\begin{table}[t]
\begin{threeparttable}
\caption{Concurrency Performance Tests }
\label{tab:concurrencyperformance}
\begin{centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{Test}&\multicolumn{3}{|c|}{\textbf{Workload Profile}}&\multicolumn{3}{|c|}{\textbf{Conc. Limit Yielding Best Perf.}} \\
\cline{2-7}
\textbf{\#} & \textbf{\textit{bloat}}$^*$ & \textbf{\textit{prime}} & \textbf{\textit{sleep}}$^*$ & \textbf{\textit{thrghpt}} & \textbf{\textit{mean lat.}} & \textbf{\textit{95th lat.}}\\
\hline
\rom{1} & - & - & - & 50 & 50 & 70 \\
\hline
\rom{2} & 128 & - & - & 30 & 30 & 30 \\
\hline
\rom{3} & 128 & 1000 & - & 10 & 10 & 10\\
\hline
\rom{4} & 128 & 10.000 & - & 30 & 30 & 10\\
\hline
\rom{5} & 128 & 100.000 & - & 10 & 10 & 10\\
\hline
\rom{6} &128 & 1000 & 1000 & 110 & 110 & 150 \\
\hline
\rom{7} & 128 & 10.000 & 1000 & 70 & 70 & 70 \\
\hline
\rom{8} & 128 & 100.000 & 1000 & 110 & 110 & 110 \\
\hline
\rom{9} & 256 & - & - & 10 & 10 & 10 \\
\hline
\rom{10} & 256 & 1000 & - & 10 & 10 & 10 \\
\hline
\rom{11} & 256 & 10.000 & - & 10 & 10 & 10 \\
\hline
\rom{12} & 256 & 100.000 & - & 10 & 10 & 10 \\
\hline
\rom{13} & 256 & 1000 & 1000 & 50 & 50 & 50 \\
\hline
\rom{14} & 256 & 10.000 & 1000 & 110 & 110 & 110 \\
\hline
\rom{15} & 256 & 100.000 & 1000 & 10 & 10 & 30 \\
\hline
\rom{16} &512 & - & - & 10 & 10 & 10\\
\hline
\rom{17} & 1024 & - & - & 30 & 30 & 30\\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item $^*$ \textit{bloat} is defined in MB and \textit{sleep} in milliseconds.
\end{tablenotes}
\label{tab1}
\end{centering}
\end{threeparttable}
\end{table}
Depending on the workload, the distance between the optimal configuration and the second best concurrency can be very small, which becomes more evident when analyzing a single test in detail.
Fig. \ref{fig:thrpt_lat} shows the result of test \#\rom{7}, which is examined representatively.
Although the individual measurement points fluctuate, clear trends are identified in the average values.
A significant increase in throughput can be observed when the concurrency limit is raised to 70.
This setting also yields the lowest value for mean latency, differing from the second-best value at concurrency 50 by only 80 milliseconds.
The distance becomes more critical when considering the tail latency of the $95^{th}$ percentile, where a request takes more than 740 milliseconds on average longer to receive a response when compared to the most effective configuration.
At a concurrency of 10, the difference amounts to almost 3 seconds, further underlining the performance variations caused by the different settings.
The greatest slowdown in tail latency in this test occurs at a concurrency of 310 with more than 3.7 seconds.
Besides, the overall performance decreases strongly when the concurrency limit exceeds a level of 210.
This tendency can be found across the majority of tests, indicating that due to the high simultaneous processing of many requests, only a limited amount of resources are available for a single request.
Further observations show that with increasing memory utilization, i.e. the bloat parameter, performance tends to drop at lower concurrency limits.
In some cases, additionally, the success ratio strongly declines.
For example in test \#\rom{16}, from a concurrency of 170 onwards, more than 10\% of the requests received non-successful responses.
In test \#\rom{17} accordingly, this output can be observed from a concurrency of 90 onwards.
\begin{figure}[t]
\centering
\includegraphics[width=7.5cm]{images/test_7_thrpt_lat.pdf}
\caption{Performance of workload test \#\rom{7}}
\label{fig:thrpt_lat}
\end{figure}
Focusing on the target metrics, the tests show that adjusting the concurrency limit to an appropriate setting can yield significant improvements in throughput and latency.
Furthermore, an inverse behavior of the measures can be observed within the tests.
For the previously considered test \#\rom{7}, the results indicate a significant negative correlation of throughput and mean latency of $-0.989$, and a similar correlation for throughput and $95^{th}$-percentile latency of $-0.916$.\footnote{For all statistical tests, Pearson correlation coefficient is used with a two-sided p-value for testing non-correlation and an alpha level of $.001$.}
This strong negative relationship between the metrics is found across all tests, with significant correlation coefficients ranging from $-0.995$ to $-0.748$.\footnote{Except for test \#\rom{1} and \#\rom{13} with significant correlations of throughput and tail latency of $-0.629$, and throughput and mean latency of $-0.677$.}
Subsequently, an improvement in throughput usually results in a lower and more favorable latency.
This finding implies that there is no need to make trade-offs between different target metrics when adjusting the concurrency.
Instead, the problem can be reduced to one objective metric, representing the others.
\begin{comment}
To explore the impact on resource consumption in more detail, CPU and memory usage was measured for different infrastructure components.
Fig. \ref{fig:metrics_cpu_avg_sum} shows the average CPU usage of a user pod, and, broken down to a more detailed level, the usage of the two containers queue-proxy and user-container, which are deployed in each user pod.
The user-container, which runs the application and processes the requests, has a linear increase in CPU consumption relative to the number of concurrent requests.
The queue-proxy, which manages the forwarding of requests to the application and limiting of the concurrency, also shows an increase, however with a lower positive slope.
Considering only the containers, this implies that the higher the concurrency, the higher the utilization of a single queue-proxy, but the lower the total usage on the cluster, due to the relatively fewer number of containers, as shown by the green dashed line in Fig. \ref{fig:metrics_cpu_avg_sum}.
Since the consumption of a user-container is relatively higher, this has no significant effect on the overall usage at pod level, see mauve colored line.
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/test7_metrics_cpu_avg_sum.pdf}
\caption{CPU util. of user-container, queue-proxy and user pod on average per container/ pod instance (avg) and the sum over all container/ pod instances (sum) based on workload test \#\rom{7}}
\label{fig:metrics_cpu_avg_sum}
\end{figure}
Similar patterns occur when analyzing memory consumption.
However, in the user-container as in the queue-proxy a lower growth rate in the average consumption with increasing concurrency can be seen, resulting in the highest total consumption of all user pods at a concurrency of 10.
To summarize, a variation of the concurrency limit can influence the performance and in particular the average resource consumption of a container.
Although performance may be affected by numerous latent factors that are difficult to control, this baseline experiment reveals the potential for performance-enhancing measures through a more precise auto-scaling mechanism.
\end{comment}
\end{document}
\subsection{Cloud Architecture}
The overall architecture of our experiment is illustrated in Fig. \ref{fig:expsetup}.
To test the auto-scaling capabilities in an isolated environment, we set up two separate Kubernetes clusters, using IBM Cloud Kubernetes Service (IKS).
On the \textit{service cluster}, the sample service used for the experiments is deployed.
The cluster contains 9 nodes with 16 vCPU and 64 GB memory each, designed to provide sufficient capacity to host all Knative components and avoid performance limitations.
The \textit{client cluster} consists of one node with 16 vCPU and 64 GB memory responsible for sending requests to the service cluster to generate load.
The agent manages the activities on both clusters, including the configuration updates of the sample service based on collected metrics, and coordinates the process flow of the experiment, taking the role of an IKS user.
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/cluster_architecture_3.pdf}
\caption{Architectural setup including the information flow}
\label{fig:expsetup}
\end{figure}
The Knative resources are installed on the service cluster (version v0.12), including the serving components explained in Section \ref{subsec:Knative}, which control the state of the deployed sample service and enable auto-scaling of additional pods on demand.
Using the trial-and-error method of RL in the second experiment, we update the concurrency configuration of the service in each iteration.
To comprehensively test the auto-scaling capability, we activated the scale-to-zero functionality in the autoscaler's configmap, which requires a cold start in each iteration.
We further increased the replica number of ingress gateways, which handle load balancing, to bypass performance issues and to focus our studies exclusively on the auto-scaling functionalities.
\subsection{Workload}
Serverless computing is used for a variety of applications, accompanied by different resource requirements.
For example, the processing of video and image material or highly-parallel analytical workloads, such as MapReduce jobs, demand considerable memory and computing power.
Other applications, such as chained API compositions or chatbots, tend to be less compute-intensive but may require longer execution or response time.
To investigate the concurrency impact of many different workloads, we generate a synthetic, stable workload profile simulating serverless applications.
We use Knative's example \emph{Autoscale-go} application for this purpose, which allows different parameters to be passed with the request to test incremental variations of the workload characteristics and thus emulate varying CPU- and memory- intensive workloads \cite{knativeautoscale-go}.
The three application parameters are \emph{bloat}, \emph{prime} and \emph{sleep}, wherein the first is used to specify the number of megabytes to be allocated and the second to calculate the prime numbers up to the given number, to create either memory- or compute-intensive loads.
The sleep parameter pauses the request for the corresponding number of milliseconds, as in applications with certain waiting times.
\subsection{Process Flow} \label{subsec:processflow}
The basic process flow of one iteration is illustrated in Fig. \ref{fig:processflow}.
In each iteration the agent sends a concurrency update to the service cluster, which accordingly creates a new revision with the respective concurrency limit.
When the service update is complete, the agent sends the start signal to the client cluster, which begins issuing parallel requests against the service cluster.
To simulate a large number of user requests at the same time, we use the HTTP load testing tool \emph{Vegeta}, which features sending HTTP requests at a constant rate.
In the experiment, 500 requests are sent simultaneously over a period of 30s to ensure sufficient demand for scaling and sufficient time to provide additional instances.
After the last response is received, Vegeta outputs a report of the test results, including information on latency distribution of requests, average throughput and success ratio of responses.
The performance measures are then stored by the agent.
Additionally, the agent crawls metrics from the Knative monitoring components, exposed via a Prometheus-based HTTP API within the cluster, to get further information about resource usage at cluster, node, pod and container level.
Using this data, the concurrency update is chosen to proceed to the next iteration.
\begin{figure}[t]
\centering
\includegraphics[width=8.5cm]{images/Process_flow2.pdf}
\caption{Process flow}
\label{fig:processflow}
\end{figure}
\end{document}
\subsection{Design}
The process flow is based on the procedure from section \ref{subsec:processflow}, extended with a more sophisticated logic of the agent.
Instead of incrementally increasing the concurrency, the agent uses knowledge of the system environment (states) to test different concurrency updates (actions) and evaluates them by receiving scores (reward).
In each iteration, the environment is defined by the current state, which,
should provide a complete description of the system dynamics including all relevant information for optimal decision making.
Due to the large number of factors influencing performance, e.g. hidden cluster activities or network utilization, this is neither traceable nor computationally feasible in the used Q-learning algorithm.
Therefore, we break down our state space $S$ into three key features.
We define $S$ at time step $i$ as the combination of the state variables $s_i$ = $(conc_i, cpu_i, mem_i)$, where $conc_i$ depicts the concurrency limit, $cpu_i$ is the average CPU utilization per user-container and $mem_i$ is the average memory utilization per user-container.
The selection of the features is aligned with related research, with $conc_i$ as the equivalent of the number of VMs in VM auto-scaling approaches \cite{barrett2013applying, bitsakos2018derp}.
$cpu_i$ and $mem_i$ serve as a direct source of information about the resource utilization of a respective workload.
Since both CPU and memory utilization are continuous numbers, we discretize them into bins of equal size.
In each state $s_i \in S$, we define $A(s_i)$ as the set of valid actions, where $A$ is the set of all actions.
The agent can choose between decreasing, maintaining or increasing the concurrency limit by 20, i.e. $A$ = $\{-20,0,20\}$.
If the agent reaches the endpoints on the concurrency scale, i.e. the minimum or maximum concurrency, the action space in this state is reduced accordingly by the non-executable action.
After each iteration, the agent receives an immediate reward according to the performance achieved through the action.
In related literature, the reward is often based on the distance or ratio between the performance measure and a certain \textit{Service Level Agreement}, such as a throughput or response time target value \cite{horovitz2018efficient, dutreilh2011using}.
Since there is no target level to be achieved nor prior information about the performance given in our problem definition, we define an artificial reference value $ref\_value$ as the best value obtained to date.
Due to the permanent, albeit minor fluctuations in the measures, we propose a tolerance band around the reference value to avoid weighting minor non-relevant deviations.
Furthermore, the results from the preliminary study have shown a highly negative correlation between throughput and latency, i.e. higher throughput usually leads to lower and therefore better latency.
This relation in turn allows to focus exclusively on throughput ($thrghpt$) as one single objective.
The calculation of the reward $r$ in time step $i$ is as follows.
\[
r_i =
\begin{cases}
\frac{thrghpt_i}{ref\_{value}}
& \begin{array}{r@{}}
\text{if $thrghpt_i$} \leq \text{ref\_value} \cdot {0.95} \\
\text{or $thrghpt_i$} \geq \text{ref\_value} \cdot {1.05}
\end{array}\\
1 & \begin{array}{r@{}}
\text{else}
\end{array}\\
\end{cases}
\]
Q-learning is initiated with the following parameters.
A learning rate $\alpha$ = 0.5 is chosen to balance newly acquired and existing information, a discount factor $\gamma$ = 0.9 to ensure that the agent strives for a long-term high return.
To encourage the exploration of actions at the beginning of training, we implement a decaying $\epsilon$-greedy policy starting at iteration 50 with $\epsilon$ = 1 and then slowly decrease over time by a decay factor of 0.995 per iteration.
The minimum exploration probability is set to $\epsilon_{min}$ = 0.1, to allow for the detection of possible changes in the system.
The knowledge the agent acquires is stored in a Q-table and updated each iteration.
To examine whether the model can effectively learn the concurrency values identified in section \ref{sec:baseline} as high throughput configurations, the results are analyzed representatively based on workload test \#\rom{7} and \#\rom{10}.
The former test showed high performance at a concurrency limit of 70, while the second reached the best test results at an edge concurrency of 10.
\subsection{Results}
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/qlearning_10_thrpt_conc_epsilon.pdf}
\caption{Performance of Q-learning model, workload \#\rom{10}}
\label{fig:qlearning1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/qlearning_7_thrpt_conc_epsilon.pdf}
\caption{Performance of Q-learning model, workload \#\rom{7}}
\label{fig:qlearning2}
\end{figure}
First, we analyze the results to examine the suitability of the proposed Q-learning-based model to fine-tune the auto-scaling.
Second, we evaluate the performance of the approach in terms of throughput improvements compared to Knative's default auto-scaling configuration.
Based on workload profile \#\rom{10}, Fig. \ref{fig:qlearning1} shows in detail how the agent applies RL logic to incrementally change the concurrency and to adjust it as the training progressed.
Beginning at concurrency 170, random exploration leads to a moderate decline of the concurrency limit in the first 30 iterations.
This results in an improvement in throughput in this test, captured by the rewards and corresponding Q-values for each state-action combination.
The most effective scaling policy of 10 parallel requests per container is first reached in iteration 121.
Nevertheless, due to the $\epsilon$-greedy strategy, exploratory actions are chosen, which might differ from the down-scaling decision and cause the agent to deviate from a good strategy.
As training progresses, a trend towards performance-enhancing concurrency configurations can be observed, indicating the agent is more likely to exploit the optimal decision rather than exploring.
After 330 iterations, the concurrency stabilizes at a limit of 10 parallel requests per container, implying the agent has learned the correct scaling policy, according to the results from section \ref{sec:baseline}.
Due to the minimum $\epsilon$ = 0.1, exploration still rarely occurs to ensure the agent can respond to changes in the environment.
A different learning process of the proposed Q-learning approach can be observed for workload \#\rom{7}, depicted in Fig. \ref{fig:qlearning2}.
The varying concurrency curve shows the initial strategy of the agent exploring first the higher state space before proceeding with lower concurrency limits.
After 250 iterations the exploitation phase outweighs and the concurrency gradually levels off.
In comparison to workload \#\rom{10}, where the algorithm's scaling policy converges to a single concurrency limit, the configuration here fluctuates, mainly between 50 and 70, and retains this pattern.
Further differences between the test results arise from the throughput metric, which shows strong fluctuations between 230 and 415 RPS across all iterations.
The deviations, which also appear within one concurrency setting, considerably impair the agent's ability to evaluate suitable state-action-pairs via the reward function.
Nevertheless, the agent is able to narrow down the scaling range to a limited number of values at which it identified the best outcomes in terms of throughput, and which agrees with the result of the baseline experiment in Section \ref{sec:baseline}.
\begin{comment}
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/qlearning_10_reward_thrpt.pdf}
\caption{Throughput and reward of workload \#\rom{10}}
\label{fig:qlearning10reward}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/qlearning_7_reward_thrpt.pdf}
\caption{Throughput and reward of workload \#\rom{7}}
\label{fig:qlearning7reward}
\end{figure}
\end{comment}
To evaluate the proposed scaling policies, we benchmark the average performance of the Q-learning-based approach with the static default setting.
For this purpose, the same experimental setup is used as in the Q-learning test, except for the auto-scaling configuration, where the original setting of a concurrency target of 100 is applied \cite{knativedocsautoscaling}.\footnote{Additionally, the container target percentage is set to 0.7 as in the default configmap.}
Fig. \ref{fig:qlearning_avgthrpt} depicts the average throughput up to the respective iteration of the Q-learning model and the default configuration for the two considered workloads.
Both result in the Q-learning model outperforming the test based on Knative's standard settings.
Considering workload \#\rom{7} first, the model requires approximately 150 iterations until the average performance reaches default-level.
Subsequently, the throughput increases to an average of 400 RPS providing a minor advantage of 20 RPS compared to the standard system.
A more significant enhancement shows workload \#\rom{10}.
While in the first 10 iteration the default settings alternate between 350 and 440 RPS, converging to an average of about 393 RPS, the performance of our model is initially lower.
However, with ongoing learning the average throughput improves and excels already from iteration 10 onwards.
After 600 iterations, the presented Q-learning based model reaches an average throughput of 740 RPS, hence achieving more than 80\% of the performance of the default setting, which stabilizes at 390 RPS on average.
To summarize the results, the proposed model learned within finite time a scaling policy that outperforms the default Knative configuration in terms of throughput, proving the Q-learning-based approach is well-feasible to refine the auto-scaling mechanism.
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{images/qlearning_avgthroughput.pdf}
\caption{Comparison of average throughput of the Q-learning model and the Knative default auto-scaling setting}
\label{fig:qlearning_avgthrpt}
\end{figure}
\end{document}
\subsection{Serverless computing}
With the growing number of serverless computing offerings, there has been an increasing interest of the academic community in comparing different solutions, with scalability being one of the key elements of evaluation \cite{li2019understanding}.
In multiple works, different propriety serverless platforms were benchmarked, including their ability to scale, focusing on Amazon Lambda, Microsoft Azure Functions \cite{lloyd2018serverless}, along with Google Cloud Functions \cite{wang2018peeking} and additionally IBM Cloud Functions \cite{lee2018evaluation}.
Similar studies have been carried out in the area of open-source serverless frameworks, with greater attention paid to the auto-scaling capabilities.
Mohanty et al. \cite{mohanty2018evaluation} evaluated Fission, Kubeless, and OpenFaaS and concluded that Kubeless provides the most consistent performance in terms of response time.
Another comparison of both qualitative and quantitative features of Kubeless, OpenFaas, Apache Openwhisk, and Knative, comes to the same conclusion, albeit generally indicating the limited user control over custom Quality of Service (QoS) requirements \cite{palade2019evaluation}.
These studies solely consider the default auto-scaler Kubernetes HPA.
Possible adjustments to the auto-scaling mechanism itself are not further examined.
Li et al. \cite{li2019understanding} propose a more concrete distinction between resource-based and workload-based scaling policies.
The authors compare the performance of different workload scenarios using the tuning capability of concurrency levels in Knative and clearly suggest further investigation of the applicability of this auto-scaling capability, which also motivates this research.
\subsection{Auto-scaling}
As elasticity is one of the main characteristics of the increasing adaption of cloud computing, the automatic, on-demand provisioning and de-provisioning of cloud resources have been the subject of intensive research in recent years \cite{singh2019research}.
We discuss related work under two aspects: first, the underlying theories on which auto-scaling is built with a focus on RL, and second, the entities being scaled.
To classify numerous techniques at the algorithmic level, different taxonomies were proposed, where the predominant categories are threshold-based rules, queuing theory and RL \cite{lorido2014review,singh2019research}.
In the former, scaling decisions are made on predefined thresholds and are most popular among public cloud providers, e.g. Amazon ECS \cite{AWSservicescaling}.
Despite the simplistic implementation, identifying suitable thresholds requires expert knowledge \cite{singh2019research}, or explicit application understanding \cite{dutreilh2010data}.
Queuing theory has been used to mathematically model applications \cite{lorido2014review}.
As they usually impose a stationary system, the models are
less reactive towards changes \cite{lorido2014review}.
In contrast, RL offers an interesting approach through online learning of the most suitable scaling action and without the need for any a-priori knowledge \cite{lorido2014review}.
Many authors have therefore investigated the applicability of model-free RL algorithms, such as Q-learning, in recent years \cite{rossi2019horizontal}.
Dutreihl et al. \cite{dutreilh2010data} show that although Q-learning based VM controlling requires an extensive learning phase and adequate system integration, it can lead to significant performance improvements compared to threshold-based auto-scaling, since thresholds are often set too tightly while seeking for the optimal resource allocation.
To combine the advantages of both, Q-learning itself can be used to automatically adapt thresholds to a specific application \cite{horovitz2018efficient}.
In terms of the entity being scaled, RL has been mostly applied to policies for VM allocation and provisioning, e.g. in \cite{bitsakos2018derp}.
With the emergence of container-based applications, this field has become a greater focus of research \cite{rossi2019horizontal}.
In both areas, the scope of action is concentrated mainly on horizontal (scale-out/-in) \cite{horovitz2018efficient}, vertical scaling (scale-up/-down) \cite{rao2011distributed}, or the combination of both \cite{rossi2019horizontal}.
However, little research has been done in areas that extend the classic auto-scaling problem of VM or container configuration.
As a novel approach we investigate the applicability of Q-learning to request-based auto-scaling in a serverless environment.
Differently from the existing work on direct vertical or horizontal scaling with RL, we propose a model that learns an effective scaling policy by adapting the level of concurrent requests per container instance to a specific workload.
\end{document} | 825c4d58557fb8919ee6a4076118818b260a6709 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Machine learning (ML) \cite{bishop_pattern_2011,cover_elements_1991,hastie2009elements,hundred} refers to a broad field of study, with multifaceted applications of cross-disciplinary breadth. ML {is a subset of Artificial Intelligence (AI) which} ultimately aims at developing computer algorithms that improve automatically through experience. The core idea is that systems can learn from data, so as to identify distinctive patterns and make consequently decisions, with minimal human intervention. The range of applications of ML methodologies is extremely vast \cite{sutton2018reinforcement,graves2013speech,sebe2005machine,grigorescu2020survey}, and still growing at a steady pace due to the pressing need to cope with the
efficiently handling of big data \cite{chen2014big}. Biomimetic approaches to sub-symbolic AI \cite{rosenblatt1961principles} inspired the design of powerful algorithms. These latter sought to reproduce the unconscious process underlying fast perception, the neurological paths for rapid decision making, as e.g. employed for faces \cite{meyers2008using} or spoken words \cite{caponetti2011biologically} recognition.
An early example of a sub-symbolic brain inspired AI was the perceptron \cite{rosenblatt1958perceptron}, the influential ancestor of deep neural networks (NN) \cite{bengio2007greedy,Goodfellow-et-al-2016}. The perceptron is indeed an algorithm for supervised learning of binary classifiers. It is a linear classifier, meaning that its forecasts are based on a linear prediction function which combines a set of weights with the feature vector. Analogous to neurons, the perceptron adds up its input: if the resulting sum is above a given threshold the perceptron fires (returns the output the value 1) otherwise it does not (and the output equals zero). Modern multilayer perceptrons, account for multiple hidden layers with non-linear activation functions. The learning is achieved via {minimizing the classification error.}
{Single or multilayered perceptrons should be trained by examples \cite{bengio2007greedy,hinton2006fast,rumelhart1988learning}. Supervised learning requires indeed a large set of positive and negative examples, the training set, labelled with their reference category.}
The perceptrons' acquired ability to perform classification is eventually stored in a finite collection of numbers, the weights and thresholds that were learned during the successive epochs of the supervised training. To date, it is not clear how such a huge collection of numbers (hundred-millions of weights in state of the art ML applications) are synergistically interlaced for the deep networks
to execute the assigned tasks, with an exceptional degree of robustness and accuracy \cite{xie2020explainable,hinton2015distilling,erhan2010understanding}.
Starting from these premises, the aims of this paper are multifold. On the one side, we will develop a novel learning scheme which is anchored on reciprocal space. Instead of {iteratively} adjusting the weights of the edges that define the connection among nodes, we will modify {the spectra of a collection of suitably engineered matrices that bridge adjacent layers. To eventually recover a multilayered feedforward architecture in direct space, we postulate a nested indentation of the associated eigenvectors. These latter act as the effective gears of a processing device operated in reciprocal space. The directed indentation between stacks of adjacent eigenvectors yield a compression of the activation pattern, which is eventually delivered to the detection nodes.}
{As a starting point, assume eigenvectors are frozen to a reference setting which fulfills the prescribed conditions. The learning is hence solely restricted to the eigenvalues, a choice which amounts to performing a {\it global} training, targeted to identifying key collective modes, the selected eigen-directions, for carrying out the assigned classification task. The idea of conducting a global training on a subset of parameters has been also proposed in other works \cite{frankle2020training, Gabri__2019}. This is at odd with the usual approach to machine learning where {\it local} adjustments of pairwise weights are implemented in direct space. As we shall prove, by tuning the eigenvalues, while freezing the eigenvectors, yields performances superior to those reached with usual (local) techniques bound to operate with an identical number of free parameters, within an equivalent network architecture. Eigenvalues are therefore identified as key target of the learning process, proving more fundamental than any other set of identical cardinality, allocated in direct space. Remarkably, the distribution of weights obtained when applying the spectral learning technique restricted to the eigenvalues is close to that recovered when training the neural network in direct space, with no restrictions on the parameters to be adjusted. In this respect, spectral learning bound to the eigenvalues could provide a viable strategy for pre-training of deep neural networks. Further, the set of trainable eigenvalues can be expanded at will by inserting linear processing units between the adjacent layers of a non-linear multilayered perceptron. Added linear layers act as veritable booms of a {\it telescopic neural network}, which can be extracted during the learning phase and retracted in operational mode, yielding compact networks with improved classification skills. The effect of the linear expansion is instead negligible, if applied to neural learning of standard conception. The entries of the indented eigenvectors can be also trained resulting in enhanced performance, as compared to the setting where eigenvalues are exclusively modulated by the learning algorithm. To demonstrate the principles which underly spectral training, we employ the MNIST database, a collection of handwritten digits to be classified. The examined problem is relatively simple: a modest number of tunable parameters is indeed necessary for achieving remarkable success rates. When allowing for the simultaneous training of the eigenvalues and (a limited fraction of ) eigenvectors, the neural networks quickly saturates to accuracy scores which are indistinguishable from those obtained via conventional approaches to supervised learning. More challenging tasks should be probably faced to fully appreciate the role played by a
progressive optimization of the eigenmodes, the collective directions in reciprocal space where information flows. As remarked above, the eigenvectors have been here constructed so as to yield a feedforward multi-layered architecture in direct space. By relaxing this assumption, comes to altering the network topology and thus exporting the spectral learning strategy to other frameworks, as e.g. reservoir computing. In general terms, working in the spectral domain corresponds to optimizing a set of {\it non orthogonal} directions (in the high dimensional space of the nodes) and associated weights (the eigenvalues), a global outlook which could contribute to shed novel light on the theoretical foundations of supervised learning.}
{\section{Linear and non-linear spectral learning}}
To introduce and test the proposed method we will consider a special task, i.e. recognition of handwritten digits. To this end, we will make use of the MNIST database \cite{lecun1998mnist} which has a training set of 60,000 examples, and a test set of 10,000 examples. Each image is made of $N_1=28 \times 28$ pixels and each pixel bears an 8-bit numerical intensity value, see Fig. \ref{fig1}. A deep neural
network can be trained using standard backpropagation \cite{bengio2007greedy} algorithms to assign the weights that link the nodes (or perceptrons) belonging to consecutive layers. The first layer has $N_1$
nodes and the input is set to the corresponding pixel's intensity. The highest error rate reported on the original website of the database \cite{lecun1998mnist} is 12 \%, which is achieved using a simple linear classifier, with no preprocessing. In early 2020, researchers announced 0.16 \% error \cite{byerly2020branching} with a deep neural network made of branching and merging convolutional networks. Our goal here is to contribute to the analysis with a radically different approach to the learning, {rather than joining the efforts to break current limit in terms of performance and classification accuracy. More specifically, and referring to the MNIST database as a benchmark application, we will assemble a network made of $N$ nodes, organized in successive $\ell$ layers, tying the training to reciprocal space.}
{Directed connections between nodes belonging to consecutive layers are encoded in a set of $\ell-1$, $N \times N$ adjacency matrices. The eigenvectors of these latter matrices are engineered so as to favour the information transfer from the reading frame to the output layer, upon proper encoding. The associated eigenvalues represent the primary target of the novel learning scheme. In the following we will set up the method, both with reference to its linear and non-linear versions. Tests performed on the MNIST database are discussed in the next Section.}
{\subsection{Linear spectral learning: Single-layer perceptron trained in reciprocal space}}
Assume $N_i$ to label the nodes assigned to layer $i$, {and define} $N=\sum_{i=1}^{\ell} N_i$. {For the specific case here inspected the} output layer is composed by ten nodes ($N_{\ell}=10$), where recognition takes eventually place. Select one image from the training set and be $n_1$ $(=0,1,2..,9)$ the generic number therein displayed. We then construct a column vector $\vec{n}_1$, of size $N$, whose first $N_1$ entries are the intensities displayed on the pixels of the selected image (from the top-left to the bottom-right, moving horizontally), as illustrated in Fig. \ref{fig1}. {All other entries are initially set to zero. As we shall explain in the following, our goal is to transform the input $\vec{n}_1$ into an output vector with same dimensions. The last $N_{\ell}$ elements of this latter vector represent the output nodes where reading is eventually performed.}
\begin{figure}
\centering
\includegraphics[scale=0.3]{Fig1.png}
\caption{\it Each image of the training set is mapped into a column vector $\vec{n}_1$, of size $N$, whose first $N_1 = 28 \times$ 28 entries are the intensities displayed on the pixels of the image.}
\label{fig1}
\end{figure}
To set the stage, we begin by reporting on a simplified scenario that, as we shall prove in the following, yields a single layer perceptron. The extension to multi-layered architectures will be discussed {right after}.
{Consider the entry layer made of $N_1$ nodes and the outer one composed of $N_2$ elements. In this case $N=N_1+N_2$. The input vector $\vec{n}_1$ undergoes a linear transformation to yield $\vec{n}_2=\mathbf{A}_1 \vec{n}_1$ where $\mathbf{A}_1$ is a $N\times N$ matrix that we shall characterize in the following. Introduce matrix $\Phi_1$: this is the identity matrix $\mathbb{1}_{N \times N}$ modified by the inclusion of a sub-diagonal block $N_{2} \times N_{1}$, e.g. filled with uniformly distributed random numbers, defined in a bounded interval, see Fig. \ref{fig2}. The columns of $\Phi_1$, hereafter $\left(\vec{\phi}_1\right)_k$ with $k=1,...,N$, define a basis of the $N$ dimensional space to which $\vec{n}_1$ and $\vec{n}_2$ belong. Then, we introduce the diagonal matrix $\Lambda_1$. The entries of $\Lambda_1$ are set to random (uniform) numbers spanning a suitable interval. A straightforward calculation returns $\left(\Phi_1\right)^{-1}=2 \mathbb{1}_{N \times N}-\Phi_1$. We hence define $\mathbf{A}_1= \Phi_1 \Lambda_1 \left(2 \mathbb{1}_{N \times N}-\Phi_1\right)$ as the matrix that transforms $\vec{n}_1$ into $\vec{n}_2$. Because of the specific structure of the input vector, and owing the nature of $\mathbf{A}_1$, the information stored in the first $N_1$ elements of $\vec{n}_1$ is passed to the $N_2$ successive entries of $\vec{n}_2$, in a compactified form which reflects both the imposed eigenvectors' indentation and the chosen non trivial eigenvalues. }
{To see this more clearly, expand the $N$-dimensional input vector $\vec{n}_1$ on the basis made of $\left( \vec{\phi}_1 \right)_k$ to yield $\vec{n}_1=\sum_{k=1}^N c_k \left( \vec{\phi}_1 \right)_k$ where $c_k$ stands for the coefficients of the expansion. The first $N_1$ vectors are necessarily engaged to explain the non zero content of $\vec{n}_1$ and, because of the imposed indentation, rebound on the successive $N_2$ elements of the basis. These latter need to adjust their associated weights $c_k$ to compensate for the echoed perturbation. The action of matrix $\mathbf{A}_1$ on the input vector $\vec{n}_1$ can be exemplified as follows:}
\begin{equation}
\vec{n}_2 = \mathbf{A}_1 \vec{n}_1 = \mathbf{A}_1 \sum_{k=1}^N c_k \left( \vec{\phi}_1 \right)_k = \sum_{k=1}^{N_1+N_2} c_k \left( \Lambda_1 \right)_k \left( \vec{\phi}_1 \right)_k
\end{equation}
{where $\left( \Lambda_1 \right)_k$ are the element of matrix $\Lambda_1$. In short, the entries of $\vec{n}_2$ from position $N_1+1$ to position $N_1+N_2$ represent a compressed (if $N_2<N_1$) rendering of the supplied input signal, the key to decipher the folding of the message being stored in the $N_{2} \times N_{1}$ sub-diagonal block of $\Phi_1$, (i.e. the eigenvector indentation) and in the first set of $N=N_1+N_2$ eigenvalues $\left( \Lambda_1 \right)_k$. The key idea is to propagate this message passing scheme, from the input to the output in a multi-layer setting, and adjust (a subset of) the spectral parameters involved so as to optimize the encoding of the information.}
\begin{figure}
\centering
\includegraphics[scale=0.5]{LineareAdiacenza.pdf}
\caption{\it {Panel (a): the structure of matrix $\Phi_k$ is schematically depicted. The diagonal entries of $\Phi_k$ are unities. The sub-diagonal block of size
$N_{k+1} \times N_{k}$ for $k=1,\ell-1$ is filled with uniform random numbers in $[a,b]$, with $a,b \in \mathbb{R}$. These blocks yields an effective indentation between successive stacks of linearly independent eigenvectors. The diagonal matrix of the eigenvalues $\Lambda_k$ is also represented. The sub-portions of $\Phi_k$ and $\Lambda_k$ that get modified by the training performed in spectral domain are highlighted (see legend). In the experiments reported in this paper the initial eigenvectors entries are uniform random variables distributed in $[-0.5,0.5]$. The eigenvalues are uniform random numbers distributed in the interval $[-0.01,0.01]$. Optimizing the range to which the initial guesses belong (for both eigenvalues and eigenvectors) is an open problem that we have not tackled.
Panel (b): a $(N_1 + N_\ell) \times (N_1 + N_\ell)$ matrix $\mathcal{A}_c$ can be obtained from $\mathcal{A}=\left( \Pi_{k=1}^{\ell-1} \mathbf{A}_{k} \right) $, which provides the weights for a single layer perceptron, that maps the input into the output, in direct space.}}
\label{fig2}
\end{figure}
{ To this end, we introduce the $N \times N$ matrix operator $\Phi_k$, for $k=2,...,\ell-1$. In analogy with the above, $\Phi_k$ is the identity matrix $\mathbb{1}_{N \times N}$ modified with a sub-diagonal block $N_{k+1} \times N_{k}$, which extends from rows $N_k$ to $N_k+N_{k+1}$, and touches tangentially the diagonal, as schematically illustrated in Fig. \ref{fig2} (a). Similarly, we introduce $\Lambda_k$, for $k=2,...,\ell-1$, which is obtained from the identity matrix $\mathbb{1}_{N \times N}$ upon mutating to uniformly distributed random entries the diagonal elements that range from $\sum_{i=1}^k N_i$ (not included) to $\sum_{i=1}^{k+1} N_i$ (included). Finally, we define $\mathbf{A}_k= \Phi_k \Lambda_k \left(2 \mathbb{1}_{N \times N}-\Phi_k\right)$, as the matrix that transforms $\vec{n}_k$ into $\vec{n}_{k+1}$, with $k=2,...,\ell-1$. In principle, both non trivial eigenvalues' and eigenvectors' input can be self-consistently adjusted by the envisaged learning strategy. The input signal $\vec{n}_1$ is hence transformed into an output vector $\vec{n}_{\ell}$ following a cascade of linear transformations implemented via matrices $\mathbf{A}_k$. In formulae:}
\begin{equation}
\vec{n}_{\ell} = \mathbf{A}_{\ell-1} ...\mathbf{A}_{1} \vec{n}_1 = \left( \Pi_{k=1}^{\ell-1} \Phi_k \Lambda_k \left(2 \mathbb{1}_{N \times N}-\Phi_k\right)\right) \vec{n}_1
\end{equation}
{where in the last step we made use of the representation of $\mathbf{A}_k$ in dual space. The generic vector $\vec{n}_{k+1}$, for $k=1,..., \ell-1$ is obtained by applying matrix $\mathbf{A}_k$ to $\vec{n}_k$. The first
$N_1+N_2+...+N_k$ components of $\vec{n}_{k+1}$ coincide with the corresponding entries of $\vec{n}_k$, namely $\left[ \vec{n}_{k+1} \right]_m \equiv \left[ \vec{n}_{k} \right]_m$ for $m<N_1+N_2+...+N_k$. Here,
$\left[ \left(\vec{\cdot}\right) \right]_m$ identifies the $m$-th component of the vector $\left(\vec{\cdot}\right)$. Recall that, by construction, $\left[ \vec{n}_{k} \right]_m=0$ for $m> N_1+N_2+...+N_k$. On the contrary, the components
$\left[ \vec{n}_{k+1} \right]_m$ with $N_1+N_2+...+N_k+1<m<N_1+N_2+...+N_k+N_{k+1}$ are populated by non trivial values which reflect the eigenvectors indentation, as well as the associated eigenvalues. This observation can be mathematically proven as follows. Write $\vec{n}_k$ on the basis formed by the eigenvectors
$\left( \vec{\phi}_k \right)_l$ to eventually get:}
\begin{equation}
\vec{n}_k = \sum_{l=1}^{N_1+N_2+...+N_{k+1}} c_l \left( \vec{\phi}_k \right)_l \equiv \sum_{l=1}^{N_1+N_2+...+N_k} c_l \vec{e}_l
\end{equation}
{where $\left(\vec{e}_1, \vec{e}_2...\right)$ stand for the canonical basis and the last inequality follows the specific structure of the eigenvectors (remark that the leftmost sum in the above equation includes $N_{k+1}$ more elements than the second). By definition:}
\begin{equation}
\vec{n}_{k+1} = \mathbf{A}_k \vec{n}_k = \sum_{l=1}^{N_1+N_2..+N_{k+1}} c_l \left( \Lambda_k \right)_l \left( \vec{\phi}_k \right)_l
\end{equation}
{From the above relation, one gets for $m \le N_1+N_2+...+N_k$}
\begin{equation}
\left[ \vec{n}_{k+1} \right]_m = \sum_{l=1}^{N_1+N_2..+N_{k}} c_l \left[ \vec{e}_l \right]_m \equiv \left[ \vec{n}_{k} \right]_m
\end{equation}
{where the first equality sign follows from the observation that $\left( \vec{\phi}_k \right)_l$ coincides with $\vec{e}_l$ and $\left( \Lambda_k \right)_l=1$,
over the explored range of $m$. For $N_1+N_2+...+N_k+1 \le m \le N_1+N_2+...+N_k+N_{k+1}$, we obtain instead:}
\begin{equation}
\left[\vec{n}_{k+1}\right]_m = \sum_{l=N_1+N_2..+N_{k-1}}^{N_1+N_2..+N_{k+1}} c_l \left( \Lambda_k \right)_l \left[ \left( \vec{\phi}_k \right)_l \right]_m
\end{equation}
{Finally, it is immediate to show that $\left[\vec{n}_{k+1}\right]_m=0$ for $m> N_1+N_2+...+N_k+N_{k+1}$, because of the specific form of the employed eigenvectors. In short, the information contained in the last non trivial $N_k$ entries of $\vec{n}_k$ rebound on the successive $N_{k+1}$ elements of $\vec{n}_{k+1}$, funnelling the information downstream from the input to the output.
The successive information processing relies on the indented (non orthogonal) eigenvectors and the associated eigenvalues, which hence define the target of the training in reciprocal space.
}
{ To carry out the learning procedure one needs to introduce a loss function $L(\vec{n}_1)$. For illustrative purposes this latter can be written as:}
\begin{equation}
\label{LF}
L(\vec{n}_1) = \left\lVert l(\vec{n}_1) - \sigma\left[ \left( \Pi_{k=1}^{\ell} \Phi_k \Lambda_k \left(2 \mathbb{1}_{N \times N}-\Phi_k\right) \right) \vec{n} _1\right ] \right\lVert^2
\end{equation}
{ where $\sigma(\cdot)$ is the softmax operation applied to the last entries of the $\ell$-th image of the input vector $\vec{n}_1$. In the above expression, $l(\vec{n}_1)$ stands for the label attached to $\vec{n}_1$ depending on its category. More into details, the $k$-th entry of $l(\vec{n}_1)$ is equal unit (and the rest identically equal to zero) if the number supplied as an input is identical to $k$, with
$k=0,1,...,9$. The loss function can be minimized by acting on the free parameters of the learning scheme. Specifically, the learning can be restricted to the set of $N$ non trivial eigenvalues, split in $\ell$ distinct groups, each referred to one of the $\mathbf{A}_{k}$
matrices (i.e. $N_1+N_2$ eigenvalues of $\mathbf{A}_{1}$, $N_3$ eigenvalues of $\mathbf{A}_{2}$,...., $N_{\ell}$ eigenvalues of $\mathbf{A}_{\ell-1}$). In addition, the
sub-diagonal block entries of $\Phi_k$, the elements of the basis which dictate the successive indentation between adjacent layers, can be adjusted as follows the training scheme. In the following section we will report about the performance of the method, implemented in its different modalities, against those obtained with a classical approach to the learning anchored in direct space. In the actual implementation we have chosen to deal with a categorical cross-entropy loss function.}
{Before ending this section a few remarks are mandatory. Introduce $\mathcal{A} = \Pi_{k=1}^{\ell} \mathbf{A}_{k}$. The linear transformation that links the input vector $\vec{n}_1$ to the generated output $\vec{n}_{\ell}$, can be compactly expressed as $\vec{n}_{\ell} = \mathcal{A} \vec{n}_1$. Then, recall that the classification relies on examining the last $N_\ell$ entries of $\vec{n}_{\ell}$.
Hence, for the specific setting here examined, where the mapping is obtained as a cascade of linear transformations, one can imagine to recast the whole procedure in a space of reduced dimensionality. Be
$\vec{z}$ a column vector made of $N_1 + N_\ell$ elements. The first $N_1$ entries of $\vec{z}$ are the intensities on the pixels of the selected image, as for the homologous $\vec{{n}}_1$ quantity. The other elements are set to zero. Then, consider the $(N_1 + N_\ell) \times (N_1 + N_\ell)$ matrix $\mathcal{A}_c$ (the label $c$ stands for {\it compact}), constructed from $\mathcal{A}$ by trimming out all the information that pertain to the intermediate layers, as introduced in the reciprocal space (see Fig. \ref{fig2}(b)). Stated differently, matrix $\mathcal{A}_c$ provides the weighted links that feed from the input to the output layer in direct space, via the linear transformation $\mathcal{A}_c \vec{z}$: this is a single layer perceptron, shown in Fig. \ref{fig2}(b), which was trained by endowing reciprocal space with an arbitrary number of additional dimensions, the intermediate stacks responsible for the sequential embedding of the information. Intermediate layers can be literally extracted, during the training phase, and subsequently retracted in operational mode. The importance to allowing for additional layers, and so provide the neural network of a telescopic attribute, will be assessed in the forthcoming sections.}
{From the algorithmic point of view the process outlined above can be rephrased in simpler, although equivalent terms. For all practical purposes, one could take the (column) input vector $\vec{{n}}_1$ to have $N_1+N_2$ elements. Following the scheme depicted above, the first $N_1$ entries are the intensities on the pixels of the selected image, while the remaining $N_2$ elements are set to zero. We now introduce a $(N_1+N_2) \times (N_1+N_2)$ matrix $\mathbf{A}_{1}$. This is the identity matrix $\mathbb{1}_{(N_1+N_2) \times (N_1+N_2)}$ with the inclusion of a sub-diagonal block $N_{2} \times N_{1}$, which handles the information processing that will populate the second $N_2$ elements of the output vector $\vec{{n}}_2= \mathbf{A}_{1} \vec{{n}_1}$. Then, we formally replace the $(N_1+N_2)$ column vector $\vec{{n}}_2$ with a column vector made of $(N_2+N_3)$ elements, termed $\vec{{n}}_{2t}$, whose first $N_2$ elements are the final entries of $\vec{{n}}_2$. The remaining $N_3$ elements of $\vec{{n}}_{2t}$ are set to zero. Now, rename $\vec{{n}}_{2t}$ as $\vec{{n}}_2$ and presents it as the input of a $(N_2+N_3) \times (N_2+N_3)$ matrix $\mathbf{A}_{2}$, with a non trivial sub-diagonal $N_{3} \times N_{2}$ block. This latter maps the first $N_2$ elements of the input vector, into the successive $N_3$ of the output one, by completing the second step of an algorithmic scheme which can be iteratively repeated. In analogy with the above, each $(N_k+N_{k+1}) \times (N_k+N_{k+1})$ matrix $\mathbf{A}_{k}$ can be written as $\mathbf{A}_k= \Phi_k \Lambda_k \left(2 \mathbb{1}_{(N_k+N_{k+1}) \times (N_k+N_{k+1})}-\Phi_k\right)$, where now the column vectors of $\Phi_k$ are the eigevenctors of $\mathbf{A}_k$ and form a non-orthogonal basis of the $(N_k+N_{k+1})$
space where input and output vectors belong. $\Lambda_k$ is a diagonal matrix of the eigenvalues: the first $N_k$ are set to one, while the other $N_{k+1}$ are non trivial entries to be adjusted self-consistently via the learning scheme. Framing the process in the augmented space of $N$ dimensions, as done earlier, allows us to avoid adapting the dimensions of the involved vectors at each iteration.
On the contrary, this is a convenient procedure to be followed when aiming at a numerical implementation of the envisaged scheme. Notice that to discuss the algorithmic variant of the method, we made use of the same symbols employed earlier. The notation clash is however solely confined to this paragraph.}
{In the following, we will discuss how these ideas extend to the more general setting of non-linear multi-layered neural networks.}
{\subsection{Training non-linear multi-layered neural networks in the spectral domain}}
{ In analogy with the above, the image to be processes is again organized in a $N\times1$ column vector $\vec{n}_1$. This latter is transformed into $\vec{n}_2=\mathbf{A}_1 \vec{n}_1$, where matrix
${N \times N}$ matrix $\mathbf{A}_1$ is recovered from its spectral properties, respectively encoded in $\Phi_1$ and $\Lambda_1$. The output vector $\vec{n}_2$ is now filtered via a suitable {\it non-linear} function $f(\cdot)$. This step marks a distinction between, respectively, the linear and non-linear versions of the learning schemes. For the applications here reported we have chosen to work with a rectified linear unit (ReLU) $f(\cdot)= max(0, \cdot)$. Another possibility is to set $f(\cdot, \beta_1)=\tanh[\beta_1(\cdot)]$, where $\beta_1$ is a control parameter which could be in principle self-consistently adjusted all along the learning procedure. We are now in a position to iterate the same reasoning carried out in the preceding section, adapted to the case at hand. More specifically, we introduce the generic $N\times N$ matrix $\mathbf{A}_k= \Phi_k \Lambda_k \left(2 \mathbb{1}_{N \times N}-\Phi_k\right)$ which transforms $\vec{n}_k$ into $\vec{n}_{k+1}$, with $k=2,...,\ell-1$. The outcome of this linear transformation goes through the non-linear filter. The loss function $L(\vec{n})$ generalizes to:}
\begin{equation}
\label{LF}
L(\vec{n}) = \left\lVert l(\vec{n}_1) - \sigma\left( f\left(\mathbf{A}_{\ell-1}.... f\left (\mathbf{A}_2 f \left (\mathbf{A}_1 \vec{n_1},\beta_1 \right), \beta_2 \right),\beta_{\ell-1} \right) \right) \right\rVert^2
\end{equation}
{with an obvious meaning of the involved symbols. In the set of experiments reported below we assume, in analogy with the above, a categorical cross-entropy loss function. The loss function is minimized upon adjusting the free parameters of the learning scheme: the $\ell-1$ blocks of tunable eigenvalues, the elements that define the successive indentation of the nested basis which commands the transfer of the information (and e.g. the quantities $\beta_k$, if the sigmoidal hyperbolic function is chosen as a non-linear filter). This eventually yields a fully trained network, in direct space, which can be unfolded into a layered architecture to perform pattern recognition (see Fig. \ref{fig3}). Remarkably, self-loop links are also present. The limit of a linear single layer perceptron is recovered when silencing the non-linearities: a $(N_1 + N_\ell) \times (N_1 + N_\ell)$ matrix $\mathcal{A}_c$ can be generated from the $N \times N$ matrices $\mathbf{A}_{k}$, following the same strategy outlined above. A sequence of linear layers can be also interposed between two consecutive non-linear stacks. The interposed layers allow to enlarge the space of parameters employed in the learning scheme, and can be retracted when operating the deep neural network after completion of the learning stage. Their role is {\it de facto} encapsulated in the entries of the linear operator that bridges the gap between the adjacent non-linear stacks, as explained above when referring to the telescopic operational modality.}
\begin{figure}
\centering
\includegraphics[scale=0.5]{EmbeddingProgressivo.pdf} \\
\caption{\it {The non-linear version of the training scheme returns a multi-layered architecture with self-loops links in direct space. Linear and non-linear transformation can be combined at will, matrices $\mathbf{A}_k$ providing the connection between successive layers. Linear layers can be retracted in operational mode, following a straightforward variant of the compactification procedure described in the main body of the paper. }}
\label{fig3}
\end{figure}
{\section{Results}}
To build and train the aforementioned models we used TensorFlow and created a custom spectral layer matrix that could be integrated in virtually every TensorFlow or Keras model. That allowed us to leverage on the automatic differentiation capabilities and the built-in optimizers of TensorFlow. Recall that we aim at training {just a a portion of the diagonal of $\Lambda_k$ and a block of $\Phi_k$}. To reach this goal we generated two fully trainable matrices, for each layer in the spectral domain, and applied a suitably designed mask to filter out the sub-parts of the matrices to be excluded from the training. This is easy to implement and, although improvable from the point of view of computational efficiency, it works perfectly, given the size of the problem to be handled. We then trained all our models with the AdaMax optimizer \cite{kingma2014adam} by using a learning rate of $0.03$ for the linear case and $0.01$ for the non-linear one. The training proceeded for {about $20$ epochs} and during each epoch the network was fed with batches of images of {different size, ranging from $300$ to $800$.}. These hyperparameters have been chosen so as to improve on GPU efficiency, accuracy and stability. However, we did not perform a systematic study to look for the optimal setting. All our models have been trained on a virtual machine hosted by Google Colaboratory. Standard neural networks have been trained on the same machine using identical software and hyperparameters, for a fair comparison. Further details about the implementation, as well as a notebook to reproduce our results, can be found in the public repository of this project \cite{gitrepo}.
{We shall start by reporting on the performance of the linear scheme. The simplest setting is that of a perceptron made of two layers: the input layer with $N_1=28 \times 28 = 784$ nodes and the output one made of $N_2=10$ elements. The perceptron can be trained in the spectral domain by e.g. tuning the $N=N_1+N_2=794$ eigenvalues of $\mathbf{A}_{1}$, the matrix that links the input ($\vec{n}_1$) and output ($\vec{n}_2$) vectors. The learning restricted to the eigenvalues returns a perceptron which performs the sought classification task with an accuracy (the fraction of correctly recognized images in the test-set) of $(82 \pm 2) \%$ (averaging over $5$ independent runs). This figure is to be confronted with the accuracy of a perceptron trained with standard techniques in direct space. For a fair comparison, the number of adjustable weights should be limited to $N$. To this aim, we randomly select a subset of weights to be trained and carry out the optimization on these latter. The process is repeated a few ($5$ in this case) times and, for each realization, the associated accuracy computed. Combining the results yields an average performance of $(79 \pm 3) \%$ , i.e. a slightly smaller score (although compatible within error precision) than that achieved when the learning takes place in the spectral domain. When the training extends to all the $N_1 \times N_2$ weights (plus $N_1+N_2$ bias), conventional learning yields a final accuracy of $(92.7 \pm 0.1) \%$. This is practically identical to the score obtained in the spectral domain, specifically $(92.5 \pm 0.2) \%$, when the sub-diagonal entries of the eigenvectors matrix are also optimized (for a total of $N_1+N_2+N_1 \times N_2$ free parameters). The remarkable observation is however that the distribution of the weights as obtained when the learning is restricted on the eigenvalues (i.e using about the 10 \% of the parameters employed for a full training in direct space) matches quite closely that retrieved by means of conventional learning schemes, see Fig. \ref{fig4} . This is not the case when the learning in direct space acts on a subset of $N$, randomly selected, weights (data not shown). Based on the above, it can be therefore surmised that optimizing the eigenvalues constitutes a rather effective pre-training strategy, which engages a modest computational load.}
\begin{figure}
\centering
\includegraphics[scale=0.7]{Percettrone.pdf} \\
\caption{\it {Distribution of the weights of a perceptron. The red line follows the spectral training limited the $N_1+N_2$ eigenvalues. The black line follows the training in direct space. Here, $N_1 \times N_2$ parameters are adjusted in the space of the nodes. The distribution are very similar, but the spectral learning employs about $10 \%$ of the parameters used in direct space. The distributions obtained
when forcing the training in direct space to operate on a subset of $N_1+N_2$ weights are very different from the one displayed (for every choice of the randomly selected family of weights to be trained).}}
\label{fig4}
\end{figure}
{To further elaborate on the potentiality of the proposed technique, we modify the simple two-layers perceptron, with the inclusion of supplementary computing layers. As explained above the newly added layers plays an active role during the learning stage, but can be retracted in operating mode so as to return a two-layers perceptron. The weights of this latter bear however an imprint of the training carried out for the linear network in the expanded configuration. Two alternative strategies will be in particular contemplated. On the one side, we will consider a sole additional layer, endowed with $N_2$ nodes, interposed between the input and output layers made of, respectively, $N_1=784$ and $N_{\ell} \equiv N_3=10$ nodes. We will refer to this as to the {\it wide linear} configuration. The performance of the method can be tested by letting $N_2$ to progressively grow. On the other side, the {\it deep linear} configuration is obtained when interposing a sequence of successive (linear) stacks between the input ($N_1=784$) and the output ($N_{\ell} =10$) layers.}
{ In Fig. \ref{fig5}, we report on the performance of the wide learning scheme as a function of $N_2+N_3$. As we shall clarify, this latter stands for the number of trained parameters for (i) the spectral learning acted on a subset of the tunable eigenvalues and for (ii) the conventional learning in direct space restricted to operate on a limited portion of the weights. The red line in the main panel of Fig. \ref{fig5} refers to the simplified scheme where a subset of the eigenvalues are solely tuned (while leaving the eigenvectors fixed at the random realization set by the initial condition). We have in particular chosen to train the second bunch of $N_2$ eigenvalues of the transfer matrix $\mathbf{A}_{1}$ and the $N_3=10$ non trivial eigenvalues of matrix $\mathbf{A}_{2}$, in line with the prescriptions reported in the preceding Section. The blue line reports on the accuracy of the neural network trained in direct space: the target of the optimization is a subset of cardinality $N_2+N_3$ of the $N_1 N_2+N_2 N_3$ weights which could be in principle adjusted in the space of the nodes. The performance of the spectral method proves clearly superior, as it can be readily appreciated by visual inspection of Fig. \ref{fig5}. The black line displays the accuracy of the linear neural network when the optimization acts on the full set of $N_1 N_2+N_2 N_3$ trainable parameters. No improvement is detectable when increasing the size of the intermediate layer: the displayed accuracy is substantially identical to that obtained for the basic perceptron trained with $N_1 N_2=7840$ parameters. The spectral learning allows to reach comparable performance already at $N_2=1000$ ($13 \%$ of the parameters used for the standard two layers perceptron with $N_1 \times N_2$ parameters, as discussed above).
In the inset of Fig. \ref{fig5}, the distribution of the entries of matrix $\mathcal{A}_c$, the equivalent perceptron, is depicted in red for the setting highlighted in the zoom. The black line refers to the two-layers equivalent of the neural network trained in direct space, employing the full set of trainable parameters (black dot enclosed in the top-left dashed rectangle drawn in the main panel of Fig. \ref{fig5}). The two distributions look remarkably close, despite the considerable reduction in terms of training parameters, as implemented in the spectral domain (for the case highlighted, $0.13 \%$ of the parameters employed under the standard training). Similarly to the above, the distribution obtained when forcing the training in direct space to act on a subset of $N_1+N_2$ weights are just a modest modulation of the initially assigned profile, owing to the {\it local} nature of the learning in the space of the nodes.}
{In Fig. \ref{fig6}, we report the results of the tests performed when operating under the deep linear configuration. Symbols are analogous to those employed in Fig. \ref{fig5}. In all inspected cases, the entry layer is made of $N_1=784$ elements and the output one has $N_{\ell}=10$ nodes. The first five points, from left to right, refer to a three layers (linear) neural network. Hence, $\ell=3$ and the size of the intermediate layer is progressively increased, $N_2=20,80,100,500,800$. The total number of trained eigenvalues is $N_2+N_3$, and gets therefore larger as the size of the intermediate layer grows.
The successive four points of the collections are obtained by setting $\ell=4$. Here, $N_2=800$ while $N_3$ is varied ($=100,200,400,600$). The training impacts on $N_2+N_3+N_4$ parameters. Finally the last point in each displayed curve is obtained by working with a five layers deep neural network, $\ell=5$. In particular $N_2=800$, $N_3=600$ and $N_4=500$, for a total of $N_2+N_3+N_4+N_5$ tunable parameters. Also in this case, the spectral algorithm performs better than conventional learning schemes constrained to operate with an identical number of free parameters. Similarly, the distribution of the weights of an equivalent perceptron trained in reciprocal space matches that obtained when operating in the space of the nodes and resting on a considerably larger number of training parameters. To sum up, eigenvalues are parameters of key importance for neural networks training, way more strategic than any other set of equivalent cardinality in the space of the nodes. As such, they allow for a global approach to the learning, with significant reflexes of fundamental and applied interest. In all cases here considered, the learning can extend to the eigenvectors: an optimized indentation of the eigen-directions contribute to enhance the overall performance of the trained device.}
\begin{figure}
\centering
\includegraphics[scale=0.5]{WideLinearBox.pdf} \\
\caption{\it {A three layers neural network is considered. The accuracy of the neural network is plotted as a function of the number of parameters that we chose to train with the spectral algorithm, $N_2+N_3$. The red line reports on the performance of the spectral training. The blue line refers to the neural network trained in direct space: the optimization runs on $N_2+N_3$ parameters, a subset of the total number of adjustable weights $N_1 N_2+N_2 N_3$. The black line stands for the accuracy of the linear neural network when training the full set of $N_1 N_2+N_2 N_3$ parameters. Notice that the
reported accuracy is comparable to that obtained for a standard two layers perceptron. Inset:
the distribution of the entries of the equivalent perceptrons are plotted. The red curve refer to the spectral learning restricted to operate on the eigenvalues; the black profile to the neural network
trained in direct space, employing the full set of adjustable parameters. In both cases, the weights refer to the two layers configuration obtained by retracting the intermediate linear layer employed during the learning stage. }}
\label{fig5}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.5]{DeepLinearBox.pdf} \\
\caption{\it { The performance of the spectral algorithm are tested for a multi-layered linear configuration. Symbols are chosen in analogy to Fig. \ref{fig5}. In all cases, the input layer is made of $N_1=784$ elements and the output layer has $N_{\ell}=10$ nodes. The first five points, from left to right in each of the curves depicted in the main panel, refer to a three layers (linear) neural network. The size of the intermediate layer is progressively increased, as $N_2=20,80,100,500,800$. The total number of trained eigenvalues is $N_2+N_3$. The subsequent four points are obtained by considering a four layers architecture. In particular, $N_2=800$ while $N_3$ takes values in the interval ($100,200,400,600$). The training acts on $N_2+N_3+N_4$ eigenvalues. The final point in each curve is obtained with a four layers deep neural network. Here, $N_2=800$, $N_3=600$ and $N_3=500$, for a total of $N_2+N_3+N_4+N_5$ tunable parameters in the spectral setting. Inset:
the distribution of the entries of the equivalent perceptrons are displayed, with the same color code adopted in Fig. \ref{fig5}. Also in this case, the weights refer to the two layers configuration obtained by retracting the intermediate linear layers employed in the learning stage.
}}
\label{fig6}
\end{figure}
{We now turn to considering a non-linear architecture. More specifically, we will assume a four layers network with, respectively, $N_1=784, N_2,N_3=120, N_4=10$. The non-linear ReLU filter acts on the third layer of the collection, while the second is a linear processing unit. As in the spirit of the wide network configuration evoked above, we set at testing the performance of the neural network for increasing $N_2$. For every choice of $N_2$, the linear layer can be retracted yielding a three-layered effective non-linear configurations. We recall however that training the network in the enlarged space where the linear unit is present leaves a non trivial imprint in the weights that set the strength of the links in direct space.}
{In Fig \ref{fig7}, we plot the computed accuracy as a function of $N_2$, the size of the linear layer. In analogy with the above analysis, the red curve refers to the training restricted to
$N_2+N_3+N_4$ eigenvalues; the blue profile is obtained when the deep neural network is trained in direct space by adjusting an identical number of inter-nodes weights. As for the case of a fully linear architecture, by adjusting the eigenvalues yields better classification performances. The black line shows the accuracy of the neural network when the full set of $N_1 N_2+N_2 N_3+N_3 N_4$ is optimized in direct space. The green line refer instead to the spectral learning when the eigenvalues and eigenvectors are trained simultaneously. The accuracies estimated for these two latter settings agree within statistical error, even if the spectral scheme seems more robust to overfitting (the black circles declines slightly when increasing $N_2$, while the collection of green points appears rather stable).}
\begin{figure}
\centering
\includegraphics[scale=0.5]{NonLinearAccuracy.pdf} \\
\caption{\it {The accuracy of the non-linear deep neural network is tested. We assume a four layers network with, respectively, $N_1=784, N_2,N_3=120, N_4=10$; $N_2$ is changed so as to enlarge the set of parameters to be trained. The red line refers to the spectral training, with $N_2+N_3+N_4$ adjusted eigenvalues. The blue line stands for a neural network trained in direct space,
the target of the optimization being a subset made of $N_2+N_3+N_4$ weights, randomly selected from the available pool of $N_1 N_2+N_2 N_3+N_3 N_4$ tunable parameters.
The black line reports the accuracy of the linear neural network when training the full set of $N_1 N_2+N_2 N_3+N_3 N_4$ weights. The green line refer to the spectral learning when eigenvalues and eigenvectors are simultaneously trained.}}
\label{fig7}
\end{figure}
\section{Conclusions}
Summing up, we have here proposed a novel approach to the training of deep neural networks which is bound to the spectral, hence reciprocal, domain. The {eigenvalues and eigenvectors} of the adjacency matrices that connects consecutive layers via directed feed-forward links are trained, instead of adjusting the weights that bridge each pair of nodes of the collection, as it is customarily done in the framework of conventional ML approaches.
{The first conclusion of our analysis is that optimizing the eigenvalues, when freezing the eigenvectors, yields performances which are superior to those attained with conventional methods {\it restricted} to a operate with an identical number of free parameters. It is therefore surmised that eigenvalues are key target parameters for neural networks training, in that they allow for a {\it global} handling of the learning. This is at variance with conventional approaches which seek at modulating the weights of the links among mutually connected nodes. Secondly, the spectral learning restricted to the eigenvalues yields a distribution of the weights which resembles quite closely that obtained with conventional algorithms bound to operate in direct space. For this reason, the proposed method could be used in combination with existing ML algorithms for an effective (and computationally advantageous) pre-training of deep neural networks. We have also shown that linear processing units inserted in between consecutive, non-linearly activated layers produce an enlargement of the learning parameters space, with beneficial effects in terms of performance of the trained device.
Extending the learning so as to optimize the eigenvectors enhances the ability of the network to operate the sought classification. In the proposed implementation, and to recover a feed-forward architecture in direct space, we have assumed a nested indentation of the eigenvectors. Entangling the eigenvectors referred to successive stacks is the key for a recursive processing of the data, from the input to the output layer. Employing other non-orthogonal basis could eventually allow to challenge different topologies in direct space and shed novel light on the surprising ability of deep networks to cope with the assigned tasks.}
{In future perspective, it would interesting to characterize the solutions attained with the spectral method, following the strategy outlined in \cite{feizi2017porcupine}.
Further, it could be interesting to combine the spectral approach to other existing schemes which have been devised to improve the computational performance of deep neural networks,
without significant loss in final recognition accuracy \cite{6638949, frankle2020training}.}
\bibliographystyle{unsrt}
| 56d318bdb92253ff4f13a9e08991fbaedbe40c4d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{V}{ideo} dominates the Internet. In North America, Netflix and YouTube alone account for more than fifty percent of downstream traffic, and there are many other significant video service providers. Improving the efficiency of video encoding, storage, and streaming over communication networks is a principle goal of video sharing and streaming platforms. One relevant and essential research direction is the perceptual optimization of rate-distortion tradeoffs in video encoding and streaming, where distortion (or quality) is usually modeled using video quality assessment (VQA) algorithms that can predict human judgements of video quality. This has motivated years of research on the topics of perceptual video and image quality assessment (VQA/IQA).
VQA research can be divided into two closely related categories: subjective video quality studies and objective video quality modeling. Subjective video quality research usually requires substantial resources devoted to time- and labor-consuming human studies to obtain valuable and reliable subjective data. The datasets obtained from subjective studies are invaluable for the development, calibration, and benchmarking of objective video quality models that are consistent with subjective mean opinion scores (MOS).
\begin{table*}[!t]
\setlength{\tabcolsep}{2.5pt}
\renewcommand{\arraystretch}{1.1}
\centering
\begin{threeparttable}
\caption{Evolution of popular public video quality assessment databases: from legacy lab studies of synthetically distorted video sets to large-scale crowdsourced user-generated content (UGC) video datasets with authentic distortions}
\label{table:db_comp}
\begin{tabular}{llllllllp{3.5cm}lllll}
\toprule
\textsc{Database} & \textsc{Year} & \textsc{\#Cont} & \textsc{\#Total} & \textsc{Resolution} & \textsc{FR} & \textsc{Len} & \textsc{Format} & \textsc{Distortion Type} & \textsc{\#Subj} & \textsc{\#Rates} & \textsc{Data} & \textsc{Env} \\
\hline\\[-1.em]
LIVE-VQA & 2008 & 10 & 160 & 768$\times$432 & 25/50 & 10 & YUV+264 & Compression, transmission & 38 & 29 & DMOS+$\sigma$ & In-lab \\
EPFL-PoliMI & 2009 & 12 & 156 & CIF/4CIF & 25/30 & 10 & YUV+264 & Compression, transmission & 40 & 34 & MOS & In-lab \\
VQEG-HDTV & 2010 & 49 & 740 & 1080i/p & 25/30 & 10 & AVI & Compression, transmission & 120 & 24 & RAW & In-lab \\
IVP & 2011 & 10 & 138 & 1080p & 25 & 10 & YUV & Compression, transmission & 42 & 35 & DMOS+$\sigma$ & In-lab \\
TUM 1080p50 & 2012 & 5 & 25 & 1080p & 50 & 10 & YUV & Compression & 21 & 21 & MOS & In-lab \\
CSIQ & 2014 & 12 & 228 & 832$\times$480 & 24-60 & 10 & YUV & Compression, transmission & 35 & N/A & DMOS+$\sigma$ & In-lab \\
CVD2014 & 2014 & 5 & 234 & 720p, 480p & 9-30 & 10-25 & AVI & Camera capture (authentic) & 210 & 30 & MOS & In-lab \\
MCL-V & 2015 & 12 & 108 & 1080p & 24-30 & 6 & YUV & Compression, scaling & 45 & 32 & MOS & In-lab \\
MCL-JCV & 2016 & 30 & 1560 & 1080p & 24-30 & 5 & MP4 & Compression & 150 & 50 & RAW-JND & In-lab \\
KoNViD-1k & 2017 & 1200 & 1200 & 540p & 24-30 & 8 & MP4 & Diverse distortions (authentic) & 642 & 114 & MOS+$\sigma$ & Crowd \\
LIVE-Qualcomm & 2018 & 54 & 208 & 1080p & 30 & 15 & YUV & Camera capture (authentic) & 39 & 39 & MOS & In-lab \\
LIVE-VQC & 2018 & 585 & 585 & 1080p-240p & 19-30 & 10 & MP4 & Diverse distortions (authentic) & 4776 & 240 & MOS & Crowd \\
YouTube-UGC & 2019 & 1380 & 1380 & 4k-360p & 15-60 & 20 & MKV & Diverse distortions (authentic) & $>$8k & 123 & MOS+$\sigma$ & Crowd \\
\bottomrule
\end{tabular}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item \textsc{\#Cont}: Total number of unique contents.
\item \textsc{\#Total}: Total number of test sequences, including reference and distorted videos.
\item \textsc{Resolution}: Video resolution (p: progressive).
\item \textsc{FR}: Framerate.
\item \textsc{Len}: Video duration/length (in seconds).
\item \textsc{Format}: Video container.
\item \textsc{\#Subj}: Total number of subjects in the study.
\item \textsc{\#Rates}: Average number of subjective ratings per video.
\item \textsc{Env}: Subjective testing environment. In-lab: study was conducted in a laboratory. Crowd: study was conducted by crowdsourcing.
\end{tablenotes}
\end{threeparttable}
\end{table*}
Hence, researchers have devoted considerable efforts on the development of high-quality VQA datasets that benefit the video quality community. Table \ref{table:db_comp} summarizes the ten-year evolution of popular public VQA databases. The first successful VQA database was the LIVE Video Quality Database \cite{seshadrinathan2010study}, which was first made publicly available in 2008. It contains $10$ pristine high-quality videos subjected to compression and transmission distortions. Other similar databases targeting simulated compression and transmission distortions have been released subsequently, including EPFL-PoliMI \cite{de2010h}, VQEG-HDTV \cite{vqeg_hdtv}, IVP \cite{ivp}, TUM 1080p50 \cite{keimel2012tum}, CSIQ \cite{vu2014vis3}, MCL-V \cite{lin2015mcl}, and MCL-JCV \cite{wang2016mcl}. All of the above mentioned datasets are based on a small set of high-quality videos, dubbed ``pristine'' or ``reference,'' then synthetically distorting them in a controlled manner. We will refer to these kinds of synthetically-distorted video sets as \textit{legacy} VQA databases. Legacy databases are generally characterized by only a small number of unique contents, each simultaneously degraded by only one or at most two synthetic distortions. For most practical scenarios, these are too simple to represent the great variety of real-world videos, and hence, VQA models derived on these databases may be insufficiently generalizable to large-scale realistic commercial VQA applications.
Recently, there has been tremendous growth in social media, where huge volumes of user-generated content (UGC) is shared over the media platforms such as YouTube, Facebook, and TikTok. Advances in powerful and affordable mobile devices and cloud computing techniques, combined with significant advances in video streaming have made it easy for most consumers to create, share, and view UGC pictures/videos instantaneously across the globe. Indeed, the prevalence of UGC has started to shift the focus of video quality research from \textit{legacy} synthetically-distorted databases to newer, larger-scale authentic UGC datasets, which are being used to create solutions to what we call the \textbf{\textsf{UGC-VQA problem}}. UGC-VQA studies typically follow a new design paradigm whereby: 1) All the source content is consumer-generated instead of professional-grade, thus suffers from unknown and highly diverse impairments; 2) they are only suitable for testing and comparing no-reference models, since reference videos are unavailable; 3) the types of distortions are authentic and commonly intermixed, and include but are not limited to capture impairments, editing and processing artifacts, compression, transcoding, and transmission distortions. Moreover, compression artifacts are not necessarily the dominant factors affecting video quality, unlike legacy VQA datasets and algorithms. These unpredictable perceptual degradations make perceptual quality prediction of UGC consumer videos very challenging.
Here we seek to address and gain insights into this new challenge (UGC-VQA) by first, conducting a comprehensive benchmarking study of leading video quality models on several recently released large-scale UGC-VQA databases. We also propose a new fusion-based blind VQA (BVQA) algorithm, which we call the VIDeo quality EVALuator (VIDEVAL), which is created by the processes of feature selection from existing top-performing VQA models. The empirical results show that a simple aggregation of these known models can achieve state-of-the-art (SOTA) performance. We believe that our expansive study will inspire and drive future research on BVQA modeling for the challenging UGC-VQA problem, and also pave the way towards deep learning-based solutions.
The outline of this paper is as follows: Section \ref{sec:ugc_db} reviews and analyzes the three most recent large-scale UGC-VQA databases, while Section \ref{sec:nr_vqa} briefly surveys the development of BVQA models. We introduce the proposed VIDEVAL model in Section \ref{sec:fs}, and provide experimental results in Section \ref{sec:exp}. Finally, concluding remarks are given in Section \ref{sec:conclud}.
\begin{figure*}[!t]
\def0.6{0.197}
\centering
\subfloat[LIVE-VQC][{LIVE-VQC}]{\includegraphics[height=0.6\textwidth]{figs/LIVE-VQC-thumbs.pdf}
\label{fig:snapshota}}
\hspace{-0.284em}
\subfloat[KoNViD-1k][{KoNViD-1k}]{\includegraphics[height=0.6\textwidth]{figs/KONVID-1K-thumbs.pdf}
\label{fig:snapshotb}}
\hspace{-0.284em}
\subfloat[YouTube-UGC][{YouTube-UGC}]{\includegraphics[height=0.6\textwidth]{figs/YOUTUBE-UGC-thumbs.pdf}
\label{fig:snapshotc}}
\caption{Sample frames of the video contents contained in the three large scale UGC-VQA databases: (a) LIVE-VQC \cite{sinno2018large}, (b) KoNViD-1k \cite{hosu2017konstanz}, and (c) YouTube-UGC \cite{wang2019youtube}. LIVE-VQC includes only natural contents captured by mobile devices, while KoNViD-1k and YouTube-UGC comprise of both natural videos, animations, and gaming sources. Note that YouTube-UGC video set is categorized whereas the others are not.}
\label{fig:snapshot}
\end{figure*}
\begin{table*}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Public large-scale user-generated content video quality assessment (UGC-VQA) databases compared: KoNViD-1k \cite{hosu2017konstanz}, LIVE-VQC \cite{sinno2018large}, and YouTube-UGC \cite{wang2019youtube}}
\label{table:ugc_db_comp}
\begin{tabular}{lccc}
\toprule
\textsc{Database Attribute} & KoNViD-1k & LIVE-VQC & YouTube-UGC \\
\hline\\[-1.em]
Number of contents & 1200 & 585 & 1380 \\
Video sources & YFCC100m (Flickr) & Captured (mobile devices) & YouTube \\
Video resolutions & 540p & 1080p,720p,480p,etc. & 4k,1080p,720p,480p,360p \\
Video layouts & Landscape & Landscape,portrait & Landscape,portrait \\
Video framerates & 24,25,30 fr/sec & 20,24,25,30 fr/sec & 15,20,24,25,30,50,60 fr/sec \\
Video lengths & 8 seconds & 10 seconds & 20 seconds \\
Audio track included & Yes (97\%) & Yes & No \\
Testing methodology & Crowdsourcing (CrowdFlower) & Crowdsourcing (AMT) & Crowdsourcing (AMT) \\
Number of subjects & 642 & 4,776 & $>$8,000 \\
Number of ratings & 136,800 (114 votes/video) & 205,000 (240 votes/video) & 170,159 (123 votes/video) \\
Rating scale & Absolute Category Rating 1-5 & Continuous Rating 0-100 & Continuous Rating 1-5 \\
\multirow[t]{6}{*}{Content remarks} & \multirow[t]{6}{4.6cm}{Videos sampled from YFCC100m via a feature space of blur, colorfulness, contrast, SI, TI, and NIQE; Some contents irrelevant to quality research; Content was clipped from the original and resized to 540p.} & \multirow[t]{6}{4.6cm}{Videos manually captured by certain people; Content including many camera motions; Content including some night scenes that are prone to be outliers; Resolutions not uniformly distributed.} & \multirow[t]{6}{4.6cm}{Videos sampled from YouTube via a feature space of spatial, color, temporal, and chunk variation; Contents categorized into $15$ classes, including HDR, screen content, animations, and gaming videos.} \\ \\ \\ \\ \\ \\
\multirow[t]{5}{*}{Study remarks} & \multirow[t]{5}{4.6cm}{Study did not account for or remove videos on which stalling events occurred when viewed; test methodology prone to unreliable individual scores.} & \multirow[t]{5}{4.6cm}{Distribution of MOS values slightly skewed towards higher scores; standard deviation statistics of MOS were not provided.} & \multirow[t]{5}{4.6cm}{Distribution of MOS values slightly skewed towards higher values; three additional chunk MOS scores with standard deviation were provided.} \\ \\ \\ \\ \\
\bottomrule
\end{tabular}
\end{table*}
\section{UGC-VQA Databases}
\label{sec:ugc_db}
The first UGC-relevant VQA dataset containing authentic distortions was introduced as the Camera Video Database (CVD2014) \cite{nuutinen2016cvd2014}, which consists of videos with in-the-wild distortions from 78 different video capture devices, followed by the similar LIVE-Qualcomm Mobile In-Capture Database \cite{ghadiyaram2017capture}. These two databases, however, only modeled (camera) capture distortions on small numbers of not very diverse unique contents. Inspired by the first successful massive online crowdsourcing study of UGC picture quality \cite{ghadiyaram2015massive}, the authors of \cite{hosu2017konstanz} created the KoNViD-1k video quality database, the first such resource for UGC videos. It consists of 1,200 public-domain videos sampled from the YFCC100M dataset \cite{thomee2015yfcc100m}, and was annotated by 642 crowd-workers. LIVE-VQC \cite{sinno2018large} was another large-scale UGC-VQA database with 585 videos, crowdsourced on Amazon Mechanical Turk to collect human opinions from 4,776 unique participants. The most recently published UGC-VQA database is the YouTube-UGC Dataset \cite{wang2019youtube} comprising 1,380 20-second video clips sampled from millions of YouTube videos, which were rated by more than 8,000 human subjects. Table \ref{table:ugc_db_comp} summarizes the main characteristics of the three large-scale UGC-VQA datasets studied, while Figure \ref{fig:snapshot} shows some representative snapshots of the source sequences for each database, respectively.
\subsection{Content Diversity and MOS Distribution}
As a way of characterizing the content diversity of the videos in each database, Winkler \cite{winkler2012analysis} suggested three quantitative attributes related to spatial activity, temporal activity, and colorfulness. Here we expand the set of attributes to include six low-level features including brightness, contrast, colorfulness \cite{hasler2003measuring}, sharpness, spatial information (SI), and temporal information (TI), thereby providing a larger visual space in which to plot and analyze content diversities of the three UGC-VQA databases. To reasonably limit the computational cost, each of these features was calculated on every 10th frame, then was averaged over frames to obtain an overall feature representation of each content. For simplicity, we denote the features as $\{\mathrm{C}_i\},i=1,2,...,6$. Figure \ref{fig:ind_feat_dis} shows the fitted kernel distribution of each selected feature. We also plotted the convex hulls of paired features, to show the feature coverage of each database, in Figure \ref{fig:feat_cvx_hull}. To quantify the coverage and uniformity of these databases over each defined feature space, we computed the relative range and uniformity of coverage \cite{winkler2012analysis}, where the relative range is given by:
\begin{equation}
\label{eq:relative_range}
\mathrm{R}_i^k=\frac{\max(\mathrm{C}_i^k)-\min(\mathrm{C}_i^k)}{\max_{k} (\mathrm{C}_i^k)},
\end{equation}
where $\mathrm{C}_i^k$ denotes the feature distribution of database $k$ for a given feature dimension $i$, and $\max_{k}(\mathrm{C}_i^k)$ specifies the maximum value for that given dimension across all databases.
Uniformity of coverage measures how uniformly distributed the videos are in each feature dimension. We computed this as the entropy of the $\mathrm{B}$-bin histogram of $\mathrm{C}_i^k$ over all sources for each database indexed $k$:
\begin{equation}
\label{eq:uniformity}
\mathrm{U}^k_i=-\sum_{b=1}^\mathrm{B} p_b \log_\mathrm{B} p_b,
\end{equation}
where $p_b$ is the normalized number of sources in bin $b$ at feature $i$ for database $k$. The higher the uniformity the more uniform the database is. Relative range and uniformity of coverage are plotted in Figure \ref{fig:relative_range} and Figure \ref{fig:uniformity}, respectively, quantifying the intra- and inter-database differences in source content characteristics.
We also extracted 4,096-dimensional VGG-19 \cite{simonyan2014very} deep features and embedded these features into 2D subspace using t-SNE \cite{maaten2008visualizing} to further compare content diversity, as shown in Figure \ref{fig:tsne}. Apart from content diversity expressed in terms of visual features, the statistics of the subjective ratings are another important attribute of each video quality database. The main aspect considered in the analysis here is the distributions of mean opinion scores (MOS), as these are indicative of the quality range of the subjective judgements. The analysis of standard deviation of MOS is not presented here since it is not provided in LIVE-VQC. Figure \ref{fig:mos_dist} displays the histogram of MOS distributions for the three UGC-VQA databases.
\begin{figure*}[!t]
\captionsetup[subfigure]{justification=centering}
\centering
\def0.32{0.133}
\def0ex{-0.4em}
\subfloat[Brightness][{Brightness}]{\includegraphics[height=0.32\textwidth]{figs/brightness_mu_dist.pdf}
\label{fig:ind_feat_disa}}
\hspace{0ex}
\subfloat[Contrast][{Contrast}]{\includegraphics[height=0.32\textwidth]{figs/contrast_mu_dist.pdf}
\label{fig:ind_feat_disb}}
\hspace{0ex}
\subfloat[Colorfulness][{Colorfulness}]{\includegraphics[height=0.32\textwidth]{figs/colorfulness_mu_dist.pdf}
\label{fig:ind_feat_disc}}
\hspace{0ex}
\subfloat[Sharpness][{Sharpness}]{\includegraphics[height=0.32\textwidth]{figs/sharpness_mu_dist.pdf}
\label{fig:ind_feat_disd}}
\hspace{0ex}
\subfloat[SI][{SI}]{\includegraphics[height=0.32\textwidth]{figs/si_mu_dist.pdf}
\label{fig:ind_feat_dise}}
\hspace{0ex}
\subfloat[TI][{TI}]{\includegraphics[height=0.32\textwidth]{figs/ti_std_mu_dist.pdf}
\label{fig:ind_feat_disf}}
\caption{Feature distribution comparisons among the three considered UGC-VQA databases: KoNViD-1k, LIVE-VQC, and YouTube-UGC.}
\label{fig:ind_feat_dis}
\end{figure*}
\begin{figure}[!t]
\centering
\def0.32{0.12}
\def0ex{-0.em}
\def0.255{0.255}
\def1pt{1pt}
\def3pt{3pt}
\footnotesize
\setlength{\tabcolsep}{1.5pt}
\renewcommand{\arraystretch}{1.0}
\begin{tabular}{ccc}
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/KONVID_1K_BRxCT.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/KONVID_1K_CFxSR.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/KONVID_1K_SIxTI.pdf} \\[1pt]
\multicolumn{3}{c}{(a) {KoNViD-1k} } \\[3pt]
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/LIVE_VQC_BRxCT.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/LIVE_VQC_CFxSR.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/LIVE_VQC_SIxTI.pdf} \\[1pt]
\multicolumn{3}{c}{(b) LIVE-VQC} \\[3pt]
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/YOUTUBE_UGC_BRxCT.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/YOUTUBE_UGC_CFxSR.pdf} &
\includegraphics[ height=0.255\linewidth, keepaspectratio]{figs/YOUTUBE_UGC_SIxTI.pdf} \\[1pt]
\multicolumn{3}{c}{(c) YouTube-UGC} \\
\end{tabular}
\caption{Source content (blue `x') distribution in paired feature space with corresponding convex hulls (orange boundaries). Left column: BR$\times$CT, middle column: CF$\times$SR, right column: SI$\times$TI.}
\label{fig:feat_cvx_hull}
\end{figure}
\subsection{Observations}
\label{ssec:observation}
We make some observations from the above plots. As may be seen in Figures \ref{fig:ind_feat_disa} and \ref{fig:ind_feat_disb}, and the corresponding convex hulls in Figure \ref{fig:feat_cvx_hull}, KoNViD-1k and YouTube-UGC exhibit similar coverage in terms of brightness and contrast, while LIVE-VQC adheres closer to middle values. Regarding colorfulness, KoNViD-1k shows a skew towards higher scores than the other two datasets, which is consistent with the observations that Flickr users self-characterize as either professional video/photographers or as dedicated amateurs. On the sharpness and SI histograms, YouTube-UGC is spread most widely, while KoNViD-1k is concentrated on lower values. Another interesting finding from the TI statistics: LIVE-VQC is distributed more towards higher values than YouTube-UGC and KoNViD-1k, consistent with our observation that videos in LIVE-VQC were captured in the presence of larger and more frequent camera motions. We will revisit this interesting aspect of TI when evaluating the BVQA models in Section \ref{sec:exp}. The visual comparison in Figure \ref{fig:tsne} shows that YouTube-UGC and KoNViD-1k span a wider range of VGG-19 feature space than does LIVE-VQC, indicating significant content diversity differences. Figure \ref{fig:mos_dist} shows the MOS distributions: all three databases have right-skewed MOS distributions, with KoNViD-1k less so, and LIVE-VQC and YouTube-UGC more so. The overall ranges and uniformity comparisons in Figures \ref{fig:relative_range}, \ref{fig:uniformity}, and \ref{fig:tsne} suggest that constructing a database by crawling and sampling from a large content repository is likely to yield a more content-diverse, uniformly-distributed dataset than one created from pictures or videos captured directly from a set of user cameras. Both cases may be argued to be realistic in some scenario.
\begin{figure}[!t]
\def0.6{0.823}
\centering
\includegraphics[width=0.6\linewidth]{figs/relative_range.pdf}
\caption{Relative range $\mathrm{R}^k_i$ comparisons of the selected six features calculated on the three UGC-VQA databases: KoNViD-1k, LIVE-VQC, and YouTube-UGC.}
\label{fig:relative_range}
\end{figure}
\begin{figure}[!t]
\def0.6{0.823}
\centering
\includegraphics[width=0.6\linewidth]{figs/coverage_uniformity.pdf}
\caption{Comparison of coverage uniformity $\mathrm{U}^k_i$ of the selected six features computed on the three UGC-VQA databases: KoNViD-1k, LIVE-VQC, and YouTube-UGC.}
\label{fig:uniformity}
\end{figure}
\begin{figure*}[!t]
\centering
\def0.32{0.16}
\def0ex{10pt}
\subfloat[KoNViD-1k][{KoNViD-1k}]{\includegraphics[height=0.32\textwidth]{figs/KONVID_1K_MOS_hist.pdf}
\label{fig:mos_dist-a}}
\hspace{0ex}
\subfloat[LIVE-VQC][{LIVE-VQC}]{\includegraphics[height=0.32\textwidth]{figs/LIVE_VQC_MOS_hist.pdf}
\label{fig:mos_dist-b}}
\hspace{0ex}
\subfloat[YouTube-UGC][{{YouTube-UGC}}]{\includegraphics[height=0.32\textwidth]{figs/YOUTUBE_UGC_MOS_hist.pdf}
\label{fig:mos_dist-c}}
\caption{MOS histograms and the fitted kernel distributions of the three UGC-VQA databases: KoNViD-1k, LIVE-VQC, and YouTube-UGC.}
\label{fig:mos_dist}
\end{figure*}
\begin{figure}[!t]
\def0.6{0.6}
\centering
\includegraphics[width=0.6\linewidth]{figs/tSNE_vgg19.pdf}
\caption{VGG-19 deep feature embedding via t-SNE \cite{maaten2008visualizing} on KoNViD-1k, LIVE-VQC, and YouTube-UGC, respectively.}
\label{fig:tsne}
\end{figure}
\section{UGC-VQA Models}
\label{sec:nr_vqa}
The goal of subjective video quality studies is to motivate the development of automatic objective video quality models. Conventionally, objective video quality assessment can be classified into three main categories: full-reference (FR), reduced-reference (RR), and no-reference (NR) models. FR-VQA models require the availability of an entire pristine source video to measure visual differences between a target signal and a corresponding reference \cite{wang2004video, vmaf,sheikh2005information, chen2020perceptual}, while RR-VQA models only make use of a limited amount of reference information \cite{wang2005reduced,soundararajan2012video}. Some popular FR-VQA models, including PSNR, SSIM \cite{wang2004image}, and VMAF \cite{vmaf} have already been successfully and massively deployed to optimize streaming and shared/uploaded video encoding protocols by leading video service providers. NR-VQA or BVQA models, however, rely solely on analyzing the test stimuli without the benefit of any corresponding ``ground truth'' pristine signal. It is obvious that only BVQA models are appropriate for the UGC-VQA problem. Here we briefly review the evolution of BVQA models, from conventional handcrafted feature-based approaches, on to convolutional neural network-based models.
\subsection{Conventional Feature-Based BVQA Models}
Almost all of the earliest BVQA models have been `distortion specific,' meaning they were designed to quantify a specific type of distortion such as blockiness \cite{wang2000blind}, blur \cite{marziliano2002no}, ringing \cite{ feng2006measurement}, banding \cite{ wang2016perceptual, tu2020bband, tu2020adaptive}, or noise \cite{amer2005fast, norkin2018film} in distorted videos, or to assess multiple specific coincident distortion types caused by compression or transmission impairments \cite{caviedes2017no, keimel2009no}. More recent top-performing BVQA models are almost exclusively learning-based, leveraging a set of generic quality-aware features, combined to conduct quality prediction by machine learning regression \cite{ moorthy2011blind, mittal2012no, saad2014blind, kundu2017no, ghadiyaram2017perceptual, korhonen2019two, ye2012unsupervised, pei2015image, tu2021rapique}. Learning-based BVQA models are more versatile and generalizable than `distortion specific' models, in that the selected features are broadly perceptually relevant, while powerful regression models can adaptively map the features onto quality scores learned from the data in the context of a specific application.
The most popular BVQA algorithms deploy perceptually relevant, low-level features based on simple, yet highly regular parametric bandpass models of good-quality scene statistics \cite{ruderman1994statistics}. These natural scene statistics (NSS) models predictably deviate in the presence of distortions, thereby characterizing perceived quality degradations \cite{sheikh2006image}. Successful blind picture quality assessment (BIQA) models of this type have been developed in the wavelet (BIQI \cite{moorthy2010two}, DIIVINE \cite{moorthy2011blind}, C-DIIVINE \cite{zhang2014c}), discrete cosine transform (BLIINDS \cite{saad2010dct}, BLIINDS-II \cite{saad2012blind}), curvelet \cite{liu2014no}, and spatial intensity domains (NIQE \cite{mittal2012making}, BRISQUE \cite{mittal2012no}), and have further been extended to video signals using natural bandpass space-time video statistics models \cite{li2016spatiotemporal,mittal2015completely,saad2014blind, sinno2019spatio}, among which the most well-known model is the Video-BLIINDS \cite{saad2014blind}. Other extensions to empirical NSS include the joint statistics of the gradient magnitude and Laplacian of Gaussian responses in the spatial domain (GM-LOG \cite{xue2014blind}), in log-derivative and log-Gabor spaces (DESIQUE \cite{zhang2013no}), as well as in the gradient domain of LAB color transforms (HIGRADE \cite{kundu2017no}). The FRIQUEE model \cite{ghadiyaram2017perceptual} has been observed to achieve SOTA performance both on UGC/consumer video/picture databases like LIVE-Challenge \cite{ghadiyaram2015massive}, CVD2014 \cite{nuutinen2016cvd2014}, and KoNViD-1k \cite{hosu2017konstanz} by leveraging a bag of NSS features drawn from diverse color spaces and perceptually motivated transform domains.
Instead of using NSS-inspired feature descriptors, methods like CORNIA \cite{ye2012unsupervised} employ unsupervised learning techniques to learn a dictionary (or codebook) of distortions from raw image patches, and was further extended to Video CORNIA \cite{xu2014no} by applying an additional temporal hysteresis pooling \cite{seshadrinathan2011temporal} of learned frame-level quality scores. Similar to CORNIA, the authors of \cite{xu2016blind} proposed another codebook-based general-purpose BVQA method based on High Order Statistics Aggregation (HOSA), requiring only a small codebook, yet yielding promising performance.
A very recent handcrafted feature-based BVQA model is the ``two level'' video quality model (TLVQM) \cite{korhonen2019two}, wherein a two-level feature extraction mechanism is adopted to achieve efficient computation of a set of carefully-defined impairment/distortion-relevant features. Unlike NSS features, TLVQM selects a comprehensive feature set comprising of empirical motion statistics, specific artifacts, and aesthetics. TLVQM does require that a large set of parameters (around 30) be specified, which may affect performance on datasets or application scenarios it has not been exposed to. The model currently achieves SOTA performance on three UGC video quality databases, CVD2014 \cite{nuutinen2016cvd2014}, KoNViD-1k \cite{hosu2017konstanz}, and LIVE-Qualcomm \cite{ghadiyaram2017capture}, at a reasonably low complexity, as reported by the authors.
\subsection{Deep Convolutional Neural Network-Based BVQA Models}
Deep convolutional neural networks (CNNs or ConvNets) have been shown to deliver standout performance on a wide variety of low-level computer vision applications. Recently, the release of several ``large-scale'' (in the context of IQA/VQA research) subjective quality databases \cite{ghadiyaram2015massive, hosu2017konstanz} have sped the application of deep CNNs to perceptual quality modeling. For example, several deep learning picture-quality prediction methods were proposed in \cite{kang2014convolutional, kim2017deep, ying2019patches, chen2020proxiqa}. To conquer the limits of data scale, they either propose to conduct patch-wise training \cite{kang2014convolutional, bosse2016deep, kim2017deep} using global scores, or by pretraining deep nets on ImageNet \cite{deng2009imagenet}, then fine tuning. Several authors report SOTA performance on legacy synthetic distortion databases \cite{sheikh2006statistical,ponomarenko2013color} or on naturally distorted databases \cite{ghadiyaram2015massive, hosu2020koniq}.
Among the applications of deep CNNs to blind video quality prediction, Kim \cite{kim2018deep} proposed a deep video quality assessor (DeepVQA) to learn the spatio-temporal visual sensitivity maps via a deep ConvNet and a convolutional aggregation network. The V-MEON model \cite{liu2018end} used a multi-task CNN framework which jointly optimizes a 3D-CNN for feature extraction and a codec classifier using fully-connected layers to predict video quality. Zhang \cite{zhang2018blind} leveraged transfer learning to develop a general-purpose BVQA framework based on weakly supervised learning and a resampling strategy. In the VSFA model \cite{li2019quality}, the authors applied a pre-trained image classification CNN as a deep feature extractor and integrated the frame-wise deep features using a gated recurrent unit and a subjectively-inspired temporal pooling layer, and reported leading performance on several natural video databases \cite{nuutinen2016cvd2014, hosu2017konstanz, ghadiyaram2017capture}. These SOTA deep CNN-based BVQA models \cite{kim2018deep, liu2018end, zhang2018blind, li2019quality} produce accurate quality predictions on legacy (single synthetic distortion) video datasets \cite{seshadrinathan2010study, vu2014vis3}, but struggle on recent in-the-wild UGC databases \cite{nuutinen2016cvd2014, ghadiyaram2017capture,hosu2017konstanz}.
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Summary of the initial feature set and the finalized VIDEVAL subset after feature selection.}
\label{table:feat_sum}
\begin{threeparttable}
\begin{tabular}{llcc}
\toprule
\textsc{Feature Name} & \textsc{Feature Index} & \textsc{\#($\mathcal{F}_\mathrm{INIT}$)} & \textsc{\#($\mathrm{VIDEVAL}$)} \\ \hline\\[-1.em]
BRISQUE\textsubscript{avg} & $f_1-f_{36}$ & 36 & 3 \\
BRISQUE\textsubscript{std} & $f_{37}-f_{72}$ & 36 & 1 \\
GM-LOG\textsubscript{avg} & $f_{73}-f_{112}$ & 40 & 4 \\
GM-LOG\textsubscript{std} & $f_{113}-f_{152}$ & 40 & 5 \\
HIGRADE-GRAD\textsubscript{avg} & $f_{153}-f_{188}$ & 36 & 8 \\
HIGRADE-GRAD\textsubscript{std} & $f_{189}-f_{224}$ & 36 & 1 \\
FRIQUEE-LUMA\textsubscript{avg} & $f_{225}-f_{298}$ & 74 & 4 \\
FRIQUEE-LUMA\textsubscript{std} & $f_{299}-f_{372}$ & 74 & 8 \\
FRIQUEE-CHROMA\textsubscript{avg} & $f_{373}-f_{452}$ & 80 & 10 \\
FRIQUEE-CHROMA\textsubscript{std} & $f_{453}-f_{532}$ & 80 & 1 \\
FRIQUEE-LMS\textsubscript{avg} & $f_{533}-f_{606}$ & 74 & 1 \\
FRIQUEE-LMS\textsubscript{std} & $f_{607}-f_{680}$ & 74 & 0 \\
FRIQUEE-HS\textsubscript{avg} & $f_{681}-f_{684}$ & 4 & 0 \\
FRIQUEE-HS\textsubscript{std} & $f_{685}-f_{688}$ & 4 & 0 \\
TLVQM-LCF\textsubscript{avg} & $f_{689}-f_{710}$ & 22 & 5 \\
TLVQM-LCF\textsubscript{std} & $f_{711}-f_{733}$ & 23 & 3 \\
TLVQM-HCF & $f_{734}-f_{763}$ & 30 & 6 \\ \hline\\[-1.em]
$\mathcal{F}_\mathrm{ALL}$ & $f_1-f_{763}$ & 763 & 60 \\
\bottomrule
\end{tabular}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item $^\star$All the spatial features are calculated every two frames and aggregated into a single feature vector within 1-sec chunks. The overall feature vector for the whole video is then obtained by averaging all the chunk-wise feature vectors. Subscript \textit{avg} means within-chunk average pooling, whereas subscript \textit{std} means within-chunk standard deviation pooling.
\end{tablenotes}
\end{threeparttable}
\end{table}
\section{Feature Fused VIDeo Quality EVALuator (VIDEVAL)}
\label{sec:fs}
We have just presented a diverse set of BVQA models designed from a variety of perspectives, each either based on scene statistics, or motivated by visual impairment heuristics. As might be expected, and as we shall show later, the performances of these models differ, and also vary on different datasets. We assume that the features extracted from different models may represent statistics of the signal in different perceptual domains, and henceforce, a selected fusion of BVQA models may be expected to deliver better consistency against subjective assessment, and also to achieve more reliable performance across different databases and use cases. This inspired our new feature fused VIDeo quality EVALuator (VIDEVAL), as described next.
\begin{figure}[!t]
\def0.32{0.8}
\def1{1}
\centering
\includegraphics[width=0.32\linewidth]{figs/ALL_COMBINED_feat_sel_results_plot.pdf}
\caption{{Feature selection performance (PLCC) of three selected algorithms as a function of $k$ on the All-Combined\textsubscript{c} dataset. The shaded error bar denotes the standard deviation of PLCC over 10 iterations.}}
\label{fig:feat_sel}
\end{figure}
\begin{figure}[!t]
\def0.32{0.8}
\def1{1}
\centering
\includegraphics[width=0.32\linewidth]{figs/ALL_COMBINED_model_based_svr_feat_importance.pdf}
\caption{Visualization of the second step in feature selection: frequency of each feature being selected over 100 iterations of train-test splits using SVR importance selection method with fixed $k=60$.}
\label{fig:feat_import}
\end{figure}
We begin by constructing an initial feature set on top of existing high-performing, compute-efficient BVQA models and features, distilled through a feature selection program. The goal of feature selection is to choose an optimcal or sub-optimal feature subset $\mathcal{F}_{k}\in \mathbb{R}^k$ from the initial feature set $\mathcal{F}_{\mathrm{INIT}}\in \mathbb{R}^\mathrm{N}$ (where $k<{N}$) that achieves nearly top performance but with many fewer features.
\subsection{Feature Extraction}
\label{ssec:feat_extract}
We construct an initial feature set by selecting features from existing top-performing BVQA models. For practical reasons, we ignore features with high computational cost, e.g., certain features from DIIVINE, BLIINDS, C-DIIVINE, and V-BLIINDS. We also avoid using duplicate features in different models, such as the BRISQUE-like features in HIGRADE, and the C-DIIVINE features in V-BLIINDS. This filtering process yields the initial feature candidates, which we denote as BRISQUE, GM-LOG, HIGRADE-GRAD, FRIQUEE-LUMA, FRIQUEE-CHROMA, FRIQUEE-LMS, FRIQUEE-HS, TLVQM-LCF, and TLVQM-HCF.
Inspired by the efficacy of standard deviation pooling as first introduced in GMSD \cite{xue2013gradient} and later also used in TLVQM \cite{korhonen2019two}, we calculate these spatial features every second frame within each sequentially cut non-overlapping one-second chunk, then we enrich the feature set by applying average and standard deviation pooling of frame-level features within each chunk, based on the hypothesis that the variation of spatial NSS features also correlates with the temporal properties of the video. Finally, all the chunk-wise feature vectors are average pooled \cite{tu2020comparative} across all the chunks to derive the final set of features for the entire video. Table \ref{table:feat_sum} indexes and summarizes the selected features in the initial feature set, yielding an overall 763-dimensional feature vector, $\mathcal{F}_\mathrm{INIT}\in\mathbb{R}^{763}$.
\begin{figure}[!t]
\centering
\def0.32{0.4}
\def0ex{0ex}
\subfloat[][{}]{\includegraphics[height=0.32\linewidth]{figs/inlsa_before.pdf}
\label{fig:inlsa-a}}
\hspace{0ex}
\subfloat[][{}]{\includegraphics[height=0.32\linewidth]{figs/inlsa_after.pdf}
\label{fig:inlsa-b}}
\caption{Scatter plots of MOS versus NIQE scores (a) before, and (b) after INLSA calibration \cite{pinson2003objective} using YouTube-UGC as the reference set.}
\label{fig:inlsa}
\end{figure}
\subsection{Feature Selection}
\label{ssec:feat_select}
We deploy two types of feature selection algorithms to distill the initial feature set. The first method is a model-based feature selector that utilizes a machine learning model to suggest features that are important. We employed the popular random forest (RF) to fit a regression model and eliminate the least significant features sorted by permutation importance. We also trained a support vector machine (SVM) with the linear kernel to rank the features, as a second model selector. Another sub-optimal solution is to apply a greedy search approach to find a good feature subset. Here we employed Sequential Forward Floating Selection (SFFS), and used SVM as the target regressor with its corresponding mean squared error between the predictions and MOS as the cost function. The mean squared error is calculated by cross-validation measures of predictive accuracy to avoid overfitting.
One problem with feature selection is that we do not know \textit{a priori} what $k$ to select, i.e., how many features are needed. Therefore, we conducted a two-step feature selection procedure. First, we evaluated the feature selection methods as a function of $k$ via 10 train-test iterations, to select the best algorithm with corresponding optimal $k$. Figure \ref{fig:feat_sel} shows the median PLCC (defined in Section \ref{ssec:eval_proto}) performance with respect to $k$ for different feature selection models, based on which we finally chose the SVM importance method with $k=60$ in our next experiments. In the second step, we applied the best feature selection algorithm with the fixed best $k$ over 100 random train-test splits. On each iteration, a subset is selected from the feature selector, based on which the frequency of each feature over the iterations is counted, and the $j$ most frequently occurring features are included into the final feature set. Figure \ref{fig:feat_import} shows the frequency of each feature being selected over 100 random splits in the second step. This selection process is implemented on a combined dataset constructed from three independent databases, as described in Section \ref{ssec:eval_proto}. Table \ref{table:feat_sum} summarizes the results of the feature selection procedure (SVR importance with $k=60$), yielding the final proposed VIDEVAL model.
\begin{figure}[!t]
\centering
\def0.32{0.98}
\includegraphics[width=0.32\linewidth]{figs/ALL_COMBINED_boxplot_new.pdf}
\caption{Box plots of PLCC, SRCC, and KRCC of evaluated learning-based BVQA algorithms on the All-Combined\textsubscript{c} dataset over 100 random splits. For each box, median is the central box, and the edges of the box represent 25th and 75th percentiles, while red circles denote outliers.}
\label{fig:all_boxplot}
\end{figure}
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Performance comparison of evaluated opinion-unaware ``completely blind'' BVQA models.}
\label{table:compete-blind}
\begin{tabular}{llcccc}
\toprule
\textsc{Dataset} & \textsc{Model} \textbackslash\ \textsc{Metric} & SRCC$\uparrow$ & KRCC$\uparrow$ & PLCC$\uparrow$ & RMSE$\downarrow$ \\
\hline\\[-1.em]
\multirow{3}{*}{KoNViD} & NIQE (1 fr/sec) & 0.5417 & 0.3790 & 0.5530 & 0.5336 \\
& ILNIQE (1 fr/sec) & 0.5264 & 0.3692 & 0.5400 & 0.5406 \\
& VIIDEO & 0.2988 & 0.2036 & 0.3002 & 0.6101 \\
\hline\\[-1.em]
\multirow{3}{*}{LIVE-C} & NIQE (1 fr/sec) & 0.5957 & 0.4252 & 0.6286 & 13.110 \\
& ILNIQE (1 fr/sec) & 0.5037 & 0.3555 & 0.5437 & 14.148 \\
& VIIDEO & 0.0332 & 0.0231 & 0.2146 & 16.654 \\
\hline\\[-1.em]
\multirow{3}{*}{YT-UGC}
& NIQE (1 fr/sec) & 0.2379 & 0.1600 & 0.2776 & 0.6174 \\
& ILNIQE (1 fr/sec) & 0.2918 & 0.1980 & 0.3302 & 0.6052 \\
& VIIDEO & 0.0580 & 0.0389 & 0.1534 & 0.6339 \\
\hline\\[-1.em]
\multirow{3}{*}{All-Comb}
& NIQE (1 fr/sec) & 0.4622 & 0.3222 & 0.4773 & 0.6112 \\
& ILNIQE (1 fr/sec) & 0.4592 & 0.3213 & 0.4741 & 0.6119 \\
& VIIDEO & 0.1039 & 0.0688 & 0.1621 & 0.6804 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table*}[!t]
\setlength{\tabcolsep}{2.7pt}
\renewcommand{\arraystretch}{1.1}
\centering
\begin{threeparttable}
\caption{Performance comparison of evaluated BVQA models on the four benchmark datasets. The \underline{\textbf{underlined}} and \textbf{boldfaced} entries indicate the best and top three performers on each database for each performance metric, respectively.}
\label{table:eval_svr}
\begin{tabular}{lcccccccccc}
\toprule
\textsc{Dataset} & \multicolumn{4}{c}{KoNViD-1k} & & \multicolumn{4}{c}{LIVE-VQC} \\ \cline{2-5}\cline{7-10}\\[-1.em]
\textsc{Model} \textbackslash\ \textsc{Metric} & \textsc{SRCC$\uparrow$ (std)} & \textsc{KRCC$\uparrow$ (std)} & \textsc{PLCC$\uparrow$ (std)} & \textsc{RMSE$\downarrow$ (std)} & & \textsc{SRCC$\uparrow$ (std)} & \textsc{KRCC$\uparrow$ (std)} & \textsc{PLCC$\uparrow$ (std)} & \textsc{RMSE$\downarrow$ (std)} \\
\hline\\[-1.em]
BRISQUE (1 fr/sec) & 0.6567 (.035) & 0.4761 (.029) & 0.6576 (.034) & 0.4813 (.022) & & 0.5925 (.068) & 0.4162 (.052) & 0.6380 (.063) & 13.100 (.796) \\
GM-LOG (1 fr/sec) & 0.6578 (.032) & 0.4770 (.026) & 0.6636 (.031) & 0.4818 (.022) & & 0.5881 (.068) & 0.4180 (.052) & 0.6212 (.063) & 13.223 (.822) \\
HIGRADE (1 fr/sec) & 0.7206 (.030) & 0.5319 (.026) & 0.7269 (.028) & 0.4391 (.018) & & 0.6103 (.068) & 0.4391 (.054) & 0.6332 (.065) & 13.027 (.904) \\
FRIQUEE (1 fr/sec) & 0.7472 (.026) & 0.5509 (.024) & 0.7482 (.025) & 0.4252 (.017) & & 0.6579 (.053) & 0.4770 (.043) & 0.7000 (.058) & 12.198 (.914) \\
CORNIA (1 fr/sec) & 0.7169 (.024) & 0.5231 (.021) & 0.7135 (.023) & 0.4486 (.018) & & 0.6719 (.047) & 0.4849 (.039) & 0.7183 (.042) & 11.832 (.700) \\
HOSA (1 fr/sec) & 0.7654 (.022) & 0.5690 (.021) & 0.7664 (.020) & 0.4142 (.016) & & 0.6873 (.046) & 0.5033 (.039) & {\textbf{0.7414 (.041)}} & {\textbf{11.353 (.747)}} \\
VGG-19 (1 fr/sec) & {\textbf{0.7741 (.028)}} & {\textbf{0.5841 (.027)}} & {\textbf{0.7845 (.024)}} & {\textbf{0.3958 (.017)}} & & 0.6568 (.053) & 0.4722 (.044) & 0.7160 (.048) & 11.783 (.696) \\
ResNet-50 (1 fr/sec) & \textbf{\underline{0.8018} (.025)} & {\textbf{\underline{0.6100} (.024)}} & {\textbf{\underline{0.8104} (.022)}} & {\textbf{\underline{0.3749} (.017)}} & & 0.6636 (.051) & 0.4786 (.042) & 0.7205 (.043) & 11.591 (.733) \\
KonCept512 (1 fr/sec) & 0.7349 (.025) & 0.5425 (.023) & 0.7489 (.024) & 0.4260 (.016) & & 0.6645 (.052) & 0.4793 (.045) & 0.7278 (.046) & 11.626 (.767) \\
PaQ-2-PiQ (1 fr/sec) & 0.6130 (.032) & 0.4334 (.026) & 0.6014 (.033) & 0.5148 (.019) & & 0.6436 (.045) & 0.4568 (.035) & 0.6683 (.044) & 12.619 (.848) \\
V-BLIINDS & 0.7101 (.031) & 0.5188 (.026) & 0.7037 (.030) & 0.4595 (.023) & & {\textbf{0.6939 (.050)}} & {\textbf{0.5078 (.042)}} & 0.7178 (.050) & 11.765 (.828) \\
TLVQM & 0.7729 (.024) & 0.5770 (.022) & 0.7688 (.023) & 0.4102 (.017) & & {\textbf{\underline{0.7988} (.036)}} & {\textbf{\underline{0.6080} (.037)}} & {\textbf{\underline{0.8025} (.036)}} & {\textbf{\underline{10.145} (.818)}} \\
VIDEVAL & \textbf{0.7832 (.021)} & \textbf{0.5845 (.021)} & \textbf{0.7803 (.022)} & \textbf{0.4026 (.017)} & & \textbf{0.7522 (.039)} & \textbf{0.5639 (.036)} & \textbf{0.7514 (.042)} & \textbf{11.100 (.810)} \\
\midrule
\textsc{Dataset} & \multicolumn{4}{c}{YouTube-UGC} & & \multicolumn{4}{c}{All-Combined\textsubscript{c}\hyperlink{all_exp}{$^\dagger$}} \\ \cline{2-5}\cline{7-10}\\[-1.em]
\textsc{Model} \textbackslash\ \textsc{Metric} & \textsc{SRCC$\uparrow$ (std)} & \textsc{KRCC$\uparrow$ (std)} & \textsc{PLCC$\uparrow$ (std)} & \textsc{RMSE$\downarrow$ (std)} & & \textsc{SRCC$\uparrow$ (std)} & \textsc{KRCC$\uparrow$ (std)} & \textsc{PLCC$\uparrow$ (std)} & \textsc{RMSE$\downarrow$ (std)} \\
\hline\\[-1.em]
BRISQUE (1 fr/sec) & 0.3820 (.051) & 0.2635 (.036) & 0.3952 (.048) & 0.5919 (.021) & & 0.5695 (.028) & 0.4030 (.022) & 0.5861 (.027) & 0.5617 (.016) \\
GM-LOG (1 fr/sec) & 0.3678 (.058) & 0.2517 (.041) & 0.3920 (.054) & 0.5896 (.022) & & 0.5650 (.029) & 0.3995 (.022) & 0.5942 (.030) & 0.5588 (.014) \\
HIGRADE (1 fr/sec) & {\textbf{0.7376 (.033)}} & {\textbf{0.5478 (.028)}} & {\textbf{0.7216 (.033)}} & {\textbf{0.4471 (.024)}} & & 0.7398 (.018) & 0.5471 (.016) & {0.7368 (.019)} & {0.4674 (.015)} \\
FRIQUEE\hyperlink{fri_exp}{$^{\star}$} (1 fr/sec) & {\textbf{{0.7652} (.030)}} & {\textbf{{0.5688} (.026)}} & {\textbf{{0.7571} (.032)}} & {\textbf{{0.4169} (.023)}} & & {\textbf{{0.7568} (.023)}} & {\textbf{{0.5651} (.021)}} & {\textbf{0.7550 (.022)}} & {\textbf{0.4549 (.018)}} \\
CORNIA (1 fr/sec) & 0.5972 (.041) & 0.4211 (.032) & 0.6057 (.039) & 0.5136 (.024) & & 0.6764 (.021) & 0.4846 (.017) & 0.6974 (.020) & 0.4946 (.013) \\
HOSA (1 fr/sec) & 0.6025 (.034) & 0.4257 (.026) & 0.6047 (.034) & 0.5132 (.021) & & 0.6957 (.018) & 0.5038 (.015) & 0.7082 (.016) & 0.4893 (.013) \\
VGG-19 (1 fr/sec) & 0.7025 (.028) & 0.5091 (.023) & 0.6997 (.028) & 0.4562 (.020) & & 0.7321 (.018) & 0.5399 (.016) & 0.7482 (.017) & {0.4610 (.013)} \\
ResNet-50 (1 fr/sec) & 0.7183 (.028) & 0.5229 (.024) & 0.7097 (.027) & 0.4538 (.021) & & {\textbf{0.7557 (.017)}} & {\textbf{0.5613 (.016)}} & {\textbf{{0.7747} (.016)}} & {\textbf{{0.4385} (.013)}} \\
KonCept512 (1 fr/sec) & 0.5872 (.039) & 0.4101 (.030) & 0.5940 (.041) & 0.5135 (.022) & & 0.6608 (.022) & 0.4759 (.018) & 0.6763 (.022) & 0.5091 (.014) \\
PaQ-2-PiQ (1 fr/sec) & 0.2658 (.047) & 0.1778 (.032) & 0.2935 (.049) & 0.6153 (.019) & & 0.4727 (.029) & 0.3242 (.021) & 0.4828 (.029) & 0.6081 (.015) \\
V-BLIINDS & 0.5590 (.049) & 0.3899 (.036) & 0.5551 (.046) & 0.5356 (.022) & & 0.6545 (.023) & 0.4739 (.019) & 0.6599 (.023) & 0.5200 (.016) \\
TLVQM & 0.6693 (.030) & 0.4816 (.025) & {0.6590} ({.030}) & {0.4849} ({.022}) & & {0.7271 (.018)} & {0.5347 (.016)} & {0.7342 (.018)} & {0.4705 (.013)} \\
VIDEVAL\hyperlink{fri_exp}{$^{\star}$} & \textbf{\underline{0.7787} (.025)} & \textbf{\underline{0.5830} (.023)} & \textbf{\underline{0.7733} (.025)} & \textbf{\underline{0.4049} (.021)} & & {\textbf{\underline{0.7960} (.015)}} & {\textbf{\underline{0.6032} (.014)}} & {\textbf{\underline{0.7939} (.015)}} & {\textbf{\underline{0.4268} (.015)}} \\
\bottomrule
\end{tabular}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item \hypertarget{fri_exp}{$^\star$}FRIQUEE and VIDEVAL were evaluated on a subset of 1,323 color videos in YouTube-UGC, denoted YouTube-UGC\textsubscript{c}, since it yields numerical errors when calculating on the remaining $57$ grayscale videos. For the other BVQA models evaluated, no significant difference was observed when evaluated on YouTube-UGC\textsubscript{c} versus YouTube-UGC, and hence we still report the results on YouTube-UGC. \\
\item \hypertarget{all_exp}{$^\dagger$}For a fair comparison, we only combined and calibrated (via INLSA \cite{pinson2003objective}) all the color videos from these three databases to obtain the combined dataset, i.e., All-Combined\textsubscript{c} (3,108)\! $=$\! KoNViD-1k (1,200)\! $+$\! LIVE-VQC (585)\! $+$\! YouTube-UGC\textsubscript{c} (1,323).
\end{tablenotes}
\end{threeparttable}
\end{table*}
\begin{figure*}[!t]
\captionsetup[subfigure]{justification=centering}
\centering
\def0ex{0em}
\def0.32{0.188}
\subfloat[BRISQUE]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_BRISQUE_kfCV_corr.pdf}
\label{fig1a}}
\hspace{0ex}
\subfloat[GM-LOG]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_GMLOG_kfCV_corr.pdf}
\label{fig1b}}
\hspace{0ex}
\subfloat[HIGRADE]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_HIGRADE1_kfCV_corr.pdf}
\label{fig1c}}
\hspace{0ex}
\subfloat[FRIQUEE]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_FRIQUEEALL_kfCV_corr.pdf}
\label{fig1d}}
\subfloat[CORNIA]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_CORNIA10K_kfCV_corr.pdf}
\label{fig1e}} \\[-2ex]
\subfloat[HOSA]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_HOSA_kfCV_corr.pdf}
\label{fig1f}}
\hspace{0ex}
\subfloat[VGG-19]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_vgg19_kfCV_corr.pdf}
\label{fig1g}}
\hspace{0ex}
\subfloat[ResNet-50]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_resnet50_kfCV_corr.pdf}
\label{fig1h}}
\hspace{0ex}
\subfloat[KonCept512]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_KONCEPT512_kfCV_corr.pdf}
\label{fig1i}}
\hspace{0ex}
\subfloat[PaQ-2-PiQ]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_PAQ2PIQ_kfCV_corr.pdf}
\label{fig1j}} \\ [-2ex]
\subfloat[V-BLIINDS]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_VBLIINDS_org_kfCV_corr.pdf}
\label{fig1k}}
\hspace{0ex}
\subfloat[TLVQM]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_TLVQM_kfCV_corr.pdf}
\label{fig1l}}
\hspace{0ex}
\subfloat[{VIDEVAL}]{\includegraphics[width=0.32\textwidth]{figs/ALL_COMBINED_FFVIQE_release_kfCV_corr.pdf}
\label{fig1m}}
\caption{Scatter plots and nonlinear logistic fitted curves of VQA models versus MOS trained with a grid-search SVR using $k$-fold cross-validation on the All-Combined\textsubscript{c} set. (a) BRISQUE (1 fr/sec), (b) GM-LOG (1 fr/sec), (c) HIGRADE (1 fr/sec), (d) FRIQUEE (1 fr/sec), (e) CORNIA (1 fr/sec), (f) HOSA (1 fr/sec), (g) VGG-19 (1 fr/sec), (h) ResNet-50 (1 fr/sec), (i) KonCept512 (1 fr/sec), (j) PaQ-2-PiQ (1 fr/sec), (k) V-BLIINDS, (l) TLVQM, and (m) VIDEVAL.}
\label{fig:kvcv_draw}
\end{figure*}
\section{Experimental Results}
\label{sec:exp}
\subsection{Evaluation Protocol}
\label{ssec:eval_proto}
\textbf{UGC Dataset Benchmarks.} To conduct BVQA performance evaluation, we used the three UGC-VQA databases: KoNViD-1K \cite{hosu2017konstanz}, LIVE-VQC \cite{sinno2018large}, and YouTube-UGC \cite{wang2019youtube}. We found that the YouTube-UGC dataset contains 57 grayscale videos, which yield numerical errors when computing the color model FRIQUEE. Therefore, we extracted a subset of 1,323 color videos from YouTube-UGC, which we denote here as the YouTube-UGC\textsubscript{c} set, for the evaluation of color models. In order to study overall model performances on all the databases, we created a large composite benchmark, which is referred to here as All-Combined\textsubscript{c}, using the iterative nested least squares algorithm (INLSA) suggested in \cite{pinson2003objective}, wherein YouTube-UGC is selected as the anchor set, and the objective MOS from the other two sets, KoNViD-1k and LIVE-VQC, are linearly mapped onto a common scale ($[1,5]$). Figure \ref{fig:inlsa} shows scatter plots of MOS versus NIQE scores before (Figure \ref{fig:inlsa-a}) and after (Figure \ref{fig:inlsa-b}) INLSA linear mapping, calibrated by NIQE \cite{mittal2012making} scores. The All-Combined\textsubscript{c} (3,108) dataset is simply the union of KoNViD-1k (1,200), LIVE-VQC (575), and YouTube-UGC\textsubscript{c} (1,323) after MOS calibration:
\begin{equation}
\label{eq:mos_cal_kon}
y_\mathrm{adj}=5-4\times\left[(5-y_\mathrm{org})/4\times1.1241-0.0993\right]
\end{equation}
\begin{equation}
\label{eq:mos_cal_live}
y_\mathrm{adj}=5-4\times\left[(100-y_\mathrm{org})/100\times0.7132+0.0253\right]
\end{equation}
where (\ref{eq:mos_cal_kon}) and (\ref{eq:mos_cal_live}) are for calibrating KoNViD-1k and LIVE-VQC, respectively. $y_\mathrm{adj}$ denotes the adjusted scores, while $y_\mathrm{org}$ is the original MOS.
\textbf{BVQA Model Benchmarks.} We include a number of representative BVQA/BIQA algorithms in our benchmarking evaluation as references to be compared against. These baseline models include NIQE \cite{mittal2012making}, ILNIQE \cite{zhang2015feature}, VIIDEO \cite{mittal2015completely}, BRISQUE \cite{mittal2012no}, GM-LOG \cite{xue2014blind}, HIGRADE \cite{kundu2017no}, FRIQUEE \cite{ghadiyaram2017perceptual}, CORNIA \cite{ye2012unsupervised}, HOSA \cite{xu2016blind}, KonCept512 \cite{hosu2020koniq}, PaQ-2-PiQ \cite{ying2019patches}, V-BLIINDS \cite{saad2014blind}, and TLVQM \cite{korhonen2019two}. Among these, NIQE, ILNIQE, and VIIDEO are ``completely blind'' (opinion-unaware (OU)), since no training is required to build them. The rest of the models are all training-based (opinion-aware (OA)) and we re-train the models/features when evaluating on a given dataset. We also utilized the well-known deep CNN models VGG-19 \cite{simonyan2014very} and ResNet-50 \cite{he2016deep} as additional CNN-based baseline models, where each was pretrained on the ImageNet classification task. The fully-connected layer (4,096-dim) from VGG-19 and average-pooled layer (2,048-dim) from ResNet-50 served as deep feature descriptors, by operating on 25 227$\times$227 random crops of each input frame, then average-pooled into a single feature vector representing the entire frame \cite{kim2017deep}. Two SOTA deep BIQA models, KonCept512 \cite{hosu2020koniq} and PaQ-2-PiQ \cite{ying2019patches}, were also included in our evaluations. We implemented the feature extraction process for each evaluated BVQA model using its initial released implementation in MATLAB R2018b, except that VGG-19 and ResNet-50 were implemented in TensorFlow, while KonCept512\footnote{\url{https://github.com/ZhengyuZhao/koniq-PyTorch}} and PaQ-2-PiQ\footnote{\url{https://github.com/baidut/paq2piq}} were implemented in PyTorch. All the feature-based BIQA models extract features at a uniform sampling rate of one frame per second, then temporally average-pooled to obtain the overall video-level feature.
\textbf{Regression Models.} We used a support vector regressor (SVR) as the back-end regression model to learn the feature-to-score mappings, since it achieves excellent performance in most cases \cite{korhonen2019two, saad2014blind, kim2017deep, ghadiyaram2017perceptual, mittal2012no, xu2014no}. The effectiveness of SVR, however, largely depends on the selection of its hyperparameters. As recommended in \cite{chang2011libsvm}, we optimized the SVR parameter values $(C,\gamma)$ by a grid-search of $10\times 10$ exponentially growing sequences (in our experiments, we used a grid of $C=2^1,2^2,...,2^{10},\gamma=2^{-8},2^{-7},...,2^{1}$) using cross-validation on the training set. The pair $(C,\gamma)$ yielding the best cross-validation performance, as measured by the root mean squared error (RMSE) between the predicted scores and the MOS, is picked. Afterward, the selected model parameters are applied to re-train the model on the entire training set, and we report the evaluation results on the test set. This kind of cross-validation procedure can prevent over-fitting, thus providing fair evaluation of the compared BVQA models. We chose the linear kernel for CORNIA, HOSA, VGG-19, and ResNet-50, considering their large feature dimension, and the radial basis function (RBF) kernel for all the other algorithms. We used Python 3.6.7 with the scikit-learn toolbox to train and test all the evaluated learning-based BVQA models.
\textbf{Performance Metrics.} Following convention, we randomly split the dataset into non-overlapping training and test sets ($80\%/20\%$), where the regression model was trained on the training set, and the performance was reported on the test set. This process of random split was iterated 100 times and the overall median performance was recorded. For each iteration, we adopted four commonly used performance criteria to evaluate the models: The Spearman Rank-Order Correlation Coefficient (SRCC) and the Kendall Rank-Order Correlation Coefficient (KRCC) are non-parametric measures of prediction monotonicity, while the Pearson Linear Correlation Coefficient (PLCC) with corresponding Root Mean Square Error (RMSE) are computed to assess prediction accuracy. Note that PLCC and RMSE are computed after performing a nonlinear four-parametric logistic regression to linearize the objective predictions to be on the same scale of MOS \cite{seshadrinathan2010study}.
\subsection{Performance on Individual and Combined Datasets}
\label{ssec:performance_diff_datasets}
Table \ref{table:compete-blind} shows the performance evaluation of the three ``completely blind'' BVQA models, NIQE, ILNIQE, and VIIDEO on the four UGC-VQA benchmarks. None of these methods performed very well, meaning that we still have much room for developing OU ``completely blind'' UGC video quality models.
Table \ref{table:eval_svr} shows the performance evaluation of all the learning-based BVQA models trained with SVR on the four datasets in our evaluation framework. For better visualization, we also show box plots of performances as well as scatter plots of predictions versus MOS on the All-Combined\textsubscript{c} set, in Figures \ref{fig:all_boxplot} and \ref{fig:kvcv_draw}, respectively. Overall, VIDEVAL achieves SOTA or near-SOTA performance on all the test sets. On LIVE-VQC, however, TLVQM outperformed other BVQA models by a notable margin, while it significantly underperformed on the more recent YouTube-UGC database. We observed in Section \ref{ssec:observation} that LIVE-VQC videos generally contain more (camera) motions than KoNViD-1k and YouTube-UGC, and TLVQM computes multiple motion relevant features. Moreover, the only three BVQA models containing temporal features (V-BLIINDS, TLVQM, and VIDEVAL) excelled on LIVE-VQC, which suggests that it is potentially valuable to integrate at least a few, if not many, motion-related features into quality prediction models, when assessing on videos with large (camera) motions.
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Performances on different resolution subsets: 1080p (427), 720p (566), and $\le$480p (448).}
\label{table:resolution_breakdown}
\begin{tabular}{lccccccccccccccccccccccccc}
\toprule
\textsc{Subset} & \multicolumn{2}{c}{1080p} & & \multicolumn{2}{c}{720p} & & \multicolumn{2}{c}{$\le$480p} \\ \cline{2-3}\cline{5-6}\cline{8-9}\\[-1.em]
\textsc{Model} & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC \\ \hline\\[-1.em]
BRISQUE & 0.4597 & 0.4637 & & 0.5407 & 0.5585 & & 0.3812 & 0.4065 \\
GM-LOG & 0.4796 & 0.4970 & & 0.5098 & 0.5172 & & 0.3685 & 0.4200 \\
HIGRADE & 0.5142 & 0.5543 & & 0.5095 & 0.5324 & & 0.4650 & 0.4642 \\
FRIQUEE & 0.5787 & 0.5797 & & 0.5369 & 0.5652 & & 0.5042 & 0.5363 \\
CORNIA & 0.5951 & \textbf{0.6358} & & 0.6212 & 0.6551 & & 0.5631 & 0.6118 \\
HOSA & 0.5924 & 0.6093 & & \textbf{\underline{0.6651}} & \textbf{0.6739} & & \textbf{0.6514} & \textbf{0.6652} \\
VGG-19 & \textbf{0.6440} & 0.6090 & & 0.6158 & \textbf{0.6568} & & \textbf{0.5845} & \textbf{0.6267} \\
ResNet-50 & \textbf{\underline{0.6615}} & \textbf{\underline{0.6644}} & & \textbf{0.6645} & \textbf{\underline{0.7076}} & & \textbf{\underline{0.6570}} & \textbf{\underline{0.6997}} \\
{KonCept512} & \textbf{0.6332} & \textbf{0.6336} & & 0.6055 & 0.6514 & & 0.4271 & 0.4612 \\
{PaQ-2-PiQ} & 0.5304 & 0.5176 & & 0.5768 & 0.5802 & & 0.3646 & 0.4748 \\
V-BLIINDS & 0.4449 & 0.4491 & & 0.5546 & 0.5719 & & 0.4484 & 0.4752 \\
TLVQM & 0.5638 & 0.6031 & & \textbf{0.6300} & 0.6526 & & 0.4318 & 0.4784 \\
VIDEVAL & 0.5805 & 0.6111 & & {0.6296} & {0.6393} & & 0.5014 & 0.5508 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Performances on different content subsets: screen content (163), animation (81), and gaming (209).}
\label{table:content_breakdown}
\begin{tabular}{lccccccccccccccccccccccccc}
\toprule
\textsc{Subset} & \multicolumn{2}{c}{Screen Content} & & \multicolumn{2}{c}{Animation} & & \multicolumn{2}{c}{Gaming} \\ \cline{2-3}\cline{5-6}\cline{8-9}\\[-1.em]
\textsc{Model} & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC \\ \hline\\[-1.em]
BRISQUE & 0.2573 & 0.3954 & & 0.0747 & 0.3857 & & 0.2717 & 0.3307 \\
GM-LOG & 0.3004 & 0.4244 & & 0.2009 & 0.4129 & & 0.3371 & 0.4185 \\
HIGRADE & 0.4971 & 0.5652 & & 0.1985 & 0.4140 & & 0.6228 & 0.6832 \\
FRIQUEE & \textbf{0.5522} & \textbf{0.6160} & & 0.2377 & 0.4574 & & \textbf{0.6919} & \textbf{0.7193} \\
CORNIA & 0.5105 & 0.5667 & & 0.1936 & 0.4627 & & 0.5741 & 0.6502 \\
HOSA & 0.4667 & 0.5255 & & 0.1048 & 0.4489 & & 0.6019 & \textbf{0.6998} \\
VGG-19 & 0.5472 & 0.6229 & & 0.1973 & 0.4700 & & 0.5765 & 0.6370 \\
ResNet-50 & \textbf{\underline{0.6199}} & \textbf{\underline{0.6676}} & & \textbf{0.2781} & \textbf{0.4871} & & \textbf{0.6378} & 0.6779 \\
{KonCept512} & 0.4714 & 0.5119 & & \textbf{0.2757} & \textbf{0.5229} & & 0.4780 & 0.6240 \\
{PaQ-2-PiQ} & 0.3231 & 0.4312 & & 0.0208 & 0.4630 & & 0.2169 & 0.3874 \\
V-BLIINDS & 0.3064 & 0.4155 & & 0.0379 & 0.3917 & & 0.5473 & 0.6101 \\
TLVQM & 0.3843 & 0.4524 & & 0.2708 & 0.4598 & & 0.5749 & 0.6195 \\
VIDEVAL & \textbf{{0.6033}} & \textbf{{0.6610}} & & \textbf{\underline{0.3492}} & \textbf{\underline{0.5274}} & & \textbf{\underline{0.6954}} & \textbf{\underline{0.7323}} \\
\bottomrule
\end{tabular}
\end{table}
It is also worth mentioning that the deep CNN baseline methods (VGG-19 and ResNet-50), despite being trained as picture-only models, performed quite well on KoNViD-1k and All-Combined\textsubscript{c}. This suggests that transfer learning is a promising technique for the blind UGC-VQA problem, consistent with conclusions drawn for picture-quality prediction \cite{kim2017deep}. Deep models will perform even better, no doubt, if trained on temporal content and distortions.
{The two most recent deep learning picture quality models, PaQ-2-PiQ, and KonCept512, however, did not perform very well on the three evaluated video datasets. The most probable reason would be that these models were trained on picture quality datasets \cite{ying2019patches, hosu2020koniq}, which contain different types of (strictly spatial) distortions than UGC-VQA databases. Models trained on picture quality sets do not necessarily transfer very well to UGC video quality problems. In other words, whatever model should be either trained or fine-tuned on UGC-VQA datasets in order to obtain reasonable performance. Indeed, if temporal distortions (like judder) are present, they may severely underperform if the frame quality is high \cite{madhusudana2020subjective}.}
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Performances on different quality subsets: low quality (1558) and high quality (1550).}
\label{table:quality_breakdown}
\begin{tabular}{lccccccccccccccccccccccccc}
\toprule
\textsc{Subset} & \multicolumn{2}{c}{Low Quality} & & \multicolumn{2}{c}{High Quality} \\ \cline{2-3}\cline{5-6}\\[-1.em]
\textsc{Model} & SRCC & PLCC & & SRCC & PLCC \\ \hline\\[-1.em]
BRISQUE & 0.4312 & 0.4593 & & 0.2813 & 0.2979 \\
GM-LOG & 0.4221 & 0.4715 & & 0.2367 & 0.2621 \\
HIGRADE & 0.5057 & 0.5466 & & 0.4714 & 0.4799 \\
FRIQUEE & \textbf{0.5460} & \textbf{0.5886} & & \textbf{0.5061} & \textbf{0.5152} \\
CORNIA & 0.4931 & 0.5435 & & 0.3610 & 0.3748 \\
HOSA & \textbf{0.5348} & \textbf{0.5789} & & 0.4208 & 0.4323 \\
VGG-19 & 0.3710 & 0.4181 & & 0.3522 & 0.3614 \\
ResNet-50 & 0.3881 & 0.4250 & & 0.2791 & 0.3030 \\
{KonCept512} & 0.3428 & 0.4497 & & 0.2245 & 0.2597 & \\
{PaQ-2-PiQ} & 0.2438 & 0.2713 & & 0.2013 & 0.2252 \\
V-BLIINDS & 0.4703 & 0.5060 & & 0.3207 & 0.3444 \\
TLVQM & 0.4845 & 0.5386 & & \textbf{0.4783} & \textbf{0.4860} \\
VIDEVAL & \textbf{\underline{0.5680}} & \textbf{\underline{0.6056}} & & \textbf{\underline{0.5546}} & \textbf{\underline{0.5657}} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!t]
\setlength{\tabcolsep}{3.5pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Best model in terms of SRCC for cross dataset generalization evaluation.}
\label{table:cross_dataset_srcc}
\begin{tabular}{lcccccccc}
\toprule
\textsc{Train}\textbackslash\textsc{Test} & LIVE-VQC & KoNViD-1k & YouTube-UGC\textsubscript{c} \\
\hline\\[-1.em]
LIVE-VQC & - & ResNet-50 (0.69) & ResNet-50 (0.33) \\
KoNViD-1k & ResNet-50 (0.70) & - & VIDEVAL (0.37) \\
YouTube-UGC\textsubscript{c} & HOSA (0.49) & VIDEVAL (0.61) & - \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!t]
\setlength{\tabcolsep}{3.5pt}
\renewcommand{\arraystretch}{1.}
\centering
\caption{Best model in terms of PLCC for cross dataset generalization evaluation.}
\label{table:cross_dataset_plcc}
\begin{tabular}{lcccccccc}
\toprule
\textsc{Train}\textbackslash\textsc{Test} & LIVE-VQC & KoNViD-1k & YouTube-UGC\textsubscript{c} \\
\hline\\[-1.em]
LIVE-VQC & - & ResNet-50 (0.70) & VIDEVAL (0.35) \\
KoNViD-1k & ResNet-50 (0.75) & - & VIDEVAL (0.39) \\
YouTube-UGC\textsubscript{c} & HOSA (0.50) & VIDEVAL (0.62) & - \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Performance Evaluation on Categorical Subsets}
We propose three new categorical evaluation methodologies - resolution, quality, and content-based category breakdown. These will allow us to study the compared BVQA models from additional and practical aspects in the context of real-world UGC scenarios, which have not been, nor can it be accounted in previous legacy VQA databases or studies.
\begin{table*}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{{Performance comparison of a total of eleven temporal pooling methods using TLVQM and VIDEVAL as testbeds on KoNViD-1k, LIVE-VQC, and YouTube-UGC. The three best results along each column are \textbf{boldfaced}.}}
\label{table:pooling}
\begin{tabular}{lccccccccccccccccccccccc}
\toprule
\textsc{Database} & \multicolumn{5}{c}{KoNViD-1k} & & \multicolumn{5}{c}{LIVE-VQC} & & \multicolumn{5}{c}{YouTube-UGC} \\ \cline{2-6}\cline{8-12}\cline{14-18}\\[-1.em]
\textsc{Model} & \multicolumn{2}{c}{TLVQM} & & \multicolumn{2}{c}{VIDEVAL} & & \multicolumn{2}{c}{TLVQM} & & \multicolumn{2}{c}{VIDEVAL} & & \multicolumn{2}{c}{TLVQM} & & \multicolumn{2}{c}{VIDEVAL} \\
\cline{2-3}\cline{5-6}\cline{8-9}\cline{11-12}\cline{14-15}\cline{17-18}\\[-1.em]
\textsc{Pooling} & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC & & SRCC & PLCC \\ \hline\\[-1.em]
Mean & \textbf{0.7511} & \textbf{0.7475} & & 0.7749 & \textbf{0.7727} & & \textbf{0.7917} & \textbf{0.7984} & & \textbf{0.7396} & 0.7432 & & \textbf{0.6369} & \textbf{0.6310} & & 0.7447 & 0.7332\\
Median & 0.7483 & 0.7437 & & 0.7650 & 0.7698 & & 0.7708 & 0.7887 & & 0.7236 & 0.7308 & & 0.6127 & 0.6090 & & 0.7452 & \textbf{0.7448} \\
Harmonic & 0.7458 & 0.7392 & & \textbf{0.7772} & 0.7681 & & 0.7845 & 0.7890 & & 0.7312 & 0.7250 & & 0.6119 & 0.6038 & & 0.7449 & 0.7318 \\
Geometric & 0.7449 & 0.7461 & & 0.7566 & 0.7592 & & \textbf{0.7878} & \textbf{0.7964} & & \textbf{0.7412} & 0.7487 & & 0.6347 & 0.6236 & & \textbf{0.7508} & \textbf{0.7437} \\
Minkowski & 0.7498 & \textbf{0.7481} & & \textbf{0.7775} & \textbf{0.7727} & & 0.7863 & 0.7908 & & 0.7371 & \textbf{0.7558} & & \textbf{0.6368} & \textbf{0.6311} & & \textbf{0.7542} & \textbf{0.7508} \\
Percentile & 0.7078 & 0.7000 & & 0.7161 & 0.7049 & & 0.7378 & 0.7313 & & 0.6596 & 0.6576 & & 0.4871 & 0.4996 & & 0.6443 & 0.6465 \\
VQPooling & 0.7240 & 0.7196 & & 0.7366 & 0.7296 & & 0.7696 & 0.7895 & & 0.7240 & 0.7311 & & 0.5654 & 0.5618 & & 0.6942 & 0.6862 \\
Primacy & 0.7456 & 0.7451 & & 0.7711 & 0.7700 & & 0.7751 & 0.7851 & & 0.7349 & \textbf{0.7523} & & 0.5734 & 0.5692 & & 0.7221 & 0.7156 \\
Recency & \textbf{0.7528} & 0.7470 & & 0.7683 & 0.7677 & & 0.7715 & 0.7857 & & \textbf{0.7405} & \textbf{0.7584} & & 0.5821 & 0.5695 & & 0.7176 & 0.7116 \\
Hysteresis & 0.7434 & 0.7430 & & 0.7612 & 0.7554 & & 0.7856 & 0.7901 & & 0.7226 & 0.7433 & & 0.6092 & 0.6109 & & 0.7370 & 0.7306 \\
EPooling & \textbf{0.7641} & \textbf{0.7573} & & \textbf{0.7831} & \textbf{0.7867} & & \textbf{0.7925} & \textbf{0.7917} & & 0.7371 & 0.7372 & & \textbf{0.6452} & \textbf{0.6592} & & \textbf{0.7517} & 0.7379 \\
\toprule
\end{tabular}
\end{table*}
For resolution-dependent evaluation, we divided the All-Combined\textsubscript{c} set into three subsets, based on video resolution: (1) 427 1080p-videos (110 from LIVE-VQC, 317 from YouTube-UGC), (2) 566 720p-videos (316 from LIVE-VQC, 250 from YouTube-UGC), and (3) 448 videos with resolution $\le$480p (29 from LIVE-VQC, 419 from YouTube-UGC), since we are also interested in performance on videos of different resolutions. We did not include 540p-videos, since those videos are almost exclusively from KoNViD-1k. Table \ref{table:resolution_breakdown} shows the resolution-breakdown evaluation results. Generally speaking, learned features (CORNIA, HOSA, VGG-19, KonCept512, and ResNet-50) outperformed hand-designed features, among which ResNet-50 ranked first.
\begin{figure}[!t]
\centering
\footnotesize
\def0.32{0.32}
\def0ex{0ex}
\def0.3264\linewidth{0.3264\linewidth}
\begin{tabular}{cc}
\multirow{1}{*}[0.3264\linewidth]{\includegraphics[height=0.32\linewidth]{figs/Vlog_2160P-408f_plot.jpg}} &
{\includegraphics[height=0.3668\linewidth]{figs/resolution_nss.pdf}} \\
(a) Vlog\_2160P-408f.mkv & (b) MSCN distributions \\
\end{tabular}
\caption{{(a) An examplary 2160p video from YouTube-UGC and (b) the mean-subtracted contrast-normalized (MSCN) distributions of its downscaled versions: 2160p, 1440p, 1080p, 720p, 480p, and 360p.}}
\label{fig:nss}
\end{figure}
{Here we make two arguments to try to explain the observations above: (1) video quality is intrinsically correlated with resolution; (2) NSS features are implicitly \textit{resolution-aware}, while CNN features are not. The first point is almost self-explanatory, no matter to what degree one agrees. To further justify this, we trained an SVR only using resolution (height, width) as features to predict MOS on YouTube-UGC, which contains balanced samples across five different resolutions. This yielded surprisingly high values $0.576 / 0.571$ for SRCC$/$PLCC, indicating the inherent correlation between video quality and resolution. Secondly, we selected one 2160p video from YouTube-UGC, namely `Vlog2160P-408f.mkv,' and plotted, in Figure \ref{fig:nss}, the mean-subtracted contrast-normalized (MSCN) distributions of its downscaled versions: 2160p, 1440p, 1080p, 720p, 480p, and 360p. It may be observed that resolution can be well separated by MSCN statistics, based on which most feature-based methods are built. We may infer, from these two standpoints, that including various resolutions of videos is favorable to the training of NSS-based models, since NSS features are resolution-aware, and resolution is further well correlated with quality. In other words, the resolution-breakdown evaluation shown in Table \ref{table:resolution_breakdown}, which removes this important implicit feature (resolution), would possibly reduce the performance of NSS-based models, such as FRIQUEE and VIDEVAL.}
\begin{table*}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\begin{threeparttable}
\caption{Feature description, dimensionality, computational complexity, and average runtime comparison (in seconds evaluated on twenty $1080p$ videos from LIVE-VQC) among MATLAB-implemented BVQA models.}
\label{table:complexity}
\begin{tabular}{llp{5.5cm}cp{5.5cm}c}
\toprule
\textsc{Class} & \textsc{Model} & \textsc{Feature Description} & \textsc{Dim} & \textsc{Computational Complexity} & \textsc{Time (sec)} \\
\hline\\[-1.em]
\multirow{15}{*}{\shortstack{IQA}} & NIQE (1 fr/sec) & Spatial NSS & 1 & $\mathcal{O}(d^2 NT)$ $d$: window size & 6.3 \\
& ILNIQE (1 fr/sec) & {Spatial NSS, gradient, log-Gabor, and color statistics} & 1 & {$\mathcal{O}((d^2+h+gh)NT)$ $d$: window size; $h$: filter size; $g$: log-Gabor filter size} & 23.3 \\
& BRISQUE (1 fr/sec) & Spatial NSS & 36 & $\mathcal{O}(d^2 NT)$ $d$: window size & 1.7 \\
& GM-LOG (1 fr/sec) & Joint statistics of gradient magnitude and laplacian of gaussian coefficients & 40 & {$\mathcal{O}(((h+k)NT)$ $d$: window size; $k$: probability matrix size} & 2.1 \\
& HIGRADE (1 fr/sec) & Spatial NSS, and gradient magnitude statistics in LAB color space & 216 & {$\mathcal{O}(3(2d^2+k)NT)$ $d$: window size; $k$: gradient kernel size} & 11.6 \\
& FRIQUEE (1 fr/sec) & Complex streerable pyramid wavelet, luminance, chroma, LMS, HSI, yellow channel, and their transformed domain statistics & 560 & $\mathcal{O}((fd^2 N+4N(\log(N)+m^2))T)$ $d$: window size; $f$: number of color spaces; $m$: neighborhood size in DNT & 701.2 \\
& CORNIA (1 fr/sec) & Spatially normalized image patches and max min pooling & 10k & $\mathcal{O}(d^2 KNT)$ $d$: window size $K$: codebook size & 14.3 \\
& HOSA (1 fr/sec) & Local normalized image patches based on high order statistics aggregation & 14.7k & $\mathcal{O}(d^2 KNT)$ $d$: window size $K$: codebook size & 1.2 \\ \\[-1.1em]
\hline\\[-1.em]
\multirow{10}{*}{\shortstack{VQA}} & VIIDEO & Frame difference spatial statistics, inter sub-band statistics & 1 & $\mathcal{O}(N\log(N)T)$ & 674.8 \\
& V-BLIINDS & Spatial NSS, frame difference DCT coefficient statistics, motion coherency, and egomotion & 47 & $\mathcal{O}((d^2 N+\log(k)N+k^2 w^3)T)$ $d$: window size; $k$: block size; $w$: motion vector tensor size & 1989.9 \\
& TLVQM & Captures impairments computed at two computation levels: low complexity and high complexity features & 75 & $\mathcal{O}((h_1^2 N+k^2 K)T_1+(\log(N)+h_2^2) NT_2))$ $h_1,h_2$: filter size; $k$: motion estimation block size; $K$: number of key points & 183.8 \\
& VIDEVAL & Selected combination of NSS features in multiple perceptual spaces and using visual impairment features from TLVQM & 60 & $\mathcal{O}((fh_1^2N+k^2K)T_1+h_2^2NT_2)$ $h_1,h_2$: filter size; $f$: number of color spaces; $k$: motion estimation block size; $K$: number of key points & 305.8 \\
\bottomrule
\end{tabular}
\begin{tablenotes}[para,flushleft]
\item $N$: number of pixels per frame; $T$: number of frames computed for feature extraction. Note that for VIIDEO and V-BLIINDS, $T$ is the total number of frames, whereas for IQA models, $T$ equals the total number of frames sampled at 1 fr/sec. For TLVQM and VIDEVAL, $T_1$ is total number of frames divided by 2, while $T_2$ is the number of frames sampled at 1 fr/sec.
\end{tablenotes}
\end{threeparttable}
\end{table*}
We also divided the All-Combined\textsubscript{c} into subsets based on content category: Screen Content (163), Animation (81), Gaming (209), and Natural (2,667) videos. We only reported the evaluation results on the first three subsets in Table \ref{table:content_breakdown}, since we observed similar results on the Natural subset with the entire combined set. The proposed VIDEVAL model outperformed over all categories, followed by ResNet-50 and FRIQUEE, suggesting that VIDEVAL features are robust quality indicatives across different content categories.
The third categorical division is based on quality scores: we partitioned the combined set into Low Quality (1,558) and High Quality (1,550) halves, using the median quality value 3.5536 as the threshold, to see the model performance only on high/low quality videos. Performance results are shown in Table \ref{table:quality_breakdown}, wherein VIDEVAL still outperformed the other BVQA models on both low and high quality partitions.
\subsection{Cross Dataset Generalizability}
We also performed a cross dataset evaluation to verify the generalizability of BVQA models, wherein LIVE-VQC, KoNViD-1k, and YouTube-UGC\textsubscript{c} were included. That is, we trained the regression model on one full database and report the performance on another. To retain label consistency, we linearly scaled the MOS values in LIVE-VQC from raw $[0,100]$ to $[1,5]$, which is the scale for the other two datasets. We used SVR for regression and adopted $k$-fold cross validation using the same grid-search as in Section \ref{ssec:eval_proto} for hyperparameter selection. The selected parameter pair were then applied to re-train the SVR model on the full training set, and the performance results on the test set were recorded. Table \ref{table:cross_dataset_srcc} and \ref{table:cross_dataset_plcc} show the best performing methods with cross domain performances in terms of SRCC and PLCC, respectively.
We may see that the cross domain BVQA algorithm generalization between LIVE-VQC and KoNViD-1k was surprisingly good, and was well characterized by pre-trained ResNet-50 features. We also observed better algorithm generalization between KoNViD-1k and YouTube-UGC than LIVE-VQC, as indicated by the performances of the best model, VIDEVAL. This might be expected, since as Figure \ref{fig:tsne} shows, YouTube-UGC and KoNViD-1k share overlapped coverage of content space, much larger than that of LIVE-VQC. Therefore, we may conclude that VIDEVAL and ResNet-50 were the most robust BVQA models among those compared in terms of cross domain generalization capacities.
\subsection{Effects of Temporal Pooling}
Temporal pooling is one of the most important, unresolved problems for video quality prediction \cite{park2012video, tu2020comparative,seshadrinathan2011temporal, korhonen2019two, bampis2018recurrent}. In our previous work \cite{tu2020comparative}, we have studied the efficacy of various pooling methods using scores predicted by BIQA models. Here we extend this to evaluate on SOTA BVQA models. For practical considerations, the high-performing TLVQM and VIDEVAL were selected as exemplar models. Since these two models independently extract features on each one-second block, we applied temporal pooling of chunk-wise quality predictions. A total of eleven pooling methods were tested: three Pythagorean means (arithmetic, geometric, and harmonic mean), median, Minkowski ($p=2$) mean, percentile pooling ($20\%$) \cite{moorthy2009visual}, VQPooling \cite{park2012video}, primacy and recency pooling \cite{murdock1962serial}, hysteresis pooling \cite{seshadrinathan2011temporal}, and our previously proposed ensemble method, EPooling \cite{tu2020comparative}, which aggregates multiply pooled scores by training a second regressor on top of mean, Minkowski, percentile, VQPooling, variation, and hysteresis pooling. We refer the reader to \cite{tu2020comparative} for detailed algorithmic formulations and parameter settings thereof.
It is worth noting that the results in Table \ref{table:pooling} are only \textit{self-consistent}, meaning that they are not comparable to any prior experiments - since we employed chunk-wise instead of previously adopted video-wise quality prediction to be able to apply temporal quality pooling, which may affect the base performance. Here we observed yet slightly different results using BVQA testbeds as compared to what we observed on BIQA \cite{tu2020comparative}. Generally, we found the mean families and ensemble pooling to be the most reliable pooling methods. Traditional sample mean prediction may be adequate in many cases, due to its simplicity. Pooling strategies that more heavily weight low-quality parts, however, were not observed to perform very well on the tested BVQA, which might be attributed to the fact that not enough samples ($8\sim 20$) can be extracted from each video to attain statistically meaningful results.
\subsection{Complexity Analysis and Runtime Comparison}
The efficiency of a video quality model is of vital importance in practical commercial deployments. Therefore, we also tabulated the computational complexity and runtime cost of the compared BVQA models, as shown in Tables \ref{table:complexity}, \ref{table:time_dl}. The experiments were performed in MATLAB R2018b and Python 3.6.7 under Ubuntu 18.04.3 LTS system on a Dell OptiPlex 7080 Desktop with Intel Core i7-8700 [email protected], 32G RAM, and GeForce GTX 1050 Graphics Cards. The average feature computation time of MATLAB-implemented BVQA models on 1080p videos are reported in Table \ref{table:complexity}. The proposed VIDEVAL method achieves a reasonable complexity among the top-performing algorithms, TLVQM, and FRIQUEE. We also present theoretical time complexity in Table \ref{table:complexity} for potential analytical purposes.
{We also provide in Table \ref{table:time_dl} an additional runtime comparison between MATLAB models on CPU and deep learning models on CPU and GPU, respectively. It may be observed that top-performing BVQA models such as TLVQM and VIDEVAL are essentially slower than deep CNN models, but we expect orders-of-magnitude speedup if re-implemented in pure $\text{C}/\text{C}\texttt{++}$. Simpler NSS-based models such as BRISQUE and HIGRADE (which only involve several convolution operations) still show competitive efficiency relative to CNN models even when implemented in MATLAB. We have also seen a $5\sim 10$ times speedup switching from CPU to GPU for the CNN models, among which KonCept512 with PyTorch-GPU was the fastest since it requires just a single pass to the CNN backbone, while the other three entail multiple passes for each input frame.}
Note that the training/test time of the machine learning regressor is approximately proportional to the number of features. Thus, it is not negligible compared to feature computation given a large number of features, regardless of the regression model employed. The feature dimension of each model is listed in Table \ref{table:complexity}. As may be seen, codebook-based algorithms (CORNIA (10$k$) and HOSA (14.7$k$)) require significantly larger numbers of features than other hand-crafted feature based models. Deep ConvNet features ranked second in dimension (VGG-19 (4,080) and ResNet-50 (2,048)). Our proposed VIDEVAL only uses 60 features, which is fairly compact, as compared to other top-performing BVQA models like FRIQUEE (560) and TLVQM (75).
\begin{table}[!t]
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{{Run time comparison of feature-based and deep learning BVQA models (in seconds evaluated on twenty $1080p$ videos from LIVE-VQC). Model loading time for deep models are excluded}.}
\label{table:time_dl}
\begin{tabular}{lrrccccccccccccccccccccccc}
\toprule
\textsc{Model} & & \textsc{Time (Sec)} \\ \hline\\[-1.em]
BRISQUE (1 fr/sec) & MATLAB-CPU & 1.7 \\
HOSA (1 fr/sec) & MATLAB-CPU & 1.2 \\
TLVQM & MATLAB-CPU & 183.8 \\
VIDEVAL & MATLAB-CPU & 305.8 \\
\hline\\[-1.em]
VGG-19 (1 fr/sec) & TensorFlow-CPU & 27.8 \\
& TensorFlow-\textit{GPU} & 5.7 \\
ResNet-50 (1 fr/sec) & TensorFlow-CPU & 9.6 \\
& TensorFlow-\textit{GPU} & 1.9 \\
\hline\\[-1.em]
KonCept512 (1 fr/sec) & PyTorch-CPU & 2.8 \\
& PyTorch-\textit{GPU} & 0.3 \\
PaQ-2-PiQ (1 fr/sec) & PyTorch-CPU & 6.9 \\
& PyTorch-\textit{GPU} & 0.8 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{{Ensembling VIDEVAL with Deep Features}}
\label{ssec:ensemble}
We also attempted a more sophisticated ensemble fusion of VIDEVAL and deep learning features to determine whether this could further boost its performance, which could give insights on the future direction of this field. Since PaQ-2-PiQ aimed for local quality prediction, we included the predicted $3\times 5$ local quality scores as well as a single global score, as additional features. For KonCept512, the feature vector (256-dim) immediately before the last linear layer in the fully-connected head was appended. Our own baseline CNN models, VGG-19 and ResNet-50, were also considered, because these are commonly used standards for downstream vision tasks.
The overall results are summarized in Table \ref{table:fusion}. We may observe that ensembling VIDEVAL with certain deep learning models improved the performance by up to $\sim 4\%$ compared to the vanilla VIDEVAL, which is very promising. Fusion with either ResNet-50 or KonCept512 yielded top performance. It should be noted that the number of fused features is also an essential aspect. For example, blending VIDEVAL (60-dim) with VGG-19 (4,096-dim) may not be recommended, since the enormous number of VGG-19 features could possibly dominate the VIDEVAL features, as suggested by some performance drops in Table \ref{table:fusion}.
\begin{table}[!t]
\setlength{\tabcolsep}{3.5pt}
\renewcommand{\arraystretch}{1.1}
\centering
\caption{Performance of the ensemble VIDEVAL models fused with additional deep learning features.}
\label{table:fusion}
\begin{tabular}{llcccc}
\toprule
\textsc{Dataset} & \textsc{Model} \textbackslash\ \textsc{Metric} & SRCC & KRCC & PLCC & RMSE \\
\hline\\[-1.em]
\multirow{5}{*}{KoNViD} &
VIDEVAL & 0.7832 & 0.5845 & 0.7803 & 0.4024 \\
& VIDEVAL$+$VGG-19 & 0.7827 & 0.5928 & 0.7913 & 0.3897 \\
& VIDEVAL$+$ResNet-50 & 0.8129 & 0.6212 & \textbf{0.8200} & \textbf{0.3659} \\
& VIDEVAL$+$KonCept512 & \textbf{0.8149} & \textbf{0.6251} & 0.8169 & 0.3670 \\
& VIDEVAL$+$PaQ-2-PiQ & 0.7844 & 0.5891 & 0.7793 & 0.4018 \\
\hline\\[-1.em]
\multirow{5}{*}{LIVE-VQC} &
VIDEVAL & 0.7522 & 0.5639 & 0.7514 & 11.100 \\
& VIDEVAL$+$VGG-19 & 0.7274 & 0.5375 & 0.7717 & 10.749 \\
& VIDEVAL$+$ResNet-50 & 0.7456 & 0.5555 & 0.7810 & 10.385 \\
& VIDEVAL$+$KonCept512 & \textbf{0.7849} & \textbf{0.5953} & \textbf{0.8010} & \textbf{10.145} \\
& VIDEVAL$+$PaQ-2-PiQ & 0.7677 & 0.5736 & 0.7686 & 10.787 \\
\hline\\[-1.em]
\multirow{5}{*}{YT-UGC} &
VIDEVAL & 0.7787 & 0.5830 & 0.7733 & 0.4049 \\
& VIDEVAL$+$VGG-19 & 0.7868 & 0.5930 & 0.7847 & 0.3993 \\
& VIDEVAL$+$ResNet-50 & \textbf{0.8085} & 0.6128 & \textbf{0.8033} & \textbf{0.3837} \\
& VIDEVAL$+$KonCept512 & 0.8083 & \textbf{0.6139} & 0.8028 & 0.3859 \\
& VIDEVAL$+$PaQ-2-PiQ & 0.7981 & 0.6015 & 0.7941 & 0.3959 \\
\hline\\[-1.em]
\multirow{5}{*}{All-Comb} &
VIDEVAL & 0.7960 & 0.6032 & 0.7939 & 0.4268 \\
& VIDEVAL$+$VGG-19 & 0.7859 & 0.5912 & 0.7962 & 0.4202 \\
& VIDEVAL$+$ResNet-50 & 0.8115 & \textbf{0.6207} & \textbf{0.8286} & \textbf{0.3871} \\
& VIDEVAL$+$KonCept512 & \textbf{0.8123} & 0.6193 & 0.8168 & 0.4017 \\
& VIDEVAL$+$PaQ-2-PiQ & 0.7962 & 0.5991 & 0.7934 & 0.4229 \\
\toprule
\end{tabular}
\end{table}
\subsection{Summarization and Takeaways}
Finally, we briefly summarize the experimental results and make additional observations:
\begin{enumerate}
\item Generally, spatial distortions dominated quality prediction on Internet UGC videos like those from YouTube and Flickr, as revealed by the remarkable performances of picture-only models (e.g., HIGRADE, FRIQUEE, HOSA, ResNet-50) on them. Some motion-related features (as in TLVQM) may not apply as well in this scenario.
\item On videos captured with mobile devices (e.g., those in LIVE-VQC), which often present larger and more frequent camera motions, including temporal- or motion-related features can be advantageous (e.g., V-BLIINDS, TLVQM, VIDEVAL).
\item Deep CNN feature descriptors (VGG-19, ResNet-50, etc.) pre-trained for other classical vision tasks (e.g. image classification) are transferable to UGC video quality predictions, achieving very good performance, suggesting that using transfer learning to address the general UGC-VQA problem is very promising.
\item It is still a very hard problem to predict UGC video quality on non-natural or computer-generated video contents: screen contents, animations, gaming, etc. Moreover, there are no sufficiently large UGC-VQA datasets designed for those kinds of contents.
\item A simple feature engineering and selection implementation built on top of current effective feature-based BVQA models is able to obtain excellent performance, as exemplified by the compact new model (VIDEVAL).
\item Simple temporal mean pooling of chunk-wise quality predictions by BVQA models yields decent and robust results. Furthermore, an ensemble pooling approach can noticeably improve the quality prediction performance, albeit with higher complexity.
\item Ensembling scene statistics-based BVQA models with additional deep learning features (e.g., VIDEVAL plus KonCept512) could further raise the performance upper bound, which may be a promising way of developing future BVQA models.
\end{enumerate}
\section{Conclusion}
\label{sec:conclud}
We have presented a comprehensive analysis and empirical study of blind video quality assessment for user-generated content (the \textbf{\textsf{UGC-VQA problem}}). We also proposed a new fusion-based BVQA model, called the VIDeo quality EVALuator (VIDEVAL), which uses a feature ensemble and selection procedure on top of existing efficient BVQA models. A systematic evaluation of prior leading video quality models was conducted within a unified and reproducible evaluation framework and accordingly, we concluded that a selected fusion of simple distortion-aware statistical video features, along with well-defined visual impairment features, is able to deliver state-of-the-art, robust performance at a very reasonable computational cost. The promising performances of baseline CNN models suggest the great potential of leveraging transfer learning techniques for the UGC-VQA problem. We believe that this benchmarking study will help facilitate UGC-VQA research by clarifying the current status of BVQA research and the relative efficacies of modern BVQA models. To promote reproducible research and public usage, an implementation of VIDEVAL has been made available online: \textbf{\url{https://github.com/vztu/VIDEVAL}}. In addition to the software, we are also maintaining an ongoing performance leaderboard on Github: \textbf{\url{https://github.com/vztu/BVQA_Benchmark}}.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
| 8e79f98a492be9949f16670481fa2fbd75ee10f5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $\mathcal{A}=\{H_1,\ldots,H_s\}$ be a central hyperplane arrangement in $V=\mathbb{K}^n,$ where $\mathbb{K}$ is a field of characteristic zero. If $x_1,\ldots,x_n$ is a basis for $V^*$, then $H_i=V(l_i)$, $i=1,\ldots,s,$ for some linear forms $l_i$ in $S:=\mathrm{Sym}(V^*)=\mathbb{K}[x_1,\ldots,x_n].$
The module of the derivations of $S$ has the structure $\mathrm{Der}(S)=\oplus_{i=1}^{n}S\frac{\partial}{\partial{x_i}}$, that is, it is free of rank $n$. A \emph{logarithmic derivation} of $\mathcal{A}$ is an element $\theta \in \mathrm{Der}(S)$, such that $\theta(l_i) \in \langle l_i \rangle$, for all $i= 1,\ldots, s.$ The set of the logarithmic derivations forms an $S$-module, usually detonated by $\mathrm{Derlog}(\mathcal{A})$; the standard grading of $S$ induces a graded module structure on this $S$-module. Therefore, by the structure of the module of the derivations we can write $\theta = \sum_{i=1}^{n}P_i\frac{\partial}{\partial{x_i}}$ with $P_i$ homogeneous polynomials of the same degree, so we can define $\mathrm{deg}(\theta)=\mathrm{deg}(P_i)$.
Consider $F:=\Pi_{i=1}^{s}l_i$ the defining polynomial of $\mathcal{A}$. Since $F$ is a homogeneous polynomial the structure of module of the logarithmic derivations is well known, $$\mathrm{Derlog}(\mathcal{A})= \mathrm{Syz}(J_F)\oplus S\theta_E,$$ where $\mathrm{Syz}(J_F)$ is the first module of the syzygies on the Jacobian ideal of $F$, that is, the ideal of $S$ generated by the partial derivatives of the $F$ and $\theta_E=\sum_{i=1}^{n}x_i\frac{\partial}{\partial{x_i}}$ is the Euler derivation. In general, when $\mathrm{Syz}(J_F)$ is free, $V(F)$ is called a \emph{free divisor} (and therefore the hyperplane arrangement $\mathcal A$ is called {\em free}).
Maybe one of the most important conjecture in the field of hyperplane arrangement is Terao's Conjecture which states that if two hyperplane arrangements have isomorphic intersection lattices, then if one is free, then so is the other. So a deeper understanding of the generators of $\mathrm{Syz}(J_F)$ is of utmost importance to tackle this conjecture: one would like to describe these generators, or their degrees, only from the combinatorics of $\mathcal A$, if possible. In the past 5-6 years, there has been a lot of work on the {\em minimal degree of a Jacobian relation}, which is the minimal degree of a syzygy of $J_F$ (and here $V(F)$ can be any divisor, not necessarily a hyperplane arrangement):
$$r(\mathcal A)=mdr(\mathcal A):= \min_{r\in\mathbb Z}\{r|(\mathrm{Syz}(J_F))_r\neq 0\}.$$
Finding degrees of syzygies (and more generally, the shapes of the graded minimal free resolutions) has been one of the most fundamental topics in commutative/homological algebra; and the literature about this is extensive, being impossible to list it all here without leaving out some of it. Same is true also if one is interested in syzygies of Jacobian ideals of divisors (and therefore analyzing their singular locus), with a particular focus on plane projective divisors. Specifically to line arrangements in $\mathbb P^2$ (i.e., central rank 3 hyperplane arrangements in $\mathbb K^3$), we must mention the work of Dimca, Dimca-Sticlaru, and their coauthors: most of the time starting from results in \cite{duWa} where the invariant {\em mdr} seems to show up for the first time in the context of our project, they find connections between this invariant and the Tjurina number of a projective plane divisor (i.e., the degree of the Jacobian ideal), which in turn give restrictions on constructing line arrangements with prescribed combinatorics: \cite{TaDiSti,Di, DiSti, DiSti2}, and the citations therein.
Our goal is the following: in the spirit of \cite{To}, we will classify all line arrangements in $\mathbb P^2$ with $r(\mathcal A)=3$, also analyzing the shape of the corresponding cubic logarithmic derivation towards finding some structures in this shape. For the theoretical results we found essential the addition-deletion results for the minimal degree of logarithmic derivations of arrangements in \cite{TaDiSti}. At the time of our work, this article was in ``preprint'' stage, yet the results that we are citing are correct: they are a structured and clear presentation of the underlying aspects of the results in \cite[Chapter 4.3]{OrTe} that concern the minimal degree of a Jacobian relation. As an appendix, at the end we briefly look at this invariant for graphic arrangements.
Below we summarize the results in \cite{TaDiSti}; for more definitions and notations clarifications, we refer the reader to \cite{OrTe}. Let $\mathcal A$ be a finite collection of (linear) hyperplanes in $\mathbb K^{\ell}$ with $\cap_{H\in\mathcal A}H=\{0\}$ (i.e., we say that $\mathcal A$ is central essential hyperplane arrangement of rank $\ell$). For $H\in\mathcal A$, let $$\mathcal A''=\mathcal A^{H}:=\{H\cap L|L\in\mathcal A\setminus\{H\}\}$$ be the {\em restriction}, and let $$\mathcal A':=\mathcal A\setminus\{H\}$$ be the {\em deletion}. Also, denote
$$r:=r(\mathcal A),\, r':=r(\mathcal A'), \, r'':=r(\mathcal A'').$$
\begin{thm} \label{additiondeletion} Let $\ell\geq 2$. With the above notations we have:
\begin{itemize}
\item[(1)] \cite[Proposition 2.12]{TaDiSti}:
\begin{itemize}
\item[(1a)] $r'\leq r\leq r'+1$.
\item[(1b)] If $|\mathcal A|-|\mathcal A''|>r$, then $r'=r$.
\item[(1c)] If $|\mathcal A'|-|\mathcal A''|>r'$, then $r=r'$.
\end{itemize}
\item[(2)] \cite[Theorem 2.14]{TaDiSti}: If $r'<r''$, then $r=r'+1$.
\item[(3)] \cite[Proposition 2.15]{TaDiSti}: If $r=r'$, then $r''\leq r$.
\end{itemize}
\end{thm}
Observe that the statement (3) is the counterpositive of (2), combined with (1a), so basically they are the same. But the flavors are different: (2) is the ``addition-deletion'' part, meaning that if we have information about $r'$ and $r''$, then we will obtain some information about $r$; whereas (3) is the ``restriction'' part, meaning that if we have information about $r$ and $r'$, then we will obtain information about $r''$.
\begin{rem}\label{upperbound} Let $\mathcal A\subset\mathbb P^{\ell-1}$ be a hyperplane arrangement of rank $\ell$. Let $X$ be a coatom, meaning that $X$ is a flat of rank $\ell-1$ in the lattice of intersection of $\mathcal A$. The {\em closure of $X$} is $cl(X):=\{H\in\mathcal A| X\in H\}$, i.e., the set of hyperplanes of $\mathcal A$ that contain $X$. The {\em multiplicity of $X$} is $\nu(X):=|cl(X)|$, and define $M=M(\mathcal A):=\max\{\nu(X)|X \mbox{ coatom}\}$.
Suppose $|\mathcal A|=s$ and suppose $M\leq s-2$. Then, using the same (classical) trick as in the proof of \cite[Theorem 1.2]{Di}, we have $$r(\mathcal A)\leq s-M.$$
Here is how: let $X$ be a coatom with $\nu(X)=M$. We can suppose $X=[0,\ldots,0,1]$. Then, the defining polynomial of $\mathcal A$ is $F=gh$, where $g,h\in S:=\mathbb K[x_1,\ldots,x_{\ell}]$ are products of the defining linear forms of the hyperplanes of $\mathcal A$, with $g(X)=0$ and $h(X)\neq 0$; and so $\deg(g)=M$ and $\deg(h)=s-M$.
For any homogeneous polynomial $P\in S$, denote $\displaystyle P_{x_i}:=\frac{\partial P}{\partial x_i}$. Then, $F_{x_{\ell}}=g\cdot h_{x_{\ell}}$. Together with Euler's formula $sF=x_1F_{x_{1}}+\cdots+x_{\ell-1}F_{x_{\ell-1}}+x_{\ell}F_{x_{\ell}}$, we get the Jacobian relation
$$(x_1h_{x_{\ell}})F_{x_{1}}+\cdots+(x_{\ell-1}h_{x_{\ell}})F_{x_{\ell-1}}+(x_{\ell}h_{x_{\ell}}-sh)F_{x_{\ell}}=0.$$ The only issue that can occur here is that the logarithmic derivation corresponding to this syzygy may be a multiple of Euler's derivation. If that were the case, since $\deg(h)=s-M\geq 2$, then there should exist a linear factor dividing both $h_{x_{\ell}}$ and $h$; an obvious contradiction.
\end{rem}
\begin{rem}\label{lowerbound} Same as above, let $\mathcal A\subset\mathbb P^{\ell-1}$ be a hyperplane arrangement of rank $\ell$. Define $0-Sing(\mathcal A):=\{X|X\mbox{ coatom}\}$ to be the set of coatoms of $\mathcal A$. This is a (reduced) set of points in $\mathbb P^{\ell-1}$. We define $\alpha_0(\mathcal A)$ to be the minimal degree of a hypersurface containing $0-Sing(\mathcal A)$. Then $$\alpha_0(\mathcal A)-1\leq r(\mathcal A).$$
To see this, we use the same trick as in \cite[Proposition 3.6]{To0}. Let $\displaystyle\theta=P_1\frac{\partial}{\partial x_1}+\cdots+P_{\ell} \frac{\partial}{\partial x_{\ell}}$ be logarithmic derivation of degree $r(\mathcal A)$, that is not a multiple of the Euler derivation.
Let $X$ be a coatom, and suppose $X=H_1\cap\cdots\cap H_{\ell-1}$, for some $H_i\in\mathcal A$. So $X=[a_1,\ldots,a_\ell]$, and if $H_i=V(l_i),i=1,\ldots,\ell-1$, then $l_i(X)=0$, for all $i$. Suppose $l_i=c_i^1x_1+\cdots+c_i^{\ell}x_{\ell}, c_i^j\in \mathbb K$. Then, for each $i=1,\ldots,\ell-1$, $\theta(l_i)\in\langle l_i\rangle$, which evaluated at $X$ gives:
$$c_i^1P_1(X)+\cdots+c_i^{\ell}P_{\ell}(X)=0.$$
Either $P_j(X)=0$, for all $j=1,\ldots,\ell$, or the point $[P_1(X),\ldots,P_{\ell}(X)]\in V(l_1,\ldots,l_{\ell-1})$. In the later case, we must have that $[P_1(X),\ldots,P_{\ell}(X)]=X$ as projective points, hence $$P_j(X)=qa_j, j=1,\ldots,\ell, q\in\mathbb K\setminus\{0\}.$$ This means that for any $1\leq j<k\leq \ell$, $(x_kP_j-x_jP_k)(X)=0$. Since $x_kP_j-x_jP_k$ cannot be the zero polynomial for all $j<k$ (since $\theta$ is not a multiple of $\theta_E$), put together the two cases we have that all coatoms $X$ belong to a hypersurface of degree $\leq r(\mathcal A)+1$.
\end{rem}
\section{Line arrangements with cubic minimal log derivations}
Let us put the general concepts in our perspective. Let $\mathcal{A}=\{V(l_1),\ldots,V(l_s)\}\subset \mathbb{P}^2$ be a line arrangement of rank 3. So $V=\mathbb K^3$, and if $x,y,z$ is a basis for $V^*$, $l_1,\ldots,l_s$ are linear forms in $S:={\rm Sym}(V^*)=\mathbb K[x,y,z]$, with three of them being linearly independent. Also let $F=l_1\cdots l_s$ be the defining polynomial of $\mathcal A$.
{\em The singular locus} of $\mathcal A$, denoted $Sing(\mathcal A)$, is the set of the intersection points of lines of $\mathcal A$. For $P\in Sing(\mathcal A)$, {\em the multiplicity} of $P$, denoted $m_P$, is the number of lines of $\mathcal A$ intersecting at $P$. Combinatorially, $Sing(\mathcal A)$ is the set of rank 2 flats in the intersection lattice of $\mathcal A$, and $m_P=\nu(P)=\mu(P)+1$, where $\mu(P)$ is the value of the M\"{o}bius function at $P$. Also, denote $$m=m(\mathcal A):=\max\{m_P|P\in Sing(\mathcal A)\}.$$ There are two very important formulas (with immediate inductive proofs on $s\geq 2$):
$$(\star) \hspace{2cm} \mbox{If } H\in \mathcal A,\mbox{ then }\sum_{P\in Sing(\mathcal A)\cap H}(m_P-1)=s-1,$$ and $$(\star\star) \hspace{2.5cm}\sum_{P\in Sing(\mathcal A)}{{m_P}\choose{2}}={{s}\choose{2}}.\hspace{2.7cm}$$
For $H\in\mathcal A$, we will denote $$|H|:=|\mathcal A''|=|Sing(\mathcal A)\cap H|,$$ the number of singularities of $\mathcal A$ lying on the line $H$. With this in mind we have the following crucial result:
\begin{thm} \label{lines} Let $\mathcal A$ be a line arrangement in $\mathbb P^2$ of rank 3, and let $H\in\mathcal A$. Then, using the notations above, we have:
\begin{itemize}
\item[(i)] \cite[Theorem 3.3]{TaDiSti}: If $|H|\geq r'+2$, then $r=r'+1$.
\item[(ii)] \cite[Theorem 3.4]{TaDiSti}: If $|H|\geq r+2$, then $r'=r-1$.
\end{itemize}
\end{thm}
The above theorem is a corollary to Theorem \ref{additiondeletion}: its proof uses the fact that $r''=|H|-1$. Based on the same fact we have
\begin{cor}\label{cor_lower} Let $\mathcal A\subset \mathbb P^2$ be a line arrangement of rank 3, with $|\mathcal A|=s$. Let $r:=r(\mathcal A)$ and $m:=m(\mathcal A)$. Then
$$r=s-m \mbox{ or } r\geq \min\{|\mathcal A^H|\, |\, H\in\mathcal A\}-1.$$
\end{cor}
\begin{proof} Denote $t:=\min\{|\mathcal A^H|\, |\, H\in\mathcal A\}$.
We prove the result by induction on $s\geq 3$.
If $s=3$, then $t=2$, $m=2$, and $r=1$, so the claim is true.
Suppose $s\geq 4$. From Remark \ref{upperbound}, we have $r\leq s-m$. Suppose $r\leq s-m-1$. We want to show $r\geq t-1$.
Let $H\in \mathcal A$ with $|\mathcal A^H|=t$. As before, $r'=r(\mathcal A\setminus\{H\})$, and $r''=r(\mathcal A^H)$; so $r''=t-1$. If rank of $\mathcal A'$ is 2, then $t=2$, and obviously $r\geq 2-1=1$.
Suppose rank of $\mathcal A'$ is 3, and suppose to the contrary that $r\leq t-2=r''-1$. Since $r'\leq r\leq r''-1$, by Theorem \ref{additiondeletion} (2), we get $r=r'+1$.
By induction, if $m':=m(\mathcal A')$, we have $r'=(s-1)-m'$, or $r'\geq |(\mathcal A')^L|-1$, for all $L\in\mathcal A'$.
If $r'=(s-1)-m'$, then $m'=m$ (the other option is $m'=m-1$, which contradicts $r'\leq r\leq s-m-1$). So $r'=r=s-m-1$; also a contradiction. So $r'\leq s-m-2$, and $r'\geq |(\mathcal A')^L|-1$, for all $L\in\mathcal A'$. So $r\geq |(\mathcal A')^L|$, for all $L\in\mathcal A'$.
If $H$ and $L$ intersect in a simple (i.e., double) point, then $|(\mathcal A')^L|=|\mathcal A^L|-1$; otherwise $|(\mathcal A')^L|=|\mathcal A^L|$. So, none the less we obtain $r\geq |(\mathcal A)^L|-1\geq t-1$. A contradiction.
\end{proof}
From this corollary we can conclude that $$\min\{|H|\, |\, H\in\mathcal A\}-1\leq r\leq s-m.$$ In the perspective of Remark \ref{lowerbound}, we ask what is the relation between $\alpha_0(\mathcal A)$ and $\min\{|H|\,|\, H\in\mathcal A\}$, for an arbitrary line arrangement $\mathcal A\subset \mathbb{P}^2$ of rank 3.
\medskip
\begin{rem} \label{recursion_remark} Suppose we want to classify all line arrangements $\mathcal A\subset\mathbb P^2$ with $r(\mathcal A)=d$ for some $d\geq 2$. Let $H\in\mathcal A$, and let $\mathcal A'=\mathcal A\setminus\{H\}$. By Theorem \ref{additiondeletion} (1a), $r'=d-1$, or $r'=d$. Then, by Theorem \ref{lines} (i) and (ii), we have either:
\begin{itemize}
\item[(a)] $\mathcal A$ is obtained recursively from the classification of all line arrangements $\mathcal B\subset\mathbb P^2$ with $r(\mathcal B)=d-1$ by adding a line $H$ that intersects all other lines of $\mathcal B$ in $\geq (d-1)+2=d+1$ points, or
\item[(b)] every line of $\mathcal A$ has $\leq d+1$ points on it.
\end{itemize}
\end{rem}
\medskip
Case (b) imposes some restrictions on the possible types of line arrangements. Let $P\in Sing(\mathcal A)$, with $m_P=m=m(\mathcal A)$. Since the rank of $\mathcal A$ is 3, there must exist a line $H\in\mathcal A$ that doesn't pass through $P$. So $$m\leq |H|\leq d+1,$$ and this is true for any line that doesn't pass through $P$.
If furthermore, $m=d+1$, then $\mathcal A$ is a supersolvable arrangement with modular point $P$. Hence it is free with exponents ${\rm exp}(\mathcal A)=(1,d, s-(d+1))$, where $|\mathcal A|=s$.
\medskip
\begin{exm} \label{example1} If $\mathcal A\subset\mathbb P^2$ is a line arrangement of rank 3, then Remarks \ref{upperbound} and \ref{lowerbound} give $$\alpha_0(\mathcal A)-1\leq r(\mathcal A)\leq |\mathcal A|-m(\mathcal A).$$ Let us look at Ziegler (\cite{Zi}) - Yuzvinsky (\cite{Yu}) example:
$$\mathcal A_1 = V(xyz(x+y+z)(2x+y+z)(2x+3y+z)(2x+3y+4z)(3x+5z)(3x+4y+5z))$$
$$\mathcal A_2 = V(xyz(x+y+z)(2x+y+z)(2x+3y+z)(2x+3y+4z)(x+3z)(x+2y+3z)).$$ These two arrangements have the same combinatorics, yet $r_1:=r(\mathcal A_1)=6$, and $r_2:=r(\mathcal A_2)=5$. Also for both arrangements $|\mathcal A_i|-m(\mathcal A_i)=9-3=6$, and $\alpha_0(\mathcal A_i)-1=6-1=5$. So $r_1$ achieves the upper bound, and $r_2$ achieves the lower bound.
Also, in both examples $|H|=6\leq r_i+1$, for any $H\in\mathcal A_i$. Therefore both options in the statement of Corollary \ref{cor_lower} are satisfied.
\end{exm}
\subsection{The case of $r=2$.} Remark \ref{recursion_remark} can shortcut quite a bit the calculations done in \cite{To} in order to obtain the classification of rank 3 line arrangements with a quadratic minimal logarithmic derivation. In the above discussion, $d=2$.
For case (a), either by using \cite[Proposition 4.29]{OrTe}, or \cite[Section 2.1]{To}, the line arrangement $\mathcal B$ with $r(\mathcal B)=d-1=1$ is a pencil of $\geq d+1=3$ lines, and an extra line not part of the pencil; or a triangle of 3 lines. There are only two different ways one can add another line to $\mathcal B$ to obtain $\mathcal A$, with at least 3 points on it, and these give \cite[Theorem 2.4 parts (1) and (2)]{To}.
For case (b), $m$ can be only 2, or 3. If $m=2$, then $\mathcal A$ is the generic line arrangement of $s$ lines, and since in this case $Sing(\mathcal A)$ is a star configuration, its defining ideal is generated in degree $s-1$ (see \cite{GeHaMi}). From Remark \ref{lowerbound} we require $s-1\leq d+1=3$, then $s\leq 4$, leading to $s=4$ which is a special case of \cite[Theorem 2.4 (2)]{To}. If $m=3$, then $\mathcal A$ is a supersolvable line arrangement with exponents $(1,2,s-3)$. At the same time, by \cite[Theorem 3.2 (1)]{DiSti}, we have $2\geq (s-2)/2$, leading to $s\leq 6$. Therefore $\mathcal A$ is the supersolvable arrangement with exponents $(1,2,3)$ which is exactly \cite[Theorem 2.4 (3)]{To}, or $(1,2,2)$ which is a special case of \cite[Theorem 2.4 (1)]{To}.
In summary this is the main result of \cite{To}; parts (1), (2), (3), correspond to (I), (II), and respectively, (III).
\begin{thm}[\cite{To} Theorem 2.4]\label{classification}
If $\mathcal{A}$ has a minimal quadratic syzygy on its Jacobian ideal, but not a linear syzygy, then, up to a change of coordinates, $\mathcal{A}$ is one of the following three types of arrangements with defining polynomials:
\begin{itemize}
\item [(I)] $F=xyz(x+y)\Pi_{j=4}^{s}(t_jy+z),~t_j\neq 0.$
\item [(II)] $F=xyz(x+y+z)\Pi_{j=4}^{s}(t_jy+z),~t_j\neq 0,1.$
\item [(III)] $F=xyz(x+y+z)(x+z)(y+z).$
\end{itemize}
See their affine pictures below:
\medskip
$$\stackrel{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (3, 0);
\draw[thick] (-0.75, -3/8) -- (3, 1.5);
\draw[thick] (-0.75/2, -0.75) -- (1.5, 3);
\draw[thick] (0.3, -0.75) -- (345/275, 3);
\draw[thick] (0.75, -0.75) -- (12/11, 3);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 3);
\draw[thick] (2.4, -0.75) -- (135/275, 3);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}{\stackrel{}{(\mathrm{I})}}
\hspace{1cm}\stackrel{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (3, 0);
\draw[thick] (-0.75, -3/8) -- (3, 1.5);
\draw[thick] (0.3, -0.75) -- (345/275, 3);
\draw[thick] (0.75, -0.75) -- (12/11, 3);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 3);
\draw[thick] (2.4, -0.75) -- (135/275, 3);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}{\stackrel{}{(\mathrm{II})}}
\hspace{1cm}\stackrel{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (3, 0);
\draw[thick] (-0.75/2, -0.75) -- (1.5, 3);
\draw[line width=1pt] (1, -0.75) -- (1, 3);
\draw[thick] (2.4, -0.75) -- (135/275, 3);
\draw[thick] (-5/6, -0.5) -- (2.5, 1.5);
\draw[thick] (43/15, -0.5) -- (-29/55, 1.5);
\end{tikzpicture}}{\stackrel{}{(\mathrm{III})}}$$
\end{thm}
\medskip
\subsection{The case of $r=3$.} As in the previous situation we apply Remark \ref{recursion_remark}, but when $d=3$. Then the classification is the following:
\begin{thm} \label{cubic} Let $\mathcal A\subset \mathbb P^2$ be a line arrangement of rank 3, with $r(\mathcal A)=3$. Then, up to a change of coordinates, $\mathcal A$ has one of the following defining polynomials (and corresponding (affine) pictures):
\begin{itemize}
\item[(Ia)] $ F=xyz(x+y)\textcolor{blue}{(bx+y)} \prod_{j=6}^s(t_jy+z), t_j\neq 0, s\geq 7.$
\item[(Ib)] $ F=xyz(x+y)\textcolor{blue}{(bx+z)}\prod_{j=6}^s(t_jy+z), t_j\neq 0, s\geq 6.$
\item[(Ic)] $ F=xyz(x+y)(y+z)\textcolor{blue}{(-x+z)}\prod_{j=7}^s(t_jy+z), t_j\neq 0,1, s\geq 7.$
\item[(Id)] $ F=xyz(x+y)\textcolor{blue}{(ax+by+z)}\prod_{j=6}^s(t_jy+z), t_j\neq 0, s\geq 5.$
\item[(IIa)]$ F=xyz(x+y+z)\textcolor{blue}{(ax+by+z)}\prod_{j=6}^s(t_jy+z), t_j\neq 0,~s\geq 5.$
\item[(IIb)]$ F=xyz(x+y+z)\textcolor{blue}{(bx+y)}\prod_{j=6}^s(t_jy+z), t_j\neq 0, s\geq 6.$
\item[(IIIa)]$ F=xyz(x+z)(y+z)(x+y+z)\textcolor{blue}{(ax+by+z)}.$
\item[(IIIb)]$ F=xyz(x+z)(y+z)(x+y+z)\textcolor{blue}{(ax+ay+z)}.$
\item[(IIIc)]$ F=xyz(x+z)(y+z)(x+y+z)\textcolor{blue}{(x+y+2z)}.$
\item[(IV)]$ F=xy(x-z)(y-z)(x-2z)(y-2z)(x-y).$
\item[(Va)]$ F=xyz(x^2-y^2)(x^2-z^2)(y^2-z^2).$
\item[(Vb)]$ F=xyz(x^2-z^2)(y^2-z^2)(x-y).$
\end{itemize}
The parameters $a$ and $b$ are any elements of $\mathbb K$ chosen such that we do not repeat any of the lines already selected, nor we obtain undesirable concurrencies; see the corresponding pictures.
\begin{center}
\begin{tiny}
\begin{tabular}{|c|c|c|}
\hline $\stackrel{\displaystyle{\Large(\mathrm{Ia})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.75, -0.55) -- (3, 1.1);
\draw[thick] (-0.75/2, -0.75) -- (1.5, 1.7);
\draw[thick] (0.2, -0.75) -- (4.4375/3.4375, 1.7);
\draw[thick] (0.5, -0.75) -- (13/11, 1.7);
\draw[thick] (0.9, -0.75) -- (57/55, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-0.6, -0.61) -- (2.5, 1.4);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{Ib})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.85, -0.6) -- (3, 1.1);
\draw[thick] (-0.75/2, -0.75) -- (1.5, 1.7);
\draw[thick] (0.2, -0.75) -- (4.4375/3.4375, 1.7);
\draw[thick] (0.5, -0.75) -- (13/11, 1.7);
\draw[thick] (0.9, -0.75) -- (57/55, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-1, -0.56) -- (3,0.44);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{Ic})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.35, -0.45) -- (2.5, 0.8);
\draw[thick] (-211/1700 , -0.75) -- (6586/4675, 1.7);
\draw[thick] (0.2, -0.75) -- (4.4375/3.4375, 1.7);
\draw[thick] (0.4, -0.75) -- (67/55, 1.7);
\draw[thick] (0.9, -0.75) -- (57/55, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-1/3, -0.7) -- (104/45,1);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ \\
\hline $\stackrel{\displaystyle{\Large(\mathrm{Id})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.75, -0.55) -- (3, 1.1);
\draw[thick] (-0.75/2, -0.75) -- (1.5, 1.7);
\draw[thick] (0.5, -0.75) -- (13/11, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-0.9, -0.48) -- (3,0.07);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{IIa})}\hspace{3.2cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.75, -0.55) -- (3, 1.1);
\draw[thick] (0.3, -0.75) -- (345/275, 1.7);
\draw[thick] (0.75, -0.75) -- (12/11, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-0.5, 0.15) -- (2.8, -0.55);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{IIb})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, -0.2) -- (3, -0.2);
\draw[thick] (-0.75, -0.55) -- (2.2, 1.1);
\draw[thick] (-0.1, -0.75) -- (7/5, 1.7);
\draw[thick] (0.3, -0.75) -- (345/275, 1.7);
\draw[thick] (0.75, -0.75) -- (12/11, 1.7);
\draw[thick] (1.8, -0.75) -- (24.375/34.375, 1.7);
\draw[thick] (2.4, -0.75) -- (135/275, 1.7);
\draw[blue, thick] (-1, -0.5) -- (2.5, 0.7);
\draw[thick,dotted] (1.1, -0.5) -- (1.35, -0.5);
\end{tikzpicture}}$ \\
\hline $\stackrel{\displaystyle{\Large(\mathrm{IIIa})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (4, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (145.299/186.190, 1.7) -- (356.811/186.190, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (497/4810, 1.5) -- (2045/962,-0.4);
\draw[thick,blue] (-3/13,1.5) -- (4,-0.185);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{IIIb})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (4, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (145.299/186.190, 1.7) -- (356.811/186.190, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (497/4810, 1.5) -- (2045/962,-0.4);
\draw[thick,blue] (-3/13,1.55) -- (3.2,-0.185);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large(\mathrm{IIIc})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (4, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (145.299/186.190, 1.7) -- (356.811/186.190, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (497/4810, 1.5) -- (2045/962,-0.4);
\draw[thick,blue] (-3/13,1.3) -- (4,-0.185);
\end{tikzpicture}}$ \\
\hline
$\stackrel{\displaystyle{\Large(\mathrm{IV})}\hspace{3.3cm}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (3.5, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (465/866, 1.7) -- (2565/866, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (-1/3, 1.1) -- (95/27, -0.2);
\draw[thick] (115.987/911.898, 1.7) -- (1484.200/455.949, -0.2);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large\hspace{0.5cm}(\mathrm{Va})}\hspace{1.8cm}\mathrm{projective}~\mathrm{picture}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (4, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (145.299/186.190, 1.7) -- (356.811/186.190, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (497/4810, 1.5) -- (2045/962,-0.4);
\draw[thick] (1093/2598, 1.6) -- (10093/2598, -0.2);
\draw[thick] (-9/26, 1) -- (4, -0.13);
\draw[thick] (-3/13,1.3) -- (4,-0.185);
\draw[thick] (3.25,1) -- (3.25,2);
\draw[thick] (3.5,1) -- (3.5,2);
\draw[thick] (3.75,1) -- (3.75,2);
\draw[thick] (3,1.75) -- (4,1.75);
\draw[thick] (3,1.5) -- (4,1.5);
\draw[thick] (3,1.25) -- (4,1.25);
\draw[thick] (3.1,1.1) -- (3.9,1.9);
\draw[thick] (3.1,1.9) -- (3.9,1.1);
\draw (3.5, 1.5) circle (.6);
\end{tikzpicture}}$ & $\stackrel{\displaystyle{\Large\hspace{0.5cm}(\mathrm{Vb})}\hspace{1.8cm}\mathrm{projective}~\mathrm{picture}}{
\begin{tikzpicture}
\draw[thick] (-1, 0) -- (4, 0);
\draw[thick] (-833/866,-0.4) -- (1267/866, 1.7);
\draw[thick] (145.299/186.190, 1.7) -- (356.811/186.190, -0.4);
\draw[line width=1pt] (1, 2) -- (1, -0.5);
\draw[thick] (-1,-13/60) -- (59/26, 1.2);
\draw[thick] (1093/2598, 1.6) -- (10093/2598, -0.2);
\draw[thick] (-9/26, 1) -- (4, -0.13);
\draw[thick] (-3/13,1.3) -- (4,-0.185);
\draw[thick] (3.25,1) -- (3.25,2);
\draw[thick] (3.5,1) -- (3.5,2);
\draw[thick] (3.75,1) -- (3.75,2);
\draw[thick] (3,1.75) -- (4,1.75);
\draw[thick] (3,1.5) -- (4,1.5);
\draw[thick] (3,1.25) -- (4,1.25);
\draw[thick] (3.1,1.1) -- (3.9,1.9);
\draw (3.5, 1.5) circle (.6);
\end{tikzpicture}}$ \\
\hline \end{tabular}
\end{tiny}
\end{center}
\end{thm}
\medskip
\begin{proof} The cases (Ia)-(IIIc) are obtained by adding a line (depicted here in blue) with at least $d+1=4$ intersection points to a line arrangement $\mathcal B$, with $r(\mathcal B)=2$ (so a part of the classification in Theorem \ref{classification}).
The arrangements not obtained recursively are treated in case (b) of Remark \ref{recursion_remark}.
\medskip
If $m=2,3$, then $\mathcal A$ is an arrangement with only double and triple points (i.e., for any $P\in Sing(\mathcal A)$, $m_P=2,3$), and therefore, by \cite[Theorem 3.2 (1)]{DiSti}, we have $$r(\mathcal A)=3\geq\frac{s-2}{2},$$ so $s\leq 8$. Also, every line has at most $d+1=4$ points on it. Suppose $T_H$ is the number of triple points on a line $H\in \mathcal A$, and $D_H$ is the number of double points on the same line $H$. So $$D_H+T_H\leq 4.$$ From formula $(\star)$ we also have:
$$D_H+2T_H=s-1.$$ Of course, if $s\geq 6$, then $T_H\geq 1$ for all $H\in\mathcal A$, because otherwise (i.e., $T_H=0$) $D_H\geq 5$ which is impossible to happen for line $H$.
$\bullet$ If $T_H=1$, then $D_H\leq 3$, so $s\leq 3+2\cdot 1+1=6$.
$\bullet$ If $T_H=2$, then $D_H\leq 2$, so $s\leq 2+2\cdot 2 +1=7$.
$\bullet$ If $T_H=3$, then $D_H\leq 1$, and so $s\leq 1+2\cdot 3 +1=8$.
$\bullet$ If $T_H\geq 4$, then $s\geq 9$, contradiction.
With all of this information, we can apply \cite[Theorem 4.18]{TaDiSti}. Below, $n_2$ is the number of double points, and $n_3$ is the number of triple points. From $(\star\star)$ we have $\displaystyle n_2+3n_3={{s}\choose{2}}$.
\begin{itemize}
\item[i.] $n_3=0$, meaning that $m=2$. Then $\mathcal A$ is a generic line arrangement of $s=3+2=5$ lines. This is the special case of (IIa) with $s=5$.
\item[ii.] $1\leq n_3\leq 3$. Then $s=3+3=6$. When $n_3=1$, we have the special case (IIa) with $s=6$; when $n_3=2$, we have the special cases (Id) and (IIb) with $s=6$; when $n_3=3$, we have the special case (Ib) with $s=6$.
\item[iii.] $n_3=4$. If $s=6$, then $n_2=15-12=3$. Since $D_H+2T_H=5$, we can have $T_H=2, D_H=1$, or $T_H=1, D_H=3=n_2$. This second case cannot happen, because otherwise, for all other lines $D_H=0$; contradiction. So $T_H=2$ and $D_H=1$ for all six lines $H$ of $\mathcal A$. This is the case of $\mathcal A(2,2,3):=V((x^2-y^2)(x^2-z^2)(y^2-z^2))$, which has $r=2$ (i.e., situation (III) in Theorem \ref{classification}). \\ Then \cite[Theorem 4.18 (3)]{TaDiSti} is saying that we must have $s=r+4=7$, and that $\mathcal A$ is obtained by adding a generic line $H$ to the line arrangement $\mathcal A(2,2,3)$ (i.e., situation (IIIa) in Theorem \ref{cubic}). But in this case $|H|=6$, a contradiction with our setup for case (b).
\item[iv.] $n_3=5$. \cite[Theorem 4.18 (4)(A)]{TaDiSti} implies that $T_L=1$ which by the first bullet above, leads to $s=6$. So $D_L=5-2=3$, yet $n_2=15-15=0$. Contradiction. So $s=7$, and $\mathcal A$ is obtained by adding the line $L$ to one of the double points of $\mathcal A(2,2,3)$ (i.e., situation (IIIb) in Theorem \ref{cubic}). But this contradicts our case (b): $|L|=5\nleq 4$. \\ So we are left with case (4)(B) of that theorem, with $s=7$: we are adding $7-7=0$ generic lines to the arrangement denoted there by $\mathcal B$; this is the same arrangement as $\mathcal A_4:=V(xy(x-z)(y-z)(x-2z)(y-2z)(x-y))$ in \cite[Section 4]{To1}. So a new type, denoted with (IV).
\item[v.] $n_3=6$. In this instance $s=7, 8$, and $T_H=2,3$ for all $H$. Through any triple point there must be a line $H$ with $T_H=2$ (otherwise we will have $n_3\geq 7$). But from the second bullet above, we must have $s=7$. Hence, $n_2=21-18=3$, and $D_H=2$. If $L$ is a line with $T_L=3$, then $D_H=0$. Because $n_2=3$, every triple point will have one line with $T=2$ and two lines with $T=3$ through it. Therefore $\mathcal A$ is the non-Fano arrangement (which is free with exponents $(1,3,3)$); this is type (IIIc).
\item[vi.] $n_3=7$. Then every line through a triple point will have $T\leq 3$ (see the forth bullet above). If $s=7$, then $n_2=0$, and such an arrangement is the Fano plane, which is realizable only over a field of characteristic 2. \\ So $s=8$, and from the third bullet above, for every $H\in \mathcal A$, we must have $D_H=1$, and $T_H=3$. Also, $n_2=28-21=7$. But this is impossible, since every double point belongs to exactly two lines, and every line has exactly one double point, hence $n_2=8/2=4$.
\item[vii.] $n_3\geq 8$. Then, $s=8$, $n_2=28-24=4$ (here $n_3=8$), or $n_2=28-27=1$ (here $n_3=9$). Also, from the third bullet $T_H=3$, and $D_H=1$ for all $H\in\mathcal A$. Therefore $n_3=9$ cannot happen; just look at two lines passing through the same triple point. \\ So $n_3=8$. But there is only one such arrangement, namely the deleted Hasse arrangement: $\mathcal A=V((x^2+xy+y^2)(x^3-z^3)(y^3-z^3))$. Computations with Macaulay 2 (\cite{EiGrSt}) show that $r(\mathcal A)=4$, so contradiction.
\end{itemize}
\bigskip
Now suppose $m=3+1=4$. So $\mathcal A$ is supersolvable with exponents $(1,3,s-4)$. If $s-4=3$, then $s=7$, and we can have two special cases of types (Ia) and (Ic) in Theorem \ref{cubic}.
Suppose $s\geq 8$. By \cite[Theorem 1.1]{AbDi}, we have $s\leq 3m-3=9$.
From \cite[Corollary 2]{HaHa}, if there is another modular point of multiplicity $m'<m=4$, then $s=m+m'-1\leq 6$. So in order to proceed with our situation, we must consider the case of {\em $m-$homogeneous} supersolvable line arrangements (i.e., all modular points have the same multiplicity $m$). As it is worked out in \cite[Section 3]{HaHa}, and very well summarized in \cite[Theorem 1.2]{AbDi}, any 4-homogeneous supersolvable arrangement, with $n_4\geq 3$ is, after a change of variables, $\mathcal A(2,1,3):=V(xyz(x^2-y^2)(x^2-z^2)(y^2-z^2))$; this is type (Va) in our list.
If $n_4=2$, let $P,Q\in Sing(\mathcal A)$ be with $m_P=m_Q=m=4$. Let $H:=\overline{PQ}$. Then $D_H+2T_H+2\cdot (4-1)=s-1\leq 8$.
\begin{itemize}
\item[i.] If $s=8$, then $D_H=1$ and $T_H=0$. If $L$ is the line that intersects $H$ at the double point, then $|L|=4$, and therefore $T_L=3$ and $D_L=1$. So $\mathcal A$ has defining polynomial $$xyz(x^2-z^2)(y^2-z^2)(x-y),$$ which is type (Vb) in our list.
\item[ii.] If $s=9$, for real arrangements, as \cite[Section 3.2.3]{HaHa} shows, this is impossible. For complex arrangements, we proceed with the same ideas as above. We can have $D_H=2$ and $T_H=0$, or $D_H=0$ and $T_H=1$. In the first case, if $L$ is a line that intersects $H$ at one of the double points, then $|L|=D_L+T_L=4$, and $D_L+2T_L=9-1=8$. So $T_L=4$, giving that $D_L=0$, contradiction with the fact that $L$ has at least one double point. In the second case, any line $L\neq H$ through the one triple point $R$ has, by the same calculation we just did, $T_L=4$, and $D_L=0$. By \cite[Conjecture 3.3]{AnTo}, $\mathcal A$ is not realizable over the field of complex numbers.\footnote{The conjecture is shown for $s\leq 12$ in \cite[Corollary 3.5]{AnTo}, and to our overwhelming excitement, the conjecture has been proven in general by Takuro Abe in \cite{Ab}.} Therefore no other type is added to the list.
\end{itemize}
\end{proof}
\section{Computational approach}
Let $\mathcal A=\{V(l_1),\ldots,V(l_s)\}\subset \mathbb P^2$ be a rank 3 line arrangement. Suppose we fix the equations of the defining linear forms: $l_i=a_ix+b_iy+c_iz\in S:=\mathbb K[x,y,z]$ for each $i\in \{1,\ldots,s\}$. Let $\theta=P\frac{\partial}{\partial x}+Q\frac{\partial}{\partial y}+R\frac{\partial}{\partial z}$ be a degree $d$ logarithmic derivation of $\mathcal{A}$, that is, $P,Q,R \in S_d.$
For practical purpose we can suppose $l_1=x,~l_2=y,~l_3=z$, then we can rewrite $$\theta=P'x\frac{\partial}{\partial x}+Q'y\frac{\partial}{\partial y}+R'z\frac{\partial}{\partial z}~ \mathrm{with}~ P',Q',R' \in S_{d-1}.$$
Using the Euler derivation, we can write $\theta=P'\theta_E+(Q'-P')y\frac{\partial}{\partial y}+(R'-P')z\frac{\partial}{\partial z}.$ Therefore, we can suppose $P'=0$, and in the following sections we will work with a logarithmic derivation with the form $$\theta=Q'y\frac{\partial}{\partial y}+R'z\frac{\partial}{\partial z}~ \mathrm{with}~ Q',R' \in S_{d-1}.$$
\begin{rem}\label{divisonlines}
Since $l_1=x,~l_2=y,~l_3=z$, for each $i\geq 4$ we have two of $a_i,b_i,c_i$ being nonzero. Reordering, if necessary, we can group the lines $l_1,\ldots,l_s$ in four cases:
\begin{itemize}
\item[i)] $a_4,b_4,\ldots,a_{\mathfrak{j}},b_{\mathfrak{j}} \neq 0$ and $c_4,\ldots,c_{\mathfrak{j}} = 0;$
\item[ii)] $a_{\mathfrak{j}+1},c_{\mathfrak{j}+1},\ldots,a_{\mathfrak{l}},c_{\mathfrak{l}} \neq 0$ and $b_{\mathfrak{j}+1},\ldots,b_{\mathfrak{l}} = 0;$
\item[iii)] $b_{\mathfrak{l}+1},c_{\mathfrak{l}+1},\ldots,b_{\mathfrak{n}},c_{\mathfrak{n}} \neq 0$ and $a_{\mathfrak{l}+1},\ldots,a_{\mathfrak{n}} = 0;$
\item[iv)] $a_{\mathfrak{n}+1},b_{\mathfrak{n}+1},c_{\mathfrak{n}+1},\ldots,a_{s},b_{s},c_{s} \neq 0,$ with $4\leq \mathfrak{j}< \mathfrak{l}< \mathfrak{n}< s.$
\end{itemize}
For our purposes we also can suppose $b_j=1,$ for $j\in \{4,\ldots,\mathfrak{j} \}$ in case i), and $c_i=1$ for for cases ii), iii), and iv) (i.e., for $i\in\{\mathfrak{j}+1,\ldots,s\}$).
\end{rem}
The main goal is to analyze the shape of such a logarithmic derivation, towards obtaining conditions for a line arrangement to possess a minimal logarithmic derivation to degrees 2 and 3 through a matrix argument.
\subsection{Degree 2 logarithmic derivations}
First we review the beginning of \cite[Section 2.3]{To} while having a more organized approach. Let $\theta=Q'y\frac{\partial}{\partial y}+R'z\frac{\partial}{\partial z},~ \mathrm{with}~ Q',R' \in S_1,$ be a degree 2 logarithmic derivation of $\mathcal{A}$. For each $i\in \{1,\ldots,s\}$ we have $\theta(l_i)=l_iT_i$ for some $T_i\in S_1$. Let $Q'=q_1x+q_2y+q_3z,~R'=r_1x+r_2y+r_3z,~T_i=t^i_1x+t^i_2y+t^i_3z\in S_1.$ Therefore,
$$b_iyQ'+c_izR'=(a_ix+b_iy+c_iz)T_i.$$
Comparing coefficients, the following six relations are obtained
\begin{eqnarray}\label{EQdegree2}
\displaystyle \Big\{\begin{array}{ccc}
0=a_it^i_1\hspace{1.5cm} & \hspace{0.5cm}b_iq_2=b_it^i_2\hspace{1.2cm} &\hspace{0.5cm} c_ir_3=c_it^i_3\hspace{2.1cm}\\
b_iq_1=a_it^i_2+b_it^i_1& \hspace{0.5cm}c_ir_1=a_it^i_3+c_it^i_1&\hspace{0.5cm} b_iq_3+c_ir_2=b_it^i_3+c_it^i_2
\end{array}
\end{eqnarray}
\begin{rem}\label{comput-quadratic} If we projectively eliminate the parameters $t_1^i,t_2^i,t_3^i$ in (\ref{EQdegree2}), we obtain that $\mathcal A$ has a quadratic logarithmic derivation that is not a multiple of $\theta_E$ if and only if the points dual to the lines of $\mathcal A$ are contained in the variety $$V(xy(yq_1-xq_2), xz(zr_1-xr_3), yz[y(q_3-r_3)-z(q_2-r_2)]),$$ for some constants $q_i,r_j$, not all zero. In a snapshot, this is \cite[Corollary 2.2]{To}, which is used to prove, via some challenging computations, the classification of line arrangements with quadratic minimal logarithmic derivations.
As we mentioned in Section 2.0.1, in this project our approach has been simplified, and it goes on the opposite direction of \cite{To}: with the powerful tools provided by the addition-deletion results in \cite{TaDiSti}, we first find the classification, and as we will see below, we find the corresponding minimal quadratic logarithmic derivations (i.e., the constants $q_i,r_j$).
\end{rem}
\bigskip
Applying Remark \ref{divisonlines} to (\ref{EQdegree2}) and eliminating the parameters $t_1^i,t_2^i,t_3^i$, we have the following equations for $q_1,q_2,q_3,r_1,r_2,r_3$ and their associated matrices $\mathcal{N}_{\mathrm{i}},\mathcal{N}_{\mathrm{ii}},\mathcal{N}_{\mathrm{iii}},\mathcal{N}_{\mathrm{iv}}$ respectively.
$\bullet$ \underline{Situation i)}: for $4\leq j \leq \mathfrak{j}$ we have $q_3=0$ and $q_1-a_jq_2=0$ with matrix
$$\mathcal{N}_{\mathrm{i}}=\left[\begin{array}{cccccc}
0&0&1&0&0&0\\
1&-a_4&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
1&-a_j&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
1&-a_{\mathfrak{j}}&0&0&0&0\\
\end{array}\right].$$
$\bullet$ \underline{Situation ii)}: for $\mathfrak{j}+1\leq l \leq \mathfrak{l}$ we have $r_2=0$ and $r_1-a_lr_3=0$ with matrix
$$\mathcal{N}_{\mathrm{ii}}=\left[\begin{array}{cccccc}
0&0&0&0&1&0\\
0&0&0&1&0&-a_{\mathfrak{j}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&0&-a_l\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&0&-a_{\mathfrak{l}}\\
\end{array}\right].$$
$\bullet$ \underline{Situation iii)}: for $\mathfrak{l}+1\leq n \leq \mathfrak{n}$ we have $q_1-r_1=0$ and $(q_2-r_2)-b_n(q_3-r_3)=0$ with matrix
$$\mathcal{N}_{\mathrm{iii}}=\left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&-b_{\mathfrak{l}+1}&0&-1&b_{\mathfrak{l}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-b_n&0&-1&b_n\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-b_{\mathfrak{n}}&0&-1&b_{\mathfrak{n}}\\
\end{array}\right].$$
$\bullet$ \underline{Situation iv)}: for $\mathfrak{n}+1\leq i \leq s$ we have $b_iq_1-a_iq_2=0$, $r_1-a_ir_3=0$, and $(q_2-r_2)-b_i(q_3-r_3)=0$ with matrix $\mathcal{N}_{\mathrm{iv}}$ obtained by concatenating the following matrices:
\begin{tiny} $$\iota=\left[\begin{array}{cccccc}
b_{\mathfrak{n}+1}&-a_{\mathfrak{n}+1}&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
b_i&-a_i&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
b_s&-a_s&0&0&0&0\\
\end{array}\right]
\varsigma=\left[\begin{array}{cccccc}
0&0&0&1&0&-a_{\mathfrak{n}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&0&-a_i\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&0&-a_s\\
\end{array}\right]
\mathfrak{i}=\left[\begin{array}{cccccc}
0&-1&b_{\mathfrak{n}+1}&0&1&-b_{\mathfrak{n}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&-1&b_i&0&1&-b_i\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&-1&b_s&0&1&-b_s\\
\end{array}\right].$$
\end{tiny}
We can represent all the equations above by using the following null product of representative matrix:
$$\mathcal{M}_{(3s-2\mathfrak{n})\times 6}\cdot \left[\begin{array}{cccccc}
q_1&q_2&q_3&r_1&r_2&r_3
\end{array}\right]_{6\times 1}^{t}=0,$$
where the matrix $\mathcal{M}:=\mathcal{M}_{(3s-2\mathfrak{n})\times 6}$ is the concatenation of the matrices $\mathcal{N}_{\mathrm{i}},\mathcal{N}_{\mathrm{ii}},\mathcal{N}_{\mathrm{iii}},\mathcal{N}_{\mathrm{iv}}$. This linear system has non trivial solution if and only if $\mathrm{rank}(\mathcal{M})<\min\{3s-2\mathfrak{n},6\}.$
Using the classification of all line arrangements in $\mathbb{P}^2$ with $r(\mathcal{A})=2$ by (\ref{classification}) we can determine the shape of each degree 2 logarithmic derivation associated to each the defining polynomial.
\begin{cor} \label{quadratic-log}
The degree 2 logarithmic derivation associated to each the defining polynomial obtained in (\ref{classification}) is, respectively,
\begin{itemize}
\item [(I)] If $s\geq 5,$ then $\theta=(x+y)(y\frac{\partial}{\partial y}+z\frac{\partial}{\partial z}).$ \\
If $s=4$ then $\theta=\alpha(x+y)(y\frac{\partial}{\partial y}+z\frac{\partial}{\partial z})+\beta(t_4y+z)z\frac{\partial}{\partial z}$, for $\alpha,\beta \in \mathbb K$, not both equal to zero.
\item [(II)] $\theta=(x+y+z)(y\frac{\partial}{\partial y} + z\frac{\partial}{\partial z}).$
\item [(III)] $\theta=(x+y+2z)y\frac{\partial}{\partial y} + (x+z) z\frac{\partial}{\partial z}$.
\end{itemize}
\end{cor}
\begin{proof}
\begin{itemize}
\item [(I)] For this type the associated matrix $\mathcal{M}$ has the following shape:
$$\mathcal{M}_\mathrm{I}= \left[\begin{array}{cccccc}
0&0&1&0&0&0\\
1&-1&0&0&0&0\\
1&0&0&-1&0&0\\
0&1&-t_4&0&-1&t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-t_j&0&-1&t_j\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-t_s&0&-1&t_s\\
\end{array}\right]_{s\times 6}\Longleftrightarrow \left[\begin{array}{cccccc}
0&0&1&0&0&0\\
1&-1&0&0&0&0\\
1&0&0&-1&0&0\\
0&1&0&0&-1&t_4\\
0&0&0&0&0&t_5-t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&t_j-t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&t_s-t_4\\
\end{array}\right]_{s\times 6}.$$ This leads to the corresponding shape of the logarithmic derivation.
\pagebreak
For $s\geq 5$, since $t_j-t_4 \neq 0 ~\forall~j$, we have: \hspace{3cm}For $s=4$, we have:
$\mathcal{M}_\mathrm{I}= \left[\begin{array}{cccccc}
0&0&1&0&0&0\\
1&-1&0&0&0&0\\
1&0&0&-1&0&0\\
0&1&0&0&-1&0\\
0&0&0&0&0&1\\
0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0\\
\end{array}\right]_{s\times 6}.\hspace*{1.5cm}\mathcal{M}_\mathrm{I}= \left[\begin{array}{cccccc}
0&0&1&0&0&0\\
1&-1&0&0&0&0\\
1&0&0&-1&0&0\\
0&1&0&0&-1&t_4\\
\end{array}\right]_{5\times 6}.$ This leads to the corresponding shape of the logarithmic derivation.
\item [(II)] For this type the associated matrix $\mathcal{M}$ has the following shape:
$$\mathcal{M}_\mathrm{II}= \left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&-t_4&0&-1&t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-t_j&0&-1&t_j\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&-t_s&0&-1&t_s\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&-1&1&0&1&-1\\
\end{array}\right]_{(s+1)\times 6}\Longleftrightarrow \left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&-t_4&0&-1&t_4\\
0&0&t_4-t_5&0&0&t_5-t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&t_4-t_j&0&0&t_j-t_4\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&t_4-t_s&0&0&t_s-t_4\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&-1&1&0&1&-1\\
\end{array}\right]_{(s+1)\times 6}.$$
For $s\geq 5$, since $t_j-t_4 \neq 0$ for all $j$, we have $\mathcal{M}_\mathrm{II}= \left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&0&0&-1&0\\
0&0&-1&0&0&1\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0\\
\end{array}\right]_{(s+1)\times 6}.$
For $s=4$:\\ $\mathcal{M}_\mathrm{II}= \left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&-t_4&0&-1&t_4\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&-1&1&0&1&-1\\
\end{array}\right]_{5\times 6} \Longleftrightarrow \left[\begin{array}{cccccc}
1&0&0&-1&0&0\\
0&1&-t_4&0&-1&t_4\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&0&1-t_4&0&0&t_4-1\\
\end{array}\right]_{5\times 6}.$ Since $t_4\neq 1$, we obtain the corresponding shape of the logarithmic derivation.
\item [(III)] For this type the associated matrix $\mathcal{M}$ has the following shape:
$$\mathcal{M}_\mathrm{III}= \left[\begin{array}{cccccc}
0&0&0&0&1&0\\
0&0&0&1&0&-1\\
1&0&0&-1&0&0\\
0&1&-1&0&-1&1\\
1&-1&0&0&0&0\\
0&0&0&1&0&-1\\
0&-1&1&0&1&-1\\
\end{array}\right]_{7\times 6}\Longleftrightarrow\left[\begin{array}{cccccc}
0&0&0&0&1&0\\
0&0&0&1&0&-1\\
1&0&0&-1&0&0\\
0&1&-1&0&0&1\\
1&-1&0&0&0&0\\
0&0&0&0&0&0\\
0&0&0&0&0 &0\\
\end{array}\right]_{7\times 6}.$$ This leads to the corresponding shape of the logarithmic derivation.
\end{itemize}
\end{proof}
\subsection{Degree 3 logarithmic derivations}
In this section we reproduce the same argument as in the previous section, but for degree 3 logarithmic derivations. Let $\theta=Q'y\frac{\partial}{\partial y}+R'z\frac{\partial}{\partial z}~ \mathrm{with}~ Q',R' \in S_2$ be a degree 3 logarithmic derivation of $\mathcal{A}$. For each $i\in \{1,\ldots,s\}$ we have $\theta(l_i)=l_iT_i$ for some $T_i\in S_2$.
Let $Q'=q_1x^2+q_2y^2+q_3z^2+q_4xy+q_5xz+q_6yz,R'=r_1x^2+r_2y^2+r_3z^2+r_4xy+r_5xz+r_6yz,~T_i=t^i_1x^2+t^i_2y^2+t^i_3z^2+t^i_4xy+t^i_5xz+t^i_6yz \in k[x,y,z]_2.$ Therefore,
$$b_iyQ'+c_izR'=(a_ix+b_iy+c_iz)T_i.$$
Comparing coefficients, the following ten relations are obtained
\begin{eqnarray}\label{EQdegree3}
\displaystyle \Biggl\{\begin{array}{ccc}
&&\\
0=a_it^i_1 & \hspace{1.1cm}b_iq_4=a_it^i_2+b_it^i_4 & b_iq_3+c_ir_6=b_it^i_3+c_it^i_6 \\
b_iq_2=b_it^i_2& \hspace{1.1cm}c_ir_5=a_it^i_3+c_it^i_5 & b_iq_6+c_ir_2=b_it^i_6+c_it^i_2 \\
c_ir_3=c_it^i_3& \hspace{1.1cm}b_iq_1=a_it^i_4+b_it^i_1 & \hspace{1.1cm}b_iq_5+c_ir_4=a_it^i_6+b_it^i_5+c_it^i_4 \\
& \hspace{1.1cm}c_ir_1=a_it^i_5+c_it^i_1 &
\end{array}
\end{eqnarray}
If we projectively eliminate $t_1^i,t_2^i,t_3^i$ in (\ref{EQdegree3}), we obtain the following result which is similar to Remark \ref{comput-quadratic}.
\begin{lem}\label{dual-points} Let $\mathcal A\subset\mathbb P^2$ be a line arrangement with defining linear forms $l_1=x$, $l_2=y$, $l_3=z$, and $l_i=a_ix+b_iy+c_iz, i\geq 4$. Let $l_1^{\vee}:=[1,0,0], l_2^{\vee}:=[0,1,0], l_3^{\vee}:=[0,0,1], l_i^{\vee}:=[a_i,b_i,c_i], i\geq 4$ be the points dual to the lines of $\mathcal A$. Let $\mathcal A^{\vee}:=\{l_1^{\vee},\ldots,l_s^{\vee}\}$.
Then, $\mathcal A$ has a cubic logarithmic derivation that is not a multiple of the Euler derivation if and only if $\mathcal A^{\vee}$ is included in $V(F_1,F_2,F_3,F_4)$, where
\begin{eqnarray}
F_1&=& xy(q_2x^2-q_4xy+q_1y^2)\nonumber\\
F_2&=& xz(r_3x^2-r_5xz+r_1z^2)\nonumber\\
F_3&=& yz[(q_3-r_3)y^2-(q_6-r_6)yz+(q_2-r_2)z^2]\nonumber\\
F_4&=&xyz[(q_3-2r_3)xy^2-(q_2-r_6)xyz-(q_5-r_5)y^2z+(q_4-r_4)yz^2],\nonumber
\end{eqnarray} where $q_i,r_j\in\mathbb K$ are not all zero, and they will be the coefficients of $Q'$ and $R'$ shown above.
\end{lem}
\begin{exm} \label{example2} Let us consider the following two examples that are presented at the beginning of \cite[Section 4]{To1}.\footnote{They are examples denoted there with $\mathcal A_2$, and $\mathcal A_6$; in the later we performed a change of variables $x+y+z\leftrightarrow x, y\leftrightarrow y, z\leftrightarrow z$, so it will fit into the statement of Lemma \ref{dual-points}.}
$$\mathcal B_1:=V(xyz(x-z)(x-2z)(y-z)(y-2z))$$
$$\mathcal B_2:=V(xyz(x-2y-2z)(2x+2y-z)(2x-y+2z)(3x-12y-4z)).$$ $\mathcal B_1$ has two minimal cubic logarithmic derivations (actually it is free with exponents $(1,3,3)$), whereas $\mathcal B_2$ has six minimal degree five logarithmic derivations (generating the first syzygy module of the Jacobian ideal), so no cubic logarithmic derivation. But the minimal graded free resolutions of the ideals of $\mathcal B_1^{\vee}$ and of $\mathcal B_2^{\vee}$, respectively, are the same:
$$0\longrightarrow S^2(-5)\longrightarrow S(-2)\oplus S^2(-4).$$
In Example \ref{example1}, though there are no cubic logarithmic derivations other than multiples of $\theta_E$, it is worth mentioning that $I(\mathcal A_1^{\vee})$ and $I(\mathcal A_2^{\vee})$ are both ideal complete intersection of two cubic forms. To have the same minimal graded free resolutions is expected to happen, since the two arrangements have the ``same pictures'' (same combinatorics). Is it possible to find an example of two line arrangements having the same combinatorics, but different graded minimal free resolutions of the ideals of the points dual to the lines?
\end{exm}
\bigskip
Applying Remark \ref{divisonlines} to (\ref{EQdegree3}) and eliminating the parameters $t_1^i,t_2^i,t_3^i$, we have the following equations for $q_1,\ldots,q_6,r_1,\ldots,r_6$ and their associated matrices $\mathcal{N}_{\mathrm{i}},\mathcal{N}_{\mathrm{ii}},\mathcal{N}_{\mathrm{iii}},\mathcal{N}_{\mathrm{iv}}$ respectively.
$\bullet$ \underline{Situation i)}: for $4\leq j \leq \mathfrak{j}$ we have $q_3=0,~q_1+a_j^2q_2-a_jq_4=0$ and $q_5-a_jq_6=0$ with matrix
$$\mathcal{N}_{\mathrm{i}}=\left[\begin{array}{cccccccccccc}
0&0&1&0&0&0&0&0&0&0&0&0\\
1&a_4^2&0&-a_4&0&0&0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
1&a_j^2&0&-a_j&0&0&0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
1&a_{\mathfrak{j}}^2&0&-a_{\mathfrak{j}}&0&0&0&0&0&0&0&0\\
0&0&0&0&1&-a_4&0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&1&-a_j&0&0&0&0&0&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&1&-a_{\mathfrak{j}}&0&0&0&0&0&0
\end{array}\right].$$
$\bullet$ \underline{Situation ii)}: for $\mathfrak{j}+1\leq l \leq \mathfrak{l}$ we have $r_2=0,~r_1+a_l^2r_3-a_lr_5=0$ and $r_4-a_lr_6=0$ with matrix
$$\mathcal{N}_{\mathrm{ii}}=\left[\begin{array}{cccccccccccc}
0&0&0&0&0& 0&0&1&0&0&0&0\\
0&0&0&0&0&0&1& 0&a_{\mathfrak{j}+1}^2&0&-a_{\mathfrak{j}+1}&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0&1&0&a_l^2&0&-a_l&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0&1&0&a_{\mathfrak{l}}^2&0&-a_{\mathfrak{l}}&0\\
0&0&0&0&0&0&0&0&0&1&0&-a_{\mathfrak{j}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0&0&0&0&1&0&-a_l\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&0&0&0&0&0&0&1&0&-a_{\mathfrak{l}}\\
\end{array}\right].$$
$\bullet$ \underline{Situation iii)}: for $\mathfrak{l}+1\leq n \leq \mathfrak{n}$ we have $q_1-r_1=0,~q_2-r_2+b_n^2(q_3-r_3)-b_n(q_6-r_6)=0$ and $(q_4-r_4)-b_n(q_5-r_5)=0$ and with matrix
$$\mathcal{N}_{\mathrm{iii}}=\left[\begin{array}{cccccccccccc}
1&0&0&0&0&0&-1&0&0& 0&0&0\\
0&1&b_{\mathfrak{l}+1}^2&0&0&-b_{\mathfrak{l}+1}&0&-1&-b_{\mathfrak{l}+1}^2& 0&0&b_{\mathfrak{l}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&b_{n}^2&0&0&-b_{n}&0&-1&-b_{n}^2& 0&0&b_{n}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&b_{\mathfrak{n}}^2&0&0&-b_{\mathfrak{n}}&0&-1&-b_{\mathfrak{n}}^2& 0&0&b_{\mathfrak{n}}\\
0&0&0&1&-b_{\mathfrak{l}+1}&0&0&0&0&-1&b_{\mathfrak{l}+1}&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&-b_n&0&0&0&0&-1&b_n&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&0&0&1&-b_{\mathfrak{n}}&0&0&0&0&-1&b_{\mathfrak{n}}&0
\end{array}\right].$$
$\bullet$ \underline{Situation iv)}: for $\mathfrak{n}+1\leq i \leq s$ we have $b_i^2q_1+a_i^2q_2-a_ib_iq_4=0,~r_1+a_i^2r_3-a_ir_5=0,~q_2-b_i^2r_3-b_i(q_6-r_6)=0$ and $-a_iq_2-2a_ib_i^2r_3+a_ib_ir_6+b_i(q_4-r_4)-b_i^2(q_5-r_5)=0$ with matrix $\mathcal{N}_{\mathrm{iv}}$ obtained by concatenating the following matrices:
$$\mathcal{T}=\left[\begin{array}{cccccccccccc}
b_{\mathfrak{n}+1}^2&a_{\mathfrak{n}+1}^2&0&-a_{\mathfrak{n}+1}b_{\mathfrak{n}+1}&0&\cdots&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&&\vdots\\
b_i^2&a_i^2&0&-a_ib_i&0&\cdots&0\\
\vdots&\vdots&\vdots&\vdots&\vdots&&\vdots\\
b_{s}^2&a_{s}^2&0&-a_{s}b_{s}&0&\cdots&0\\
\end{array}\right],~\mathcal{U}=\left[\begin{array}{cccccccccccc}
0&\cdots&0&1& 0&a_{\mathfrak{n}+1}^2&0&-a_{\mathfrak{n}+1}&0\\
\vdots&&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&\cdots&0&1&0&a_i^2&0&-a_i&0\\
\vdots&&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&\cdots&0&1&0&a_{s}^2&0&-a_{s}&0\\
\end{array}\right],$$
$$\mathcal{V}=\left[\begin{array}{cccccccccccc}
0&1&0&0&0&-b_{\mathfrak{n}+1}&0&0&-b_{\mathfrak{n}+1}^2& 0&0&b_{\mathfrak{n}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&0&0&0&-b_{i}&0&0&-b_{i}^2& 0&0&b_{i}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&1&0&0&0&-b_{s}&0&0&-b_{s}^2& 0&0&b_{s}\\
\end{array}\right],$$
$$ \mathcal{W}= \left[\begin{array}{cccccccccccc}
0&-a_{\mathfrak{n}+1}&0&b_{\mathfrak{n}+1}&-b_{\mathfrak{n}+1}^2&0&0&0&-2a_{\mathfrak{n}+1}b_{\mathfrak{n}+1}^2&-b_{\mathfrak{n}+1}&b_{\mathfrak{n}+1}^2&a_{\mathfrak{n}+1}b_{\mathfrak{n}+1}\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&-a_i&0&b_i&-b_i^2&0&0&0&-2a_ib_i^2&-b_i&b_i^2&a_ib_i\\
\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots\\
0&-a_s&0&b_s&-b_s^2&0&0&0&-2a_sb_s^2&-b_s&b_s^2&a_sb_s\\
\end{array}\right].$$
\vspace{0.3cm}
We can represent all the equations above by using the following null product of representative matrix:
$$\mathcal{M}_{(3(s-1)-\mathfrak{n})\times 12}\cdot \left[\begin{array}{llllllllllll}
q_1&q_2&q_3&q_4&q_5&q_6&r_1&r_2&r_3&r_4&r_5&r_6
\end{array}\right]_{12\times 1}^{t}=0,$$
where the matrix $\mathcal{M}:=\mathcal{M}_{(3(s-1)-\mathfrak{n})\times 12}$ is the concatenation of the matrices $\mathcal{N}_{\mathrm{i}},\mathcal{N}_{\mathrm{ii}},\mathcal{N}_{\mathrm{iii}},\mathcal{N}_{\mathrm{iv}}$. This linear system has non trivial solution if and only if $\mathrm{rank}(\mathcal{M})<\min\{3(s-1)-\mathfrak{n},12\}.$
Using the classification of all line arrangements in $\mathbb{P}^2$ with $r(\mathcal{A})=3$ by (\ref{cubic}) we can determine the shape of each degree 3 logarithmic derivation associated to each the defining polynomial.
This comes handy especially when dealing with cases (IIIa)-(Vb), because $s$ has specific values. For example, the associated matrix $\mathcal{M}_{\mathrm{Vb}}$ to the type (Vb) $F=xyz(x+y)(x+z)(-x+z)(y+z)(-y+z)$ (here we changed $x\mapsto -x$) is
{\small$$\left[\begin{array}{cccccccccccc}
0&0&1&0&0&0&0&0&0&0&0&0\\
1&1&0&-1&0&0&0&0&0&0&0&0\\
0&0&0&0&1&-1&0&0&0&0&0&0\\
0&0&0&0&0& 0&0&1&0&0&0&0\\
0&0&0&0&0&0&1& 0&1&0&-1&0\\
0&0&0&0&0&0&1&0&1&0&1&0\\
0&0&0&0&0&0&0&0&0&1&0&-1\\
0&0&0&0&0&0&0&0&0&1&0&1\\
1&0&0&0&0&0&-1&0&0& 0&0&0\\
0&1&1&0&0&-1&0&-1&-1& 0&0&1\\
0&1&1&0&0&1&0&-1&-1& 0&0&-1\\
0&0&0&1&1&0&0&0&0&1&-1&0\\
0&0&0&1&-1&0&0&0&0&1&1&0\\
\end{array}\right]\Longleftrightarrow\left[\begin{array}{cccccccccccc}
0&0&1&0&0&0&0&0&0&0&0&0\\
1&1&0&0&0&0&0&0&0&0&0&0\\
0&0&0&1&0&0&0&0&0&0&0&0\\
0&0&0&0&1&0&0&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0&0&0&0\\
1&0&0&0&0&0&-1&0&0&0&0&0\\
0&0&0&0&0&0&0&1&0&0&0&0\\
0&0&0&0&0&0&1&0&1&0&0&0\\
0&1&0&0&0&0&0&0&-1&0&0&0\\
0&0&0&0&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&0&0&0&1&0\\
0&0&0&0&0&0&0&0&0&0&0&1\\
0&0&0&0&0&0&0&0&0&0&0&0\\
\end{array}\right].$$}
Therefore, the shape of the logarithmic derivation of this type of arrangement is $$\theta=(-x^2+y^2)y\frac{\partial}{\partial y} + (-x^2+z^2)z\frac{\partial}{\partial z}.$$
\subsubsection{Non recursive cubic logarithmic derivations.}
For types (Ia) - (IIIc), the corresponding line arrangements $\mathcal A$ are obtained by adding an extra line $H=V(l)$ to line arrangements $\mathcal B$ with minimal quadratic logarithmic derivation. The minimal cubic logarithmic derivations will be obtained by simply multiplying by $l$ the logarithmic derivations obtained in Corollary \ref{quadratic-log}, except when $\mathcal B$ has also a minimal cubic logarithmic derivation.
We are interested in this exceptional case. So it may happen that $$\theta = l'\rho_2+\rho_3,$$ where $l'\in S_1$ and $\rho_2$ is a minimal quadratic logarithmic derivation, and $\rho_3$ is a minimal cubic logarithmic derivation, both of $\mathcal B$. Also, since $r(\mathcal A)=3$, then $\rho_2(l)\notin \langle l\rangle$.
From Corollary \ref{quadratic-log}, we know the shape of $\rho_2$. So we are left to finding how $\rho_3$ looks like. Now we can use the shape of the defining polynomials of $\mathcal B$ as expressed in Theorem \ref{classification}, together with the matrix approach we have been discussing in this section.
\medskip
$\bullet$ If $\mathcal B$ is of type (I), then it is supersolvable, and we want it to have exponents $(1,2,3)$. Then its defining polynomial is $xyz(x+y)(t_4y+z)(t_5y+z)$. After some change of variables, we can assume this polynomial is $xyz(x+y)(y+z)(ty+z)$, where $t\neq 0,1$. Modulo linear multiples of $\rho_2$, we have $\displaystyle\rho_3=(ty+z)(y+z)z\frac{\partial}{\partial_z}$.
After a change of coordinates, we can assume that the defining polynomial of $\mathcal A$ in Theorem \ref{cubic} is $F=xyz(x+y)(y+z)(ty+z)l$, where $l=bx+y$ for (Ia), $l=bx+z$ for (Ib) and (Ic), $l=ax+by+z$ for (Id).
The question is if we can find $l'=\alpha x+\beta y +\gamma z$ such that $l'\rho_2(l)+\rho_3(l)\in\langle l\rangle$.
\begin{itemize}
\item[(Ia)] If $l=bx+y$, then we can take $l'=l$.
\item[(Ib)] If $l=bx+z$, $b\neq -1,$ then, in order to have such $\theta$, $b=-t$. But with this, we get the picture of (Ic), since we have three triple points, instead of just two: $\{x,y,x+y\},\{x,z,-tx+z\},\{x+y,-tx+z,ty+z\}$.
\item[(Ic)] If $l=-x+z$, then we can take $l'=x-ty-2z$.
\item[(Id)] If $l=ax+by+z$, then $b=t$ or $b=1$. But in this case we get two triple points instead of just one: $\{x,y,x+y\},\{x,ax+by+z,by+z\}$, so not a type (Id) arrangement.
\end{itemize}
\medskip
$\bullet$ If $\mathcal B$ is of type (II), then, after some simple calculation with Macaulay 2 (\cite{EiGrSt}), its defining polynomial is $xyz(x+y+z)(ty+z), t\neq 0,1$. In this case $\mathcal B$ has two minimal cubic logarithmic derivations: $$\rho_3=(ty+z)(x+y+z)(\alpha y\frac{\partial}{\partial y}+\beta z\frac{\partial}{\partial z}).$$ We also know that $\rho_2=(x+y+z)(y\frac{\partial}{\partial y}+z\frac{\partial}{\partial z})$.
If $l=ax+by+cz$, then we need to have $$l'(x+y+z)(by+cz)+(ty+z)(x+y+z)(\alpha by+\beta cz)=(ax+by+cz)Q, Q\in S_2.$$ Since $\gcd(l,x+y+z)=1$, then $Q=(x+y+z)P$, where $P\in S_1$.
\begin{itemize}
\item[(IIb)] If $c=0$ and $b=1$, then we have $$l'y+(ty+z)\alpha y=(ax+y)P,$$ leading to $P=\delta y$. Then $\alpha=0$ and $l'=\delta(ax+y)$.
\item[(IIa)] If $c=1$, and $a$ and $b$ are general enough ($a,b\neq 0$, $b\neq t$), we have $$l'(by+z)+(ty+z)(\alpha by+\beta z)=(ax+by+cz)P.$$ After some calculations, $\beta=\alpha$, and we can choose $l'=ax+(b-\alpha t)y+(1-\alpha)z$ (in this instance $P=by+z$).
\end{itemize}
\medskip
$\bullet$ If $\mathcal B$ is of type (III), as we mention before we can obtain all the minimal cubic logarithmic derivations by matrix computations.
\section{Appendix: graphic arrangements}
Let $G=(V,E)$ be a simple (no loops, no multiple edges) undirected graph with $V=\{1,\ldots,n\}$, $n\geq 2$. The {\em graphic arrangement corresponding to $G$} is $\mathcal A(G):=\{V(x_i-x_j)|(i,j)\in E\}$. The rank of $\mathcal A(G)$ equals $n$ minus the number of connected components of $G$.
Consider the derivation $$\theta=x_1^2\frac{\partial}{\partial x_1}+\cdots+x_n^2\frac{\partial}{\partial x_n}.$$ Obviously, $\theta(x_i-x_j)=x_i^2-x_j^2=(x_i-x_j)(x_i+x_j)$, so $\theta$ is a logarithmic derivation of $\mathcal A(G)$, not a multiple of the Euler derivation. So $$r(G):=r(\mathcal A(G))\leq 2.$$ We are interested when $r(G)=1$.
Below, $\delta(G)$ denotes the minimum degree of a vertex of $G$. Suppose $G$ is connected. $S\subset V$ is called {\em vertex-cut set}, if $G-S$ is disconnected. The smallest cardinality of a vertex-cut set is called {\em vertex-connectivity} of $G$, and it is denoted $k(G)$. By convention, $k(K_n)=n-1$.
\begin{flushleft}
$\begin{array}{lc}
\hspace{0.25cm}\textrm{A vertex}~v\in V~\textrm{such that}~ G-\{v\}~\textrm{is disconnected}~(\textrm{hence}~k(G)=1)~\textrm{is} \\
\textrm{called \emph{articulated vertex}, and}~G~\textrm{with}~k(G)=1~\textrm{is called {\emph {articulated graph}}}.\\
\\
\end{array}\hspace*{0.3cm}
\begin{tikzpicture}
\draw[thick](0, 0) ellipse (1 and 0.5) (2, 0) ellipse (1 and 0.5);
\draw [thick,fill] (1, 0) circle (0.05);
\end{tikzpicture}$
\end{flushleft}
\begin{lem}\label{bowtie} Let $G=(V,E)$ be a simple undirected graph with $|E|\geq 2$ and $\delta(G)\geq 1$. If $G$ is disconnected, or if $G$ is articulated, then $r(G)=1$.
\end{lem}
\begin{proof} To prove the lemma, we will appeal to the discussion in \cite[Section 2.1]{To}, where it says that a hyperplane arrangement $\mathcal A$ has $r(\mathcal A)=1$ if and only if $\mathcal A$ is reducible (i.e., $\mathcal A=\mathcal A_1\times\mathcal A_2$).
If $G$ is disconnected with components $G_1$ and $G_2$, then the defining polynomial of $\mathcal A(G)$ is the product of the defining polynomials of $\mathcal A(G_1)$ and $\mathcal A(G_2)$, and each of these is expressed in different sets of variables. Hence $\mathcal A(G)$ is reducible.
\medskip
Suppose $G$ is articulated at a vertex $v$, with the two induced (connected) subgraphs $G_1$ and $G_2$, such that $G=G_1\cup G_2$ and $G_1\cap G_2=\{v\}$.
Let $T$ be a spanning tree for $G$. Then we can make a change of variables such that for each $(i,j)$ edge of $T$, $x_i-x_j$ gets assigned a new variable. Since $T$ is obtained by removing efficiently edges from $G$ to ``destroy'' cycles, then all forms $x_a-x_b$ corresponding to other edges $(a,b)$ of $G$ will be linear combinations of these new variables.
But $T$ consists of a spanning tree of $G_1$ and a spanning tree of $G_2$ glued at the vertex $v$. Using the change of variables associated to spanning trees as above, the claim follows (there is no cycle of $G$ consists of edges from both $G_1$ and $G_2$).
\end{proof}
The converse is the following. But first, we summarize \cite[Proposition 2.87]{OrTe}: Let $e$ be an edge of $G$, and let $G':=G-e$ (the deletion of the edge $e$), and let $G'':= G/e$ (the contraction w.r.t. the edge $e$). If $H\in \mathcal A:=\mathcal A(G)$ is the hyperplane corresponding to $e$, then $$\mathcal A':=\mathcal A\setminus\{H\} = \mathcal A(G') \mbox{ and }\mathcal A'':=\mathcal A^H=\mathcal A(G'').$$
\begin{lem} Let $G=(V,E)$ be a simple connected graph with $s:=|E|\geq 2$. If $r(G)=1$, then $G$ is articulated.
\end{lem}
\begin{proof} We will prove the result by induction on $s\geq 2$. Of course, if $s=2$, then $G$ is a path of length 2, hence it is articulated.
Suppose $s\geq 3$. If $G$ has a leaf, then the neighbor of that leaf is an articulated vertex of $G$. Suppose $\delta(G)\geq 2$. Let $e\in E$, and consider $G'=G-e$. Then, $\delta(G')\geq 1$. By Theorem \ref{additiondeletion}, since the rank of $\mathcal A(G)$ is at least 2, we have $r(G')=1$.
$G'$ is disconnected or, by induction, $G'$ is articulated; i.e, $G'=G_1'\cup G_2'$, with $G_1'\cap G_2'=\emptyset$, or $G_1'\cap G_2'=\{v\}$.
In the first case, since $G$ is connected, then $e$ must be a bridge connecting $G_1'$ and $G_2'$. But then, any of the end vertices of $e$ is an articulated vertex for $G$.
In the second case, suppose $e$ connects the two halves $G_1'$ and $G_2'$ at vertices $w_1$ and $w_2$ respectively, and it is not incident to $v$ (for any other placement of $e$, $G$ is articulated at $v$).
If $G$ is the triangle $vw_1w_2$, then $G=K_3$, so $r(G)=2$. Contradiction. Otherwise, if one looks at $G'':=G/e$, we have: \begin{itemize}
\item $G''$ is connected.
\item $\mathrm{rank}(\mathcal{A}(G''))\geq 2.$
\item $G''$ has less than or equal $s-1$ edges.
\end{itemize}
If $G''$ is articulated, then $G$ is articulated. If $G''$ is not articulated, by induction, $r(G'')=2$. But then, by Theorem \ref{additiondeletion} (2), $r(G)=r(G')+1=2$; a contradiction.
\end{proof}
We end by mentioning that since $r(G)\leq 2$ for any graphic arrangement, it becomes more appealing to analyze ``the maximal degree of a Jacobian relation'' (so a maximal degree of a generating syzygy of the Jacobian ideal), and \cite[Corrolary 4.6]{Wa} finds an interesting lower bound of this invariant, in terms of the maximum number of new triangles that can be formed by adding an edge to $G$.
\vskip 0.3in
\noindent {\bf Acknowledgment.} Ricardo Burity is grateful to CAPES for funding his one year stay at University of Idaho.
\bigskip
\renewcommand{\baselinestretch}{1.0}
\small\normalsize
\bibliographystyle{amsalpha}
| e6d444b55405eea393a91d00063b7b5694d58adc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{section one}
The symmetric simple exclusion process (SSEP) with a slow bond was introduced in \cite{francogn2013slowbond} by Franco, Gon{\c{c}}alves and Neumann to derive from microscopic systems PDEs with boundary conditions, which has become a popular topic recently \cite{franco2016scaling,franco2019hydrodynamic,baldasso2017exclusion}. The process evolves on the discrete ring with $N$ sites, where $N$ is the scaling parameter. There is at most one particle per site. Particles cross each bond at rate $N^2$ except one particular bond, where the rate is $N$.
The hydrodynamic limit of the SSEP with a slow bond has been well understood \cite{francogn2013slowbond,francogn2015slowbond}. The hydrodynamic equation turns out to be the heat equation with Robin's boundary conditions:
\begin{equation}\label{hydroEqn}
\left\{\begin{array}{ll}\partial_{t} \rho\,(t, u)=\partial_{u}^{2} \rho\,(t, u), &\quad t>0, \,u \in \mathbb{T} \backslash\{0\}, \\ \partial_{u} \rho\left(t, 0^{+}\right)=\partial_{u} \rho\left(t, 0^{-}\right)=\rho\left(t, 0^{+}\right)-\rho\left(t, 0^{-}\right), &\quad t>0, \\ \rho(0, u)=\gamma(u), &\quad u \in \mathbb{T},\end{array}\right.
\end{equation}
where $\mathbb{T}$ is the continuous ring, $0^+$ and $0^-$ denote respectively the right limit and left limit at site $0$, and $\gamma (\cdot)$ is the initial density profile.
Then it is natural to consider the equilibrium fluctuations and large deviations from the hydrodynamic limit. Equilibrium fluctuations have been studied in \cite{francogn2013slowbondfluc} and large deviations in \cite{francon2017largedeviationslowbond} by Franco, Gon{\c{c}}alves and Neumann.
To better understand the SSEP with a slow bond, we consider the moderate deviations from the hydrodynamic limit, which gives asymptotic behavior of the model between the central limit theorem and the large deviation. As far as we know, the only paper concerned about moderate deviations from hydrodynamics is \cite{gao2003moderate} authored by Gao and Quastel, where the classic SSEP was considered. For literatures about theories of moderate deviations, see References \cite{Borovkov1978, deAcosta1998, Dembo1997, Gao1996, Wang2016, WangR2015, Wu1995} and so on.
\,
Next, we introduce the SSEP with a slow bond and main results. The process evolves on $\mathbb{T}_{N}=\{0,1,\ldots,N-1\}$ the ring with $N$ sites, with the convention $N \equiv 0$. Therefore, the state space is $\{0,1\}^{\mathbb{T}_N}$. For each configuration $\eta \in \{0,1\}^{\mathbb{T}_N}$, $\eta (x)=1$ means site $x$ is occupied by a particle, and $\eta (x) = 0$ means site $x$ is vacant. The infinitesimal generator $\mathcal{L}_N$ of the process is
\begin{equation*}
\mathcal{L}_N f (\eta) = N [f (\eta^{-1,0}) - f (\eta)] + N^2 \sum_{x \in \mathbb{T}_N, \atop x \neq -1} [f (\eta^{x,x+1}) - f(\eta)],
\end{equation*}
where
\[
\eta^{x,y}(u)=
\begin{cases}
\eta(u) & \text{~if~}u\neq x, y,\\
\eta(y) & \text{~if~}u=x,\\
\eta(x) & \text{~if~}u=y
\end{cases}
\]
for any $x\neq y$. Denote by $\{\eta_t\}_{t\geq 0}$ the process with generator $\mathcal{L}_N$. We suppress the dependence of the process $\{\eta_t\}_{t\geq 0}$ on $N$ for short.
Equivalently, we can define the process in the following way. For each $i\neq -1$, let $\{Y_i(t)\}_{t\geq 0}$ be a Poisson process with rate $N^2$ and $\{Y_{-1}(t)\}_{t\geq 0}$ be a Poisson process with rate $N$. Assume that all these Poisson processes are independent. Then at any event moment of $Y_i (\cdot)$, $\eta(i)$ and $\eta(i+1)$ exchange their values.
The SSEP with a slow bond has a family of invariant measures indexed by the particle density. To be precise, let $\nu_\rho,\,\rho \in [0,1],$ be the product measure on $\mathbb{T}_N$ with marginals given by
$$
\nu_\rho \{\eta: \eta (x) = 1\} = \rho, \,\,\forall x \in \mathbb{T}_N.
$$
Then, it can be checked easily that $\nu_\rho, \, \rho \in [0,1],$ are reversible measures for the process $\{\eta_t\}_{t\geq 0}$.
\,
To define the empirical density and rate functions, we need to introduce some definitions and notations and then discuss some topological issues. We identify $\mathbb{T}$ with $[0,1)$, and thus $0^+$ with $0$ and $0^-$ with $1$.
By the boundary conditions imposed on the hydrodynamic equation \eqref{hydroEqn}, it is reasonable to consider test functions $G\in C^1[0,1]$ with the property
\begin{equation}\label{equ special function}
G^\prime (0) = G^\prime (1) = G (0) - G (1).
\end{equation}
The result of this paper relies heavily on the above kind of functions, especially trigonometric functions satisfying \eqref{equ special function}.
Define $\mathscr{G}_0$ as
\[
\mathscr{G}_0 := {\rm span}\left(\left\{\sin\left(k_n\left(x-\frac{1}{2}\right)\right)\right\}_{n\geq 1}\,\bigcup\, \left\{\cos\big(2n\pi x\big)\right\}_{n\geq 0}\right),
\]
where $k_n$ is the unique solution to the equation $-\frac{x}{2}=\tan\frac{x}{2}$ in ${\big((2n-1)\pi, (2n+1)\pi\big)}$ for each $n \geq 1$. It can be checked easily that any $G \in \mathscr{G}_0$ satisfies \eqref{equ special function}. According to \cite[Theorem 1]{Franco2009moderate} given by Franco and Landim, we can prove the set of the above trigonometric functions is a basis in $L^2 [0,1]$, which is crucial to construct the topology of this paper.
\begin{lemma}\label{lemma 1.1.1 topology}
The set
$\left\{\sin\big(k_n(x-1/2)\big)\right\}_{n\geq 1}\,\bigcup \,\left\{\cos\big(2n\pi x\big)\right\}_{n\geq 0}$ is an orthogonal basis of $L^2[0,1]$.
\end{lemma}
We put the proof of Lemma \ref{lemma 1.1.1 topology} in the appendix.
\,
Let $\mathscr{M}$ be the space of linear (not necessarily bounded) functionals on $\mathscr{G}_0$ endowed with the following topology: for any $\mathscr{A}_n \in \mathscr{M},\, n \geq 1,$ and $\mathscr{A} \in \mathscr{M}$,
\[
\lim_{n\rightarrow+\infty}\mathscr{A}_n=\mathscr{A} \quad \text{in $\mathscr{M}$} \qquad \text{if and only if} \qquad \lim_{n\rightarrow+\infty}\mathscr{A}_n(\theta_k)=\mathscr{A}(\theta_k) \quad \text{for all integers $k$},
\]
where $\theta_n(x)=\sin\big(k_n(x-1/2)\big)$ for $n\geq 1$ and $\theta_{-n}(x)=\cos\big(2n\pi x\big)$ for $n\geq 0$. The above topology is metrizable and the metric $d\,(\cdot,\cdot)$ is given by
\[
d\big(\mathscr{A}_1, \mathscr{A}_2\big)=\sum_{-\infty<n<+\infty}\frac{1}{2^{|n|}}\frac{|\mathscr{A}_1(\theta_n)-\mathscr{A}_2(\theta_n)|}{1+|\mathscr{A}_1(\theta_n)-\mathscr{A}_2(\theta_n)|},\quad
\mathscr{A}_1, \,\mathscr{A}_2\in \mathscr{M}.
\]
It can be checked directly that the space $\mathscr{M}$ is complete and separable under the above metric. Note that a bounded signed measure $\mu$ on $[0,1]$ can be identified with an element in $\mathscr{M}$ in the sense that $\mu(f)=\int_{[0,1]}f(x)\mu(dx)$ for any $f\in \mathscr{G}_0$.
\,
\begin{remark}\label{remark 1.2}
We construct the above topology for technical reasons. Mainly, we can not show the uniqueness or existence of the weak solution to a PDE arising from hydrodynamics of the SSEP with a slow bond under a Girsanov's transformed measure. However, if we do not distinguish two measures $\mu_1$ and $\mu_2$ satisfying $\mu_1(\theta_n)=\mu_2(\theta_n)$ for all $n$, the above PDE can be reduced to an ODE on $\mathscr{M}$, the existence and uniqueness of the solution to which can be rigorously proved. For mathematical details, see Section \ref{section of lower bound} and appendix.
\end{remark}
\,
In the following, we will fix a horizonal time $T > 0$. Let $D\big([0,T], \mathscr{M} \big)$ be the space of c\`{a}dl\`{a}g functions from $[0,T]$ to $\mathscr{M}$ endowed with the Skorohod topology. Define the rescaled central empirical density $\mu^N_t (d u)$ as
\begin{equation*}
\mu^N_t (d u) := \frac{1}{a_N} \sum_{x \in \mathbb{T}_N} (\eta_t (x) - \rho) \delta_{x/N} (d u),
\end{equation*}
where $\sqrt{N} \ll a_N \ll N$. We will regard $\mu^N :=\{\mu_t^N\}_{0\leq t\leq T}$ as a random element taking values in $D\big([0,T], \mathscr{M} \big)$.
Let $\mathscr{G}$ be the family of functions $G :[0,T] \times [0,1] \rightarrow \mathbb{R}$ with the following forms: there exist $M \in \mathbb{N}$ and $b_m (t) \in C^1 ([0,T])$, $- M \leq m \leq M$ such that
\begin{equation*}
G (t,u) = \sum_{m=-M}^M b_m (t)\, \theta_m (u), \quad (t,u) \in [0,T] \times [0,1].
\end{equation*}
Then for any $G \in \mathscr{G}$,
\begin{equation}\label{bc}
\partial_u G (t,0) = \partial_u G (t,1) = G (t,0) - G (t,1),\quad \forall\, t \in [0,T].
\end{equation}
We sometimes write $G_t (u)$ for $G (t,u)$. For $G \in \mathscr{G}$, define the extended Laplacian $\tilde{\Delta}$ as
\[
\tilde{\Delta} G_t \,(u) = \begin{cases}
\partial_u^2 G_t\, (u) &\text{if $u \neq 0$},\\
\partial_u^2 G_t\, (0^+) &\text{if $u = 0$}.
\end{cases}
\]
Fix a density $\rho \in (0,1)$. Denote by $Q^N_\rho$ the law of $\{\mu^N_t\}_{0 \leq t \leq T}$ with initial distribution $\nu_\rho$. Let $\mathbb{P}^N_\rho$ be the law of the process $\{\eta_t\}_{0 \leq t \leq T}$ with initial distribution $\nu_\rho$, and $\mathbb{E}^N_\rho$ the corresponding expectation. Let $\rm{E}_{\nu_\rho}$ be the expectation with respect to $\nu_\rho$. For $\mu \in D ([0,T],\mathscr{M})$, define
\begin{equation}
\begin{aligned}
&I (\mu) := I_{ini} (\mu_0) + I_{dyn} (\mu),\\
&I_{ini}(\mu_0) :=\sup_{\gamma \in \mathscr{G}_0} \left\{\mu_0 (\gamma) -\frac{\rho(1-\rho)}{2}\int_0^1 \gamma^2(u) \,du \right\},\\
&I_{dyn} (\mu) := \sup_{G \in \mathscr{G}} \left\{ \ell_T (\mu,G) - \rho (1-\rho) \int_0^T \left(G_t (0) - G_t (1) \right)^2 dt \right. \\
&\left. \qquad \qquad \qquad \qquad \qquad \qquad
- \rho (1-\rho) \int_0^T \int_0^1 \left( \partial_u G_t (u) \right)^2 \,d u \,d t\right\},
\end{aligned}
\end{equation}
where
\begin{align}\label{l_T}
\ell_T (\mu,G) := \mu_T\left(G_T\right) - \mu_0 \left(G_0\right) - \int_0^T \mu_t \left((\partial_t + \tilde{\Delta}) G_t\right) \,d t.
\end{align}
Now we are ready to state the main result of the paper.
\,
\begin{theorem}\label{t1 main}
For any closed set $C$ of $D \left( [0,T], \mathscr{M} \right) $,
\begin{equation}\label{eqnupperbound}
\limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log Q^N_\rho\, [C] \leq - \inf_{\mu \in C} I (\mu),
\end{equation}
and for any open set $O$ of $D \left( [0,T], \mathscr{M} \right)$,
\begin{equation}\label{equ lower bound}
\liminf_{N \rightarrow \infty} \frac{N}{a_N^2} \log Q^N_\rho\, [O] \geq - \inf_{\mu \in O} I (\mu).
\end{equation}
\end{theorem}
\,
\begin{remark}\label{remark 1.4}
We recall the large deviation principle of the SSEP with a slow bond established in \cite{francon2017largedeviationslowbond} by Franco and Neumann for a comparison. Note that definitions and notations in this remark are not utilized elsewhere. Let
\[
\pi^N_t (d u)= \frac{1}{N} \sum_{x \in \mathbb{T}_N} \eta_t (x) \delta_{x/N} (d u),
\quad
\pi^N=\{\pi^N_t\}_{0\leq t\leq T},
\]
then it was shown in \cite{francon2017largedeviationslowbond} that, roughly speaking,
\[
P(\pi^N\approx\pi)\approx \exp\left\{ -N J(\pi) \right\}
\]
assuming uniqueness for the weak solution to hydrodynamic equation associated to the perturbed process, where
\begin{align*}
J(\pi)=\sup_{H\in C^{1,2}\left([0, T]\times [0,1]\right)}\{\widehat{\ell}_H(\pi)-\Phi_H(\pi)\}
\end{align*}
with $\widehat{\ell}_H(\pi)$ given by
\begin{align*}
\widehat{\ell}_H(\pi)=&\langle \rho_T, H_T\rangle-\langle\rho_0, H_0\rangle-\int_0^T\langle \rho_t, \left(\partial_t+\Delta \right)H_t\rangle \,dt\\
&-\int_0^T \left\{\rho_t(0)\partial_uH_t(0)-\rho_t(1)\partial_uH_t(1)\right\}\,dt+\int_0^T \left(\rho_t(0)-\rho_t(1)\right)\delta H_t(0)\,dt
\end{align*}
and $\Phi_H(\pi)$ given by
\begin{align*}
\Phi_H(\pi)=&\int_0^T\langle\chi(\rho_t), \left(\partial_uH_t\right)^2\rangle\, dt+\int_0^T\rho_t(1)\left(1-\rho_t(0)\right)\psi\left(\delta H_t(0)\right)\,dt\\
&+\int_0^T\rho_t(0)\left(1-\rho_t(1)\right)\psi\left(-\delta H_t(0)\right)\,dt,
\end{align*}
where $\rho_t$ is the Radon-Nikodym derivative of $\pi_t$ with respect to the Lebesgue measure, $\psi(x)=e^x-x-1$, $\chi(\rho)=\rho(1-\rho)$ and $\delta H_t(0)=H_t(0)-H_t(1)$. Since $e^x-x-1=\frac{x^2}{2}+o(x^2)$ as $|x|$ decreases to $0$ and
$\widehat{\ell}_H(\pi)$ equals $\ell_T(\pi, H)$ when $H$ satisfies \eqref{bc}, the rate function $I_{dyn}$ can be intuitively considered as the quadratic part of $J$ about its minimum, which is a common relationship between large and moderate deviations for many models in statistical physics.
\end{remark}
\,
{\bf Notation.} For deterministic positive sequences $\{b_n\}_{n\geq 1}$, $\{c_n\}_{n\geq 1}$ and random sequence $\{X_n\}_{n\geq 1}$, we write $b_n=o(c_n)$ if $
\limsup_{n \rightarrow \infty}\, b_n/c_n = 0
$ and $b_n = \mathcal{O} (c_n)$ if
$
\limsup_{n \rightarrow \infty}\, b_n/c_n < C
$
for some constant $C$ independent of $n$. We also write $b_n = \mathcal{O}_G (c_n)$ to stress the dependence on some parameter $G$ of the constant $C$.
We write $X_n=o_p(c_n)$ if $X_n / c_n \rightarrow 0$ in probability as $n \rightarrow \infty$, and $X_n=o_{\exp}(c_n)$ if
\[
\limsup_{n\rightarrow+\infty}\,\frac{1}{c_n}\,\log P\big(|X_n|>\epsilon\big)=-\infty, \quad \forall \epsilon > 0.
\]
We remark on these last points that the constant throughout the paper may be different from line to line.
\,
The rest of the paper is devoted to the proof of Theorem \ref{t1 main}. In Section \ref{secsed} we give several super-exponential estimates that are necessary in the proof of upper and lower bounds as a preparation. Moderate upper bounds are proved in Section \ref{sec upper}. Our proof follows a strategy similar with that introduced in \cite{gao2003moderate}, except for some details modified due to technical reasons caused by the slow bond. First, as introduced above, we have to choose a proper topology and to consider the empirical density as a random element taking values in the linear functional space $\mathscr{M}$, instead of the dual of Schwartz functions. Second, an extra super-exponential estimate (Lemma \ref{lem1}) is needed. Third, because of the topology constructed, we have to use a different version of Minimax Theorem (Theorem \ref{thm minimax}) from the one in \cite{gao2003moderate}. Moderate lower bounds are proved in Section \ref{section of lower bound}. A crucial step in the proof is the utilizing of a generalized Girsanov's theorem to give the hydrodynamic equation of the model under a transformed measure.
\section{Super-exponential Decay}\label{secsed}
In this section, we mainly present three super-exponential estimates that are critical when making some replacements and proving exponential tightness.
\,
\begin{lemma}\label{lem1}
For any continuous function $G: [0,T] \rightarrow \mathbb{R}$ and any $\delta,\,t > 0$,
\begin{equation}\label{l2}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho \left[ \left|\int_0^t \left(\eta_s(0) (1-\eta_s(-1)) - \rho (1-\rho)\right) G_s \, ds\right| > \delta\right] = - \infty.
\end{equation}
The same result holds with $\eta_s (0) (1-\eta_s (-1))$ replaced by $\eta_s (-1) (1-\eta_s (0))$.
\end{lemma}
\begin{proof}
We only present the proof of \eqref{l2} since the rest is the same. For any integer $M > 0$ and $x \in \mathbb{T}_N$, define $\eta^{M,\rm{R}} (x)$ (resp. $\eta^{M,\rm{L}} (x)$) as the average density over the box of size $M$ to the right (resp. left) of site $x$ ,
\begin{equation*}
\eta^{M,\rm{R}} (x) = \frac{1}{M} \sum_{y=x}^{x+M-1} \eta(y), \quad
\eta^{M,\rm{L}} (x) = \frac{1}{M} \sum_{y=x-M+1}^{x} \eta(y).
\end{equation*}
Note that for every integer $M > 0$,
\begin{align*}
\eta (0) (1-\eta (-1)) &- \rho (1-\rho) =
\left(\eta(0) - \eta^{M,\rm{R}} (0) \right) \left(1-\eta(-1)\right) \\
&+ \eta^{M,\rm{R}} (0) (\eta^{M,\rm{L}} (-1) - \eta(-1))
+ \left(\eta^{M,\rm{R}} (0) - \rho \right)\left(1-\eta^{M,\rm{L}} (-1) \right)\\
&+ \rho \left(\rho - \eta^{M,\rm{L}} (-1) \right).
\end{align*}
Since for any positive sequences $\{b_N\}_{N \geq 1}$ and $\{c_N\}_{N \geq 1}$,
$$
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log (b_N+c_N) \leq \max \left\{ \limsup_{N \rightarrow \infty} \frac{1}{a_N} \log b_N,\, \limsup_{N \rightarrow \infty} \frac{1}{a_N} \log c_N \right\},
$$
to prove \eqref{l2}, we only need to prove for any $\delta > 0$,
\begin{equation}\label{a1}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho \left[ \left|\int_0^t \left(\eta_s(0) - \eta_s^{M,\rm{R}} (0) \right) \left(1-\eta_s(-1)\right) G_s \,ds\right| > \delta\right] = - \infty,
\end{equation}
\begin{equation*}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho \left[ \left|\int_0^t \eta_s^{M,\rm{R}} (0) (\eta_s^{M,\rm{L}} (-1) - \eta_s(-1)) G_s \,ds\right| > \delta\right] = - \infty,
\end{equation*}
\begin{equation}\label{a2}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho \left[ \left|\int_0^t \left(\eta_s^{M,\rm{R}} (0) - \rho \right)\left(1-\eta_s^{M,\rm{L}} (-1) \right) G_s \,ds\right| > \delta\right] = - \infty,
\end{equation}
and
\begin{equation*}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho \left[ \left|\int_0^t \rho \left(\rho - \eta_s^{M,\rm{L}} (-1) \right) G_s \,ds\right| > \delta\right] = - \infty.
\end{equation*}
We only prove \eqref{a1} and \eqref{a2}, since the remaining two terms are similar.
\,
For any $A > 0$, by Chebyshev's inequality, the formula on the left-hand side of \eqref{a1} is bounded from above by
\begin{equation}\label{a3}
-\frac{A \delta}{a_N} + \frac{1}{a_N} \log \mathbb{E}^N_\rho \left[ \exp \left\{ A \left|\int_0^t \left(\eta_s(0) - \eta_s^{M,\rm{R}} (0) \right) \left(1-\eta_s(-1)\right) G_s \,ds\right|\right\} \right].
\end{equation}
Since $e^{|x|} \leq e^x + e^{-x}$, we can remove the modulus in the expectation above. By the Feynman-Kac formula (see \cite[Lemma A.1.7.2]{klscaling} by Kipnis and Landim for example), the second term in \eqref{a3} is bounded by
\begin{equation*}
\frac{1}{a_N} \int_0^t d s \sup_{\text{$f$ density}} \left\{ A G_s \int \left(\eta (0) - \eta^{M,\rm{R}} (0) \right) \left(1-\eta(-1)\right) f (\eta) \,d \nu_\rho - \mathcal{D}_N \left( f; \nu_\rho \right) \right\},
\end{equation*}
where $\mathcal{D}_N \left( f; \nu_\rho \right)$ is the Dirichlet form of $f$ associated with $\nu_\rho$ given by
\begin{equation*}
\begin{aligned}
\mathcal{D}_N \left( f; \nu_\rho \right) &:= \left\langle \sqrt{f}, (- \mathcal{L}_N) \sqrt{f}\, \right\rangle_{\nu_\rho} \\
& = N \left[\sqrt{f (\eta^{-1,0})} - \sqrt{f (\eta)}\right]^2
+ N^2 \sum_{x \in \mathbb{T}_N, \atop x \neq -1} \left[\sqrt{f (\eta^{x,x+1})} - \sqrt{f(\eta)}\right]^2.
\end{aligned}
\end{equation*}
We first write $\eta (0) - \eta^{M,\rm{R}} (0)$ as a telescope sum,
\begin{equation*}
\eta (0) - \eta^{M,\rm{R}} (0) = \frac{1}{M} \sum_{x=0}^{M-1} \sum_{y=0}^{x-1} \left( \eta(y) - \eta(y+1) \right).
\end{equation*}
Making the transformations $\eta \rightarrow \eta^{y,y+1}$, by Cauchy-Schwartz inequality, we obtain that there exists a constant $C$ only depending on $G$ such that for any $B > 0$,
\begin{align*}
&A G_s \int \left(\eta (0) - \eta^{M,\rm{R}} (0) \right) \left(1-\eta(-1)\right) f (\eta) \,d \nu_\rho \\
& = \frac{A G_s}{2 M} \sum_{x=0}^{M-1} \sum_{y=0}^{x-1} \int \left( \eta(y) - \eta(y+1) \right)
\left(1-\eta(-1)\right) \left(f(\eta) - f(\eta^{y,y+1})\right) d \nu_\rho\\
& \leq \frac{A B ||G||_{\infty}}{4 M} \sum_{x=0}^{M-1} \sum_{y=0}^{x-1} \int \left(\sqrt{f(\eta)} - \sqrt{f(\eta^{y,y+1})}\right)^2 d \nu_\rho \\
&\quad \quad \quad \quad \quad \quad + \frac{A ||G||_{\infty}}{4 B M} \sum_{x=0}^{M-1} \sum_{y=0}^{x-1} \int \left(\sqrt{f(\eta)} + \sqrt{f(\eta^{y,y+1})}\right)^2 d \nu_\rho\\
&\leq C \left( \frac{A B }{ N^2} \mathcal{D}_N \left( f; \nu_\rho \right) + \frac{A M}{B} \right).
\end{align*}
Taking $B = N^2 A^{-1} C^{-1}$, we bound \eqref{a3} by
\begin{equation*}
\inf_{A > 0} \,\left\{-\frac{A \delta}{a_N} + \frac{ A^2 C^2 t M}{N^2 a_N}\right\} = - \frac{\delta^2 N^2}{4 C^2 t M a_N}.
\end{equation*}
We prove \eqref{a1} by choosing $M$ such that $M a_N \ll N^2$.
\,
As above, for any $A > 0$, the formula on the left-hand side of \eqref{a2} is bounded by
\begin{align*}
-\frac{A \delta}{a_N} + \frac{1}{a_N} \log \mathbb{E}^N_\rho \left[ \exp \left\{ A \left| \int_0^t \left(\eta_s^{M,\rm{R}} (0) - \rho \right)\left(1-\eta_s^{M,\rm{L}} (-1) \right) G_s \,ds \right|\right\} \right].
\end{align*}
As before, we can first remove the modulus. By Jensen's inequality and the invariance of the measure $\nu_\rho$, we bound the above formula by
\begin{equation}\label{a4}
-\frac{A \delta}{a_N} + \frac{1}{a_N} \log \left(\frac{1}{t} \int_0^t d s \,\rm{E}_{\nu_\rho} \left[ \exp \left\{ A t G_s \left(\eta^{M,\rm{R}} (0) - \rho \right)\left(1-\eta^{M,\rm{L}} (-1) \right) \right\} \right] \right).
\end{equation}
By Taylor's expansion, the expectation in the above formula is less than or equal to
\begin{equation*}
\begin{aligned}
&\sum_{k \geq 0} \frac{A^{2k} t^{2k} ||G||_\infty^{2k}}{(2k)!} \,\rm{E}_{\nu_\rho} \left[ \left(\eta^{M,\rm{R}} (0) - \rho \right)^{2k}\right] \\
&\quad \quad + \sum_{k \geq 0} \frac{A^{2k+1} t^{2k+1} ||G||_\infty^{2k+1}}{(2k+1)!} \,\rm{E}_{\nu_\rho} \left[ \left|\eta^{M,\rm{R}} (0) - \rho \right|^{2k+1}\right] \\
& \leq (1+ At||G||_\infty)\sum_{k \geq 0} \frac{A^{2k} t^{2k} ||G||_\infty^{2k}}{(2k)!} \,\rm{E}_{\nu_\rho} \left[ \left(\eta^{M,\rm{R}} (0) - \rho \right)^{2k}\right].
\end{aligned}
\end{equation*}
Some computations (see the appendix) show that there exists a constant $ C (\rho)$ such that
\begin{equation}\label{equ control of 2k moments}
{\rm E}_{\nu_\rho} \left[ \left(\eta^{M,{\rm R}} (0) - \rho \right)^{2k} \right] \leq \frac{C(\rho)^k\, k!}{M^k}.
\end{equation}
Since $2^k( k! )^2 \leq (2k)!$, the expectation in \eqref{a4} is bounded by $C A \exp\{C A^2/M\}$ for some constant $C = C (t,G,\rho)$. Therefore, we bound \eqref{a4} by
\begin{equation*}
-\frac{A \delta}{a_N} + \frac{C A^2}{M a_N} + \frac{\log C + \log A}{a_N}.
\end{equation*}
We finish the proof by taking $M = [N/2]$ and $A= M \delta / (2 C)$.
\end{proof}
\,
\begin{lemma} \label{lem2}
For any $G \in C \left([0,T]\times [0,1]\right)$ and any $\delta,\,t > 0$,
\begin{equation*}
\begin{aligned}
\limsup_{N \rightarrow \infty} \frac{1}{a_N} \log \mathbb{P}^N_\rho
\left[\left|\int_0^t \frac{ 1}{N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} \left[ (\eta_s (x) - \eta_s (x+1) )^2- 2 \rho(1-\rho) \right] G_s \left(\frac{x}{N}\right) \,d s \right| > \delta\right]
= - \infty.
\end{aligned}
\end{equation*}
\end{lemma}
\begin{comment
\begin{proof}
Fix $A > 0$. The probability estimated in the lemma is bounded form above by
\begin{equation*}
e^{-A \delta} \, \mathbb{E}^N_\rho \left[ \exp \left\{ \left|\int_0^t \frac{ A}{N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} \left( \eta_s (x)( 1- \eta_s (x+1)) - \rho(1-\rho) \right) G_s \left(\frac{x}{N}\right) d s \right| \right\} \right].
\end{equation*}
Since $e^{|x|} \leq e^x + e^{-x}$, we can remove the modulus in the expectation above. Moreover, by Jensen's inequality, we can bound the expectation by
\begin{equation*}
\rm{E}_{\nu_\rho} \left[ \exp \left\{ \frac{ At||G||_\infty}{N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} \left( \eta (x)( 1- \eta (x+1)) - \rho(1-\rho) \right) \right\} \right].
\end{equation*}
By the Cauchy-Schwartz inequality, the above formula is less than or equal to
\begin{align*}
\left(\rm{E}_{\nu_\rho} \left[ \exp \left\{ \frac{ 2At ||G||_\infty}{N} \sum_{\text{$x$ is odd},\atop x \neq -1} \left( \eta (x)( 1- \eta (x+1)) - \rho(1-\rho) \right) \right\} \right]\right)^{1/2} \\
\times \left(\rm{E}_{\nu_\rho} \left[ \exp \left\{ \frac{ 2At ||G||_\infty}{N} \sum_{\text{$x$ is even}} \left( \eta (x)( 1- \eta (x+1)) - \rho(1-\rho) \right) \right\} \right]\right)^{1/2}.
\end{align*}
Note that under $\nu_\rho$, $\{\eta(x) (1 - \eta(x+1)),\,\text{$x$ is odd}\}$ (resp. $\{\eta(x) (1 - \eta(x+1)),\,\text{$x$ is even}\}$)are independent random variables. Using the inequality $e^x \leq 1 + x + x^2/2 + |x|^3e^{|x|}$, we bound the above formula by
\begin{equation*}
\left(1 + \frac{C(t,G,\rho) A^2}{N^2} + \frac{C(t,G,\rho) A^3}{N^3} \exp\{C(t,G,\rho) A/N\}\right)^{N}.
\end{equation*}
Therefore, the formula in \eqref{l1} is bounded by
\begin{equation*}
-\frac{A \delta}{a_N} + \frac{C(t,G,\rho) A^2}{N a_N} + \frac{C(t,G,\rho) A^3}{N^2 a_N} \exp \left\{C (t,G,\rho) A/N \right\}.
\end{equation*}
We conclude the proof by taking $A = \sqrt{N a_N}$.
\end{proof}.
\end{comment}
\,
\begin{lemma}\label{lem3}
Let $G \in C [0,1]$. Then for any $t > 0$,
\begin{equation}
\limsup _{A \rightarrow \infty}\, \limsup _{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \left[ \sup_{0 \leqslant t \leqslant T} \left| \int_0^t \langle \mu^N_s, G\rangle \,d s \right| > A \right] = -\infty,
\end{equation}
and for any $\epsilon,\,t > 0$,
\begin{equation}
\limsup_{\delta \rightarrow 0} \,\limsup_{N \rightarrow \infty}\, \frac{N}{a_N^2} \log \mathbb{P}^N_\rho \left[\sup_{|t-s| \leq \delta} \left| \int_s^t \left\langle \mu^N_u, G \right\rangle \,d u \right| > \epsilon \right] = - \infty.
\end{equation}
\end{lemma}
\,
The proof of \cite[Lemmas 2.1 and 2.2]{gao2003moderate} also applies to the above two lemmas. The main ingredients are the invariance of the Bernoulli product measure $\nu_\rho$. For that reason we omit the proof.
\section{Upper Bound}\label{sec upper}
In this section, we prove \eqref{eqnupperbound} the moderate deviations upper bound. The strategy is first proving upper bound over compact sets, and then extending to closed sets, which follows from the exponential tightness.
\,
\begin{comment}
For integers $m,\,n \geq 0$, let $C^{m,n} \left([0,T] \times [0,1]\right)$ be the space of functions which are $m$-th continuously differentiable at the time variable and $n$-th at the space variable. Endow $C^{m,n} \left([0,T] \times [0,1]\right)$ with the metric $d\,(\cdot,\cdot)$ defined as
$$
d\, (f,g) := \sum_{i=0}^m \, \sum_{j=0}^n\, \sup_{0 \leq t \leq T,\atop
0 \leq u \leq 1}
\left|\left(\partial_t^i \, \partial_u^j \, f \right) \,(t,u)- \left(\partial_t^i \, \partial_u^j \,g\right) \,(t,u)\right|,\quad f,g\in C^{m,n} \left([0,T] \times [0,1]\right).
$$
It is easy to check that the space $C^{m,n} \left([0,T] \times [0,1]\right)$ is complete and separable under the metric $d(\cdot,\cdot)$ introduced above. Let $\bar{\mathscr{G}}$ be the closure of $\mathscr{G}$ in $C^{1,2} \left([0,T] \times [0,1]\right)$. Then $\bar{\mathscr{G}}$ is also complete and separable. Moreover, for any $G \in \bar{\mathscr{G}}$, for any $0 \leq t \leq T$,
\begin{equation}
\partial_u G_t (0) = \partial_u G_t (1) = G_t (0) - G_t (1).
\end{equation}
\end{comment}
Fix $G \in {\mathscr{G}}$. By Feynman-Kac formula (see \cite[A.1.7]{klscaling}),
\begin{equation}\label{martingale1}
M_t^N (G):= \frac{f(t,\eta_t)}{f(0,\eta_0)} \exp \left\{ - \int_0^t \frac{\partial_s f + \mathcal{L}_N f }{f} (s,\eta_s) \,d s\right\}
\end{equation}
is a positive mean-one martingale, where
\begin{equation*}
f (t,\eta) := f_G (t,\eta) := \exp \left\{ \frac{a_N}{N} \sum_{x \in \mathbb{T}_N} \left(\eta (x) - \rho\right) G_t \left(x/N\right)\right\}.
\end{equation*}
Notice that
$$
f (t,\eta_t) = \exp \left\{ \frac{a_N^2}{N} \left\langle \mu^N_t,G_t \right\rangle \right\}.
$$
A simple calculation yields that
\begin{align*}
(\partial_s f + \mathcal{L}_N f) (s,\eta_s) &= f (s,\eta_s) \Bigg( \frac{a_N^2}{N} \langle \mu^N_s,\partial_s G_s \rangle \\
&+ N \left[ \exp \left\{\frac{a_N}{N} (\eta_s (-1) - \eta_s (0)) \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right)\right\} -1 \right]\\
&\left. + \sum_{x \in \mathbb{T}_N,\atop x \neq -1} N^2 \left[ \exp \left\{\frac{a_N}{N} (\eta_s (x) - \eta_s (x+1)) \left( G_s \left(\frac{x+1}{N}\right) - G_s \left(\frac{x}{N}\right) \right)\right\} -1 \right] \right).
\end{align*}
By Taylor's expansion,
\begin{align*}
&N \left[ \exp \left\{\frac{a_N}{N} (\eta_s (-1) - \eta_s (0)) \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right)\right\} -1 \right]\\
&= a_N (\eta_s (-1) - \eta_s (0)) \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right) \\
&+ \frac{a_N^2}{2 N} (\eta_s (-1) - \eta_s (0))^2 \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right)^2 + \mathcal{O}_G \left(\frac{a_N^3}{N^2}\right),
\end{align*}
and for $x \neq -1$,
\begin{align*}
&N^2 \left[ \exp \left\{\frac{a_N}{N} (\eta_s (x) - \eta_s (x+1)) \left( G_s \left(\frac{x+1}{N}\right) - G_s \left(\frac{x}{N}\right) \right)\right\} -1\right]\\
&= N a_N (\eta_s (x) - \eta_s (x+1)) \left( G_s \left(\frac{x+1}{N}\right) - G_s \left(\frac{x}{N}\right) \right)\\
&+ \frac{a_N^2}{2} (\eta_s(x) - \eta_s(x+1))^2 \left( G_s \left(\frac{x+1}{N}\right) - G_s \left(\frac{x}{N}\right)^2 \right) + \mathcal{O}_G \left(\frac{ a_N^3}{N^4}\right).
\end{align*}
Using the summation by parts formula,
\begin{align}
M_t^N (G) &= \exp \frac{a_N^2}{N} \bigg\{ \langle \mu^N_t, G_t\rangle - \langle \mu^N_0, G_0\rangle - \int_0^t \langle \mu^N_s, (\partial_s + \tilde{\Delta}) G_s\rangle ds \label{r0}\\
&- \int_0^t \frac{N}{a_N} (\eta_s (-1) - \eta_s (0)) \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right) d s\label{r1} \\
&- \int_0^t \frac{1}{2} (\eta_s (-1) - \eta_s (0))^2 \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right)^2 ds \label{r2}\\
&- \int_0^t \frac{N}{a_N} \left( (\eta_s (0) - \rho) \nabla_N G_s \left(\frac{0}{N}\right) - (\eta_s (-1) - \rho) \nabla_N G_s \left(\frac{-2}{N}\right)\right) ds \label{r3}\\
&- \int_0^t \frac{1}{2 N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} (\eta_s (x) - \eta_s (x+1))^2 \left(\nabla_N G_s \left(\frac{x}{N}\right)\right)^2 d s \label{r4}\\
&+ \mathcal{O}_G \left(\frac{a_N}{N}\right) + \mathcal{O}_G \left(\frac{ 1}{a_N}\right) \bigg\},\label{r5}
\end{align}
where $\nabla_N$ is the discrete space derivative, $\nabla_N G_s (x/N) := N \left[ G_s \left((x+1)/N\right) - G_s (x/N) \right]$. By the boundary condition \eqref{bc} imposed on $G$, the sum of \eqref{r1} and \eqref{r3} is of order $\mathcal{O}_G \left( a_N^{-1} \right)$. Therefore,
\begin{equation*}
\begin{aligned}
M_t^N (G) &= \exp \frac{a_N^2}{N} \left\{ \ell_t (\mu^N) - \int_0^t \frac{1}{2} (\eta_s (-1) - \eta_s (0))^2 \left( G_s \left(\frac{0}{N}\right) - G_s \left(\frac{-1}{N}\right) \right)^2 ds\right.\\
&\left.- \int_0^t \frac{1}{2 N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} (\eta_s (x) - \eta_s (x+1))^2 \left(\nabla_N G_s \left(\frac{x}{N}\right)\right)^2 d s + \mathcal{O}_G \left(\frac{ 1}{a_N}\right) \right\}.
\end{aligned}
\end{equation*}
\,
\begin{lemma}[Upper bounds over compact sets]
For any compact set $K \subset D \left( [0,T], \mathscr{M} \right) $,
\begin{equation}
\limsup_{N \rightarrow \infty} \, \frac{N}{a_N^2} \log Q^N_\rho [K] \leq - \inf_{\mu \in K} I(\mu).
\end{equation}
\end{lemma}
\begin{proof}
For any $\delta > 0$ and any $G \in {\mathscr{G}}$, let
\begin{equation*}
\begin{aligned}
B_{N,\delta} &= \left\{ \left| \int_0^T \sum_{x \in \mathbb{T}_N,\atop x \neq -1} \frac{1}{2 N} (\eta_t (x) - \eta_t (x+1))^2 \left(\nabla_N G_t \left(\frac{x}{N}\right)\right)^2 d t - \int_0^T \int_0^1 \rho (1-\rho) \left(\nabla G_t (u)\right)^2 \,dt \,du\right| < \delta \right\}\\
&\bigcap \left\{ \left| \int_0^T \left[ \frac{1}{2} (\eta_t (-1) - \eta_t (0))^2 - \rho (1-\rho) \right] \left( G_t \left(\frac{0}{N}\right) - G_t \left(\frac{-1}{N}\right) \right)^2 d t \right| < \delta \right\}.
\end{aligned}
\end{equation*}
By Lemmas \ref{lem1},\,\ref{lem2} and the assumption $a_N \ll N$,
\begin{equation*}
\limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{P}^N_\rho [B_{N,\delta}^c] = - \infty.
\end{equation*}
Therefore, for any $\gamma \in \mathscr{G}_0$,
\begin{equation*}
\begin{aligned}
\limsup_{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^N_\rho [\mu^N \in K]
& \leq \limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{P}^N_\rho \left[\left\{\mu^N \in K \right\} \cap B_{N,\delta} \right]\\
&= \limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{E}^N_\rho \left[ \left(M_T^N (G)\right)^{-1} M_T^N (G) \mathbf{1}_{ \left\{\mu^N \in K\right\} \bigcap B_{N,\delta} } \right]\\
&\leq \sup_{\mu \in K} \left\{- \ell_T (\mu) + \rho (1-\rho) \int_0^T \left(G_t (0) - G_t (1) \right)^2 dt \right.\\
&\qquad \qquad \left. + \rho (1-\rho) \int_0^T \int_0^1 \left( \partial_u G_t (u) \right)^2 \,d u \,d t - \mu_0 (\gamma ) \right\} + \mathcal{O} (\delta)\\
& + \limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{E}^N_\rho \left[ M_T^N (G) \exp \left\{ \frac{a_N^2}{N} \langle \mu^N_0,\gamma \rangle \right\}\right].
\end{aligned}
\end{equation*}
Because $\{M_t^N (G)\}$ is a mean one martingale and $\nu_\rho$ is a product measure, direct calculations yield that
\begin{equation*}
\limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{E}^N_\rho \left[ M_T^N (G) \exp \left\{ \frac{a_N^2}{N} \langle \mu^N_0,\gamma \rangle \right\}\right] = \frac{\rho (1-\rho)}{2} \int_0^1 \gamma (u)^2 \,d u.
\end{equation*}
Letting $\delta \rightarrow 0$,
and then minimizing over $G \in {\mathscr{G}}, \gamma \in \mathscr{G}_0$,
\begin{equation*}
\begin{aligned}
&\limsup_{N \rightarrow \infty} \frac{N}{a_N^2} \log \mathbb{P}^N_\rho [\mu^N \in K]
\leq \inf_{G \in {\mathscr{G}}, \atop \gamma \in \mathscr{G}_0} \,\sup_{\mu \in K}\,
\left\{ - \ell_T (\mu) + \rho (1-\rho) \int_0^T \left(G_t (0) - G_t (1) \right)^2 dt \right.\\
&\qquad \left. + \rho (1-\rho) \int_0^T \int_0^1 \left( \partial_u G_t (u) \right)^2 \,d u \,d t - \mu_0 (\gamma) + \frac{\rho (1-\rho)}{2} \int_0^1 \gamma (u)^2 \,d u \right\}.
\end{aligned}
\end{equation*}
\,
In order to exchange the supremum and infimum above, we use the following version of Minimax Theorem proved by Nikaid{\^o}.
\,
\begin{theorem}[Minimax Theorem, {\cite[Theorem 1]{nikaido1953minimax}}]\label{thm minimax}
Let $\mathbf{X}$ be a linear space endowed with separative topology and $\mathbf{Y}$ a linear space. Moreover, assume $\mathbf{X}$ is compact. Let $f: \mathbf{X} \times \mathbf{Y} \rightarrow \mathbb{R}$ satisfy that $f (x,y)$ is convex in $y$ for each fixed $x$, and concave in $x$ for each fixed $y$. Furthermore, $f (x,y)$ is continuous in $x$ for each fixed $y$. If $\sup_{x \in \mathbf{X}} \, \inf_{y \in \mathbf{Y}} \, f (x,y)$ is finite, then
\begin{equation*}
\sup _{x \in \mathbf{X}} \, \inf _{y \in \mathbf{Y}} \, f(x, y)=\inf _{y \in \mathbf{Y}} \, \sup_{x \in \mathbf{X}} \, f(x, y).
\end{equation*}
\end{theorem}
\,
We finish the proof by taking $\mathbf{X} = K \subset D \left( [0,T], \mathscr{M}\right),$ \,$\mathbf{Y} = \mathscr{G} \times \mathscr{G}_0$ and
\begin{align*}
f\left(\mu, (G, \gamma)\right)&= - \ell_T (\mu) + \rho (1-\rho) \int_0^T \left(G_t (0) - G_t (1) \right)^2 dt \\
&+\rho (1-\rho) \int_0^T \int_0^1 \left( \partial_u G_t (u) \right)^2 \,d u \,d t - \mu_0 (\gamma) + \frac{\rho (1-\rho)}{2} \int_0^1 \gamma (u)^2 \,d u
\end{align*}
for any $\mu\in \mathbf{X}$ and $(G, \gamma)\in \mathbf{Y}$.
\end{proof}
\,
To extend the moderate deviations upper bound to any closed set, it suffices to show the exponential tightness of the sequence $\{Q_N\}_{ N \geq 1}$, which follows from the following Lemma as in \cite{gao2003moderate}.
\,
\begin{lemma}\label{exptight}
For any $G \in \mathscr{G}_0$,
\begin{equation}\label{b1}
\limsup _{A \rightarrow \infty} \, \limsup _{N \rightarrow \infty}\, \frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho}\left(\sup _{0 \leqslant t \leqslant T}\left|\left\langle\mu_{t}^{N}, G\right\rangle\right|>A\right)=-\infty,
\end{equation}
and for any $\epsilon > 0$,
\begin{equation}\label{u0}
\limsup _{\delta \rightarrow 0} \, \limsup _{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho}\left(\sup _{0<t \leqslant \delta}\left|\left\langle\mu_{t}^{N}-\mu_{0}^{N}, G\right\rangle\right|>\varepsilon\right)=-\infty.
\end{equation}
\end{lemma}
\,
We first explain why the above lemma implies exponential tightness. For any $m \in \mathbb{N},\, k \in \mathbb{Z} $ and any $\delta, A > 0$, define
\begin{align*}
B_{k,A} = \left\{ \sup_{0 \leq t \leq T} \left| \mu_t (\theta_k ) \right| \leq A \right\}, \quad
B_{k,m,\delta} = \left\{ \sup_{0 \leq |t -s| \leq \delta} \left| (\mu_t - \mu_s) (\theta_k) \right| \leq \frac{1}{m} \right\}.
\end{align*}
Then by Lemma \ref{exptight}, for any $n > 0$, there exist $A = A (n,k)$ and $\delta = \delta (m,k,n)$ such that
\begin{equation*}
\sup_{N \geq 1} Q^N_\rho \left[ B_{k,A}^c \right] < e^{- (a_N^2/N) n k},\quad
\sup_{N \geq 1} Q^N_\rho \left[ B_{k,m,\delta}^c \right] < e^{- (a_N^2/N) n k m}.
\end{equation*}
Let
$$
\mathcal{K}_n = \left\{ \bigcap_{k \geq 1} B_{k,A(n,k)} \right\} \bigcap
\left\{ \bigcap_{k,\,m \geq 1} B_{k,m,\delta(m,k,n)} \right\}.
$$
It can be checked that $\mathcal{K}_n$ is a compact set for each $n \geq 1$. Moreover, $Q^N_\rho \, [\mathcal{K}_n^c]$ is bounded by a multiple of $\exp \{ - (a_N^2 / N) n\}$. This proves the exponential tightness.
\,
\begin{proof}[Proof of Lemma \ref{exptight}]
We first prove \eqref{b1}. Since \eqref{r2} and \eqref{r4} are bounded, we only need to show that
\begin{equation*}
\limsup _{A \rightarrow \infty}\, \limsup _{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \,\left[ \sup_{0 \leqslant t \leqslant T} \left| \frac{N}{a^2_N} \log M_t^N (G) + \langle \mu^N_0,G\rangle + \int_0^t \langle \mu^N_s, \tilde{\Delta} G\rangle d s \right| > A\right] = - \infty,
\end{equation*}
which is a consequence of
\begin{align}
\limsup _{A \rightarrow \infty} \,\limsup _{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \left[ \sup_{0 \leqslant t \leqslant T} \left| \frac{N}{a_N^2} \log M^N_t (G) \right| > A/3 \right] = -\infty,\label{b4}\\
\limsup _{A \rightarrow \infty}\, \limsup _{N \rightarrow \infty} \,\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \left[ \frac{1}{a_N} \left| \sum_{x \in \mathbb{T}_N} \left( \eta_0 (x) - \rho \right) G \left( x/N \right) \right| > A/3 \right] = -\infty,\label{b5}
\end{align}
and
\begin{equation}
\limsup _{A \rightarrow \infty} \,\limsup _{N \rightarrow \infty}\, \frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \left[ \sup_{0 \leqslant t \leqslant T} \left| \int_0^t \langle \mu^N_s, \tilde{\Delta} G\rangle d s \right| > A/3 \right] = -\infty.\label{b6}
\end{equation}
Notice that \eqref{b6} follows from Lemma \ref{lem3}. To prove \eqref{b4}, without loss of generality, we first remove the modulus since otherwise we can replace $G$ by $- G$. Then
\begin{align*}
&\mathbb{P}^{N}_{\rho} \left[ \sup _{0\leqslant t \leqslant T} \frac{N}{ a_N^2} \log M^N_t (G) > A/3 \right]
= \mathbb{P}^{N}_{\rho} \left[ \sup _{0\leqslant t \leqslant T} M^N_t (G) > \exp \left\{ \frac{A a_N^2 }{3 N} \right\} \right]\\
& \leq 4 \exp \left\{ - \frac{A a_N^2 }{3 N} \right\} \mathbb{E}^N_\rho \left[\left( M_T^N (G) \right)^2\right] \leq 4 \exp \left\{ C \left(||G||_\infty^2 + ||G^\prime||_\infty^2\right) T - \frac{A a_N^2 }{3 N} \right\}.
\end{align*}
This proves \eqref{b4}. For \eqref{b5}, removing the modulus inside the probability as before and then by Chebyshev's inequality,
\begin{align*}
&\frac{N}{a_N^2} \log \mathbb{P}^{N}_{\rho} \left[ \frac{1}{a_N} \sum_{x \in \mathbb{T}_N} \left( \eta_0 (x) - \rho \right) G \left( x/N \right) > A/3 \right] \\
& \leq - A/3 + \frac{N}{a_N^2} \log \mathbb{E}^{N}_{\rho} \left[ \exp \left\{\frac{a_N}{N} \sum_{x \in \mathbb{T}_N} \left( \eta_0 (x) - \rho \right) G \left( x/N \right) \right\} \right] \\
&= - A/3 + \frac{N}{a_N^2} \sum_{x \in \mathbb{T}_N} \log \left( 1 + \frac{C (\rho) a_N^2}{N^2} G(x/N)^2 + \mathcal{O}_G \left( \frac{a_N^3}{N^3} \right) \right)\\
&\leq - A/3 + \frac{C (\rho)}{N} \sum_{x \in \mathbb{T}_N} G(x/N)^2 + \mathcal{O}_G \left( \frac{a_N}{N} \right).
\end{align*}
This proves \eqref{b5} by letting $N \rightarrow \infty$ and then $A \rightarrow \infty$.
Next we prove \eqref{u0}. Fix $A > 0$, which will converge to infinity after $\delta \rightarrow 0,\, N \rightarrow \infty$. From Equations \eqref{r0}-\eqref{r5} with $G$ replaced by $A G$, we only need to prove \eqref{u0} for the following four terms:
$$
\frac{N}{A a_N^2} \log M^N_t (A G), \quad \int_0^t \langle \mu^N_s, \tilde{\Delta} G\rangle ds,
$$
$$ A \int_0^t \frac{1}{2} (\eta_s (-1) - \eta_s (0))^2 \left( G \left(\frac{0}{N}\right) - G \left(\frac{-1}{N}\right) \right)^2 ds,
$$
and
$$
A \int_0^t \frac{1}{2 N} \sum_{x \in \mathbb{T}_N,\atop x \neq -1} (\eta_s (x) - \eta_s (x+1))^2 \left(\nabla_N G \left(\frac{x}{N}\right)\right)^2 d s.
$$
Notice that the proof of \eqref{b4} also applies to the martingale term. The second one follows from Lemma \ref{lem3}. For the last two terms, notice that they are both bounded by $C (G) \delta A$. The proof is complete.
\end{proof}
\section{Lower bound}\label{section of lower bound}
For $f,\,g \in \mathscr{G}_0$, we define
\[
\langle f | g\rangle=2\rho(1-\rho)\left[\big(f(0) - f(1)\big)\big(g(0) - g(1)\big)+ \int_0^1 \partial_u f\,(u) \,\partial_u g \,(u) \,d u\right].
\]
For $f,\,g\in \mathscr{G}$ and $0\leq t\leq T$, we define
\begin{align*}
\langle\langle f, g \rangle\rangle_t = \int_0^t \langle f_s | g_s\rangle \,ds.
\end{align*}
For simplicity, we write $\langle\langle f, g\rangle\rangle_T$ as $\langle\langle f, g\rangle\rangle$. To make $\langle\langle \cdot, \cdot\rangle\rangle$ an inner product, we write $f\simeq g$ if and only if $\langle\langle f-g, f-g \rangle\rangle=0$ and then define $\mathscr{H}$ as the Hilbert space which is the completion of $\mathscr{G}/_{\simeq}$.
For locally square integrable martingales $\{M_t\}_{t\geq 0}$ and $\{N_t\}_{t\geq 0}$, we use $\{\langle M, N\rangle_t\}_{t\geq 0}$ to denote the predictable quadratic-covariation process which is continuous and use $\{[M,N]_t\}_{t\geq 0}$ to denote the optional quadratic-covariation process which satisfies
\[
[M, N]_T= \lim_{\sup(t_{i+1}-t_i)\rightarrow 0}\,\sum_i\big(M_{t_{i+1}}-M_{t_i}\big)\big(N_{t_{i+1}}-N_{t_i}\big) \quad \text{in $L^2$},
\]
where the limit is over all partitions $\{t_i\}$ of $[0, T]$. Note that $[M, N]=\langle M, N \rangle$ when $M$ and $N$ are continuous. For any $H\in C^{1,2}([0, T]\times \{0,1\}^{\mathbb{T}_N})$, by Dynkin's martingale formula,
\begin{equation}\label{martingale2}
\Lambda_t^N(H) :=H(t, \eta_t)-H(0, \eta_0)-\int_0^t (\mathcal{L}_N+\partial_s)H(s, \eta_s)\,ds
\end{equation}
is a martingale and for any $H_1, H_2\in C^{1,2}([0, T]\times \{0,1\}^{\mathbb{T}_N})$,
\begin{align}\label{equ 5.0}
\langle \Lambda^N(H_1), \Lambda^N(H_2) \rangle_t =\int_0^t\mathcal{L}_N\big(H_1H_2\big)-H_1\mathcal{L}_NH_2
-H_2\mathcal{L}_NH_1 \,ds.
\end{align}
The following lemma gives clear expressions of $I_{dyn}$ and $I_{ini}$.
\,
\begin{lemma}\label{lemma 5.3}
(i) If $I_{dyn}\, (\mu)<+\infty$, then there exists $\psi\in \mathscr{H}$ such that $\ell_T(\mu, G)=\langle\langle G,\psi \rangle\rangle$ for any $G\in \mathscr{G}$ and $I_{dyn}\, (\mu)=\frac{1}{2}\langle\langle \psi,\psi \rangle\rangle$.\\
(ii) If $I_{ini}\,(\nu)<+\infty$ for $\nu \in \mathscr{M}$, then there exists $\phi\in L^2[0,1]$ such that $\nu(G)=\langle \phi,G\rangle$ for any $G \in \mathscr{G}_0$ and
\[
I_{ini}\,(\nu)=\frac{\int_0^1\phi^2(u)\,du}{2\rho(1-\rho)}.
\]
\end{lemma}
\proof
The proofs of the two parts follow the same strategy, hence we only give the proof of $(i)$. According to the definition of $I_{dyn}$,
\[
I_{dyn} (\mu)=\sup_{G\in \mathscr{G}} \left\{\ell_{T}(\mu, G)-(1/2)\langle\langle G,G \rangle\rangle \right\}.
\]
If $\ell_{T}(\mu, G)\neq 0$ for some $G$ such that $\langle\langle G,G \rangle\rangle=0$, then
\[
I_{dyn}\,(\mu)\geq \sup_{c\in \mathbb{R}} \left\{ \ell_{T}(\mu, cG)-\frac{1}{2} \langle\langle cG, cG \rangle\rangle \right\}=\sup_{c\in \mathbb{R}} \left\{ c \, \ell_{T}(\mu, G) \right\}=+\infty,
\]
which is contradictory. Therefore, $\ell_{T}(\mu, \cdot)$ is well defined on $\mathscr{G}/_{\simeq}$. For $G\in \mathscr{G}/_{\simeq}$ such that $G\neq 0$,
$\ell_{T}(\mu, cG)-(1/2)\langle\langle cG, cG\rangle\rangle$ obtains maximum $\frac{\ell^2_T(\mu,G)}{2\langle\langle G,G\rangle\rangle}$ at $c=\frac{\ell_T(\mu, G)}{\langle\langle G,G\rangle\rangle}$. Therefore,
\[
I_{dyn} \,(\mu)=\sup_{G\in \mathscr{G}/_{\simeq}, G\neq 0}\, \frac{\ell^2_T(\mu,G)}{2\langle\langle G,G \rangle\rangle}.
\]
Since $I_{dyn} (\mu)<+\infty$, $\ell_T(\mu, \cdot)$ can be extended to a bounded linear function on $\mathscr{H}$. As a result, the existence of $\psi$ follows from Riesz's representation theorem and $I_{dyn} \,(\mu)=\frac{1}{2}\langle\langle\psi, \psi\rangle\rangle$ follows from Cauchy-Schwartz's inequality.
\qed
\,
For $\phi\in \mathscr{G}_0$ and sufficiently large $N$, denote by $\nu_{_{N,\phi}}$ the product measure on $\{0,1\}^{\mathbb{T}_N}$ with marginals given by
\[
\nu_{_{N,\phi}}\,\{\eta: \eta(x)= 1\}=\rho+\frac{a_N}{N}\,\phi \left(\frac{x}{N} \right),\quad x \in \mathbb{T}_N,
\]
and by $\mathbb{P}^N_\phi$ the law of the process $\{\eta_t\}_{t\geq 0}$ with initial distribution $\nu_{_{N,\phi}}$. For any $G\in \mathscr{G}$, denote by $\widehat{\mathbb{P}}^N_{\phi, G}$ the probability measure on $D \left([0,T],\{0,1\}^{\mathbb{T}_N}\right)$ such that
\[
\frac{d\, \widehat{\mathbb{P}}^N_{\phi, G}}{d\,\mathbb{P}^N_\phi}=M_T^N(G).
\]
\,
\begin{lemma}\label{lemma 5.4}
For any $G\in \mathscr{G}$ and any $\phi\in \mathscr{G}_0$, $\{\mu^N_t\}_{0\leq t\leq T}$ converges in $\widehat{\mathbb{P}}^N_{\phi, G}$-probability to $\mu^G=\{\mu_t^G\}_{0\leq t\leq T}$ as $N \rightarrow \infty$, where $\mu^{G}$ is the unique element in $D\big([0,T], \mathscr{M}\big)$ such that
\begin{equation}\label{equ ODE for linear operator}
\begin{cases}
&\frac{d}{dt}\mu_t^G(h)=\mu_t^G(\tilde{\Delta}h)+\langle h|G_t\rangle, \\
&\mu_0^G(h)=\int_0^1\phi(u)h(u)\, du
\end{cases}
\end{equation}
for any $h\in \mathscr{G}_0$ and $0\leq t\leq T$.
\end{lemma}
\begin{remark}
Intuitively but not rigorously, integrating by parts, $\mu^G_t$ given by \eqref{equ ODE for linear operator} should be a signed measure such that $\mu^G_t(du)=\rho(t,u)\,du$, where $\rho(t,u)$ is the solution to the PDE
\[
\begin{cases}
&\partial_t\rho(t,u)=\tilde{\Delta}\rho(t,u)-2\rho(1-\rho)\tilde{\Delta}G_t(u),\\
&\rho(0,u)=\phi(u), ~0\leq u\leq 1, \\
&\rho(t, \cdot)\in \mathscr{G}_0, ~0\leq t\leq T.
\end{cases}
\]
However, as we have discussed in Remark \ref{remark 1.2}, we do not manage to prove the uniqueness or existence to this PDE. That's why we only consider $\mu^G$ as the solution to an equation on $\mathscr{M}$ the space of linear functionals on $\mathscr{G}_0$, the uniqueness and existence of which we can show rigorously.
\end{remark}
\,
To prove Lemma \ref{lemma 5.4}, we need some preparation. For $h\in \mathscr{G}$, we write
$$
F_h(t, \eta)=\frac{1}{a_N}\sum_{x \in \mathbb{T}_N}\,(\eta(x)-\rho)\,h_t\left(\frac{x}{N}\right)
$$
and hence
\[
\Lambda^N_t(F_h)=\langle\mu^N_t, h_t\rangle-\langle\mu^N_0, h_0\rangle-\int_0^t (\mathcal{L}_N+\partial_s)\langle\mu_s^N, h_s\rangle \,ds.
\]
\,
\begin{lemma}\label{lemma 5.5}
For any $\phi\in \mathscr{G}_0$, $G, h\in \mathscr{G}$,
\begin{equation}\label{equ 5.3}
[\Lambda^N(F_h), \Lambda^N(F_h)]_T=o_{exp}(a_N^2)
\end{equation}
under both $\mathbb{P}^N_\phi$ and $\widehat{\mathbb{P}}^N_{\phi, G}$.
\end{lemma}
\proof
We first show that Equation \eqref{equ 5.3} holds under $\mathbb{P}^N_\phi$. According to the definition of $\Lambda^N_t(F_h)$,
\[
[\Lambda^N(F_h), \Lambda^N(F_h)]_T=\sum_{0\leq s\leq T}\big[\langle \mu_s^N, h_s\rangle-\langle\mu_{s-}^N, h_{s-}\rangle\big]^2.
\]
Recall that $\{Y_i (\cdot)\}_{i \in \mathbb{T}_N}$ are independent Poisson processes. If $s$ is an event moment of $Y_i(\cdot)$, then
\[
\langle\mu_s^N, h_s\rangle-\langle\mu_{s-}^N, h_{s-}\rangle=\frac{1}{a_N}\left(h_{s-}\left(\frac{i+1}{N}\right)-h_{s-}\left(\frac{i}{N}\right)\right)\big(\eta_{s-}(i)-\eta_{s-}(i+1)\big).
\]
Consequently, let $C_h=\sup\limits_{0\leq t\leq T, \atop 0\leq u \leq 1}|h(t,u)|$ and $D_h=\sup\limits_{0\leq t\leq T, \atop 0\leq u\leq 1}|\partial_u h\,(t,u)|$, then
\[
[\Lambda^N(F_h), \Lambda^N(F_h)]_T\leq \frac{4}{a_N^2}\sum_{i\neq -1}\frac{Y_i(T)}{N^2}\,D_h^2+\frac{16}{a_N^2}\,C_h^2\,Y_{-1}(T)
\]
according to Lagrange's mean value theorem and the fact that there is at most one particle per site. By Chebyshev's inequality, for any $\theta>0$,
\begin{align*}
& \mathbb{P}^N_\phi\left([\Lambda^N(F_h), \Lambda^N(F_h)]_T\geq \epsilon\right) \leq \exp \left\{-\theta a_N^2\epsilon\right\} \,E\,\left[ \exp \left\{4\theta\sum_{i\neq -1}\frac{Y_i(T)}{N^2}D_h^2
+16\, \theta \,C_h^2\,Y_{-1}(T) \right\} \right] \\
&=\exp \left\{-\theta a_N^2\epsilon\right\} \left[\exp \left\{ N^2T \left(\exp \left\{\frac{4\theta D_h^2}{N^2}\right\}-1 \right) \right\}\right]^{N-1} \exp \left\{NT \left( \exp \{16\theta C_h^2\}-1 \right) \right\}.
\end{align*}
Then
\[
\limsup_{N\rightarrow+\infty}\frac{1}{a_N^2}\log \mathbb{P}^N_\phi\big([\Lambda^N(F_h), \Lambda^N(F_h)]_T\geq \epsilon\big)\leq -\theta \epsilon.
\]
This proves Equation \eqref{equ 5.3} under $\mathbb{P}^N_\phi$ since $\theta$ is arbitrary.
Now we only need to show that Equation \eqref{equ 5.3} holds under $\widehat{\mathbb{P}}^N_{\phi, G}$. According to the definition of $\widehat{\mathbb{P}}^N_{\phi, G}$ and Cauchy-Schwartz inequality, for any $\epsilon>0$,
\[
\widehat{\mathbb{P}}^N_{\phi, G}\big([\Lambda^N(F_h), \Lambda^N(F_h)]_T\geq \epsilon\big)
\leq \sqrt{\mathbb{P}^N_\phi\big([\Lambda^N(F_h), \Lambda^N(F_h)]_T\geq \epsilon\big)}\sqrt{\mathbb{E}_\phi^N\left[\left(M_T^N(G)\right)^2\right]}.
\]
Recall the expressions of $M_t^N (G)$ given in \eqref{r0}-\eqref{r5}. It is not difficult to check that there exists a finite constant $C$ independent of $N$ such that $M_T^N(G)\leq e^{C\,a_N}$ for sufficiently large $N$. Therefore, Equation \eqref{equ 5.3} also holds under $\widehat{\mathbb{P}}^N_{\phi, G}$.
\qed
\,
\proof[Proof of Lemma \ref{lemma 5.4}]
The existence and uniqueness of Equation \eqref{equ ODE for linear operator} are given in the appendix. It remains to show that $\mu^N$ converges weakly under $\widehat{\mathbb{P}}^N_{\phi, G}$ to this unique solution $\mu^G$ as $N \rightarrow \infty$. To achieve this purpose, we need to investigate the martingale $\{M_t^N(G)\}_{t\geq 0}$ in \eqref{martingale1} and to utilize a generalized version of Girsanov's theorem introduced in \cite{Schuppen1974} by Schuppen and Wong.
Recall the definition of $\Lambda_t^N(f)$ in \eqref{martingale2} and that for any $G \in \mathscr{G}$,
$$
f_G (t,\eta) = \exp \left\{\frac{a_N}{N}\sum_{i \in \mathbb{T}_N} (\eta(i)-\rho)G_t\left(\frac{i}{N}\right)\right\}.
$$
According to Ito's formula,
\[
dM_t^N(G)=f_G (0, \eta_0)^{-1} \exp \left\{-\int_0^t\frac{(\partial_s+\mathcal{L}_N)f_G}{f_G} (s,\eta_s) \,ds\right\} d\Lambda_t^N(f_G).
\]
For any $t\geq 0$, let
\[
\widetilde{\Lambda}_t^N(f_G)=\int_0^t \frac{1}{f_G (s-, \eta_{s-})}\,d\Lambda_s^N(f_G),
\]
then
\begin{equation}\label{equ 5.5}
dM_t^N(G)=M_{t-}^N(G)\,d\widetilde{\Lambda}_t^N(f_G).
\end{equation}
\,
For any local martingale $\{M_t\}_{t\geq 0}$ under $\mathbb{P}_\phi^N$, let
$$
\widehat{M}_t=M_t- \left\langle M, \widetilde{\Lambda}^N(f_G) \right\rangle_t.
$$
By Equation \eqref{equ 5.5} and the generalized version of Girsanov's theorem \cite{Schuppen1974}, $\{\widehat{M}_t\}_{t\geq 0}$ is a local martingale under $\widehat{\mathbb{P}}^N_{\phi, G}$ and $[\widehat{M}, \widehat{M}] =[M, M]$ under both $\mathbb{P}_\phi^N$ and $\widehat{\mathbb{P}}^N_{\phi, G}$. Therefore, for any $h\in \mathscr{G}$,
\begin{align*}
\langle \mu_t^N, h_t\rangle=&\langle \mu_0^N, h_0\rangle+\int_0^t (\partial_s+\mathcal{L}_N)\langle\mu_s^N, h_s\rangle ds+\widehat{\Lambda_t^N(F_h)}+\langle\Lambda^N(F_h), \widetilde{\Lambda}^N(f_G)\rangle_t,
\end{align*}
where $\left\{\widehat{\Lambda_t^N(F_h)}\right\}_{t\geq 0}$ is a local martingale under $\widehat{\mathbb{P}}^N_{\phi, G}$ with
\[
\left[\widehat{\Lambda^N(F_h)}, \widehat{\Lambda^N(F_h)}\right]=\left[\Lambda^N(F_h), \Lambda^N(F_h)\right].
\]
Then, by Lemma \ref{lemma 5.5} and Doob's inequality, $\widehat{\Lambda_t^N(F_h)}=o_p(1)$ under $\widehat{\mathbb{P}}^N_{\phi, G}$ and hence
\[
\langle \mu_t^N, h_t\rangle=\langle \mu_0^N, h_0\rangle+\int_0^t (\partial_s+\mathcal{L}_N)\langle\mu_s^N, h_s\rangle ds+o_p(1)+\langle \Lambda^N(F_h), \widetilde{\Lambda}^N(f_G) \rangle_t
\]
under $\widehat{\mathbb{P}}^N_{\phi, G}$.
\,
Next we calculate $\langle \Lambda^N(F_h), \widetilde{\Lambda}^N(f_G) \rangle_t$. According to the definition of $\widetilde{\Lambda}^N_t(f_G)$ and \eqref{equ 5.0},
\[
d \left\langle \Lambda^N(F_h), \widetilde{\Lambda}^N(f_G) \right\rangle_t=\frac{1}{f_G(t, \eta_t)}d \left\langle \Lambda^N(F_h), \Lambda^N(f_G) \right\rangle_t,
\]
where
\begin{align*}
d \langle \Lambda^N(F_h), \Lambda^N(f_G) \rangle _t=\big(\mathcal{L}_N\big(F_hf_G\big)-f_G\mathcal{L}_NF_h-F_h\mathcal{L}_Nf_G\big)\,dt.
\end{align*}
By direct calculations,
\begin{align*}
\frac{1}{f_G}\Big(\mathcal{L}_N\big(F_hf_G\big)-f_G\mathcal{L}_NF_h-F_h\mathcal{L}_Nf_G\Big)={\rm \uppercase\expandafter{\romannumeral1}}_N
+{\rm \uppercase\expandafter{\romannumeral2}}_N,
\end{align*}
where
\[
\begin{aligned}
&{\rm \uppercase\expandafter{\romannumeral1}}_N
=\frac{N^2}{a_N}\sum_{i\neq-1}\big(\eta_t(i+1)-\eta_t(i)\big)\left(h_t\left(\frac{i}{N}\right)-h_t\left(\frac{i+1}{N}\right)\right) \\
& \quad \quad \quad \times \left(\exp \left\{\frac{a_N}{N}\big(\eta_t(i+1)-\eta_t(i)\big)\left(G_t\left(\frac{i}{N}\right)-G_t\left(\frac{i+1}{N}\right)\right)\right\}-1\right)
\end{aligned}
\]
and
\[
\begin{aligned}
&{\rm \uppercase\expandafter{\romannumeral2}}_N
=\frac{N}{a_N}\big(\eta_t(0)-\eta_t (-1)\big) \left(h_t\left(\frac{-1}{N}\right)-h_t\left(\frac{0}{N}\right)\right) \\
& \quad \quad \quad \times \left(\exp \left\{\frac{a_N}{N}\big(\eta_t(0)-\eta_t(-1)\big)\left(G_t\left(\frac{-1}{N}\right)-G_t\left(\frac{0}{N}\right)\right)\right\}-1\right).
\end{aligned}
\]
By Taylor's expansion formula up to second order,
\[
{\rm \uppercase\expandafter{\romannumeral1}}_N=\frac{1}{N}\sum_{i\neq -1}\big(\eta_t(i)-\eta_t(i+1)\big)^2\,\partial_uh_t\left(\frac{i}{N}\right)\partial_uG_t\left(\frac{i}{N}\right)+o(1)
\]
and
\[
{\rm \uppercase\expandafter{\romannumeral2}}_N=\big(\eta_t(-1)-\eta_t(0)\big)^2\big(h_t(0)-h_t(1)\big)\big(G_t(0)-G_t(1)\big)+o(1).
\]
Since
$
\big(\eta_t(i)-\eta_t(i+1)\big)^2=\eta_t(i)(1-\eta_t(i+1))+\eta_t(i+1)(1-\eta_t(i))
$,
Lemmas \ref{lem1} and \ref{lem2} control the errors when we replace $\big(\eta_t(i)-\eta_t(i+1)\big)^2$ by $2\rho(1-\rho)$ in ${\rm \uppercase\expandafter{\romannumeral1}}_N$ and ${\rm \uppercase\expandafter{\romannumeral2}}_N$. To be precise, under $\mathbb{P}_\rho^N$,
\begin{equation}\label{equ 5.6}
\int_0^T {\rm \uppercase\expandafter{\romannumeral1}}_N \,dt
=2\rho(1-\rho)\int_0^T\,\int_0^1\partial_uh_t\,(u)\,\partial_uG_t(u)\,du\,dt+o(1)+o_{exp}(a_N)
\end{equation}
and
\begin{equation}\label{equ 5.7}
\int_0^T {\rm \uppercase\expandafter{\romannumeral2}}_N\,dt=2\rho(1-\rho)\int_0^T \big(h_t(0)-h_t(1)\big)\big(G_t(0)-G_t(1)\big)dt+o(1)+o_{exp}(a_N).
\end{equation}
By Taylor's expansion formula, it is not difficult to show that there exists a finite constant $C$ independent of $N$ such that
$\frac{d\,\nu_{_{N,\phi}}}{d\,\nu_\rho}\leq e^{C a_N}$ for sufficiently large $N$. Therefore, $\frac{d\widehat{\mathbb{P}}^N_{\phi, G}}{d\mathbb{P}^N_\rho}\leq e^{C a_N}$ for large $N$. By Cauchy-Schwartz inequality, Equations \eqref{equ 5.6} and \eqref{equ 5.7} also hold under $\widehat{\mathbb{P}}^N_{\phi, G}$. As a result, under $\widehat{\mathbb{P}}^N_{\phi, G}$,
\begin{align*}
&\langle \mu_t^N, h_t\rangle
=o_p(1)+\langle \mu_0^N, h_0\rangle+\int_0^t (\partial_s+\mathcal{L}_N)\langle\mu_s^N, h_s\rangle ds\\
&+2\rho(1-\rho)\Bigg(\int_0^t \big(h_s(0)-h_s(1)\big)\big(G_s(0)-G_s(1)\big)ds
+\int_0^t\,\int_0^1\,\partial_uh_s\,(u)\,\partial_uG_s\,(u)\,du\,ds\Bigg).
\end{align*}
\,
Now we calculate $(\partial_s+\mathcal{L}_N)\langle\mu_s^N, h_s\rangle$. By direct calculations,
\begin{align*}
\mathcal{L}_N\langle\mu_s^N, h_s\rangle=&\frac{N^2}{a_N}\sum_{i=0}^{N-1}(\eta_s(i)-\rho)
\left(h_s\left(\frac{i+1}{N}\right)+h_s\left(\frac{i-1}{N}\right)-2h_s\left(\frac{i}{N}\right)\right)\\
&+\frac{N-N^2}{a_N}\big(\eta_s(0)-\eta_s(-1)\big) \left(h_s\left(\frac{-1}{N}\right)-h_s(0)\right)={\rm \uppercase\expandafter{\romannumeral3}}_N+{\rm \uppercase\expandafter{\romannumeral4}}_N,
\end{align*}
where
\[
{\rm \uppercase\expandafter{\romannumeral3}}_N=\frac{N^2}{a_N}\sum_{i\neq 0,-1}\big(\eta_s(i)-\rho\big)\left(h_s\left(\frac{i+1}{N}\right)+h_s\left(\frac{i-1}{N}\right)-2h_s\left(\frac{i}{N}\right)\right)
\]
and
\begin{align*}
{\rm \uppercase\expandafter{\romannumeral4}}_N=&\frac{N}{a_N}(\eta_s(0)-\eta_s(-1))\left(h_s\left(\frac{-1}{N}\right)-h_s(0)\right)
+\frac{N^2}{a_N}(\eta_s(0)-\rho)\left(h_s\left(\frac{1}{N}\right)-h_s(0)\right)\\
&+\frac{N^2}{a_N}(\eta_s(-1)-\rho)\left(h_s\left(\frac{-2}{N}\right)-h_s(-1)\right).
\end{align*}
By Taylor's expansion formula up to third order,
$$
{\rm \uppercase\expandafter{\romannumeral3}}_N=\langle\mu_s^N, \tilde{\Delta}h_s\rangle+o(1).
$$
Since $h\in \mathscr{G}$, it is not difficult to check that ${\rm \uppercase\expandafter{\romannumeral4}}_N=o(1)$.
\,
In conclusion, we have shown that under $\widehat{\mathbb{P}}^N_{\phi, G}$,
\begin{align*}
&\langle \mu_t^N, h_t\rangle
=o_p(1)+\langle \mu_0^N, h_0\rangle+\int_0^t \langle \mu_s^N, (\partial_s+\tilde{\Delta})h_s\rangle ds\\
&+2\rho(1-\rho)\Bigg(\int_0^t \big(h_s(0)-h_s(1)\big)\big(G_s(0)-G_s(1)\big)dt
+\int_0^t\,\int_0^1\,\partial_uh_s(u)\,\partial_uG_s(u)\,du\,ds\Bigg)\\
&=o_p(1)+\langle \mu_0^N, h_0\rangle+\int_0^t \langle \mu_s^N, (\partial_s+\tilde{\Delta})h_s\rangle ds+\int_0^t\langle h_s|G_s\rangle ds.
\end{align*}
Specially, when $h\in \mathscr{G}_0$,
\[
\langle \mu_t^N, h\rangle
=o_p(1)+\langle \mu_0^N, h\rangle+\int_0^t \langle \mu_s^N, \tilde{\Delta}h\rangle \,ds+ \int_0^t\langle h|G_s\rangle \,ds
\]
for all $0\leq t\leq T$. Note that although the $o_p(1)$ term in the above equation is given for each $t$, it is easy to check that this $o_p(1)$ term can be chosen uniformly for $0\leq t\leq T$.
Since
\[
\mu_t^G(h)=\int_0^1 \phi(u)h(u)\,du+\int_0^t \mu_s^G(\tilde{\Delta}h)\,ds+\int_0^t\langle h|G_s\rangle \,ds,
\]
by Grownwall's inequality,
\[
\big|\langle \mu_t^N, \theta_m\rangle-\mu^G(\theta_m)\big|\leq \Big(o_p(1)+\big|\int_0^1 \phi(x)\theta_m(x)dx-\langle \mu_0^N, \theta_m\rangle\big|\Big)e^{|e_m|t}
\]
for all $0\leq t\leq T$ and $m\geq 1$.
Therefore, to show that $\mu^N$ converges in $\widehat{\mathbb{P}}^N_{\phi, G}$-probability to $\mu^G$ in $D\left([0,T],\mathscr{M}\right)$, we only need to show that
\begin{equation}\label{equ 5.8}
\langle \mu_0^N, h\rangle=\int_0^1 \phi(u)h(u)\,du+o_p(1)
\end{equation}
under $\widehat{\mathbb{P}}^N_{\phi, G}$ for any $h\in \mathscr{G}_0$. According to the definition of $\nu_{_{N, \phi}}$ and Chebyshev's inequality, it is easy to check that Equation \eqref{equ 5.8} holds under $\mathbb{P}_\phi^N$. Since $M_0^N(G)=1$, $\mu_0^N$ has the same distribution under $\mathbb{P}_\phi^N$ and $\widehat{\mathbb{P}}^N_{\phi, G}$. This finishes the proof.
\qed
\,
\proof[Proof of the lower bound]
If $\inf_{\mu \in O} I (\mu) =+\infty$, then Equation \eqref{equ lower bound} holds trivially. So we only need to deal with the case where
$\inf_{\mu \in O} I (\mu)<+\infty$. For given $\epsilon>0$, there exists $\mu^\epsilon\in O$ such that
\[
I_{ini}(\mu^\epsilon_0)+I_{dyn} (\mu^\epsilon)\leq \inf_{\mu \in O}\, I (\mu)+\epsilon.
\]
By Lemma \ref{lemma 5.3}, there exists $\phi^\epsilon\in L^2[0,1]$ and $\psi^\epsilon\in \mathscr{H}$ such that
$$
\mu_0^\epsilon\,(G)=\langle\phi^\epsilon,G\rangle, \,\forall G \in \mathscr{G}_0, \quad I_{ini}\,(\mu_0^\epsilon)=\frac{\int_0^1(\phi^\epsilon(u))^2\,du}{2\rho(1-\rho)},
$$
and
$$
\ell_{T}(\mu^\epsilon, G)=\langle\langle G, \psi^\epsilon \rangle\rangle,\,\forall G\in \mathscr{G}, \quad
I(\mu^\epsilon)=\frac{1}{2}\langle\langle\psi^\epsilon, \psi^\epsilon\rangle\rangle.
$$
Let $G\in \mathscr{G}$ such that $G_t=b_th$ for some $h\in \mathscr{G}_0$ and $b\in C^1[0,T]$. By the above formula and \eqref{l_T},
\[
b_T\mu_T^\epsilon(h)-b_0\mu_0^\epsilon(h)-\int_0^T b^\prime(s)\mu^\epsilon_s(h)ds=\int_0^Tb (s)\Big(\mu_s^\epsilon(\tilde{\Delta}h)+\langle h|\psi_s^\epsilon\rangle\Big) ds.
\]
Since $b$ is arbitrary, according to the formula of integration by parts, $\{\mu_t^\epsilon(h)\}_{0\leq t\leq T}$ is absolutely continuous and
\begin{equation}\label{equ 5.10}
\begin{cases}
&\frac{d}{dt}\mu_t^\epsilon(h)=\mu_t^{\epsilon}(\tilde{\Delta}h)+\langle h|\psi_t^\epsilon\rangle,\\
&\mu_0^\epsilon(h)=\int_0^1\phi^\epsilon(u)h(u) \,du
\end{cases}
\end{equation}
for any $h\in \mathscr{G}_0$.
\,
Since $\mathscr{G}_0$ is dense in $L^2[0,1]$ by Lemma \ref{lemma 1.1.1 topology} and $\mathscr{G}$ is dense in $\mathscr{H}$, there exist $\phi_n\in \mathscr{G}_0$ and $\psi_n\in \mathscr{G}$ such that
$\phi_n$ converges to $\phi^\epsilon$ in $L^2[0,1]$ and $\psi_n$ converges to $\psi^\epsilon$ in $\mathscr{H}$ as $n\rightarrow\infty$. Let $\mu_n\in D\big([0,T], \mathscr{M}\big)$ such that $\mu_{n,0}(G)=\langle \phi_n,G\rangle$ for any $G \in \mathscr{G}_0$, and $\ell_{T}(\mu_n, G)=\langle \langle G, \psi_n\rangle \rangle$ for any $G\in \mathscr{G}$. According to an analysis similar with that leading to Equation \eqref{equ 5.10}, $\mu_n$ is the solution to the Equation
\begin{equation}\label{equ 5.11}
\begin{cases}
&\frac{d}{dt}\mu_{n,t}(h)=\mu_{n,t}(\tilde{\Delta}h)+\langle h|\psi_{n,t}\rangle, \\
&\mu_{n,0}(h)=\int_0^1 \phi_n(u)h(u) \,du
\end{cases}
\end{equation}
for any $h\in \mathscr{G}_0$.
By Lemma \ref{lemma 5.3}, $$
I_{ini}(\mu_{n,0})=\frac{\int_0^1(\phi_n(u))^2\,du}{2\rho(1-\rho)}
$$
and
$$
I_{dyn}\,(\mu_n)=\frac{1}{2} \langle\langle \psi_n, \psi_n \rangle\rangle =\ell_T(\mu_n, \psi_n)-\frac{1}{2} \langle\langle \psi_n, \psi_n \rangle\rangle.
$$
By \eqref{equ 5.10}, \eqref{equ 5.11} and Grownwall's inequality, for any $0\leq t\leq T$ and any integer $k$,
\[
\big|\mu_{n,t}(\theta_k)-\mu_t^\epsilon(\theta_k)\big|\leq \Big|\int_0^1 (\phi^\epsilon(u)-\phi_n(u))\theta_k(u)\,du+\langle \langle \theta_k, \psi^\epsilon-\psi^n\rangle\rangle\Big|e^{|e_k|t}.
\]
Consequently, $\mu_n$ converges to $\mu^\epsilon$ in $D \left([0,T],\mathscr{M}\right)$ and
\[
\lim_{n\rightarrow+\infty}\big(I_{dyn}\,(\mu_n)+I_{ini}(\mu_{n,0})\big)=I_{dyn}\,(\mu^\epsilon)+I_{ini}(\mu_0^\epsilon).
\]
Hence, there exists $m\geq 1$ such that $\mu_m\in O$ and
\[
I_{dyn}\,(\mu_m)+I_{ini}(\mu_{m,0})\leq I_{dyn}\,(\mu^\epsilon)+I_{ini}(\mu_0^\epsilon)+\epsilon.
\]
Let $D_{\epsilon}=\big\{\mu:~|\ell_T(\mu, \psi_m)-\ell_T(\mu_m, \psi_m)|<\epsilon\big\}\bigcap O$, then by Lemma \ref{lemma 5.4} and Equation \eqref{equ 5.11}, $\mu^N$ converges in $\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}$-probability to $\mu_m$ as $N\rightarrow+\infty$ and hence
\begin{equation*}
\lim_{N\rightarrow+\infty}\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}\big(\mu^N\in D_{\epsilon}\big)=1.
\end{equation*}
According to the expression of $M_T^N(G)$ given in Equation \eqref{r5} and Lemmas \ref{lem1} and \ref{lem2},
\[
M_T^N(\psi_m)=\exp \left\{\frac{a_N^2}{N}\Big(\ell_T(\mu^N, \psi_m)-\frac{1}{2} \langle\langle \psi_m, \psi_m \rangle\rangle +o(1)+\widehat{\varepsilon}_N\Big)\right\},
\]
where $\widehat{\varepsilon}_N=o_{exp}(a_N)$ under $\mathbb{P}_\rho^N$.
As we have shown above, $\frac{d\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}}{d\mathbb{P}_\rho^N}\leq e^{C a_N}$ for sufficiently large $N$, hence $\widehat{\varepsilon}_N=o_{exp}(a_N)$ under $\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}$.
According to the definition of $\nu_{_{N,\phi_m}}$, Chebyshev's inequality and Taylor's expansion formula up to second order, it is not difficult to show that
\[
\frac{d\mathbb{P}_\rho^N}{d\mathbb{P}^N_{\phi_m}}=\exp\left\{-\frac{a_N^2}{N}\left(\frac{\int_0^1 \phi_m^2(u)\,du}{2\rho(1-\rho)}+\widetilde{\varepsilon}_N\right)\right\}
=\exp\left\{-\frac{a_N^2}{N}\left(I_{ini}(\mu_{m,0})+\widetilde{\varepsilon}_N\right)\right\},
\]
where $\widetilde{\varepsilon}_N=o_p(1)$ under $\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}$. Consequently, let
\[
\widehat{D}_{N, \epsilon}=\{\mu^N\in D_{\epsilon}\}\bigcap \{|\widehat{\varepsilon}_N|<\epsilon, |\widetilde{\varepsilon}_N|<\epsilon\},
\]
then
\begin{equation}\label{equ 5.9}
\lim_{N\rightarrow+\infty}\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}\big(\widehat{D}_{N, \epsilon}\big)=1.
\end{equation}
For sufficiently large $N$, on $\widehat{D}_{N, \epsilon}$,
\begin{align*}
&\frac{d\mathbb{P}_\rho^N}{d\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}}=\frac{d\mathbb{P}_\rho^N}{d\mathbb{P}^N_{\phi_m}}\frac{d\mathbb{P}^N_{\phi_m}}{d\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}}=\frac{d\mathbb{P}_\rho^N}{d\mathbb{P}^N_{\phi_m}}\frac{1}{M_T^N(\psi_m)} \\
&\geq \exp\Big\{-\frac{a_N^2}{N}\Big(I_{ini}(\mu_{m,0})+\ell_T(\mu_m, \psi_m)-\frac{1}{2}\langle\langle \psi_m, \psi_m \rangle\rangle+3\epsilon\Big)\Big\} \\
&=\exp\Big\{-\frac{a_N^2}{N}\Big(I_{ini}(\mu_{m,0})+I_{dyn}(\mu_m)+3\epsilon\Big)\Big\}\geq \exp\Big\{-\frac{a_N^2}{N}\Big(I_{ini}(\mu^\epsilon_0)+I_{dyn}(\mu^\epsilon)+4\epsilon\Big)\Big\}\\
&\geq \exp\Big\{-\frac{a_N^2}{N}\Big(\inf_{\mu\in O}\big(I_{ini}(\mu_0)+I_{dyn}(\mu)\big)+5\epsilon\Big)\Big\}.
\end{align*}
Therefore, by Equation \eqref{equ 5.9},
\begin{align*}
&\liminf_{N \rightarrow \infty} \frac{N}{a_N^2} \log Q^N_\rho\, [O]=\liminf_{n\rightarrow+\infty}\,\frac{N}{a_N^2}\log \mathbb{P}_\rho^N\big(\mu_N\in O\big)\geq \liminf_{n\rightarrow+\infty}\,\frac{N}{a_N^2}\log \mathbb{P}_\rho^N\big(\widehat{D}_{N, \epsilon}\big)\\
&=\liminf_{n\rightarrow+\infty}\,\frac{N}{a_N^2}\log \widehat{\mathbb{E}}^N_{\phi_m, \psi_m}\left[\frac{d\mathbb{P}_\rho^N}{d\widehat{\mathbb{P}}^N_{\phi_m, \psi_m}}\mathbf{1}_{\widehat{D}_{N, \epsilon}}\right]\geq -\inf_{\mu\in O}\big(I_{ini}(\mu_0)+I_{dyn}(\mu)\big)-5\epsilon.
\end{align*}
Since $\epsilon$ is arbitrary, the proof is complete.
\qed
\,
| 1129358972a7d12aca17d18df2010dee6b136ed3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
\label{sec:intro}
Planetary nebulae (PNe) represent a short-lived configuration of the
circumstellar medium of evolved low- and intermediate-mass stars (1~M$_{\odot} \lesssim M_\mathrm{i} \lesssim$8~M$_{\odot}$).
The interacting stellar winds model of formation predicts that low- and intermediate-mass stars evolve through the asymptotic giant branch (AGB) phase producing a dense and slow wind, which eventually will be swept and compressed by the future fast stellar wind of the post-AGB star and subsequently ionised by the newly developed UV flux \citep[see][]{Kwok1978,Balick1987}.
The wide variety of PN morphologies \citep[round, bipolar, multipolar, irregular, etc; see][]{Sahai2011} has led to the suggestion that several agents might contribute to the PN shaping, including binarity \citep[or multiple systems;][]{Akashi2017}, jets and magnetic fields \citep[see the review by][]{Zijlstra2015}.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.95\linewidth]{Fig3_RGB}
\label{fig:NGC2371_RGB}
\caption{Colour-composite images of NGC\,2371. The wide-field image
(left panel) was produced by combining the CFHT and {\it HST}
observations. The right panels were obtained by only taking the {\it
HST} observations. In all panels red, green and blue correspond to
[N\,{\sc ii}], H$\alpha$ and [O\,{\sc iii}], respectively.}
\end{center}
\end{figure*}
Central stars of PNe (CSPNe) commonly show the presence of hydrogen in their surfaces.
However, about 10\% of CSPNe do not exhibit hydrogen features, i.e., they are hydrogen-deficient, rather displaying emission features in their spectra similar to those of the massive Wolf-Rayet (WR) stars of the carbon sequence \citep[see][and references therein]{Acker2003,Gorny2001}. Accordingly, they are classified in the same way as WR stars, just using square brackets to denote that they are CSPNe instead \citep[][]{Tylenda1993,Crowther1998,Acker2003}.
It is interesting to note that only a few PNe have been reported to harbor
hydrogen-poor [WR] CSPN of the nitrogen sequence: the transitional [WN/WC]-type PB\,8
\citep[][]{Todt2010} and the [WN]-type CSPNe of Abell\,48 \citep[][]{Todt2013a} and IC\,4663 \citep[][]{Misza2012}.
In general, abundance determination of PNe with [WR]-type CSPNe do not present major differences compared to those classical PNe with hydrogen-rich CSPNe \citep[e.g.,][]{Pena1998,Pena2001}.
Hydrogen-deficient material ejected inside the old, hydrogen-rich PN has been identified
only around a few [WR]-type CSPN \citep[see][and references therein]{Guerrero2018}. These are known as \emph{born-again} PNe and it is accepted that the CSPN experienced a very late thermal pulse (VLTP) ejecting hydrogen-depleted and carbon-rich processed material \citep[e.g.,][]{Herwig1999,MillerBertolami2006}.
Although this evolutionary path has been suggested to explain the formation of
hydrogen-poor [WR] stars, it might not apply to all PNe harboring [WR] stars \citep[see][]{Gorny2000}.
Statistical studies of PNe around [WR] CSPNe have shed light into the relation between the nebular properties and the evolutionary sequence of the progenitor stars. For example, the decrease in electron density ($n_\mathrm{e}$) with decreasing spectral subtype has been suggested to be a signature of an evolutionary sequence from the late to early [WC]-type \citep{Acker1996,Gorny2000}. Furthermore, the N/O abundance ratio vs.\ [WC] class presented by \citet{Pena2001} confirmed the claim that stars with different masses evolve through the same [WC] state
\citep[e.g.,][]{Pena1998,DeMarco1999}.
Comparisons between PNe harboring non-emission line CSPNe with those hosting [WR] CSPNe
have unveiled subtle differences.
PNe with [WR] CSPNe are generally more centrally condensed, expand faster, and
exhibit a higher degree of turbulence than PNe with non-emission line CSPNe
\citep[e.g.,][and references therein]{Gesicki2006,Medina2006,Jacob2013}.
They also seem to have higher N and C abundances, implying that they have evolved
from stars with initial masses $\sim$4~M$_{\odot}$ \citep{GarciaRojas2013}.
This is consistent with their preferential location at low latitudes
in the Galactic disk \citep{Pena2013}.
Finally, it has been noted recently that PNe with [WC] CSPNe have a higher
occurrence of fast collimated outflows than PNe with weak emission line CSPNe or non-emission line CSPNe \citep{Rechy2017}.
These statistical studies rely on averaged values of the nebular physical properties
(electron density and temperature), abundances, and velocity fields
that might hide important clues on the evolution of PNe with [WR]-type CSPNe.
For example, dedicated studies of compact [WR] PNe have unveiled the effects of
fast collimated outflows in their early shaping \citep[][]{Rechy2017,Rechy2019}.
Detailed studies of individual PNe with [WR] CSPNe are clearly needed to understand this unique evolutionary path of low- and intermediate-mass stars. We have therefore started a series of detailed studies of PNe harboring [WR] CSPNe.
In this first paper we present a study of NGC\,2371
\citep[a.k.a. PNG\,189.1+19.8;][]{Curtis1918}, which
harbors the [WR] CSPN WD\,0722$+$295.
NGC\,2371 has been classified as a barrel-like PNe of Peimbert type II \citep[see table~1 in][]{Henry2018}.
Morphological studies performed with the \emph{Hubble Space Telescope} (\emph{HST}), \emph{Spitzer} and ground-based telescopes have unveiled a variety of morphological features \citep[see Fig.~1;][]{Sabbadin1982,RamosLarios2012}:
two-bipolar lobes extending $\gtrsim$1~arcmin from the CSPN towards the NW and SE directions with position angle PA$\approx -60^{\circ}$, a main central cavity with an ellipsoidal shape with an extension of 28$^{\prime\prime}\times$38$^{\prime\prime}$ that exhibits a couple of blowout features aligned with the outer lobes, and a collection of dense knots mainly detected in
[N\,{\sc ii}] and [S\,{\sc ii}] narrow-band filter images (see next section).
The two brightest knots are aligned in the NE to SW direction with a PA$\approx60^{\circ}$, but the line that connects them is not perfectly aligned with the CSPN. \citet{RamosLarios2012} presented a detailed discussion on the properties of the dense NE and SW knots. Their \emph{HST} images show that these are actually conglomerates of clumps (see also Fig.~1).
A secondary pair of smaller size clumps is observed at a PA=5$^\circ$.
It is worth noting that the low-ionisation clumps do not have a uniform
distribution around the CSPN nor are they aligned with the bipolar lobes.
Optical \citep{TorresPeimbert1977,Aller1979} and \emph{International Ultraviolet Explorer} (\emph{IUE}) UV \citep{Pottasch1981} observations of NGC\,2371 have been used to estimate an averaged electron temperatures ($T_\mathrm{e}$) in the range 9,000--23,000~K, depending on the line ratio used or the adopted electron density ($n_\mathrm{e}$), which is found to be 800--2,500~cm$^{-3}$ \citep[e.g.,][]{Pena2001}.
\citet{Kaler1993} presented the analysis of the optical spectrum in the
blue region ($\sim$3700--5000~\AA) and showed it to be dominated by the
O\,{\sc vi}~$\lambda\lambda$3811,3834 doublet with contributions from the
He\,{\sc ii}~$\lambda$4686, C\,{\sc iii}~$\lambda$4650 and
C\,{\sc iv}~$\lambda$4658.
These authors estimated an effective temperature ($T_\mathrm{eff}$) for
WD\,0722$+$295 of about $T_\mathrm{eff}$=[1--1.2]$\times$10$^{5}$~K,
consistent with the analysis of \emph{IUE} observations \citep{Pottasch1981}.
Since its early classification as a [WR] star of the oxygen sequence \citep{Smith1969},
its spectral type has swung from early [WC3] \citep[][]{Heap1982} to [WO1] \citep{Acker2003}.
In this work, we intend to unveil the physical structure of the different components
of NGC\,2371 and to produce an accurate model of its central star.
We have obtained Isaac Newton Telescope (INT) spectroscopic observations to study the
physical properties (density and temperature) and abundances to
describe the ionisation structure of the different morphological
components of NGC\,2371.
In conjunction with archival \emph{IUE} observations, the INT spectra are also used to perform a detailed characterisation of its CSPN.
The paper is organized as follows.
In Section~2 we describe the observations used in the present work.
Section~3 presents the analysis and atmospheric model of WD\,0722$+$295.
In Section~4 we present our spectral analysis of NGC\,2371 as well as the estimates of the physical structure.
In Section~5 we discuss our results and, finally, a summary is presented in Section~6.
\section{Observations}
\label{sec:obs}
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=0.95\linewidth]{NGC2371_slits_LR}
\label{fig:NGC2371_slits}
\caption{CFHT colour-composite ([O\,{\sc iii}] - blue, H$\alpha$ - green, [N\,{\sc ii}] - red) image of NGC\,2371. The (green) dashed-line regions represent the slits positions obtained from our INT IDS observations. The PA for the slits are 65$^{\circ}$ (Slit\,A - minor axis) and 128$^{\circ}$ (Slit\,B - major axis). Different extraction regions are labeled and shown with different colours. The green and yellow velocity measurements correspond to the A2, A7 and
B1-B9 extraction regions. See text for details.}
\end{center}
\end{figure}
Narrow-band images of NGC\,2371 were downloaded from the archives of the Canada-France-Hawaii Telescope (CFHT)\footnote{\url{http://www.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/en/}} and \emph{HST} Legacy Archive\footnote{\url{https://hla.stsci.edu/}}. The CFHT images were acquired with the multi-object spectrograph (MOS) camera.
These images were obtained through the H$\alpha$, [O\,{\sc iii}] and [N\,{\sc ii}] filters with total exposure times of 1200~s on each filter on December 18, 2002 (Prop.\,ID.: 02BC17; PI: S.\,Kwok).
The {\it HST} images were obtained through the F502N, F656N, and F658N filters (hereafter [O\,{\sc iii}], H$\alpha$ and [N\,{\sc ii}]) on 2007 November 15 with exposure times on each filter of 1600 s (Prop.ID.\,11093, PI: K.\,Noll). A colour-composite image obtained by combining all these optical images is shown in Figure~1.
Long-slit spectroscopic observations were obtained at the INT of the Observatorio de El Roque de los Muchachos (La Palma, Spain) using the Intermediate Dispersion Spectrograph (IDS) in low- and high-dispersion spectral modes (PI: M.~A.\,Guerrero).
The high-resolution spectra were obtained on 2018 October 24
using the EEV10 CCD camera with the R1200U grating.
This configuration has a dispersion of 0.48~\AA~pixel$^{-1}$ over the 3200--4500 \AA\
wavelength range and a plate scale of 0\farcs44~pixel$^{-1}$.
The low-resolution spectra were acquired on 2018 November 15
using the RED$+$2 CCD camera with the R400V grating.
This configuration has a dispersion of 1.55~\AA~pixel$^{-1}$ over the 3500--8000 \AA\
wavelength range and a similar plate scale of 0\farcs44~pixel$^{-1}$.
The slit width was set at 1\farcs1 in both observing runs.
Two slits at PA=65$^\circ$ (Slit~A) and PA=128$^\circ$ (Slit~B) were used with both spectral configurations. Slit B is centered on WD\,0722$+$295, whereas Slit A probes the low-ionisation knots and therefore it does not goes exactly across the CSPN. The positions of these slits are overlaid on the colour-composite image of NGC\,2371 in Figure~2. The total integration time on each position and spectral configuration was 900~s.
All spectra were analysed following {\sc iraf} standard routines.
The wavelength calibration was performed using CuAr+CuNe lamps.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=1.0\linewidth]{fig32020.pdf}~
\label{fig:NGC2371_spec_star}
\caption{INT IDS low-resolution, normalised spectrum of the CSPN of
NGC\,2371 obtained from slit\,B. The most prominent WR features and nebular lines are marked.}
\end{center}
\end{figure*}
In order to study the ionisation structure of NGC\,2371, we have
extracted spectra from different regions defined on each slit
using the {\sc iraf} task {\it apall}.
Those extracted from Slit~A were used to study the central main cavity of
NGC\,2371 and the dense knots located $\gtrsim$16$^{\prime\prime}$ from
the CSPN, and those from Slit~B to study the outermost lobes and different
regions of the central cavity along the NW-SE direction.
All regions, labeled from A1--A9 and B0--B9, are shown in Figure~2
and their sizes are indicated in Tables 3 and 4, respectively.
We first extracted the 1-D spectrum of the CSPN (regions A4 and B5) tracing the
stellar continuum along the 2-D spectrum.
This information was then used as a reference to trace the nebular spectra (regions A1--A3,
A5--A9, B0--B4, and B6--B9) along the 2-D spectra, as these regions does not show continuum emission.
Since Slit~A misses some stellar flux, the spectrum extracted for region A4
has not been used for stellar modelling.
Finally, we note that all extracted spectra were corrected for extinction by
using the $c$(H$\beta$) value estimated from the Balmer decrement method.
We assume intrinsic Balmer decrement ratio corresponding to a case B
photoionised nebula of $T_{\rm e}$ = 10,000~K and $n_{\rm e}$ = 100~cm$^{-3}$ \citep[see][]{Osterbrock2006} and the reddening curve of \cite{Cardelli1989}.
\section{WD\,0722$+$295 - the CSPN of NGC\,2371}
The optical spectrum of WD\,0722$+$295 obtained from our INT IDS observations
is presented in Figure~3.
The stellar spectrum exhibits the classic WR blue and red bumps (BB and RB) at
$\sim$4686~\AA\ and $\sim$5806~\AA, respectively, as well as a broad O\,{\sc vi}
feature at $\sim$3820~\AA\ \citep[see also figure~1 in][]{Kaler1993}.
Other emission lines related to WR features such as those reported in
\citet{Acker2003} -- namely, O\,{\sc vi} 5290~\AA, C\,{\sc iv} 5470~ \AA\
and 7060~\AA, and He\,{\sc ii} 5412~\AA\ -- are also present in the optical spectrum.
Furthermore, we also detect the O\,{\sc vi} 6200~\AA\, as a broad emission line.
All identified WR features as well as some contributing nebular lines are marked in Figure~3.
We note the presence of narrow emission lines at $\approx$5665~\AA\ and
$\approx$6066~\AA\ marked with an asterisk in Figure~3 and Figure~4.
There has been controversy in the identification of these lines, which can be
assigned either to the O\,{\sc vii} $\lambda$5666 and O\,{\sc viii}
$\lambda$6068 lines or to the Ne\,{\sc vii} $\lambda$5666 and Ne\,{\sc viii}
$\lambda$6068 lines.
This is discussed in Section 5.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=\linewidth]{Fig4}
\label{fig:NGC2371_spec_star_2}
\caption{Multi-component Gaussian fits to the O-bump ({\it top left}), blue bump ({\it top center}) and red bump ({\it top right}) panels. The bottom panels show other WR features in the spectrum of WD\,0722$+$295. The (blue) dashed lines represents the fits to the broad (stellar) features, the (red) dotted-line the fits to the nebular emission and the (black) line the sum. Residuals are shown at the bottom of each panel.}
\end{center}
\end{figure*}
To provide an appropriate description of the spectrum and spectral classification of WD\,0722$+$295, the contribution from nebular lines blended with the broad WR features needs to be accounted for. This was done by applying the analysis described by \citet{GomezGonzalez2020}. This method consists on fitting the broad WR features with multi-Gaussian components using a tailor-made code that uses the {\sc idl} routine {\sc lmfit}\footnote{The {\sc lmfit} function (lmfit.pro) performs a non-linear least squares fit to a function with an arbitrary number of parameters. It uses the Levenberg-Marquardt algorithm, incorporated in the routine {\it mrqmin} from \citet{Press1992}.}. As a result, fluxes, central wavelength, FWHM and equivalent widths (EW) from the WR spectral features as well as contributing nebular lines have been estimated. All these parameters are listed in Table~1. We note that those lines listed as WR features have FWHM larger than those of the nebular narrow lines, that is, FWHM$>4.5$~\AA.
\begin{table}
\begin{center}
\caption{\label{tab:elines} Parameters for the emission lines considered for the
multi-Gaussian fitting of the WR features of the central star in NGC\,2371.}
\setlength{\tabcolsep}{0.6\tabcolsep}
\begin{tabular}{cllrrrrrr}
\hline
ID & &
\multicolumn{1}{c}{Ion} &
\multicolumn{1}{c}{$\lambda_0$} &
\multicolumn{1}{c}{F} &
\multicolumn{1}{c}{FWHM} &
\multicolumn{1}{c}{EW} &
\multicolumn{1}{c}{$\lambda_\mathrm{c}$} &
\multicolumn{1}{c}{$F/F_\mathrm{RB}$} \\
\multicolumn{1}{c}{(1)} &
\multicolumn{1}{c}{(2)} &
\multicolumn{1}{c}{(3)} &
\multicolumn{1}{c}{(4)} &
\multicolumn{1}{c}{(5)} &
\multicolumn{1}{c}{(6)} &
\multicolumn{1}{c}{(7)} &
\multicolumn{1}{c}{(8)} &
\multicolumn{1}{c}{(9)} \\
\hline
BB & & & & & & & & \\
1 & WR & He\,{\sc ii} & 4686 & 11.4 &28.2 &11.2 & 4686.0 & 712.5 \\
2 & WR & C\,{\sc iv} & 4658 & 7.3 &15.9 & 7.0 & 4656.9 & 456.3 \\
3 & Neb& He\,{\sc ii} & 4686 & 5.8 & 4.5 & 5.7 & 4685.2 & $\dots$~~ \\
4 & Neb& Ar\,{\sc iv} & 4711 & 0.9 & 4.5 & 0.9 & 4711.5 & $\dots$~~ \\
RB & & & & & & & & \\
5 & WR & C\,{\sc iv} & 5801 & 1.0 & 7.1 & 3.1 & 5799.7 & $\dots$~~ \\
6 & WR & C\,{\sc iv} & 5812 & 0.6 & 5.4 & 1.9 & 5810.5 & $\dots$~~ \\
OB & & & & & & & & \\
7 & WR & O\,{\sc vi} & 3820 &117.6 &57.5 &53.2 & 3818.7 & 7350.0 \\
8 & Neb& O\,{\sc vi} & 3811 & 8.0 & 4.2 & 3.6 & 3810.9 & $\dots$~~ \\
9 & Neb& O\,{\sc vi} & 3834 & 5.7 & 4.1 & 2.6 & 3834.0 & $\dots$~~ \\
other & & & & & & & & \\
10 & WR & O\,{\sc vi} & 5290 & 1.7 & 9.6 & 3.2 & 5288.7 & 106.3 \\
11 & Neb& He\,{\sc ii} & 5412 & 0.3 & 3.6 & 0.7 & 5411.1 & $\dots$~~ \\
12 & WR & C\,{\sc iv} & 5470 & 0.1 & 4.6 & 0.3 & 5470.4 & 6.3 \\
13 & WR & Ne\,{\sc vii}$^\dagger$& 5666 & 0.3 & 8.2 & 0.8 & 5664.6 & 18.8 \\
14 & WR & Ne\,{\sc viii}$^\dagger$& 6068& 0.4 & 6.4 & 1.6 & 6066.0 & 25.0 \\
15 & WR & O\,{\sc vi} & 6200 & 0.4 &18.0 & 1.4 & 6200.0 & 25.0 \\
16 & WR & C\,{\sc iv} & 7060 & 3.1 &30.3 & 3.5 & 7062.5 & 193.8 \\
\hline
\end{tabular}\\
(1) Identification number of the Gaussian components in the WR features;
(2) Nature of the contributing emission line: WR (broad) or nebular (narrow);
(3) Ion responsible for the line;
(4) Rest wavelength in \AA;
(5) Flux in units of $10^{-14}$~erg~cm$^{-2}$~s$^{-1}$;
(6) Full Width at Half Maximum (FWHM) [\AA];
(7) Equivalent Width (EW) [\AA];
(8) Observed center of the line;
(8) Line fluxes normalised to the RB. The ratio has been computed adopting a $F$(RB)=100.
$^\dagger$This identification is discussed in Section~5.
\end{center}
\end{table}
Figure~4 shows the fits to the different WR features of the WD\,0722$+$295. Two of these bumps are composed of several blended emission lines and need a careful modelling in order to separate their broad and narrow contributions. The O-bump feature is made of a broad O\,{\sc vi} (FWHM of 57.5~\AA), two O\,{\sc vi} narrow lines (FWHM$<$4.5~\AA) and several Ne lines at the red wing (see Table 1).
It is clear that the broad O\,{\sc vi}~3820~\AA\ has a stellar origin and, although the other two O\,{\sc vi} lines have FWHM that suggest a nebular origin, some contribution from the star is expected according to our stellar atmosphere model (see Section~3.1 and Fig.~5). The Gaussian-fitting method shows that the O\,{\sc vi} lines at 3811~\AA\ and 3834~\AA\ contribute about 10\% to the O-bump. Thus, this can be considered as a upper limit to the nebular contribution. Other narrow emission lines also contribute to the BB. For example, the He\,{\sc ii}, which is blended with WR features contributes $\simeq$30\% of the total flux of the BB (He\,{\sc ii}~4686~\AA$+$C\,{\sc iv}~4658~\AA; see Table~2). There is no contribution of nebular lines to the RB (C\,{\sc iv}~5801~\AA\ + C\,{\sc iv}~5812~\AA).
With reliable flux estimates for each WR feature, we can assess the spectral
classification of WD\,0722$+$295.
For this, we estimated the line ratios of oxygen, carbon and helium WR features
relative to the RB (Column~9 of Table~1) and compared them to those defining
the quantitative classification scheme of [WR] stars listed in table~2 of
\citet{Acker2003}. According to this scheme, WD\,0722$+$295 fulfills the criteria for a [WO1]-type star, but we note that the FWHM of the RB is around half of that suggested for this spectral type. One could argue that using dereddened flux ratios is model dependent and less accurate than using EW ratios. For this, we also compared our results listed in Table~1 with the classification scheme proposed in \citet{Crowther1998} and found that the EW ratio of the O\,{\sc vi}~3820~\AA\ WR feature over the RB is 73.5, above the limit of $\sim$1.6 suggested by these authors for a [WO1] spectral type.
\subsection{NLTE analysis of WD\,0722$+$295}
To obtain a more quantitative description of the properties of WD\,0722$+$295, we have analysed optical and UV spectra by means of the updated version of the Potsdam Wolf-Rayet (PoWR)\footnote{\url{http://www.astro.physik.uni-potsdam.de/~wrh/PoWR}} NLTE stellar atmosphere code \citep[][]{Grafener2002,Hamann2004}. Details of the computing scheme can be found in \citet{Todt2015} and the most recent
example of the capabilities of the PoWR code performed by our group can be found in \citet{Toala2019} for the case of the [WR]-type CSPN of NGC\,40.
Available UV \emph{Far Ultraviolet Spectroscopic Explorer} (\emph{FUSE}) and \emph{IUE} observations of WD\,0722$+$295 were retrieved from the Mikulski Archive for Space Telescopes\footnote{\url{https://archive.stsci.edu/hst/}}. The UV observations were analysed in conjunction with the optical INT IDS spectra with the PoWR stellar atmosphere code. The \emph{FUSE} observations correspond to Obs.\,ID. p1330301000 (PI: L.\ Bianchi) and were obtained on 2000 February 26 with the LWRS aperture for a total exposure time of 5259~s. The {\it IUE} data, taken on 1979 April 7, correspond to the Obs.\,ID. SWP04883 and LRW04210 (Program ID SP127; PI: S.\,R.\,Pottasch) with exposure times of 1560~s and 2700~s, respectively. The spectra imply a reddening of $E(B-V)=0.08$~mag using the Cardelli extinction law which is consistent with the value estimated by \citet[][]{Herald2004}.
\begin{table}
\caption{Parameters of WD\,0722$+$295 obtained with PoWR.}
\label{tab:parameters}
\begin{tabular}{lcl}
\hline
Parameter & Value & Comment\\
\hline
$T_\mathrm{eff}$ [kK] & 130 & Defined at $\tau_\text{Ross}=20$ \\
$d$ [kpc] & 1.75 & \citet{bailerjones2018}\\
$\log(L_{\star}/L_\odot)$ & 3.45 & \\
$R_{\star}$ [$R_\odot$] & 0.105 & \\
$R_\text{t}$ [$R_\odot$] & 20\\
$D$ & 10 & Density contrast\\
$\log(\dot{M}/\mathrm{M}_\odot\,\text{yr}^{-1})$ & $\mathrm{-7.75}$ & for $D$=10 \\
$v_{\infty}$ [km\,s$^{-1}$] & 3700 & \\
$M_{\star}$ [M$_\odot$] & 0.6 & Stellar mass adopted \\
\hline
\multicolumn{3}{l}{Chemical abundances (mass fraction)}\\
\hline
He & 0.71 & \\
C & 0.20 & \\
N & 0.001 & \\
O & 0.06 & \\
Ne & 0.03 & \\
Fe & 1.4$\times10^{-3}$ & \\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.85\linewidth]{NGC2371_PoWR_model2.pdf}
\label{fig:PoWR}
\caption{Comparison between our PoWR model (red dashed line) with optical and UV observations (blue solid line).}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.85\linewidth]{NGC2371_PoWR_model3.pdf}
\label{fig:PoWR_2}
\caption{Comparison between our PoWR model (red line) with observed spectra in the optical and UV and IR photometry (blue).}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.85\linewidth]{fig7_2.pdf}
\label{fig:spec_examples}
\caption{Examples of INT IDS spectra of NGC\,2371. The panels show the
spectra extracted from regions A3, A7, B1 and B4. The most prominent lines
are labeled. See Tables~ 3 and 4 for details on the fluxes.}
\end{center}
\end{figure*}
The parameters of our best-fit model are listed in Table~2.
We note that the stellar mass was adopted to be $M_{\star}=0.6$~M$_{\odot}$, a typical value for CSPNe \citep[e.g.,][and references therein]{MillerBertolami2016}. The resultant temperature, defined at $\tau_\text{Ross}=20$, which is constrained by the relative strength of the emission lines, is $T_\mathrm{eff}$=130~kK. A luminosity of $L_{\star}=2820$~L$_{\odot}$ is estimated by using a distance $d=1.75$~kpc \citep{bailerjones2018} and the resultant mass-loss rate is $\dot{M}=1.78\times10^{-8}$~M$_{\odot}$~yr$^{-1}$.
We note that the estimated value for the mass-loss rate is somewhat smaller than those listed for other [WR] stars in \citet{Todt2013b}, who used a canonical luminosity of $L=5000\,L_\odot$ to calculate the mass-loss rates from the $R_\text{t}$ values obtained from the analysis. Taking the lower luminosity of NGC\,2371 into account, our value of $\dot{M}$ is comparable with the one derived for Hen\,2-55. Furthermore, our $R_\text{t}=20\,R_\odot$ is fully compatible with the value of $R_\text{t}=15^{+10}_{-5}\,R_\odot$ by \citet{Herald2004}. They used the same luminosity as in our analysis (based on the distance of $d=1.5\,$kpc), but assumed a smooth wind with $D=1$, where we adopted a density contrast of $D=10$ for a clumped wind. Therefore, their mass-loss rate is about a factor $\sqrt{10}$ larger than ours.
Our best-fit model to the P Cygni profiles in the UV spectra resulted in a stellar wind velocity $v_{\infty}=3700$~km~s$^{-1}$ with a micro-turbulence of less than 3\%. The micro-clumping parameter $D$, which is defined as the
density contrast between a smooth wind and a clumpy wind, is estimated to be 10. Figure~5 shows a comparison between the PoWR model of WD\,0722$+$295 and the optical and UV spectra, and Figure~6 presents the synthetic spectral energy distribution (SED) from the far-UV to the IR in comparison with UV and optical spectra and IR photometry obtained from public archives.
The abundances for WD\,0722$+$295 derived from our best-fit model are also listed in Table~2. Whereas the stellar parameters agree with those reported by \citet[][see their table~6]{Herald2004}, our abundance determinations are at variance, with our He
abundance 30\% larger and our N/O ratio half their value.
The differences can be attributed to the use of different stellar atmosphere codes, as \citet{Herald2004} use a relatively old version of the CMFGEN code \citep[][]{Hillier1998,Hillier1999}, but we note that our abundance determination is improved by the simultaneous modelling of high-quality optical spectra and UV data.
\section{Physical conditions, excitation and chemical abundances of NGC\,2371}
\begin{table*}
\centering
\caption[]{Reddeding-corrected line intensities for selected regions of Slit\,A of NGC\,2371}
\setlength{\columnwidth}{0.4\columnwidth}
\setlength{\tabcolsep}{0.7\tabcolsep}
\begin{tabular}{cccccccccc}
\hline
\multicolumn{1}{c}{$\lambda_{0}$}&
\multicolumn{1}{c}{line}&
\multicolumn{1}{c}{A1}&
\multicolumn{1}{c}{A2}&
\multicolumn{1}{c}{A3}&
\multicolumn{1}{c}{A5}&
\multicolumn{1}{c}{A6}&
\multicolumn{1}{c}{A7}&
\multicolumn{1}{c}{A8}&
\multicolumn{1}{c}{A9}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(5$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(4$''$)}&
\multicolumn{1}{c}{(5$''$)}&
\multicolumn{1}{c}{(5$''$)}&
\multicolumn{1}{c}{(5$''$)}&
\multicolumn{1}{c}{(5$''$)}&
\multicolumn{1}{c}{(10$''$)}\\
\hline
3726.2 & [O\,{\sc ii}] & 33.2$\pm$5.4 & 35.3$\pm$1.8 & 11.6$\pm$2.4 & 16.5$\pm$3.0 & 24.1$\pm$1.7 & 201.2$\pm$4.9 & 141.7$\pm$8.1 & \dots \\
3797.9 & H$\theta$ & 8.4$\pm$3.2 & 6.5$\pm$0.7 & 6.9$\pm$1.7 & 7.2$\pm$1.5 & 5.7$\pm$0.6 & 8.6$\pm$0.8 & \dots & \dots \\
3835.4 & H$\eta$ & 13.5$\pm$3.9 & 9.6$\pm$0.8 & 12.3$\pm$2.1 & 12.2$\pm$1.8 & 9.9$\pm$0.6 & 9.3$\pm$0.7 & \dots & \dots \\
3868.8 & [Ne\,{\sc iii}]& 99.4$\pm$7.0 & 81.3$\pm$1.9 & 32.6$\pm$2.2 & 33.5$\pm$2.2 & 74.7$\pm$1.2 & 157.4$\pm$2.3 & 97.5$\pm$6.1 & \dots \\
3889.1 & H$\zeta$ & 16.5$\pm$3.5 & 16.0$\pm$0.9 & 12.6$\pm$1.7 & 14.9$\pm$1.5 & 16.2$\pm$0.6 & 21.9$\pm$0.8 & 15.6$\pm$2.7 & \dots \\
3967.5 & [Ne\,{\sc iii}]& 49.8$\pm$5.0 & 44.3$\pm$1.4 & 24.4$\pm$2.2 & 29.5$\pm$2.0 & 43.7$\pm$1.0 & 72.1$\pm$1.4 & 46.1$\pm$3.7 & \dots \\
3970.0 & H$\epsilon$ & \dots & \dots & \dots & \dots & \dots & \dots & \dots & \dots \\
4026.2 & He\,{\sc i} & 4.6$\pm$2.6 & 2.0$\pm$0.4 & 1.2$\pm$0.8 & 2.7$\pm$1.1 & 2.2$\pm$0.4 & 2.6$\pm$0.3 & 1.1$\pm$0.9 & \dots \\
4068.6 & [S\,{\sc ii}] & 2.3$\pm$1.5 & 2.8$\pm$0.5 & 2.2$\pm$1.0 & \dots & \dots & \dots & \dots & \dots \\
4101.7 & H$\delta$ & 31.9$\pm$3.4 & 32.7$\pm$1.0 & 27.2$\pm$1.8 & 31.9$\pm$1.5 & 34.1$\pm$0.8 & 33.3$\pm$0.7 & 26.5$\pm$2.4 & \dots \\
4200.0 & He\,{\sc ii} & \dots & \dots & \dots & \dots & 2.3$\pm$0.2 & 1.2$\pm$0.2 & \dots & \dots \\
4340.5 & H$\gamma$ & 54.0$\pm$3.5 & 57.2$\pm$1.0 & 49.9$\pm$1.8 & 54.9$\pm$1.6 & 56.6$\pm$1.0 & 57.4$\pm$0.9 & 46.2$\pm$3.3 & \dots \\
4363.2 & [O\,{\sc iii}] & 19.6$\pm$2.2 & 17.1$\pm$0.6 & 10.0$\pm$1.0 & 9.7$\pm$0.8 & 17.0$\pm$0.6 & 18.3$\pm$0.4 & 16.0$\pm$1.8 & \dots \\
4471.5 & He\,{\sc i} & \dots & \dots & \dots & \dots & \dots & 3.7$\pm$0.2 & 3.8$\pm$1.1 & \dots \\
4540.0 & He\,{\sc ii} & 3.8$\pm$1.1 & 4.3$\pm$0.4 & 3.4$\pm$0.8 & 4.1$\pm$0.7 & 4.4$\pm$0.5 & 2.4$\pm$0.3 & 2.4$\pm$0.8 & \dots \\
4686.0 & He\,{\sc ii} & 112.8$\pm$3.6 & 116.4$\pm$1.6 & 122.8$\pm$2.9 & 133.0$\pm$2.5 & 118.5$\pm$1.4 & 65.2$\pm$0.9 & 78.0$\pm$2.9 & \dots \\
4711.4 & [Ar\,{\sc iv}] & 22.6$\pm$1.8 & 24.2$\pm$0.6 & 20.9$\pm$1.3 & 20.4$\pm$1.0 & 25.2$\pm$0.6 & 11.2$\pm$0.4 & 14.5$\pm$1.4 & \dots \\
4740.2 & [Ar\,{\sc iv}] & 14.5$\pm$1.4 & 17.7$\pm$0.6 & 14.8$\pm$1.1 & 14.5$\pm$0.9 & 18.5$\pm$0.6 & 8.1$\pm$0.4 & 10.3$\pm$1.2 & \dots \\
4861.4 & H$\beta$ & 100$\pm$2.2 & 100$\pm$0.9 & 100$\pm$1.8 & 100$\pm$1.6 & 100$\pm$0.7 & 100$\pm$0.5 & 100$\pm$2.6 & 100$\pm$ \\
4958.9 & [O\,{\sc iii}] & 305.2$\pm$2.2 & 213.6$\pm$8.1 & 122.4$\pm$1.0 & 118.7$\pm$0.6 & 210.4$\pm$9.3 & 392.3$\pm$11.9 & 283.1$\pm$27.2& 453.1$\pm$152.8 \\
5006.8 & [O\,{\sc iii}] & 907.0$\pm$5.7& 638.5$\pm$5.7 & 366.5$\pm$2.6 & 354.2$\pm$1.5 & 628.5$\pm$6.8 & 1174.7$\pm$8.6 & 849.4$\pm$27.9 & 1130.4$\pm$419.3\\
5197.9 & [N\,{\sc i}] & \dots & \dots & \dots & \dots & \dots & 1.8$\pm$0.1 & 1.1$\pm$0.3 & \dots \\
5411.0 & He\,{\sc ii} & 6.2$\pm$0.7 & 8.1$\pm$0.2 & 7.4$\pm$0.5 & 8.1$\pm$0.4 & 8.3$\pm$0.2 & 4.6$\pm$0.1 & 4.6$\pm$0.6 & \dots \\
5517.7 & [Cl\,{\sc iii}]& \dots & 1.0$\pm$0.1 & \dots & \dots & 1.0$\pm$0.1 & 1.4$\pm$0.1 & \dots & \dots \\
5537.9 & [Cl\,{\sc iii}]& \dots & 0.7$\pm$0.1 & \dots & \dots & 0.8$\pm$0.1 & 1.1$\pm$0.1 & \dots & \dots \\
5754.6 & [N\,{\sc ii}] & \dots & 0.4$\pm$0.1 & \dots & \dots & \dots & 2.3$\pm$0.1 & 2.2$\pm$0.6 & \dots \\
5875.6 & He\,{\sc i} & 0.9$\pm$0.2 & 2.0$\pm$0.1 & 1.7$\pm$0.2 & 1.7$\pm$0.2 & 1.8$\pm$0.1 & 6.8$\pm$0.2 & 4.3$\pm$0.6 & \dots \\
6101.0 & [K\,{\sc iv}] & \dots & 0.5$\pm$0.1 & \dots & \dots & 0.7$\pm$0.1 & 0.2$\pm$0.1 & 0.6$\pm$0.4 & \dots \\
6234.0 & He\,{\sc ii} & \dots & \dots & \dots & \dots & \dots & \dots & \dots & \dots \\
6312.1 & [S\,{\sc iii}] & 3.8$\pm$0.4 & \dots & 2.1$\pm$0.4 & 1.9$\pm$0.3 & 4.2$\pm$0.1 & 5.6$\pm$0.1 & 3.9$\pm$0.7 & \dots \\
6363.0 & [O\,{\sc i}] & \dots & 0.4$\pm$0.1 & \dots & \dots & \dots & 3.3$\pm$0.1 & 1.8$\pm$0.5 & \dots \\
6406.0 & He\,{\sc ii} & \dots & 1.1$\pm$0.2 & \dots & 1.0$\pm$0.5 & 1.2$\pm$0.2 & 0.5$\pm$0.2 & \dots & \dots \\
6435.1 & [Ar\,{\sc v}] & 2.1$\pm$0.6 & 3.8$\pm$0.3 & 5.6$\pm$0.7 & 6.7$\pm$0.6 & 4.0$\pm$0.2 & 0.9$\pm$0.1 & 1.9$\pm$0.9 & \dots \\
6548.1 & [N\,{\sc ii}] & 4.5$\pm$0.2 & 8.3$\pm$0.1 & 2.8$\pm$0.1 & 1.8$\pm$0.0 & 4.4$\pm$0.1 & 66.2$\pm$1.1 & 52.0$\pm$3.4 & \dots \\
6562.8 & H$\alpha$ & 287.0$\pm$5.5 & 287.0$\pm$1.6 & 287.0$\pm$4.1 & 287.0$\pm$3.1 & 287.0$\pm$1.4 & 287.0$\pm$1.5 & 286.5$\pm$6.4 & 248.5$\pm$64.3 \\
6583.5 & [N\,{\sc ii}] & 15.8$\pm$1.3 &25.0$\pm$0.5 &7.9$\pm$0.7 &6.0$\pm$0.4 &15.3$\pm$0.3 & 191.6$\pm$1.2 &149.9$\pm$4.5 & \dots \\
6678.2 & He\,{\sc i} & \dots & \dots & \dots & \dots & \dots & 4.5$\pm$0.3 & 2.4$\pm$0.9 & \dots \\
6716.5 & [S\,{\sc ii}] & 8.0$\pm$1.6 & 7.9$\pm$0.4 & 3.3$\pm$1.0 & 2.2$\pm$0.8 & 6.1$\pm$0.3 & 46.0$\pm$0.8 & 51.7$\pm$3.3 & \dots \\
6730.8 & [S\,{\sc ii}] & 6.6$\pm$1.5 & 8.2$\pm$0.4 & 2.5$\pm$0.9 & 1.8$\pm$0.9 & 5.9$\pm$0.3 & 58.2$\pm$0.9 & 50.4$\pm$3.2 & \dots \\
7005.9 & [Ar\,{\sc v}] & 15.6$\pm$4.0 & 25.6$\pm$1.3 & 41.9$\pm$4.0 & 43.6$\pm$3.2 & 29.0$\pm$1.0 & 7.5$\pm$0.5 & 15.4$\pm$4.5 & \dots \\
7065.2 & He\,{\sc i} & \dots & 2.3$\pm$0.6 & \dots & \dots & 2.8$\pm$0.5 & 9.2$\pm$0.5 & \dots & \dots \\
7135.8 & [Ar\,{\sc iii}]& 78.7$\pm$6.6 & 84.2$\pm$1.7 & 40.0$\pm$3.7 & 30.0$\pm$2.4 & 75.8$\pm$1.3 & 158.2$\pm$1.7 & 99.0$\pm$9.1 & \dots \\
\hline
log(H$\beta$) &[erg~cm$^{-2}$~s$^{-1}$] & $\mathrm{-13.26}$& $\mathrm{-12.55}$ &$\mathrm{-13.45}$&$\mathrm{-13.14}$& $\mathrm{-12.68}$&$\mathrm{-12.73}$& $\mathrm{-13.65}$&$\mathrm{-15.24}$ \\
$c$(H$\beta$) & & 0.38 & 0.16 & 0.06 & 0.22 & 0.17 & 0.08 & 0 & \dots \\
A$_{\rm V}$ & & 0.80 & 0.32 & 0.13 & 0.47 & 0.35& 0.16 & 0 & 0\\
\hline
$T_\mathrm{e}$([O\,{\sc iii}]) & [K] &
16000$\pm$800 & 17800$\pm$200 & 18000$\pm$900 &
18000$\pm$800 & 18000$\pm$200 & 13800$\pm$100 &
15000$\pm$500 & \dots \\
$T_\mathrm{e}$([N\,{\sc ii}]) & [K] & \dots & 11000$\pm$1100& \dots & \dots & \dots & 9330$\pm$120&10300$\pm$1100& \dots \\
$n_\mathrm{e}$ ([S\,{\sc ii}]) & [cm$^{-3}$] &
250$\pm$60 & 850$\pm$50 & 100$\pm$80 & 280$\pm$200 & 680$\pm$40 & 1640$\pm$40 & 620$\pm$50 & \dots \\
$n_\mathrm{e}$ ([Cl\,{\sc iii}])&[cm$^{-3}$]& \dots & \dots & \dots & \dots &270$\pm$220 & 880$\pm$80 & \dots & \dots \\
$n_\mathrm{e}$ ([Ar\,{\sc iv}]) &[cm$^{-3}$]& \dots &420$\pm$80 & \dots &50$\pm$170 &440$\pm$90 & 330$\pm$130 &100$\pm$180 & \dots \\
\hline
\end{tabular}
\vspace{0.4cm}
\label{sample}
\end{table*}
\begin{table*}
\centering
\caption[]{Reddeding-corrected line intensities for selected regions of Slit\,B of NGC\,2371}
\setlength{\columnwidth}{0.3\columnwidth}
\setlength{\tabcolsep}{0.4\tabcolsep}
\begin{tabular}{ccccccccccc}
\hline
\multicolumn{1}{c}{$\lambda_{0}$}&
\multicolumn{1}{c}{line}&
\multicolumn{1}{c}{B1}&
\multicolumn{1}{c}{B2}&
\multicolumn{1}{c}{B3}&
\multicolumn{1}{c}{B4}&
\multicolumn{1}{c}{B5}&
\multicolumn{1}{c}{B6}&
\multicolumn{1}{c}{B7}&
\multicolumn{1}{c}{B8}&
\multicolumn{1}{c}{B9}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(15$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(10$''$)}&
\multicolumn{1}{c}{(15$''$)}\\
\hline
3726.2 & [O\,{\sc ii}] & 258.7$\pm$41.7 & \dots & \dots & \dots & \dots & \dots & \dots & \dots & 145.9$\pm$32.4 \\
3868.8 & [Ne\,{\sc iii}] & 190.1$\pm$35.2 & 100.6$\pm$37.2 & 100.1$\pm$29.5 & 36.8$\pm$3.8 & \dots & 41.4$\pm$2.6 & \dots & \dots & 175.8$\pm$35.9 \\
3888.7 & He\,{\sc i} & \dots & \dots & \dots & 12.2$\pm$3.1 & \dots & 14.1$\pm$2.0 & \dots & \dots & \dots \\
3967.5 & [Ne\,{\sc iii}] & 73.9$\pm$18.9 & \dots & \dots & 29.8$\pm$3.7 & \dots & 27.9$\pm$2.3 & \dots & \dots & 73.3$\pm$21.6 \\
4101.7 & H$\delta$ & 31.4$\pm$11.6 & \dots & \dots & 30.7$\pm$3.0 & 26.6$\pm$2.4 & 34.2$\pm$2.1 & 38.4$\pm$13.7 & \dots & 30.2$\pm$11.6 \\
4340.5 & H$\gamma$ & 53.0$\pm$10.8 & 57.9$\pm$18.7 & 36.9$\pm$12.3 & 56.5$\pm$3.2 & 90.0$\pm$2.9 & 58.5$\pm$2.0 & 51.5$\pm$14.3 & 72.5$\pm$23.4 & 55.7$\pm$10.7 \\
4363.2 & [O\,{\sc iii}] & 21.6$\pm$7.7 & 44.5$\pm$17.1 & \dots & 11.7$\pm$2.0 & 10.6$\pm$3.1 & 12.0$\pm$1.2 & \dots & \dots & 14.4$\pm$6.5 \\
4540.0 & He\,{\sc ii} & \dots & \dots & \dots & \dots & \dots & 3.8$\pm$0.8 & \dots & \dots & \dots \\
4686.0 & He\,{\sc ii} & 96.9$\pm$13.3 & 107.6$\pm$19.9 & 107.6$\pm$15.4 & 130.8$\pm$4.0 & \dots & 133.1$\pm$3.2 & 126.1$\pm$12.0 & 138.4$\pm$22.7 & 103.3$\pm$12.9 \\
4711.4 & [Ar\,{\sc iv}] & \dots & \dots & 15.4$\pm$8.6 & 22.4$\pm$2.0 & \dots & 20.9$\pm$1.1 & \dots & \dots & \dots \\
4740.2 & [Ar\,{\sc iv}] & \dots & \dots & 12.6$\pm$7.9 & 16.1$\pm$1.7 & 14.6$\pm$1.6 & 17.3$\pm$0.9 & \dots & \dots & \dots \\
4861.4 & H$\beta$ & 100$\pm$9.6 & 100$\pm$16.9 & 100$\pm$12.2 & 100$\pm$3.4 & 100$\pm$1.3 & 100$\pm$1.8 & 100$\pm$8.8 & 100$\pm$15.6& 100$\pm$10.6 \\
4958.9 & [O\,{\sc iii}] & 533.3$\pm$21.2 & 407.2$\pm$28.1 & 278.2$\pm$17.7 & 135.7$\pm$3.8 & 154.7$\pm$1.9 & 135.1$\pm$2.1 & 253.5$\pm$12.5 & 329.5$\pm$23.5 & 478.5$\pm$22.3 \\
5006.8 & [O\,{\sc iii}] & 1549.7$\pm$31.1& 1191.3$\pm$45.1 & 807.0$\pm$26.0 & 401.9$\pm$6.0 & 448.3$\pm$3.1 & 394.9$\pm$3.4 & 711.3$\pm$21.5 & 954.6$\pm$41.3 & 1385.8$\pm$38.6\\
5411.0 & He\,{\sc ii} & 5.7$\pm$2.2 & \dots & \dots & 6.5$\pm$0.6 & 8.1$\pm$0.9 & 7.5$\pm$0.4 & 7.4$\pm$2.5 & \dots & 5.3$\pm$2.1 \\
5875.6 & He\,{\sc i} & \dots & \dots & \dots & 2.0$\pm$0.4 & \dots & 2.1$\pm$0.2 & \dots & \dots & \dots \\
6101.0 & [K\,{\sc iv}] & \dots & \dots & \dots & \dots & \dots & \dots & \dots & \dots & \dots \\
6312.1 & [S\,{\sc iii}] & 5.8$\pm$2.9 & \dots & \dots & 1.7$\pm$0.6 & \dots & 2.0$\pm$0.3 & \dots & \dots & 4.8$\pm$1.9 \\
6435.1 & [Ar\,{\sc v}] & \dots & \dots & \dots & 4.3$\pm$0.9 & 3.8$\pm$1.7 & 4.2$\pm$0.4 & \dots & \dots & \dots \\
6548.1 & [N\,{\sc ii}] & 46.0$\pm$6.3& 24.5$\pm$11.1& 18.2$\pm$2.8 & 3.4$\pm$0.2 & \dots & 1.0$\pm$0.1& \dots & \dots & 31.8$\pm$5.1 \\
6562.8 & H$\alpha$ & 287.0$\pm$17.7 & 287.0$\pm$26.3 & 287.0$\pm$17.4 & 287.0$\pm$6.8 & 287.0$\pm$3.9 & 287.0$\pm$4.3 & 287.0$\pm$17.6 & 287.0$\pm$32.2 & 287.0$\pm$15.0 \\
6583.5 & [N\,{\sc ii}] & 159.0$\pm$13.0 & 49.3$\pm$11.7 &43.0$\pm$7.0 & 6.2$\pm$1.3 & 12.5$\pm$2.7 & 3.7$\pm$0.5 & 9.4$\pm$4.1 & 21.0$\pm$10.0 & 106.6$\pm$9.0 \\
6716.5 & [S\,{\sc ii}] & 46.6$\pm$9.9 & \dots & 17.4$\pm$7.9 & 2.8$\pm$1.6 & 2.5$\pm$1.8 & 2.0$\pm$0.6 & \dots & \dots & 33.6$\pm$6.6 \\
6730.8 & [S\,{\sc ii}] & 41.4$\pm$9.4 & \dots & 18.0$\pm$8.2 & 2.8$\pm$1.5 & 2.9$\pm$2.3 & 2.2$\pm$0.6 & \dots & \dots & 31.1$\pm$6.5 \\
7005.9 & [Ar\,{\sc v}] & \dots & \dots & \dots & 28.0$\pm$5.1 & 23.1$\pm$4.8 & 27.8$\pm$2.8 & \dots & \dots & \dots \\
7135.8 & [Ar\,{\sc iii}]& 115.7$\pm$2.1 & 127.3$\pm$55.3 & 62.5$\pm$28.4 & 49.5$\pm$5.9 & 37.3$\pm$5.9 & 28.1$\pm$2.6 & 50.0$\pm$17.8 & \dots & 98.9$\pm$22.5 \\
\hline
log(H$\beta$) &[erg~cm$^{-2}$~s$^{-1}$] & $\mathrm{-13.38}$&$\mathrm{-14.10}$& $\mathrm{-13.87}$&$\mathrm{-13.13}$& $\mathrm{-12.88}$&$\mathrm{-12.66}$&$\mathrm{-13.67}$& $\mathrm{-14.14}$& $\mathrm{-13.43}$\\
$c$(H$\beta$) & & 0.49 & 0.41 & 0.46 & 0.35 & 0.56 & 0.54 & 0.54 & 0.44 & 0.47 \\
A$_{\rm V}$ & & 1.02& 0.86& 0.98& 0.74& 1.17& 1.14& 1.13& 0.91 & 0.99\\
\hline
$T_\mathrm{e}$([O\,{\sc iii}]) & [K] &
13100$\pm$1700 & 22000$\pm$5000 & \dots & 18600$\pm$1600 & 15100$\pm$2300 & 19000$\pm$1000 & \dots & \dots & 11800$\pm$1800 \\
$n_\mathrm{e}$([S\,{\sc ii}]) & [cm$^{-3}$]& 380$\pm$30 & \dots &690: & 810$\pm$360 & 1190$\pm$240 & 1070$\pm$30 & \dots & \dots &450$\pm$30 \\
$n_\mathrm{e}$ ([Ar\,{\sc iv}])&[cm$^{-3}$]& \dots & \dots & 1600$\pm$500 & 140$\pm$200 & \dots & 2160$\pm$40 & \dots & \dots & \dots \\
\hline
\end{tabular}
\vspace{0.4cm}
\label{sample}
\end{table*}
Examples of INT IDS spectra extracted from Slits A and B are shown in Figure~7.
The complete list of lines detected in the different extraction regions
and their dereddened intensities relative to an arbitrary value 100 for
H$\beta$ are presented in Tables~3 and 4.
The spectra extracted from Slit~B, particularly those from regions B2, B3,
B7, and B8, show fewer emission lines than those extracted from Slit~A, in
agreement with the lower surface brightness of the major axis probed by
Slit~B (see Fig.~1).
Intensity uncertainties at 1-$\sigma$ are provided in these tables, together
with the values derived for $c$(H$\beta$) and the corresponding fluxes of
the H$\beta$ line for each extraction region.
Different line intensity ratios have been used to determine the physical conditions of the different regions in NGC\,2371 using the {\sc iraf} task {\it temden} \citep{Tody1993}.
$T_\mathrm{e}$ was estimated using the [O\,{\sc iii}] emission lines for
most regions and the [N\,{\sc ii}] emission lines whenever possible, i.e.,
in regions A2, A7, and A8.
$n_\mathrm{e}$ was derived using the
[S\,{\sc ii}] $\lambda\lambda$6717,6731,
[Cl\,{\sc iii}] $\lambda\lambda$5517,5537, and
[Ar\,{\sc iv}] $\lambda\lambda$4711,4740 doublets whenever available.
The values of $T_\mathrm{e}$ and $n_\mathrm{e}$, together with their 1-$\sigma$
uncertainties are listed in the bottom rows of Tables~3 and 4.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=1.0\linewidth]{ionisation_ratio3}
\label{fig:ionisation}
\caption{Line intensity ratios derived for spectra extracted from the regions in Slit~A (bullets) and Slit~B (stars), and from different regions of \emph{HST} flux-calibrated ratio maps as described by the coloured points. Different colours represent line ratios extracted from the regions defined in Figure~9 left panel.
The dashed line in the bottom panel marks the theoretical limit
$\log$([O\,{\sc iii}]/H$\alpha$) = 1.89 $\times\;\log$([S\,{\sc ii}]/H$\alpha$)\;+\;2.46
(see Section~4 for details) between photoionization (below the line) and shock-excitation
(above the line).}
\end{center}
\end{figure}
The values of $T_\mathrm{e}$([O\,{\sc iii}]) listed in Tables~3 and 4
reveal notable temperature gradients in the central cavity of NGC\,2371.
The innermost regions A2, A3, A5, A5, B4, and B6 immediately around the
CSPN show consistently high values of $T_\mathrm{e}$([O\,{\sc iii}]) in
the range 18,000--19,000~K,
that decrease to $\approx$15,000--16,000~K
in the outermost regions A1 and A8,
and have the notably lowest value of $\approx$13,800~K
in the region A7,
which probes the SW knot.
Such high values of $T_\mathrm{e}$ in the close vicinity of the CSPN of
NGC\,2371 arise from its 130~kK high $T_\mathrm{eff}$ (see Section~3.1).
Interestingly, $T_\mathrm{e}$([N\,{\sc ii}]) is only computed on
regions that probe the NE and SW knots.
The much lower values of $T_\mathrm{e}$([N\,{\sc ii}]) in these regions
with respect to those of $T_\mathrm{e}$([O\,{\sc iii}]) indicate that
these apertures probe a mix of low- and high-excitation material.
As for $n_\mathrm{e}$([S\,{\sc ii}]), it varies in Tables~3 and 4 from
100~cm$^{-3}$ to 1640~cm$^{-3}$,
with the densest region corresponding
to the core of the SW knot A7, whose head A6 and tail A8 have intermediate density values 600--700~cm$^{-3}$.
The NE knot seems to have lower density estimates than the SW know, but we note that Slit~A does not goes exactly across the NE knot.
Furthermore, the NE knot is smaller and appears more fragmented than the
SW knot (see Fig.~1 right panels).
These facts will have consequences for the abundance determination too
(see below).
Whereas the regions A3, A5, B4 and B6 immediately around the CSPN share
similar values of $T_{\rm e}$, the densities $n_{\rm e}$([S\,{\sc ii}])
are larger for regions B4 and B6 ($n_\mathrm{e}$= 810--1100~cm$^{-3}$)
than for regions A3 and A5 (100--280~cm$^{-3}$).
The latter seem to probe a hollow region around the CSPN.
Regions B1 and B9 extracted from the outer edges of the major
axis of NGC\,2371 have the lowest temperature values
($T_\mathrm{e}$([O\,{\sc iii}])$\simeq$11,800--13,100~K), but
they have higher densities
($n_\mathrm{e}$([S\,{\sc ii}])$\approx400$~cm$^{-3}$) than
regions within the inner main cavity.
In order to unveil differences in excitation from different morphological features in NGC\,2371, we plot in Figure~8 line ratios of key emission lines derived from the INT IDS spectroscopy \citep[see][and references therein]{Akras2016}, including
[N\,{\sc ii}]~6583/H$\alpha$ vs.\ [O\,{\sc iii}]~5007/H$\alpha$ (top panel)
and [S\,{\sc ii}]~(6717$+$6731)/H$\alpha$ vs.\ [O\,{\sc iii}]~5007/H$\alpha$ (bottom panel).
Differences in excitation are clear: regions close to the star (inner regions -- A3, A5, B4, B5, B6) show the lowest [O\,{\sc iii}]/H$\alpha$ and [N\,{\sc ii}]/H$\alpha$ line ratios, whereas other regions, particularly the outermost regions B1 and B9, and the regions A7 and A8 of the SW knot show the highest [O\,{\sc iii}]/H$\alpha$ and [N\,{\sc ii}]/H$\alpha$.
In Figure~8 bottom panel we show the theoretical limit between photoionised and shocked material $\log$([O\,{\sc iii}]/H$\alpha$) = 1.89~$\log$([S\,{\sc ii}]/H$\alpha$) + 2.46, as estimated by \citet{Danehkar2018} based on the models presented by \citet{Raga2008}. Line ratios below this limit show gas excited by ionisation, while regions above present shocked excited gas. Most morphological features in NGC\,2371 are located below this theoretical limit, i.e., they are excited by photoionisation.
On the other hand, the outermost B1 and B9 regions are located above the dashed line, suggesting that they are excited by shocks as could be expected if these structures were expanding at high speed into the interstellar medium.
The spectra of regions A7 and A8 extracted from the head of the SW dense knot also present line intensity ratios typical of shock-excitation.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=\linewidth]{ionisation_ratio}
\label{fig:ionisation}
\caption{Line intensity ratio maps obtained from calibrated {\it HST} images. ({\it Left}) The (black) dashed-line, red line, blue line, and green line rectangles represent different regions used to study the ionisation structure of NGC\,2371 in Figure~8. The three panels have the same field of view and orientation.}
\end{center}
\end{figure*}
The \emph{HST} WFC3 images of NGC\,2371 have also been used to provide a complementary
view of its ionisation structure.
Ratio maps were obtained by dividing the [O\,{\sc iii}], [N\,{\sc ii}] and [S\,{\sc ii}] calibrated images by the H$\alpha$ image (see Fig.~9).
We note that the \emph{HST} [S\,{\sc ii}] image does not have adequate signal-to-noise ratio in the NW and SE lobes of NGC\,2371. Thus, the line ratio calculations were only performed for the innermost regions of NGC\,2371.
We defined four different rectangular regions (see Fig.~9 left panel) to study the dominant excitation mechanisms:
a large region that includes most of the main inner cavity of NGC\,2371,
two regions covering the SW and NE dense knots, and
a region that includes only gas around the central star.
We computed the pixel-by-pixel line ratios and these are plotted in Figure~8 as colour-coded regions alongside the results found for regions extracted from Slit~A and B. The results are consistent with those obtained from the spectroscopic observations.
Finally, the abundances were computed using the extensively tested code
{\sc PyNEB} developed by \citet[][]{Luridiana2015}.
PyNEB computes the physical conditions ($T_\mathrm{e}$ and $n_\mathrm{e}$)
and ionic and total abundances.
Ionic abundances were calculated with their corresponding temperature
and averaged density based on the ionization potential (IP) of the ion.
We note that the $T_\mathrm{e}$ and $n_\mathrm{e}$ values obtained from PyNEB
(not presented here) are consistent within $<$2\% with those listed in Tables~3
and 4 obtained in {\sc iraf}.
For the calculation of ionic abundances, we only considered emission
lines with uncertainties smaller than 50\%, resulting in abundances
with typical uncertainties of a few percent.
The total abundances of He, O, N, Ne, Ar, S and Cl, listed in Table~5 for
apertures A1--A3 and A5--A8, were computed adopting the ionization correction
factors (ICFs) provided by \citet{Delgado2014}.
The error budget of the abundances of heavy elements in
Table~5 is mostly dominated by uncertainties in the
computation of those ICFs, which are notably large
for high excitation nebulae.
\section{Discussion}
NGC\,2371 has a multi-component morphology with
(1) a pair of outer lobes aligned along PA$=128^{\circ}$,
(2) a high-excitation main shell with apparent elliptical morphology aligned along the same direction with two lower excitation caps at its tips, and
(3) two ensembles of dense low-ionisation knots
in the equatorial region of the main shell, although
misaligned with the previous structures.
We provide next a description of the overall physical structure of NGC\,2371 using kinematical information to help us interpret its different components.
\subsection{On the kinematic structure of NGC\,2371}
By Gaussian fitting of bright emission lines, their centroid can be routinely
determined with an accuracy $\sim$10 times better than the spectral resolution,
which is $\sim$200 km~s$^{-1}$ for our observations with the R400V grating and
$\sim$70 km~s$^{-1}$ for those with the R1200U grating.
Thus, our low resolution spectra allow us to investigate the kinematics of
structures that have velocities $\geq$20 km~s$^{-1}$.
This is illustrated in Figure~10, which presents three sections of the R400V 2D-spectrum obtained with the Slit B along the major nebular axis.
These spectral sections include the H$\alpha+$[N~{\sc ii}], [O~{\sc ii}] and [O~{\sc iii}] emission lines. Velocity gradients are clearly seen in these (as well as other) spectral lines as an S-shaped pattern in the position-velocity (PV) map along the nebular axis, but tilted along the opposite direction for the main cavity.
These PV maps have been used to estimate the systemic radial velocities of the different morphological features in NGC\,2371 (i.e., the difference of the radial velocities of a feature with respect to the average radial velocity of the nebula) labeled in Figure~2 next to their corresponding extraction regions.
The outermost structures B1 and B9 have systemic radial velocities of $\pm$30 km~s$^{-1}$, but the tips of the inner cavity B3 and B7 have larger systemic radial velocities of $\pm$70~km~s$^{-1}$.
Meanwhile, regions B4 and B6 inside the main nebular shell have systemic radial velocities of $\pm$40 km~s$^{-1}$ with opposite sign than those of the adjacent tips B3 and B7.
The fast expansion of the bipolar lobes in NGC\,2371 is consistent with
the faster expansion of PNe with [WR]-type CSPNe compared to other PNe
\citep{Pena2003}.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=\linewidth]{lines_vel.pdf}
\label{fig:lines_vel}
\caption{Sections of the INT IDS spectrum of the major axis of NGC\,2371 (Slit~B) centered on the H$\alpha+$[N\,{\sc ii}], [O\,{\sc ii}] and [O\,{\sc iii}] lines. The red vertical lines show the position of the rest wavelength of each line. The horizontal axis represents wavelength while the vertical the spatial profile.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=1.0\linewidth]{velocity_struc.pdf}
\label{fig:vel_stru}
\caption{Velocity structure of NGC\,2371 suggested by the radial velocities obtained from our spectral analysis.}
\end{center}
\end{figure}
The overall expansion patterns described above are consistent with those reported by \citet{Ayala2005} using longslit high-dispersion echelle data. With that information at hand, we can envisage the simple model of the physical structure along the main nebular axis of NGC\,2371 sketched in Figure~11. The physical structure of NGC\,2371 shares many similarities with that of NGC\,650-1 \citep[see][]{RamosLarios2018}, with two pairs of bipolar lobes tilted with the line of sight along different inclination angles,
and a toroidal or barrel-like central cavity.
In this geometrical model, the reversal of systemic radial velocities between
the toroidal structure and the tips of the innermost bipolar lobes B3 and B7
can be explained if the toroidal structure is orthogonal to the bipolar lobes.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=1.0\linewidth]{NGC2371_compare}
\label{fig:compare}
\caption{{\it GALEX} UV images of NGC\,2371. ({\it Left}) {\it GALEX} far-UV gray scale image with the natural {\it GALEX} pixel size of 1.5~arcsec. ({\it Right}) Colour-composite UV image using the near- and far-UV filters. The dashed-line circle in both panels has an angular radius of 70~arcsec. North is up and east to the left.}
\end{center}
\end{figure*}
The present data allow a basic description of NGC\,2371, but we note that the NW and SE lobes have a rich morphology that suggests a more complex structure than that proposed in Figure~11.
A complete analysis of the physical structure of NGC\,2371 using information obtained along different slit positions using the Manchester Echelle Spectrograph (MES) at the 2.1m telescope of the Observatorio Astron\'omico Nacional de San Pedro M\'artir (SPM-OAN) in conjunction with a morpho-kinematic model is underway (V\'{a}zquez et al., in preparation) and will certainly produce a detailed view of the physical structure of this PN to peer further into its formation history. Yet, the available data still have an interesting piece of information.
The NE and SW low-ionisation knots move at systemic radial velocities
$\pm7$~km~s$^{-1}$ considerably smaller than that of the toroidal
structure of the main nebula where they seem to be embedded.
These velocities, which are consistent with those reported by \citet{Sabbadin1982}, imply that the knots are being overtaken by the nebular material as suggested by \citet{RamosLarios2012}. Furthermore, the arc-like features observed around these structures (see Fig.~1 right panels) seems to suggest that they are interacting with the current fast wind from the CSPN, similarly to what is observed in {\it HST} images of Orion proplyds \citep[see][]{GA2001}.
\begin{table*}
\centering
\caption[]{Total abundances for different regions in NGC\,2371}
\setlength{\columnwidth}{0.2\columnwidth}
\setlength{\tabcolsep}{0.55\tabcolsep}
\begin{tabular}{ccccccccc}
\hline
\multicolumn{1}{l}{} &
\multicolumn{1}{c}{A1}&
\multicolumn{1}{c}{A2}&
\multicolumn{1}{c}{A3}&
\multicolumn{1}{c}{A5}&
\multicolumn{1}{c}{A6}&
\multicolumn{1}{c}{A7}&
\multicolumn{1}{c}{A8}&
\multicolumn{1}{c}{Solar}\\
\multicolumn{1}{l}{} &
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{\citep{Lodders2010}}\\
\hline
He & 0.106$\pm$0.013 & 0.116$\pm$0.005 & 0.122$\pm$0.005 & 0.130$\pm$0.006 & 0.118$\pm$0.003 & 0.102$\pm$0.002 & 0.098$\pm$0.007 & 0.084 \\
O & (7.1$\pm$3.0)$\times10^{-4}$ & (3.8$\pm$1.3)$\times10^{-4}$ & (1.6$\pm$0.6)$\times10^{-4}$ & (1.7$\pm$0.7)$\times10^{-4}$ & (2.8$\pm$0.9)$\times10^{-4}$ & (6.8$\pm$2.4)$\times10^{-4}$ & (4.9$\pm$1.7)$\times10^{-4}$ & 5.37$\times10^{-4}$ \\
N & (2.2$\pm$1.4)$\times10^{-4}$ & (1.1$\pm$0.4)$\times10^{-4}$ & (9.2$\pm$3.2)$\times10^{-5}$ & (4.7$\pm$1.7)$\times10^{-5}$ & (1.3$\pm$0.4)$\times10^{-4}$ & (2.2$\pm$0.8)$\times10^{-4}$ & (2.1$\pm$0.7)$\times10^{-4}$ & 7.24$\times10^{-5}$ \\
Ne & (3.4$\pm$1.4)$\times10^{-4}$ & (3.4$\pm$1.0)$\times10^{-4}$ & (8.1$\pm$2.4)$\times10^{-5}$ & (1.1$\pm$0.4)$\times10^{-4}$ & (1.6$\pm$0.5)$\times10^{-4}$ & (1.4$\pm$0.4)$\times10^{-5}$ & (1.8$\pm$0.5)$\times10^{-4}$ & 1.12$\times10^{-4}$ \\
Ar & 3.5$\times10^{-5}$: & 1.5$\times10^{-5}$: & 1.1$\times10^{-5}$ & 5.0$\times10^{-6}$: & 1.2$\times10^{-5}$: & 1.3$\times10^{-5}$: & 9.3$\times10^{-6}$: & 3.16$\times10^{-6}$ \\
S & (2.9$\pm$1.5)$\times10^{-5}$ & (7.2$\pm$2.9)$\times10^{-6}$& (8.5$\pm$3.4)$\times10^{-6}$ & (4.4$\pm$1.8)$\times10^{-6}$ & (1.5$\pm$0.6)$\times10^{-5}$ & (1.7$\pm$0.7)$\times10^{-5}$ & (1.7$\pm$0.7)$\times10^{-5}$ & 1.45$\times10^{-5}$ \\
Cl & \dots & (1.9$\pm$0.7)$\times10^{-7}$ & \dots & \dots & (3.6$\pm$1.4)$\times10^{-7}$ & (1.5$\pm$0.6)$\times10^{-7}$ & \dots & 1.78$\times10^{-7}$\\
\hline
He/He$_\odot$ & 1.26 & 1.38 & 1.45 & 1.55 & 1.41 & 1.21 & 1.17 & $\dots$ \\
N/O & 0.31 & 0.29 & 0.58 & 0.28 & 0.46 & 0.32 & 0.43 & 0.13 \\
Ne/O & 0.48 & 0.89 & 0.51 & 0.65 & 0.57 & 0.21 & 0.37 & 0.21 \\
Ar/O & 0.05: & 0.04: & 0.05: & 0.03: & 0.04: & 0.02: & 0.02: & 0.006 \\
S/O & 0.04 & 0.02 & 0.05 & 0.03 & 0.05 & 0.03 & 0.03 & 0.03 \\
\hline
\end{tabular}
\vspace{0.4cm}
\label{sample}
\end{table*}
\subsection{A comprehensive view of NGC\,2371}
The spatio-kinematic model of NGC\,2371 described above can be compared to its excitation structure and spatially varying chemical abundances. Despite its complex morphology, NGC\,2371 is quite symmetric, with an ellipsoidal main nebular shell with polar blowouts and a pair of bipolar lobes whose symmetry axis is aligned along a slightly tilted direction. The tips of the bipolar lobes show clear evidence of shock-excitation, which can arise as the nebular material expands at relatively high speed and interacts with the ISM. On the other hand, the main nebular shell and very notably the regions close to the CSPN exhibit very high excitation, with the lowest [N~{\sc ii}]/H$\alpha$ and highest He\,{\sc ii}/H$\beta$ line ratios. Indeed, excitation in these innermost regions is so high that the [O~{\sc iii}]/H$\alpha$ line ratio is damped as O$^{++}$ is photo-ionised to higher ionisation levels. Furthermore, several He\,{\sc ii} emission lines can be seen in the optical spectrum shown in Figure~5. Along with C~{\sc iv} and [Ne~{\sc iv}], it has also been reported to be present in the \emph{IUE} UV nebular spectrum of NGC\,2371 \citep[see fig.~2 in][]{Pottasch1981}. Since the He~{\sc ii} $\lambda$1640 \AA\ is the dominant line in the wavelength range between 1400 \AA\ and 1800 \AA, the \emph{GALEX} far-UV image in Figure~12 (left panel) can be used to probe the spatial location of high-excitation material in NGC\,2371. The emission in this image traces the main morphological features of NGC\,2371, including the main cavity, the dense knots, and the outer edges of the NW and SE lobes. In addition, there is a diffuse halo that extends to distances of $\sim70$~arcsec from the CSPN, which is not detected in the near-UV {\it GALEX} image (see Fig.~12 right panel). When comparing with mid-IR \emph{Spitzer} presented in \citet{RamosLarios2012} and optical CFHT [O~{\sc iii}] images, there is notable emission in the \emph{Spitzer} images close to the CSPN, most likely associated with high-excitation [O~{\sc iv}] emission lines. The high excitation of the inner regions of NGC\,2371 results from the strong UV flux of WD\,0722$+$295, whereas the high-excitation of the outer halo might result from the "hardening" of the stellar radiation, as high energy photons leak from the main nebular shell of NGC\,2371 \citep[see][]{Guerrero1999,Gesicki2003}.
The low-ionisation dense SW and NE knots are aligned along a direction different from the main nebular axis and they do not share the expansion velocity of the main nebula, with lower expansion velocities ($\pm$7~km~s$^{-1}$).
Their morphology, with bow-shocks pointing towards the ionisation source,
also precludes a jet interpretation for these features, but they are more
consistent with clumps surrounded by photo-evaporated material.
Thus, whereas the main nebula of NGC\,2371 has very high-excitation, it
seems that the face of these knots pointing towards the CSPN shields their cometary tails
from the UV stellar flux, producing emission from low ionisation
species at their tails.
Indeed, the spectra extracted from these structures present emission from [O\,{\sc i}] and [N\,{\sc i}] (see A2, A7 and A8 in Tab.~2) suggesting them
to be low-ionisation structures (LIS).
The analysis of LIS in PNe presented by \citet{Akras2016} suggests
$T_\mathrm{e}$=10,000--14,700~K, consistent with our estimates for
the SW and NE knots, but lower electron densities than the surrounding
ionised gas, contrary to the SW and NE knots of NGC\,2371, which are
much denser than the nebular material where they are embedded (see Tab.~2).
In this sense, the SW and NE knots of NGC\,2371 are not typical LIS.
The chemical abundances of different apertures along the minor axis
of NGC\,2371 are listed in Table~5.
The mean value of the He abundances is 0.113$\pm$0.012
\cite[$\simeq$1.4 times the solar value in][]{Lodders2010},
with most individual values within 1-$\sigma$ uncertainty.
This might also be the case for the O abundances, with a mean value of
(4.1$\pm$2.3)$\times$10$^{-4}$, which is mostly consistent within 1-$\sigma$
uncertainty with that of individual apertures.
This mean value suggests slightly subsolar O abundances,
(O/H)$\simeq$0.75 (O/H)$_\odot$.
On the contrary, the N abundances are clearly larger than solar,
with a mean N/O ratio of 0.38$\pm$0.11, i.e., (N/O)$\approx$3
(N/O)$_\odot$.
The chemical abundances (He/H, O/H, N/O) of NGC\,2371 are basically
consistent with those reported for comprehensive samples of PNe with
[WC] central stars \citep{Pena2001,GarciaRojas2013}.
The other chemical abundances in Table~5 have larger uncertainties,
particularly the Ar abundances, for which no error bar is provided.
The values of the chemical abundances of Ne, Ar, S, and Cl are
also consistent with those of PNe with [WC] central stars.
The chemical abundances found in different regions are in general consistent
among them,
within 2$\sigma$ up most from their mean values,
but we note that the abundances
of the main elements He, N, and O show some
differences from
one region to another.
It might be argued that the O abundances from apertures probing the
dense low-ionisation structures are higher than those of the main cavity enclosing them, whereas the N/O ratios are lower.
We believe this is not the case, because a
close inspection reveals that the O abundances are mostly
anti-correlated with $T_\mathrm{e}$, underlying the difficulties to
determine the chemical abundances in highly excited nebulae and from
regions where low- and high-excitation material coexists.
Finally, we would like to remark that \citet{Herald2004} suggested
WD\,0722$+$295 to be in the same evolutionary stage as the CSPN of
the born-again PN A\,78, a [WR]-PG 1159 star.
Although their analysis of the stellar atmospheres of those two CSPNe with
their version of the CMFGEN code resulted in similar atmospheres, our
best-fit models of WD\,0722$+$295 and our recent analysis of the CSPN of
A\,78 \citep[see][]{Toala2015} show these two stars to be different.
Moreover, the presence of hydrogen-deficient material inside
A\,78 suggests a different evolutionary path compared to the
CSPN of NGC\,2371.
The spectral analysis of WD\,0722$+$295 presented here represents an
improvement over other similar studies due to the high-quality of the
optical spectrum and the combination with available UV data used for
the fit.
The PoWR model presented here is not able to reproduce the lines identified
in Section~3 at $\approx$5665 \AA\ and $\approx$6066 \AA\ as O\,{\sc vii}
$\lambda$5666 and O\,{\sc viii} $\lambda$6068, because the ionisation
potentials of these species, 739.3~eV and 871.4~eV, respectively, are too
high to be produced by the stellar atmosphere.
Although it can be argued that such high-temperature gas can be produced by the X-ray emission of WD\,0722$+$296 detected by {\it Chandra} \citep[][]{Montez2015}, it has been long discussed whether these emission lines are due to oxygen, but to neon.
\citet{Werner2007} presented stellar atmosphere models of pre-WD stars and
demonstrated that most of the emission lines from the O\,{\sc vii} and
O\,{\sc viii} high-ionisation states of oxygen might be misidentified with
those of Ne\,{\sc vii} and Ne\,{\sc viii},
whose ionisation potentials, 207.3~eV and 239.1~eV, respectively,
are considerably lower.
This does not affect the spectral classification of WD\,0722$+$296 as
part of the [WO] sequence, because the O\,{\sc vi} $\lambda$3820,
O\,{\sc vi} $\lambda$5290, and O\,{\sc vi} $\lambda$6200 lines are still
present in its spectrum (see Table~1).
Indeed, the Ne\,{\sc vii} and Ne\,{\sc viii} lines require a high
temperature for the central star.
\section{Summary}
We presented spatially-resolved longslit INT IDS spectroscopic observations of NGC\,2371, a PN with a [WR]-type CSPN with an apparent complex morphology.
Our spectral observations, in conjunction with \emph{HST}, {\it GALEX} and CFHT images, have allowed us to characterise the central star of NGC\,2371 as well as to unveil its high-ionisation structure. Our findings can be summarised as follows:
\begin{itemize}
\item We studied the spectral properties of the CSPN of NGC\,2371, WD\,0722$+$295, by analysing the IDS observations. A Gaussian-fitting procedure was performed to measure the true contribution of the WR spectral lines and to assess the spectral sub-type of the CSPN. Line ratios of WR features with that of the RB suggest a [WO1]. In combination with the UV data we used the stellar atmosphere code PoWR to estimate the stellar parameters. Although we found very similar results as those obtained in previous works, the differences are due to the assumed clumping factor and a more accurate {\it Gaia} distance.
\item We studied the physical properties of different regions of NGC\,2371. In accordance with previous works on NGC\,2371 the electron density ranges between 100~cm$^{-3}$ and 1,640~cm$^{-4}$. The densest regions are the SW and NE low-ionisation knots. The electron temperatures vary from $T_\mathrm{e}\sim$12,000~K at the outer edges to $T_\mathrm{e}\approx18,000$~K for regions close to the star. The outer bipolar (NW and SE) lobes have electron densities $n_\mathrm{e}\approx400$~cm$^{-3}$ and $T_\mathrm{e}\approx12,000-13,000$~K.
\item The analysis of emission line ratios from the IDS spectra as well as those extracted from the {\it HST} calibrated images demonstrate the powerful effect of the UV flux from the CSPN of NGC\,2371.
Most emission from this PNe is dominated by ionisation, but the
low-ionisation SW and NE knots and the outer regions B1 and B9,
where shock-excitation is important.
We suggest that B1 and B9 are shock-excited because they trace the shock
due to the expansion of NGC\,2371.
\item Inspection of {\it GALEX} images unveiled the presence of a halo extended to 70~arcsec from WD\,0722$+$295. The halo is detected in the far-UV channel and is mainly dominated by He\,{\sc ii} emission. We suggest that this halo has formed due to high-energy photon leakage.
\item
We found that the bow shock-like structure around the CSPN-facing head
of the low-ionisation dense knots correspond to photoevaporation flows.
The head of the knots shield their tails producing emission from low-ionisation species such as [O\,{\sc i}] and [N\,{\sc i}]. This reinforces the idea proposed by \citet{Herald2004} and \citet{RamosLarios2012} that molecular material might be present in the vicinity of the CSPN of NGC\,2371, probably in these structures similar to LIS in other PNe \citep[][]{Akras2017}. However, we note that these clumps do not share the same physical properties as classic LIS described in the literature.
\item We estimated the chemical abundances for different regions of
NGC\,2371.
The abundances are typical of PNe with [WC] central stars.
Although some variations may exist among different regions, the abundances
of the dense knots lay within those of adjacent regions, suggesting that
they were not ejected as the result of a VLTP.
Instead, their kinematics and detailed morphology suggest they were
ejected before the formation of the main nebular shell.
\item The relatively high-resolution of the optical longslit spectra presented here allowed us to suggest a possible kinematic structure of NGC\,2371. The outer lobes have the typical expansion velocity of evolved PNe ($\sim$30~km~s$^{-1}$) whilst the inner shell has an expansion velocity of $\sim$70~km~s$^{-1}$ in accordance with previous analysis of high-resolution echelle observations. On the other hand, the dense knots have a slow radial velocity. We suggest that NGC\,2371 has a bipolar shape with each lobe presenting a double-structure protruding from a barrel-like central region.
\end{itemize}
We propose that the densest material around WD\,0722$+$295 might be the
relic of an early ejection of material along a ring-like structure.
The ejection of the main nebula is more recent and can
be directly related to the current fast stellar wind of the progenitor star
of NGC\,2371.
It seems to have expanded towards the low-density region, the polar zone
of the barrel-like structure.
We emphasise that the analysis of high-resolution echelle observations
such as those obtained with the San Pedro M\'{a}rtir MES are most
needed in order to construct a detailed view of the morpho-kinematic
structure of NGC\,2371.
\section*{Acknowledgements}
The authors thank the referee for a critical reading of the manuscript and valuable suggestions that improved the presentation of the paper. The authors would like to thank V.\,G\'{o}mez Llanos Sandoval for helping them in using PyNEB. VMAGG, JAT, MAG and HT are funded by UNAM DGAPA PAPIIT project
IA100318. GRL acknowledges support from Fundaci\'on Marcos Moshinsky, CONACyT and PRODEP (Mexico).
MAG acknowledges support of the Spanish Ministerio de Ciencia, Innovación y Universidades grant PGC2018-102184-B-I00, co-funded by FEDER funds.
LS thanks support from UNAM PAPIIT grant IN101819. YDM thanks CONACyT for the research grant CB-A1-S-25070. This work has make extensive use of the NASA's Astrophysics Data System. This paper also presents data obtained with the MOS camera at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Sciences de l'Univers of the Centre National de la Recherche Scientifique of France, and the University of Hawaii. The authors thank E.\,Santamar\'{i}a for helping produce the velocity sketch of NGC\,2371. Based on observations made with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555.
| 9cf82b0e4729de4dc3bf671099d699a42f70c5c6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The emergence of COVID-19 in early 2020 and its
subsequent outbreak have affected and changed the world dramatically.
According to the World Health Organization (WHO), by mid-May 2020, the number of confirmed COVID-19 cases has reached 5 millions with death toll over 300,000 world wid
.
Several mandatory rules have been introduced by the government to prevent the spread of the coronavirus, such as social distancing, bans on social gatherings, store closures and school closures. Despite their positive effects on slowing the spread of the pandemaic, they neverthless caused severe impacts on the society, the economy and people's everyday life. There have been anti-lockdown and anti-social-distancing protests in many places around the world. Given these difficult situations, it is crucial for policy-makers to understand people's opinions toward the pandemic so that they can (1) balance the concerns of stoping the pandemic on the one hand and keeping people in good spirits on the other hand and (2) anticipate people's reactions to certain events and policy so that the policy-makers can prepare in advance. More generally, a close look at the public affect during the time of COVID-19 could help us understand people's reaction and thoughts in the face of extreme crisis, which sheds light on humanity in moments of darkness.
People constantly post about the pandemic on social media such as Twitter, Weibo and Facebook. They express their
attitudes and feelings regarding various aspects of the pandemic, such as the medical treatments, public policy, their worry, etc. Therefore, user-generated content on social media provides an important source for understanding
public emotions and concerns.
In this paper, we provide a comprehensive analysis on the affective trajectories of American people and Chinese people based on Twitter and Weibo posts between January 20th, 2020 and May 11th 2020.
We identify fine-grained emotions (including anger, disgust, fear, happiness, sadness, surprise) expressed
on social media based on the user-generated content.
Additionally,
we build NLP taggers to
extract the triggers of different emotions, e.g., why people are angry or surprised, what they are worried, etc.
We also contrast public emotions between China and the Unites States,
revealing sharp differences in public reactions towards COVID-19 related issues in different countries.
By tracking the change of public sentiment and emotion over time,
our work sheds light on the evolution of public attitude towards this global crisis.
This work contributes to the growing body of research on social media content in the time of COVID-19.
Our study provides a way to extracting public emotion towards the pandemic in real-time, and could potentially lead to better decision-making and the development of wiser interventions to fight this global crisis.
The rest of this paper is organized as follows:
we briefly go through some related work in Section 2.
We then present the analyses on topic trends in Weibo and Twitter (section 3), the extracted emotion trajectories (section 4) and triggers of those emotions (section 5). We finally conclude this paper in Section 6.
\section{Related Work}
\subsection{Analyses on Social Media Content about COVID-19 }
\begin{comment}
There has been a decade history of using NLP tools to analyze user-generated content on social media regarding variety of aspects, such as
epidemic detection \cite{achrekar2011predicting,paul2011you,ji2012epidemic,li2013early},
natural disaster detection \cite{sakaki2010earthquake,sakaki2012tweet},
topic trend detection \cite{hong2010empirical,o2010tweetmotif,zhao2011comparing},
sentiment and emotion analysis \cite{go2009twitter,o2010tweets,kouloumpis2011twitter,bollen2011modeling,balabantaray2012multi,severyn2015twitter,kumar2015emotion},
entity and event extraction \cite{ritter2011named,ritter2012open,li2014major,zhou2014simple,deng2015overview}.
\end{comment}
At the time of writing, analyses on people's discussions and behaviors on social media in the context of COVID-19 has attracted increasing attention.
\cite{alshaabi2020world}
analyzed tweets concerning
COVID-19 on
Twitter by
selecting important 1-grams based on rank-turbulence divergence
and compare languages used in early 2020 with the ones used a year ago.
The authors observed the first peak of public attention to COVID-19 around January 2020 with the first wave of infections in China, and the
second peak later when the outbreak hit many western countries.
\cite{chen2020covid} released the first COVID-19 Twitter dataset.
\cite{kleinberg2020measuring}
provided a ground truth corpus by annotating 5,000
texts (2,500 short + 2,500 long texts) in UK and
showed
people's worries about their families and economic situations.
\cite{gao2020mental} viewed emotions and sentiments on social media as indicators of mental
health issues, which result from self-quarantining and social isolation.
\cite{schild2020go} revealed increasing amount of hateful speech and conspiracy theories
towards specific ethnic groups such as Chinese
on Twitter and 4chan’s.
Other researchers started looking at the spread of misinformation on social media \cite{gallotti2020assessing,ferrara2020covid}. \cite{cinelli2020covid}
provide an in-depth analysis on the diffusion of misinformation concerning COVID-19
on five different social platforms.
\subsection{Analyses of Emotions and Sentiments on Social Media}
{\it Discrete Emotion Theory } \cite{plutchik1980general,frijda1988laws,parrott2001emotions} think that all humans have an innate set of distinct basic emotions.
Paul Ekman and his colleagues \cite{ekman1992argument} proposed that
the six basic emotions of humans are anger, disgust, fear, happiness, sadness, and surprise.
Ekman explains that different emotions have particular characteristics expressed in varying degrees.
Researchers have debated over the exact categories of discreate emotions. For instance, \cite{raghavan1975number}
proposed eight classes for emotions including love, mirth, sorrow, anger, energy, terror, disgust and astonishment.
Automatically detecting sentiments and emotions in text is a crucial problem in NLP and there has been a large body of
work on annotating texts based on sentiments and building machine tools to automatically identify emotions
and sentiments \cite{yang2007emotion,pang2008opinion,mohammad2016sentiment,tang2015document}.
\cite{mohammad2017wassa} created the first annotated dataset for four classes of emotions, anger, fear, joy, and sadness, in which
each text is annotated with not only a label of emotion category, but also the intensity of the emotion expressed based on the Best–Worst Scaling (BWS) technique \cite{louviere1991best}.
A follow-up work by \cite{mohammad2018semeval} created a more comprehensively annotated dataset from tweets in English, Arabic, and Spanish. The dataset covers five different sub-tasks including
emotion classification, emotion intensity
regression, emotion intensity ordinal classification, valence regression and valence ordinal classification.
There has been a number of studies on
extracting
aggregated public mood and emotions from social media \cite{mishne2006capturing,liu2007arsa,dodds2010measuring,gilbert2010widespread}.
Facebook introduced Gross National Happiness (GNH) to estimate the aggregated mood of the
public using
the LIWC dictionary. Results show a clear weekly cycle of public mood.
\cite{mitchell2013geography}
and \cite{li2014nasty}
specially investigate the influence of geographic places and weather on public mood from Twitter data.
The mood indicators extracted from tweets are very predictive and robust \cite{dodds2010measuring,dodds2011temporal}. Therefore,
they have been used to predict real-world outcomes
such as economic trends \cite{gilbert2010widespread,rao2012tweetsmart,rao2012using,zhang2011predicting}, stock market \cite{bollen2011twitter,chung2011predicting},
influenza outbreak \cite{li2013early},
and political events \cite{hopkins2010method,o2010tweets,tumasjan2010predicting,larsson2012studying}.
\section{General Trends for COVID-19 Related Posts}
\input{figs/general.tex}
In this section,
we present the general trends for COVID19-related posts
on Twitter and Weibo.
We first present the semi-supervised models we used to detect COVID-19 related tweets.
Next we present the analysis on the topic trends on
the two social media platforms.
\subsection{Retrieving COVID-19 Related Posts}
For Twitter, we first obtained 1\% of all tweets that are written in English and published within the time period between January 20th, 2020 and May 11th 2020.
The next step is to select tweets related to COVID-19.
The simplest way, as in \cite{chen2020covid,ferrara2020covid}, is to use a handcrafted keyword list to obtain tweets containing words found in the list.
However,
this method leads to lower values in both precision and recall:
for precision, user-generated content that contains the mention of a keyword is not necessarily related to COVID-19.
For example, the keyword list used in \cite{chen2020covid} include the word {\it China}, and it is not suprising that a big proportion of the posts containing ``China" is not related to COVID-19; for recall, keywords for COVID-19 can change over time and might be missing in the keyword list.
To tackle this issue,
we adopt a bootstrapping
approach. The
bootstrapping approach is related to previous work
on semi-supervised data harvesting methods \cite{li2014weakly,davidov2007fully,kozareva2010learning}, in which
we build
a model that
recursively uses seed examples
to extract patterns, which are then used to harvest
new examples. Those new examples are further used as new
seeds to get new patterns.
To be specific, we first obtained a starting seed keyword list by (1) ranking words
based on tf-idf scores
from eight COVID-19 related wikipedia articles; (2)
manually examining the ranked word list, removing those words that are apparently not COVID-19 related, and use the top 100 words in the remaining items.
Then we retrieved tweets with the mention of those keywords.
Next, we randomly sampled 1,000 tweets from the collection and manually labeled
them as either COVID-19 related or not.
The labeled dataset is split into the training, development and test sets with ratio 8:1:1.
A binary classification model is trained on the labeled dataset
to classify whether a post with the mention of COVID-related keywords is actually COVID-related.
The model
is trained using BERT \cite{devlin2018bert} and optimized using Adam \cite{kingma2014adam}.
Hyperparameters such as the batch size, learning rate are tuned on the development set.
Next, we obtain a new seed list by picking the most salient words that contribute to the positive category in the binary classification model
based on the first-order derivative saliency scores \cite{erhan2009visualizing,simonyan2013deep,li2015visualizing}.
This marks the end of the first round of the bootstrapping.
Next we used the new keyword list to re-harvest a new dataset with the mention of the keyword, 1,000 of which is selected and labeled to retrain the binary classification model.
We repeat this process for three times.
F1 scores for the three rounds of binary classification are 0.74, 0.82, 0.86 respectively.
After the final round of bootstrapping,
we collected a total number of 78 million English tweets concerning the topic of COVID-19.
We used this strategy to retrieve COVID-related posts on Weibo and collected a total number of 16 million posts.
\subsection{Analyses}
We report the intensity scores for Weibo and Twitter in Figure \ref{general}.
We split all tweets by date, where $X_t$ denotes all tweets published on day $t$.
The value of intensity is the number of posts classified as COVID-related
divided by the total number of retrieved posts, i.e., $|X_t|$.
On Weibo, we observe a peak in late January and February, then a drop, followed by another rise in March, and a gradual decline afterwards.
The trend on Chinese social media largely reflects the progress of the pandemic in China:
the outbreak of COVID-19 and the spread from Wuhan to the rest of the country corresponds to the first peak.
The subsequent drop reflects the promise in containing the virus, followed by a minor relapse in March.
For Twitter, we observe a small peak that is aligned
with the news from China about the virus.
The subsequent drop reflects the decline of the attention to the outbreak in China. The curve progressively went up since March, corresponding to the outbreak in the US.
Upon the writing of this paper, we have not observed a sign of drop in the intensity score of COVID19-related posts.
\section{The Evolution of Public Emotion}
In this section, we present the analyses on the evolution of public emotion in the time of COVID-19.
We first present the algorithms we used to identify the emotions expressed in a given post. Next we present the results of the analyses.
\subsection{Emotion Classification}
We adopted the well-established emotion theory by Paul Ekman \cite{ekman1992argument}, which groups human emotions into 6 major categories,
i.e., anger, disgust, worry, happiness, sadness, and surprise.
Given a post from a
social network
user, we assign one or multiple emotion labels to it \cite{buechel2016emotion,demszky2020goemotions}.
This setup is quite common in text classification \cite{ikonomakis2005text,aggarwal2012survey,zhang2015character,zhou2015c,joulin2016bag}.
For emotion classification of English tweets, we
take the advantage of labeled datasets from
the SemEval-2018 Task 1e \cite{mohammad2018semeval}, in which a tweet was associated with either
the ``neutral'' label or with one or multiple emotion
labels by human evaluators.
The SemEval-2018 Task 1e contains eleven emotion categories in total, i.e., anger, anticipation, disgust, fear, joy, love, optimism, pessimism,
sadness, surprise and trust, and we only use the datasets for a six-way classification, i.e.,
anger, disgust, fear, happiness, sadness, and surprise.
Given that
the domain of the dataset used in \cite{mohammad2018semeval} covers all kinds of tweets, and our domain of research covers only COVID-related tweets, there is a gap between the two domains.
Therefore, we additionally labeled 15k COVID-related tweets following the guidelines in \cite{mohammad2018semeval}, where each tweet can take either the neural label or one/multiple emotion
labels.
Since one tweet can take on multiple emotion labels,
the task is formalized as a
a multi-label classification task, in which six binary (one vs. the rest) classifiers are trained.
We used the description-based BERT model \cite{chai2020description} as the backbone, which achieves current SOTA performances on a wide variety of text classification tasks.
More formally, let us consider
a to-be-classified tweet
$x=\{x_1,\cdots,x_L\}$, where $L$ denotes the length of the text $\bm{x}$.
Each $x$ will be tagged with one or more class labels $y \in \mathcal{Y} = [1,N]$, where $N=6$ denotes the number of the predefined emotion classes (the six emotion categories). To compute the probability $p(y|\bm{x})$, each input text $\bm{x}$ is concatenated with the description $\bm{q}_y$ to generate $\{\text{[CLS]};\bm{q}_y;\text{[SEP]};\bm{x}\}$, where $\text{[CLS]}$ and $\text{[SEP]}$ are special tokens.
The description $q_y$ is the Wikipedia description for each of the emotions. For example, $q_y$ for the category {\it anger} is {\it ``Anger, also known as wrath or rage, is an intense emotional state involving a strong uncomfortable and hostile response to a perceived provocation, hurt or threat."}
Next, the concatenated sequence
is fed to the BERT model, from which we obtain the contextual representations $h_\text{[CLS]}$.
$h_\text{[CLS]}$ is then transformed to a real value between 0 and 1 using the sigmoid function, representing the probability of assigning
the emotion
label $y$ to the input tweet $\bm{x}$:
\begin{equation}
p(y|\bm{x})=\text{sigmoid}(W_2\text{ReLU}(W_1h_\text{[CLS]}+b_1)+b_2)
\end{equation}
where $W_1,W_2,b_1,b_2$ are some parameters to optimize.
Classification performances for different models are presented in
Table \ref{en-emotion}.
\begin{table}
\center
\begin{tabular}{cccc}\hline
Model & Acc & micro F1 & macro F1 \\\hline
SVM biagram & 51.4 & 63.0 & 52.7 \\
BERT \cite{devlin2018bert} & 65.0 & 75.2 & 66.1 \\
BERT-description \cite{chai2020description} & 66.8 & 77.0 & 68.3 \\\hline
\end{tabular}
\caption{ Results for the multi-label emotion classification for English tweets. }
\label{en-emotion}
\end{table}
\subsection{Analyses}
For emotion $y$, its intensity score $S(t,y)$ for day $t$ is the average probability (denoted by $P(y|x)$)
of assigning label $y$ to all
the texts in that day $X_t$. For non COVID-related texts, $P(y|x)$ is automatically set to 0. We thus have:
\begin{equation}
S(t,y) = \frac{1}{|X_t|} \sum_{x\in X_t} p (y|x)
\end{equation}
For Chinese emotion classification, we used the labeled dataset in \cite{lai2019fine}, which contains 15k labeled microblogs from weibo\footnote{The original dataset contains 7 categories of emotions, and we used only six of them.}.
In addition to the dataset provided by \cite{lai2019fine}, we labeled COVID-related 20k microblogs.
The combined dataset is then used to train a multi-label classification model based on the description-BERT model \cite{chai2020description}.
Everyday emotion scores for Weibo are computed in the same way as for Twitter.
The time series of intensity scores of six different emotions, i.e., sadness, anger, disgust, worry, happiness, surprise,
for Weibo and Twitter
are shown in Figures \ref{ch-emotion} and \ref{en-emotion}, respectively.
For Weibo,
as can be seen, the trend of {\it worry} is largely in line with the trend of the general intensity of the COVID-related posts. It reached a peak in late January, and then gradually went down, followed by a small relapse in mid-March.
For {\it anger}, the intensity first went up steeply at the initial stage of the outbreak, staying high for two weeks,
and then had another sharp increase
around February 8th. The peak on February 8th was due to the death of Wenliang Li,
a Chinese ophthalmologist
who issued the warnings about the virus.
The intensity for anger then gradually decreased, with no relapse afterwards.
The intensity for
{\it disgust} remained relatively low across time.
For {\it sadness}, the intensity reached the peak at the early stage of the outbreak, then gradually died out with no
relapse.
For {\it surprise}, it went up first, mostly because of the fact that the public was surprised by the new virus and the unexpected outbreak, but then gradually went down.
The intensity for {\it happiness} remained relatively low across time, with a small peak in late April, mostly because the countrywide lockdown was over.
For Twitter, the intensity for {\it worry} went up shortly in late January, followed by a drop.
The intensity then went up steeply in mid-March in response to the pandemic breakout in the States, reaching a peak around March 20th, then decreased a little bit and remained steady afterwards.
The intensity for {\it anger} kept going up after the outbreak in mid-March, with no drop observed.
The trend for {\it sadness} is mostly similar to that of the overall intensity.
For {\it surprise}, the curve went up first after the breakout in early March, reaching a peak around Mar 20th, then dropped, and remained steady afterwards.
For {\it happiness}, the intensity remained low over time.
\begin{table*}
\small
\center
\begin{tabular}{|c|c|c|c|c|c|}\hline
sadness & anger & disgust &worry &happiness & surprise \\\hline
passed away &realdonaldtrump & covid / covid-19 &covid / covid-19 & healthy &covid / covid-19 \\
\multirow{2}{*}{died} & \multirow{2}{*}{lockdown} & \multirow{2}{*}{realdonaldtrump}& \multirow{2}{*}{job} & \multirow{2}{*}{help} &human-to-human \\
& & & &&transmission \\
deaths & government & chinese &kids &recover &outbreak \\
fever &quarantine& trump&bill& check & lockdown\\
unemployed & wuhan& masks & unemployed& return to work& test\\
parents& lies & virus & food &reopening& deaths \\
test positive & close & pence & crisis & vaccine &pandemic\\
family & stayhome & chinks & parents & money &confirmed cases\\
patients & WHO & china & economy & treatment &total numbers\\
isolation & trump & hospitals & families & friends & conspiracy \\\hline
\end{tabular}
\caption{Top 10 extracted trigger spans regarding different emotions on Twitter.}
\label{top-entity}
\end{table*}
\begin{table*}
\small
\center
\begin{tabular}{|c|c|c|c|c|}\hline
China& lockdown &Trump & Hospitals &Increasing Cases and Deaths \\\hline
China& quarantine& realdonaldtrump & hospital& case \\
Chinese& stay & donald & patients & deaths\\
Wuhan& close&lies & test & report\\
bat-eating&home &republicans &doctor & confirmed\\
chink& stayhome & hydroxychloroquin & case & report\\
chinesevirus & boarder & american & healthcare& york \\
sinophobia&shutdown&media & drug &us \\
chingchong & distancing &governor & vaccine & total \\
hubei& coronalockdown&pence & ventilator & government\\ \hline
\end{tabular}
\caption{Top mentions of different subcategories for {\it anger} on Twitter.}
\label{anger}
\end{table*}
\begin{table*}
\small
\center
\begin{tabular}{|c|c|c|c|c|}\hline
syndrome and being infected& families &finance and economy & jobs and food &Increasing Cases and Deaths \\\hline
fever& parents& money & jobs & deaths \\
hospital & mother & stock & unemployed& spread\\
cough& children&financial &food &poll number \\
test positive&mom &price & money & death toll \\
icu& families & loan &work & confirmed \\
doctor &father &business& starve & rise\\
bed&kids& debt & unemployment & official \\
confimed & daughter &market & check & number \\
sick& father-in-law& crash& layoff & safe\\ \hline
\end{tabular}
\caption{Top mentions of different subcategories for {\it worry} anger on Twitter.}
\label{worry}
\end{table*}
\input{figs/ch_emotion.tex}
\input{figs/en_emotion.tex}
\section{ Emotional Triggers}
For a given emotion, we would like to dive deeper into its different
subcategories. For example, for {\it worry}, we wish to know
what the public is worried about, and how these triggers of worry change over time.
In this section, we first present our methods for extracting triggers/subcategories for different emotions, followed by
some analyses and visualization on the Twitter data.
\subsection{Extracting the Triggers for Different Emotions}
In order to extract the emotional triggers from Twitter’s noisy
text, we first annotate a corpus of tweets.
For the ease of annotation, each emotion is associated with only a single trigger:
the person/entity/event that a
user has a specific emotion towards/with/about.
A few examples are shown as follows with target triggers surrounded by brackets.
\begin{itemize}
\item
{\it Angry protesters are traveling 100's of miles to join organized rallies over COVID-19 [lockdown]$_{attr\_anger}$. }
\item
{\it Feeling awfully tired after a 5.30am start for work today. Worried too about [the early return to school]$_{attr\_worry}$, my grandchildren are so very dear to me . I could not bear to lose them to covid. }
\item
{\it All Americans are very angry with
[@realDonaldTrump]$_{attr\_anger}$
81,647 dead Americans would be very angry as well. If they weren’t dead.}
\item
{\it
Fucking [bat-eating chinks]$_{{attr\_anger}}$, go die in a hole, far away from us. }
\item
{\it Well, I am ANGRY as hell at Trump$_{attr\_anger }$}.
\item
{\it With great sadness we report the sad [loss of two dear Friends]$_{{attr\_sadness}}$ }
\item
{\it
The lockdown$_{{attr\_anger}}$ was implemented when there were hardly any cases and now it is above lakhs and people are acting so carelessly. Making me so angry. }
\end{itemize}
In order to build an emotional trigger tagger, we annotated 2,000 tweets in total, and split them into training, development and test sets with ratio 8:1:1.
We treat the problem as a sequence labeling task, using Conditional Random Fields for learning and inference with
BERT-MRC features \cite{li2019unified}.
Comparing with vanilla BERT tagger \cite{devlin2018bert},
the BERT-MRC tagger has the strength of encoding the description of the to-be-extracted entities, e.g., {\it what they are worried about}.
As this description provides the prior knowledge about the entities, it has been shown to outperform vanilla BERT even when less training data is used.
In addition to the representation features from BERT-MRC, we
also considered the Twitter-tuned POS features \cite{ritter2011named},
the dependency features
from a Twitter-tuned dependency parsing model \cite{kong2014dependency} and the Twitter event features \cite{ritter2012open}.
The precision and recall for segmenting emotional triggers on English tweets are reported in Table \ref{trigger}.
\begin{table*}[!ht]
\center
\begin{tabular}{lccc}\hline
Model & Pre & Rec & F1 \\\hline
BERT & 0.41 & 0.60 & 0.48\\
BERT-MRC & 0.54 & 0.66 &0.59 \\
CRF with BERT-MRC features & 0.53 &0.68&0.60\\
CRF with BERT-MRC features, POS, event and parse tree features & 0.58 & 0.78 & 0.66 \\\hline
\end{tabular}
\caption{Performances of different models on emotional trigger extraction from Tweets.}
\label{trigger}
\end{table*}
The precision and recall for segmenting triggering event phrases are
reported in Table 3. We observe a significant performance boost with linguistic features such as POS and dependency features. This is mainly due to the small size of the labeled dataset. The best model achieves an F1 score of 0.66.
\input{figs/angry.tex}
\input{figs/worry.tex}
\subsection{Clustering Trigger Mentions}
Since different extracted tokens may refer to the same concept or topic, we would like to cluster the extracted trigger mentions.
The method of supervised classification is unsuitable for this purpose since (1) it is hard to predefine a list of potential triggers to people's anger or worry; (2) it is extremely labor-intensive to annotate tweets with worry types or anger types and (3) these types may change over time.
For these reasons, we decided to use semi-supervised approaches that will automatically induce worry/anger types that match the data.
We adopt an approach based on LDA \cite{blei2003latent}. It was inspired by work on unsupervised information extraction \cite{chambers2011template,ritter2012open,li2014major}.
We use the emotion {\it anger} to illustrate how trigger mentions are clustered.
Each extracted trigger mentions for {\it anger} is modeled as a mixture of {\it anger} types.
Here we use subcategory, type, and topic interchangeablely, all referring to the cluster of similar mentions.
Each topic is characterized by a distribution over triggers, in addition to a distribution over dates on which a user talks about the topic.
Taking dates into account encourages triggers that are mentioned on the same date to be assigned to the same topic.
We used collapsed Gibbs Sampling \cite{griffiths2004finding} for inference.
For each emotion, we ran Gibbs Sampling with 20 topics for 1,000 iterations, obtaining the hidden variable assignments in the last iteration.
Then we manually inspected the top mentions for different topics and abandoned the incoherent ones.
The daily intensity score for a given subcategory $k$ belonging to emotion $y$ is given as follows:
\begin{equation}
S(t,y, k) = \frac{1}{|X_t|} \sum_{x\in X_t} p (k|x) {\bf I}(y_x = y)
\end{equation}
where $p(k|x)$ is computed based on the parameters of the latent variable model.
\subsection{Analyses}
We report the top triggers for different emotions in Table \ref{top-entity}.
We adopt a simple strategy of reporting the most frequent triggers for different emotions.
For {\it sadness}, the most frequent triggering events and topics are being test positive, and the death of families and friends.
For {\it anger}, the top triggers are shutdown, quarantine and other mandatory rules.
People also express their anger towards
public figures such as
President Donald Trump,
Mike Pence,
along with China and Chinese.
For {\it worry}, the top triggers include jobs, getting the virus, payments and families.
For {\it happiness}, the top triggers are recovering from the disease, city reopening and returning to work.
For {\it surprise}, the public are mostly surprised by the virus itself, its spread and the mass deaths it caused.
Next we report the results of the mention clustering for {\it anger} and {\it worry} in Tables \ref{anger} and \ref{worry}, respectively.
The unsupervised clustering reveals clearer patterns in the triggering events: top subcategories for {\it anger} include China with racist words such as {\it chink} and {\it chingchong}; lockdown and social distancing;
public figures like President Donal Trump and Mike Pence;
treatments in hospitals, and the increasing cases and deaths;
Table \ref{anger} displays the change of intensity scores for the subcategories of {\it anger}.
We observe a sharp increase in public anger toward China and Chinese around March 20th,
in coincidence with President Donald Trump calling coronavirus 'Chinese virus' in his tweets.
Public anger towards the lockdown sharply escalated in mid-March, but decreased a bit after late April when some places started to reopen.
Top subcategories for {\it worry} include syndromes for COVID-19, finance and economy, families, jobs and food, and increasing cases and deaths.
Table \ref{worry} displays the change of intensity scores for subcategories of {\it worry}.
People increasingly worried about families over time.
It is interesting to see that the worry about finance and economy started going up in mid-February, earlier than other subcategories.
\section{Conclusion}
In this paper, we
perform analyses on
topic trends, sentiments and emotions of the public in the time of COVID-19 on social media.
By tracking the change of public emotions over time,
our work reveals how the general public reacts to different events and government policy.
Our study provides a computational approach to understanding public affect towards the pandemic in real-time, and could help create better solutions and interventions
for fighting this global crisis.
| e060203c53a612a109f28bd7559c97390ba16379 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{S}{everal} applications use information that is encoded
as vectors of integers, either directly or indirectly. Furthermore,
these vectors are affected by noise that may increase or decrease
entries of the vectors by a limited amount. We mention a few of these
examples: In high-density magnetic recording channels, information is
stored in the lengths of runs of $0$'s. Various phenomena may cause
the reading process to shift the positions of $1$'s (peak-shift
error), thereby changing the length of adjacent runs of $0$'s by a
limited amount (e.g., see \cite{KuzVin93,LevVin93}). In flash
memories, information is stored in the charge levels of cells in an
array. However, retention (slow charge leakage), and inter-cell
interference, may cause charge levels to move, usually, by a limited
amount (e.g., see~\cite{CasSchBohBru10}). More recently, in some
DNA-storage applications, information is stored in the lengths of
homopolymer runs. These however, may end up shorter or longer than
planned, usually by a limited amount, due to variability in the
molecule-synthesis process (see~\cite{JaiFarSchBru20}).
In all of the applications mentioned above, an integer vector
$\mathbf{v}\in\mathbb{Z}^n$ encodes information. If at most $t$ of its entries suffer
an increase by as much as $k_+$, or a decrease by as much as $k_-$, we
can write the corrupted vector as $\mathbf{v}+\mathbf{e}$, where $\mathbf{e}$ resides
within a shape we call the \emph{$(n,t,k_+,k_-)$-error-ball}, and is
defined as
\begin{equation}
\label{eq:ball}
{\mathcal B}(n,t,\kp,\km)\triangleq \set*{\mathbf{x}=(x_1,x_2,\ldots,x_n)\in \mathbb{Z}^n ; -k_- \leq x_i \leq k_+ \text{ and } \wt(\mathbf{x}) \leq t },
\end{equation}
where $\wt(\mathbf{x})$ denotes the Hamming weight of $\mathbf{x}$.
It now follows that an error-correcting code in this setting is
equivalent to a packing of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$, a covering code is
equivalent to a covering of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$, and a perfect code is
equivalent to a tiling of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$. An example of
$\mathcal{B}(3,2,2,1)$ is shown in Fig.~\ref{fig:qc}.
\begin{figure}
\begin{center}
\includegraphics{qc.eps}
\end{center}
\caption{A depiction of $\mathcal{B}(3,2,2,1)$ where each point in
$\mathcal{B}(3,2,2,1)$ is shown as a unit cube.}
\label{fig:qc}
\end{figure}
A significant amount of works has been devoted to
tiling/packing/covering of $\mathbb{Z}^n$ by these shapes, albeit, almost
exclusively for the case of $t=1$. When packing and tiling are
concerned, the \emph{cross}, $\mathcal{B}(n,1,k,k)$, and semi-cross,
$\mathcal{B}(n,1,k,0)$ have been extensively researched, e.g., see
\cite{Ste84,HamSte84,HicSte86,SteSza94,KloLuoNayYar11} and the many
references therein. This was extended to \emph{quasi-crosses},
$\mathcal{B}(n,1,k_+,k_-)$, in \cite{Sch12}, creating a flurry of activity on
the subject
\cite{YarKloBos13,Sch14,ZhaGe16,ZhaZhaGe17,ZhaGe18,YeZhaZhaGe20}. To
the best of our knowledge,
\cite{Ste90,KloLuoNayYar11,BuzEtz12,WeiSchwartz} are the only works to
consider $t\geq 2$. \cite{Ste90,BuzEtz12} considered tiling a notched
cube (or a ``chair''), which for certain parameters becomes
$\mathcal{B}(n,n-1,k,0)$, while \cite{KloLuoNayYar11} considered packing the
same ball $\mathcal{B}(n,n-1,k,0)$. Among others, \cite{WeiSchwartz} recently
studied the tiling problem in the most general case, i.e., tiling
${\mathcal B}(n,t,\kp,\km)$ for $t \geq 2$. Covering problems have also been studied,
though only when $t=1$,~\cite{CSW2014,Klove2016,KloveSchwartz2014}.
The main goal of this paper is to study packing and covering of $\mathbb{Z}^n$
by ${\mathcal B}(n,t,\kp,\km)$ when $t\geq 2$. We provide explicit constructions for both
packings and coverings, as well some non-constructive existence
results. In particular, we demonstrate the existence of packings with
asymptotic packing density $\Omega(1)$ (as $n$ tends to infinity) for
some sets of $(t,k_+,k_-)$, and the existence of coverings with
density $O(1)$ for any given $(t,k_+,k_-)$. Additionally, we
generalize the concept of packing to $\lambda$-packing, which works in
conjunction with the list-decoding framework and list size $\lambda$.
We show the existence of $\lambda$-packings with density
$O(n^{-\epsilon})$ for any $(t,k_+,k_-)$ and arbitrarily small
$\epsilon>0$, while maintaining a list size
$\lambda=O(\epsilon^{-t})$, which does not depend on $n$. Our results
are summarized at the end of this paper, in Table~\ref{tab:summary}.
The paper is organized as follows. We begin, in
Section~\ref{sec:prelim}, by providing notation and basic known
results used throughout the paper. Section~\ref{sec:packing} is
devoted to the study of packings. This is generalized in
Section~\ref{sec:genpack} to $\lambda$-packings. In
Section~\ref{sec:covering} we construct coverings. Finally, we
conclude in Section~\ref{sec:conclusion} by giving a summary of the
results as well as some open problems.
\section{Preliminaries}
\label{sec:prelim}
For integers $a\leq b$ we define $[a,b]\triangleq\set*{a,a+1,\dots,b}$ and
$[a,b]^*\triangleq [a,b]\setminus\set*{0}$. We use $\mathbb{Z}_m$ to denote the
cyclic group of integers with addition modulo $m$, and $\mathbb{F}_q$ to
denote the finite field of size $q$.
A \emph{lattice} $\Lambda\subseteq\mathbb{Z}^n$ is an additive subgroup of
$\mathbb{Z}^n$ (sometimes called an \emph{integer lattice}). A lattice
$\Lambda$ may be represented by a matrix $\mathcal{G}(\Lambda)\in\mathbb{Z}^{n\times
n}$, the span of whose rows (with integer coefficients) is
$\Lambda$. From a geometric point of view, when viewing $\Lambda$ inside
$\mathbb{R}^n$, a \emph{fundamental region} of $\Lambda$ is defined as
\[ \Pi(\Lambda)\triangleq \set*{ \sum_{i=1}^n c_i\mathbf{v}_i ; c_i\in\mathbb{R}, 0\leq c_i < 1 },\]
where $\mathbf{v}_i$ is the $i$-th row of $\mathcal{G}(\Lambda)$. It is well known
that the volume of $\Pi(\Lambda)$ is $\abs{\det(\mathcal{G}(\Lambda))}$, and
is independent of the choice of $\mathcal{G}(\Lambda)$. We therefore denote
\[ \vol(\Lambda)\triangleq \vol(\Pi(\Lambda)) = \abs*{\det(\mathcal{G}(\Lambda))}.\]
In addition, if $\vol(\Lambda)\neq 0$ then
\[ \vol(\Lambda)=\abs*{\mathbb{Z}^n / \Lambda}.\]
We say $\mathcal{B}\subseteq\mathbb{Z}^n$ \emph{packs} $\mathbb{Z}^n$ by $T\subseteq\mathbb{Z}^n$, if
the translates of $\mathcal{B}$ by elements from $T$ do not intersect,
namely, for all $\mathbf{v},\mathbf{v}'\in T$, $\mathbf{v}\neq\mathbf{v}'$,
\[ (\mathbf{v}+\mathcal{B})\cap(\mathbf{v}'+\mathcal{B})=\varnothing.\]
We say $\mathcal{B}$ \emph{covers} $\mathbb{Z}^n$ by $T$ if
\[ \bigcup_{\mathbf{v}\in T} (\mathbf{v}+\mathcal{B}) = \mathbb{Z}^n.\]
If $\mathcal{B}$ both packs and covers $\mathbb{Z}^n$ by $T$, then we say $\mathcal{B}$
\emph{tiles} $\mathbb{Z}^n$ by $T$. The \emph{packing density} (or
\emph{covering density}, respectively) of $\mathcal{B}$ by $T$ is defined as
\[ \delta\triangleq \lim_{\ell\to\infty} \frac{\abs*{[-\ell,\ell]^n\cap T}\cdot\abs{\mathcal{B}}}{\abs*{[-\ell,\ell]^n}}.\]
When $T=\Lambda$ is some lattice, we call these lattice packings and
lattice coverings, respectively. The density then takes on a simpler form
\[\delta = \frac{\abs{\mathcal{B}} }{\vol(\Lambda)}.\]
Throughout the paper, the object we pack and cover $\mathbb{Z}^n$ with, is the
error ball, ${\mathcal B}(n,t,\kp,\km)$, defined in \eqref{eq:ball}. We conveniently
observe that for all integers $n\geq 1$, $0\leq t\leq n$, $0\leq
k_-\leq k_+$, we have
\[ \abs*{{\mathcal B}(n,t,\kp,\km)}=\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i.\]
\subsection{Lattice Packing/Covering/Tiling and Group Splitting}
Lattice packing, covering, and tiling of $\mathbb{Z}^n$ with
$\mathcal{B}(n,t,k_+,k_-)$, in connection with group splitting, has a long
history when $t=1$ (e.g., see \cite{Ste67}), called lattice tiling by
crosses if $k_+=k_-$ (e.g., \cite{Ste84}), semi-crosses when $k_-=0$
(e.g., \cite{Ste84,HamSte84,HicSte86}), and quasi-crosses when
$k_+\geqk_-\geq 0$ (e.g., \cite{Sch12,Sch14}). For an excellent
treatment and history, the reader is referred to~\cite{SteSza94} and
the many references therein. Other variations, keeping $t=1$
include~\cite{Tam98,Tam05}. More recent results may be found
in~\cite{YeZhaZhaGe20} and the references therein.
For $t\geq 2$, an extended definition of group splitting in connection
with lattice tiling is provided in \cite{WeiSchwartz}. In the
following, we modify this definition to distinguish between lattice
packings, coverings, and tilings.
\begin{definition}
\label{def:split}
Let $G$ be a finite Abelian group, where $+$ denotes the group
operation. For $m\in\mathbb{Z}$ and $g\in G$, let $mg$ denote $g+g+\dots+g$
(with $m$ copies of $g$) when $m>0$, which is extended in the
natural way to $m\leq 0$. Let $M\subseteq\mathbb{Z}\setminus\set*{0}$ be a
finite set, and $S=\set*{s_1,s_2,\dots,s_n}\subseteq G$.
\begin{enumerate}
\item
If the elements $\mathbf{e}\cdot (s_1,\dots,s_n)$, where $\mathbf{e}\in
(M\cup\set{0})^n$ and $1\leq \wt(\mathbf{e})\leq t$, are all distinct and
non-zero in $G$, we say the
set $M$ partially $t$-splits $G$ with splitter set $S$, denoted
\[G \geq M\diamond_t S.\]
\item
If for every $g\in G$ there exists a vector $\mathbf{e}\in
(M\cup\set{0})^n$, $\wt(\mathbf{e})\leq t$, such that $g=\mathbf{e}\cdot
(s_1,\dots,s_n)$, we say the set $M$ completely $t$-splits $G$
with splitter set $S$, denoted
\[G \leq M\diamond_t S.\]
\item
If $G \geq M\diamond_t S$ and $G \leq M\diamond_t S$ we say $M$
$t$-splits $G$ with splitter set $S$, and write
\[ G = M\diamond_t S.\]
\end{enumerate}
\end{definition}
In our context, since we are interested in packing and covering with
${\mathcal B}(n,t,\kp,\km)$, then in the previous definition, we need to take $M\triangleq
[-k_-,k_+]^*$. Thus, the following two theorems show the equivalence
of partial $t$-splittings with $M$ and lattice packings of ${\mathcal B}(n,t,\kp,\km)$,
summarizing Lemma 3 and Lemma 4 in\cite{BuzEtz12}, and similarly for
complete $t$-splittings and lattice coverings.
\begin{theorem}
\label{th:lattotile}
Let $G$ be a finite Abelian group, $M\triangleq [-k_-,k_+]^*$, and
$S=\set*{s_1,\dots,s_n}\subseteq G$. Define $\phi:\mathbb{Z}^n\to G$ as
$\phi(\mathbf{x})\triangleq\mathbf{x}\cdot(s_1,\dots,s_n)$ and let
$\Lambda\triangleq\ker\phi$ be a lattice.
\begin{enumerate}
\item
If $G \geq M\diamond_t S$, then $\mathcal{B}(n,t,k_+,k_-)$ packs $\mathbb{Z}^n$ by
$\Lambda$.
\item
If $G \leq M\diamond_t S$, then $\mathcal{B}(n,t,k_+,k_-)$ covers $\mathbb{Z}^n$ by
$\Lambda$.
\end{enumerate}
\end{theorem}
\begin{IEEEproof}
For packing, see Lemma 4 in \cite{BuzEtz12}. For covering, denote
$\mathcal{B}\triangleq\mathcal{B}(n,t,k_+,k_-)$. Assume $\mathbf{x}\in\mathbb{Z}^n$. Since $G \leq
M\diamond_t S$, there exists a vector $\mathbf{e}\in\mathcal{B}$ such that
$\phi(\mathbf{x})=\phi(\mathbf{e})$. Then $\mathbf{v}\triangleq \mathbf{x}-\mathbf{e}\in\Lambda$, and
$\mathbf{x}\in\mathbf{v}+\mathcal{B}$.
\end{IEEEproof}
In the theorem above, for $G \geq M\diamond_t S$, since the quotient
group $\mathbb{Z}^n / \Lambda$ is isomorphic to the image of $\phi$, which is
a subgroup of $G$, we have $\vol(\Lambda)\leq \abs{G}$. Then the
packing density of $\Lambda$ is
\[\delta = \frac{\abs{{\mathcal B}(n,t,\kp,\km)} }{\vol(\Lambda)}\geq\frac{\abs{{\mathcal B}(n,t,\kp,\km)} }{\abs{G}} .\]
For $G \leq M\diamond_t S$, $\vol(\Lambda)= \abs{G}$, and the
covering density of $\Lambda$ is
\[\delta = \frac{\abs{{\mathcal B}(n,t,\kp,\km)} }{\vol(\Lambda)}=\frac{\abs{{\mathcal B}(n,t,\kp,\km)} }{\abs{G}} .\]
It is known that a lattice packing implies a partial splitting. While
not of immediate use to us in this paper, we do mention that an
analogous claim is also true for lattice coverings, as the following theorem
shows.
\begin{theorem}
Let $\Lambda\subseteq\mathbb{Z}^n$ be a lattice. Define $G\triangleq \mathbb{Z}^n /
\Lambda$. Let $\phi:\mathbb{Z}^n\to G$ be the natural homomorphism, namely
the one that maps any $\mathbf{x}\in\mathbb{Z}^n$ to the coset of $\Lambda$ in
which it resides, and then $\Lambda=\ker\phi$. Finally, let $\mathbf{e}_i$
be the $i$-th unit vector in $\mathbb{Z}^n$ and set $s_i\triangleq\phi(\mathbf{e}_i)$ for
all $1\leq i\leq n$ and $S\triangleq\set*{s_1,s_2,\dots,s_n}$.
\begin{enumerate}
\item If $\mathcal{B}(n,t,k_+,k_-)$ packs $\mathbb{Z}^n$ by $\Lambda$, then $G\geq M\diamond_t S$;
\item if $\mathcal{B}(n,t,k_+,k_-)$ covers $\mathbb{Z}^n$ by $\Lambda$, then $G\leq M\diamond_t S$,
\end{enumerate}
where $M\triangleq [-k_-,k_+]^*$.
\end{theorem}
\begin{IEEEproof}
The packing see Lemma 3 in \cite{BuzEtz12}. Now we prove the claim
for covering. Let $\Lambda+\mathbf{x} \in G$ be any element of $G$. Since
${\mathcal B}(n,t,\kp,\km)$ covers $\mathbb{Z}^n$ by $\Lambda$, there exist $\mathbf{v}\in\Lambda$ and
$\mathbf{e}\in{\mathcal B}(n,t,\kp,\km)$ such that $\mathbf{x}=\mathbf{v}+\mathbf{e}$. This means
\[\Lambda+\mathbf{x}=\phi(\mathbf{x})=\phi(\mathbf{v})+\phi(\mathbf{e})=\phi(\mathbf{e})=\mathbf{e}\cdot(s_1,\dots,s_n),\]
which completes the proof.
\end{IEEEproof}
Finally, a connection between perfect codes in the Hamming metric, and
lattice tilings with ${\mathcal B}(n,t,\kp,\km)$ was observed in~\cite{WeiSchwartz}. We repeat
a theorem we shall later generalize.
\begin{theorem}[Theorem~3 in \cite{WeiSchwartz}]
\label{th:perfectcode}
In the Hamming metric space, let $C$ be a perfect linear
$[n,k,2t+1]$ code over $\mathbb{F}_p$, with $p$ a prime. If $k_++k_-+1=p$,
then
\[ \Lambda\triangleq \set*{ \mathbf{x}\in\mathbb{Z}^n ; (\mathbf{x}\bmod p)\in C}\]
is a lattice, and $\mathcal{B}(n,t,k_+,k_-)$ tiles $\mathbb{Z}^n$ by
$\Lambda$.
\end{theorem}
\section{Constructions of Lattice Packings}
\label{sec:packing}
In this section we describe several constructions for packings of
${\mathcal B}(n,t,\kp,\km)$. We begin by showing how to translate codes in the Hamming
metric into lattices that pack ${\mathcal B}(n,t,\kp,\km)$. Apart from a single case,
these have vanishing density. The motivation for showing these
``off-the-shelf'' constructions is to create a baseline against which
we measure our tailor-made constructions that appear later. These use
$B_t[N;1]$ sequences, or take inspiration from constructions of sets
with no arithmetic progression, to construct codes that improve upon
the baseline.
\subsection{Constructions Based on Error-Correcting Codes}
Theorem~\ref{th:perfectcode} can be easily modified to yield the
following construction, the proof of which is the same as that of
\cite[Theorem 3]{WeiSchwartz} and we omit here to avoid unnecessary
repetition.
\begin{theorem}
\label{th:linearcode}
In the Hamming metric space, let $C$ be a linear $[n,k,2t+1]$ code
over $\mathbb{F}_p$, with $p$ a prime. If $0\leq k_++k_- < p$ are integers,
then
\[ \Lambda\triangleq \set*{ \mathbf{x}\in\mathbb{Z}^n ; (\mathbf{x}\bmod p)\in C}\]
is a lattice, and $\mathcal{B}(n,t,k_+,k_-)$ packs $\mathbb{Z}^n$ by
$\Lambda$.
\end{theorem}
Since $\mathcal{B}(n,t,k_+,k_-)$ packs $\mathbb{Z}^n$ by $\Lambda$, the lattice
$\Lambda$ is an error-correcting code over $\mathbb{Z}$ for asymmetric
limited-magnitude errors. We note that a similar construction of
error-correcting codes over a finite alphabet for asymmetric
limited-magnitude errors was presented in \cite{CasSchBohBru10} and
the decoding scheme therein can be adapted here as follows. Let
$\mathbf{x}\in\Lambda$ be a codeword, and $\mathbf{y} \in\mathbf{x}+{\mathcal B}(n,t,\kp,\km)$ be the channel
output. Denote $\boldsymbol{\psi}= \mathbf{y} \pmod{p}$. Run the decoding
algorithm of the linear $[n,k,2t+1]$ code on $\boldsymbol{\psi}$ and
denote the output as $\boldsymbol{\phi}$. Then $\boldsymbol{\phi}$ is
a codeword of the linear code over $\mathbb{F}_p$ and it is easy to see that
$\boldsymbol{\phi}=\mathbf{x} \pmod{p}$. Thus $\mathbf{y}-\mathbf{x}\equiv\boldsymbol{\psi}
- \boldsymbol{\phi} \pmod{p}$. Denote
$\boldsymbol{\epsilon}=\boldsymbol{\psi} - \boldsymbol{\phi} \pmod{p}$
and let $\mathbf{e}=(e_1,e_2,\ldots,e_n)$ where
\[
e_i \triangleq \begin{cases} \epsilon_i, \textup{\ \ if $0\leq \epsilon_i \leq k_+$;} \\
\epsilon_i-p, \textup{\ \ otherwise.}\\
\end{cases}
\]
Then $\mathbf{x}$ can be decoded as $\mathbf{x}=\mathbf{y}-\mathbf{e}$.
Now let us look at the packing density.
\begin{corollary}
\label{cor:linearcode}
Let $\Lambda$ be the lattice constructed in
Theorem~\ref{th:linearcode}. Then $\vol(\Lambda)=p^{n-k}$
and the packing density is
\[\delta=\frac{\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i}{ p^{n-k} }.\]
\end{corollary}
\begin{IEEEproof}
In Theorem~\ref{th:linearcode}, the quotient group $\mathbb{Z}^n/ \Lambda$
is isomorphic to the group $\mathbb{Z}_p^{n-k}$ (see Example 3 and Example 4
in \cite{WeiSchwartz}). The claim is then immediate.
\end{IEEEproof}
When $t$ is small, we may use BCH codes as the input to construct the
lattice packing.
\begin{theorem}[Primitive narrow-sense BCH codes {\cite[Theorem 10]{aly2007quantum}}]\label{primitiveBCH} Let $p$ be a prime.
Fix $m\ge 1$ and $2\le d\le p^{\ceil{m/2}}-1$. Set $n=p^m-1$. Then
there exists an $[n,k,d]$-code over $\mathbb{F}_p$ with
\[k=n-\ceil{(d-1)(1-1/p)}m.\]
\end{theorem}
\begin{corollary}\label{cor:lc}
Let $\psi(x)$ be the smallest prime not smaller than $x$ \footnote{It
is known that $\psi(x)\leq x+x^{21/40}$ \cite{BakHarPin01}, and
conjectured that $\psi(x)=x+O(\log x)$.} and denote
$p\triangleq\psi(k_++k_-+1)$. Let $m,t$ be positive integers such that
$2t\le p^{\ceil{m/2}}-2$, and set $n=p^m-1$. Then $\mathbb{Z}^n$ can be
lattice packed by ${\mathcal B}(n,t,\kp,\km)$ with density
\[\delta = \frac{\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i}{(n+1)^{\ceil{2t(1-{1}/p)}} }.\]
\end{corollary}
\begin{IEEEproof}
Simply combine Theorem~\ref{primitiveBCH} with
Corollary~\ref{cor:linearcode}.
\end{IEEEproof}
Note that if $k_+=1$ and $k_-=0$, then the packing density in
Corollary~\ref{cor:lc} is $\delta=\frac{\sum_{i=0}^t
\binom{n}{i}}{(n+1)^{t}}=\frac{1}{t!}+o(1)$ (when $t$ is fixed and
$n$ tends to infinity). However, for all the other values of $k_+$ and
$k_-$, namely $p\geq 3$, the density always vanishes when $n$ tends to
infinity, i.e., $\delta=\Theta(n^{t-\ceil{2t(1-{1}/p)}})$. In the remainder
of this section, we will present some constructions to provide lattice
packings of higher density.
Perfect codes were used in \cite{WeiSchwartz} obtain lattice tilings,
i.e., lattice packings with density $1$. Similarly, it is possible to
use quasi-perfect linear codes to obtain lattice packings with high
densities.
\begin{corollary}\label{cor:quasiperfectlc} Assume that $1\leq k_++k_-\leq 2$
are non-negative integers.
\begin{enumerate}
\item
Let $m$ be a positive integer and $n=(3^m+1)/2$. Then $\mathbb{Z}^n$ can
be lattice-packed by $\mathcal{B}(n,2,k_+,k_-)$ with density
\[\delta = \frac{ {n\choose 2}(k_++k_-)^2 +n(k_++k_-) +1 }{(2n-1)^2}.\]
\item
Let $m\geq 3$ be an odd integer and $n=(3^m-1)/2$. Then $\mathbb{Z}^n$ can
be lattice packed by $\mathcal{B}(n,2,k_+,k_-)$ with density
\[\delta = \frac{ {n\choose 2}(k_++k_-)^2 +n(k_++k_-) +1 }{(2n+1)^2}.\]
\end{enumerate}
\end{corollary}
\begin{IEEEproof}
For the first case, we take a $[(3^m+1)/2, (3^m+1)/2-2m,5]_3$ code
from \cite{GashkovSidelnikov1986} as the input of
Theorem~\ref{th:linearcode} to obtain the lattice packing, while for
the second, we take a $[(3^m-1)/2, (3^m-1)/2-2m,5]_3$ code from
\cite{DanevDodunekov2008} as the input.
\end{IEEEproof}
We note that \cite{LiHelleseth2016} presented some binary
quasi-perfect linear codes with minimum distance $5$, which can give
rise to packings of $\mathcal{B}(n,2,1,0)$. It has been checked out that the
corresponding densities are asymptotically the same as that in
Corollary~\ref{cor:lc}, i.e.,
$\frac{1}{2}+o(1)$. \cite{LiHelleseth2016} also studied $p$-ary
quasi-perfect linear codes with $p\geq 3$. However, the minimum
distances of those codes are no more that $4$. So they cannot be used
to obtain packings of ${\mathcal B}(n,t,\kp,\km)$ with $t \geq 2$.
The following theorem uses non-linear codes to construct non-lattice
packing, the proof of which is the same as that of \cite[Theorem
5]{CasSchBohBru10} and we omit here to avoid unnecessary repetition.
\begin{theorem}
\label{th:nonlinearcode}
In the Hamming metric space, let $C$ be a $q$-ary $(n,M,2t+1)$
code. Denote
\[ V\triangleq \set*{ \mathbf{v}\in\mathbb{Z}^n ; (\mathbf{v}\bmod q)\in C}.\]
If $k_++k_- < q$, then for any distinct $\mathbf{v},\mathbf{v}'\in V$, we have
$(\mathbf{v}+{\mathcal B}(n,t,\kp,\km)) \cap (\mathbf{v}'+{\mathcal B}(n,t,\kp,\km))=\varnothing$, namely, ${\mathcal B}(n,t,\kp,\km)$ can
pack $\mathbb{Z}^n$ by $V$.
\end{theorem}
\begin{corollary}
\label{cor:nonlinearcode}
The density of the packing of $\mathbb{Z}^n$ constructed in
Theorem~\ref{th:nonlinearcode} is
\[\delta=\frac{M\cdot\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i}{ q^{n} }.\]
\end{corollary}
\begin{IEEEproof}
Note that the set $V$ constructed in Theorem~\ref{th:nonlinearcode}
has period $q$ in each coordinate. Thus, the packing density of
$\mathbb{Z}^n$ is equal to the packing density of $\mathbb{Z}_q^n$, which is
$\frac{M}{q^n}\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i.$
\end{IEEEproof}
\begin{corollary}\label{cor:perparata}
Let $m \geq 4$ be even integer and let $n=2^m-1$. Then $\mathbb{Z}^n$ can be
packed by $\mathcal{B}(n,2,1,0)$ with density
\[\delta = \frac{ {n\choose 2} +n +1 }{(n+1)^2/2}=1-\frac{n-1}{(n+1)^2}.\]
\end{corollary}
\begin{proof}
We take a binary $(2^m-1, 2^{2^m-2m},5)$ Preparata code
\cite{Preparata1968} as the input of Theorem~\ref{th:nonlinearcode}
to obtain the packing.
\end{proof}
\subsection{Construction Based on $B_t[N;1]$ Sets for $(k_+,k_-)=(1,0)$ or $(1,1)$}
A subset $A\subseteq\mathbb{Z}$ is called a \emph{$B_t[g]$ set} if every
integer can be written in at most $g$ different ways as a sum of $t$
(not necessary distinct) elements of $A$ (e.g., see
\cite[Section~4.5]{TaoVu06} and the many references therein). In this
section, however, we require $B_t[1]$ sets with a somewhat stronger
property. Specifically, a subset $A$ of $\mathbb{Z}_N$ is called a
\emph{$B_t[N;1]$ set} if the sums of any $t$ (not necessary distinct)
elements of $A$ are all different modulo $N$. Bose and Chowla
\cite{BoseChowla} presented two classes of $B_t[N;1]$ sets.
\begin{theorem}\cite[Theorem 1 and Theorem 2]{BoseChowla}\label{thm:Btset}
Let $q$ be a prime power and $t$ be a positive integer. Let
$\alpha_0=0,\alpha_1,\alpha_2,\ldots,\alpha_{q-1}$ be all the
different elements of $\mathbb{F}_q$.
\begin{enumerate}
\item
Let $\xi$ be a primitive element of the extended field
$\mathbb{F}_{q^t}$. For each $0\leq i \leq q-1$, let $d_i \in \mathbb{Z}_{q^{t}-1}$
be such that
\[\xi^{d_i}=\xi+\alpha_i.\]
Then the set $S_1\triangleq \set*{d_i;0\leq i \leq q-1}$ is a
$B_t[q^t-1;1]$ set of size $q$.
\item
Let $\eta$ be a primitive element of the extended field
$\mathbb{F}_{q^{t+1}}$. For each $0\leq i \leq q-1$, let $\beta_i \in \mathbb{F}_q$
and $s_i \in \mathbb{Z}_{(q^{t+1}-1)/(q-1)}$ such that
\[ \beta_i\eta^{s_i}=\eta+\alpha_i.\]
Then the set $S_2\triangleq \set*{s_i;0\leq i \leq q-1} \cup \set{0}$ is a
$B_t[(q^{t+1}-1)/(q-1);1]$ set of size $q+1$.
\end{enumerate}
\end{theorem}
\begin{theorem}\label{thm:con-Btset-1}
Let $A$ be a $B_t[N;1]$ set which contains $0$. Denote $S\triangleq A\setminus\set{0}$. Then $\mathbb{Z}_N \ge \set{1} \diamond_t S$.
\end{theorem}
\begin{IEEEproof}
Suppose to the contrary that
$\set*{s_{i_1},s_{i_2},\ldots,s_{i_{\ell}}}$ and
$\set*{s_{j_1},s_{j_2},\ldots,s_{j_{r}}}$ are two distinct
subsets of $S$ such that
\[s_{i_1}+s_{i_2}+\ldots+s_{i_{\ell}}\equiv s_{j_1}+s_{j_2}+\ldots+s_{j_{r}} \pmod{N},\]
where $\ell,r\leq t$.
Then we have
\[ \underbrace{0+0+\cdots +0}_{t-\ell}+ s_{i_1}+s_{i_2}+\ldots+s_{i_{\ell}}\equiv \underbrace{0+0+\cdots +0}_{t-r}+s_{j_1}+s_{j_2}+\ldots+s_{j_{r}} \pmod{N},\]
which contradicts that $S\cup \set{0}$ is a $B_t[N;1]$ set.
\end{IEEEproof}
The following result slightly improves upon the density obtained in
Corollary~\ref{cor:lc} for lattice packings of $\mathcal{B}(n,t,1,0)$.
\begin{corollary}
\label{cor:btsets}
\begin{enumerate}
\item
If $n+1$ is a prime power, then for any $2\leq t \leq n$, there is
a lattice packing of $\mathbb{Z}^n$ by $\mathcal{B}(n,t,1,0)$ with density
\[\delta=\frac{\sum_{i=0}^t\binom{n}{i} }{(n+1)^t-1}=\frac{1}{t!}+o(1).\]
\item
If $n$ is a prime power, then for any $2\leq t \leq n$, there is a
lattice packing of $\mathbb{Z}^n$ by $\mathcal{B}(n,t,1,0)$ with
density
\[\delta=\frac{\sum_{i=0}^t\binom{n}{i}}{(n^{t+1}-1)/(n-1)}=\frac{1}{t!}+o(1).\]
\end{enumerate}
\end{corollary}
\begin{IEEEproof}
Note that if $\set*{s_1,s_2,\ldots,s_n}$ is a $B_t[N;1]$ set, then
$\set*{0,s_2-s_1,s_3-s_1,\ldots,s_n-s_1}$ is also a $B_t[N;1]$ set,
which contains $0$. Hence, combining Theorem~\ref{thm:Btset} and
Theorem~\ref{thm:con-Btset-1}, together with
Theorem~\ref{th:lattotile}, we prove the claim.
\end{IEEEproof}
In general, we do not have an efficient decoding scheme for the
lattice code obtained from Theorem~\ref{thm:con-Btset-1}. However, for
the lattice code $\Lambda_{S_2\setminus\set{0}}$ obtained from the
$B_t[(q^{t+1}-1)/(q-1);1]$ set $S_2$ in Theorem~\ref{thm:Btset}, we
have the following decoding algorithm (summarized in
Algorithm~\ref{alg:decoding}). Let $n=q$ and let
$S_2\setminus\set{0}=\set{s_0,\dots,s_{n-1}}$ be defined as in
Theorem~\ref{thm:Btset}. Let $\mathbf{x}\in\Lambda_{S_2\setminus\set{0}}$ be
a codeword and $\mathbf{y} \in \mathbf{x}+\mathcal{B}(n,t,1,0)$ be the channel output. Then
$\mathbf{x}\cdot (s_0,s_1,\ldots,s_{n-1})=0$, and $\mathbf{y}-\mathbf{x}$ is a binary
vector over $\set{0,1}$ of weight at most $t$. Let
$i_1,i_2,\ldots,i_r$ be the indices of the nonzero bits of $\mathbf{y}-\mathbf{x}$,
and denote $s=\mathbf{y}\cdot (s_0,s_1,\ldots,s_{n-1})$. We aim to recover
$i_1,i_2,\ldots,i_r$ from $s$. Since
\begin{align*}
s&=\mathbf{y}\cdot (s_0,s_1,\ldots,s_{n-1}) = \mathbf{y}\cdot (s_0,s_1,\ldots,s_{n-1}) -0 =\mathbf{y}\cdot (s_0,s_1,\ldots,s_{n-1})-\mathbf{x}\cdot (s_0,s_1,\ldots,s_{n-1}) \\
&= (\mathbf{y}-\mathbf{x})\cdot (s_0,s_1,\ldots,s_{n-1}) =\sum_{\ell=1}^r s_{i_\ell},
\end{align*}
we have that
\begin{equation}\label{eq:decode}
\parenv*{\prod_{\ell=1}^r \beta_{i_\ell}} \eta^s= \prod_{\ell=1}^r\parenv*{\beta_{i_\ell} \eta^{s_{i_\ell}}} =\prod_{\ell=1}^r(\eta+\alpha_{i_\ell}).
\end{equation}
Let $p(x)$ be the primitive polynomial of $\eta$ and $r(x) = x^s \bmod p(x)$. Then $r(\eta)=\eta^s$, and we substitute this in \eqref{eq:decode} to obtain
\[\parenv*{\prod_{\ell=1}^r \beta_\ell} r(\eta)=\prod_{\ell=1}^r(\eta+\alpha_{i_\ell}).\]
Since both the polynomials $(\prod_{\ell=1}^r \beta_{i_\ell}) r(x)$
and $\prod_{\ell=1}^r(x+\alpha_{i_\ell})$ are over $\mathbb{F}_q$ and have
degrees at most $t$, they should be the same; otherwise, $\eta$ is a
root of a nonzero polynomial of degree at most $t$, which contradicts
the fact that $\eta$ is a primitive element of $\mathbb{F}_{q^{t+1}}$. Thus,
we may solve $(\prod_{\ell=1}^r \beta_\ell) r(x)$ to find out
$\alpha_{i_1}, \alpha_{i_1},\ldots, \alpha_{i_r}$. Finally, we can subtract
$\sum_{\ell=1}^r \mathbf{e}_{i_\ell}$ from $\mathbf{y}$ to obtain $\mathbf{x}$.
\begin{algorithm}[ht]
\caption{Decoding algorithm for $\Lambda_{S_2\setminus\set{0}}$ from Theorem~\ref{thm:Btset}}
\label{alg:decoding}
\begin{algorithmic}[1]
\Statex \textbf{Input:} received vector $\mathbf{y}\in\mathbb{Z}^n$ suffering at most $t$ errors
\Statex \quad\quad\quad $S_2\setminus\set{0}=\set{s_0,\dots,s_{n-1}}$ from Theorem~\ref{thm:Btset}
\Statex \quad\quad\quad where $\eta$ is a root of a primitive polynomial $p(x)$ of degree $t+1$
\Statex \quad\quad\quad and where $\mathbb{F}_q\setminus\set{0}=\set{\alpha_1,\dots,\alpha_{q-1}}$.
\Statex \textbf{Output:} codeword $\mathbf{x}\in\Lambda_{S_2\setminus\set{0}}$ such that $\mathbf{y}\in\mathbf{x}+\mathcal{B}(n,t,1,0)$
\State $s \gets \mathbf{y} \cdot (s_0,s_1,\ldots,s_{n-1})$ \label{step:s}
\State $r(x)\gets x^s \bmod p(x)$ \label{step:r}
\For{$1\leq i\leq q-1$} \label{step:loop}
\If{$r(\alpha_i)=0$}
\State $\mathbf{y} \gets \mathbf{y} - \mathbf{e}_i$
\EndIf
\EndFor
\State \Return{$\mathbf{y}$}
\end{algorithmic}
\end{algorithm}
Let us analyze the time complexity of Algorithm~\ref{alg:decoding},
where we count the number of field operations in $\mathbb{F}_q$. The inner
product in Step~\ref{step:s} takes $O(n)$
operations. Step~\ref{step:r} is possible to compute in $O(t^2 \log
s)$ field operations (by using successive squaring and multiplication
by $x$ as necessary, taking a modulo $p(x)$ after each
iteration). Since $s \in \mathbb{Z}_{(q^{t+1}-1)/(q-1)}$ and $n=q$, it is
$O(t^3 \log n)$. Finally, the root search loop starting in
Step~\ref{step:loop} takes $O(tn)$ operations. Thus, in total, the
time complexity $O(tn+t^3\log n)$ field operations. If $t$ is
constant, then this is linear in the code length. As a final comment,
we point out that $q=O(n)$, and thus the basic field operations of
addition and multiplication may be realized in $O(\mathrm{polylog}(n))$ time.
We now move from packing $\mathcal{B}(n,t,1,0)$ to packing $\mathcal{B}(n,t,1,1)$. In
general, we note that one can use a $B_{h}[N;1]$ set with
$h=t(k_++k_-)$ to obtain a lattice packing of ${\mathcal B}(n,t,\kp,\km)$ in $\mathbb{Z}_n$ for
$k_++k_- \geq 2$. However, in this case, the density is
$O(n^{t(1-{k_++k_-})})$, which vanishes when $n$ tends to
infinity. Similarly, the lattice packing from Corollary~\ref{cor:lc}
also has vanishing density $O(n^{t-\ceil{2t(1-{1}/p)}})$. In the
following, we give a modified construction which uses a $B_{t}[N;1]$
set to obtain a lattice packing of $\mathcal{B}(n,t,1,1)$ with density
$\Omega(1)$.
\begin{theorem}\label{thm:con-Btset-pm1}
Let $A=\set{a_1,a_2,\dots,a_n}$ be a $B_t[N;1]$ set. In the group
$\mathbb{Z}_N \times \mathbb{Z}_{2t+1}$, construct a set
\[S\triangleq \set*{ (a_i,1) ; a_i \in A}.\]
Then $\mathbb{Z}_N \times \mathbb{Z}_{2t+1} \ge \set{-1,1} \diamond_t S$.
\end{theorem}
\begin{IEEEproof}
Suppose to the contrary that there are
$(a_{i_1},1),(a_{i_1},1),\dots,(a_{i_{\ell}},1)$ and
$(a_{j_1},1),(a_{j_2},1),\dots,(a_{j_{r}},1)$ in $S$ such that
\begin{equation}\label{eq:con-Bset}
\sum_{m=1}^{\ell'} (a_{i_m},1)-\sum_{m=\ell'+1}^{\ell}(a_{i_m},1) =
\sum_{m=1}^{r'} (a_{j_m},1)-\sum_{m=r'+1}^{r}(a_{j_m},1),
\end{equation}
where $0\leq \ell'\leq \ell \leq t$ and $0\leq r'\leq r \leq t$, and
the addition is over the group $\mathbb{Z}_N \times \mathbb{Z}_{2t+1}$.
The second coordinate of the equation above implies that
\[\ell-2\ell'\equiv r-2r' \pmod{2t+1}.\]
Since $0\leq \ell'\leq \ell \leq t$ and $0\leq r'\leq r \leq t$, we
have $\ell-2\ell', r-2r'\in [-t,t]$. It follows that $\ell-2\ell'=
r-2r',$ and so, $\ell'+r-r'=r'+\ell-\ell'$. Let
$\tau\triangleq\ell'+r-r'$. Then
\[\tau = \frac{\ell'+r-r'+r'+\ell-\ell'}{2} =\frac{\ell+r}{2} \leq t.\]
Rearranging the terms in the first coordinate of the equation
\eqref{eq:con-Bset}, we have
\begin{equation}\label{eq:con-Bset-2}
a_{i_1}+ a_{i_2}+\cdots+ a_{i_{\ell'}} + a_{j_{r'+1}} + \cdots + a_{j_{r}} \equiv a_{j_1}+ a_{j_2}+\cdots+ a_{j_{r'}} + a_{i_{\ell'+1}} + \cdots + a_{i_{\ell}} \pmod{N}.
\end{equation}
On both side of \eqref{eq:con-Bset-2}, there are $\tau$ terms. This contradicts the fact that $A$ is a $B_{t}[N;1]$ (and hence a $B_{\tau}[N;1]$ set for any $\tau \leq t$).
\end{IEEEproof}
Combining Theorem~\ref{thm:Btset} and Theorem~\ref{thm:con-Btset-pm1}, together with Theorem~\ref{th:lattotile}, we have the following result.
\begin{corollary}
\label{cor:kpm1}
If $n$ is a prime power, then for any $2\leq t \leq n$, there is a
lattice packing of $\mathbb{Z}^n$ by $\mathcal{B}(n,t,1,1)$ with density
\[\delta=\frac{\sum_{i=0}^t\binom{n}{i}2^i }{(2t+1)(n^t-1)}=\frac{2^t}{t!(2t+1)}+o(1).\]
If $n-1$ is a prime power, then for any $2\leq t \leq n$, there is a
lattice packing of $\mathbb{Z}^n$ by $\mathcal{B}(n,t,1,1)$ with density
\[\delta=\frac{\sum_{i=0}^t\binom{n}{i} 2^i}{(2t+1)((n-1)^{t+1}-1)/(n-2)}=\frac{2^t}{t!(2t+1)}+o(1).\]
\end{corollary}
\subsection{Constructions for $t=2$}
Whereas in the previous section we considered unconstrained $t$ but
only small values of $k_+,k_-$, in this section we focus on the case
of $t=2$ but unconstrained $k_+,k_-$.
We first present a construction based on $k$-fold Sidon sets. Such
sets were first defined in~\cite{Lazebnik2003} as a generalization of
Sidon sets. We repeat the definition here. Let $k$ be a positive
integer and let $N$ be relatively prime to all elements of $[1,k]$,
i.e., $\gcd(N,k!)=1$. Fix integers $c_1,c_2,c_3,c_4\in [-k,k]$ such
that $c_1+c_2+c_3+c_4=0$, and let $\mathcal{S}$ be the collection of sets $S
\subseteq \set{1,2,3,4}$ such that $\sum_{i \in S} c_i=0$ and $c_i\neq
0$ for $i \in S$. We note that $\mathcal{S}$ always contains the empty
set. Consider the following equation over $x_1,x_2,x_3,x_4\in \mathbb{Z}_N$:
\begin{equation}\label{eq:kfoldSidonset}
c_1x_1+c_2x_2+c_3x_3+c_4x_4\equiv 0 \pmod{N}.
\end{equation}
A solution of \eqref{eq:kfoldSidonset} is \emph{trivial} if there
exists a partition of the set $\set{i ; c_i\neq 0}$ into sets $S,T\in
\mathcal{S}$ such that $x_i=x_j$ for all $i,j\in S$ and all $i,j\in T$. We now
define a \emph{$k$-fold Sidon set} to be a set $A\subseteq\mathbb{Z}_N$ such
that for any $c_1,c_2,c_3,c_4\in [-k,k]$ with $c_1+c_2+c_3+c_4=0$,
equation \eqref{eq:kfoldSidonset} has only trivial solutions in
$A$. In the special case of $k=1$, a $1$-fold Sidon set coincides with
the usual definition of a \emph{Sidon set}, which is also a $B_2[N;1]$
set.
\begin{theorem}\label{thm:con-kfoldSidonset}
Let $A=\set{a_1,a_2,\dots,a_n}\subseteq\mathbb{Z}_n$ be a $k$-fold Sidon
set. Assume that $0\leq k_-\leq k_+\leq k$ and $k_++k_-\geq 1$. In the
group $G \triangleq \mathbb{Z}_{2(k_++k_-)+1} \times \mathbb{Z}_N$, construct a set
\[S\triangleq \set*{ (1,x) ; x \in A}.\]
Then $G \ge [-k_-,k_+]^* \diamond_2 S$.
\end{theorem}
\begin{IEEEproof}
Suppose to the contrary that $G$ is not partially $2$-split by
$S$. Then there are $x_1,x_2,x_3,x_4 \in A$ and $c_1,c_2,c_3,c_4 \in
[-k_-,k_+]$ such that
\[ c_1+c_2\equiv c_3+c_4 \pmod{2(k_++k_-)+1},\]
and
\begin{equation}\label{eq:con-kfoldSidonset}
c_1 x_1+c_2 x_2\equiv c_3 x_3 + c_4 x_4 \pmod{N},
\end{equation}
where all the following hold:
\begin{enumerate}
\item
$x_1\neq x_2$
\item
$x_3\neq x_4$
\item
$x_1\neq x_3$ if $c_1=c_3$ and $c_2=c_4=0$
\item
$x_2\neq x_4$ if $c_2=c_4$ and $c_1=c_3=0$
\item
$(x_1,x_2)\neq (x_3,x_4)$ if $(c_1,c_2)=(c_3,c_4)$.
\item
$(x_1,x_2)\neq (x_4,x_3)$ if $(c_1,c_2)=(c_4,c_3)$.
\end{enumerate}
Since $-(k_++k_-) \leq a+b,c+d \leq k_++k_-$, it follows that
$c_1+c_2=c_3+c_4$, or equivalently, $c_1+c_2-c_3-c_4=0$. To avoid
contradicting the assumption that $A$ is a $k$-fold Sidon set,
$(x_1,x_2,x_3,x_4)$ should be a trivial solution of
\eqref{eq:con-kfoldSidonset}. We consider the following cases:
\textbf{Case 1.} If none of $c_1,c_2,c_3,c_4$ are $0$, we consider
the possible partitions of $\set{1,2,3,4}$. Since $x_1\neq x_2$ and
$x_3\neq x_4$, $1$ and $2$, respectively $3$ and $4$, cannot be placed
in the same set in the partition. Then the possible partitions are
$\set{\set{1,3}, \set{2,4}}$, and $\set{\set{1,4}, \set{2,3}}$. If the
partition is $\set{\set{1,3}, \set{2,4}}$, then $c_1-c_3=0$ and
$c_2-c_4=0$. It follows that $x_1=x_3$ and $x_2=x_4$, which
contradicts that $(x_1,x_2)\neq (x_3,x_4)$ when
$(c_1,c_2)=(c_3,c_4)$. The case of $\set{\set{1,4},\set{2,3}}$ is
proved symmetrically.
\textbf{Case 2.} If there is exactly one element of $c_1,c_2,c_3,c_4$
that is equal to $0$, say w.l.o.g., $c_1=0$, then the only possible
partition of $\set{2,3,4}$ is $\set{ \emptyset, \set{2,3,4}}$, which
contradicts $x_3\neq x_4$.
\textbf{Case 3.} If there are exactly two elements of
$c_1,c_2,c_3,c_4$ that are equal to $0$, w.l.o.g., we may consider the
two cases where $c_1=c_2=0$, and $c_1=c_3=0$. If $c_1=c_2=0$, the only
possible partition of $\set{3,4}$ is $\set{ \emptyset, \set{3,4} }$,
which contradicts $x_3\neq x_4$. If $c_1=c_3=0$, the only possible
partition of $\set{2,4}$ is $\set{ \emptyset, \set{2,4} }$. Then we
have $x_2=x_4$. Note that $c_1=c_3=0$ and $c_2=c_4$, and we get a
contradiction.
\textbf{Case 4.} If there are exactly three elements of
$c_1,c_2,c_3,c_4$ that are equal to $0$, assume w.l.o.g., that
$c_2=c_3=c_4=0$ and $c_1\neq 0$. Then we need to partition
$\set{1}$. However, such a partition does not exist as $c_1\neq
0$. Thus, there is no solution to (\refeq{eq:con-kfoldSidonset}).
\end{IEEEproof}
When $k=2$, a family of $2$-fold Sidon sets is constructed in \cite{Lazebnik2003} by removing some elements from Singer difference sets with multiplier $2$.
\begin{theorem}[Theorem~2.5 in \cite{Lazebnik2003}]\label{thm:2foldSidonset}
Let $m$ be a positive integer and $N=2^{2^{m+1}}+2^{2^m}+1$. Then
there exists a $2$-fold Sidon set $A\subseteq \mathbb{Z}_N$ such that
\[\abs{A} \geq \frac{1}{2} N^{1/2}-3. \]
\end{theorem}
We immediately get the following corollary.
\begin{corollary}
\label{cor:t2sidon}
Let $0\leq k_-\leq k_+\leq 2$ be integers with $k_++k_-\geq
1$. There is an infinite family of integers $n$ such that $\mathbb{Z}^n$ can
be lattice packed by $\mathcal{B}(n,2,k_+,k_-)$ with density
\[\delta=\frac{\binom{n}{2}(k_++k_-)^2 + n(k_++k_-)+1 }{(2(k_++k_-)+1)(2n+6)^2}=\frac{1}{8(2(k_++k_-)+1)}+o(1).\]
\end{corollary}
\begin{IEEEproof}
Simply combine Theorem~\ref{thm:con-kfoldSidonset} and
Theorem~\ref{thm:2foldSidonset}.
\end{IEEEproof}
Now, we present a construction for $t=2$ and $0\leq k_-\leq k_+\leq
3$, which combines Behrend's method \cite{Behrend46} and Ruzsa's
method \cite{Ruzsa93} to forbid some specified linear equations.
\begin{theorem}\label{thm:con-t=2}
Let $0\leq k_-\leq k_+\leq 3$ be integers such that $k_++k_-\geq 1$. Set
$\alpha \triangleq \max\set{2k_+^2,3}$. Let $D \geq 2$ and $K\geq 1$ be integers,
and $p\equiv\pm 5\pmod{12}$ be a prime such that $(\alpha K+1)^D
\leq p$. For each $0\leq m < DK^2$, define
\[C_m\triangleq \set*{ x= \sum_{i=0}^{D-1} x_i (\alpha K+1)^i ; 0\leq x_i\leq K, \sum_{i=0}^{D-1}x_i^2=m }. \]
Let $G \triangleq \mathbb{Z}_{3k_++2k_-+1} \times \mathbb{Z}_p\times \mathbb{Z}_p$, and construct a subset
\[ S_m\triangleq \set*{ s_x ; x \in C_m }, \textup{ where } s_x \triangleq (1,x,x^2) \in G.\]
If $k_+\leq 3$, then $G\geq M\diamond_2 S_m$ for every $0\leq m < DK^2$,
and where $M\triangleq [-k_-,k_+]^*$.
\end{theorem}
\begin{IEEEproof}
Suppose to the contrary that
$G$ is not partially $2$-split by $S_m$. We consider the following
cases.
\textbf{Case 1.} $a s_x ={\mathbf{0}}$ for some $a \in M$ and $x\in C_m$. The
first coordinate of this equation is $a\equiv 0
\pmod{3k_++2k_-+1}$. Since $-k_-\leq a \leq k_+$, necessarily $a=0$, a
contradiction.
\textbf{Case 2.} $a s_x =b s_y$ for some $a,b \in M$, $x,y \in C_m$
and $(a,x)\neq (b,y)$. Similarly to Case 1, the first coordinate implies
that $a=b$. From the second coordinate, we have $a x \equiv b y
\pmod{p}$, and so, $x\equiv y \pmod{p}$. It follows that $x =y$ as
$0\leq x,y <p$, which contradicts $(a,x)\neq (b,y)$.
\textbf{Case 3.} $a s_x + b s_y = {\mathbf{0}}$ for some $a,b \in M$, $x,y
\in C_m$ and $x\neq y$. The first coordinate implies $a+b=0$ as $-2k_-
\leq a+b\leq 2k_+$. W.l.o.g., we assume $a>0$. Then $a s_x= (-b) s_y$,
where $0< a, -b \leq k_-$, which was ruled out in Case 2.
\textbf{Case 4.} $a s_x + b s_y = c s_u$ for some $a,b,c \in M$ and
$x,y,u\in C_m$ with $x\not =y$. From the first coordinate, we have
$a+b=c$ as $-2k_- \leq a+b\leq 2k_+$ and $-k_-\leq c \leq k_+$. If
$x=u$ or $y=u$, then $b s_y=bs_u$ or $a s_x=as_u$, respectively, both
of which were ruled out in Case 2. Thus, in the following, we assume
$x,y,u$ are pairwise distinct. Furthermore, using the condition
$a+b=c$ and rearranging the terms, we may assume that $a,b,c>0$.
From the second coordinate, we have that $a x + by \equiv c u \pmod{p}$, or equivalently,
\[ \sum_{i=0}^{D-1} (ax_i+b y_i) (\alpha K+1)^i \equiv \sum_{i=0}^{D-1} c u_i (\alpha K+1)^i \pmod{p}.\]
Note that $0\leq ax_i+b y_i, cu_i\leq 2k_+ K <\alpha K+1$ and $p \geq
(\alpha K+1)^D$. It follows that $ax_i+b y_i=cu_i$ for all $0\leq
i\leq D-1$. Thus, the three distinct points $(x_0,x_1,\ldots,
x_{D-1})$, $(y_0,y_1,\ldots, y_{D-1})$, and $(u_0,u_1,\ldots,
u_{D-1})$, are collinear in $\mathbb{Z}^D$ where $D\geq 2$, which contradicts
the fact that they are on the same sphere, i.e., $\sum_i x_i^2=\sum_i
y_i^2=\sum_i u_i^2=m$.
\textbf{Case 5.} $a s_x + b s_y = c s_u +d s_v$ for some $a,b,c,d \in
M$, $x,y,u,v\in C_m$, $x\neq y$ and $u\neq v$, where $abcd$ is
negative. By rearranging the terms, we may assume w.l.o.g. that
\[a s_x + b s_y + c s_z = d s_u\] for some $0<a,b,c,d \leq k_+$ and $x,y,z,u\in C_m$ where $x,y,z,u$ are not all the same.
Note that $0<a+b+c\leq 3k_+$ and $0< d \leq k_+$. From the first
coordinate of the equation above we have $a+b+c=d$. The second
coordinate of the equation implies that
\[ \sum_{i=0}^{D-1} (ax_i+b y_i+cz_i) (\alpha K+1)^i \equiv \sum_{i=0}^{D-1} d u_i (\alpha K+1)^i \pmod{p}.\]
Since $0\leq ax_i+b y_i+ cz_i\leq 3 k_+ K < \alpha K+1$ and $0\leq
du_i\leq k_+ K <\alpha K+1$, necessarily $ax_i+by_i+cz_i=d u_i$ for
all $0\leq i \leq D-1$. Then
\begin{align*}
ax_i^2+by_i^2+cz_i^2 & = a (x_i-u_i+u_i)^2 + b (y_i-u_i+u_i)^2 +c (z_i-u_i+u_i)^2 \\
&=a (x_i-u_i)^2 +2a(x_i-u_i)u_i+au_i^2 +b (y_i-u_i)^2 +2b(y_i-u_i)u_i+bu_i^2 \\
&\ \ \ +c (z_i-u_i)^2 +2c(z_i-u_i)u_i+cu_i^2 \\
& = a (x_i-u_i)^2 +b (y_i-u_i)^2 +c (z_i-u_i)^2+2(ax_i+by_i+cz_i)u_i \\
& \ \ \ -(a+b+c) u_i^2 \\
& = a (x_i-u_i)^2 +b (y_i-u_i)^2 +c (z_i-u_i)^2 +(a+b+c) u_i^2.
\end{align*}
Note that $x,y,z,u\in C_m$, i.e., $\sum_{i=0}^{D-1}x_i^2=\sum_{i=0}^{D-1}y_i^2=\sum_{i=0}^{D-1}z_i^2=\sum_{i=0}^{D-1}u_i^2$. It follows that
\begin{align*}
(a+b+c) \sum_{i=0}^{D-1}u_i^2 &= a \sum_{i=0}^{D-1}x_i^2+b \sum_{i=0}^{D-1}y_i^2+c \sum_{i=0}^{D-1}z_i^2 \\
& = a \sum_{i=0}^{D-1}(x_i-u_i)^2 +b \sum_{i=0}^{D-1} (y_i-u_i)^2 +c \sum_{i=0}^{D-1}(z_i-u_i)^2 +(a+b+c) \sum_{i=0}^{D-1}u_i^2,
\end{align*}
which in turn implies that $x_i=y_i=z_i=u_i$ for all $0\leq i\leq D-1$, and so, $x=y=z=u$, a contradiction.
\textbf{Case 6.} $a s_x + b s_y = c s_u +d s_v$ for some $a,b,c,d \in M$, $x,y,u,v\in C_m$, $x\neq y$ and $u \neq v$, where $abcd$ is positive. Note that from the first coordinate, we have $a+b=c+d$. By rearranging the terms, we may assume that
\[a s_x + b s_y = c s_u + d s_v\] for some $0<a,b,c,d \leq k_+$, $a+b=c+d$, $x,y,u,v\in C_m$ and $x,y,u,v$ are not all the same.
The second coordinate and the third coordinate of the equation above imply that
\begin{equation}\label{eq:1order}
ax+by\equiv cu+dv \pmod{p}
\end{equation}
and
\begin{equation}\label{eq:2order}
ax^2+by^2\equiv cu^2+dv^2 \pmod{p}.
\end{equation}
We multiply \eqref{eq:2order} by $a+b$, and then subtract the square of \eqref{eq:1order}. Noting that $a+b=c+d$, the result is
\begin{equation}\label{eq:dif}
ab(x-y)^2\equiv cd (u-v)^2 \pmod{p}.
\end{equation}
If $x= y$, using \eqref{eq:1order} and \eqref{eq:dif}, it is easy to
see that $x,y,u,v$ are all the same, a contradiction; if $x=u$, then
\eqref{eq:1order} was ruled out in Case 4. Thus, we may assume that
$x,y,u,v$ are pairwise distinct, and so,
\begin{equation}\label{eq:square}
abcd\equiv c^2d^2(u-v)^2/ (x-y)^2\pmod{p},
\end{equation}
i.e., $abcd$ should be a quadratic residue modulo $p$.
Check all the possible $abcd$, where $0<a,b,c,d\leq k_+\leq 3$ and $a+b=c+d$. We have $abcd\in\set{1,2^2,3^2,4^2,6^2,9^2,3\cdot 2^2}$. Since $p\equiv \pm 5\pmod{12}$, $3$ is not a quadratic residue modulo $p$, and so, $abcd\in\set{1,2^2,3^2,4^2,6^2,9^2}$. In all of these cases, $abcd$ is a square in $\mathbb{Z}$. Denote $t=\sqrt{abcd}$. Since $0<a,b,c,d\leq k_+$, we have $0<t\leq k_+^2$. Substituting $abcd=t^2$ in \eqref{eq:square} yields
\begin{equation}\label{eq:sqrt}
\pm t(x-y)\equiv cd(u-v) \pmod{p}.
\end{equation}
Solving the system of equations \eqref{eq:1order} and \eqref{eq:sqrt}, we get
\begin{equation*}
(c^2+cd)u\equiv (\pm t +ac)x+(bc\mp t)y \pmod{p}.
\end{equation*}
Note that $a+b=c+d$. Hence,
\begin{equation}\label{eq:sol}
\sum_{i=0}^{D-1} (ac+bc)u_i (\alpha K+1)^i \equiv \sum_{i=0}^{D-1}( (\pm t +ac)x_i+(bc\mp t)y_i )(\alpha K+1)^i \pmod{p}.
\end{equation}
Since $(\pm t +ac)+(bc\mp t)=ac+bc>0$, at least one of $\pm t +ac$ and $bc\mp t$ is positive.
We proceed in the following subcases.
\begin{enumerate}
\item If $\pm t +ac=0$, we have $t=ac$, and so,
\[\sum_{i=0}^{D-1} (ac+bc)u_i (\alpha K+1)^i \equiv \sum_{i=0}^{D-1} (bc+ac)y_i (\alpha K+1)^i \pmod{p},\]
which in turn implies $u=y$, a contradiction.
Similarly, if $bc\mp t=0$, we can get $u=x$, again, a contradiction.
\item If both $\pm t +ac$ and $bc\mp t$ are positive, then $0\leq (\pm t +ac)x_i+(bc\mp t)y_i \leq (ac+bc) K \leq 2k_+^2 K\leq \alpha K$. On the other hand, $0\leq (ac+bc)u_i \leq 2k_+^2 K \leq \alpha K$. Thus it follows from \eqref{eq:sol} that
\[(ac+bc)u_i= (\pm t +ac)x_i+(bc\mp t)y_i \textup{ for all } 0\leq i \leq D-1.\] That is, the three distinct points $(x_0,x_1,\ldots, x_{D-1}), (y_0,y_1,\ldots, y_{D-1})$ and $(u_0,u_1,\ldots, u_{D-1})$ of $\mathbb{F}_p^D$ are collinear, which contradicts the fact that they are on the same sphere.
\item If $\pm t +ac$ is negative, then $bc\mp t$ is positive. Rearranging the terms in \eqref{eq:sol}, we have that
\[\sum_{i=0}^{D-1} ((ac+bc)u_i-(\pm t +ac)x_i) (\alpha K+1)^i \equiv \sum_{i=0}^{D-1}(bc\mp t)y_i (\alpha K+1)^i \pmod{p}.\]
Since
\begin{align*}
0 & \leq (ac+bc)u_i-(\pm t +ac)x_i = (ac+bc)u_i+(\mp t -ac)x_i \\
& \leq (ac+bc)K+(\mp t -ac)K = (bc \mp t) K \leq 2k_+^2 K \leq \alpha K,
\end{align*}
then
\[(ac+bc)u_i- (\pm t +ac)x_i =(bc\mp t)y_i \textup{ for all } 0\leq i \leq D-1.\]
Again we get three distinct points on the same sphere which are collinear, a contradiction.
\item If $bc\mp t$ is negative, then $\pm t +ac$ is positive. Using the same argument as above, we can get the contradiction.
\end{enumerate}
Thus we complete our proof.
\end{IEEEproof}
\begin{remark}
In the proof above, the product $abcd$ is required to be either a square of $\mathbb{Z}$ or a non quadratic residue modulo $p$. This requirement comes from Ruzsa's method, in the proof of Theorem of 7.3 of \cite{Ruzsa93}. However, for $k_+\geq 4$, this requirement cannot be satisfied: we may choose $(a,b,c,d)$ to be $(1,4,2,3), (1,3,2,2)$ or $(2,4,3,3)$, the products $24,12$ and $72$ are not squares and they cannot simultaneously be non quadratic residues modulo $p$ for any prime $p$ as, using the Legendre symbol,
\[ \parenv*{\frac{6}{p}}=\parenv*{\frac{2}{p}}\parenv*{\frac{3}{p}}.\]
\end{remark}
\begin{corollary}
\label{cor:behruz}
Let $t=2$ and $0\leq k_- \leq k_+ \leq 3$, $k_++k_-\geq 1$. There
is an infinite family of $n$ such that $\mathbb{Z}^n$ can be lattice packed
by ${\mathcal B}(n,t,\kp,\km)$ with density $\delta=\Omega(c^{-\sqrt{\ln n}})$ for some
real number $c>0$.
\end{corollary}
\begin{IEEEproof}
Let $p\equiv\pm 5\pmod{12}$ be a sufficiently large prime. Set
$D\triangleq \floor{\sqrt{\ln p}}$ and $K\triangleq
\floor{(e^D-1)/\alpha}$. Then we have $(\alpha K+1)^D \leq p$.
Consider the group $G$ and the splitting sets $S_m$, $m\in
[0,DK^2-1]$, in Theorem~\ref{thm:con-t=2}. According to the pigeonhole
principle, there exists one set $S_m$ of size $\geq
\frac{(K+1)^D}{DK^2}$. Then
\begin{align*}
n & \geq \frac{(K+1)^D}{DK^2} = \frac{\parenv*{\floor*{\frac{e^D-1}{\alpha}}+1}^D }{D\floor*{\frac{e^D-1}{\alpha}}^2} \geq \frac{\parenv*{\frac{e^D-\alpha}{\alpha}+1}^D }{D\parenv*{\frac{e^D-1}{\alpha}}^2}
\geq
\frac{\alpha^2 e^{D^2}}{\alpha^{D}D e^{2D} } \\
& \geq \frac{\alpha^2 e^{ \floor*{\sqrt{\ln p}}^2}}{\alpha^{\sqrt{\ln p}} \sqrt{\ln p} e^{2\sqrt{\ln p}} } \geq \frac{\alpha^2 e^{ \parenv*{\sqrt{\ln p}-1}^2}}{\alpha^{\sqrt{\ln p}} \sqrt{\ln p} e^{2\sqrt{\ln p}} } = \frac{e\alpha^2 p}{\alpha^{\sqrt{\ln p}} \sqrt{\ln p} e^{4\sqrt{\ln p}} }\\
& \geq \frac{p}{c_1 ^{ \sqrt{\ln p}} },
\end{align*}
for some real number $c_1>0$. Taking logarithm of both side, we have
$\ln n \geq \ln p -\sqrt{\ln p} \ln c_1$, or equivalently,
\[\ln p - \ln c_1 \sqrt{\ln p} - \ln n \leq 0.\]
Solving for $\sqrt{\ln p}$, we get
\[\sqrt{\ln p} \leq \frac{\ln c_1 +\sqrt{ (\ln c_1)^2+4\ln n } }{2},\]
and so
\[ \ln p \leq \frac{4\ln n +2(\ln c_1)^2 +2 \ln c_1 \sqrt{ (\ln c_1)^2+4\ln n } }{4}\leq \ln n +c_2 \sqrt{\ln n},\]
for some real number $c_2>0$. It follows that
\[ n\geq \frac{p}{e^{c_2\sqrt{\ln n}}},\]
and
\[\delta \geq \frac{ \binom{n}{2} (k_++k_-)^2+n(k_++k_-)+1 }{\abs{G}} = \Omega(c^{-\sqrt{\ln n}}),\]
for some real number $c>0$.
\end{IEEEproof}
\section{Generalized Packings}
\label{sec:genpack}
It is a common practice in coding theory to also consider list
decoding instead of unique decoding. In such scenarios, the channel
output is decoded to produce a list of up to $\lambda$ possible
distinct codewords, where the channel output is within the error balls
centered at each of these codewords. In this section, we therefore
generalize the concept of packing to work in conjunction with list
decoding. The trade-off we present here is that at the price of a
constant-sized list, $\lambda$, we can find lattice packings of
${\mathcal B}(n,t,\kp,\km)$ with density almost constant, $\Omega(n^{-\epsilon})$, for any
$\epsilon>0$. The proof method, however, is non-constructive, and
relies on the probabilistic method.
Given a shape $\mathcal{B} \subseteq \mathbb{Z}^n$ and a lattice $\Lambda \subseteq
\mathbb{Z}^n$, we say $\mathcal{B}$ \emph{$\lambda$-packs $\mathbb{Z}^n$ by $\Lambda$} if for
every element $\mathbf{z}\in\mathbb{Z}^n$, there are at most $\lambda$ distinct
elements $\mathbf{v}_i\in\Lambda$ such that $\mathbf{z}\in\mathbf{v}_i+\mathcal{B}$. Obviously, if
$\lambda=1$, this definition coincides with the packing defined in
Section~\ref{sec:prelim}.
Let $G$ be a finite Abelian group, $M\triangleq [-k_-,k_+]^*$, and
$S\subseteq G$. If each element of $G$ can be written in at most
$\lambda$ ways as a linear combination of $t$ elements of $S$ with
coefficients from $M \cup \set{0}$, then we say $G
\overset{\lambda}{\geq} M \diamond_t S$.
The following result is an analogue of Theorem~\ref{th:lattotile},
which relates lattice packings to Abelian groups. The proof is exactly
analogous, and we omit it.
\begin{theorem}
Let $G$ be a finite Abelian group and $M\triangleq
[-k_-,k_+]^*$. Suppose that there is a subset
$S=\set*{s_1,s_2,\ldots,s_n}\subseteq G$ such that $G
\overset{\lambda}{\geq} M \diamond_t S$. Define $\phi:\mathbb{Z}^n\to G$ as
$\phi(\mathbf{x})\triangleq\mathbf{x}\cdot(s_1,\dots,s_n)$ and let
$\Lambda\triangleq\ker\phi$ be a lattice. Then $\mathcal{B}(n,t,k_+,k_-)$
$\lambda$-lattice-packs $\mathbb{Z}^n$ by $\Lambda$.
\end{theorem}
We use the probabilistic approach detailed in~\cite{vu:2002}, and
follow some of the notation there. Let $x_1,x_2,\ldots, x_N$ be
independent $\set{0,1}$ random variables. Let
$Y=Y(x_1,x_2,\ldots,x_N)$ be a polynomial of $x_1,x_2,\ldots, x_N$.
$Y$ is \emph{normal} if its coefficients are between $0$ and $1$. A
polynomial $Y$ is \emph{simplified} if every monomial is a product of
different variables. Since we are dealing with $\set{0,1}$ random
variables, every $Y$ has a unique simplification. Given a set $A$,
let $\partial_A(Y)$ denote the partial derivative of $Y$ with respect
to $A$, and let $\partial_A^*$ be the polynomial obtained from the
partial derivative $\partial_A (Y)$ by subtracting its constant
coefficient. Define $\mathbb{E}_j^*(Y)\triangleq \max_{\abs{A}\geq j}
\mathbb{E}(\partial_A^* Y)$.
\begin{theorem}\cite[Corollary~4.9]{vu:2002}\label{thm:vucon}
For any positive constants $\alpha$ and $\beta$ and a positive
integer $d$, there is a positive constant $C=C(d,\alpha,\beta)$ such
that if $Y$ is a simplified normal polynomial of degree at most $d$
and $\mathbb{E}_0^*(Y)\leq N^{-\alpha}$, then $\Pr(Y\geq C) \leq
N^{-\beta}$.
\end{theorem}
We now use Theorem~\ref{thm:vucon} to show the existence of
generalized lattice packings with the desired parameters.
\begin{theorem}
\label{th:prob}
Let $0\leq k_-\leqk_+$ with $k_++k_-\geq 1$, and $t>0$, be
integers. Let $N$ be a sufficiently large integer such that
$\gcd(N,k_+!)=1$, and fix $G\triangleq\mathbb{Z}_N$. Then for any
$0<\epsilon<1/t$, there is a number $\lambda$ which only depends on
$t$ and $\epsilon$, and a subset $S=\set{s_1,s_2,\dots,
s_n}\subseteq G$ with $\frac{1}{2}N^{1/t-\epsilon}\leq n \leq
\frac{3}{2}N^{1/t-\epsilon}$, such that $G \overset{\lambda}{\geq} M
\diamond_t S$, where $M\triangleq[-k_-,k_+]^*$.
\end{theorem}
\begin{IEEEproof}
Set $\alpha=\epsilon t$, $\beta=2$, and $d=t$. Denote
\[p\triangleq N^{\frac{1}{t}-1-\epsilon}.\]
We construct $S$ randomly. For each $0\leq i < N$, let the event
$i\in S$ be independent with probability $p$. Let $x_i$ be the
indicator variable of the event $i \in S$. Then
$\abs{S}=\sum_{i=0}^{N-1}x_i$, and
\[\mathbb{E}(\abs{S})=Np=N^{\frac{1}{t}-\epsilon}.\]
Using Chernoff's inequality, one can show that
\begin{equation}\label{eq:sizeexpc}
\Pr\parenv*{\frac{1}{2}\mathbb{E}(\abs{S})\leq \abs{S}\leq \frac{3}{2}\mathbb{E}(\abs{S})}\geq 1- 2e^{-\mathbb{E}(\abs{S})/16}.
\end{equation}
For every $g\in G$ and $0 \leq i_1 < i_1 <\cdots <i_\ell < N$, denote
\[c(g;i_1,i_2,\ldots, i_\ell)\triangleq\abs{ \set*{(a_1,a_2,\ldots,a_\ell)\in M^\ell; g=a_1i_1+a_2i_2+\cdots a_\ell i_\ell} },\]
where addition and multiplication are in $G=\mathbb{Z}_N$. Consider the
following random variables (which are polynomials in the indicator random
variables $x_0,\dots,x_{N-1}$),
\[Y_g \triangleq \parenv*{ \sum_{0 \leq i_1 <\cdots <i_t < N } {c(g;i_1,i_2,\ldots,i_t)} x_{i_1}x_{i_2}\cdots x_{i_t}}/(k_++k_-)^t,\]
and
\[Z_g \triangleq \parenv*{\sum_{\substack{1\leq \ell \leq t-1 \\ 0 \leq i_1 <\cdots <i_\ell < N }} {c(g;i_1,i_2,\ldots,i_{\ell})} x_{i_1}x_{i_2}\cdots x_{i_\ell}}/(k_++k_-)^{t-1}.\]
Both of them are positive, and as polynomials, they are simplified,
and normal. To show that $G \overset{\lambda}{\geq} M \diamond_t S$, it
suffices to show that $(k_++k_-)^tY_g+(k_++k_-)^{t-1}Z_g \leq \lambda-1$
for every $g\in G$.
We first look at $Y_g$. Since $\gcd(N,k_+!)=1$, if we fix
$a_1,a_2,\ldots,a_t\in M^t$ and $i_1,i_2,\ldots, i_{t-1}$, then there is a
unique $i_t \in [0,N-1]$ such that
$a_1i_1+a_2i_2+\cdots+a_ti_t=g$. Hence,
\[\mathbb{E}(Y_g) \leq ((k_++k_-)^t N^{t-1} p^{t})/(k_++k_-)^t= N^{-\epsilon t} =N^{-\alpha}.\]
For the partial derivative $\partial_A (Y_g)$ with $A=\set{j_1,j_2,\ldots, j_k} \subseteq [0,N-1]$ and $k\leq t-1$,
\[ \partial_A (Y_g)= \sum_{c_1,c_2,\ldots, c_k \in M}\parenv*{ \sum_{0\leq i_1<i_2 <\cdots i_{t-k}<N} {c(g-c_1j_1-\cdots-c_kj_k;i_1,i_2,\ldots,i_{t-k})} x_{i_1}x_{i_2}\cdots x_{i_{t-k}}}/(k_++k_-)^t, \]
hence,
\begin{align*}
\mathbb{E}(\partial_A (Y_g)) & \leq (k_++k_-)^k((k_++k_-)^{t-k}N^{t-k-1}p^{t-k}/(k_++k_-)^t) \\
& = N^{-\epsilon t +k\epsilon -k/t} < N^{-\alpha}.
\end{align*}
Applying Theorem~\ref{thm:vucon}, there is a number $C$, depending on $t$ and $\epsilon$, such that
\begin{equation}\label{eq:Yupbound}
\Pr(Y_g \geq C) \leq N^{-2}.
\end{equation}
As for $Z_g$, a similar computation to the above shows that
\[\mathbb{E}(\partial_A^*(Z_g))= O(N^{t-1-k-1} p^{t-1-k}).\]
Since
\begin{align*}
N^{t-1-k-1} p^{t-1-k}= N^{-\epsilon t-(\frac{1}{t}-\epsilon)(k+1)} < N^{-\alpha},
\end{align*}
we have $\mathbb{E}(\partial_A^*(Z_g)) < N^{-\alpha}$ when $N$ is sufficiently large. Applying Theorem~\ref{thm:vucon} again, there is a $C'$ such that
\begin{equation}\label{eq:Zupbound}
\Pr(Z_g\geq C') \leq N^{-2}.
\end{equation}
Denote $\lambda = C(k_++k_-)^t +C'(k_++k_-)^{t-1}+1$. Then \eqref{eq:Yupbound} and \eqref{eq:Zupbound} imply, via a union bound, that
\[\Pr(\textup{there exists } g \in G \textup{ such that } (k_++k_-)^tY_g+(k_++k_-)^{t-1}Z_g > \lambda-1) \leq 2N^{-1}. \]
From this, together with \eqref{eq:sizeexpc}, we can see that the
random set $S$ satisfies the conditions with probability at least
$1-2N^{-1}-2e^{-\mathbb{E}(\abs{S})/16}$, which is positive for large enough
$N$. Thus, such a set exists.
\end{IEEEproof}
\begin{corollary}
\label{cor:genpack}
Let $0\leq k_-\leqk_+$ with $k_++k_-\geq 1$, and $t>0$, be
integers. Then for any real number $\epsilon>0$, there is an integer
$\lambda$ and infinitely many values of $n$ such that $\mathbb{Z}^n$ can be
$\lambda$-lattice-packed by ${\mathcal B}(n,t,\kp,\km)$ with density
$\delta=\Omega(n^{-\epsilon})$.
\end{corollary}
\begin{IEEEproof}
Fix $\epsilon'\triangleq\frac{\epsilon}{\epsilon t+ t^2}$, and observe
that $\epsilon'<1/t$. Use Theorem~\ref{th:prob} with $\epsilon'$,
noting that $N=\Theta(n^{\epsilon+t})$, to obtain a lattice
$\lambda$-packing of ${\mathcal B}(n,t,\kp,\km)$ with density
\[ \delta \geq \frac{\abs{{\mathcal B}(n,t,\kp,\km)}}{\abs{G}}=\frac{\sum_{i=0}^t \binom{n}{i}(k_++k_-)^i}{N}=\Omega(n^{-\epsilon}).\]
\end{IEEEproof}
As a final comment on the matter, we observe that a tedious
calculation shows that in the above corollary
$\lambda=O(\epsilon^{-t})$ -- a calculation which we omit.
\section{Constructions of Lattice Coverings}
\label{sec:covering}
We switch gears in this section, and focus on covering instead of
packing. We first argue that using known techniques from the theory of
covering codes in the Hamming metric, we can show the existence of
non-lattice coverings of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$. However, these have a high
density of $\Omega(n)$. We then provide a product construction to
obtain a lattice covering by ${\mathcal B}(n,t,\kp,\km)$ with density $O(1)$.
Fixing an integer $\ell\in\mathbb{N}$, we use the same argument as the one given
in \cite[Section~12.1]{CohHonLitLob97} to construct a covering
code $C\subseteq\mathbb{Z}_\ell^n$, of size
\[ \abs*{C}=\ceil*{\frac{n\ell^n \ln \ell}{\abs*{{\mathcal B}(n,t,\kp,\km)}}}.\]
We can then translate this covering of $\mathbb{Z}_\ell^n$ by ${\mathcal B}(n,t,\kp,\km)$ to a covering
of $\mathbb{Z}^n$ by using the same idea as Theorem~\ref{th:linearcode}, and
defining $C'\triangleq\set*{ \mathbf{x}\in\mathbb{Z}^n ; (\mathbf{x}\bmod \ell)\in C}$. However,
the density of the resulting covering is
\[ \delta= \frac{\abs*{C}\cdot\abs*{{\mathcal B}(n,t,\kp,\km)}}{\ell^n} = \Omega(n).\]
We therefore proceed to consider more efficient coverings using the
product construction whose details follow.
\begin{theorem}
\label{con-reccovering}
Suppose that there exist a finite Abelian group $G$ and a subset $S
\subseteq G$ such that $G \leq M \diamond_1 S$. Let $t>0$ be an
integer, and denote
\[S^{(t)}\triangleq\set*{(s,0,0,\dots, 0) ; s \in S} \cup \set*{(0,s,0,\dots,0) ; s\in S} \cup \cdots \cup\set*{(0,0,0,\dots,s) ; s \in S}.\]
Then $G^{t}\leq M \diamond_t S^{(t)}$.
\end{theorem}
\begin{IEEEproof}
For any element $g=(g_1,g_2,\ldots,g_t)\in G^{t}$, since $G\leq M
\diamond_1 S$, for each $1\leq i \leq t$, there are $s_i\in S$ and $c_i
\in M\cup \set{0}$ such that $g_i=c_i\cdot s_i$. Hence,
\[g=c_1(s_1,0,0,\ldots,0)+c_2(0,s_2,0,\ldots,0)+\cdots+c_t(0,0,0,\ldots,s_t).\]
That is, $g$ can be written as a linear combination of $t$ elements
of $S$ with coefficients from $M \cup \set{0}$, and so, $G^{t}\leq M
\diamond_t S^{(t)}$.
\end{IEEEproof}
We can now construct a lattice covering, using the previous theorem.
\begin{corollary}
\label{cor:cover}
Let $\overline{\psi}(x)$ be the largest prime not larger than $x$, and denote
$p\triangleq\overline{\psi}(k_++k_-+1)$. Let $0\leq k_-\leq k_+$, with $k_++k_-\geq
1$, and $t>0$, be integers. Define $M\triangleq [-k_-,k_+]^*$. Then for
any integer $m>0$, there exists $S\subseteq (\mathbb{Z}_{p^m})^t$,
$\abs{S}=t\cdot\frac{p^m-1}{p-1}$, such that $(\mathbb{Z}_{p^m})^t\leq M
\diamond_t S$, and thereby, a lattice covering of $\mathbb{Z}^n$, $n=\abs{S}$,
with density
\[\delta = \frac{ \sum_{i=0}^{t} \binom{n}{i} (k_++k_-)^i }{ ( n(p-1)/t+1 )^t }=\frac{(t(k_++k_-))^t}{t!(p-1)^t}+o(1).\]
\end{corollary}
\begin{IEEEproof}
According to \cite[Construction 1]{Sch12}, there is a subset
$A\subseteq \mathbb{Z}_{p^m}$ of size $\frac{p^m-1}{p-1}$ such that
$\mathbb{Z}_{p^m}=[-k_-,p-1-k_-]^*\diamond_1 A$. We may apply
Theorem~\ref{con-reccovering} to obtain a subset $S=A^{(t)}\subseteq
(\mathbb{Z}_{p^m})^t$ of size $t\cdot\frac{p^m-1}{p-1}$ such that
$(\mathbb{Z}_{p^m})^t\leq M\diamond_t S$. The calculation of the density of the
resulting lattice covering is straightforward.
\end{IEEEproof}
\section{Conclusion}
\label{sec:conclusion}
Motivated by coding for integer vectors with limited-magnitude errors,
we provided several constructions of packings of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$ for
various parameters. These are summarized in
Table~\ref{tab:summary}. While the parameter ranges of the
constructions sometimes overlap, and perhaps result in equal or
inferior asymptotic density, having more constructions allows for more
choices for fixed values of the parameters.
One main goal was to construct lattice packings, analogous to linear
codes, as these are generally easier to analyze, encode, and
decode. Thus, except for one case, all constructions we provide
lattices. The main tool in constructing these is the connection
between lattice packings of ${\mathcal B}(n,t,\kp,\km)$ and $t$-splittings of Abelian
groups. The other important goal was to have asymptotic packing
density that is non-vanishing. This is achieved in many of the cases.
We also discussed $\lambda$-packing, which allows for a small overlap
between the translates of ${\mathcal B}(n,t,\kp,\km)$ centered at the lattice points. This
is useful for list-decoding setting with a list size of $\lambda$. The
result we obtain is non-constructive, and it provides a trade-off
between the list size and the packing density. Finally, we also
addressed the problem of lattice-covering of $\mathbb{Z}^n$ by ${\mathcal B}(n,t,\kp,\km)$,
showing using the product construction, that there exist such
coverings with asymptotic constant density.
The results still leave numerous open questions, of which we mention
but a few:
\begin{enumerate}
\item
Constructions for packings of ${\mathcal B}(n,t,\kp,\km)$ with $t\geq 3$, $k_+\geq 2$,
and non-vanishing asymptotic density are still unknown.
\item
Whether asymptotic density of $1$ is attainable for all parameters
is still an open questions. Such lattice packings would be analogous
to asymptotically perfect codes.
\item
In the asymptotic regime of $t=\Theta(n)$, all of the constructions
in this paper produce packings with vanishing asymptotic rates. Such
families of packings are analogous to good codes in the Hamming
metric, and their existence and constructions would be most welcome.
\item
Efficient decoding algorithms are missing for most of the cases. In
the asymptotic regime of constant $t$,
$\abs{{\mathcal B}(n,t,\kp,\km)}=\Theta(n^t)$. Thus, we are looking for non-trivial
decoding algorithms, whose run-time is $o(n^t)$.
\item
We would also like to find constructive versions of the non-constructive
proofs for $\lambda$-packings, and covering lattices.
\end{enumerate}
\begin{table*}
\caption{A summary of the results}
\label{tab:summary}
{\renewcommand{\arraystretch}{1.2}
\begin{tabular}{cccccll}
\hline\hline
Type & $t$ & $k_+$ & $k_-$ & density & Location & Comment \\
\hline
Packing & any & 1 & 0 & $\frac{1}{t!}+o(1)$ & Corollary~\ref{cor:lc} & via BCH codes \\
Packing & any & any & any & $\Theta(n^{t-\ceil{2t(1-{1}/p)}})$ & Corollary~\ref{cor:lc} & via BCH codes, $p\geq k_++k_-+1$ prime\\
Packing & any & 1 & 0 & $\frac{1}{t!}+o(1)$ & Corollary~\ref{cor:btsets} & via $B_t[N;1]$ sets \\
Packing & any & 1 & 1 & $\frac{2^t}{t!(2t+1)}+o(1)$ & Corollary~\ref{cor:kpm1} & via $B_t[N;1]$ sets \\
Packing & 2 & 1 & 0 & $1-o(1)$ & Corollary~\ref{cor:perparata} & via Preparata codes, non-lattice packing \\
Packing & 2 & 1 & 1 & $\frac{1}{2}+o(1)$ & Corollary~\ref{cor:quasiperfectlc} & via quasi-perfect linear codes\\
Packing & 2 & 2 & 0 & $\frac{1}{2}+o(1)$ & Corollary~\ref{cor:quasiperfectlc} & via quasi-perfect linear codes\\
Packing & 2 & $\leq 2$ & $\leq 2$ & $\frac{1}{8(2(k_++k_-)+1)}+o(1)$ & Corollary~\ref{cor:t2sidon} & via $2$-fold Sidon sets \\
Packing & 2 & $\leq 3$ & $\leq 3$ & $\Omega(c^{-\sqrt{\ln n}})$ & Corollary~\ref{cor:behruz} & via Behrend's and Ruzsa's methods \\
$\lambda$-Packing & any & any & any & $\Omega(n^{-\epsilon})$ & Corollary~\ref{cor:genpack} & $\lambda=O(\epsilon^{-t})$ \\
Covering & any & any & any & $\frac{(t(k_++k_-))^t}{t!(p-1)^t}+o(1)$ & Corollary~\ref{cor:cover} & $p\leq k_++k_-+1$ prime \\
\hline\hline
\end{tabular}
}
\end{table*}
| eecb771e98f59bd2455abc547820b8dc55725a1b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Access to freshwater is of pivotal importance to humanities. Fast population growth and climate change demand increasing supply of freshwater. Water scarcity has become a threat to the sustainable development of human society \cite{mekonnen2016four,reif2015solar}. This motivates the development of utilizing saline waters from the oceans and other brackish water sources and the processes that convert saltwater into freshwater. Solar energy is now emerging as one of the most promising sustainable energy sources, as it is clean and can be supplied without any environmental pollution compared with other forms of energy \cite{kalogirou2004solar}. And the abundant solar energy makes the solar-driven evaporation one of the promising approaches for water desalination and purification.
In a conventional solar-driven evaporation system, bulk water is heated to a high temperature to generate water vapor, resulting in a slow response to sunlight and heat loss to the bulk water or the external environment \cite{kabeel2011review,neumann2012solar,tao2018solar,lenert2012optimization,hogan2014nanoparticles}. In 2011, Wang \textit{et al.,} were among the first to demonstrate the floating interfacial solar-driven evaporation structure into desalination area, which does not require significant capital investment in high-cost permanent construction and/or land use \cite{zeng2011solar}. Compared with bulk water-heating method which heats the entire body of water, the interfacial solar-driven evaporation approach mainly localizes the heat generation at the water-air interface. This method avoids heating a large volume of water, \textit{e.g.,} the ocean, which serves as a low-temperature sink. The interfacial evaporation has an ultra-fast response to sunlight and is endowed with a much higher systematic thermal efficiency \cite{wang2018emerging,wang2014bio}. Furthermore, the technique can be easily applied by floating an absorber sheet on the water surface without complex pressure control or other bulky infrastructure. Owing to all these advantages, scientists have made great efforts to develop the interfacial solar-driven evaporation.
To yield a remarkable photothermal performance, the development of interfacial evaporation structure should center around the following four key factors. (1) The ideal solar absorber should possess outstanding absorptance in the range of visible and near-infrared regions \cite{zhang2017vertically} to convert the solar radiation into heat to be used for water vapor generation; (2) It is crucial for the interfacial evaporation to feature a low thermal conductivity, thereby reducing the heat loss from the absorber to the bulk water and localizing the heat at the air/water evaporative interface; (3) High hydrophilicity and porous framework are necessary for sufficient water transport from the bulk water to the heated area; and (4) Self-floating ability is required to avoid land use and/or permanent infrastructure construction. Even with intensive efforts so far, it remains as a huge challenge to develop an approach that can satisfy these four criteria simultaneously.
To obtain a high efficiency, a multilayered structure is proposed to take into account these pivotal factors and it provides us with more options to choose alternative appropriate materials \cite{shi2017rational,lou2016bioinspired,ghasemi2014solar,ni2018salt,li2016graphene}. Ghasemi \textit{et al.,} first demonstrated the double-layered structure consisting of a carbon foam layer supporting an exfoliated graphite layer \cite{ghasemi2014solar}. The bottom layer serves as a heat barrier to minimize the heat loss to the bulk water, and its porous structure is used to transport water to the heated area, and meanwhile it supports the top layer to make it possible for self-floating. The top porous layer serves as the absorber for the vapor transport.
Recently, many similar double-layered evaporation structures have been widely adopted. In most cases, such as air-laid paper \cite{liu2015bioinspired}, polystyrene foam \cite{shi2017rational}, woodblock \cite{kim2018mesoporous,jiang2017water,chen2017highly}, polyurethane sponge \cite{zhang2019facile}, macroporous silica substrate \cite{wang2016self}, and bacterial nanocellulose aerogel \cite{jiang2016bilayered}, the bottom layer serves both as a heat barrier and a water transport medium. Under a steady working condition, when water is transported to the top photothermal area through the bottom layer, almost all of the macropores inside the heat barrier are filled with water which typically is more thermally conductive than the heat barrier. This process renders the heat barrier less effective. To avoid heat loss, a new multilayer consisting of photothermal material, closed-pore thermal insulator, and external hydrophilic materials has generated much interest. The closed-pore thermal insulator serves as a real heat barrier layer and the air inside the insulator reduces heat conduction. For example, polystyrene foam \cite{li2016graphene,ni2016steam,li2017three,ni2018salt,xu2017mushrooms,liu2019high} is a good and inexpensive thermal insulator with low thermal conductivity. External hydrophilic materials, wrapping or passing through the insulator foam, act as a capillary-driven pump with a one-/two-dimensional water path \cite{liu2017bioinspired,li2016graphene}. By and large, this modified multilayered evaporation structure has become a favorable way to further minimize heat conduction losses and improve energy conversion efficiency.
An efficient solar steam generation depends highly on the photothermal conversion materials with broadband absorptance and high photothermal properties \cite{zhu2018solar,gao2019solar}. On the basis of a multilayered structure, the diverse photothermal materials with wide light absorbance have been extensively studied including carbon-based materials \cite{zeng2011solar,ghasemi2014solar,hu2017tailoring,zhou2016macroscopic}, metals \cite{brongersma2015plasmon,richardson2009experimental,liu2017solar,zhou20163d}, and metal oxides \cite{zhu2016constructing,ye2017synthesis,wang2017high,ding2017non}. Carbon-based absorbers, such as carbon black \cite{liu2015floatable}, graphene \cite{yang2018graphene}, graphene oxide \cite{li2016graphene,li20173d,jiang2016bilayered} and carbon nanotube \cite{chen2017highly,wang2016self}, offer great advantages of high solar absorbance and thermal stability. They have become the largely approving choices, however, the utilization of elaborate technologies and the high-associated cost limit their field applications to large-scale desalination \cite{lin2018integrative,tao2018solar,li2016graphene}. Although the metal-based (\textit{e.g.,} gold nanoparticles) materials have attracted tremendous attention due to their unique optical and photothermal properties \cite{loeb2017solar,liu2015bioinspired,zhou2017self}, their industrial applications are limited because of their inferior chemical and thermal stabilities, exorbitant cost, and complex fabrication processes.
With the aim to further explore novel techniques and materials in improving evaporation efficiency and reducing cost for desalination, we have applied the blackest paint, called Black 3.0 Paint\textregistered, to the floating interfacial solar-driven evaporation device serving as the photothermal conversion material combined with a sheet of melamine foam (MF). Herein, a sheet of hot-pressed MF plays an ideal elastic skeleton material sprayed with Black 3.0 paint, serving on top as a perfect solar absorber to efficiently absorb and convert the solar radiation into heat. To achieve a perfect evaporation performance and high solar-thermal energy conversion efficiency, we demonstrate a self-floating three-layer evaporation structure with a two-dimensional water path to localize solar-thermal heat generation to the air/water interface. This entire device is made from the commercially available and low-cost materials: solar absorber uses a sheet of hot-pressed MF combined with Black 3.0 paint; two-dimensional water path is provided by the highly absorbent Webril all-cotton wipes, and thermal barrier uses the Polyvinyl chloride (PVC) foam. Under 1 sun illumination without solar concentration, this evaporation device yields an excellent evaporation rate of freshwater as high as 2.48 kg m$^{-2}$ h$^{-1}$ and a highlighted evaporation efficiency of 172.5\%. Based on this multilayered evaporation structure, the novel Black 3.0 paint demonstrates a superb performance with cost-effectiveness, operability, and durability, which can be regarded as one of the indispensable choices of the solar-absorbing materials in the future development of the solar-driven evaporation process.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\textwidth]{absorber}
\caption{ \label{fig:absorber} Preparation process and characterizations of the black-sprayed hot-pressed melamine foam (MF) based absorber layer. (a) Schematic illustration of the fab processes of an absorber layer in the evaporation device. (b) UV-Visible-Near-infrared absorptivity spectra of the absorber layer, the hot-pressed MF, and the normalized spectral solar irradiance density of air mass 1.5 (AM 1.5 G) global tilt solar spectrum. (c) The temperature of the absorber sheet over time under 1 sun. }
\end{figure}
\section{Results and Discussions}
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\textwidth]{test_absorber}
\caption{ \label{fig:test_absorber} Microscope images of (a) pristine MF, (b) pressed MF, (c) one-layer paint coated MF, and (d) three-layer paint coated MF (a-d, insets are the photographs of the pristine/hot-pressed, one-/three-layer paint coated MFs). (e) The UV-Visible-Near-infrared absorptivity spectra measurement and the chemical stability tests of the absorber layer under different harsh environments. The inset shows the absorber layer is immersed in an alkaline solution (pH \textasciitilde 14) and acidic solution (pH \textasciitilde 1) for 24 hours in the sealed bottles.}
\end{figure}
The schematic of the synthesis process of the absorber layer is shown in Fig. \ref{fig:absorber}a. The fabrication process is simple and scalable, which is summarized as three steps: hot-pressing, spraying, and drying. In this study, the MF with the dimension of 10 cm $\times$ 6 cm $\times$ 2 cm is selected as an elastic skeleton for immobilization of the Black 3.0 paint spray. MF is a commercially available low-cost polymer material, and its high porosity, low density, excellent hydrophilicity, and elasticity render it highly applicable for the solar steam generation devices \cite{lin2018integrative,li2019scalable}. Hot-pressing treatment of the MF with a compression ratio of 4 has been experimentally confirmed to be advantageous to increase its elasticity and fatigue resistance. More importantly, the compressed three-dimensional network of the foam can dramatically increase its density to effectively increase the ability of water absorption. In laboratory experiments, the pressed MF is cut into a 1 mm-thick circle with a diameter of 47 mm, fitting the size of a 100 ml beaker.
The inexpensive Black 3.0 paint has gained great attention in creative art and as an superb solar photothermal material. As the blackest and most matte acrylic paint, Black 3.0 paint exhibits nearly perfect solar absorption at up to 99\% of visible light. The paint can be easily applied via brushing or spraying onto most surfaces, such as wood, paper, canvas, and plastic. Last but not the least, the paint is light in density and very suitable to the desalination process.
Figure S2 shows the transmittance spectrum of the Black 3.0 paint hydraulic pressed with KBr powder showing corresponding wavenumber of functional groups. Then the diluted paint is sprayed on the pressed MF sheet using a spray gun. The distance between the spray head and the MF sheet is about 25 cm. Considering the porous surface of the pressed MF sheet surface, it turns out that a couple of thin coating layers are better than one the thick layer. Between each coating process, drying it thoroughly with hot air is necessary at 190$^\circ$C for 5 minutes. After thoroughly drying, the Black 3.0 paint-coated MF sheet is rinsed in deionized (DI) water for several times to remove the residual impurities. Hereafter, keep this sheet in the oven at 60$^\circ$C for drying and then use it as the solar absorber layer in the evaporation device. The absorber layer can be easily prepared requiring no complex process or expensive equipment, and therefore, most suitable for volume production. As shown in Fig. \ref{fig:absorber}b, the absorber sheet exhibits a superb absorption ranging from 95\% to 99\% at the wavelength from 0.2 $\mu$m to 2.5 $\mu$m, indicating that it acts as an efficient broadband solar radiation absorber, whereas the pure pressed MF shows poor absorptivity. Besides, this absorber also exhibits an angular-independent hemispherical absorptivity (Fig. S1). Figure \ref{fig:absorber}c shows the surface temperature of the absorber sheet in the air under 1 sun for an hour. The time-dependent temperature changes and images are captured by an infrared camera at room temperature, displaying the maximum temperature of the sheet over time. In taking the temperature measurement, the tested absorber sample is placed on a PVC foam plate to minimize the heat exchange with the base below. Upon light illumination, the surface temperature of the absorber sheet rises sharply to an equilibrium temperature around 100$^\circ$C within the initial 2.5 minutes and then floats slightly around this temperature afterward, indicating an excellent photothermal performance of the absorber sheet. The inset of Fig. \ref{fig:absorber}c demonstrates the change in surface temperature more clearly in the initial 4 minutes.
Figures \ref{fig:test_absorber}a - \ref{fig:test_absorber}d show the microscopic structures of the absorber. The pristine MF has polygon networks with a size of about 200 $\mu$m and the diameter of the fibers is around 7 $\mu$m (Fig. \ref{fig:test_absorber}a). While the geometric shape of the networks collapses with the reduced network size (Fig. \ref{fig:test_absorber}b), it retains its porous structure, enabling the transportation of water. One-layer Black 3.0 paint coating forms the beads along the fibers (Fig. \ref{fig:test_absorber}c), and the three-layer coating causes an increase in the fiber diameter from 7 $\mu$m to 50 $\mu$m (Fig. \ref{fig:test_absorber}d) and fully-covered Black 3.0 paint coated MF presents nearly unity absorptivity in solar irradiance region. The fully covered MF still keeps a highly-porous structure facilitating the water transport and vapor release, which results in the high water evaporation rate. To evaluate the stability of the absorber sheet, the absorber layers undergo severe tests, kept in the boiling water (around 90$^\circ$C) for 1 hours and immersed in the alkaline solution (pH \textasciitilde 14) and acidic solution (pH \textasciitilde 1) for 24 hours in the closed bottles. After this harsh thermal and chemical stability test, no obvious changes in the appearance are observed and their absorption spectra are basically consistent with the original one (Fig. \ref{fig:test_absorber}e).
Localizing the heat converted from solar energy to the air/water interface is the critical step to achieve high evaporation performance.
Figure \ref{fig:device_wet_salt}a demonstrates a self-floating evaporation device structure which is composed of three layers. The top layer is a solar absorber layer made of the MF sheet combined with Black 3.0 paint and it plays an essential role in an efficient absorption of the solar radiation. Beneath the absorber layer is a two-dimensional water path enabled by a cotton wipe that is cut into the shape of a circle with 4 extended strips. The round area of the cotton wipe is kept the same as the absorber area, and the bottom edges of 4 strips are in direct contact with the bulk water, which decreased the contact area between the evaporation surface and the bulk water. The cotton wipe is used to transport water due to its strong capillary force and to reduce the direct contact between the bulk water and the absorber layer. Simultaneously ensuring the efficient water supply to the solar heating area and minimizing the heat dissipation to the bulk water is crucial to achieve the high-efficiency interfacial evaporation. The interlocked fibers of cotton wipe do not come apart when soaked in water for a long time, that contributes to the durability and stability of the device. In the working process, the cotton wipe pumps the bulk water towards the heating surface through the 4 strips around the floating foam relying on its strong capillary wicking effect. The bottom layer is the thermal insulation wrapped with the cotton water path. The thermal insulation is made from PVC foam with only closed pores which are impermeable to water, and its low thermal conductivity (\textasciitilde 0.03 W m$^{-1}$ K$^{-1}$) almost reduces the downward thermal dissipation from the evaporation surface. Thanks to the low density of the thermal insulation and the simple structure of the whole device, the evaporation device can be placed on the water and move together with the waving water surface achieving self-floating for continuous operation.
The surface wettability of the evaporation device also plays an important role in the steam generation process. As shown in Fig. \ref{fig:device_wet_salt}b, the evaporation device is placed in a beaker filled with DI water at the initial temperature of 21$^\circ$C, and its wetting process is recorded from the top view. It is obvious that the water immediately reaches the surface of the absorber layer from the regional edge (the enclosed areas in the blue dashed), and then the wet area quickly expands to the central area until it covers the entire evaporation surface. The surface wettability owes much to forceful hydrophilicity of the two-dimensional water path of the cotton and the porous structure of the absorber layer, which assures ample water supply to the evaporation surface.
Avoiding salt accumulating in the evaporation device remains a significant character for a self-floating solar evaporation structure utilizing heat localization \cite{ni2018salt}. Figure \ref{fig:device_wet_salt}c shows the progression of salt dissolution which demonstrates the salt rejection ability of the three-layer device. In this experiment, a three-layer structure with a diameter of 47 mm floated in a 100 ml beaker of 132g 3.5 wt\% NaCl solution with 21$^\circ$C initial temperature, and 1.7g of additional solid NaCl is placed directly on the top of the absorber surface. Upon contact with water, the solid NaCl on top starts to dissolve due to the movement and exchange of solution inside the absorber layer and two-dimensional water path between the device surface and the bulk water below the insulation. After approximately 65 minutes, the three-layer device fully rejects the salt, that indicates its good salt rejection ability.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\textwidth]{device_wet_salt}
\caption{ \label{fig:device_wet_salt} The self-floating evaporation device with surface wettability and salt rejection ability tests. (a) Photographs of the three-layer evaporation device placed in a beaker filled with freshwater. (b) Wetting process of the evaporation device placed on the surface of the water. The blue-dash enclosed regions exhibit the wetting areas. (c) The salt rejection progress of the evaporation device. The evaporation device is placed in a beaker filled with 3.5 wt$\%$ NaCl solution, and NaCl is originally stacked on the top surface of the absorber.}
\end{figure}
The steam generation ability of the evaporation device at a laboratory scale is tested combining an irradiation system to simulate the solar radiation, a weighing system monitoring the mass change of water in a beaker during a certain period of irradiation, and an infrared camera monitoring the real-time temperature. Both 3.5 wt\% NaCl solution and DI water are prepared to keep the initial temperature at 21$^\circ$C in the 100 ml beakers. The evaporation device floated on water in the beaker filled with 127g of water, beneath the beaker is an electronic balance connected to the computer, and the simulated solar radiation was provided by a solar simulator. When the illumination began, the solar steam generation is determined by the balance, and the real-time temperature of the absorber sheet and water surface are monitored by an infrared camera. The laboratory temperature and humidity are kept at about 22$^\circ$C and 21\%, respectively.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\textwidth]{lab_experiment}
\caption{ \label{fig:lab_experiment} (a) Left: Infrared (IR) thermal images of the top surface of a water-only beaker under 1 sun illumination over exposure time. Middle: IR thermal image and the photograph of the beaker filled with DI water only under 1 sun after one hour. Right: The photograph of the beaker filled with water only placed on an electrical balance under 1 sun illumination. (b) Left: IR thermal images of the evaporation device under 1 sun illumination in the initial 5 minutes. Middle: IR thermal image of the evaporation device floated on water in the beaker under 1 sun illumination after one hour. Right: The photograph of the beaker with the evaporation device placed on an electrical balance under 1 sun illumination. (c) The change of mass of water over time of DI water only and with the evaporation device placed at the interfaces of the DI water, and 3.5 wt\% NaCl solution. (d) The maximum temperature profiles of the pure water surface and the absorber layer under 1 sun illumination over exposure time. The inset is a visible steam flow generated under 1 sun irradiation.}
\end{figure}
Figure \ref{fig:lab_experiment} compares the surface temperatures of pure water and 3.5 wt\% NaCl solution with and without the evaporation device under 1 sun illumination. The temperature data and images are captured by an infrared camera. Figure \ref{fig:lab_experiment}a shows the surface temperature distribution of the beaker filled with water without the evaporation device over the irradiation time. The surface temperature rises slowly under irradiation owing to the poor light harvesting efficiency and its maximum temperature profile shown in Fig. \ref{fig:lab_experiment}d slopes gently from the initial temperature of 21$^\circ$C to 32$^\circ$C for an hour. In contrast, upon light illumination, the temperature of the absorber layer rises sharply to an equilibrium temperature around 50$^\circ$C, indicating good photothermal performance. As shown in Fig. \ref{fig:lab_experiment}c, for the water only experiment, the steam generation rate is measured to be 0.41 kg m$^{-2}$ h$^{-1}$ under 1 sun irradiation (1 kW m$^{-2}$). In contrast, the steam generation rate of evaporation device in DI water reaches up to 2.48 kg m$^{-2}$ h$^{-1}$ under the same experimental conditions, 6.04 times higher than that of pure water. Most significantly, when the evaporation device floats on 3.5 wt\% NaCl solution, it sports a comparable steam generation rate of 2.32 kg m$^{-2}$ h$^{-1}$, 5.65 times higher than that of pure water. Evaporation efficiency, $\eta_{evap}$ = $\Dot{m}h_{fg}$/$Q_{s}$, is an important parameter to evaluate the steam generation performance, where $\Dot{m}$ is the water evaporation rate (kg m$^{-2}$ h$^{-1}$), $Q_{s}$ is the power density of irradiation (kW m$^{-2}$), and $h_{fg}$ is the total enthalpy of vaporization of water (kJ kg$^{-1}$). In particular, $h_{fg}$ is regarded as the sum of the sensible heat (121.26 kJ kg$^{-1}$) and the temperature-dependent enthalpy of vaporization (2382.7 kJ kg$^{-1}$) in this text. Thus, we obtain $\eta_{evap}$ = 172.5\% under 1 sun illumination. More detailed calculation of the evaporation efficiency and the photothermal efficiency are provided in the supplementary materials.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth]{outdoor_1}
\caption{ \label{fig:outdoor_1} Rooftop experiments with the evaporation system under natural sunlight in January 2020 in Boston, MA, USA. (a) Three-dimensional model of the evaporation system with a PMMA condensation cover in the dimension of 30 cm $\times$ 30 cm for desalination. (b) Schematic illustration of the solar desalination system. (c) Photograph of the evaporation system placed on a utility cart (left) and the dense condensed water droplets on the interior surface of the condensation cover and sidewall (right) during the experiment. (d) The contrast experiment with an evaporation device (left inset) and without (right inset) on the Day 5. The blue and red curves show the change of ambient temperature (left y-axis) and the solar intensity (right y-axis). Two bottles inset show the weight of the collected water from the contrast experiment. (e) The averaged outdoor steam generation rates of the 3.5 wt\% NaCl solutions for 5 days in January 2020. }
\end{figure}
\section{Field Test}
We have performed field test of this evaporation device to demonstrate the feasibility of scalable production of freshwater. We constructed a prototype of the evaporation system with a condensation chamber and an evaporation device of a large area of the absorber (24 cm $\times$ 24 cm ) inside, as shown in Figs. \ref{fig:outdoor_1}a and S3. The condensation chamber made of Polymethyl methacrylate (PMMA) board surrounds the entire evaporation device to capture condensate of the evaporated steam. According to the transmission spectrum of the PMMA board, shown in Fig. S4, the condensation structure is well transparent in the solar spectrum to allow the solar irradiance to reach the evaporation device, while it is opaque in the mid-infrared region to confine the infrared emission of the evaporation device inside the chamber to contribute to maintaining a relatively higher evaporation temperature. As shown in Fig. \ref{fig:outdoor_1}b, the tilted angle of the condensation cover in the dimension of 30 cm $\times$ 30 cm is fixed to be 32$^\circ$ considering the latitude of Boston city, allowing the solar flux to reach the evaporator without refraction. The photograph of Fig. \ref{fig:outdoor_1}c shows the prototype is placed on the polystyrene foam with a low thermal conductivity to reduce the thermal flux between the bottom of the prototype and the utility cart, and the condensed water is observed on the condensation cover (Fig. S5). The condensed water droplets fall to the collection tank and eventually flow to a chemical storage bottle. It has been confirmed that the sodium concentration of the desalinated water is 5 mg l$^{-1}$, that is much lower than the standard of the drinkable water specified by the World Health Organization (WHO) \cite{world2011safe}. The validated prototype is tested on the roof of Snell Engineering Center at Northeastern University, Boston, MA, USA, and water collections are measured over 5 days in January 2020. The specific experiment dates, instantaneous incident sunlight, ambient temperature, wind speed, and humidity are recorded and provided in the supplementary materials. The 3.5 wt\% NaCl solution, which simulates the average salinity of seawater all over the world, is used during the field test and the seawater holder is filled up through the water injection tube after the daily use.
The average evaporation rate of drinkable freshwater, shown in Fig. \ref{fig:outdoor_1}e, varies from 0.32 kg m$^{-2}$ h$^{-1}$ to 0.47 kg m$^{-2}$ h$^{-1}$ under various solar intensity and ambient temperature. The experiment data and detailed weather condition data are provided in Figs. S6 and S7. To demonstrate the promising solar steam generation ability of this evaporation system, the prototype, and the control group, the same condensation chamber without a three-layer evaporation device inside is conducted for a contrast experiment on day 5 (Fig. \ref{fig:outdoor_1}d). At the end of the experiment, 164g of water is collected from the prototype, while the control group collects nearly zero water in the same-day operation. In the rooftop experiments, it is well noted that the weight of collected freshwater refers to the weight of water in the chemical storage bottle, not exactly the weight of condensed water produced. The inset of Fig. \ref{fig:outdoor_1}d shows the end state of the contrast experiment. Obviously in the control group, dense condensed water droplets can be observed on the cover without forming the water dripping flow (Fig. S8). In day 4, the daily condensate rate (2.83 kg m$^{-2}$ d$^{-1}$) in winter produced from the evaporation device of the area of 0.0576 m$^{-2}$ is comparable to a previous work (2.81 kg m$^{-2}$ d$^{-1}$) in the summer \cite{ni2018salt}, which proves a reliable and scalable production of freshwater using our design in four seasons.
\section{Conclusions}
We have demonstrated a high-performance, low cost, and simple interfacial three-layer steam generation device based on the novel commercial Black 3.0 paint, which is first applied to the solar steam generation as a promising candidate for the photothermal materials. Black 3.0 paint sprayed on a sheet of MF serves as the top solar absorber layer that can widely absorb the solar radiation and efficiently convert it into heat. The absorber layer which is placed on a PVC foam plane under 1 sun irradiation shows a significant temperature gradient and reaches up to 100$^\circ$C of equilibrium temperature. The absorber is shown to sport strong thermal and chemical stabilities under harsh environment, and it exhibits a remarkable salt rejection ability. All these performance contribute to the long-term durable steam generation process. In the laboratory experiments, enabled by the assistance of the low thermal conductivity PVC foam to reduce the heat loss and the cotton wipe to provide sufficient water through two-dimensional water path to the heating region, the evaporation device has reached a striking steam generation rate of 2.48 kg m$^{-2}$ h$^{-1}$, 6 times higher than the natural evaporation, and a highlighted evaporation efficiency of 172.5\%\ under 1 sun irradiation at the room temperature, surpassing most of the reported works. Even in the rooftop experiments during a cloudy winter day in Boston, MA with the maximum environment temperature of 4$^\circ$C, 2.83 kg m$^{-2}$ d$^{-1}$ of water can be collected. This simple, low cost, and easy to manufacture evaporation device is highly beneficial to large scale water desalination and purification applications.
\section{Experimental section}
\subsection{Materials}
The commercial MF is purchased from South Street Designs company (UPC: 089902974060) with the dimension of 10 cm $\times$ 6 cm $\times$ 2 cm (\$0.3/piece). Black 3.0 paint is purchased on the Culture Hustle (\$0.146/ml). Webril pure cotton wipe is in the size of 0.2 m $\times$ 0.2 m (\$2.99/m$^{2}$). The PVC foam insulator sheet is purchased from the McMaster-Carr with the dimension of 813 mm $\times$ 1219 mm $\times$ 13 mm (\$58.7/m$^{2}$).
\subsection{Sample preparation}
Solar absorber layer is fabricated as follows: pristine MF is thoroughly washed several times with ethanol and deionized (DI) water and then put in an oven kept at 60$^\circ$C in preparation for the hot-pressing treatment. After it is totally dried, the MF is pressed at 200$^\circ$C for 6 minutes with a compression ratio of 4, which is the height ratio of pristine MF to hot-pressed one. Serving as the skeleton of the absorber, hot-pressed MF is cut into the desired shape with a thickness of about 1 mm. Black 3.0 paint is thinned with DI water under vigorous stirring for 5 minutes with a paintbrush, which helps to get a homogeneous mixture. The mass ratio of DI water to Black 3.0 paint is kept in the range of 0.35 $\sim$ 0.4 in the dilution process. Subsequently, the diluted Black 3.0 paint is sprayed onto the MF sheet by a touch-up spray gun (Paasche Airbrush, USA) with a 0.8 mm spray head at the pressure of 70 psi. The distance between the spray head and the MF sheet is about 25 cm. The golden rule is that 3 or 4 thin layers is much better than a single thick layer. And then dry it between each spraying with the hot air blower (Yihua Electronic Equipment Co., Ltd, Guangzhou, China) at a temperature of 190$^\circ$C for 5 minutes. It's worth noting that this drying time is suitable for the specific dimension of the sample in this work.
A piece of PVC foam (47 mm in diameter and 13 mm in thickness) is utilized as the thermal insulator. Webril pure cotton wipe is cut into a 47 mm circle with four extended strips of 30 mm in width and 20 mm in length. The hydrophilic cotton wipe wrapped around the PVC foam with the four strips tip soaking in the bulk water ensures that the water reaches the upper circular area due to capillary force. Then, the MF sheet is placed over the circular area of the cotton wipe.
\subsection{Solar steam generation experiments}
The steam generation experiments in the lab are carried out under a solar simulator (Newport, 94081A, class ABB) which supplies solar flux of 1 kW m$^{-2}$ with an optical filter for the standard AM 1.5 G spectrum. 127g of DI water and seawater (3.5 wt\% NaCl) are prepared at the same initial temperature of 21$^\circ$C and placed in the 100 ml beakers with a mouth diameter of 50 mm. The steam generation device is floated on the solution surface and the mass of water is accurately monitored by an electric balance (RADWAG, PS 1000.X2.NTEP) connected to a computer for recording the real-time mass change. The real-time temperature is monitored by an infrared radiation camera (FLIR, A655sc).
\subsection{Materials characterizations}
The microscope images are obtained by using Metallurgical Microscope (AmScope, ME520TC-18M3) in the darkfield mode. The reflectivity spectra (UV-Visible-Near-infrared range: 200 nm $\sim$ 2500 nm) are measured by the Jasco V770 spectrophotometer at an incident angle of 6$^\circ$ with the ISN-923 60 mm BaSO$_4$ based integrating sphere equipped with PMT and PbS detectors. The reflectivity spectra are normalized by a PTFE based reflectance standard. The transmittance spectra (Mid-infrared region: 2.5 $\mu$m $\sim$ 20 $\mu$m) are measured by the Jasco FTIR 6600 spectrometer at a normal incident angle with reference to the background spectrum of a hydraulic pressed KBr film (20 psi). The Extech EC400 ExStik salinity meter is utilized to characterized the water quality of the collected water samples.
\section*{Acknowledgements}
This project is supported partly by the Soleeva Energy Innovation Award and the National Science Foundation through grant numbers CBET-1941743.
\newpage
| cbb3635e795f7bcf5ab979448c53f4bb2d64c11d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Strictly bipartite polyhedra of type $(4,4,2k)$}
\subsection{The case $\tau_1=(4,4,6)$}
\label{sec:446}
At this point we can now assume that all 1-vertices of $P$ are of type $(4,4,6)$ and that each 2-vertex of $P$ that is incident to a 6-face is of type $(4,6,4,6)$.
In particular, $P$ contains a 2-vertex $w\in V_2$ of this type.
Since there is no $(6,6)$-edge in $P$, the two 6-faces incident to $w$ cannot be adjacent.
In other words, the faces around $w$ must occur alternatingly of type 4 and type 6, which is the reason for writing the type $(4,6,4,6)$ with alternating entries.
On the other hand, $P$ contains $(4,4)$-edges, and none of these is incident to a $(4,6,4,6)$-vertex surrounded by alternating faces.
Thus, there must be further 2-vertices of a type other than $(4,6,4,6)$, necessarily \emph{not} incident to any 6-face.
These must then be of type
$$(4^r):=(\underbrace{4,...,4}_{r}),\quad\text{for some $r\ge 3$}.$$
\begin{proposition}
$r=5$.
\begin{proof}
If there is a $(4^r)$-vertex, \cref{res:spherical_interior_angles_are_good} $(i)$ yields $\beta_2^2=2\pi/r$.
Analogously,~from~the existence of a $(4,6,4,6)$-vertex follows
$$2\beta_2^2+2\beta_2^3\overset{(i)}=2\pi\quad\implies\quad \beta_2^3=\frac{2\pi-2\beta_2^2}2 = \Big(1-\frac2r\Big)\pi.$$
\begin{samepage}%
Recall $k\beta_1^k+k\beta_2^k>2\pi(k-1)$ from \cref{res:spherical_interior_angles_are_good} $(ii)$.
Together with the previously established values for $\beta_2^2$ and $\beta_2^3$, this yields
\begin{equation}
\label{eq:4}
\begin{array}{rcl}
\beta_1^2 &\!\!\!>\!\!\!& \displaystyle \frac{2\pi(2-1)-2\beta_2^2}2= \Big(1-\frac2r\Big)\pi, \quad\text{and} \\[1ex]
\beta_1^3 &\!\!\!>\!\!\!& \displaystyle \frac{2\pi(3-1)-3\beta_2^3}3= \Big(\frac13+\frac2r\Big)\pi.
\end{array}
\end{equation}
Since the 1-vertices are of type $(4,4,6)$, \cref{res:spherical_interior_angles_are_good} $(i)$ yields
\end{samepage}
$$2\pi\overset{(i)}=2\beta_1^2+\beta_1^3 \overset{\eqref{eq:4}}> 2\Big(1-\frac2r\Big)\pi + \Big(\frac13+\frac2r\Big)\pi = \Big(\frac73-\frac2r\Big)\pi.$$
And one checks that this rearranges to $r<6$.
This leaves us with the options $r\in\{3,4,5\}$.
If $r=4$, then $\beta_2^3=\pi/2=\beta_2^2$, which~is~impossible by equation \eqref{eq:angle_relations}.
And if $r=3$, then \eqref{eq:4} yields $\beta_1^3 > \pi$, which is also impossible~for a convex face of a spherical polyhedron.
We are left with $r=5$.
\end{proof}
\end{proposition}
To summarize: $P$ is a strictly bipartite polyhedron in which~all 1-vertices are~of type $(4,4,6)$, and all 2-vertices are of types $(4,6,4,6)$ or $(4^5)$,~and both types actually occur in $P$.
This information turns out to be sufficient to uniquely determine the edge-graph of $P$, which is shown in \cref{fig:graph}.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{img/unique_graphs}
\caption{The edge-graph of the final candidate polyhedron.}
\label{fig:graph}
\end{figure}
This graph can be constructed by starting with a hexagon in the center with vertices of alternating colors (indicating the partition classes).
One then successively adds further faces (according to the structural properties determined above), layer by layer.
This process involves no choice and thus the result is unique.
As mentioned in \cref{rem:alternative_definiton}, a bipartite polyhedron has an edge in-sphere.
Thus, $P$ is a polyhedral realization of the graph in \cref{fig:graph} with an edge in-sphere.
It is known that any two such realizations are related by a projective transformation \cite{sachs1994coin}.
One representative $Q\subset\RR^3$ from this class (which we do not yet claim~to coincide with~$P$) can~be~constructed by
applying the following operation $\star$ to each vertex of the regular icosahedron:
\begin{center}
\includegraphics[width=0.5\textwidth]{img/icosahedron_transformation}
\end{center}
The operation is performed in such a way, so that
\begin{itemize}
\item the five new \enquote{outer} vertices of the new 4-gons are positioned in the centers of edges of the icosahedron.
\item the edges of each new 4-gon are tangent to a common sphere centered at the center of the icosahedron
\end{itemize}
The resulting polyhedron $Q$ looks as follows:\\[-1.5ex]
\begin{center}
\includegraphics[width=0.7\textwidth]{img/Conway_DL0D}
\end{center}
One can verify that $Q$ has indeed the desired edge-graph.
It is clear from the construction that $Q$ has an edge in-sphere, and any two of its 4-gonal or 6-gonal faces are congruent (as we would expect from a bipartite polyhedron).
Like-wise, $P$ has an edge in-sphere and the same edge-graph.
Hence, $P$ must be a projective transformation of $Q$.
However, any~projective transformation that is not just a re-orientation or a uniform rescaling will inevitably destroy the property of congruent faces.
In conclusion, we can assume that $P$ is identical to $Q$ (up to scale and orientation).
It remains to check whether $Q$ is indeed a bipartite~polyhedron. For this, recall that any two of the following properties imply the third (\shortStyle{cf.}\ \cref{rem:alternative_definiton}):
\begin{myenumerate}
\item $Q$ has an edge in-sphere.
\item $Q$ has all edges of the same length.
\item for each vertex $v\in \F_0(Q)$, the distance $\|v\|$ only depends on the partition class of the vertex.
\end{myenumerate}
Now, $Q$ satisfies \itm1 by construction, and it would need to satisfy both \itm2 and \itm3 in order to be bipartite.
The figure certainly suggests that all edges of $Q$ are of the same length.
However, as we shall show now, $Q$ cannot satisfy both \itm2 and \itm3 at the same time, and thus, can satisfy neither.
In particular, the edges must have a tiny difference in length that cannot be spotted visually, making $Q$ into a remarkable near-miss (we will quantify this below).
For what follows, let us assume that \itm2 holds, that is, that all edges of $Q$ are of the same length, in particular, that all 4-gons are rhombuses.
Our goal is to show that $\|v\|$ depends on the type of the vertex $v\in V_2$ (not only its partition class), establishing that \itm3 does not hold.
\iffalse
\newpage
\begin{theorem}
Given a 3-connected planar graph $G$, there is a unique polyhedron (up to scale and orientation) with
\begin{myenumerate}
\item $G$ as its edge-graph,
\item an edge-insphere, and
\item a centroid that agrees withe the centroid of the edge-insphere.
\end{myenumerate}
The polyhedron with these properties is called the \underline{\smash{canonical polyhedron}} of $G$.
\end{theorem}
\begin{proposition}
A bipartite polyhedron is the canonical polyhedron of its edge-graph
\begin{proof}
\msays{TODO}
\end{proof}
\end{proposition}
We construct a polyhedron with $G$ as its edge-graph.
Consider the following~polyhedron $Q$:
\begin{center}
\includegraphics[width=0.7\textwidth]{img/Conway_DL0D}
\end{center}
$Q$ is constructed from the regular icosahedron by applying the following operation~$\star$ to each of its vertices:
\begin{center}
\includegraphics[width=0.5\textwidth]{img/icosahedron_transformation}
\end{center}
Note that the \enquote{outer} vertices of each of the 4-gonal faces are supposed to be~exactly in the middle of an edge of the icosahedron.
This operation has one degree of freedom (namely, the distance of the vertex of degree five from the centroid).
We choose to performe it in such a way, so that the resulting polyhedron has an edge-insphere.
This happens when the degree-five vertex has the same distance from the origin as the \enquote{outer} vertices.
Now, $Q$ has the following properties:
\begin{itemize}
\item $G$ is the edge-graph of $Q$,
\item $Q$ has an edge-insphere, and
\item the centroid of the polyhedron agrees with the centroid of the points of tangency of the edges with the edge-insphere.
\end{itemize}
\newpage
This graph is transitive on all kinds of substructures, \shortStyle{e.g.}\ it is transitive on all faces/edges/vertices that share the same type.
This is easiest seen by considering the following polyhedron which has the above graph as its edge-graph.
\begin{center}
\includegraphics[width=0.7\textwidth]{img/Conway_DL0D}
\end{center}
This polyhedron can be constructed from the regular icosahedron by applying the following operation $\star$ to each vertex:
\begin{center}
\includegraphics[width=0.5\textwidth]{img/icosahedron_transformation}
\end{center}
The 4-gonal faces are placed in such a way so that their \enquote{outer} vertices are exactly in the middle of an edge of the icosahedron.
The symmetry properties of the~polyhedron, and hence the symmetry properties of the graph are apparent from this construction.
The operation $\star$ is not completely well-defined, but has one degree of freedom.
In particular, one can choose to \enquote{cut} each vertex in such a way, so that all edges become of the same length and the 4-gonal faces become rhombuses), or so that all 2-vertices have the same distance from the origin (or equivalently, so that the polyhedron has an edge-insphere).
The question then is, whether both can be achieved simultaneously.
Recall the definition of the \emph{canonical polyhedron} of a (planar, 3-connected) graph: it is the unique polyhedron (up to scale and orientation) which has an edge-insphere and for which the
\newpage
By way of construction (it does not matter which hexagon is put into the center), it follows that this graph is transitive on its 6-gonal faces.
Similarly, this graph is transitive on its rhombic faces, 1-vertices, 2-vertices and edges of the same type.
Similarly, piecing together polygons according to above rules, one finds that there is a unique polyhedron with this edge-graph and all edges of the same length.
While a graph can be realized in several metrically distinct ways as a polyhedron, the facts that all faces of the same type must be congruent induces a high degree of symmetry:
\begin{center}
\includegraphics[width=0.7\textwidth]{img/Conway_DL0D}
\end{center}
This polyhedron belongs to a class occasionally named \emph{symmetroheda}, and has Conway notation $\mathrm{dL_0 D}$.
Alternatively, it can be constructed by applying the following operation $\star$ to each vertex of the regular icosahedron:
\begin{center}
\includegraphics[width=0.5\textwidth]{img/icosahedron_transformation}
\end{center}
The rhombuses are placed in such a way so that their \enquote{outer} vertices are exactly in the middle of an edge of the icosahedron.
The centroid of the resulting polyhedron coincides with the one of the icosahedron.
We can therefore compute the distances of its vertices from this center to determine whether it is bipartite or not.
\iffalse
In fact, if one would try to construct the edge-graph of this hypothetical polyhedron from this information alone, one would find that there is a unique way to do so.
We will not make use of this fact, but we included the resulting image below to give the reader a headstart on some intuition for what is to come.
\begin{center}
\includegraphics[width=0.6\textwidth]{img/Conway_DL0D}
\end{center}
The image shows two perspectives of the same hypothetical strictly bipartite polyhedron.
This polyhedron belongs to a class occasionally named \emph{symmetroheda}, and it can be described using Conway notation $\mathrm{dL_0 D}$.
So the conditions on the vertex types are combinatorially feasible, and the image stands a visual test for having all edges of the same length.
In fact, it is possible to realize $P$ (its combinatorial type) with all edges of the same length (as required by \cref{def:bipartite} $(i)$), and it is also possible to realize it with vertices on exactly two spheres (as required by \cref{def:bipartite} $(iii)$).
However, it turns out that it is not possible to achieve both at the same time.
Instead, we will find that fixing the edge length results in having two \emph{slightly} distinct values for $\|v\|,v\in V_2$ depending on the type of the vertex.
The depicted polyhedron is a remarkable near-miss.
The image suggests to conjecture that $P$ has icosahedral symmetry.
This is in fact true.
$P$ can be given as the intersection of half spaces $\mathcal H(\sigma)$ defined by its faces $\sigma\in\F_2(P)$.
We show that if we drop all $\mathcal H(\sigma)$ associated to the rhombic faces, we are left with the regular icosahedron (see \cref{res:P_is_icosahedral} $(ii)$).
The other way around, adding back in the half spaces of the rhombic faces is equivalent to applying the following operation $\star$ to each vertex of the icosahedron:
\begin{center}
\includegraphics[width=0.6\textwidth]{img/icosahedron_transformation}
\end{center}
From each vertex of the icosahedron this generates a 2-vertex of type $(4^5)$.
The vertices opposite to the $(4^5)$-vertex in each rhombus must lie exactly in the center of an edge of the icosahedron in order for $P$ to not have any $(6,6)$-edges, and for the operation to work on all vertices simultaneously.
\begin{proposition} \label{res:P_is_icosahedral} \textcolor{white}{x}
\begin{myenumerate}
\item Let $P'$ be the polyhedron obtained as the convex hull of all $(4,6,4,6)$-vertices of $P$. Then $P'$ is the icosidodecahedron.
\item Let $P''$ be the polyhedron obtained as the intersection of all $\mathcal H(\sigma)$ for all 6-faces of $P$. Then $P''$ is the regular icosahedron.
\end{myenumerate}
\end{proposition}
The image below shows the regular icosahedron (left) and the icosidodecahedron (right) which can be obtained from the icosahedron by truncating its vertices up to the edge midpoints.
In particular, the triangular faces of the icosidodecahedron are in one-to-one correspondence with the faces of the icosahedron, and are defined by the same halfspaces.
Moreover, the icosidodecahedron is the uniquely determined polyhedron in which at each vertex meet two regular triangles and two regular pentagons.
\begin{center}
\includegraphics[width=0.5\textwidth]{img/icosidodecahedron}
\end{center}
\begin{proof}[Proof of \cref{res:P_is_icosahedral}]
...
\end{proof}
Assuming that all edges of $P$ are of the same length, the operation $\star$ is well-defined.
We are going to show, that the resulting polyhedron has distinct values for $\|v\|,v\in V_2$ depending on the type of $v$.
\fi
\fi
For this, start from the following well-known construction of the regular icosahedron from the cube of edge-length 2 centered at the origin.
\begin{center}
\includegraphics[width=0.95\textwidth]{img/constructing_ico}
\end{center}
The construction is as follows:
insert a line segment in the center of each face of the cube as shown in the left image. Each line segment is of length $2\varphi$, where $\varphi\approx 0.61803$ is the positive solution of $\varphi^2=1-\varphi$ (one of the numbers commonly knows as the \emph{golden ratio}).
The convex hull of these line segments gives the icosahedron with edge length $2\varphi$.
It is now sufficient to consider a single vertex of the icosahedron together with its incident faces.
The image below shows this vertex after we applied $\star$.
\begin{center}
\includegraphics[width=0.9\textwidth]{img/5_pyramid_projection}
\end{center}
The image on the right is the orthogonal projection of the configuration on the left onto the $yz$-plane.
This projection makes it especially easy to give 2D-coordinates for several important points:
\begin{center}
\includegraphics[width=0.85\textwidth]{img/5_pyramid_coordinates}
\end{center}
The points $\mathbf A$ and $\mathbf C$ are 2-vertices of $Q$ of type $(4^5)$ and $(4,6,4,6)$ respectively.
Both points and the origin $\mathbf O$ are contained in the $yz$-plane onto which we projected.
Consequently, distances between these points are preserved during the projection, and assuming that $Q$ is bipartite, we would expect to find $|\overline{\mathbf {OA}}|=|\overline{\mathbf{OC}}|=r_2$.
We shall see that this is not the case, by explicitly computing the coordinates of $\mathbf A$ and $\mathbf C$ in the new coordinate system $(y,z)$.
By construction, $\mathbf C=(0,1)$ and $|\overline{\mathbf{OC}}|=1$.
Other points with easily determined coordinates are $\mathbf P$, $\mathbf Q$, $\mathbf R$, $\mathbf S$, $\mathbf T$ (the midpoint of $\mathbf R$ and $\mathbf S$) and $\mathbf U$ (the midpoint of $\mathbf Q$ and $\mathbf S$).
By construction, the point $\mathbf B$ lies on the line segment $\overline{\mathbf{QT}}$.
The parallel projection of~a~rhombus is a (potentially degenerated) parallelogram, and thus, opposite edges in the projection are still parallel.
Hence, the gray edges in the figure are parallel.
For that reason, the segment $\overline{\mathbf{UB}}$ is parallel to $\overline{\mathbf{PQ}}$.
This information suffices to determine the coordinates of $\mathbf B$, which is now the intersection of $\overline{\mathbf{QT}}$ with the parallel of $\overline{\mathbf{PQ}}$ through $\mathbf U$.
The coordinates are given in the figure.
The rhombus containing the vertices $\mathbf A$, $\mathbf B$ and $\mathbf C$ degenerated to a line.
Its~fourth vertex is also located at $\mathbf B$.
Therefore, the segments $\overline{\mathbf{CB}}$ and $\overline{\mathbf{BA}}$ are translates of each other.
Since the point $\mathbf B$ and the segment $\overline{\mathbf{CB}}$ are known, this allows the computation of the coordinates of $\mathbf A$ as given in the figure.
We can finally compute $|\overline{\mathbf{OA}}|$.
For this, recall $(*)\,\varphi^{2n}=F_{2n-2}-\varphi F_{2n-1}$, where $F_n$ denotes the $n$-th \emph{Fibonacci number} with initial conditions $F_0=F_1=1$.
Then
\begin{align*}
|\overline{\mathbf O\mathbf A}|^2
&= (4\varphi-3)^2+(3\varphi-1)^2 \\
&= 25\varphi^2 - 30\varphi + 10 \\[-1ex]
&\overset{\mathclap{(*)}}= 25(1-\varphi)-30\varphi + 10 \\
&= 35 - 55\varphi
\\[-1ex] &= 1 + (34 - 55\varphi)\overset{\mathclap{(*)}}=1+\varphi^{10} > 1,
\end{align*}
and thus, $Q$ cannot be bipartite.
Remarkably, we find that
$$|\overline{\mathbf O\mathbf A}|=\sqrt{1+\varphi^{10}}\approx 1.00405707$$
is only about $0.4\%$ larger than $|\overline{\mathbf O\mathbf C}|=1$, and so while $Q$ is not bipartite, it~is~a~remarkable near-miss.
Since $P$ was assumed to be bipartite, but was also shown to be identical to $Q$, we reached a contradiction, which finally proves \cref{res:strictly_bipartite_polyhedra}, and the goal of the paper is achieved.
\section{}
\label{sec:appendix}
\iffalse
\subsection*{Spherical polyhedron}
There is a very useful second point of view onto a bipartite polyhedron via the associated \emph{spherical polyhedron}.
The spherical polyhedron $P^S$ is obtained as the central projection of the vertices, edges and faces of $P$ onto the unit sphere.
Formally, $P^S$ is an embedding of the edge graph of $P$ into the unit sphere obtained by projecting the vertices and edges via
$$\RR^3\setminus\{0\}\to\Sph^2,\;x\mapsto \frac x{\|x\|}.$$
\begin{center}
\includegraphics[width=0.45\textwidth]{img/spherical_polyhedron}
\end{center}
The vertices of $P^S$ are points on the unit sphere,
its edges are great circle arcs, and its faces (the regions \enquote{cut out} by the embedded graph) are \enquote{spherical polygons}.
To every vertex $v$, edge $e$ and face $\sigma$ of $P$ we denote their projections by $v^S$, $e^S$ and $\sigma^S$ respectively.
Clearly, the interior angles of a spherical face at a spherical vertex add up to $2\pi$.
In the following, let $P$ denote a bipartite polyhedron.
\begin{proposition}\label{res:spherical_edge_length}
All edges of $P^S$ are of the same length.
\begin{proof}
Choose an edge $e^S\in\F_1(P^S)$ with underlying edge $e\in \F_1(P)$.
The edge $e$ connects vertices $v_1\in V_1$ and $v_2\in V_2$.
The length of $e^S$ is the angle $\theta:=\angle(v_1,v_2)$.
This angle is uniquely determined by the law of cosine:
\begin{align*}
\|v_1-v_2\|^2 &= \|v_1\|^2 + \|v_2\|^2 - 2\|v_1\|\|v_2\|\cos\angle(v_1,v_2) \\
\ell^2 &= r_1^2 + r_2^2 - 2 r_1 r_2 \cos\angle(v_1,v_2).
\end{align*}
\end{proof}
\end{proposition}
\begin{proposition}\label{res:spherical_unique_shape}
For some spherical face $\sigma^S\in\F_2(P^S)$, any of the following two properties of $\sigma^S$ determines the other one:
\begin{myenumerate}
\item its type $2k$,
\item its interior angles (in fact, a single angle is sufficient).
\end{myenumerate}
\end{proposition}
Again, this result can be interpreted as follows: two spherical faces of $P^S$ with the same type or the same interior angles are congruent. In the spherical case it is clear that this congruence must be realized by an orthogonal map.
\begin{proof}[Proof of \cref{res:spherical_unique_shape}]
Two $2k$-faces of $P$ are congruent via an orthogonal map $T\in\Ortho(\RR^3)$.
Orthogonal maps commute with the central projection, and thus, the spherical faces (which are the projections of the polyhedral faces) are also congruent via the same map $T$.
Consider a spherical face $\sigma^S\in\F_2(P^S)$ with interior angle $\beta$ at an $i$-vertex $v^S\in\F_0(\sigma^S)$.
Let $w_1^S,w_2^S\in\F_0(\sigma^S)$ be the two neighbors of $v^S$ that surround that angle.
Since the edge between $v$ and $w_i^S$ has a known length by \cref{res:spherical_edge_length}, the \emph{spherical law of cosine}\footnote{$\cos(c)=\cos(a)\cos(b)+\sin(a)\sin(b)\cos(\beta)$.} uniquely determines the spherical distance between $w_1^S$ and $w_2^S$, which is just the angle $\angle(w_1,w_2)$.
The usual law of cosine that determines the angle of the face $\sigma$ at $v$.
By \cref{res:unique_shape}, this angle uniquely determines the type of $\sigma$, and as previously seen, the shape of $\sigma^S$ up to an orthogonal transformation.
\end{proof}
Spherical law of cosine:
$$\cos(c)=\cos(a)\cos(b) + \sin(a)\sin(b)\cos(\gamma).$$
$$\chi(\alpha):=1-\cos(\alpha).$$
\begin{align*}
\rho^2 &= \ell^2 + \ell^2 - 2\ell^2 \cos\alpha(\sigma,v) = 2\ell^2\chi(\alpha(\sigma,v)),\\
\rho^2 &= r_i^2+r_i^2-2r_i^2\cos\angle(w_1, w_2) = 2r_i^2\chi(\angle(w_1,w_2)),\\
\cos\angle(w_1^S, w_2^S) &= \cos^2\ell^S+\sin^2\ell^S\cos\alpha(\sigma^S,v^S)=1-\cos^2\ell^S\chi( \beta(\sigma^S\!,v^S)) \\
\chi(\angle(w_1,w_2))&=\cos^2\ell^S\chi(\beta(\sigma^S\!,v^S)).
\end{align*}
\iffalse
\begin{proposition}
$0\in\mathrm{int}(P)$.
\begin{proof}
Suppose that $0\not\in\mathrm{int}(P)$.
Since $P$ is the intersection of the half-spaces determined by its faces, there must be a face $\sigma$ whose associated half-space does not contain the origin.
Let $n\in\Sph^2$ be the outwards pointing normal of $\sigma$, and $v\in\F_0(P)$ be an incident 1-vertex.
Since the points of $\sigma$ maximize the functional $\<n,\free\>$, we have $\<n,w\>\le \<n,v\><\<n,0\>=0$ for all
$w\in\F_0(P)$.
Now consider the projection map
$$\pi:\RR^3\to v^\bot, w\mapsto w-v\frac{\<v,w\>}{\|v\|^2}$$
onto the 2-dimensional subspace $v^\bot$. We want to show that $\<\pi(w),\pi(n)\><0$ for all $w\in\F_0(P)$ as well.
Observe
$$\<\pi(n),\pi(w)\> = \Big\<n-v\frac{\<n,v\>}{\|v\|^2},w-v\frac{\<w,v\>}{\|v\|^2}\Big\> = \<n,w\>-\frac1{\|v\|^2}\<w,v\>\<n,v\>.$$
\end{proof}
\end{proposition}
\fi
\iffalse
\begin{proposition}
All dihedral angles of $P$ are obtuse.
\begin{proof}
Fix a vertex $v\in V_1$ with incident edges $e_1,e_2,e_3\in\F_1(P)$ and incident faces $\sigma_{12},\sigma_{23},\sigma_{31}\in F_2(P)$, so that $\sigma_{ij}$ is incident to $e_i$ and $e_j$.
Let $u_i\in\RR^3$ be the direction of the edge $e_i$ emanating from $v$.
The interior angle $\alpha_1^k$ of $\sigma_{ij}$ at $v$ is then $\angle\<u_i,u_j\>$, and since $k\ge 2$, from \cref{res:2d_shape} follows
$$\alpha_1^k >\Big(1-\frac1k\Big)\pi \ge \frac\pi2$$
and thus $\<u_i,u_j\><0$.
Suppose that the $e_i$ are enumerated in a way so that $n_{ij}:=e_i\times e_j$ is an outwards pointing normal vector of $\sigma_{ij}$.
The dihedral angle at $e_j$ is then $\pi-\angle(n_{ij},n_{jk})$ and we can compute that
\begin{align*}
\<n_{ij},n_{jk}\>
&= \<u_i\times u_j,u_j\times u_k\> \\
&= \<u_j\times(u_j\times u_k),u_i \> \\
&= \<\<u_j,u_k\> u_j-\<u_j,u_j\>u_k,u_i \> \\
&= \underbrace{\<u_j,u_k\>}_{<0} \underbrace{\<u_j,u_i\>}_{<0}-\underbrace{\<u_j,u_j\>}_{>0}\underbrace{\<u_k,u_i \>}_{<0} > 0,
\end{align*}
which means $\angle(n_{ij},n_{jk})<\pi/2$ and the dihedral angle is indeed obtuse.
Since every edge is of this form (incident to a 1-vertex), we are done.
\end{proof}
\end{proposition}
\fi
\begin{proposition}
$0\in\mathrm{int}(P)$.
\begin{proof}
The proof proceeds in several steps.
\textit{Step 1}:
Fix a 1-vertex $v\in V_1$ with neighbors $w_1,w_2,w_3\in V_2$, and let $u_i:=w_i-v_i$ be the direction of the edge $\conv\{v,w_i\}$ emanating from $v$.
Let $\sigma_{ij}\in\F_2(P)$ denote the $2k$-face containing $v,w_i$ and $w_j$.
The interior angle of $\sigma_{ij}$ at $v$ then is $\angle(u_i,u_j)$, which by \cref{res:2d_shape} and $k\ge 2$ estimates as
$$\angle(u_i,u_j)>\Big(1-\frac1k\Big)\pi \ge \frac\pi2\quad\implies\quad \<u_i,u_j\><0.$$
\textit{Step 2}:
Besides $v$, the polyhedron $P$ contains another 1-vertex $v'\in V_1$.
Then $v'\in v+\cone\{u_1,u_2,u_3\}\setminus\{v\}$, which means that there are non-negative coefficients $a_1,a_2,a_3\ge 0$, at least one positive, so that $v+a_1 u_1+ a_2u_2 + a_3u_3 = v'$.
Rearranging and applying $\<v,\free\>$ yields
\begin{align*}
(*)\quad a_1\<v,u_1\>+a_2\<v,u_2\>+a_3\<v,u_3\>
&=\<v,v'\>-\<v,v\> \\
&= r_1^2\cos\angle(v,v')-r_1^2 < 0.
\end{align*}
As seen via $r_2^2=\|w_i\|^2=\|u_i-v\|^2=\ell^2+r_1^2-2\<v,u_i\>$, the inner product $\<v,u_i\>$ is the same for all $i\in\{1,2,3\}$.
Conclusively, since there is at least positive coefficient $a_i$, from $(*)$ follows $\<v,u_i\><0$.
\textit{Step 3}:
By the previous steps, $\{v,u_1,u_2,u_3\}$ is a set of four vectors with pair-wise negative inner product.
For such an arrangement there exist positive coefficients $a_0,...,a_3>0$ with $a_0 v+a_0 v_1+a_2 v_2+a_3 v_3=0$ (for a proof, see \cref{res:sum_to_zero} in the Appendix).
In other words: $0\in v+\mathrm{int}(\cone\{u_1,u_2,u_3\})
$.
\textit{Step 4}:
If $H(\sigma)$ denotes the half-space associated to the face $\sigma\in\F_2(P)$, then
$$0\in v+\mathrm{int}(\cone\{u_1,u_2,u_3\}) = \bigcap_{\sigma\sim v} \mathrm{int}(H(\sigma)).$$
Thus, $0\in\mathrm{int}(H(\sigma))$ for all faces $\sigma$ incident to $v$.
But since every face is incident to a 1-vertex, we obtain $0\in\mathrm{int}(H(\sigma))$ for all $\sigma\in\F_2(P)$, and thus $\sigma\in\mathrm{int}(P)$ as well.
\end{proof}
\end{proposition}
\iffalse
Farka's Lemma can be stated in many forms.
We decided to cite the form which is closest to its geometric application in the context of this paper.
\begin{theorem}[Farkas' Lemma, \cite{farkas1902theorie}]
\label{res:farkas}
Given vectors $u_1,...,u_n\in\RR^d$ and a point $x\in\RR^d$, exactly one of the following is true:
\begin{myenumerate}
\item $x\in\mathrm{int}(\cone\{u_1,...,u_n\})$, that is, there are positive coefficients $a_1,...,a_n>0$ with $a_1 u_1+\cdots +a_n u_n=x$.
\item There is a non-zero vector $y\in\RR^d\setminus\{0\}$ so that $\<x,y\>\le 0$ and $\<u_i,y\>\ge0$ for all $i\in\{1,\...,n\}$.
\end{myenumerate}
\begin{proof}[Sketch of proof.]
$(i)$ states that $x$ is in the interior of a certain cone, and $(ii)$ states that $x$ can be (weakly) separated from the interior of that cone by a hyperplane.
Exactly one of these must be true.
\end{proof}
\end{theorem}
\begin{lemma}\label{res:spherical_code}
There can be at most $d+1$ vectors in $\RR^d$ with pair-wise negative inner product.
\begin{proof}
An example of $d+1$ points is given by the vertices of a regular $d$-dimensional simplex centered at the origin.
Suppose that $x_1,...,x_{d+2}\in\RR^d$ are $d+2$ points with pair-wise negative inner product and set $X=(x_1,...,x_{d+2})\in\RR^{d\times(d+2)}$ the matrix with the $x_i$ as columns.
The null space of $X$ is at least 2-dimensional.
In particular, we can find a vector $v\in\ker(X)$ with some of its components positive, say $I_+\subset\{1,...,d+2\}$, and some negative, $I_- = \{1,\...,d+2\}\setminus I_+$.
This means
$$Xv=0\quad\implies\quad \sum_{i\in I_+} v_i x_i = \sum_{i\in I_-}(-v_i)x_i.$$
This leads to a contradiction as follows:
$$
0
\le \Big\<\sum_{i\in I_+} v_i x_i,\sum_{i\in I_+} v_i x_i\Big\>
= \Big\<\sum_{i\in I_+} v_i x_i,\sum_{i\in I_-}(-v_i)x_i\Big\>
= \sum_{\substack{i\in I_+\\j\in I_-}} \underbrace{(-v_i v_j)}_{>0}\underbrace{\<x_i,x_j\>}_{<0} < 0.
$$
\end{proof}
\end{lemma}
\begin{corollary}\label{res:spherical_code_spread_out}
If $d+1$ vectors $x_0,...,x_d\in\RR^d$ have pair-wise negative inner product, then they are not on the same side of a hyperplane, that is, there is no non-zero $y\in\RR^d\setminus\{0\}$ with $\<x_i,y\>\ge 0$ for all $i\in\{0,...,d\}$.
\begin{proof}
Suppose that such a $y\in\RR^d$ exists, and set $I:=\{i\in\{0,...,d\}\mid \<y,x_i\><0\}$.
Note that $I$ cannot be empty, as otherwise $x_0,...,x_d\in y^\bot$, which would be $d+1$ vectors with pair-wise negative inner product in a $(d-1)$-dimensional space, which is impossible by \cref{res:spherical_code}.
We then set
$$y':=y + \sum_{i\in I} x_i \not=y $$
For all $j\in\{0,...,d\}$ then holds
$$\<y',x_j\> = \underbrace{\<y,x_j\>}_{\le 0} + \sum_{i\in I} \underbrace{\<x_i,x_j\>}_{<0} < 0.$$
But then, $\{x_0,...,x_d,y\}$ is a set of $d+2$ vector with pair-wise negative inner product, which is impossible by \cref{res:spherical_code}.
\end{proof}
\end{corollary}
\fi
\fi
\subsection{Geometry}
\begin{proposition}\label{res:sum_to_zero}
Given a set $x_0,...,x_d\in\RR^d\setminus\{0\}$ of $d+1$ vectors with pair-wise negative inner product,
then there are \ul{positive} coefficients $\alpha_0,...,\alpha_d>0$ with
$$\alpha_0 x_0+\cdots +\alpha_d x_d=0.$$
\iftrue
\begin{proof}
We proceed by induction.
The induction base $d=1$ which is trivially true.
Now suppose $d\ge 2$, and, W.l.o.g.\ assume $\|x_0\|=1$.
Let $\pi_0$ be the orthogonal projection onto $x_0^\bot$, that is, $\pi_0(u):= u - x_0\<x_0,u\>$.
In particular, for $i\not=j$ and $i,j>0$
$$\<\pi_0(x_i),\pi_0(x_j)\> = \underbrace{\<x_i,x_j\>}_{<0} - \underbrace{\<x_0,x_i\>}_{<0}\underbrace{\<x_0,x_j\>}_{<0} < 0.$$
Then $\{\pi(x_1),\...,\pi_0(x_d)\}$ is a set of $d$ vectors in $x_0^\bot\cong\RR^{d-1}$ with pair-wise negative inner product.
By induction assumption there are positive coefficients $\alpha_1,...,\alpha_d>0$ so that $\alpha_1\pi_0(x_1)+\cdots+\alpha_d\pi_0(x_d)=0$.
Set $\alpha_0:=-\<x_0,\alpha_1x_1+\cdots+\alpha_d x_d\>>0$.
We claim that $x:=x_0\alpha_0+\cdots+\alpha_dx_d=0$.
Since $\RR^d=\Span\{x_0\}\oplus x_0^\bot$, it suffices to check that $\<x_0,x\>=0$ as well as $\pi_0(x)=0$. This follows:
\begin{align*}
\<x_0,x\> &= \alpha_0\underbrace{\<x_0,x_0\>}_{=1} + \underbrace{\<x_0,\alpha_1x_1+\cdots+\alpha_d x_d\>}_{=-\alpha_0}=0, \\
\pi_0(x) &= \alpha_0 \underbrace{\pi_0(x_0)}_{=0} + \underbrace{\alpha_1\pi_0(x_1)+\cdots+\alpha_d\pi_0(x_d)}_{=0}=0.
\end{align*}
\end{proof}
\else
\begin{proof}
Let $X=(x_0,...,x_d)\in\RR^{d\x(d+1)}$ be the matrix with the $x_i$ as columns.
There are more vectors than dimensions, and thus, the $x_i$ must be linearly dependent, \shortStyle{i.e.,}\ there is an $a\in\RR^{d+1}$ with $Xa=0$.
We need to show that $a$ can be chosen so, that all entries are positive.
Consider the matrix $Y:=X^{\Tsymb}\! X\in\RR^{(d+1)\x(d+1)}$, which is positive semi-definite and satisfies $Ya=X^{\Tsymb}\!(Xa)=0$.
Since eigenvalues of positive semi-definite matrices are non-negative, we found that the smallest eigenvalue of $Y$ is zero, and $a$ is an associated eigenvector.
We have $Y_{ij}=\<x_i,x_j\><0$ for $i\not=j$, and thus, for an appropriately chosen $\eps>0$, $Z:=I-\eps Y$ is a matrix with only positive entries.
Each eigenvalue of $Z$ is of the form $1-\eps \lambda$, where $\lambda$ is an eigenvalue of $Y$.
In particular, the maximal eigenvalue of $Z$ is $1$, and $a$ is an associated eigenvector.
By \emph{Perron-Frobenius}, the entries of $a$ are non-zero and have all the same sign.
In particular, they can be chosen positive.
\end{proof}
\fi
\end{proposition}
\begin{proposition}\label{res:simple_vertex}
Let $P\subset\RR^3$ be a polyhedron with $v\in\F_0(P)$ a vertex of degree three.
The interior angles of the faces incident to $v$ determine the dihedral angles at the edges incident to $v$ and vice versa.
\begin{proof}
For $w_1,w_2,w_3\in\F_0(P)$ the neighbors of $v$, let $u_i:=w_i-v$ denote the direc\-tion of the edge $e_i$ from $v$ to $w_i$.
Let $\sigma_{ij}$ be the face that contains $v,w_i$ and $w_j$.
Then $\angle(u_i,u_j)$ is the interior angle of $\sigma_{ij}$ at $v$.
The set $\{u_1,u_2,u_3\}$ is uniquely determined (up to some orthogonal transformation) by the angles $\angle(u_i,u_j)$.
Furthermore, since $P$ is convex, $\{u_1,u_2,u_3\}$ forms a basis of $\RR^3$, and this uniquely determines the \emph{dual basis} $\{n_{12}, n_{23}, n_{31}\}$ for which $\<n_{ij},u_i\>=\<n_{ij},u_j\>=0$.
In other words, $n_{ij}$ is a normal vector to $\sigma_{ij}$.
The dihedral angle at the edge $e_j$ is then $\pi-\angle(n_{ij},n_{jk})$, hence uniquely determined.
The other direction is analogous, via constructing $\{u_1,u_2,u_3\}$ as the dual basis to the set of normal vectors.
\end{proof}
\end{proposition}
\subsection{Computations}
\label{sec:computations}
The edge lengths in a spherical polyhedron are measured as angles between its end vertices. Consider adjacent vertices $v_1^S,v_2^S\in\F_0(P^S)$, then the incident edge has (arc-)length $\ell^S:=\angle(v_1^S,v_2^S)=\angle(v_1,v_2)$.
It follows from \cref{res:angles_between_vertices} that these angles are completely determined by the parameters, hence the same for all edges of $P^S$.
\begin{proposition}\label{res:computation}
For a face $\sigma\in\F_2(P)$ and a vertex $v\in\F_0(\sigma)$, there is a direct relationship between the value of $\alpha(\sigma,v)$ and the value of $\beta(\sigma,v)$.
\begin{proof}
Let $w_1,w_2\in V_2$ be the neighbors of $v$ in the $2k$-face $\sigma$, and set $u_i:=w_i-v$. Then $\angle(u_1,u_2)=\alpha(\sigma,v)$.
W.l.o.g.\ assume that $v$ is a 1-vertex (the argument is equivalent for a 2-vertex).
For convenience, we introduce the notation $\chi(\theta):= 1-\cos(\theta)$.
We find that
\begin{align*}
(*)\quad 2\ell^2\cdot\chi(\alpha(\sigma,v))
&= \ell^2+\ell^2-2\ell^2\cos(\angle(u_1,u_2)) \\
&= \|u_1\|^2+\|u_2\|^2 - 2\<u_1,u_2\> \\
&= \|u_1-u_2\|^2
= \|w_1-w_2\|^2 \\
&= \|w_1\|^2+\|w_2\|^2-2\<w_1,w_2\>\\
&= r_2^2+r_2^2-r_2^2\cos\angle(w_1,w_2)
= 2r_2^2\cdot\chi(\angle(w_1,w_2)).
\end{align*}
The side lengths of the spherical triangle $w_1^Sv^Sw_2^S$ are $\angle(w_1,w_2), \ell^S$ and $\ell^S$.
By the spherical law of cosine\footnote{$\cos(c)=\cos(a)\cos(b)+\sin(a)\sin(b)\cos(\gamma)$, where $a,b$ and $c$ are the side lengths (arc-lengths) of a spherical triangle, and $\gamma$ is the interior angle opposite to the side of length $c$.} we obtain
\begin{align*}
\cos\angle(w_1,w_2)
&= \cos(\ell^S)\cos(\ell^S) + \sin(\ell^S)\sin(\ell^S)\cos(\beta(\sigma,v)) \\
&= \cos^2(\ell^S) + \sin^2(\ell^S) (\cos(\beta(\sigma,v))-1+1) \\
&= [\cos^2(\ell^S) + \sin^2(\ell^S)] + \sin^2(\ell^S) (\cos(\beta(\sigma,v))-1) \\
&= 1-\sin^2(\ell^s)\cdot\chi(\beta(\sigma,v))\\
\implies\quad \sin^2(\ell^S)\cdot \chi(\beta(\sigma,v)) &= \chi(\angle(w_1,w_2)) \overset{(*)}= \Big(\frac{\ell}{r_2}\Big)^2\cdot \chi(\alpha(\sigma,v)) .
\end{align*}
\end{proof}
\end{proposition}
\iffalse
\hrulefill
\newpage
For each pair $(\sigma, v)$, where $\sigma\in\F_2(P)$ is a face of $P$ and $v\in\F_0(\sigma)$ is an incident vertex, there are two relevant notions of angle:
\begin{myenumerate}
\item
the \emph{interior angle} $\alpha(\sigma,v)$ of the face $\sigma$ at $v$.
If $\sigma$ is of type $2k$ and is incident to vertices $v_1,...,v_s\in\F_0(\sigma)$, then by the formula for the total interior angle of a $2k$-gon holds
$$\alpha(\sigma,v_1)+\cdots+\alpha(\sigma, v_s)=2(k-1)\pi.$$
On the other hand, if $v$ is incident to faces $\sigma_1,...,\sigma_s\in\F_2(P)$, then from the convexity of $P$ follows
$$\alpha(\sigma_1,v)+\cdots + \alpha(\sigma_s,v)<2\pi.$$
\item the \emph{dihedral angle} $\beta(\sigma, v)$ of the pyramid $\Pi(\sigma):=\conv\{\sigma\cup\{0\})$ at the edge $\conv\{0,v\}$:
\begin{center}
\includegraphics[width=0.65\textwidth]{img/pyramid_over_face}
\end{center}
We shall see that $\beta(\sigma,v)>\alpha(\sigma,v)$.
For a $2k$-face $\sigma$ incident to vertices $v_1,...,v_s\in\F_0(\sigma)$ we therefore find
$$\beta(\sigma,v_1)+\cdots+\beta(\sigma, v_s)>2(k-1)\pi.$$
If $P$ contains the origin (which it always does, see \cref{res:0_in_P}, then $P$ decomposes as the union of the pyramids $\Pi(\sigma),\sigma\in\F_2(P)$ with disjoint interior.
If the vertex $v$ is incident to faces $\sigma_1,...,\sigma_v\in\F_2(P)$, then the pyramids $\Pi(\sigma_i)$ completely surround the common edge $\conv\{0,v\}$ and hence
$$\beta(\sigma_1,v)+\cdots+\beta(\sigma_s,v)=2\pi.$$
\end{myenumerate}
\begin{proposition}
$0\in\mathrm{int}(P)$.
\end{proposition}
\begin{proposition}
Either of the angles $\alpha(\sigma,v)$ and $\beta(\sigma,v)$ determines the other one, and it holds $\alpha(\sigma,v)< \beta(\sigma,v)$.
\begin{proof}
W.l.o.g.\ given a 1-vertex $v\in V_1$ of a face $\sigma\in\F_2(P)$, and $w_1,w_2\in V_2$ the neighbors of $v$ in $\sigma$.
Consider the simplex $S:=\conv\{0,v,w_1,w_2\}$.
Except for the the edge $e:=\conv\{w_1,w_2\}$, the length of every other edge of $S$ is determined by the parameters of $P$ (these lengths work out to be exactly $r_1$, $r_2$ and $\ell$).
Note that the interior angle of the face $\conv\{v,w_1,w_2\}$ at $v$ is $\alpha(\sigma,v)$, and the dihedral angle of $S$ at $\conv\{0,v\}$ is $\beta(\sigma,v)$.
The shape of $S$ is completely determined by either
\begin{myenumerate}
\item knowing the length of all edges of $S$, or
\item knowing the length of all edges of $S$, except for the length of $e$, and instead knowing the dihedral angle at the edge opposite to $e$ (which is $\conv\{0,v\}$).
\end{myenumerate}
Knowing $\alpha(\sigma,v)$ immediately determines the length of $e$ via
\begin{align*}
|e|^2
=\|w_1-w_2\|^2
&=\|w_1-v\|^2+\|w_2-v\|^2-2\<w_1-v,w_2-v\>
\\&= 2\ell^2(1-\cos\alpha(\sigma,v)).
\end{align*}
By $(i)$, this determines the shape of $S$, in particular, the value of $\beta(\sigma,v)$.
Conversely, knowing $\beta(\sigma,v)$ determines the shape of $S$ via $(ii)$, thus, the value of $\alpha(\sigma,v)$.
Let $u_i$ be the orthogonal projection of $w_i-v$ onto the plane $v^\bot$.
One checks that the length of $u_i$ is determined by the parameters of $P$ (following the example of \cref{res:angles_between_vertices}), in particular, $\|u_1\|=\|u_2\|$.
Furthermore, it holds $\<u_i,v\>=0$ and $w_i-v=u_i+h v$ for some value $h\not=0$.
Set
\begin{align*}
f(x):=\cos\angle(u_1+x v,u_2+x v) = \frac{\<u_1+x v,u_2+xv\>}{\|u_1+ xv\|\cdot\|u_2 +xv\|} = \frac{\<u_1,u_2\>+x^2}{\|u_1\|\cdot\|u_2\|+x^2}.
\end{align*}
One checks that $f(x)$ is minimized only for $x=0$.
By construction of the $u_i$, we have that $f(0)=\cos\beta(\sigma,v)$ and $f(h)=\cos\alpha(\sigma,v)$.
It follows $\cos \alpha(\sigma,v)=f(h)>f(0)=\cos\beta(\sigma,v)$, and hence $\alpha(\sigma,v)<\beta(\sigma,v)$.
\end{proof}
\end{proposition}
\begin{proposition}
There are angles $\alpha_i^k,\beta_i^k\in(0,\pi)$ so that
$$\alpha(\sigma,v)=\alpha_i^k,\qquad \beta(\sigma,v)=\beta_i^k,$$
whenever $\sigma\in\F_2(P)$ is a $2k$-face and $v\in\F_0(\sigma)$ is an $i$-vertex.
\end{proposition}
\begin{proposition}
It holds
\begin{myenumerate}
\item given $i$-vertices $v_1,v_2\in V_i$ contained in faces $\sigma_1,\sigma_2\in \F_2(P)$ respectively. If $\alpha(\sigma_1,v_1)=\alpha(\sigma_2,v_2)$, then $\sigma_1$ and $\sigma_2$ are of the same type.
\item $\alpha(\sigma,v) < \beta(\sigma,v)$.
\item Either of $\alpha(\sigma,v)$ and $\beta(\sigma,v)$ uniquely determines the other one.
\end{myenumerate}
\begin{proof}
Suppose $v$ is a 1-vertex (the argument is analogue for a 2-vertex).
For a pair $(\sigma,v)$ let $w_1,w_2\in V_2$ denote the neighbors of $v$ in $\sigma$ and define the simplex $S:=\conv\{0,v,w_1,w_2\}$.
The shape of a simplex is uniquely determined by knowing the lengths of all its edges.
Alternatively, for a fixed edge $e\in\F_1(S)$, instead of knowing its length, it is also sufficient to know the dihedral angle of $S$ at the opposite edge (the only edge not incident to $e$).
\end{proof}
\end{proposition}
$$\angle(X,Y):=\max_{x\in X}\min_{y\in Y}\angle(x,y).$$
\fi
\section{Strictly bipartite polyhedra}
\label{sec:bipartite_polyhedra}
In this section we derive the classification of strictly bipartite polyhedra.
The main goal is to show that there are only two: the rhombic dodecahedron and the rhombic triacontahedron.
From this section on, let $P\subset\RR^3$ denote a fixed \emph{strictly bipartite polyhedron} with radii $r_1< r_2$ and edge length $\ell$.
The 2-faces of $P$ will be shortly referred to as just \emph{faces} of $P$.
Since they are bipartite, they are necessarily $2k$-gons.
\begin{definition}
We use the following terminology:
\begin{myenumerate}
\item a face of $P$ is of \emph{type} $2k$ (or called a \emph{$2k$-face}) if it is a $2k$-gonal polygon.
\item an edge of $P$ is of \emph{type} $(2k_1,2k_2)$ (or called a \emph{$(2k_1,2k_2)$-edge}) if the two incident faces are of type $2k_1$ and $2k_2$ respectively.
\item a vertex of $P$ is of \emph{type} $(2k_1,...,2k_s)$ (or called a \emph{$(2k_1,...,2k_s)$-vertex}) if its incident faces can be enumerated as $\sigma_1,...,\sigma_s$ so that $\sigma_i$ is a $2k_i$-face (note, the order of the numbers does not matter).
\end{myenumerate}
We write $\tau(v)$ for the type of a vertex $v\in\F_0(P)$.
\end{definition}
\subsection{General observations}
In a given bipartite polyhedron, the type of a vertex, edge or face already determines much of its metric properties.
We prove this for faces:
\begin{proposition}\label{res:unique_shape}
For some face $\sigma\in\F_2(P)$, any of the following properties of $\sigma$ determines the other two:
\begin{myenumerate}
\item its type $2k$,
\item its interior angles $\alpha_1>\alpha_2$.
\item its height $h$ (that is, the distance of $\aff(\sigma)$ from the origin).
\end{myenumerate}
\end{proposition}
\begin{corollary}
Any two faces of $P$ of the same height, or the same type, or the same interior angles, are congruent.
\end{corollary}
\begin{proof}[Proof of \cref{res:unique_shape}]
Fix a face $\sigma\in\F_2(P)$.
Suppose that the height $h$ of $\sigma$ is known.
By \cref{res:faces_of_bipartite}, a face of $P$ of height $h$ is bipartite with radii $\rho_i^2:=r_i^2-h^2$ and edge length $\ell$.
By \cref{res:2k_gons}, these parameters then uniquely determine the shape of $\sigma$, which includes its type and its interior angles. This shows $(iii)\implies(i),(ii)$.
\iffalse
Now suppose we know the interior angles $\alpha_1>\alpha_2$ of $\sigma$ (actually, it suffices to know one of these, say $\alpha_1$).
Fix a 1-vertex $v\in V_1$ of $\sigma$ and let $w_1,w_2\in V_2$ be its two adjacent 2-vertices in $\sigma$.
Then $\angle(w_1-v,w_2-v)=\alpha_1$.
All neighbors of $v$ lie on the circle $C:=\mathbf S_\ell(v)\cap\mathbf S_{r_2}(0)$, the symmetry axis of which is $\Span\{v\}$ (consider the figure below).
Thus, there is a unique way (up to orthogonal transformation) to place $w_1$ and $w_2$ on $C$ so that $\angle(w_1-v,w_2-v)=\alpha_1$.
But this uniquely determines the distance of $\aff(\sigma)=\aff\{v,w_1,w_2\}$ from the origin (the height of $\sigma$).
This shows $(ii)\implies(iii)$.
\begin{center}
\includegraphics[width=0.45\textwidth]{img/intersecting_spheres}
\end{center}
\hrulefill
\fi
Suppose now that we know the interior angles $\alpha_1>\alpha_2$ of $\sigma$ (it actually suffices to know one of these, say $\alpha_1$).
Fix a 1-vertex $v\in V_1$ of $\sigma$ and let $w_1,w_2\in V_2$ be its two adjacent 2-vertices in $\sigma$.
Consider the simplex $S:=\conv\{0,v,w_1,w_2\}$.
The length of each edge of $S$ is already determined, either by the parameters alone, or by additionally using the known interior angles via
\begin{align*}
&\|w_1-w_2\|^2= \|w_1-v\|^2+\|w_2-v\|^2 -2\<w_1-v,w_2-v\> \\ &\phantom{\|w_1-w_2\|^2}= 2\ell^2(1-\cos\underbrace{\angle(w_1-v,w_2-v)}_{\alpha_1}).
\end{align*}
Thus, the shape of $S$ is determined.
In particular, this determines the height of the face $\conv\{v,w_1,w_2\}\subset S$ over the vertex $0\in S$.
Since $\aff\{v,w_1,w_2\}=\aff(\sigma)$, this determines the height of $\sigma$ in $P$. This proves $(ii)\implies(iii)$.
\iffalse
Finally, suppose the type $2k$ is known.
If this does not uniquely determine the height of a face, then w.l.o.g., there a is a $2k$-face $\sigma'\in\F_2(P)$ of some height $h'<h$ (where $h$ is the height of $\sigma$).
Let $\rho_1<\rho_2$ resp.\ $\rho_1'<\rho_2'$ denote the radii of $\sigma$ resp.\ $\sigma'$.
From $h>h'$ follows via \cref{res:faces_of_bipartite} $(iii)$ that $(*)\;\rho_i<\rho_i'$.
But with \cref{res:angles_between_vertices} we have
$$\cos\frac\pi k = \frac{\rho_1^2+\rho_2^2-\ell^2}2 \overset{(*)}< \frac{\rho_1'^2+\rho_2'^2-\ell^2}2 = \cos \frac\pi k.$$
This is a contradiction, and we have shown $(i)\implies(iii)$.
\fi
Finally, suppose that the type $2k$ is known.
We then want to show that the~height $h$ is uniquely determined.%
\footnote{The reader motivated to prove this himself should know the following: it is indeed possible to write down a polynomial in $h$ of degree four whose coefficients involve only $r_1, r_2$, $\ell$ and $\cos(\pi/k)$, and whose zeroes include all possible heights of any $2k$-face of $P$.
However, it turns out to be quite tricky to work out which zeroes correspond to feasible solutions.
For certain values of the coefficients there are multiple positive solutions for $h$, some of which correspond to \emph{non-convex} $2k$-faces.
There seems to be no easy way to tell them apart.}
For the sake of contradiction, suppose that the type $2k$ does \emph{not} uniquely determine the height of the face.
Then there is another $2k$-face $\sigma'\in\F_2(P)$ of some height $h'\not=h$.
W.l.o.g.\ assume $h>h'$.
Visualize both faces embedded in $\RR^2$, on top of each other and centered at the origin as shown in the figure below:
\begin{center}
\includegraphics[width=0.33\textwidth]{img/embedded_2_faces}
\end{center}
The vertices in both polygons are equally spaced by an angle of $\pi/k$ (\shortStyle{cf.}\ \cref{res:angles_between_vertices}) and we can therefore assume that the vertex $v_i$ of $\sigma$ (resp.\ $v'_i$ of $\sigma'$) is a positive multiple of $(\sin(i\pi/k),\cos(i\pi/k))\in\RR^2$ for $i\in\{1,...,2k\}$.
There are then factors $\delta_i\in\RR_+$ with $v_i'=\delta v_i$.
The norms of vectors $v_1$, $v_2$, $\delta_1 v_1$ and $\delta_2 v_2$ are the radii of the bipartite polygons $\sigma$ and $\sigma'$.
With \cref{res:faces_of_bipartite} $(iii)$ from $h>h'$ follows $\|v_1\| < \|\delta_1 v_1\|$ and~$\|v_2\|<\|\delta_2 v_2\|$, and thus, $(*)\;\delta_1,\delta_2>1$. W.l.o.g. assume $\delta_1\le\delta_2$.
Since both faces have edge length $\ell$, we have $\|v_1-v_2\| = \|\delta_1 v_1-\delta_2 v_2\| = \ell$.
Our goal is to derive the following contradiction:
$$\ell=\|v_1-v_2\| \overset{\mathclap{\smash{(*)}}}< \delta_1 \|v_1-v_2\| = \|\delta_1 v_1-\delta_1 v_2\| \,\overset{\mathclap{\smash{(**)}}}<\, \|\delta_1 v_1-\delta_2 v_2\|=\ell,$$
To prove $(**)$, consider the triangle $\Delta$ with vertices $\delta_1 v_1$, $\delta_2 v_2$ and $\delta_1 v_2$:
\begin{center}
\includegraphics[width=0.29\textwidth]{img/embedded_2_faces_angle_trick}
\end{center}
Since $\sigma$ is convex, the angle $\alpha$ is smaller than $90^\circ$. It follows that the interior angle of $\Delta$ at $\delta_1 v_2$ is obtuse (here we are using $\delta_1\le \delta_2$). Hence, by the sine law, the edge of $\Delta$ opposite to $\delta_1 v_2$ is the longest, which translates to $(**)$.
\iffalse
It remains to prove inequality $(**)$.
This inequality is trivially satisfied if $\delta_1=\delta_2$ and so we can assume $\delta_1<\delta_2$.
We now provide a chain of equivalence transformations of $(**)$ (note the use of $\delta_1-\delta_2<0$ in the third step to reverse the inequality):
\begin{align*}
\|\delta_1 v_1-\delta_1 v_2\|^2 &\le \|\delta_1 v_1-\delta_2 v_2\|^2 \\
\delta_1^2 \|v_2\|^2 - 2\delta_1^2\<v_1,v_2\> &\le \delta_2^2 \|v_2\|^2 -2\delta_1\delta_2\<v_1,v_2\> \\
(\delta_1^2 - \delta_2^2) \|v_2\|^2 &\le 2\delta_1(\delta_1-\delta_2)\<v_1,v_2\> \\
(\delta_1 + \delta_2) \|v_2\|^2 &\ge 2\delta_1\<v_1,v_2\> \\
\bar\delta \|v_2\|^2 &\ge \delta_1\<v_1,v_2\>,
\end{align*}
where $\bar\delta:=(\delta_1+\delta_2)/2$.
Since $\bar\delta >\delta_1$, it suffices to check $\|v_2\|^2\ge \<v_1,v_2\>$ in order to conclusively prove $(**)$.
Note that $\|v_2\|^2\ge \<v_1,v_2\>$ is equivalent to $\<v_2,v_2-v_1\>\ge 0$, which is equivalent to the statement that the angle $\alpha$ (see figure above) is at most $90^\circ$.
This is true since $\sigma$ is convex (the interior angle is $2\alpha\le 180^\circ$).
We therefore found a contradiction to the assumption that there are two non-congruent $2k$-faces and this proves $(i)\implies(ii),(iii)$.
\fi
\end{proof}
As a consequence of \cref{res:unique_shape}, the interior angles of a face of $P$ do only depend on the type of the face (and the parameters), and so we can introduce the notion of \emph{the} interior angle $\alpha_i^k\in(0,\pi)$ of a $2k$-face at an $i$-vertex.
Furthermore, set $\eps_k:=(\alpha_1^k-\alpha_2^k)/2\pi$.
By \cref{res:angles} we have $\eps_k> 0$ and
$$\alpha_1^k = \Big(1-\frac1k+\eps_k\Big)\pi,\qquad \alpha_2^k=\Big(1-\frac 1k-\eps_k\Big)\pi.$$
\begin{definition}
If $\tau=(2k_1,...,2k_s)$ is the type of a vertex, then define
%
$$K(\tau):=\sum_{i=1}^s \frac 1{k_i},\qquad E(\tau):=\sum_{i=1}^s \eps_{k_i}.$$
\end{definition}
Both quantities are strictly positive.
\begin{proposition}\label{res:K_E_ineq}
Let $v\in \F_0(P)$ be a vertex of type $\tau=(2k_1,...,2k_s)$.
\begin{myenumerate}
\item If $v\in V_1$, then $E(\tau)<K(\tau)-1$ and $s=3$.
\item If $v\in V_2$, then $E(\tau)>s-2-K(\tau)$.
\end{myenumerate}
\begin{proof}
Let $\sigma_1,...,\sigma_s\in\F_2(P)$ be the faces incident to $v$, so that $\sigma_j$ is a $2k_j$-face.
The interior angle of $\sigma_j$ at $v$ is $\alpha_i^{\smash{k_j}}$, and the sum of these must be smaller than $2\pi$.
In formulas
$$2\pi > \sum_{j=1}^s \alpha_i^{k_{\smash j}} = \sum_{j=1}^s \Big(1-\frac1{k_j}\pm\eps_{k_j} \Big)\pi = (s-K(\tau)\pm E(\tau))\pi,$$
where $\pm$ is the plus sign for $i=1$, and the minus sign for $i=2$.
Rearranging for $E(v)$ yields $(*)\;{\mp E(\tau)} > s - 2 - K(\tau)$.
If $i=2$, this proves $(ii)$.
If $i=1$, note that from the implication $k_j\ge 2\implies K(\tau)\le s/2$ follows
$$s \overset{(*)}< -E(\tau)+K(\tau)+2 \le 0+\frac s2+2 \quad\implies\quad s<4.$$
The minimum degree of a vertex in a polyhedron is at least three, hence $s=3$, and $(*)$ becomes $(i)$.
\end{proof}
\end{proposition}
This allows us to exclude all but a manageable list of types for 1-vertices.
Note that a vertex $v\in V_1$ has a type of some form $(2k_1,2k_2,2k_3)$.
\begin{corollary}\label{res:1_types}
For a 1-vertex $v\in V_1$ of type $\tau$ holds $K(\tau)>1+E(\tau)>1$. One checks that this leaves exactly the options in \cref{tab:1_types}.
\begin{table}[h!]
\centering
\begin{tabular}{l|l|l}
$\tau$ & $K(\tau)$ & $\Gamma$ \\
\hline
$\overset{\phantom.}(4,4,\phantom04)$ & $3/2$ & $I_1\oplus I_1\oplus I_1$ \\
$(4,4,\phantom06)$ & $4/3$ & $I_1\oplus I_2(3)$ \\
$(4,4,\phantom08)$ & $5/4$ & $I_1\oplus I_2(4)$ \\
[-0.4ex]
$\quad\;\;\vdots\;\;$ & $\;\;\vdots$ & $\quad\;\vdots$ \\[0.3ex]
$(4,4,2k)$ & $1+1/k$ & $I_1\oplus I_2(k)$ \\[0.5ex]
\hline
$\overset{\phantom.}(4,6,\phantom06)$ & $7/6$ & $A_3=D_3$ \\
$(4,6,\phantom08)$ & $13/12$ & $B_3$ \\
$(4,6,10)$ & $31/30$ & $H_3$
\end{tabular}
\caption{Possible types of 1-vertices, their $K$-values and the $\Gamma$ of the $\Gamma$-permutahedron in which all vertices have this type.}
\label{tab:1_types}
\end{table}
\end{corollary}
The types in \cref{tab:1_types} are called the \emph{possible types} of 1-vertices.
Each of the possible types is realizable in the sense that there exists a bipartite polyhedron in which all 1-vertices have this type.
Examples are provided by the $\Gamma$-permutahedra (the $\Gamma$ of that $\Gamma$-permutahedron is listed in the right column of \cref{tab:1_types}).
These are not \emph{strictly} bipartite though.
The convenient thing about $\Gamma$-permutahedra is that all their vertices are of the same type.
We cannot assume this for general strictly bipartite polyhedra, not even for all 1-vertices.
\iffalse
If two faces $\sigma_1,\sigma_2\in\F_2(P)$ are congruent, then there is a \emph{congruence map} between them, that is, a map $T\colon \RR^3\to\RR^3$ that preserves Euclidean distance (\shortStyle{i.e.,}\ an isometry) and maps $T(\sigma_1)=\sigma_2$.
It will be of use later, that this congruence map can be chosen as an \emph{orthogonal} transformation $T\in\Ortho(\RR^3)$.
\begin{proposition}\label{res:orthogonal_map}
If $\sigma_1,\sigma_2\in\F_2(P)$ are congruent faces of $P$, then the congruence map $T\:\RR^3\to\RR^3$ between them can be chosen as an orthogonal transformation.
\begin{proof}
An isometry is an orthogonal transformation if and only if it fixes the origin.
If $v\in\F_1(\sigma_1)$ is an $i$-vertex of $P$, then $\|v\|=r_i$.
By \cref{res:faces_of_bipartite}, $v$ is also an $i$-vertex of $\sigma_1$.
\cref{res:2d_shape} then states that the interior angle of $\sigma_1$ at $v$ is larger (if $i=1$) resp.\ smaller (if $i=2$) than $ (1-1/k)\pi$, and by congruence, so must be the interior angle of $\sigma_2$ at $T(v)$.
Applying \cref{res:2d_shape} in the other direction, $T(v)$ must be a $i$-vertex of $\sigma_2$, and by \cref{res:faces_of_bipartite} also an $i$-vertex of $P$.
In particular, we obtain
\begin{align}
\|0-T(v)\| &=\|T(v)\|=r_i = \|v\|.\label{eq:distance_1}
\end{align}
The face $\sigma_2$ contains an affine basis of the plane $\aff(\sigma_2)$ and so the set of equations $\|x-T(v)\|=\|v\|$ for all $v\in\F_0(\sigma_2)$ has at most two solutions in $x\in\RR^3$.
Equation \eqref{eq:distance_1} states that the origin is one such solution.
The other solution must be the orthogonal reflection of the origin on the plans $\aff(\sigma_2)$.
By
$$\|T(0)-T(v)\|=\|0-v\|=\|v\|,\qquad\text{for all $v\in\F_1(\sigma-2)$},$$
$T(0)$ is also a solution to this system of equations. If $T(0)=0$, we would be done.
So suppose that $T(0)$ is the reflection of the origin on the plane $\aff(\sigma_2)$.
Thus, if we let $R\:\RR^3\to\RR^3$ denote this reflection on $\aff(\sigma_2)$, then $R$ is an isometry that fixes $\aff(\sigma_2)$ and maps $R(T(0))=0$.
Then $R\circ T$ is the desired congruence map that fixes the origin.
\end{proof}
\end{proposition}
\fi
\subsection{Spherical polyhedra}
The purpose of this section is to define a second notion of interior angle for each face.
These angles can be defined in several equivalent~ways, one of which is via spherical polyhedra.
A \emph{spherical polyhedron} is an embedding of a planar graph into the unit sphere, so that all edges are embedded as great circle arcs, and all regions are convex\footnote{Convexity on the sphere means that the shortest great circle arc connecting any two points in the region is also contained in the region.}.
If $0\in\mathrm{int}(P)$, we can associate to $P$ a spherical polyhedron $P^S$ by applying central projection
$$\RR^3\setminus\{0\}\to\Sph_1(0),\quad x\mapsto\frac x{\|x\|}$$
to all its vertices and edges (this process is visualized below).
\begin{center}
\includegraphics[width=0.42\textwidth]{img/spherical_polyhedron}
\end{center}
The vertices, edges and faces of $P$ have spherical counterparts in $P^S$ obtained as projections onto the unit sphere.
Those will be denoted with a superscript \enquote{$S$\,}.
For example, if $e\in\F_1(P)$ is an edge of $P$, then $e^{\smash S}$ denotes the corresponding \enquote{spherical edge}, which is a great circle arc obtained as the projection of $e$ onto the sphere.
We still need to justify that the spherical polyhedron of $P$ is well-defined, by proving that $P$ contains the origin:
\begin{proposition}\label{res:0_in_P}
$0\in\mathrm{int}(P)$.
\begin{proof}
The proof proceeds in several steps.
\textit{Step 1}:
Fix a 1-vertex $v\in V_1$ with neighbors $w_1,w_2,w_3\in V_2$, and let $u_i:=w_i-v$ be the direction of the edge $\conv\{v,w_i\}$ emanating from $v$.
Let $\sigma_{ij}\in\F_2(P)$ denote the $2k$-face containing $v,w_i$ and $w_j$.
The interior angle of $\sigma_{ij}$ at $v$ is then $\angle(u_i,u_j)$, which by \cref{res:angles} and $k\ge 2$ satisfies
$$\angle(u_i,u_j)>\Big(1-\frac1k\Big)\pi \ge \frac\pi2\quad\implies\quad \<u_i,u_j\><0.$$
\textit{Step 2}:
Besides $v$, the polyhedron $P$ contains another 1-vertex $v'\in V_1$.
It then holds $v'\in v+\cone\{u_1,u_2,u_3\}$, which means that there are non-negative coefficients $a_1,a_2,a_3\ge 0$, at least one positive, so that $v+a_1 u_1+ a_2u_2 + a_3u_3 = v'$.
Rearranging and applying $\<v,\free\>$ yields
\begin{align*}
a_1\<v,u_1\>+a_2\<v,u_2\>+a_3\<v,u_3\>
&=\<v,v'\>-\<v,v\>
\tag{$*$}
\\&= r_1^2\cos\angle(v,v')-r_1^2 < 0.
\tag*{}
\end{align*}
The value $\<v,u_i\>$ is independent of $i$ (see \cref{res:angles_between_vertices}).
Since there is at least one positive coefficient $a_i$, from $(*)$ follows $\<v,u_i\><0$.\footnote{Note that this provides the formal proof mentioned in \cref{rem:alternative_definiton}, namely, that the triangle $\conv\{0,v_1,v_2\}$ is acute at $v_1$ and $v_2$.}
\textit{Step 3}:
By the previous steps, $\{v,u_1,u_2,u_3\}$ is a set of four vectors with pair-wise negative inner product.
The convex hull of such an arrangement in 3-dimensional Euclidean space does necessarily contain the origin in its interior, or equivalently, there are positive coefficients $a_0,...,a_3>0$ with $a_0 v+a_1 u_1+a_2 u_2+a_3 u_3=0$ (for a proof, see \cref{res:sum_to_zero}).
In other words: $0\in v+\mathrm{int}(\cone\{u_1,u_2,u_3\})$.
\textit{Step 4}:
If $H(\sigma)$ denotes the half-space associated with the face $\sigma\in\F_2(P)$, then
$$0\in v+\mathrm{int}(\cone\{u_1,u_2,u_3\}) = \bigcap_{\sigma\sim v} \mathrm{int}(H(\sigma)).$$
Thus, $0\in\mathrm{int}(H(\sigma))$ for all faces $\sigma$ incident to $v$.
But since every face is incident to a 1-vertex, we obtain $0\in\mathrm{int}(H(\sigma))$ for all $\sigma\in\F_2(P)$, and thus $0\in\mathrm{int}(P)$ as well.
\end{proof}
\end{proposition}
\iffalse
The next goal is to show, that like the usual interior angles $\alpha_i^k$ of a face $\sigma$, also~the spherical interior angles of $\sigma^S$ depend only on the type of the face and the class of the vertex (and the parameters).
We then denote by $\beta_i^k$
the spherical interior angle of a $2k$-face at an $i$-vertex.
\begin{proposition}\label{res:spherical_interior_angles}
The notion $\beta_i^k$ is well-defined and it holds
\begin{myenumerate}
\item $\beta_i^{k_1}=\beta_i^{k_2}$ if and only if $k_1=k_2$,
\item if $P$ has an $i$-vertex of type $(2k_1,...,2k_s)$, then
%
$\beta_i^{k_1}+ \cdots + \beta_i^{k_s}=2\pi,$
\end{myenumerate}
\begin{proof}
\iffalse
Consider a 1-vertex $v\in V_1$ (the argument is analogous for a 2-vertex) with neighbors $w_1,w_2\in V_2$, all contained in a common $2k$-face $\sigma\in \F_2(P)$.
We show that the interior angle $\alpha_1^k$ of the triangle $w_1 v w_2$ at $v$ uniquely determines the (spherical) interior angle $\beta_1^k$ of the spherical triangle $w_1^S v^S w_2^S$ at $v^S$, and vice versa.
Set $u_i:=w_i-v\in\RR^3$, so that $\alpha_i^k=\angle(u_1,u_2)$ and $\|u_i\|=\ell$.
Further, define the bijection $\chi\:(0,\pi)\to(0,2),\theta\mapsto 1-\cos(\theta)$.
We find that
\begin{align*}
(*)\quad 2\ell^2\cdot\chi(\alpha_1^k)
&= \ell^2+\ell^2-2\ell^2\cos(\angle(u_1,u_2)) \\
&= \|u_1\|^2+\|u_2\|^2 - 2\<u_1,u_2\> \\
&= \|u_1-u_2\|^2
= \|w_1-w_2\|^2 \\
&= \|w_1\|^2+\|w_2\|^2-2\<w_1,w_2\>\\
&= r_2^2+r_2^2-r_2^2\cos\angle(w_1,w_2)
= 2r_2^2\cdot\chi(\angle(w_1,w_2)).
\end{align*}
The side lengths of the spherical triangle $w_1^Sv^Sw_2^S$ are $\angle(w_1,w_2)$ for the side opposite to $\beta_1^k$, and $\ell^S$ (uniquely determined by the parameters, see \cref{res:spherical_geometry_facts} $(i)$) for the other two sides.
By the spherical law of cosine (\cref{res:spherical_geometry_facts} $(iii)$), we obtain
\begin{align*}
\cos\angle(w_1,w_2)
&= \cos(\ell^S)\cos(\ell^S) + \sin(\ell^S)\sin(\ell^S)\cos(\beta_1^k) \\
&= \cos^2(\ell^S) + \sin^2(\ell^S) (\cos(\beta_1^k)-1+1) \\
&= [\cos^2(\ell^S) + \sin^2(\ell^S)] + \sin^2(\ell^S) (\cos(\beta_1^k)-1) \\
&= 1-\sin^2(\ell^s)\cdot\chi(\beta_1^k)\\
\implies\quad \sin^2(\ell^S)\cdot \chi(\beta_1^k) &= \chi(\angle(w_1,w_2)) \overset{(*)}= \frac{\ell^2}{r_2^2}\cdot \chi(\alpha_1^k) .
\end{align*}
We found a direct relation between $\beta_1^k$ and $\alpha_1^k$, each one determining the other one uniquely.
The argument is equivalent for $\beta_2^k$.
In detail, since $\chi$ is a bijection, and all other factors are non-zero, we obtain part $(i)$ via
$$\beta_i^{k_1}=\beta_i^{k_2} \;\Longleftrightarrow\; \alpha_i^{k_1}=\alpha_i^{k_2} \;\Longleftrightarrow\; k_1=k_2.$$
Part $(ii)$ is now an immediate consequence of \cref{res:spherical_geometry_facts} $(ii)$.
\else
Consider a 1-vertex $v\in V_1$ (analogous for a 2-vertex) with neighbors $w_1,w_2\in V_2$, all contained in a common face $\sigma\in \F_2(P)$.
Let $\Delta:=\conv\{v,w_1,w_2\}$ be a triangle, and $\Delta^S$ its spherical projection.
Suppose that $\sigma$ is a $2k$-face, and hence $\angle(w_1-v,w_2-v)=\alpha_1^k$ is determined by the parameters.
The length of each edge of the simplex $S:=\conv\{0,v,w_1,w_2\}$ is then also determined by the parameters.
We repeat this only for the edge $\conv\{w_1,w_2\}$:
$$\|w_1-w_2\|^2=\|w_1-v\|^2+\|w_2-v\|^2 - 2\<w_1-v,w_2-v\> = 2\ell^2(1-\cos(\alpha_1^k)).$$
Thus, the shape of $S$ is uniquely determined, and since $0$ is a vertex of $S$, the shape of the cone $C:=\cone(S)$ is also uniquely determined.
This then further determines the shape of the spherical triangle $\Delta^S=C\cap\Sph_1(0)$ including its spherical interior angle, say $\beta_1^k$, at $v^S$.
This shows that $\beta_1^k$ indeed only depends on the type $2k$ (and the parameters), and thus, that this notion is well-defined.
Conversely, assume that the spherical interior angle of $\Delta^S$ at $v^S$ is $\beta_1^k$.
We~also know that two spherical edges incident to $v^S$ have length $\ell^S$ (see \cref{res:spherical_geometry_facts} $(i)$).
This determines the shape of $\Delta^S$, in particular, the placement of $w_1$ and $w_2$ (up to some orthogonal transformation).
Further, this determines the shape of the triangle $\Delta=\conv\{r_1 v^S, r_2 w_1^S, r_2 w_2^S\}$, in particular its interior angle $\alpha_1^k$ at $v$, and thus also $k$ by \cref{res:unique_shape}. This proves $(i)$.
Part $(ii)$ is now an immediate consequence of \cref{res:spherical_geometry_facts} $(ii)$.
\fi
\end{proof}
\end{proposition}
\fi
The main reason for introducing spherical polyhedra is that we can talk about the \emph{spherical interior angles} of their faces.
Let $\sigma\in \F_2(P)$ be a face, and $v\in\F_0(\sigma)$ one of its vertices.
Let $\alpha(\sigma,v)$ denote the interior angle of $\sigma$ at $v$, and $\beta(\sigma, v)$ the spherical interior angle of $\sigma^S$ at $v^S$.
It only needs a straight-forward computation (involving some spherical geometry) to establish a direct relation between these angles: \shortStyle{e.g.}\ if $v$ is a 1-vertex, then
$$\sin^2(\ell^S)\cdot (1-\cos\beta(\sigma,v)) = \Big(\frac{\ell}{r_2}\Big)^2\!\!\cdot (1-\cos\alpha(\sigma,v)),$$
where $\ell^S$ denotes the arc-length of an edge of $P^S$ (indeed, all edges are of the same length).
An equivalent formula exists for 2-vertices.
The details of the computation are not of relevance, but can be found in \cref{sec:computations}.
The core message is that the value of $\alpha(\sigma,v)$ uniquely determines the value of $\beta(\sigma,v)$ and vice versa.
In particular, since the value of $\alpha(\sigma,v)=\alpha_i^{\smash k}$ does only depend on the type of the face and the partition class of the vertex, so does $\beta(\sigma,v)$, and it makes sense to introduce the notion $\beta_i^k$ for the spherical interior angle of a $2k$-gonal spherical face of $P^S$ at (the projection of) an $i$-vertex.
Thus, we have
\begin{equation}\label{eq:angle_relations}
\beta_i^{k_1}=\beta_i^{k_2}\quad\Longleftrightarrow\quad\alpha_i^{k_1}=\alpha_i^{k_2} \quad\overset{\ref{res:unique_shape}}\Longleftrightarrow\quad k_1=k_2,
\end{equation}
where we use \cref{res:unique_shape} for the last equivalence.
\begin{observation}\label{res:spherical_interior_angles_are_good}
The spherical interior angles $\beta_i^k$ have the following properties:
\begin{myenumerate}
\item The spherical interior angles surrounding a vertex add up to exactly $2\pi$. That is, for an $i$-vertex $v\in\F_0(P)$ of type $(2k_1,...,2k_s)$ holds
%
$$\beta_i^{k_1}+\cdots+\beta_i^{k_s} = 2\pi.$$
\item
The sum of interior angles of a spherical polygon always exceed the interior angle sum of a respective flat polygon. That is, it holds
%
$$k \beta_1^k+k\beta_2^k > 2(k-1)\pi\quad\Longrightarrow\quad \beta_1^k+\beta_2^k > 2\Big(1-\frac1k\Big)\pi.$$
\end{myenumerate}
\end{observation}
\iffalse
\begin{observation}\label{res:spherical_geometry_facts}
We note two relevant facts:
\begin{myenumerate}
\item
The (arc-)length $\ell^S$ of a (spherical) edge of $P^S$ is exactly the angle between its end vertices.
We determined in \cref{res:angles_between_vertices}
%
$$\ell^S = \arccos\Big(\frac{r_1^2+r_2^2-\ell^2}{2r_1 r_2}\Big),$$
%
in particular, $P^S$ has all edges of the same length.
\item
A face $\sigma^S$ of $P^S$ is a convex spherical polygon with well-defined \emph{(spherical) interior angles}.
The sum of interior angles of a spherical polygon is larger than the one of a respective flat polygon, and depends on its area.
However, the sum of the interior angles surrounding a vertex is exactly $2\pi$.
This is the central property we apply later in many circumstances.
\end{myenumerate}
\end{observation}
\fi
This has some consequences for the strictly bipartite polyhedron $P$:
\begin{corollary}\label{res:incompatible_1_types}
$P$ contains at most two different types of 1-vertices, and if there are two, then one is of the form $(4,4,2k)$, and the other one is of the form $(4,6,2k')$ for distinct $k\not=k'$ and $2k'\in\{6,8,10\}$.
\begin{proof}
Each possible type listed in \cref{tab:1_types} is either of the form $(4,4,2k)$ or of the form $(4,6,2k')$ for some $2k\ge 4$ or $2k'\in\{6,8,10\}$.
If $P$ contains simultaneously 1-vertices of type $(4,4,2k_1)$ and $(4,4,2k_2)$, apply \cref{res:spherical_interior_angles_are_good} $(i)$ to see
$$
\beta_1^2+\beta_1^2+\beta_1^{k_1}
\overset{(i)}=\beta_1^2+\beta_1^2+\beta_1^{k_2}
\; \implies \;
\beta_1^{k_1}=\beta_1^{k_2}
\; \overset{\eqref{eq:angle_relations}}\implies \;
k_1=k_2.$$
If $P$ contains simultaneously 1-vertices of type $(4,6,2k_1')$ and $(4,6,2k_2')$, then
$$\beta_1^2+\beta_1^3+\beta_1^{k_1'}
\overset{(i)}=\beta_1^2+\beta_1^3+\beta_1^{k_2'}
\; \implies \;
\beta_1^{k_1'}=\beta_1^{k_2'}
\; \overset{\eqref{eq:angle_relations}}\implies \;
k_1'=k_2'.$$
Finally, if $P$ contains simultaneously 1-vertices of type $(4,4,2k)$ and $(4,6,2k')$, then
$$\beta_1^2+\beta_1^2+\beta_1^{k}\overset{(i)}=\beta_1^2+\beta_1^3+\beta_1^{k'} \; \implies \; \beta_1^k-\beta_1^{k'} = \underbrace{\beta_1^3-\beta_1^2}_{\text{$\not=0$ by \eqref{eq:angle_relations}}} \;\overset{\eqref{eq:angle_relations}}\implies \; k\not=k'.$$
\end{proof}
\end{corollary}
Since each edge of $P$ is incident to a 1-vertex, we obtain
\begin{observation}\label{res:edge_types}
If $P$ has only 1-vertices of types $(4,4,2k)$ and $(4,6,2k')$, then each edge of $P$ is of one of the types
$$\underbrace{(4,4), \; (4,2k)}_{\mathclap{\text{\textup{from a $(4,4,2k)$-vertex}}}}, \; \underbrace{(4,6), \; (4,2k') \quad\text{or}\quad (6,2k')}_{\text{\textup{from a $(4,6,2k')$-vertex}}}.$$
\end{observation}
\begin{corollary}\label{res:dihedral_angles}
The dihedral angle of an edge $e\in\F_1(P)$ of $P$ only depends on its type.
\begin{proof}
Suppose that $e$ is a $(2k_1,2k_2)$-edge.
Then $e$ is incident to a 1-vertex $v\in V_1$ of type $(2k_1,2k_2, 2k_3)$.
By \cref{res:spherical_interior_angles_are_good} $(i)$ holds $\beta_1^{\smash{k_3}}=2\pi-\beta_1^{\smash{k_1}}-\beta_1^{\smash{k_2}}$, which further determines $k_3$.
By \cref{res:unique_shape} we have uniquely determined interior angles $\alpha_1^{\smash{k_1}},\alpha_1^{\smash{k_2}}$ and $\alpha_1^{\smash{k_3}}$.
It is known that for a simple vertex (that is, a vertex of degree three) the interior angles of the incident faces already determine the dihedral angles at the incident edges (for a proof, see the Appendix, \cref{res:simple_vertex}).
Consequently, the dihedral angle at $e$ is already determined.
\end{proof}
\end{corollary}
The next result shows that $\Gamma$-permutahedra are the only bipartite polytopes in which a 1-vertex and a 2-vertex can have the same type.
\begin{corollary}\label{res:1_2_vertex_same_type}
$P$ cannot contain a 1-vertex and a 2-vertex of the same type.
\begin{proof}
Let $v\in\F_0(P)$ be a vertex of type $(2k_1,2k_2,2k_3)$.
The incident edges are of type $(2k_1,2k_2)$, $(2k_2,2k_3)$ and $(2k_3,2k_1)$ respectively.
By \cref{res:dihedral_angles} the dihedral angles of these edges are uniquely determined, and since $v$ is simple (that is, has degree three), the interior angles of the incident faces are also uniquely determined (\shortStyle{cf.}\ Appendix, \cref{res:simple_vertex}).
In particular, we obtain the same angles independent of whether $v$ is a 1-vertex or a 2-vertex.
A 1-vertex is always simple, and thus, a 1-vertex and a 2-vertex of the same type would have the same interior angles at all incident faces, that is, $\alpha_1^k=\alpha_2^k$ for each incident $2k$-face.
But this is not possible if $P$ is \emph{strictly} bipartite (by \cref{res:faces_of_bipartite} $(ii)$ and \cref{res:angles}).
\end{proof}
\end{corollary}
\iffalse
\newpage
\hrulefill
The \emph{dihedral angle} of an edge $e\in\F_1(P)$ is the angle between its incident faces.
In a bipartite polyhedron, this angle is determined by the type of $e$.
\begin{proposition}\label{res:dihedral_angle}
The dihedral angle of an $e\in\F_1(P)$ is uniquely determined by the type of the edge.
\begin{proof}
Suppose $e$ is of type $(2k_1,2k_2)$.
Let $\sigma_1,\sigma_2\in\F_2(P)$ be the faces incident to $e$, so that $\sigma_i$ is of type $2k_i$.
By \cref{res:unique_shape}, the type $2k_i$ uniquely determines the height $h_i$ of $\sigma_i$.
There are exactly two planes through $E$ of distance $h_i$ from the origin.
\msays{TODO} \msays{: Why does $P$ contain the origin?}
The affine span $\aff(\sigma_i)$ is the uniquely determined plane of distance $h_i$ from the origin that contains the line $\aff(e)$.
The dihedral angle at $e$ is the angle between these planes.
\end{proof}
\end{proposition}
\fi
\subsection{Adjacent pairs}
Given a 1-vertex $v\in V_1$ of type $\tau_1=(2k_1,2k_2,2k_3)$, for any two distinct $i,j\in\{1,2,3\}$, $v$ has a neighbor $w\in V_2$ of type $\tau_2=(2k_i,2k_j,*,...,*)$, where $*$ are placeholders for unknown entries.
The pair of types
$$(\tau_1,\tau_2)=((2k_1,2k_2,2k_3),\,(2k_i,2k_j,*,...,*))$$
is called an \emph{adjacent pair} of $P$.
It is the purpose of this section to show that certain adjacent pairs cannot occur in $P$.
Excluding enough adjacent pairs for fixed $\tau_1$ then proves that the type $\tau_1$ cannot occur as the type of a 1-vertex.
Our main tools for achieving this will be the inequalities established in \cref{res:K_E_ineq} $(i)$ and $(ii)$, that is
$$E(\tau_1)\overset{\mathclap{(i)}}<K(\tau_1)-1\quad\text{and}\quad E(\tau_2)\overset{\mathclap{(ii)}}>s-2-K(\tau_2),$$
where $s$ is the number of elements in $\tau_2$.
For a warmup, and as a template for~fur\-ther calculations, we prove that the adjacent pair $(\tau_1,\tau_2)=((4,6,8),(6,8,8))$ will not occur in $P$.
\begin{example}\label{ex:infeasible}
By \cref{res:K_E_ineq} $(i)$ we have
$$(*)\quad \eps_2+\eps_3+\eps_4=E(\tau_1)\overset{(i)}<K(\tau_1)-1 = \frac12+\frac13+\frac14-1=\frac1{12}.$$
On the other hand, by \cref{res:K_E_ineq} $(ii)$ we have
\begin{align*}
(**)\quad \frac2{12} = 3-2-\Big(\frac13+\frac14+\frac14\Big)&=s-2-K(\tau_2) \\[-1ex]
&\overset{(ii)}< E(\tau_2) = \underbrace{\eps_3+\eps_4}_{< 1/12}+\underbrace{\eps_4}_{\mathclap{<1/12}} < \frac2{12},
\end{align*}
which is a contradiction.
Hence this adjacent pair cannot occur.
Note that we used $(*)$ to upperbound certain sums of $\eps_i$ in $(**)$.
\end{example}
An adjacent pair excluded by using the inequalities from \cref{res:K_E_ineq} $(i)$ and $(ii)$ as demonstrated in \cref{ex:infeasible} will be called \emph{infeasible}.
The argument applied in \cref{ex:infeasible} will be repeated many times for many different adjacent pairs in the upcoming sections \cref{sec:468,sec:4610,sec:466,sec:442k}, and we shall therefore use a tabular form to abbreviate it.
After fixing, $\tau_1=(4,6,8)$, the argument to refute the adjacent pair $(\tau_1,\tau_2)=((4,6,8),(6,8,8))$ is abbreviated in the first row of the following table:
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ & $\overset?<$ & $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(6,8,8)$ & $2/12$ & $\not<$ & $(\eps_3+\eps_4)+\eps_4$ & $<$
$2/12$ \\
$(6,8,6,6)$ & $9/12$ & $\not<$ & $(\eps_3+\eps_4)+\eps_3+\eps_3$ & $<$ $3/12$
\end{tabular}
\end{center}
The second row displays the analogue argument for another example, namely, the pair $((4,6,8),(6,8,6,6))$, showing that it is infeasible as well.
Both rows will reappear in the table of \cref{sec:468} where we exclude $(4,6,8)$ as a type for 1-vertices entirely.
Note that the terms in the column below $E(\tau_2)$ are grouped by parenthesis to indicate which subsums are upper bounded via \cref{res:K_E_ineq} $(i)$.
In this example, if there are $n$ groups, then the sum is upper bounded by $n/12$.
The placeholders in an adjacent pair $((2k_1,2k_2,2k_3),(2k_i,2k_j,*,...,*))$ can, in theory, be replaced by an arbitrary sequence of even numbers, and each such pair has to be refuted separately.
The following fact will make this task tractable: write $\tau\subset \tau'$ if $\tau$ is a \emph{subtype} of $\tau'$, that is, a vertex type that can be obtained from $\tau'$ by removing some of its entries.
We then can prove
\begin{proposition}\label{res:subtype}
If $(\tau_1,\tau_2)$ is an infeasible adjacent pair, then the pair $(\tau_1,\tau_2')$ is infeasible as well, for every $\tau_2'\supset \tau_2$.
\begin{proof}
Suppose $\tau_2=(2k_1,...,2k_s)$, $\tau_2'=(2k_1,...,2k_s,2k_{s+1},...,2k_{s'})\supset\tau_2$, and that the pair $(\tau_1,\tau_2')$ is \emph{not} infeasible.
Then $\tau_2'$ satisfies \cref{res:K_E_ineq} $(ii)$
\begin{align*}
E(\tau_2') &> s'-2-K(\tau_2') \\[-5ex]
\implies\quad E(\tau_2) &> s - 2 - K(\tau_2) + \sum_{i=s+1}^{s'}\!\!\overbrace{\Big(1-\frac1{k_i}-\eps_{k_i}\Big)}^{\alpha_2^{k_i}/\pi > 0} > s - 2 - K(\tau_2).
\end{align*}
But this is exactly the statement that $\tau_2$ satisfies \cref{res:K_E_ineq} $(ii)$ as well, \shortStyle{i.e.,}\ that the pair $(\tau_1,\tau_2)$ is also not infeasible.
\end{proof}
\end{proposition}
By \cref{res:subtype} it is sufficient to exclude so-called \emph{minimal infeasible adjacent pairs}, that is, infeasible adjacent pairs $(\tau_1,\tau_2)$ for which $(\tau_1,\tau_2')$ is not infeasible for any $\tau_2'\subset\tau_2$.
A second potential problem is, that we know little about the values that might replace the placeholders in $\tau_2=(2k_i,2k_j,*,...,*)$.
For our immediate goal, dealing with the following special case is sufficient:
\begin{proposition}\label{res:exclusive_types}
The placeholders in an adjacent pair $((4,6,2k'),(6,2k',*,...,*))$ can only contain $4$, $6$ and $2k'$.
\begin{proof}
Suppose that $P$ contains an adjacent pair $$(\tau_1,\tau_2)=((4,6,2k'),(6,2k',2k,*,...,*))$$ induced by a 1-vertex $v\in V_1$ of type $\tau_1$ with neighbor $w\in V_2$ of type $\tau_2$.
Suppose further that $2k\not\in\{4,6,2k'\}$.
The vertex $w$ is then incident to a $2k$-face, and therefore also to a 1-vertex $u\in V_1$ of type $(4,4,2k)$ ($u$ cannot be of type $(4,6,2k)$ because of $k\not=k'$ and \cref{res:incompatible_1_types}).
This configuration is depicted below:
\begin{center}
\includegraphics[width=0.34\textwidth]{img/adjacent_pair_exclusion}
\end{center}
Note that $w$ is also incident to a $4$-face, and thus $(6,2k',2k,4)\subseteq \tau_2$.
By \cref{res:K_E_ineq} $(i)$ the existence of 1-vertices of type $(4,4,2k)$ and $(4,6,2k')$ yields inequalities
\begin{equation}
\label{eq:two_inequalities}
\eps_2+\eps_2+\eps_k<\frac1k \quad\text{and}\quad\eps_2+\eps_3+\eps_{k'}<\frac1{k'}-\frac16.
\end{equation}
Since $\tau_2$ has $\tau:=(6,2k',2k,4)$ as a subtype, by \cref{res:subtype} it suffices to show that the pair $((4,6,2k'),(6,2k',2k,4))$ is infeasible.
This follows via \cref{res:K_E_ineq} $(ii)$:
\begin{align*}
\frac76-\frac1k-\frac1{k'}
=\;&4-2-K(\tau)\\\overset{(ii)}< \;&E(\tau) = \underbrace{\eps_2+\eps_3+\eps_{k'}}_{<1/k'-1/6}+\underbrace{\eps_k}_{\mathclap{<1/k}} \!\!\overset{\eqref{eq:two_inequalities}}< \frac1k+\frac1{k'}-\frac16,
\end{align*}
which rearranges to $1/k+1/{k'}>2/3$.
Recalling $2k'\in\{6,8,10\}\implies k'\ge 3$ (from \cref{res:incompatible_1_types}) and $2k\not\in\{4,6,2k'\}\implies k\ge 4$ shows that this is not possible.
\end{proof}
\end{proposition}
\subsection{The case $\tau_1=(4,6,10)$}
\label{sec:4610}
If $P$ contains a 1-vertex of type $(4,6,10)$, then it contains an adjacent pair of the form
$$(\tau_1,\tau_2)=((4,6,10),(6,10,*,...,*)).$$
We proceed as demonstrated in \cref{ex:infeasible}.
\cref{res:K_E_ineq} $(i)$ yields $\eps_2+\eps_3+\eps_5 < 1/30$.
By \cref{res:exclusive_types} the placeholders can only take on values 4, 6 or $10$.
The following table lists the minimally infeasible adjacent pairs and proves their infeasibility.
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ &$\overset?<$& $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(6,10,6)$ & $\phantom04/30$ &$\not<$& $(\eps_3+\eps_5)+\eps_3$ & $< 2/30$ \\
$(6,10,10)$ & $\phantom08/30$ &$\not<$& $(\eps_3+\eps_5)+\eps_5$ & $< 2/30$ \\
$(6,10,4,4)$ & $14/30$ &$\not<$& $(\eps_2+\eps_3+\eps_5)+\eps_2$ & $< 2/30$ \\
\end{tabular}
\end{center}
By \cref{res:subtype} we conclude: the placeholder in $\tau_2=(6,10,*,...,*)$ can contain no 6 or 10, and at most one 4.
This leaves us with the option $\tau_2=(4,6,10)$, which is the same as $\tau_1$ and therefore not possible by \cref{res:1_2_vertex_same_type}.
Therefore, $P$ cannot contain a 1-vertex of type $(4,6,10)$.
\subsection{The case $\tau_1=(4,6,8)$}
\label{sec:468}
If $P$ contains a 1-vertex of type $(4,6,8)$, then it also contains an adjacent pair of the form
$$(\tau_1,\tau_2)=((4,6,8),(6,8,*,...,*)).$$
We proceed as demonstrated in \cref{ex:infeasible}.
\cref{res:K_E_ineq} $(i)$ yields $\eps_2+\eps_3+\eps_4 < 1/12$.
By \cref{res:exclusive_types} the placeholders can only take on values 4, 6 or 8.
The following table lists the minimally infeasible adjacent pairs and proves their infeasibility.
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ &$\overset?<$& $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(6,8,8)$ & $2/12$ & $\not<$ & $(\eps_3+\eps_4)+\eps_3$ & $< 2/12$ \\
$(6,8,4,4)$ & $5/12$ &$\not<$& $(\eps_2+\eps_3+\eps_4)+\eps_2$ & $< 2/12$ \\
$(6,8,4,6)$ & $7/12$ &$\not<$& $(\eps_2+\eps_3+\eps_4)+\eps_3$ & $< 2/12$ \\
$(6,8,6,6)$ & $9/12$ &$\not<$& $(\eps_2+\eps_3+\eps_4)+\eps_3 + \eps_3$ & $< 3/12$
\end{tabular}
\end{center}
By \cref{res:subtype} we conclude: the placeholder in $\tau_2=(6,8,*,...,*)$ can contain no 8, and at most one 4 or 6, but not both at the same time.
This leaves us with the options $\tau_2=(4,6,8)$ and $\tau_2=(6,6,8)$.
In the first case, $\tau_1=\tau_2$ which not possible by \cref{res:1_2_vertex_same_type}.
In the second case, there would be two adjacent 6-faces, but $P$ does not contain $(6,6)$-edges by \cref{res:edge_types} with $2k'=8$.
Therefore, $P$~cannot contain a 1-vertex of type $(4,6,8)$.
\subsection{The case $\tau_1=(4,6,6)$}
\label{sec:466}
If $P$ contains a 1-vertex of type $(4,6,6)$, then it also contains an adjacent pair of the form
$$(\tau_1,\tau_2)=((4,6,6),(6,6,*,...,*)).$$
We proceed as demonstrated in \cref{ex:infeasible}.
\cref{res:K_E_ineq} $(i)$ yields $\eps_2+\eps_3+\eps_3 < 1/6$.
By \cref{res:exclusive_types} the placeholders can only take on values 4 or 6.
The following table lists the minimally infeasible adjacent pairs and proves their infeasibility.
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ &$\overset?<$& $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(6,6,4,4)$ & $2/6$ & $\not<$ & $(\eps_2+\eps_3+\eps_3)+\eps_2$ & $<$ $2/6$ \\
$(6,6,6,4)$ & $3/6$ & $\not<$ & $(\eps_2+\eps_3+\eps_3)+\eps_3$ & $<$ $2/6$ \\
$(6,6,6,6)$ & $4/6$ & $\not<$ & $(\eps_3+\eps_3)+(\eps_3+\eps_3)$ & $<$ $2/6$
\end{tabular}
\end{center}
By \cref{res:subtype} we conclude: the placeholder in $\tau_2=(6,6,*,...,*)$ can contain at most one 4 or 6, but not both at the same time.
This leaves us with the options $\tau_2=(4,6,6)$ and $\tau_2=(6,6,6)$.
In the first case we have $\tau_1=\tau_2$, which is not possible by \cref{res:1_2_vertex_same_type}.
Excluding $(6,6,6)$ needs more work:
fix a 6-gon $\sigma\in\F_2(P)$.
Each edge of $\sigma$ is either of type $(4,6)$ or of type $(6,6)$ (by \cref{res:edge_types}).
Each 1-vertex of $\sigma$ (which must be of type $(4,6,6)$) is then incident to exactly one of these $(6,6)$-edges of $\sigma$.
Thus, there are exactly \emph{three} $(6,6)$-edges incident to $\sigma$ (see \cref{fig:hexagon_fail}).
On the other hand, each 2-vertex of $\sigma$ is incident to an even number of $(6,6)$-edges of $\sigma$ (since if a 2-vertex is incident to at least one $(6,6)$-edge, then we have previously shown that its type must be $(6,6,6)$, implying another incident $(6,6)$-edge). Therefore the number of $(6,6)$-edges incident to $\sigma$ must be \emph{even} (see \cref{fig:hexagon_fail}), in contradiction to the previously obtained number three of such edges.
\begin{figure}
\centering
\includegraphics[width=0.55\textwidth]{img/hexagon_fail}
\caption{Possible distributions of $(4,6)$-edges (gray) and $(6,6)$-edges (thick) around a 6-gon as discussed in \cref{sec:466}. The top row shows configurations compatible with the conditions set by 1-vertices (black), and the bottom row shows the configurations compatible with the conditions set by the 2-vertices (white).}
\label{fig:hexagon_fail}
\end{figure}
Consequently, $P$~cannot contain a 1-vertex of type $(4,6,6)$.
\begin{observation}
It is a consequence of \cref{sec:466,sec:468,sec:4610} that $P$ cannot have a 1-vertex of a type $(4,6,2k')$ for a $2k'\in\{6,8,10\}$.
By \cref{res:incompatible_1_types} this means that \emph{all} 1-vertices of $P$ are of the same type $\tau_1=(4,4,2k)$ for some fixed $2k\ge 4$.
\end{observation}
It is worth to distinguish the case $(4,4,4)$ from the cases $(4,4,2k)$ with $2k\ge 6$.
\subsection{The case $\tau_1=(4,4,4)$}
\label{sec:444}
In this case, all 2-faces are 4-gons, and all 4-gons are congruent by \cref{res:unique_shape}.
A 4-gon with all edges of the same length is known as a \emph{rhombus}, and the polyhedra with congruent rhombic faces are known as \emph{rhombic isohedra} (from german \emph{Rhombenisoeder}).
These have a known classification:
\begin{theorem}[S. Bilinksi, 1960 \cite{bilinski1960rhombic}]\label{res:rhombic_isohedra}
If $P$ is a polyhedron with congruent rhombic faces, then $P$ is one of the following:
\begin{myenumerate}
\item a member of the infinite family of rhombic hexahedra, \shortStyle{i.e.,}\ $P$ can be obtained from a cube by stretching or squeezing it along a long diagonal,
\item the rhombic dodecahedron,
\item the Bilinski dodecahedron,
\item the rhombic icosahedron, or
\item the rhombic triacontahedron.
\end{myenumerate}
\end{theorem}
The figure below depicts these polyhedra in the given order (from left to right; including only one instance from the family $(i)$):
\begin{center}
\includegraphics[width=0.9\textwidth]{img/rhombic_isohedra}
\end{center}
The rhombic dodecahedron and triacontahedron are known edge- but not vertex-transitive polytopes.
We show that the others are not even strictly bipartite.
\begin{corollary}\label{res:strictly_bipartite_444}
If $P$ is strictly bipartite with all 1-vertices of type $(4,4,4)$, then~$P$ is one of the following:
\begin{myenumerate}
\item the rhombic dodecahedron,
\item the rhombic triacontahedron.
\end{myenumerate}
\begin{proof}
The listed ones are edge-transitive but not vertex-transitive.
Also they are not inscribed.
By \cref{res:trans_is_bipartite} they are therefore \emph{strictly} bipartite.
We then have to exclude the other polyhedra listed in \cref{res:rhombic_isohedra}.
The rhombic hexahedra include the cube, which is inscribed, hence not strictly bipartite.
In all the other cases, there exist vertices where acute and obtuse angles meet (see the figure).
So this vertex cannot be assigned to either $V_1$ or $V_2$ (\shortStyle{cf.}\ \cref{res:2k_4_case}), and the polyhedron cannot be bipartite.
\end{proof}
\end{corollary}
These are the only strictly bipartite polyhedra we will find, and both are edge-transitive without being vertex-transitive.
\subsection{The case $\tau_1=(4,4,2k),2k\ge 6$}
\label{sec:442k}
If $P$ contains a 1-vertex of type $(4,4,2k)$ with $2k\ge 6$, then it also has an adjacent pair of the form
$$(\tau_1,\tau_2)=((4,4,2k),(4,2k,*,...,*)).$$
We proceed as demonstrated in \cref{ex:infeasible}.
\cref{res:K_E_ineq} $(i)$ yields $\eps_2+\eps_2+\eps_k < 1/k$.
Since $(4,4,2k)$ is the only type of 1-vertex of $P$, there are only 4-faces and $2k$-faces and~the placeholders can only take on the values 4 and $2k$ (note that we do \emph{not} use \cref{res:exclusive_types} for this).
The following table lists some inequalities derived for infeasible pairs:
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ &$\overset?<$& $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(4,2k,4,4,4)$ & $1-1/k$ & $<$ & $(\eps_2+\eps_2+\eps_k)+(\eps_2+\eps_2)$ & $< 2/k$ \\
$(4,2k,4,4,2k)$ & $3/2-2/k$ &$<$& $(\eps_2+\eps_2+\eps_k)+(\eps_2+\eps_k)$ & $< 2/k$
\end{tabular}
\end{center}
One checks that these inequalities are not satisfied for $2k\ge 6$.
\Cref{res:subtype} then states that the placeholders can contain at most two 4-s, and if exactly two, then nothing else.
Moreover, $\tau_2$ must contain at least as many 4-s as it con\-tains $2k$-s, as otherwise we would find two adjacent $2k$-faces while $P$ cannot contain a $(2k,2k)$-edge by \cref{res:edge_types}.
We are therefore left with the following options for $\tau_2$:
$$(4,4,2k),\; (4,4,4,2k) \quad\text{and}\quad (4,2k,4,2k).$$
The case $\tau_2=(4,4,2k)$ is impossible by \cref{res:1_2_vertex_same_type}.
We show that $\tau_2=(4,4,4,2k)$~is also not possible:
consider the local neighborhood of a $(4,4,4,2k)$-vertex (the highlighted~ver\-tex) in the following figure:
\begin{center}
\includegraphics[width=0.25\textwidth]{img/contradiction_4442k}
\end{center}
Since the 1-vertices (black dots) are of type $(4,4,6)$, this configuration forces on us the~existence of the two gray 6-faces.
These two faces intersect in a 2-vertex, which is then incident to two $2k$-faces and must be of type $(4,2k,4,2k)$.
But we can show that the types $(4,4,4,2k)$ and $(4,2k,4,2k)$ are incompatible by \cref{res:spherical_interior_angles_are_good} $(i)$:
$$\beta_2^2+\beta_2^2+\beta_2^2+\beta_2^k \overset{(i)}= \beta_2^2+\beta_2^k+\beta_2^2+\beta_2^k \;\implies\; \beta_2^2=\beta_2^k \;\overset{\eqref{eq:angle_relations}}\implies\; 4=2k\ge 6.$$
Thus, $(4,4,4,2k)$ cannot occur.
We conclude that every 2-vertex incident to a $2k$-face must be of type $(4,2k,4,2k)$.
Consider then the following table:
\begin{center}
\begin{tabular}{l|rcll}
$\tau_2$ & $s-2-K(\tau_2)$ &$\overset?<$& $E(\tau_2)$ & \\[0.4ex]
\hline
$\overset{\phantom.}(4,2k,4,2k)$ & $1-2/k$ &$<$& $(\eps_2+\eps_2+\eps_k)+\eps_2$ & $< 2/k$
\end{tabular}
\end{center}
The established inequality yields $2k\le 6$, and hence $2k=6$.
We found that then all 1-vertices must be of type $(4,4,6)$, and all 2-vertices incident to a 6-face must be of type $(4,6,4,6)$.
\section{Bipartite polytopes}
\label{sec:bipartite_polytopes}
From this section on let $P\subset\RR^d,d\ge 2$ denote a $d$-dimensional polytope of full dimension (\shortStyle{i.e.,}\ $P$ is not contained in a proper affine subspace).
By $\F(P)$ we denote the face lattice of $P$, and by $\F_\delta(P)\subset\F(P)$ the subset of $\delta$-dimensional faces.
\begin{definition}\label{def:bipartite}
$P$ is called \emph{bipartite}, if
\begin{myenumerate}
\item all its edges are of the same length $\ell$,
\item its edge-graph is bipartite, which we write as $G_P=(V_1\mathbin{\mathaccent\cdot\cup} V_2,E)$, and
\item there are radii $r_1\le r_2$ so that $\|v\|=r_i$ for all $v\in V_i$.
\end{myenumerate}
If $r_1<r_2$, then $P$ is called \emph{strictly bipartite}.
A vertex $v\in V_i$ is called an \emph{$i$-vertex}.
The numbers $r_1$, $r_2$ and $\ell$ are called the \emph{parameters} of a bipartite polytope.
\end{definition}
\begin{remark}
\label{rem:abuse_of_notation}
Since $P$ is full-dimensional by convention, \cref{def:bipartite} only defines \emph{full-dimensional} bipartite polytopes.
To extend this notion to not necessarily full-dimensional polytopes, we shall call a polytope \emph{bipartite} even if it is just bipartite as a subset of its affine hull where we made an appropriate choice of origin in the affine hull (note that whether a polytope is bipartite depends on its placement relative to the origin and that there is at most one such placement if the polytope is full-dimensional).
This comes in handy when we discuss faces of bipartite polytopes.
\end{remark}
\begin{remark}
\label{rem:alternative_definiton}
An alternative definition of bipartite polytope would replace \itm3 by the condition that $P$ has an \emph{edge in-sphere}, that is, a sphere that touches each edge of $P$ in a single point (this definition was used in the abstract).
\mbox{The configuration~de}\-picted below (an edge of $P$ connecting two vertices $v_1\in V_1$ and $v_2 \in V_2$) shows~how any one of the four quantities $r_1,r_2,\ell$ and $\rho$ (the radius of the edge in-sphere) is determined from the other three by solving the given set of equations:
\begin{center}
\includegraphics[width=0.35\textwidth]{img/edge_in_sphere}
\qquad\quad\raisebox{5em}{$\displaystyle
\begin{array}{rcl}
\rho^2+\ell_1^2&\!\!\!\!=\!\!\!\!\!&r_1^2\\[0.3ex]
\rho^2+\ell_2^2&\!\!\!\!=\!\!\!\!\!&r_2^2\\[0.3ex]
\ell_1+\ell_2&\!\!\!\!=\!\!\!\!\!&\ell
\end{array}
$}
\end{center}
There is a subtlety: for the edge in-sphere to actually touch the edge (rather than only its affine hull outside of the edge) it is necessary that the perpendicular projection of the origin onto the edge ends up inside the edge (equivalently, that the triangle $\conv\{0,v_1,v_2\}$ is acute at $v_1$ and $v_2$). One might regard this as intuitively clear since we are working with convex polytopes, but this will also follows formally as part of our proof of \cref{res:0_in_P} (as we shall mention there in a footnote).
This alternative characterization of bipartite polytopes via edge in-spheres will become relevant towards the end of the classification (in \cref{sec:446}).
Still, for the larger part of our investigation, \cref{def:bipartite} \itm3 is the more convenient version to work with.
\end{remark}
\subsection{General obsevations}
\begin{proposition}\label{res:trans_is_bipartite}
If $P$ is edge- but not vertex-transitive, then $P$ is bipartite.
\end{proposition}
This is a geometric analogue to the well known fact that every edge- but not vertex-transitive \emph{graph} is bipartite.
A proof of the graph version can be found in \cite{godsil1978graphs}.
The following proof can be seen as a geometric analogue:
\begin{proof}[Proof of \cref{res:trans_is_bipartite}]
Clearly, all edges of $P$ are of the same length.
Fix some edge $e\in\F_1(P)$ with end vertices $v_1,v_2\in\F_0(P)$.
Let $V_i$ be the orbit of $v_i$ under $\Aut(P)$.
We prove that $V_1\cup V_2=\F_0(P)$, $V_1\cap V_2=\eset$ and that the edge graph $G_P$ is bipartite with partition $V_1\mathbin{\mathaccent\cdot\cup} V_2$.
Let $v\in \F_0(P)$ be some vertex and $\tilde e\in\F_1(P)$ an incident edge.
By edge-transitivity, there is a symmetry $T\in\Aut(P)$ that maps $\tilde e$ onto $e$, and therefore maps $v$ onto $v_i$ for some $i\in\{1,2\}$.
Thus, $v$ is in the orbit $V_i$.
This holds for all vertices of $P$, and therefore $V_1\cup V_2=\F_0(P)$.
The orbits of $v_1$ and $v_2$ must either be identical or disjoint.
Since $V_1\cup V_2=\F_0(P)$, from $V_1=V_2$ it would follow $V_1=\F_0(P)$, stating that $P$ has a single orbit of vertices.
But since $P$ is \emph{not} vertex-transitive, this cannot be.
Thus, $V_1\cap V_2=\eset$, and therefore $V_1\mathbin{\mathaccent\cdot\cup} V_2=\F_0(P)$.
Let $\tilde e\in\F_1(P)$ be an edge with end vertices $\tilde v_1$ and $\tilde v_2$.
By edge-transitivity, $\tilde e$ can be mapped onto $e$ by some symmetry $T\in\Aut(P)$.
Equivalently $\{T\tilde v_1,T\tilde v_2\}=\{v_1,v_2\}$.
Since $v_1$ and $v_2$ belong to different orbits under $\Aut(P)$, so do $\tilde v_1$ and $\tilde v_2$.
Hence $\tilde e$ has one end vertex in $V_1$ and one end vertex in $V_2$.
This holds for all edges, and thus, $G_P$ is bipartite with partition $V_1\mathbin{\mathaccent\cdot\cup} V_2$.
It remains to determine the radii $r_1\le r_2$.
Set $r_i:=\|v_i\|$ (assuming w.l.o.g.\ that $\|v_1\|\le\|v_2\|$).
Then for every $v\in V_i$ there is a symmetry $T\in\Aut(P)\subset\Ortho(\RR^d)$ so that $Tv_i=v$, and thus
$$\|v\|=\|T v_i\| = \|v_i\|=r_i.$$
\end{proof}
Bipartite polytopes are more comfortable to work with than edge- but not vertex-transitive polytopes because their faces are again bipartite polytopes (in the sense as explained in \cref{rem:abuse_of_notation}).
Later, this will enable us to reduce the problem to an investigation in lower dimensions.
\begin{proposition}\label{res:faces_of_bipartite}
Let $\sigma\in\F(P)$ be a face of $P$. Then it holds
\begin{myenumerate}
\item if $P$ is bipartite, so is $\sigma$.
\item if $P$ is strictly bipartite, then so is $\sigma$, and $v\in\F_0(\sigma)\subseteq\F_0(P)$ is an $i$-vertex in $P$ if and only if it is an $i$-vertex in $\sigma$.
\item if $r_1\le r_2$ are the radii of $P$ and $\rho_1\le \rho_2$ are the radii of $\sigma$, then there holds
%
$$h^2 + \rho_i^2 = r_i^2,$$
%
where $h$ is the height of $\sigma$, that is, the distance of $\aff(\sigma)$ from the origin.
\end{myenumerate}
\begin{proof}
\iffalse
Properties clearly inherited by $\sigma$ are that all edges are of the same length and that the edge graph is bipartite.
By the latter, the vertex set of $\sigma$ can be partition like $W_1\mathbin{\mathaccent\cdot\cup} W_2=\F_0(\sigma)$, where $W_1$ and $W_2$ are the bipartition classes of the edge graph.
It remains to show the existence of radii $\rho_1,\rho_2>0$ (w.l.o.g.\ $\rho_1\le\rho_2$) so that $\|v-c\|=\rho_i$ for all $v\in W_i$, where $c\in\RR^d$ is a suitable center of $\sigma$ (translating $\sigma$ so that $c$ becomes the origin gives a bipartite polytope as in \cref{def:bipartite}).
By definition, all vertices of $P$ lie on $\Sph_{r_1}(0)\cup\Sph_{r_2}(0)$ (where $\Sph_r(x)$ denotes the $(d-1)$-sphere of radius $r$ centered at $x$).
The vertices of $\sigma$ therefore lie on $\aff(\sigma)\cap(\Sph_{r_1}(0)\cup\Sph_{r_2}(0))$ which consists of two concentric $(d-2)$-spheres (see the figure below).
Let $c\in\aff(\sigma)$ be the common center of these $(d-2)$-spheres, and note that $\aff(\sigma)$ is perpendicular to the vector $c$, and $\|c\|=h$ (the height of $\sigma$ as defined in $(iii)$).
\begin{center}
\includegraphics[width=0.7\textwidth]{img/aff_sphere_intersection}
\end{center}
For an $i$-vertex $v\in\F_0(\sigma)$ consider the triangle $\Delta:=\conv\{0,v,c\}$.
This is~a~right~triangle (with right angle at $c$), and
$$(*)\quad \rho_i^2:=\|v-c\|^2 = \|v\|^2-\|c\|^2 = r_i^2-h^2.$$
In particular, the distance $\rho_i$ of $v$ from $c$ does only depend on $i$, which provides the existence of the radii and proves $(i)$.
Also, $(*)$ is equivalent to $(iii)$.
From $(*)$ also follows $r_1< r_2\Leftrightarrow \rho_1< \rho_2$, which proves $(ii)$.
\hrulefill
\fi
Properties clearly inherited by $\sigma$ are that all edges are of the same length and that the edge graph is bipartite.
It remains to show the existence of the radii $\rho_1\le \rho_2$ compatible with the bipartition of the edge-graph of $\sigma$.
Let $c\in\aff(\sigma)$ be the orthogonal projection of $0$ onto $\aff(\sigma)$.
Then $\|c\|=h$, the height of $\sigma$ as defined in $(iii)$.
For any vertex $v\in\F_0(\sigma)$ which is an $i$-vertex in $P$, the triangle $\Delta:=\conv\{0,c,v\}$ has a right angle at $c$.
Set $\rho_i:=\|v-c\|$ and observe
\begin{equation}
\quad \rho_i^2:=\|v-c\|^2 = \|v\|^2-\|c\|^2 = r_i^2-h^2.
\tag{$*$}
\end{equation}
In particular, the value $\rho_i$ does only depend on $i$.
In other words, $\sigma$ is a bipartite polytope when considered as a subset of its affine hull, where the origin is chosen to be $c$ (\shortStyle{cf.}\ \cref{rem:abuse_of_notation}).
This proves $(i)$, and $(*)$ is equivalent to the equation in $(iii)$.
From $(*)$ also follows $r_1< r_2\Leftrightarrow \rho_1< \rho_2$, which proves $(ii)$.
\end{proof}
\end{proposition}
\iftrue
The following observation will be of use later on.
\begin{observation}\label{res:angles_between_vertices}
Given two adjacent vertices $v_1,v_2\in\F_0(P)$ with $v_i\in V_i$, and if $P$ has parameters $r_1,r_2$ and $\ell$, then
$$\ell^2 = \|v_1-v_2\|^2 = \|v_1\|^2+\|v_2\|^2 - 2\<v_1,v_2\> = r_1^2+r_2^2-2 r_1 r_2 \cos\angle(v_1,v_2),$$
This can be rearranged for $\cos\angle(v_1,v_2)$. While the exact value of this expression is not of relevance to us, this shows that this angle is determined by the parameters and does not depend on the choice of the adjacent vertices $v_1$ and $v_2$.
\end{observation}
\subsection{Bipartite polygons}
The easiest to describe (and to explicitly construct) are the bipartite \emph{polygons}.
Foremost, the edge-graph is bipartite, and thus, a bipartite polygon must be a $2k$-gon for some $k\ge 2$.
One can show that the bipartite polygons are exactly the edge-transitive $2k$-gons (\shortStyle{cf.}\ \cref{fig:2n_gons}), and that such one is \emph{strictly} bipartite if and only if it is \emph{not} vertex-transitive (or equivalently, not regular).
We will not make use of these symmetry properties of bipartite polygons.
The parameters $r_1, r_2$ and $\ell$ uniquely determine a bipartite polygon, as can be seen by explicit construction:
\begin{center}
\includegraphics[width=0.98\textwidth]{img/polygon_construction}
\end{center}
One starts with an arbitrary 1-vertex $v\in\RR^2$ placed on the circle $\Sph_{r_1}(0)$.
Its~neigh\-boring vertices are then uniquely determined as the intersections $\Sph_{r_2}(0)\cap\Sph_\ell(v)$.
The procedure is repeated with the new vertices until the edge cycle closes (which only happens if the parameters are chosen appropriately).
The procedure also makes clear that the interior angle $\alpha_i\in(0,\pi)$ at an $i$-vertex only depends on $i$, but not on the chosen vertex $v\in V_i$.
\begin{corollary}\label{res:2k_gons}
A bipartite polygon $P\subset\RR^2$ is a $2k$-gon with alternating interior angles $\alpha_1,\alpha_2\in(0,\pi)$ ($\alpha_i$ being the interior angle at an $i$-vertex), and its shape is uniquely determined by its parameters (up to congruence).
\end{corollary}
The exact values for the interior angles are not of relevance.
Instead, we only need the following properties:
\begin{proposition}\label{res:angles}
The interior angles $\alpha_1,\alpha_2\in(0,\pi)$ satisfy
\begin{equation}\label{eq:angles}
\alpha_1+\alpha_2=2\alpha_{\mathrm{reg}}^k
\quad\text{and}\quad
\alpha_2\le \alpha_{\mathrm{reg}}^k\le\alpha_1,
\end{equation}
where $\alpha_{\mathrm{reg}}^k:=(1-1/k)\pi$ is the interior angle of a regular $2k$-gon, and the inequalities are satisfied with equality if and only $r_1=r_2$.
\begin{proof}
The sum of interior angles of a $2k$-gon is $2(k-1)\pi$, and thus
$k \alpha_1+k\alpha_2=2(k-1)\pi$, which, after division by $k$, yields the first part of \eqref{eq:angles}.
For two adjacent vertices $v_1,v_2\in \F_0(P)$ (where $v_i\in V_i$), consider the triangle $\Delta:=\conv\{0,v_1,v_2\}$ whose edge lengths are $r_1, r_2$ and $\ell$, and whose interior angles at $v_1$ resp.\ $v_2$ are $\alpha_1/2$ resp.\ $\alpha_2/2$.
From $r_1\le r_2$ (resp.\ $r_1<r_2$) and the law of sine follows $\alpha_1 \ge \alpha_2$ (resp.\ $\alpha_1 > \alpha_2$).
With $\alpha_1+\alpha_2=2\alpha_{\mathrm{reg}}^k$ this yields the second part of \eqref{eq:angles}.
\end{proof}
\end{proposition}
\begin{observation}\label{res:2k_4_case}
For later use (in \cref{res:strictly_bipartite_444}), consider \cref{res:angles} with $2k=$ $4$.
In this case we find,
$$\alpha_2\le \frac\pi 2\le \alpha_1,$$
that is, $\alpha_1$ is never acute, and $\alpha_2$ is never obtuse.
\end{observation}
\iffalse
\begin{definition}
Let $v,w\in\F_0(P)$ be any two adjacent vertices in a bipartite polygon $P$.
The triangle $\Delta(P):=\conv\{0,v,w\}$ is called the \emph{fundamental triangle} of $P$.
\end{definition}
The edge lengths of $\Delta(P)$ are $r_1, r_2$ and $\ell$.
Thus the parameters completely determine the shape of the triangle, and (concerning its shape and not it position) it is justified to speak of \emph{the} fundamental triangle.
\begin{proposition}
If $P$ is a bipartite polygon, then $0\in\mathrm{int}(P)$.
\begin{proof}
The interior angle sum of a $2k$-gon is $2(k-1)\pi$, and thus, $P$ has an interior angle of size at least
$$\frac{2(k-1)\pi}{2k}=\Big(1-\frac1k\Big)\pi \ge \Big(1-\frac12\Big)\pi=\frac\pi2.$$
If all interior angles are of size $\pi/2$, then $P$ is a square, and it is easy to see that it must contain the origin in order to be bipartite.
We can therefore assume that $P$ contains an interior angle strictly larger than $\pi/2$.
Let $v\in \F_0(P)$ be such a vertex with interior angle $>\pi/2$, and let $w_1,w_2\in \F_0(P)$ its two neighbors.
Set $u_i:=w_i-v$.
Then $\angle(u_1,u_2)$ is the interior angle of $P$ at $v$, hence $\<u_1,u_2\>\le 0$.
Similarly, we see
$$\<u_i,v\>=\<w_i,v\>-\<v,v\> = $$
\end{proof}
\end{proposition}
We start with an observation whose most direct consequences are for polygons, but which we state for general dimensions for later use:
\begin{observation}\label{res:angles_between_vertices}
Given two adjacent vertices $v_1,v_2\in\F_0(P)$ of a bipartite polytope $P$ so that $v_i\in V_i$. If $P$ has parameters $r_1,r_2$ and $\ell$, then
$$\ell^2 = \|v_1-v_2\|^2 = \|v_1\|^2+\|v_2\|^2 - 2\<v_1,v_2\> = r_1^2+r_2^2-2 r_1 r_2 \cos\angle(v_1,v_2),$$
which rearranges to
$$\angle(v_1,v_2) = \arccos\Big(\frac{r_1^2+r_2^2-\ell^2}{2r_1 r_2}\Big).$$
In particular, the angle can be computed from the parameters alone.
In the case of a bipartite $n$-gon, where the edges form a closed cycle, this implies that the vectors of the vertices are equally spread out by an angle of $2\pi/n$.
That is, for any two adjacent vertices $v_1,v_2\in\F_0(P)$ holds
$$\angle(v_1,v_2)=\frac{2\pi}n.$$
\end{observation}
\begin{proposition}
The shape of a bipartite polygon $P\subset\RR^2$ is uniquely determined by its parameters. In particular:
\begin{myenumerate}
\item there are uniquely determined angles $\alpha_1,\alpha_2\in(0,\pi)$ so that the interior angle of $P$ at an $i$-vertex is $\alpha_i$.
\item it holds
%
$\alpha_2\le \alpha_{\mathrm{reg}}^{\smash k}\le \alpha_1,$
%
where $\alpha_{\mathrm{reg}}^{\smash k}:=(1-1/k)\pi$ is the interior angle of a regular $2k$-gon.
\end{myenumerate}
\begin{proof}
The shape of a polygon is uniquely determined by its edge lengths and interior angles (and their exact order), and thus it suffices to determine the latter ones.
The neighbors of a vertex $v\in V_1$ are exactly the two points in the intersection $\Sph_{r_2}(0)\cap\Sph_\ell(v)$.
\begin{center}
\includegraphics[width=0.4\textwidth]{img/neighbors}
\end{center}
This uniquely determines the interior angle at $v$.
An elementary computation would yield
$$\alpha_1=2\arccos\Big(\frac{\ell^2+r_1^2-r_2^2}{2\ell r_1}\Big).$$
In particular, the value does not depend on the exact vertex $v\in V_1$.
Equivalently for $v\in V_2$.
$P$ is a polygon with all edges of length $\ell$ and alternating interior angles $\alpha_1$ and $\alpha_2$. Its shape is therefore uniquely determined.
The interior angle sum of a $2k$-gon is $2(k-1)\pi$, and thus
$$k\alpha_1+ k\alpha_2 = 2(k-1)\pi\quad\implies\quad (*)\;\;\alpha_1+\alpha_2=2\Big(1-\frac1k\Big)\pi = 2\alpha_{\mathrm{reg}}^k.$$
\end{proof}
\end{proposition}
\begin{proposition}\label{res:2d_shape}
Let $P\subset\RR^2$ be a bipartite polygon with parameters $r_1\le r_2$ and $\ell$.
It holds:
\begin{myenumerate}
\item $P$ has alternating interior angles, that is, there are $\alpha_1,\alpha_2\in(0,\pi)$ so that the interior angle of $P$ at an $i$-vertex is $\alpha_i$.
\item The shape of $P$ (number of vertices, edge lenghts and interior angles) is uniquely determined by~the~para\-meters of $P$.
\item If $\alpha_{\mathrm{reg}}^k:=(1-1/k)\pi$ denotes the interior angle of a regular $2k$-gon, then $\alpha_2\le \alpha_{\mathrm{reg}}^{\smash{k}}\le\alpha_1$, with equality (in either place) if and only if $r_1=r_2$.
\end{myenumerate}
\begin{proof}
Let $v_1,...,v_{2k}\in\F_0(P)$ be the vertices of $P$ (in this order).
The polygons then decomposes into triangles $\Delta_i:=\conv\{0,v_i,v_{i+1}\}$ for $i\in\{1,...,2k\}$ (all indices are considered modulo $2k$; see the figure below). \msays{Why are the triangles disjoint?}
\begin{center}
\includegraphics[width=0.37\textwidth]{img/bipartite_polygon}
\end{center}
The sides of each triangle have lengths $r_1$, $r_2$ and $\ell$, and thus, all the triangles are congruent and their shape is determined by the parameters.
In particular, they have the same interior angles.
Denote by $\theta_1$ (resp.\ $\theta_2$) the interior angle opposite to the~side of length $r_2$ (resp.\ $r_1$; not that $\theta_i$ is \emph{not} opposite to $r_i$, but the other way around).
The interior angle of $\Delta_i$ at the origin is $\pi/k$ (as determined in \cref{res:angles_between_vertices}).
And since this angle is also determined by the parameters, so must be $k$.
An $i$-vertex $v\in F_0(P)$ lies in two triangles.
In both triangles, the interior angle at $v$ is $\theta_i$.
Thus, the interior angle of $P$ at $v$ is $\alpha_i=2\theta_i$.
In particular, $\alpha_i$ does only depend on $i$ (this proves $(i)$) and the parameters (since also $k$ is determined by the parameters, this proves $(ii)$).
As well-known, the interior angles of a $2k$-gon add up to $2(k-1)\pi$, which yields
$$k\alpha_1+ k\alpha_2 = 2(k-1)\pi\quad\implies\quad (*)\;\;\alpha_1+\alpha_2=2\Big(1-\frac1k\Big)\pi = 2\alpha_{\mathrm{reg}}^k.$$
By the sine theorem applied to some triangle $\Delta_i$ we have
$$\frac{\sin(\alpha_1/2)}{\sin(\alpha_2/2)} = \frac{r_2}{r_1},$$
and from this follows $r_1<r_2\Leftrightarrow \alpha_1>\alpha_2$ and $r_1=r_2\Leftrightarrow \alpha_1=\alpha_2$.
Together with $(*)$ this proves $(iii)$.
\end{proof}
\end{proposition}
\fi
\subsection{The case $r_1=r_2$}
We classify the inscribed bipartite polytopes, that is, those with coinciding radii $r_1=r_2$.
This case is made especially easy by a classification result from \cite{winter2019classification}.
We need the following definition:
\begin{definition}\label{def:permutahedron}
Let $\Gamma\subset\Ortho(\RR^d)$ be a finite reflection group and $v\in \RR^d$ a \emph{generic} point \shortStyle{w.r.t.}\ $\Gamma$ (\shortStyle{i.e.,}\ $v$ is not fixed by a non-identity element of $\Gamma$).
The orbit~polytope
$$\Orb(\Gamma,v):=\conv\{Tv\mid T\in\Gamma\}\subset\RR^d$$
is called a \emph{$\Gamma$-permutahedron}.
\end{definition}
The relevant result then reads
\begin{theorem}[Corollary 4.6.\ in \cite{winter2019classification}]
\label{res:classification_inscribed_zonotopes}
If $P$ has only centrally symmetric 2-dimen\-sional faces (that is, it is a zonotope), has all vertices on a common sphere and all edges of the same length, then $P$ is a $\Gamma$-permutahedron.
\end{theorem}
This provides a classification of bipartite polytopes with $r_1=r_2$.
\begin{theorem}
If $P\subset\RR^d$ is bipartite with $r_1=r_2$, then it is a $\Gamma$-permutahedron.
\begin{proof}
If $r_1=r_2$, then all vertices are on a common sphere (that is, $P$ is inscribed).
By definition, all edges are of the same length.
Both statements then also hold for the faces of $P$, in particular, the 2-dimensional faces.
An inscribed polygon with a unique edge length is necessarily regular.
With \cref{res:2k_gons} the 2-faces are then regular $2k$-gons, therefore centrally symmetric.
Summarizing, $P$ is inscribed, has all edges of the same length, and all 2-dimen\-sio\-nal faces of $P$ are centrally symmetric.
By \cref{res:classification_inscribed_zonotopes}, $P$ is a $\Gamma$-permutahedron.
\end{proof}
\end{theorem}
$\Gamma$-permutahedra are vertex-transitive by definition, hence do not provide examples of edge- but not vertex-transitive polytopes.
\subsection{Strictly bipartite polytopes}
It remains to classify the \emph{strictly} bipartite~poly\-topes.
This problem is divided into two independent cases: dimension $d=3$, and dimension $d\ge 4$.
The detailed study of the case $d=3$ (which turns out to be the actual hard work) is postponed until \cref{sec:bipartite_polyhedra}, the result of which is the following theorem:
\begin{theorem}\label{res:strictly_bipartite_polyhedra}
If $P\subset\RR^3$ is strictly bipartite, then $P$ is the rhombic dodecahedron or the rhombic triacontahedron.
\end{theorem}
Presupposing \cref{res:strictly_bipartite_polyhedra}, the case $d\ge 4$ is done quickly.
\begin{theorem}\label{res:3_dim_suffices}
There are no strictly bipartite polytopes in dimension $d\ge 4$.
\begin{proof}
It suffices to show that there are no strictly bipartite polytopes in dimension $d=4$, as any higher-dimensional example has a strictly bipartite 4-face (by \cref{res:faces_of_bipartite}).
Let $P\subset\RR^4$ be a strictly bipartite 4-polytope.
Let $e\in\F_1(P)$ be an edge of $P$.
Then there are $s\ge 3$ cells (aka.\ 3-faces) $\sigma_1,...,\sigma_s\in\F_3(P)$ incident to $e$, each of which is again strictly bipartite (by \cref{res:faces_of_bipartite}).
By \cref{res:strictly_bipartite_polyhedra} each $\sigma_i$ is a rhombic dodecahedron~or~rhom\-bic triacontahedron.
The dihedral angle of the rhombic dodecahedron resp.\ triacontahedron is $120^\circ$ resp.\ $144^\circ$ at every edge \cite{coxeter1973regular}.
However, the dihedral angles meeting at $e$ must sum up to less than $2\pi$.
With the given dihedral angles this is impossible.
\end{proof}
\end{theorem}
\section{Conclusions and open questions}
\label{sec:conclusions}
In this paper we have shown that any edge-transitive (convex) polytope in four or more dimensions is necessarily vertex-transitive.
We have done this by classifying all polytopes which simultaneously have all edges of the same length, an edge in-sphere and a bipartite edge graph (which we named \emph{bipartite} polytopes).
The obstructions we~derived for being edge-transitive without being vertex-transitive have been mainly geometrical and less a matter of symmetry (a detailed investigation of the Euclidean symmetry groups was not necessary, but it might be interesting to view the problem from this perspective).
We suspect that dropping convexity or considering combinatorial symmetries instead of geometrical ones will quickly lead to further examples of just edge-transitive structures.
For example, it is easy to find embeddings of graphs into $\RR^d$ with these properties.
Slightly stronger than being simultaneously vertex- and edge-transitive, is being transitive on~\emph{arcs}, that is, on incident vertex-edge pairs.
This additional degree of symmetry allows an edge to be not only mapped onto any other edge, but also onto itself with inverted orientation.
While there are graphs that are vertex- and edge-transitive without being arc-transitive (the so-called \emph{half-transitive} graphs, see \cite{holt1981graph}), we believe it is unlikely that this distinction is necessary for convex polytopes.
\begin{question}
Is there a polytope $P\subset\RR^d$ that is edge-transitive and vertex-tran\-sitive,~but~not arc-transitive?
\end{question}
In a different direction, the questions of this paper naturally generalize to faces of higher dimensions. In general, the interactions between transitivities of faces of different dimensions have been little investigated.
For example, already the~following question seems to be open:
\begin{question}
For fixed $k\in\{2,...,d-3\}$, are there convex $d$-polytopes for arbitrarily large $d\in\NN$ that are transitive on $k$-dimensional faces without being transitive on either vertices or facets?
\end{question}
Of course, any such question could be attacked by attempting to classify the $k$-face-transitive (convex) polytopes for some $k\in\{1,...,d-2\}$.
It seems to be~un\-clear for which $k$ this problem is tractable (for comparison, $k=0$ is intractable, see \cite{babai1977symmetry}), and it appears that there are no techniques applicable to all (or many) $k$ at the same time.
\section{Introduction}
\label{sec:introduction}
A $d$-dimensional (convex) polytope $P\subset\RR^d$ is the convex hull of finitely many points.
The polytope $P$ is \emph{vertex-transitive} resp.\ \emph{edge-transitive} if its (orthogonal) symmetry group $\Aut(P)\subset\Ortho(\RR^d)$ acts transitively on its vertices resp.\ edges.
It has long been known that there are exactly \emph{nine} edge-transitive polyhedra in $\RR^3$ (see \shortStyle{e.g.}\ \cite{grunbaum1987edge}).
These are the five Platonic solids (tetrahedron, cube, octahedron, icosahedron and \mbox{dodecahedron}) together with the cuboctahedron, the icosidodecahedron, and their duals, the \emph{rhombic dodecahedron} and the \emph{rhombic triacontahedron} (depicted below in this roder):
\begin{center}
\includegraphics[width=0.8\textwidth]{img/edge_transitive_polyhedra}
\end{center}
Little is known about the analogous question in higher dimensions.
Branko Grünbaum writes in \enquote{Convex Polytopes} \cite[p.\ 413]{grunbaum2013convex}
\begin{quote}
No serious consideration seems to have been given to polytopes in dimension $d\ge 4$ about which transitivity of the symmetry group is assumed only for faces of suitably low dimensions, [...].
\end{quote}
Even though families of higher-dimensional edge-transitive polytopes have been studied, to the best of our knowledge, no classification of these has been achieved so far.
Equally striking, all the known examples of such polytopes in dimension at least four are simultaneously \emph{vertex-transitive}.
In dimension up to three, certain polygons (see \cref{fig:2n_gons}), as well as the rhombic dodecahedron and rhombic triacontahedron are edge- but \emph{not} vertex-transitive.
No higher dimensional example of this kind has been found.
In this paper we prove that this is not for lack of trying:
\begin{theorem}\label{res:edge_implies_vertex}
In dimension $d\ge 4$, edge-transitivity of convex polytopes implies vertex-transitivity.
\end{theorem}
As immediate consequence, we obtain the classification of all polytopes that are edge- but not vertex-transitive.
The list is quite short:
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{img/2n_gons}
\caption{Some examples of edge-transitive $2n$-gons with $2n\in\{4,6,8\}$ (the same works for all $n$). The polygons depicted with black boundary are not vertex-transitive.}
\label{fig:2n_gons}
\end{figure}
\begin{corollary}
If $P\subset\RR^d,d\ge 2$ is edge- but not vertex-transitive, then $P$ is one of the following:
\begin{myenumerate}
\item a non-regular $2k$-gon (see \cref{fig:2n_gons}),
\item the rhombic dodecahedron, or
\item the rhombic triacontahedron.
\end{myenumerate}
\end{corollary}
\Cref{res:edge_implies_vertex} is proven by embedding the class of edge- but not vertex-transitive polytopes in a larger class of polytopes, defined by geometric regularities instead of symmetry.
In \cref{res:trans_is_bipartite} we show that a polytope $P\subset\RR^d$ which is edge- but not vertex-transitive must have all of the following properties:
\begin{myenumerate}
\item all edges are of the same length,
\item it has a bipartite edge-graph $G_P=(V_1\mathbin{\mathaccent\cdot\cup} V_2,E)$, and
\item there are radii $r_1\le r_2$, so that $\|v\|=r_i$ for all $v\in V_i$.
\end{myenumerate}
We compile this into a definition: a polytope that has these three properties shall be called \emph{bipartite} (\shortStyle{cf.}\ \cref{def:bipartite}).
The edge- but not vertex-transitive polytopes then form a subclass of the bipartite polytopes, but the class of bipartite polytopes is much better behaved.
For example, faces of bipartite polytopes are bipartite (\cref{res:faces_of_bipartite}), something which is not true for edge/vertex-transitive polytopes\footnote{For example, consider a vertex-transitive but not uniform antiprism. Its faces are non-regular triangles, which are thus not vertex-transitive. Alternatively, consider the $(n,n)$-duoprism, $n\not=4$, that is, the cartesian product of a regular $n$-gon with itself. This polytope is edge-transitive, but its facets are $n$-gonal prisms (the cartesian product of a regular $n$-gon with an edge), which are not edge-transitive.}.
\mbox{Our quest is then to classify all bipartite polytopes.
The~sur}\-prising result: already being bipartite is very restrictive:
\begin{theorem}
If $P\subset\RR^d,d\ge 2$ is bipartite, then $P$ is one of the following:
\begin{myenumerate}
\item an edge-transitive $2k$-gon (see \cref{fig:2n_gons}),
\item the rhombic dodecahedron,
\item the rhombic triacontahedron, or
\item a $\Gamma$-permutahedron for some finite reflection group $\Gamma\subset\Ortho(\RR^d)$ (see \cref{def:permutahedron}; some 3-dimensional examples are shown in \cref{fig:permutahedron}).
\end{myenumerate}
\end{theorem}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{img/permutahedron}
\caption{From left to right: the $A_3$-, $B_3$ and $H_3$-permutahedron.}
\label{fig:permutahedron}
\end{figure}
The $\Gamma$-permutahedra are vertex-transitive, and all the other entries in the list are of dimension $d\le 3$. This immediately implies \cref{res:edge_implies_vertex}.
Remarkably, despite the definition of bipartite polytope being purely geometric, all bipartite polytopes are highly symmetric, that is, at least vertex- or facet-transitive, and sometimes even edge-transitive.
\subsection{Overview}
In \cref{sec:bipartite_polytopes} we introduce the central notion of \emph{bipartite \mbox{polytope}} and prove its most relevant properties: that being bipartite generalizes being edge- but not vertex-transitive, and that all faces of bipartite polytopes are again bipartite.
We then investigate certain subclasses of bipartite polytopes: bipartite polygons and inscribed bipartite polytopes.
We prove that the latter coincide with the $\Gamma$-permutahedra, a class of vertex-transitive polytopes.
It therefore remains to classify the non-inscribed cases, the so-called \emph{strictly} bipartite polytopes.
We show that the classification of these reduces to the classification of bipartite \emph{polyhedra}, \shortStyle{i.e.,}\ the case $d=3$.
From \cref{sec:bipartite_polyhedra} on the investigation is focused on the class of strictly bipartite polyhedra.
We successively determine restrictions on the structure of such, \shortStyle{e.g.}\ the degrees of their vertices and the shapes of their faces.
This quite elaborate process uses many classical geometric results and techniques, including spherical polyhedra, the classification of rhombic isohedra and the realization of graphs as edge-graphs of polyhedra.
As a result, we can exclude all but two cases, namely, the rhombic dodecahedron, and the rhombic triacontahedron.
Additionally, we shall find a remarkable near-miss, that is, a polyhedron which fails to be bipartite only by a tiny (but quantifiable) amount.
\section{Edge-transitive zonotopes}
\label{sec:zonotopes}
Recall the several definitions of a zonotope:
\begin{definition}
A \emph{zonotope} $Z\subset\RR^d$ is a polytope which satisfies any of the following properties:
\begin{myenumerate}
\item $Z$ is the projection of a $d$-cube,
\item $Z$ has only centrally symmetric faces,
\item $Z$ has only centrally symmetric 2-faces,
\item $Z$ can be written as the Minkowski sum of line segments.
\end{myenumerate}
\end{definition}
We give a full classification of the edge-transitive zonotopes, which will turn out to be the following:
\begin{theorem}\label{res:edge_transitive_zonotopes}
If $Z\subset\RR^d$ is an edge-transitive zonotope, then $Z$ is one of the following:
\begin{myenumerate}
\item a regular $2k$-gon,
\item a certain non-regular $4k$-gon,
\item a $d$-cube,
\item a $(2k,\...,2k)$-hyperprism, that is, the $k$-fold cartesian product of the regular $2k$-gon with itself,
\item the rhombic dodecahedron, or
\item the rhombic triacontahedron.
\end{myenumerate}
\end{theorem}
While the $(4,...,4)$-hyperprism is also just the $d$-cube for an appropriate \emph{even} $d\ge 2$, the $d$-cubes actually exist also in odd dimensions and are therefore listed separately.
The theorem follows from the following observation.
\begin{proposition}\label{res:edge_transitive_permutahedra}
If $P\subseteq\RR^d$ is an edge-transitive $\Gamma$-permutahedron, then $P$ is one of the following:
\begin{myenumerate}
\item a $d$-cube,
\item a $(2k,\...,2k)$-hyperprism.
\end{myenumerate}
\begin{proof}
\msays{TODO}
\end{proof}
\end{proposition}
\begin{proof}[Proof of \cref{res:edge_transitive_zonotopes}]
\msays{TODO}
\end{proof}
| 8df90945d4b58d9c358b26c306497b936cab84dc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this paper we explore the novel concept of deficiency. Deficiency is a characteristic unique to signed graphs from which arises a multitude of interesting questions, many of which are yet unanswered.
A \emph{signed graph} is a graph in which every edge has an associated sign. We write a signed graph $\Sigma$ as the triple $(V, E, \sigma)$ where $V$ is the vertex set, $E$ is the edge set, and $\sigma: E \to \{+,-\}$ is the \emph{signature}. Our graphs are \emph{signed simple graphs}: no loops and no multiple edges.
We define signed-graph coloring as in \cite{zaslavsky}, and chromatic number as in \cite{macajova}. A \emph{proper coloration} of a signed graph $\Sigma$ is a function, $\kappa : V \to \{\pm1, \pm 2, \ldots, \allowbreak \pm k, 0\},$ such that for any edge $e_{ab} \in E$, $\kappa(a) \neq \sigma(e) \kappa(b).$ The \emph{chromatic number} of $\Sigma$, written $\chi(\Sigma)$, is the size of the smallest set of colors which can be used to properly color $\Sigma.$ A graph with chromatic number $k$ is called \emph{k-chromatic}. A coloration is \emph{minimal} if it is proper and uses a set of colors of size $\chi(\Sigma).$ If $\chi = 2k,$ then a minimal \emph{color set} is $\{\pm1, \pm2, \ldots, \pm k\},$ and if $\chi = 2k+1,$ then a minimal color set is $\{\pm1, \pm2, \ldots, \pm k, 0\}.$ If $\chi = 2k+1$, there must be at least one vertex colored 0 in every minimal coloration. The \emph{deficiency} of a coloration, $\operatorname{def}(\kappa)$, is the number of unused colors from the color set of $\kappa.$ The \emph{deficiency set}, $\operatorname{D}(\kappa)$, is the set of unused colors.
In depictions of signed graphs we use solid lines for positive edges and dashed lines for negative edges.
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.7
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node (A) at (0,0) [acteur, label = below: 1]{};
\node (B) at (2,0) [acteur, label = below: $-1$]{};
\node (C) at (1,1.7) [acteur, label = above: 0]{};
\node (a) at (4,0) [acteur, label = below: 1]{};
\node (b) at (6,0) [acteur, label = below: 0]{};
\node (c) at (5,1.7) [acteur, label = above: 1]{};
\draw[thick] (A) to (B);
\draw[thick, dashed] (A) -- (C);
\draw[thick, dashed] (B) -- (C);
\draw[thick] (a) to (b);
\draw[thick, dashed] (a) to (c);
\draw[thick, dashed] (b) to (c);
\end{tikzpicture}
\caption[A 3-chromatic graph colored in two ways]{A 3-chromatic graph colored in two ways.}
\label{deficiency example}
\end{figure}
In Figure \ref{deficiency example} the coloration on the left has deficiency 0 while the coloration on the right has deficiency 1. The deficiency set of the coloration on the right is $\{-1\}.$
The \emph{maximum deficiency} of a graph, $\operatorname{M}(\Sigma) $, is $\max\{\operatorname{def}(\kappa) \mid \kappa$ is a minimal proper coloration of $\Sigma \}.$ The \emph{minimum deficiency}, $\operatorname{m}(\Sigma)$, is $\min\{\operatorname{def}(\kappa) \mid \kappa$ is a minimal proper coloration of $\Sigma\}.$ The \emph{deficiency range} of a graph, $\mathcal{L}(\Sigma)$, is $\{\operatorname{def}(\kappa) \mid \kappa$ s a minimal proper coloration of $\Sigma\}.$
The concept of deficiency arose when considering the chromatic number of joins of signed graphs.
\begin{thm}[\cite{mattern}]
Let $\Sigma_1$ and $\Sigma_2$ be signed graphs with maximum deficiencies $M_1$ and $M_2,$ respectively. Assume that $M_1 \geq M_2.$ Then, with one exception,
$$\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi(\Sigma_1 \vee_- \Sigma_2) = \max\{\chi_1 + \chi_2 - M_1 - M_2, \chi_1\}.$$
Exception: If $\Sigma_1$ and $\Sigma_2$ both have even chromatic number, exactly one of $M_1$ and $M_2$ is odd, and both $\Sigma_1$ and $\Sigma_2$ are exceptional graphs, then
$$\chi(\Sigma_1 \vee_+ \Sigma_2) = \chi(\Sigma_1 \vee_- \Sigma_2) = \max\{\chi_1 + \chi_2 - M_1 - M_2 +1, \chi_1\}.$$
\end{thm}
When considering the broad question of possible deficiencies of a signed graph, it makes sense to first focus on the maximum and minimum deficiencies. The simplest difficult case is that of 3-chromatic signed graphs. There are four decision problems about the maximum and minimum deficiencies of a 3-chromatic signed graph, where deficiency can only be 0 or 1.
\begin{enumerate}
\item Is the minimum deficiency 0?
\item Is the maximum deficiency 1?
\item Is the minimum deficiency 1?
\item Is the maximum deficiency 0?
\end{enumerate}
Questions 1 and 2 are clearly in class NP. Questions 3 and 4, on the other hand, are neither obviously in nor not in class NP. Furthermore, if the answer to either question 3 or 4 is "yes," then questions 1 and 2 are also solved. The main result of this paper is to show that questions 2 and 4 are both in class P by providing a polynomial-time algorithm for deciding the maximum deficiency of a 3-chromatic graph.
\section{Introductory Results}
\begin{thm}
\label{2-chromatic}
Let $\Sigma$ be a 2-chromatic signed graph. The deficiencies of $\Sigma$ can be classified as follows:
Case 1: $\operatorname{M}(\Sigma) = 1 = \operatorname{m}(\Sigma)$ if and only if $\Sigma$ is connected and all negative.
Case 2: $\operatorname{M}(\Sigma) = 0 = \operatorname{m}(\Sigma)$ if and only if $\Sigma$ contains a positive edge.
Case 3: $\operatorname{M}(\Sigma) = 1$ and $\operatorname{m}(\Sigma) = 0$ if and only if $\Sigma$ is disconnected and all negative.
\end{thm}
We leave the proof of Theorem \ref{2-chromatic} to the reader.
A subset of the vertices, $A \subseteq V$, is \emph{stable} if the subgraph induced by $A$ contains no edges.
\begin{proposition}
\label{bipartite and stable}
Let $\Sigma$ be a 3-chromatic signed graph. Then $\operatorname{M}(\Sigma) = 1$ if and only if $\Sigma^+$ is bipartite and there exists a bipartition of $V(\Sigma^+)$ with both parts stable in $\Sigma^+$ and one part stable in $\Sigma$.
\end{proposition}
\begin{proof}
Let $\Sigma$ be a 3-chromatic signed graph. Then there must exist at least one positive edge, meaning $\Sigma^+ \neq \emptyset.$
$(\Rightarrow)$ Suppose $\operatorname{M}(\Sigma) = 1.$ Then there exists a coloration of $\Sigma$ using only the colors 0 and 1. Thus, exactly two colors are used on $\Sigma^+$, meaning $\Sigma^+$ is bipartite. Furthermore, one part is colored 0, so it must be stable in $\Sigma.$
$(\Leftarrow)$ Suppose $\Sigma^+$ is bipartite and there exists a bipartition of $V(\Sigma^+)$ with both parts stable in $\Sigma^+$ and one part stable in $\Sigma$. Then color the part that is stable in $\Sigma$ with the color 0 and all other vertices in $\Sigma$ with the color 1. This coloration is proper since the 0-color set is stable and the 1-color set induces an all-negative subgraph. Thus $\operatorname{M}(\Sigma) =1.$
\end{proof}
\section{Maximum Deficiency Algorithm}
\label{s4.1}
A \emph{directed graph} is a pair $G = (V,A)$ where $V$ is the vertex set and $A$ is the multi-set of \emph{directed edges} consisting of ordered pairs of vertices. A \emph{cycle} is a directed path that starts and ends at the same vertex. The \emph{outdegree} of a vertex, the \emph{indegree} of a vertex, the minimum and maximum outdegree, and the minimum and maximum indegree are written $\delta^+(v),$ $\delta^-(v),$ $\delta^+(G),$ $\Delta^+(G)$, $\delta^-(G)$ and $\Delta^-(G)$, respectively.
\begin{lem}
If $G$ is a directed graph with $\delta^+(G) \geq 1$, then $G$ contains a cycle.
\label{digraph cycles}
\end{lem}
Before presenting the maximum deficiency algorithm, we rephrase Proposition \ref{bipartite and stable} to give an alternate way to think about maximum deficiency of a 3-chromatic signed graph. A \emph{vertex cover} of $F$ is a set of vertices of $\Sigma$ such that every edge in $F$ is incident with at least one vertex in the set. For $\alpha \in \{+,-\}$ we use the following two notations to represent specific subsets of the edge and vertex sets: $E^\alpha = \{e \in E \mid \sigma(e) = \alpha\}$ and $V^\alpha = \{v \in V \mid \text{there exists an edge of sign $\alpha$ incident with } v\}.$
\begin{lem}
For a 3-chromatic signed graph $\Sigma$, $\operatorname{M}(\Sigma) = 1$ if and only if there exists a vertex cover of $E^+$ that is stable in $\Sigma$.
\label{vertex cover}
\end{lem}
\subsection{MaxDef}
Given a 3-chromatic, not necessarily simple, signed graph, $\Sigma,$ we provide an algorithm that decides whether the maximum deficiency of $\Sigma$ is 1 or 0. We call the algorithm MaxDef. We provide a worked example of MaxDef in section \ref{ss4.2.3}.
The input for MaxDef is a 3-chromatic signed graph. We assume the graph is input as a pair of adjacency lists; one array of lists contains a list of the positive adjacencies for each vertex, and the other array of lists does the same for the negative adjacencies. Note that there must be at least one positive edge because the graph is 3-chromatic. The output of MaxDef is the maximum deficiency of the graph, either a 0 or a 1, along with a stable vertex cover of $E^+$ if the maximum deficiency is 1.
We build a partial stable cover throughout this process and end either in the creation of a stable cover of $E^+$ or in the non-existence of such a cover. In order to produce a stable cover, if one exists, we store recovery information about the identification of vertices. We use the symbol $\rightarrow$ to indicate replacement. For example, $A \cup B \rightarrow A$ means replace object $A$ with object $A \cup B.$ We use $N_-(v)$ to indicate the set of vertices of $\Sigma$ that are negatively adjacent to $v.$ Let $S$ be the partial stable cover, starting with $S = \emptyset.$ Let $B$ be a list of vertices that are forbidden from being in $S,$ starting with $B = \emptyset.$
{\bf Step 1:} Delete all vertices of $\Sigma$ not in $V^+.$ Every vertex incident to only negative edges does not affect the existence of a stable cover of $E^+$, and therefore, by Lemma \ref{vertex cover}, the existence of a deficiency-1 coloration. Call this set of vertices $A.$ Then let $\Sigma_{01} = \Sigma \setminus A.$
{\bf Step 2:} Let $H_1, H_2, \ldots, H_r$ be the connected components of $\Sigma^+_{01}$. For each connected component of $\Sigma^+_{01},$ check to see if it is bipartite.
If $H_i$ is not bipartite, then $\operatorname{M}(\Sigma_{01}) = 0$ and thus, $\operatorname{M}(\Sigma) = 0$ and the algorithm ends here.
If $H_i$ is bipartite, let its unique vertex bipartition have parts $A_i$ and $B_i$. Start to create a new graph from $\Sigma_{01}$ by collapsing $A_i$ to a vertex $a_i$ and each $B_i$ to a vertex $b_i$ and eliminating multiple edges of the same sign. This makes $H_i$ into a positive edge with endpoints $a_i$ and $b_i.$ If there exists a negative edge $a_ib_i$, delete it.
If all $H_i$ are bipartite, then at the end of Step 2 we have a flattened graph, call it $\Sigma_{02}$, composed of a positive perfect matching with all other edges negative. For a matched pair $(a_i, b_i)$ we use $x_i$ to represent one vertex of the pair and $\bar x_i$ to represent the other.
\begin{lem}
\label{1-1}
There is a one-to-one correspondence between stable covers of $E^+$ and stable covers of $E^+_{02}.$
\end{lem}
\begin{proof}
For each $i$, every stable cover of $E^+$ uses \textit{all} the vertices of either $A_i$ or $B_i.$ This corresponds to having exactly one vertex of the pair $(a_i, b_i)$ in every stable cover of $E^+_{02}$. In addition, the removal of any negative $x_i \bar x_i$ edge does not affect the 0/1 outcome of MaxDef.
\end{proof}
Steps 1 and 2 are executed only once during the algorithm. The remainder of the steps are part of a recursive process and may be executed multiple times. Let $\Sigma_1 = \Sigma_{02}.$
{\bf Step 3:} Look for an $i$ such that $x_i$ and $\bar{x}_i$ are both in $B.$ If such an $i$ exists, then $\operatorname{M}(\Sigma) = 0.$ If no such $i$ exists, then continue to Step 4.
{\bf Step 4:} Look for an $i$ such that $x_i$ is in $B.$ If such an $i$ exists, then $S \cup \{\bar x_i\} \rightarrow S,$ $(B \cup N_-(\bar x_i)) \setminus \{x_i\} \rightarrow B,$ and $\Sigma_1 \setminus \{x_i, \bar x_i\} \rightarrow \Sigma_1.$ Return to Step 3. If no such $i$ exists, then continue to Step 5.
{\bf Step 5:} Look for an $i$ such that $x_i$ and $\bar x_i$ both have loops. If such an $i$ exists, then $\operatorname{M}(\Sigma) = 0.$ If no such $i$ exists, continue to Step 6.
{\bf Step 6:} Look for an $i$ such that $x_i$ has a loop. If such an $i$ exists, then $S \cup \{\bar x_i \} \rightarrow S,$ $B \cup N_-(\bar x_i) \rightarrow B$, and $\Sigma_1 \setminus \{x_i, \bar x_i\} \rightarrow \Sigma_1.$ Return to Step 3. If no such $i$ exists, continue to Step 7.
{\bf Step 7:} Look for an $i$ such that $x_i$ is adjacent to both $x_j$ and $\bar x_j$. If such an $i$ exists, then $S \cup \{\bar x_i\} \rightarrow S,$ $B \cup N_-(\bar x_i) \rightarrow B,$ and $\Sigma_1 \setminus \{x_i, \bar x_i\} \rightarrow \Sigma_1.$ Return to Step 3. If no such $i$ exists, then continue to Step 8.
{\bf Step 8:} Look for an $i$ and $j,$ $i \neq j,$ such that $x_i$ is adjacent to $x_j$ and $\bar x_i$ is adjacent to $\bar x_j$. If such an $i$ and $j$ exist, then identify vertex $x_i$ to $\bar x_j$, and vertex $\bar x_i$ to $x_j.$ We are identifying vertices that must always be in a stable cover together. So $\Sigma_1$ (with the above stated identifications) $\rightarrow \Sigma_1.$ Delete negative edge $x_i \bar x_i$, and return to Step 3. If no such $i$ and $j$ exist, then continue to Step 9.
{\bf Step 9:} Look for a vertex of degree one. If there exists such a vertex, call it $x_i,$ then $S \cup \{x_i\} \rightarrow S,$ $B \cup N_-(x_i) \rightarrow B,$ and $\Sigma_1 \setminus \{x_i, \bar x_i\} \rightarrow \Sigma_1.$ If both $x_i$ and $\bar x_i$ are vertices of degree 1, then we only place one into $S$; it doesn't matter which one. Return to Step 3. If no such vertex exists, then continue to Step 10.
Step 9 is the only step where the vertex we add to $S$ is not a forced addition. We show that we can add it to $S$ without affecting the 0/1 outcome of MaxDef.
\begin{lem} Let $x_i$ be a vertex of degree one not in $B$. Then there exists a stable cover of $E_1^+$ extending $S$ if and only if there exists a stable cover extending $S$ which contains $x_i.$
\end{lem}
\begin{proof} Suppose $\Sigma_1$ has a vertex of degree one that is not in $B$. Call this vertex $x_i.$ Recall that $E_1^+$ is a perfect matching. Suppose $S_1$ is a stable cover of $E_1^+$ that extends $S.$ If $S_1$ contains $x_i$, then the lemma holds. So suppose $S_1$ does not contain $x_i;$ then $S_1$ must contain $\bar x_i.$ Thus, $S_1$ is a stable cover of $E_1^+$ if and only if $S_1' = S_1 \setminus \{\bar x_1\}$ is a stable cover of $E^+$ in $\Sigma_1 \setminus \{x_i, \bar x_i\}.$ Because $x_i$ has degree one and is not in $B$, it has no neighbors in $S_1'.$ Therefore, $S_1'$ is stable in $\Sigma_1 \setminus \{x_i, \bar x_i\}$ if and only if $S_2 = S_1' \cup \{x_i\}$ is stable in $\Sigma_1.$ Moreover, since $x_i$ covers edge $x_i \bar x_i,$ $S_1'$ is a cover of $E^+$ in $\Sigma_1 \setminus \{x_i, \bar x_i\}$ if and only if $S_2$ is a cover of $E_1^+$.
\end{proof}
{\bf Step 10:} If $\Sigma_1$ is the empty graph, then $\operatorname{M}(\Sigma) = 1.$ Use the recovery information to produce a stable cover of $E^+$. If $\Sigma_1$ is not the empty graph, then continue to Step 11.
Upon reaching Step 11, $\Sigma_1$ is simple, is composed of a positive perfect matching with all other edges negative, contains no loops and no vertices of degree 1, and has at most one edge between every two matched pairs $(a_i, b_i)$ and $(a_j, b_j).$
{\bf Step 11:} Create the \emph{forcing graph} of $\Sigma_1$, called $F(\Sigma_1)$. We define $F(\Sigma_1)$ as a directed graph with $V = V_1$ and $E = \{x_ix_j \mid x_i\bar x_j \in E_1^-\}.$ Every edge in $F(\Sigma_1)$ represents a forced decision about which vertices of $\Sigma_1$ must appear together in every stable cover of $E^+_1.$ For example, if $x_ix_j \in E$, then $x_i\bar x_j \in E_1^-$. This means that if $x_i$ is in a stable cover of $E^+_1$, then $\bar x_j$ cannot be, and so $x_j$ must also be in the stable cover.
Observe that $x_ix_j \in E$ if and only if $\bar x_j\bar x_i \in E.$ Also, $\delta_+(F(\Sigma_1)) \geq 1,$ because there are no vertices of degree one in $\Sigma_1.$ Thus, there must exist a cycle in $F(\Sigma_1).$ By our first observation, there must in fact be two cycles, possibly not disjoint; if one cycle uses vertices $x_1, \ldots, x_r,$ then the other must use vertices $\bar x_1, \ldots, \bar x_r.$ Choosing a single vertex for the stable cover from a cycle forces the rest of the vertices in the cycle to also be chosen.
{\bf Step 12:} If the cycles found in Step 11 are not disjoint, then $\operatorname{M}(\Sigma) = 0.$ If they are, identify the vertices of $\Sigma_1$ that belong to each of the cycles from the pair found in Step 11. Eliminate multiple edges of the same sign and delete any negative $x_i \bar x_i$ edge. Then $\Sigma_1$ (with the above identifications) $\rightarrow \Sigma_1$ and return to Step 3.
This ends the MaxDef algorithm.
\begin{thm}
The MaxDef algorithm correctly decides the maximum deficiency of $\Sigma$ and is a finite process.
\end{thm}
\begin{proof}
One possible way for MaxDef to end is with $\operatorname{M}(\Sigma) = 0$ in Steps 2, 3, 5, and 12. The other possibility is to end with $\operatorname{M}(\Sigma) =1$ in Step 10. If the graph does not satisfy the ending conditions of Steps 2, 3, 5, or 12, at least one of the following Steps must be completed: 4, 6, 7, 8, 9, 10, or 12. In fact, Step 12 is not reached unless the graph does not satisfy the conditions of Steps 3--10. In the completion of Steps 4, 6, 7, 8, 9, or 12, the graph is reduced by at least one matched pair. Thus, since our graph is finite, if it never satisfies the ending conditions of Steps 2, 3, 5, or 12, it must eventually be an empty graph and satisfy the ending condition of Step 10.
Now we show that MaxDef correctly decides the maximum deficiency. We first prove if MaxDef returns 1, then $\operatorname{M}(\Sigma) = 1.$
Suppose MaxDef returns 1. Then MaxDef ends with Step 10. Let $S$ be the set of vertices produced, and let $S'$ be $S$ restricted to $\Sigma_{02}$. Observe that $S'$ is stable in $\Sigma_{02}.$ Indeed, any vertex added to the partial stable cover (through Steps 4, 6, 8, and 9) was guaranteed by Steps 3 and 4 to be allowed in the partial stable cover. Furthermore, any vertices that were identified through Steps 8 and 12 were a stable set, as guaranteed by Step 7 and the definition of the forcing graph. Also observe that $S'$ is a cover of $E^+_{02}$ since exactly one of every matched pair is present in $S'$ by definition of the algorithm. By Lemma \ref{1-1}, if $S'$ is a stable cover of $E^+_{02}$, then $S$ is a stable cover of $E^+_{01}$. Thus, $S$ is a stable cover of $E^+$ since $E^+ = E^+_{01},$ and $\Sigma_{01}$ is an induced subgraph of $\Sigma.$ Therefore, by Lemma \ref{vertex cover} $\operatorname{M}(\Sigma) = 1.$
Suppose MaxDef returns 0. Then the algorithm ended with Step 2, 3, 5, or 12.
Suppose MaxDef ended with Step 2. Then there exists a connected component of $\Sigma_{01}^+$ that is not bipartite. Thus, it is also a connected component of $\Sigma^+$ that is not bipartite. By Lemma \ref{bipartite and stable}, $\operatorname{M}(\Sigma) = 0.$
Suppose MaxDef ended with Step 3. Then at some point in the algorithm, there exists a matched pair of vertices that were both in $B$. Since all additions to the partial stable set are either forced or of degree one---and therefore cannot be a neighbor of either vertices in the matched pair---this means a stable cover of $E^+_1$ does not exist. Thus, a stable cover of $E^+_{01}$, and therefore of $E^+,$ does not exist.
Suppose MaxDef ended with Step 5. Then after completing either Step 2 or Step 12,
there exists a matched pair of vertices that both have loops. Supposed the matched pair is $(a_1, b_1).$ Loops in $\Sigma_1$ directly after Step 2 come from either pre-existing loops, or internal edges in $A_1$ and $B_1$. Loops in $\Sigma_1$ after completing Step 12 come from sets of vertices with internal negative edges being identified. In either case, choosing either $a_1$ or $b_1$ results in choosing an unstable set of vertices for the stable cover.
Finally, suppose MaxDef ended with Step 12. Then there existed a pair of cycles in the forcing graph that were not disjoint. Thus, choosing a single vertex from either cycle forces the rest of the vertices in both cycles to be chosen. Therefore, a matched pair, $x_i$ and $\bar x_i$, must both be chosen and it is impossible to create a stable cover.
\end{proof}
\subsection{Example of MaxDef}
\label{ss4.2.3}
Let the following graph be our original $\Sigma.$ We start with $S = \emptyset,$ $B = \emptyset,$ and a recovery array $R = \emptyset.$
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (0,0) [acteur, label = below: $a_1$]{};
\node at (0,2) [acteur, label = above: $b_1$]{};
\node at (2,0) [acteur, label = below: $a_2$]{};
\node at (2,2) [acteur, label = left: $b_2$]{};
\node at (4,0) [acteur, label = below: $a_3$]{};
\node at (4,2) [acteur, label = right: $b_3$]{};
\node at (6,0) [acteur, label = below: $a_4$]{};
\node at (6,2) [acteur, label = above: $b_4$]{};
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\node at (10,0) [acteur, label = below: $a_6$]{};
\node at (10,2) [acteur, label = above: $b_6$]{};
\node at (12,0) [acteur, label = below: $a_7$]{};
\node at (12,2) [acteur, label = above: $b_7$]{};
\draw[thick] (0,0) -- (0,2);
\draw[thick] (2,0) -- (2,2);
\draw[thick] (4,0) -- (4,2);
\draw[thick] (6,0) -- (6,2);
\draw[thick] (8,0) -- (8,2);
\draw[thick] (10,0) -- (10,2);
\draw[thick] (12,0) -- (12,2);
\draw[thick, dashed] (0,0) -- (2,0);
\draw[thick, dashed] (0,2) to[bend left] (6,2);
\draw[thick, dashed] (2,0) -- (8,2);
\draw[thick, dashed] (2,2) -- (4,2);
\draw[thick, dashed] (2,2) to[bend left] (6,2);
\draw[thick, dashed] (4,0) -- (6,0);
\draw[thick, dashed] (6,2) -- (8,0);
\draw[thick, dashed] (8,2) -- (10,2);
\draw[thick, dashed] (10,0) -- (12,2);
\draw[thick, dashed] (10,2) -- (12,0);
\end{tikzpicture}
\caption[$\Sigma$ is the starting graph for our example]{$\Sigma$ is a 3-chromatic graph.}
\label{original}
\end{figure}
Because $\Sigma$ is composed of a positive perfect matching with all other edges negative we skip Steps 1 and 2. At this point $\Sigma_1 = \Sigma$, $S = \emptyset, B = \emptyset,$ and $R$ has an empty list for each vertex.
{\bf Step 8:} Using Step 8 we can identify vertices $b_6$ and $b_{7}$, calling this vertex $b_6$ and vertices $a_6$ and $a_{7},$ calling this vertex $a_6.$ We update $R$ to get the table below.
{\centering
\begin{tabular}{| c | c | c | c | c | c | c | c | c | c | c | c | c | c |}
\hline
$a_1$ & $b_1$ & $a_2$ & $b_2$ & $a_3$ & $b_3$ & $a_4$ & $b_4$ & $a_5$ & $b_5$ &
$a_6$ & $b_6$ \\ \hline
$\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{\}$ & $\{a_7\}$ & $\{b_7\}$ \\ \hline
\end{tabular} \par
}
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (0,0) [acteur, label = below: $a_1$]{};
\node at (0,2) [acteur, label = above: $b_1$]{};
\node at (2,0) [acteur, label = below: $a_2$]{};
\node at (2,2) [acteur, label = left: $b_2$]{};
\node at (4,0) [acteur, label = below: $a_3$]{};
\node at (4,2) [acteur, label = right: $b_3$]{};
\node at (6,0) [acteur, label = below: $a_4$]{};
\node at (6,2) [acteur, label = above: $b_4$]{};
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\node at (10,0) [acteur, label = below: $a_6$]{};
\node at (10,2) [acteur, label = above: $b_6$]{};
\draw[thick] (0,0) -- (0,2);
\draw[thick] (2,0) -- (2,2);
\draw[thick] (4,0) -- (4,2);
\draw[thick] (6,0) -- (6,2);
\draw[thick] (8,0) -- (8,2);
\draw[thick] (10,0) -- (10,2);
\draw[thick, dashed] (0,0) -- (2,0);
\draw[thick, dashed] (0,2) to[bend left] (6,2);
\draw[thick, dashed] (2,0) -- (8,2);
\draw[thick, dashed] (2,2) -- (4,2);
\draw[thick, dashed] (2,2) to[bend left] (6,2);
\draw[thick, dashed] (4,0) -- (6,0);
\draw[thick, dashed] (6,2) -- (8,0);
\draw[thick, dashed] (8,2) -- (10,2);
\end{tikzpicture}
\caption[$\Sigma_1$ is the starting graph for our example]{$\Sigma_1$ after executing Step 8.}
\label{original}
\end{figure}
{\bf Step 9:} Vertex $a_6$ has degree one, so we update $S$ to be $S = \{a_6\}.$
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=. 6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (0,0) [acteur, label = below: $a_1$]{};
\node at (0,2) [acteur, label = above: $b_1$]{};
\node at (2,0) [acteur, label = below: $a_2$]{};
\node at (2,2) [acteur, label = left: $b_2$]{};
\node at (4,0) [acteur, label = below: $a_3$]{};
\node at (4,2) [acteur, label = right: $b_3$]{};
\node at (6,0) [acteur, label = below: $a_4$]{};
\node at (6,2) [acteur, label = above: $b_4$]{};
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\draw[thick] (0,0) -- (0,2);
\draw[thick] (2,0) -- (2,2);
\draw[thick] (4,0) -- (4,2);
\draw[thick] (6,0) -- (6,2);
\draw[thick] (8,0) -- (8,2);
\draw[thick, dashed] (0,0) -- (2,0);
\draw[thick, dashed] (0,2) to[bend left] (6,2);
\draw[thick, dashed] (2,0) -- (8,2);
\draw[thick, dashed] (2,2) -- (4,2);
\draw[thick, dashed] (2,2) to[bend left] (6,2);
\draw[thick, dashed] (4,0) -- (6,0);
\draw[thick, dashed] (6,2) -- (8,0);
\end{tikzpicture}
\caption[$\Sigma_1$ is the starting graph for our example]{$\Sigma_1$ after executing Step 9.}
\label{original}
\end{figure}
{\bf Step 11:} Consider the forcing graph of $\Sigma_1.$
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\tikzset{edger/.style={decoration={
markings,
mark=at position #1 with {\arrow{angle 45}}},postaction={decorate}, color = red}}\tikzset{edgeb/.style={decoration={
markings,
mark=at position #1 with {\arrow{angle 45}}},postaction={decorate}, color = blue}}
\tikzset{edge/.style={decoration={
markings,
mark=at position #1 with {\arrow{angle 45}}},postaction={decorate}}}
\node at (0,0) [acteur, label = right: $a_1$]{};
\node at (0,2) [acteur, label = left: $b_1$]{};
\node at (2,0) [acteur, label = below: $a_2$]{};
\node at (2,2) [acteur, label = left: $b_2$]{};
\node at (4,0) [acteur, label = right: $a_3$]{};
\node at (4,2) [acteur, label = right: $b_3$]{};
\node at (6,0) [acteur, label = above: $a_4$]{};
\node at (6,2) [acteur, label = below: $b_4$]{};
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\draw [edge=.8] (0,0) to (2,2);
\draw [edge=.8] (2,0) to (0,2);
\draw [edge=.7] (6,2) to[bend right = 90, looseness = 1.3] (0,0);
\draw [edge=.7] (0,2) to[bend right = 90, looseness = 1.3] (6,0);
\draw [edge=.85] (2,2) to (4,0);
\draw [edge=.8] (4,2) to (2,0);
\draw [edge=.5] (2,0) to[bend right] (8,0);
\draw [edge=.5] (8,2) to[bend right] (2,2);
\draw [edge=.8] (2,2) to (6,0);
\draw [edge=.45] (6,2) to (2,0);
\draw [edge=.7] (4,0) to (6,2);
\draw [edge=.45] (6,0) to (4,2);
\draw [edge=.6] (6,2) to (8,2);
\draw [edge=.6] (8,0) to (6,0);
\end{tikzpicture}
\caption[The forcing graph of $\Sigma_1$]{The forcing graph of $\Sigma_1$.}
\label{forcing graph 1}
\end{figure}
{\bf Step 12:} Identify the vertices, $a_1, b_2, a_3,$ and $b_4,$ and call this vertex $a_1.$ Identify the vertices, $a_4, b_3, a_2,$ and $b_1,$ and call this vertex $b_1.$ Our recovery array is updated to be the following:
{\centering
\begin{tabular}{| c | c | c | c | c | c |}
\hline
$a_1$ & $b_1$ & $a_5$ & $b_5$ & $a_6$ & $b_6$ \\ \hline
$\{b_2,a_3,b_4\}$ & $\{a_2,b_3,a_4\}$ & $\{\}$ & $\{\}$ & $\{a_7\}$ & $\{b_7\}$ \\ \hline
\end{tabular} \par}
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (6,0) [acteur, label = below: $a_1.$]{};
\node at (6,2) [acteur, label = above: $b_1$]{};
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\draw[thick] (6,0) -- (6,2);
\draw[thick] (8,0) -- (8,2);
\draw[thick, dashed] (6,0) to[in= 180, out =220, loop] (6,0);
\draw[thick, dashed] (6,0) -- (8,0);
\draw[thick, dashed] (6,2) -- (8,2);
\end{tikzpicture}
\caption[After identifying the cycles of the forcing graph]{$\Sigma_1$ is the result of identifying the vertices in each of the two cycles.}
\label{sigma 1}
\end{figure}
{\bf Step 6:} Vertex $b_1$ must be in every stable cover since its matched partner has a loop. Now $S = \{a_6, b_1\}$ and $B = \{b_5\}.$
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (8,0) [acteur, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\draw[thick] (8,0) -- (8,2);
\end{tikzpicture}
\caption[Step 6 example]{$\Sigma_1$ after removing vertices $a_1$ and $b_1$ with Step 6.}
\label{sigma 1.1}
\end{figure}
{\bf Step 4:} Vertex $b_5$ is in $B$. Thus, by Step 4, vertex $a_5$ must be in $S.$ So $S = \{a_6, b_1, a_5\}$ and $B = \emptyset.$
{\bf Step 10:} Our graph has no more vertices. Thus, $\operatorname{M}(\Sigma) = 1$. If we use $R$ to expand the vertices of $S$ into their original pre-identified vertices we get $S = \{a_6, b_1, a_5\} = \{a_6, a_7, a_4, b_3, a_2, b_1, a_5\}.$ Figure \ref{stable cover} shows the resulting stable cover of $E^+$.
\begin{figure}[H]
\centering
\begin{tikzpicture}[
thick,
acteur/.style={
circle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
},
sq/.style={
rectangle,
fill=black,
thick,
inner sep=2pt,
minimum size=0.2cm
}, scale=.6
]
\tikzset{every loop/.style={min distance=12mm,looseness=12}}
\node at (0,0) [acteur, label = below: $a_1$]{};
\node at (0,2) [sq, label = above: $b_1$]{};
\node at (2,0) [sq, label = below: $a_2$]{};
\node at (2,2) [acteur, label = left: $b_2$]{};
\node at (4,0) [acteur, label = below: $a_3$]{};
\node at (4,2) [sq, label = right: $b_3$]{};
\node at (6,0) [sq, label = below: $a_4$]{};
\node at (6,2) [acteur, label = above: $b_4$]{};
\node at (8,0) [sq, label = below: $a_5$]{};
\node at (8,2) [acteur, label = above: $b_5$]{};
\node at (10,0) [sq, label = below: $a_6$]{};
\node at (10,2) [acteur, label = above: $b_6$]{};
\node at (12,0) [sq, label = below: $a_7$]{};
\node at (12,2) [acteur, label = above: $b_7$]{};
\draw[thick] (0,0) -- (0,2);
\draw[thick] (2,0) -- (2,2);
\draw[thick] (4,0) -- (4,2);
\draw[thick] (6,0) -- (6,2);
\draw[thick] (8,0) -- (8,2);
\draw[thick] (10,0) -- (10,2);
\draw[thick] (12,0) -- (12,2);
\draw[thick, dashed] (0,0) -- (2,0);
\draw[thick, dashed] (0,2) to[bend left] (6,2);
\draw[thick, dashed] (2,0) -- (8,2);
\draw[thick, dashed] (2,2) -- (4,2);
\draw[thick, dashed] (2,2) to[bend left] (6,2);
\draw[thick, dashed] (4,0) -- (6,0);
\draw[thick, dashed] (6,2) -- (8,0);
\draw[thick, dashed] (8,2) -- (10,2);
\draw[thick, dashed] (10,0) -- (12,2);
\draw[thick, dashed] (10,2) -- (12,0);
\end{tikzpicture}
\caption[The resulting stable cover]{The square vertices of $\Sigma$ make up the stable cover $\{b_1, a_2, b_3, a_4, a_5, a_6, a_7\}$ of $E^+$.}
\label{stable cover}
\end{figure}
\section{Complexity}
\subsection{Data Format}
Our algorithm requires five different objects to be stored. First, we store the signed graph using adjacency lists. The adjacency lists will be two arrays of lists, one for positive edges and one for negative edges. Since each edge of $\Sigma$ appears twice in these arrays, the size of these adjacency lists is at its largest $2E.$ After Step 2, we no longer need the positive adjacency lists and stop updating them.
Second, we need to store the current forcing graph adjacency lists. Every negative edge in our signed graph gives rise to two directed edges. Therefore the size of these adjacency lists is at its largest $2E^-.$
Third, we need to store the partial stable set $S$. This is stored simply as a list of size at most $\frac{1}{2}V.$
Fourth, we need a list of vertices forbidden from the partial stable set $S.$ This is stored simply as a list of size at most $V-1$.
Finally, we store recovery information in order to produce a stable cover of the original graph if the maximum deficiency is 1. The recovery information is stored as an array of lists, one list for each group of identified vertices, and has size at most $V$.
\subsection{Time Complexity}
\begin{thm}
\label{time complexity}
If $\Sigma$ is without multiple edges of the same sign, then MaxDef has running time $\mathcal O(V^5).$
\end{thm}
\subsubsection{Updating Lists}
Suppose vertex $x_i$ is added to the partial stable set. Then the sets $S$ and $B$ and the adjacency lists of our current version of $\Sigma$ need to be updated. This happens in several steps of the algorithm, so we calculate the time complexity of updating these three objects separately.
In order to update $S,$ we simply append element $x_i$ to $S.$ This takes constant time.
We update the adjacency lists and $B$ together. We look through the negative adjacency lists for both $x_i$ and $\bar x_i.$ Looking through the lists takes time $\mathcal O(E^-).$ We delete every occurrence of $\bar x_i.$ We need to delete vertex $\bar x_i$ at most $V-1$ times so deleting $\bar x_i$ from the adjacency lists takes time $\mathcal O(V).$ We delete every occurrence of $x_i$ and add the vertices in whose lists $x_i$ appeared to $B$. Since we need to delete $x_i$ at most $V-1$ times, the time complexity of deleting and updating $B$ is $\mathcal O(V).$
Finally, we need to possibly delete vertex $\bar x_i$ from $B.$ This requires looking through $B$, which takes time $\mathcal O(V),$ and then removing $\bar x_i$ if it appears, which takes constant time.
Therefore, the total time complexity for updating all three objects is $\mathcal O(V + E^-).$
\subsubsection{Steps 1 Through 12}
{\bf Step 1:} We find vertices that are not in $V^+$ by looking for empty lists in the positive edge adjacency lists. Finding these vertices takes time $\mathcal O(V).$
In order to delete vertices from $\Sigma,$ we first remove the lists associated with those vertices. Deleting a list takes constant time and we delete at most $V-2$ lists; thus, deleting the full lists takes time $\mathcal O(V).$ We then delete any appearance of the vertices themselves in an adjacency list. Since the vertices being deleted only have negative edges, and there are at worst $V-2$ vertices being deleted, the time complexity is $\mathcal O(VE^-).$ Therefore the total time complexity for Step 1 is $\mathcal O(VE^-).$
{\bf Step 2:} In order to decide whether a given $H_i$ is bipartite, we can employ a depth-first search (DFS) algorithm. The DFS algorithm assigns a color to a vertex that is different than the color of its parent in the depth-first search tree. For an $H_i$ with $s$ vertices and $t$ edges, this algorithm takes time $\mathcal O(s + t),$ \cite{Cormen}. Therefore, checking all of the $H_i$ gives us a total time complexity of $\mathcal O(V + E^+).$
To collapse an $H_i$, we use the bipartition $A_i,$ $B_i$ from the DFS algorithm. Suppose $A_i$ is composed of vertices $v_1, \ldots, v_r$ and $B_i$ is composed of vertices $v_{r+1}, \ldots, v_{r+s}.$ Then we create a list for $a_i$ and a list for $b_i$ by moving every element in the negative adjacency lists of vertices $v_1, \ldots, v_r$ to the negative adjacency list for $a_i$ and every element in the negative adjacency lists of vertices $v_{r+1}, \ldots, v_{r+s}$ to the negative adjacency list for $b_i.$ Moving an element takes constant time, and we do it at most $2E^-$ times.
We delete the negative edge $a_ib_i,$ if it exists, and remove duplicate elements from the $a_i$ and $b_i$ lists. This requires looking though the lists at most $2E^-$ times, once for each element of the lists, for a total time complexity of $\mathcal O((E^-)^2).$
Finally, we initialize the recovery array with enough positions for each $a_i$ and $b_i$, then add the lists for two new elements, one for vertex $a_i$ and one for vertex $b_i.$ Because each $H_i$ results in two lists for the recovery array, we put the list of vertices from $A_i$ in position $2i-1$ and the list of vertices from $B_i$ in position $2i.$ This takes constant time.
Thus, collapsing one $H_i$ takes time $\mathcal O((E^-)^2).$ We must do this process at most $V/2$ times, as each $H_i$ has at least 2 vertices. Thus, the entire collapsing process takes time $\mathcal O(V(E^-)^2).$ Therefore, the total time complexity of Step 2 is $\mathcal O(V(E^-)^2 + V + E^+).$
{\bf Step 3:} We look through $B$ once for each element of $B$. Thus, Step 3 takes time $\mathcal O(V^2).$
{\bf Step 4:} We check to see if $B$ is empty; this takes constant time. If $B$ is not empty, we add a new vertex to $S$ and update the various lists. As shown above, updating the lists is of time complexity $\mathcal O(V + E^-).$ Therefore, Step 4 takes time $\mathcal O(V + E^-).$
{\bf Step 5:} We look through the negative adjacency lists to find a matched pair of lists that both contain their own vertex. This takes time $\mathcal O(E^-).$
{\bf Step 6:} We look through the negative adjacency lists to find a list that contains its own vertex. This has time complexity $\mathcal O(E^-).$ If we find such a vertex, we must update the lists; this takes time $\mathcal O(V + E^-).$ Therefore, the total time complexity of Step 6 is $\mathcal O(V + E^-).$
{\bf Step 7:} We look through each negative adjacency list to see if it contains an $x_i, \bar x_i$ pair. This takes time $\mathcal O(V^2)$. Since we must do this for each vertex, the total time complexity for checking the lists is $\mathcal O(V^3).$ Then we must update the various lists, taking time $\mathcal(V + E^-).$ Therefore, the total time complexity of Step 7 is $\mathcal O(V^3 + E^-).$
{\bf Step 8:} We look through the negative adjacency lists once for each element of the negative adjacency lists. Thus, the time complexity of finding such vertices is $\mathcal O((E^-)^2).$
If we find such vertices, we must identify them. Suppose we are identifying vertices $x_i$ and $\bar x_j$ and vertices $\bar x_i$ and $x_j.$ We describe the process for identifying $x_i$ and $\bar x_j.$ Without loss of generality, assume $i < j.$
First, we combine the negative adjacency lists for $x_i$ and $\bar x_j.$ Concatenating the two lists takes time at most $\mathcal O(V)$. We then need to delete edge $x_i \bar x_i$ and remove duplicate elements from the newly combined lists, which as shown in Step 2 of this proof takes time $\mathcal O((E^-)^2).$
Second, we change every occurrence of $\bar x_j$ in an adjacency list to $x_i.$ Changing the element takes constant time, and looking through the adjacency lists has time complexity $\mathcal O(E^-).$
Finally, we update the recovery information. This requires concatenating the $\bar x_j$ list to the $x_i$ list, and removing the $\bar x_j$ list. Concatenating takes time at most $\mathcal O(V)$, and deleting the $\bar x_i$ list takes constant time.
Therefore, identifying vertex $x_i$ to vertex $\bar x_j$ has total time complexity $\mathcal O((V + E^-)^2).$ We must do this process twice, once for each identification. Therefore, Step 8 has total time complexity of $\mathcal O(V + (E^-)^2).$
{\bf Step 9:} This step again requires looking through the negative adjacency lists and then updating. Therefore, the time complexity of Step 9 is $\mathcal O(V + E^-).$
{\bf Step 10:} This step requires looking through the negative adjacency lists, so takes time $\mathcal O(E^-).$ If the current version of $\Sigma$ is empty, we use the recovery information to produce a stable cover of the original graph. This requires looking through $S,$ and then for each element of $S$, looking through the recovery array. Since $S$ is size at most $\frac{1}{2}V$ and the recovery array is size at most $V,$ producing a stable cover takes time $\mathcal O(V^2).$ Therefore, Step 10 takes time $\mathcal O(V^2 + E^-).$
{\bf Step 11:} To create the forcing graph, we must look through the negative adjacency lists and create a new set of adjacency lists. For every endpoint $x_i$ in the adjacency list of vertex $v_1$ we add the vertex $\bar x_i$ to the adjacency list of $v_1$ in the forcing graph. The addition of a new element to the forcing graph adjacency lists takes constant time, and there are $2E^-$ elements of the original adjacency lists. Therefore, Step 11 takes time $\mathcal O(E^-).$
{\bf Step 12:} For this step, we first find a cycle---recall the cycles come in pairs, so we need only look for one cycle. This can be done with a DFS algorithm, so takes time $\mathcal O(V + E^-)$ \cite{Cormen}. Let the cycles found be $C_1$ and $C_2.$
We then check to see whether $C_1$ and $C_2$ are disjoint. This requires reading through the list of vertices in $C_2$, once for each vertex in $C_1.$ This takes time $\mathcal O(V^2).$
If the cycles are disjoint, we identify the vertices of each cycle. Let one cycle be made up of vertices $x_1, x_2, \ldots, x_r.$ We first take all the elements of the negative adjacency lists of vertices $x_2, \ldots, x_r$ and add them to the negative adjacency list for vertex $x_1.$ Each addition takes constant time, and there are at most $E^- - 2$ additions. Next, we delete the empty lists for vertices $x_2, \ldots, x_r.$ Each deletion again takes constant time, and there are at most $V-2$ deletions. We then look through all the negative adjacency lists to change every appearance of $x_2, \ldots, x_r$ to $x_1.$ Each change takes constant time, and we must look through $2E^-$ elements. Next we look through the adjacency lists and delete every repeated element in a list to get rid of multiple edges. We also delete the negative edge $x_1 \bar x_1$ if it exists. The deletions take constant time, and we must look through $2E^-$ elements. Finally, we concatenate the lists for $x_2, \ldots, x_r$ to the recovery list for $x_1.$ Each addition takes constant time and there are at most $V-1$ lists to append. Therefore, identifying the vertices of one cycle takes time $\mathcal O(V + E^-).$ We do the identification process twice, after completing the depth first search; thus, the time complexity of Step 12 is $\mathcal O(V^2 + E^-).$
\subsubsection{Proof of Theorem \ref{time complexity}}
\begin{proof}
Let $\Sigma$ be a 3-chromatic graph. The algorithm completes Steps 1 and 2 exactly once each. This gives a time complexity of $\mathcal O(V(E^-)^2 + V + E^+).$
Every pass through Steps 3--9 consists of Step 3, at worst checks on all the conditions of Steps 4--9, and the completion of at most one step. Step 3 takes time $\mathcal O(V^2).$ Checking for vertices that satisfy the conditions of Steps 4--9 is bounded by $\mathcal O((E^-)^2).$ The completion of a single step from the Steps 4--9 is bounded by $\mathcal O(V + (E^-)^2).$ Each time the algorithm returns to Step 3, the graph has been reduced by at least one matched pair of vertices. Thus, the algorithm must pass through Steps 3--9 at most $V/2$ times. Therefore, Steps 3--9 are of total time complexity $\mathcal O(V^3 + V(E^-)^2).$
Steps 10, 11, and 12 are repeated at most $V/6$ times. This happens if every pair of cycles the algorithm finds in the forcing graph are composed of 3 vertices each. The cycles cannot be smaller than 3 vertices since Step 7 removes all possible cycles of length 2, and there are no loops in our forcing graphs. Thus, Steps 10, 11, and 12 have total time complexity $\mathcal O(V^3 + VE^-).$
Therefore, the algorithm takes time
$$\mathcal O(V^3 + V(E^-)^2 + V + E^+) = \mathcal O(V^3 + VE^2 + V + E) = \mathcal O(V^5). \qedhere$$
\end{proof}
\begin{corollary}
MaxDef is a polynomial-time algorithm if the signed graph is without multiple edges of the same sign and is input as a vertex list and two edge lists.
\end{corollary}
\begin{proof}
The vertex list is of size $V$ and the two edge lists are of size $E^+ + E^- = E.$ Thus, by Theorem \ref{time complexity}, the time for MaxDef is bounded by a polynomial in the size of our input.
\end{proof}
\begin{rmk}
A more compact way to input the graph is to input the number of vertices and an edge list. This input is size $E + 1$. If the graph is input in this more compact way, the MaxDef bound of $\mathcal O(V^5)$ is not necessarily polynomial. For example, if $E$ was of size $o(V)$, $\mathcal O(V^5)$ could be exponential or worse. Fortunately, if $\Sigma$ is connected, then $V = \mathcal O(E)$, meaning MaxDef is still a polynomial-time algorithm in the more compact input format.
\end{rmk}
\section{Switching Deficiency}
Let $A$ be a set of vertices of $\Sigma.$ \emph{Switching} $A$ is negating the signs of the edges with exactly one endpoint in $A$. Two signed graphs are \emph{switching equivalent} if they are related by switching. The chromatic number of a signed graph is the same as the chromatic number of the switched graph. A minimal coloration of the switched graph is simply a byproduct of switching the graph; given a minimal coloration of $\Sigma$, the signs of the colors on $A$ are negated during the switching process. Two colorations, $\kappa$ and $\kappa^*,$ are \emph{switching equivalent} if they are related by switching. If $\kappa^*$ is switching equivalent to $\kappa$, then $\kappa^*$ colors $\Sigma^*.$
In general, allowing switching makes the questions surrounding deficiency straightforward to solve. The \emph{switching deficiency range of $\kappa$}, $\mathcal{L}^*(\kappa)$, is $\{\operatorname{def}(\kappa^*) \mid \kappa^* $ is switching equivalent to $\kappa \}.$ The \emph{switching deficiency range of $\Sigma$}, $\mathcal{L}^*(\Sigma)$, is $\{\operatorname{def}(\kappa^*) \mid \kappa^* $ is a minimal coloration of a graph that is switching equivalent to $\Sigma.\}.$
\begin{thm}
Let $\Sigma$ be a signed graph and let $\kappa$ be a minimal coloration of $\Sigma.$ Then $\mathcal{L}^*(\kappa) = \mathcal{L}^*(\Sigma) = \left[0, \floor{\frac{1}{2}\chi(\Sigma)}\right].$
\label{switching def main}
\end{thm}
\begin{proof}
Let $\Sigma$ be a signed graph and let $\kappa$ be a minimal coloration of $\Sigma$.
First we show that $\operatorname{M*}(\Sigma) = \floor{\frac{1}{2}\chi}.$ We may assume that $\operatorname{D}(\kappa)$ contains only negative colors. Define $\kappa'$ to be the coloration obtained by switching all vertices with color less than 0. Then $\operatorname{def}(\kappa') = \floor{ \frac{1}{2}\chi}$ and thus, $\operatorname{M*}(\Sigma) = \floor{\frac{1}{2}\chi}$.
Next we show that $\operatorname{m*}(\Sigma) = 0.$
Let $\Sigma$ be a signed graph and let $\kappa$ be a minimal coloration of $\Sigma.$ We prove that for each $i \in D(\kappa)$ there exist at least two vertices colored $-i.$
Suppose $\chi(\Sigma) = 2k$ for some non-negative integer $k$. Suppose there exists an $i \in \operatorname{D}(\kappa)$ such that no negative edge of $\Sigma$ has both endpoints colored $-i$. Then $\{v \mid \kappa(v) = -i\}$ is stable. Thus, recoloring all such vertices 0 results in a proper coloration. But this coloration would use $2k-1$ colors.
Now suppose $\chi(\Sigma) = 2k + 1$ for some non-negative integer $k$. First note that $0 \not \in \operatorname{D}(\kappa)$ and there must at least one vertex colored $-i$ for each $i \in \operatorname{D}(\kappa)$, otherwise $\chi$ would be even. Suppose there exists $i \in \operatorname{D}(\kappa)$ such that exactly one vertex of $\Sigma$ is colored $-i.$ Call this vertex $w.$ Let $\kappa'$ be the coloration defined by
$$\kappa'(v) = \begin{cases}
i & \text{ if } \kappa(v) = 0 \text{ and } v \text{ is positively adjacent to } w, \\
-i & \text{ if } \kappa(v) = 0 \text{ and } v \text{ is negatively adjacent to } w, \\
-i & \text{ if } \kappa(v) = 0 \text{ and } v \text{ is not adjacent to } w, \\
\kappa(v) & \text{ otherwise.}
\end{cases}$$
The 0-color set of $\kappa$ is stable, so recoloring these vertices $i$ and $-i$ is proper. Also, every vertex that was recolored $i$ is a positive neighbor of $w$. Finally, every vertex that was recolored $-i$ is either negatively adjacent to $w$ or not adjacent to $w$. But then $\kappa'$ is a proper coloration using $2k$ colors.
Switch $\Sigma$ at exactly one of the vertices colored $-i$ for each $i \in D(\kappa),$
and call this new coloration $\kappa'$. Then every color in the color set is being used, so $\operatorname{def}(\kappa') = 0$ and therefore, $\operatorname{m*}(\Sigma) = 0.$
Finally, we show that every switching deficiency between 0 and $\floor{\frac{1}{2}\chi}$ can be attained. We may assume $\operatorname{def}(\kappa) = 0.$ Let $r \in [0, \operatorname{M*}(\kappa)].$ Choose $r$ positive colors in the color set of $\kappa;$ call them $c_1, c_2, \ldots, c_r.$ Let $\kappa'$ be the coloration defined by switching the vertices colored $c_i$ for each $i \in [1, r].$ Then $\kappa'$ does not have any vertices colored $c_i$ for each $i \in [1,r].$ Thus $\operatorname{def}(\kappa') = \operatorname{def}(\kappa) + r = r.$ Since every integer in $\mathcal{L}^*(\kappa)$ is an achievable value of switching deficiency for every minimal $\kappa,$ every integer in $\mathcal{L}^*(\Sigma)$ is achievable as well.
\end{proof}
| cc9f2071317bdd2d47ade92cc6b0633d761ac44f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Twisted conjugacy and Reidemeister numbers} \label{sec:twistedConjugacy}
\addtocounter{subsection}{1}
Let $G$ be a group and $\phi: G \to G$ be an automorphism. For $x, y \in G$, we say that $x$ and $y$ are \emph{$\phi$-conjugate} and write $x \conj{\phi} y$ if there exists a $g \in G$ such that $x = g y \inv{\phi(g)}$. The equivalence class of $x$ is denoted by $[x]$, or $[x]_\phi$ for clarity if there are multiple automorphisms involved.
We define $\mathcal{R}[\phi]$ to be the set of all $\conj{\phi}$-equivalence classes and the \emph{{Reidemeister} number $R(\phi)$ of $\phi$} as the cardinality of $\mathcal{R}[\phi]$. Note that $R(\phi) \in \mathbb{N}_0 \cup \{\infty\}$. Finally, we define the \emph{Reidemeister spectrum} to be
\(
\Spec_R(G) := \{ R(\phi) \mid \phi \in \Aut(G)\}.
\)
We say that \emph{$G$ has the $R_\infty$-property}, also denoted as $G \in R_\infty$, if $\Spec_R(G) = \{\infty\}$. We say that $G$ has \emph{full Reidemeister spectrum} if $\Spec_R(G) = \mathbb{N}_0 \cup \{\infty\}$.
The notion of Reidemeister number arises from Nielsen fixed-point theory, where its topological analog serves as a count of the number of fixed point classes of a continuous self-map, and is strongly related to the algebraic one defined above, see \cite{Jiang}.
It has been proven for several (classes of) groups that they possess the \(R_\infty\)-property, e.g.\ Baumslag-Solitar groups \cite{BaumslagSolitar} and their generalisations \cite{Levitt}, extensions of \(\SL(n, \mathbb{Z})\) and \( \GL(n, \mathbb{Z})\) by a countable abelian group \cite{MubeenaSankaran}, and Thompson's group \cite{ThompsonGroup}. We refer the reader to \cite{FelshtynNasybullov} for a more exhaustive list of groups having the \(R_\infty\)-property. \newline
In this article, we study the Reidemeister spectrum of right-angled Artin groups, RAAGs for short. Given a graph \(\Gamma\) with the set of vertices \(V\), the RAAG associated to it is the group
\[
A_{\Gamma} = \grp{V}{[v, w] \text{ if $v, w \in V$ are joined by an edge in $\Gamma$}}.
\]
Extreme cases of RAAGs include free groups and free abelian groups, coming from edgeless and complete graphs, respectively. From \cite[Theorem 3]{Felshtyn} (see also \cite{DekimpeGoncalves}), it readily follows that all non-abelian free groups of finite rank have the \(R_\infty\)-property. On the other hand, \(\Spec_{R}(\mathbb{Z}) = \{2, \infty\}\) and \(\Spec_{R}(\mathbb{Z}^{n})=\mathbb{N}_0\cup \{\infty\}\) for \(n \geq 2\) (see e.g.\ \cite{Romankov}). For groups closely related to right-angled Artin groups, several results regarding the \(R_\infty\)-property have been obtained, e.g.\ A.\ Juh\'asz has proven that certain Artin groups which are not right-angled Artin groups possess the \(R_\infty\)-property \cite{Juhasz} and recently, T.\ K.\ Naik, N.\ Nanda and M.\ Singh have showed that twin groups, a subfamily of the right-angled Coxeter groups, all possess the \(R_\infty\)-property \cite{NaikNandaSingh}.
We suspect that, amongst all RAAGs, only the free abelian ones do not possess the \(R_\infty\)-property:
\begin{conjecture}
Let $\Gamma(V, E)$ be a finite non-complete graph, i.e.\ $V$ is finite and there are two (distinct) vertices not joined by an edge. Then \(A_{\Gamma}\) has the \(R_\infty\)-property.
\end{conjecture}
We first reduce the conjecture to graphs belonging to three specific classes, after which we prove the conjecture for one of these classes and for several subclasses of the other two.
We start by recalling two ways of proving that an automorphism has infinite Reidemeister number.
\begin{defin}
Let $G, H$ be two groups. If $H \cong G / N$ for some characteristic subgroup $N$ of $G$, we call $H$ a \emph{characteristic quotient of $G$}.
\end{defin}
The following result is well-known, see e.g.\ \cite[Lemma 2.1]{MubeenaSankaran}.
\begin{lemma} \label{cor:characteristicQuotientRinf2}
Let $G, H$ be two groups. If $H$ is a characteristic quotient of $G$ and $H \in R_\infty$, then $G \in R_\infty$.
\end{lemma}
Reidemeister numbers also behave nicely under conjugation. For elements \(a, b\) of a group \(G\), we put \(a^{b} := \inv{b} a b\).
\begin{prop} \label{prop:conjugateEndomorphisms}
Let $G$ be a group and $\phi, \psi \in \Aut(G)$. Then $R(\phi) = R(\phi^\psi)$.
\end{prop}
\begin{proof}
Define $\rho: \mathcal{R}[\phi] \to \mathcal{R}[\phi^\psi]: [g]_\phi \to [\inv{\psi}(g)]_{\phi^\psi}$. As $\inv{\psi}$ is surjective, so is $\rho$, and
\begin{align*}
x \conj{\phi} y &\iff \exists g \in G: x = gy\inv{\phi(g)} \\
&\iff \exists g \in G: \inv{\psi}(x) = \inv{\psi}(g) \inv{\psi}(y) \inv{\inv{\psi}(\phi(\psi(\inv{\psi}(g))))} \\
&\iff \exists g \in G: \inv{\psi}(x) = \inv{\psi}(g) \inv{\psi}(y) \inv{\phi^\psi(\inv{\psi}(g))} \\
&\iff \inv{\psi}(x) \conj{\phi^\psi} \inv{\psi}(y),
\end{align*}
which shows well-definedness and injectivity of $\rho$.
\end{proof}
Next, we recall the definition and some properties of the lower central series of a group.
\begin{defin}
Let \(G\) be a group. The \emph{lower central series} of $G$ is defined as follows: put $\gamma_1(G) = G$ and inductively define \(\gamma_{i + 1}(G) = [\gamma_i(G), G]\) for $i \geq 1$.
\end{defin}
Each term in the lower central series is a characteristic subgroup, the quotients \(\gamma_i(G) / \gamma_{i + 1}(G)\) are all abelian and we will refer to them as the \emph{factors of the lower central series}. We can put these together to form the associated Lie ring of \(G\):
\begin{defin}
Let $G$ be a group. The \emph{Lie ring associated to $G$} is the abelian group
\[
L(G) := \bigoplus_{i = 1}^\infty L_i(G), \quad \text{ where } L_i(G) = \frac{\gamma_i(G)}{\gamma_{i + 1}(G)},
\]
and $L(G)$ is equipped with the following Lie bracket $\liebarg{\cdot}{\cdot}{}_{L}$: for $g \in \gamma_i(G)$ and $h \in \gamma_j(G)$, we define
\[
[g\gamma_{i + 1}(G), h\gamma_{j +1}(G)]_L := [g, h]\gamma_{i + j + 1}(G),
\]
and extend it by linearity to the whole of $L(G)$. Here, $[g, h]$ is the usual commutator bracket in $G$, i.e.\ \([g, h] = \inv{g} \inv{h} gh\).
\end{defin}
\begin{remark}
For $i$-fold commutators, we work with left-normed commutators, i.e.\ $[x_1, x_2, x_3] := [[x_1, x_2], x_3]$ and inductively, \([x_1, x_2, \ldots, x_n] := [[x_1, \ldots, x_{n - 1}], x_n].\)
\end{remark}
\begin{remark}
We write cosets multiplicatively, i.e.\ $g \gamma_{i}(G)$ and operations with cosets additively, i.e.\ $g\gamma_{i}(G) + h \gamma_{i}(G)$.
\end{remark}
Each automorphism $\phi$ of $G$ induces an automorphism $\phi_*$ of $L(G)$ and if, moreover, each factor of the lower central series is finitely generated and torsion-free, we can talk about the eigenvalues of $\phi_*$: these are simply the eigenvalues of each $\phi_i$. The following result is then immediate from \cite[Lemma~2.2]{DekimpeGoncalves}
\begin{theorem} \label{theo:usingL(G)ToEstablishRinf}
Suppose $G$ is a finitely generated group such that all factors of the lower central series are torsion-free. Let $\phi \in \Aut(G)$ be an automorphism. Denote by $\phi_*$ the induced automorphism on $L(G)$ and by $\phi_i$ the restriction of $\phi_*$ to $L_i(G)$.
If $\phi_i$ has eigenvalue $1$ for some $i$, then $R(\phi) = \infty$.
\end{theorem}
\section{Right-angled Artin groups} \label{sec:RAAGs}
In this section, we briefly recall the necessary definitions and results regarding graphs and right-angled Artin groups, including isomorphisms between two related RAAGs and a generating set for the automorphism group of a RAAG.
\subsection{Definitions and examples}
By a \emph{graph}, we mean a finite simple non-emtpy graph \(\Gamma(V, E)\) with vertices \(V\) and edges \(E\), although for technical reasons, we will sometimes mention the empty graph.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. The \emph{right-angled Artin group $A_\Gamma$ (or RAAG) associated to $\Gamma$} is defined by the presentation
\[
\pres{V}{[v, w], vw \in E}.
\]
The graph $\Gamma$ is called the \emph{defining graph of $A_{\Gamma}$}.
\end{defin}
From the definition of a RAAG, it is clear that the complete graph $K^n$ (i.e.\ every two vertices are connected by an edge) on $n$ vertices corresponds to the free abelian group $\mathbb{Z}^n$ and that the edgeless graph $\overline{K^n}$ on $n$ vertices corresponds to the free group $F_n$.
\medskip
There are two operations on the level of graphs that correspond to natural operations on the level of groups and that will play an important role.
\begin{defin}
Let $\Gamma_i(V_i, E_i)$, $i = 1, 2$, be two graphs. The \emph{disjoint union} of $\Gamma_1$ and $\Gamma_2$ is the graph
\[
\Gamma_1 \sqcup \Gamma_2(V_1 \sqcup V_2, E_1 \sqcup E_2)
\]
and the \emph{simplicial join} of $\Gamma_1$ and $\Gamma_2$ is the graph $\Gamma_1 * \Gamma_2$ with vertices $V_1 \sqcup V_2$ and edges
\[
E_{\Gamma_1 * \Gamma_2} = E_1 \sqcup E_2 \cup \{v_1v_2 \mid v_i \in V_i\}.
\]
We write $\sqcup_{n} \Gamma$ and $*_{n} \Gamma$ for the $n$-fold disjoint union, respectively, simplicial join of $\Gamma$ with itself.
\end{defin}
The following result follows readily from the definitions of a RAAG, direct product and free product.
\begin{prop} \label{prop:directAndFreeProductRAAGs}
Let $\Gamma_1, \Gamma_2$ be two graphs. Then $A_{\Gamma_1 \sqcup \Gamma_2} \cong A_{\Gamma_1} * A_{\Gamma_2}$ and $A_{\Gamma_1 * \Gamma_2} \cong A_{\Gamma_1} \times A_{\Gamma_2}$.
\end{prop}
\subsection{General isomorphisms of RAAGs}
In view of \cref{cor:characteristicQuotientRinf2}, it can be useful to transform one RAAG into another one by either deleting vertices or adding edges, and determining when this quotient is in fact characteristic. In this section, we make the first two claims more precise, in the next one, we discuss the characteristic quotients.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. We say that a subgraph $\Gamma'(V', E') \subseteq \Gamma$ is a \emph{full subgraph} or \emph{induced subgraph}, if $E'$ is given by
\(
\{vw \in E \mid v, w \in V'\}.
\)
Similarly, for $V' \subseteq V$, the \emph{subgraph induced on $V'$} is the graph $\Gamma'(V', E')$ where
\(
E' = \{vw \in E \mid v, w \in V'\}.
\)
We write $\Gamma(V')$.
\end{defin}
The RAAG associated to an induced subgraph $\Gamma(V')$ can be viewed as a subgroup of $A_\Gamma$ in a natural way.
\begin{lemma}[{\cite[Proposition~3.1]{DromsThesis}}]
Let $\Gamma$ be a graph and $\Gamma' := \Gamma(V')$ an induced subgraph. The map
\[
i: A_{\Gamma'} \to A_{\Gamma}: v' \mapsto v', \quad \text{ for $v' \in V'$}
\]
is well-defined and injective.
\end{lemma}
We now formulate precisely how we can `delete vertices'. Given a subset \(S\) of a group \(G\), we denote the \emph{normal closure of \(S\) in \(G\)} by \(\normcl{S}_{G}\) or simply \(\normcl{S}\) if \(G\) is clear from the context.
\begin{prop} \label{prop:eliminatingGenerators}
Let $\Gamma$ be a graph, $\Gamma_1$ an induced subgraph and view $A_{\Gamma_1}$ as a subgroup of $A_\Gamma$. Let $\Gamma_2$ be the induced subgraph on $V_2 = V \setminus V_1$. We also write $\Gamma_2 = \Gamma \setminus \Gamma_1$. Define
\[
\phi: A_\Gamma \to A_{\Gamma_2}: v \mapsto \begin{cases} 1 & \mbox{if } v \in V_1 \\
v & \mbox{otherwise}
\end{cases}.
\]
Then $\phi$ is a well-defined homomorphism, $\normcl{A_{\Gamma_1}}_{A_\Gamma} = \ker \phi$ and
\[
\frac{A_\Gamma}{\normcl{A_{\Gamma_1}}_{A_\Gamma}} \cong A_{\Gamma_2}.
\]
\end{prop}
For a proof, we refer the reader to Fontelles master thesis \cite[Proposition~2.1]{FontellesThesis} or the master thesis \cite[Proposition~3.2.6]{SendenThesis} of one of the authors.
To make notations less heavy, we introduce the following definition:
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $A_\Gamma$ its associated RAAG. For a subset $V' \subseteq V$, the normal subgroup $\normcl{A_{\Gamma(V')}}$ is called the \emph{normal subgroup generated by $V'$} and we denote it by $N(V')$. Any subgroup $N(V')$ is called a \emph{normal vertex-subgroup}.
\end{defin}
As we will never consider the subgroups $A_{\Gamma(V')}$ but only their normal closures, we simply refer to $N(V')$ as \emph{vertex-subgroups}.
Note that $N(V_1)N(V_2) = N(V_1 \cup V_2)$ for all $V_1, V_2 \subseteq V$. Indeed, both $N(V_1)$ and $N(V_2)$ lie in $N(V_1 \cup V_2)$, so their product does too. Conversely, it is clear that $A_{\Gamma(V_1 \cup V_2)}$ lies in $N(V_1)N(V_2)$, and as this product group is normal, we have that $N(V_1 \cup V_2) \leq N(V_1)N(V_2)$.
The second isomorphism result for RAAGs tells us how we can `add edges'.
\begin{lemma} \label{lem:imageOfNormalClosure}
Let $G, H$ be groups, $S \subseteq G$ a subset and $\phi: G \to H$ a surjective homomorphism. Then $\phi(\normcl{S}) = \normcl{\phi(S)}$.
\end{lemma}
\begin{prop} \label{prop:addingEdges}
Let $\Gamma_1 \subseteq \Gamma_2$ be two graphs on $n$ vertices. Denote by $R_i$ the relators in the presentation of $A_{\Gamma_i}$ and put $N = \normcl{R_2}_{A_{\Gamma_1}}$. Then
\[
\frac{A_{\Gamma_1}}{N} \cong A_{\Gamma_2}.
\]
\end{prop}
\begin{proof}
Recall that $A_{\Gamma_1} \cong \frac{F_n}{\normcl{R_1}_{F_n}}$. Let $\pi: F_n \to A_{\Gamma_1}$ be the natural projection and put $\tilde{N} = \inv{\pi}(N)$. We claim that $\tilde{N} = \normcl{R_2}_{F_n}$. As $\normcl{R_1}_{F_n}$ lies in both subgroups, it is sufficient to prove that $\pi(\tilde{N}) = N$, by the correspondence theorem. This follows from \cref{lem:imageOfNormalClosure}, as $\pi$ is surjective. Hence, by the third isomorphism theorem, we have that
\[
\frac{A_{\Gamma_1}}{N} = \frac{F_n / \normcl{R_1}_{F_n}}{\normcl{R_2}_{F_n} / \normcl{R_1}_{F_n}} \cong \frac{F_n}{\normcl{R_2}_{F_n}} \cong A_{\Gamma_2}. \qedhere
\]
\end{proof}
\begin{cor} \label{cor:abelianizationIsFreeAbelian}
Let $\Gamma$ be a graph on $n$ vertices and $A_{\Gamma}$ its associated RAAG. Then $A_{\Gamma} / [A_{\Gamma}, A_{\Gamma}]$ is a free abelian group of rank $n$.
\end{cor}
\subsection{Generating set for \texorpdfstring{$\Aut(A_\Gamma)$}{}}
Naturally, to determine whether or not a RAAG has the $R_\infty$-property, we need to understand its automorphism group.
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $v \in V$. A vertex $w$ is called a \emph{neighbour of $v$}, or \emph{adjacent to $v$}, if $vw \in E$.
We then define the \emph{link of $v$} as the set of all vertices adjacent to $v$, and it is denoted by $lk(v)$. The \emph{star of $v$} is $lk(v) \cup \{v\}$, and it is denoted by $st(v)$.
If $v \ne w$ are vertices, we say that \emph{$w$ dominates $v$} if $lk(v) \subseteq st(w)$, and write $v \leq w$.
\end{defin}
There are four basic types of automorphisms: \emph{graph automorphisms}, \emph{inversions}, \emph{transvections} and \emph{partial conjugations}.
\begin{itemize}
\item Graph automorphisms of \(\Gamma\) extend to automorphisms of \(A_{\Gamma}\).
\item Inversions \(\imath_{a}\) send one generator \(a\) to its inverse and leave all other generators fixed.
\item Transvections are maps \(\tau_{ab}\) sending a generator \(a\) to \(ab\) and leaving all other generators fixed, where \(b\) is another (different) generator with \(a \leq b\).
\item Partial conjugations are maps \(\gamma_{b, C}\) where \(b\) is a generator and \(C\) is a union of connected components of \(\Gamma \setminus \Gamma(st(b))\), sending every generator \(a\) in \(C\) to \(a^{b}\) and leaving the other generators fixed.
\end{itemize}
We will refer to these as \emph{automorphisms of basic type}. M.\ Laurence and H.\ Servatius have shown that the automorphisms of basic type generate $\Aut(A_\Gamma)$:
\begin{theorem}[\cite{Laurence95, Servatius89}]
For every graph $\Gamma$, $\Aut(A_\Gamma)$ is generated by its graph automorphisms, inversions, transvections and partial conjugations.
\end{theorem}
\begin{lemma} \label{lem:vertexSubgroupsArePartiallyCharacteristic}
Let $\Gamma$ be a graph and $\Gamma_1$ an induced subgraph. Let ${\phi \in \Aut(A_\Gamma)}$ be a composition of inversions and partial conjugations. Then $\phi(\normcl{A_{\Gamma_1}}) \leq \normcl{A_{\Gamma_1}}$.
\end{lemma}
\begin{proof}
Since \(\phi(N) \leq N\) and \(\psi(N) \leq N\) together imply \((\phi \circ \psi)(N) \leq N\) for any group \(G\) with normal subgroup \(N\) and automorphisms \(\phi\) and \(\psi\), it suffices to prove the statement for $\phi$ equal to an inversion or partial conjugation. If $\phi$ is an inversion, then $\phi(A_{\Gamma_1}) = A_{\Gamma_1}$, hence \(\phi(\normcl{A_{\Gamma_1}}) \leq \normcl{A_{\Gamma_1}}\). If $\phi$ is a partial conjugation, then $\phi(A_{\Gamma_{1}}) \subseteq \normcl{A_{\Gamma_{1}}}$ and thus \(\phi(\normcl{A_{\Gamma_1}}) \leq \normcl{A_{\Gamma_1}}\).
\end{proof}
So, whenever we need to prove that $\normcl{A_{\Gamma_1}}$ is characteristic in $A_\Gamma$, we only need to prove it is preserved under graph automorphisms and transvections.
\section{Characteristic vertex-subgroups} \label{sec:characteristicVertexSubgroups}
We begin the section by classifying all subsets of vertices that induce characteristic vertex-subgroups. After that, we treat two special cases, the vertices of maximal degree and the so-called transvection-free vertices, each of which implies a strong result regarding the $R_\infty$-property of RAAGs.
\subsection[Classification of characteristic vertex-subgroups]{Classification of characteristic vertex-subgroups}
For a graph $\Gamma(V, E)$ and a vertex $v \in V$, start with the set $V_0(v)$ containing only $v$. Add to $V_0(v)$ all vertices $w$ such that $\tau_{vw}$ is a well-defined transvection, to obtain $V_1(v)$. In symbols,
\[
V_1(v) = \{w \in V \mid v \leq w\}.
\]
Inductively, given $V_i(v)$, put
\[
V_{i + 1}(v) = \{w \in V \mid \exists v' \in V_i(v): v' \leq w\}.
\]
Define $V_\omega(v) = \bigcup_{i \in \mathbb{N}} V_i(v)$. Finally, put
\[
\Vtext{char}(v) = \bigcup_{\phi \in \Aut(\Gamma)} \phi(V_\omega(v)).
\]
\begin{defin}
The set $\Vtext{char}(v)$ defined above is called the \emph{characteristic closure of $v$ in $V$}.
\end{defin}
With this terminology, we then have the following result (which also explains the name \emph{characteristic closure}):
\begin{theorem} \label{theo:classificationCharacteristicVertexSubgroups}
Let $\Gamma$ be a graph and $A_\Gamma$ its associated RAAG. For vertices $v_1, \ldots, v_n \in V$, we have that
\[
N\left(\bigcup_{i = 1}^n \Vtext{char}(v_i)\right)
\]
is characteristic in $A_\Gamma$.
Conversely, if $N(V')$ is a characteristic vertex-subgroup for some $V' \subseteq V$, then $V'$ is a union of characteristic closures.
\end{theorem}
\begin{proof}
As
\[
N\left(\bigcup_{i = 1}^n \Vtext{char}(v_i)\right) = N(\Vtext{char}(v_1)) \ldots N(\Vtext{char}(v_n))
\]
and the product of characteristic subgroups is characteristic, it is sufficient to prove the theorem for $n = 1$. For simplicity, we write $V_{i}$ and $V_{\omega}$ for $V_{i}(v)$ and $V_{\omega}(v)$ as above.
Let $\phi \in \Aut(A_\Gamma)$ be of basic type. By \cref{lem:vertexSubgroupsArePartiallyCharacteristic}, we only need to consider the case where $\phi$ is a graph automorphism or a transvection. If $\phi$ is a graph automorphism, then $\phi(w) \in \Vtext{char}(v)$ for all $w \in \Vtext{char}(v)$, by construction, hence $\phi(N(\Vtext{char}(v))) \leq N(\Vtext{char}(v))$. Similarly, if $\phi = \tau_{v'w}$ is a transvection with $v' \in \Vtext{char}(v)$, then $v' = \psi(v'')$ for some $\psi \in \Aut(\Gamma)$ and $v'' \in V_\omega$. Note that $v'' \in V_i$ for some $i \in \mathbb{N}$. As $lk(v') \subseteq st(w)$, it follows that $lk(v'') \subseteq st(\inv{\psi}(w))$, hence $\inv{\psi}(w) \in V_{i + 1} \subseteq V_\omega$, and thus $w \in \psi(V_\omega) \subseteq \Vtext{char}(v)$. It follows that $\phi(N(\Vtext{char}(v))) \leq N(\Vtext{char}(v))$.
Conversely, suppose $N(V')$ is characteristic in $A_\Gamma$ for some $V' \subseteq V$. For each $v \in V'$, we will prove that $\Vtext{char}(v) \subseteq V'$; the result then follows from the equality
\[
V' = \bigcup_{v \in V'} \Vtext{char}(v).
\]
First, we prove by induction on $i$ that for all $i \in \mathbb{N}$, $V_i$ as defined above is a subset of $V'$. Clearly, $V_0 \subseteq V'$. So suppose $V_i \subseteq V'$. Let $w \in V_{i + 1}$ be arbitrary. Then there is a $v' \in V_i$ such that $v' \leq w$. Then $\phi = \tau_{v'w}$ is a well-defined transvection, and as $N(V')$ is characteristic, $\phi(N(V')) \leq N(V')$. In particular, $v'w \in N(V')$ and as $v' \in V'$ by induction hypothesis, $w \in V'$. Consequently, $V_{i + 1} \subseteq V'$.
From this, it follows that $V_\omega = \bigcup_{i \in \mathbb{N}} V_i \subseteq V'$. Now, let $\phi \in \Aut(\Gamma)$ and $v' \in V_\omega$. As $\phi(N(V')) \leq N(V')$, we see that $\phi(v') \subseteq V'$. Hence, $\phi(V_\omega) \subseteq V'$ for all $\phi \in \Aut(\Gamma)$. We conclude that
\[
\Vtext{char}(v) = \bigcup_{\phi \in \Aut(\Gamma)} \phi(V_\omega) \subseteq V'. \qedhere
\]
\end{proof}
\subsection{Vertices of maximal degree}
In this section and the following, we will treat two particular instances of characteristic vertex-subgroups and each of them will have an important consequence in establishing the $R_\infty$-property of RAAGs. We begin by recalling the definition of the degree of a vertex.
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $v \in V$ a vertex. The \emph{degree of $v$}, denoted by $\deg(v)$, is the number of adjacent vertices of $v$.
The \emph{maximal degree of $\Gamma$} is
\[
\Delta(\Gamma) := \max \{ \deg(v) \mid v \in V\}.
\]
\end{defin}
\begin{lemma}
Let $\Gamma$ be a graph and $v \in V$. Then for all $w \in \Vtext{char}(v)$, $\deg(w) \geq \deg(v)$.
\end{lemma}
\begin{proof}
First, we prove by induction that for all $i \in \mathbb{N}$ and $w \in V_i(v)$, we have $\deg(w) \geq \deg(v)$. Again, put $V_i = V_i(v)$. It clearly holds for $i = 0$. Suppose it holds for $i$. Let $w \in V_{i + 1}$. Then there is a $v' \in V_i$ such that $lk(v') \subseteq st(w)$. From this inclusion, it readily follows that $\deg(v') \leq \deg(w)$. As $\deg(v') \geq \deg(v)$ by induction hypothesis, we find that $\deg(w) \geq \deg(v)$.
By taking the union over all $V_i$, it follows that $\deg(w) \geq \deg(v)$ for all $w \in V_\omega$. To show that it holds for all $w \in \Vtext{char}(v)$, recall that
\[
\Vtext{char}(v) = \bigcup_{\phi \in \Aut(\Gamma)} \phi(V_\omega)
\]
and note that every automorphism preserves the degree of a vertex, so if $w = \phi(w')$ for some $\phi \in \Aut(\Gamma)$ and $w' \in V_\omega$, then $\deg(w) = \deg(w') \geq \deg(v)$.
\end{proof}
\begin{cor}
Let $\Gamma$ be a graph. Then $N(\Vtext{max})$ is characteristic in $A_\Gamma$, where
\[
\Vtext{max} = \{v \in V \mid \deg(v) = \Delta(\Gamma)\}.
\]
Moreover, $A_\Gamma / N(\Vtext{max}) \cong A_{\Gamma \setminus \Gamma(\Vtext{max})}$.
\end{cor}
\begin{proof}
The first claim follows from \cref{theo:classificationCharacteristicVertexSubgroups} together with the equality
\[
\Vtext{max} = \bigcup_{v \in \Vtext{max}} \Vtext{char}(v),
\]
which holds by the previous lemma. The isomorphism is a direct application of \cref{prop:eliminatingGenerators}.
\end{proof}
The true power of the previous corollary lies in the fact that for `almost all' graphs, $\Vtext{max}$ will be non-trivial. Indeed, as each graph considered is assumed to be finite, $\Vtext{max} \ne \emptyset$, and $\Vtext{max} = V$ if and only if $\Gamma$ is \emph{regular}, i.e.\ all vertices have the same degree. Therefore, every RAAG associated to a non-regular graph has a non-trivial characteristic vertex-subgroup, which provides a powerful tool in answering the conjecture. We elaborate further on this.
The next lemma is straightforward.
\begin{lemma} \label{cor:characteristicQuotientsTransitive}
Let $G$ be a group. If $G_{1}$ is a characteristic quotient of $G$ and $G_{2}$ one of $G_{1}$, then $G_{2}$ is a characteristic quotient of $G$.
\end{lemma}
\begin{lemma}[Simplification Lemma] \label{lem:simplificationLemma}
Let $\Gamma$ be a non-complete graph. Then $A_{\Gamma}$ has a RAAG $A_{\Gamma'}$ as a characteristic quotient where $\Gamma'$ is either
\begin{itemize}
\item disconnected;
\item connected, regular and non-complete;
\item non-regular, connected and such that $\Gamma'(V \setminus \Vtext{max})$ is complete.
\end{itemize}
\end{lemma}
\begin{proof}
Put $\Gamma_{0} = \Gamma$ and define inductively $\Gamma_{i + 1} = \Gamma_{i}(V_{i} \setminus V_{i, \mathrm{max}})$. Then each $A_{\Gamma_{i + 1}}$ is a characteristic quotient of $A_{\Gamma_{i}}$, and by \cref{cor:characteristicQuotientsTransitive}, also a characteristic quotient of $A_{\Gamma}$.
Since $\Gamma$ is finite and for each $i$ we have $|V_{i + 1}| < |V_{i}|$, there is an $m$ such that $\Gamma_{m + 1}$ is the empty graph. This implies that $\Gamma_{m}$ is regular. If $\Gamma_{m}$ is disconnected, or connected and non-complete, we are done, so suppose $\Gamma_{m}$ is complete. Then $\Gamma_{m - 1}$ cannot be regular and $\Gamma_{m - 1}(V_{m - 1} \setminus V_{m - 1, \mathrm{max}})$ is complete. If $\Gamma_{m - 1}$ is disconnected, we are in the first case, otherwise we are in the third case.
\end{proof}
We give a name to the third type of graph arising in the Simplification Lemma.
\begin{defin}
Let $\Gamma(V, E)$ be a connected non-regular graph. If the induced subgraph $\Gamma(V \setminus \Vtext{max})$ is complete, we call $\Gamma$ a \emph{(maximal degree)-by-(free abelian) graph}. We also say that $\Gamma$ is \emph{max-by-abelian}.
\end{defin}
\begin{remark}The inspiration for this terminology comes from group theory, where a group $G$ is called $\mathcal{P}$-by-$\mathcal{Q}$ if $G$ has a normal subgroup $N$ having property $\mathcal{P}$ such that $G / N$ has property $\mathcal{Q}$. Here, if $\Gamma$ is max-by-abelian, then the normal subgroup $N(\Vtext{max})$ of $A_\Gamma$ is generated by the vertices of \emph{maximal degree}, and the quotient $A_\Gamma / N(\Vtext{max})$ is a \emph{free abelian} group.
\end{remark}
The Simplification Lemma thus states that we only need to consider three types of graphs to establish the $R_\infty$-property for all non-abelian RAAGs. By demanding max-by-abelian graphs to be non-regular, there is no overlap between these three types of graphs. Note also that these types of graphs all have more structure than arbitrary graphs: although disconnectedness is not a very strong property from a graph theoretical point of view, it will be enough to ensure the $R_\infty$-property for the associated RAAG. For the second type, we have regularity, whereas the structure of max-by-abelian graphs is less obvious and less simple than that of the other two, as we will see later.
The Simplification Lemma proves the existence of at least one characteristic quotient of a particular form. It is of course possible for a RAAG to have multiple characteristic quotients, where for instance one quotient has a disconnected graph as defining graph and the other one a max-by-abelian graph.
\subsection{Transvection-free vertices}
Whereas $\Vtext{max}$ provided us with the Simplification Lemma, the set of all so-called transvection-free vertices will be in some sense more constructive towards establishing the $R_\infty$-property for RAAGs: if the aforementioned set contains all vertices, then $A_\Gamma$ has the $R_\infty$-property.
\subsubsection{Definition and statement of the main theorem}
Let $\Gamma$ be a graph and $v \in V$ a vertex. In the construction of $\Vtext{char}(v)$, we started with looking at which vertices $w \in V$ induce a well-defined transvection $\tau_{vw}$ on $A_\Gamma$. We consider now the vertices for which no such $w$ exist.
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $v \in V$ a vertex. We call $v$ \emph{transvection-free} if the set \(\{w \in V \mid v \leq w\} \) only contains $v$. The set of all transvection-free vertices is denoted by $\Vtext{\overline{\tau}}$.
\end{defin}
\begin{remark}
As $lk(v) \subseteq st(v)$, this definition makes sense.
\end{remark}
The following result is quite straightforward.
\begin{prop} \label{prop:transvectionFreeVerticesCharacteristic}
For every graph $\Gamma$, the subgroup $N(\Vtext{\overline{\tau}})$ is characteristic in $A_\Gamma$.
\end{prop}
\begin{proof}
Let $v \in V$ be transvection-free. It immediately follows that $V_\omega(v) = \{v\}$, thus
\[
\Vtext{char}(v) = \{\phi(v) \mid \phi \in \Aut(\Gamma)\}.
\]
If \(\phi \in \Aut(\Gamma)\) and $lk(\phi(v)) \subseteq st(w)$, then $lk(v) \subseteq st(\inv{\phi}(w))$, hence if $v$ is transvection-free, so is $\phi(v)$. Consequently,
\[
\Vtext{\overline{\tau}} = \bigcup_{v \in \Vtext{\overline{\tau}}} \Vtext{char}(v)
\]
and we conclude by \cref{theo:classificationCharacteristicVertexSubgroups}.
\end{proof}
Although the previous proposition provides us with a characteristic subgroup of $A_\Gamma$ differing (in general) from $N(\Vtext{max})$, the most interesting situation occurs when $\Vtext{\overline{\tau}} = V$. In that case, no generator of $A_\Gamma$ admits transvections, so the group $\Aut(A_\Gamma)$ will be significantly smaller. In fact, it turns out that in that case, $A_\Gamma$ has the $R_\infty$-property, but even more is true:
\begin{restatable}{theorem}{transvectionFreeAutomorphism} \label{theo:transvectionFreeAutomorphism}
Let $\Gamma$ be a graph and denote by $\Autnottrans(A_{\Gamma})$ the subgroup of $\Aut(A_\Gamma)$ generated by all graph automorphisms, inversions and partial conjugations. If $\Gamma$ is not complete, then all $\phi \in \Autnottrans(A_{\Gamma})$ have infinite Reidemeister number, i.e.\ $R(\phi) = \infty$ for all $\phi \in \Autnottrans(A_{\Gamma})$.
\end{restatable}
To prove this theorem, we need to understand the quotient groups $\gamma_i(A_\Gamma) / \gamma_{i + 1}(A_\Gamma)$ for $i = 1, 2, 3$. In order to do so, we study the associated Lie ring of a RAAG.
\subsubsection{Lyndon elements}
Lyndon elements can be used to show that the factors of the lower central series of a RAAG are free abelian groups, and they even provide us with a basis of these factors. This section is based on \cite{Wade}. Throughout this section, $\Gamma(V, E)$ is a graph, $V = \{v_1, \ldots, v_n\}$ and $A_\Gamma$ is its associated RAAG.
Define $W(V)$ to be the set of all words in the letters $v_1, \ldots, v_n$. The trivial word is denoted by $1$. The length of a word $w \in W(V)$ is denoted by $\len(w)$. For $w, w' \in W(V)$, we write $w \leftrightarrow w'$ if there are words $w_1, w_2 \in W(V)$ and vertices $v, v' \in V$ such that $vv' \in E$ and
\[
w = w_1 vv' w_2 \quad \text{ and } \quad w' = w_1 v'v w_2.
\]
We then define an equivalence relation on $W(V)$: we write $w \sim w'$ if there are words $w_1, \ldots, w_k \in W(V)$ such that
\[
w = w_1 \leftrightarrow w_2 \leftrightarrow \ldots \leftrightarrow w_k = w'.
\]
It is clear that if $w \sim w'$ and $\tilde{w} \sim \tilde{w}'$, then $w\tilde{w} \sim w' \tilde{w}'$. This allows us to define a multiplication on the set $M := W(V) / \sim$. Also, the length of a word is preserved under the relation `$\leftrightarrow$', hence under the equivalence relation `$\sim$'. Thus, we can define the length of an element $m \in M$, which we still denote by $\len(m)$.
We put a total order on $W(V)$ as follows: for all $w \in W(V) \setminus \{1\}$, $1 < w$. Then, for $w \ne w' \in W(V) \setminus \{1\}$, write $w = v_i w_1$ and $w' = v_j w_2$ for $v_i, v_j \in V$ and $w_1, w_2 \in W(V)$. We put $w < w'$ if either $i < j$, or $i = j$ and $w_1 < w_2$.
We now proceed to define \emph{Lyndon elements} in $M$. Denote by $\pi: W(V) \to M$ the projection map linked to the equivalence relation defining $M$.
\begin{defin}
For $m \in M$, we define the \emph{standard representative of $m$ in $W(V)$} to be the largest element of $\inv{\pi}(m)$ with respect to the total order. It is denoted by $std(m)$.
\end{defin}
Note that this is indeed well-defined: every element in $\inv{\pi}(m)$ has the same length, and as the alphabet $V$ is finite, so is the set of words of a fixed length. It follows that $\inv{\pi}(m)$ is finite, and as the order on $W(V)$ is total, this set has a largest element.
The standard representative of an element of $M$ allows us to define an order on $M$: for $m \ne m' \in M$, we put $m < m'$ if and only if $std(m) < std(m')$. From the totality of the order on $W(V)$, it follows that this order on $M$ is also total.
\begin{defin}
Let $m, m' \in M$. We say that $m$ and $m'$ are \emph{transposed} if there exist $x, y \in M$ such that $m = xy$ and $m' = yx$.
\end{defin}
In $M$, being transposed is not an equivalence relation: if we take $\Gamma$ to be the graph with $V = \{v_1, v_2, v_3\}$ and $E = \{v_2v_3\}$, then putting $m_1 = v_2v_1v_3$, $m_2 = v_1v_3v_2 = v_1v_2v_3$ and $m_3 = v_3v_1v_2$, we find that $m_1$ and $m_2$ are transposed, as are $m_2$ and $m_3$, but $m_1$ cannot be transformed into $m_3$ using only one transposition.
To circumvent this, we do a similar trick as in the definition of $M$.
\begin{defin}
Let $m, m' \in M$. We say that $m$ and $m'$ are \emph{conjugate} if there exist $m_1, \ldots, m_k$ such that $m = m_1$, $m' = m_k$, and $m_i$ and $m_{i + 1}$ are transposed for $1 \leq i \leq k - 1$.
\end{defin}
It is clear that this does define an equivalence relation on $M$.
\begin{defin}
An element $m \in M$ is called \emph{primitive} if for every $x, y \in M$, the equality $m = xy = yx$ implies that $x = 1$ or $y = 1$.
An element $m \in M$ is called a \emph{Lyndon element} if it is non-trivial, primitive and minimal in its conjugacy class w.r.t.\ the order $<$.
\end{defin}
\begin{defin}
Let $m \in M$. Then $init(m)$ is the set of all vertices in $V$ that can occur as the initial letter of a word in $\inv{\pi}(m)$.
\end{defin}
\begin{lemma}[{\cite[Corollary 3.2]{Lalonde}}]
If $m \in M$ is a Lyndon element, then $init(m)$ consists of a single vertex.
\end{lemma}
For a Lyndon element $m$, we consider $init(m)$ to be the single vertex it contains rather than the singleton.
\begin{defin}
For $m \in M$, we define $\zeta(m)$ to be
\[
supp(m) \cup \{v_j \mid \exists v_i \in supp(m): [v_i, v_j] \ne 1 \text{ in $A_\Gamma$}\}.
\]
\end{defin}
\begin{remark}
In the rest of this section, when writing $[v_i, v_j] \ne 1$, we mean $[v_i, v_j] \ne 1$ in $A_\Gamma$.
\end{remark}
Checking if a given element is a Lyndon element by means of the definition is quite cumbersome. Luckily, D.\ Krob and P.\ Lalonde \cite{Lalonde} proved the following equivalence.
\begin{theorem}[{\cite[Propositions 3.5 and 3.7]{Lalonde}}] \label{theo:equivalenceLyndonElements}
For $m \in M$, the following are equivalent:
\begin{enumerate}[(i)]
\item $m$ is a Lyndon element.
\item For all $x, y \in M \setminus \{1\}$ such that $m = xy$, we have $m < y$.
\item Either $\len(m) = 1$ or there exists Lyndon elements $x, y$ with $x < y$ and $init(y) \in \zeta(x)$ such that $m = xy$.
\end{enumerate}
\end{theorem}
We finish by determining the Lyndon elements of length at most $3$.
\begin{prop} \label{prop:LyndonElementsLength<=3}
The sets
\begin{align*}
LE_1 &= \{v_i \mid 1 \leq i \leq n\} \\
LE_2 &= \{v_i v_j \mid 1 \leq i < j \leq n, [v_i, v_j] \ne 1\} \\
LE_3 &= \{v_i v_i v_k \mid 1 \leq i < k \leq n, [v_i, v_k] \ne 1\} \\
&\quad \cup \{v_i v_j v_k \mid 1 \leq i < j < k \leq n, [v_i, v_j] \ne 1 \ne [v_i, v_k] \} \\
& \quad \cup \{v_i v_j v_j \mid 1 \leq i < j \leq n, [v_i, v_j] \ne 1\} \\
& \quad \cup \{v_i v_j v_k \mid 1 \leq i < j \ne k \leq n, i < k, [v_i, v_j] \ne 1, ([v_i, v_k] \ne 1 \text{ or } [v_j, v_k] \ne 1)\}.
\end{align*}
are precisely the Lyndon elements of length $1$, $2$ and $3$, respectively.
\end{prop}
\begin{proof}
For $LE_1$, this is clear. For $LE_2$, suppose $m$ is a Lyndon element of length $2$. Then $m = xy$ with $x < y$ both Lyndon elements and $init(y) \in \zeta(x)$. Note that $\len(x) = \len(y) = 1$, hence $x = v_i$ and $y = v_j$ for some $i, j$. As $x < y$, we have that $i < j$. Moreover, as $\zeta(x) = \{v_i\} \cup \{v_k \in V \mid [v_i, v_k] \ne 1\}$, we find that $[v_i, v_j] \ne 1$. Hence, $m = v_iv_j \in LE_2$. A similar argument shows that every element in $LE_2$ is a Lyndon element.
Suppose now that $m$ is a Lyndon element of length $3$. Write $m = v_iv_jv_k$. We distinguish two cases.
\emph{Case 1: $v_i < v_jv_k$ are both Lyndon elements}. Then $v_jv_k \in LE_2$, hence $[v_j, v_k] \ne 1$ and $j < k$. As $v_i < v_jv_k$ and $std(v_i) = v_i$, $std(v_jv_k) = v_jv_k$, we find that $i \leq j$. Also, $v_j = init(v_jv_k) \in \zeta(v_i) = \{v_i\} \cup \{v_l \mid [v_i, v_l] \ne 1\}$. Hence, either $i = j$, or $i < j$ and $[v_i, v_j] \ne 1$.
\emph{Case 2: $v_iv_j < v_k$ are both Lyndon elements}. Then $v_iv_j \in LE_2$, hence $i < j$ and $[v_i, v_j] \ne 1$. Similarly as in the previous case, $v_iv_j < v_k$ implies that $i < k$. Now, we have that $init(v_k) = v_k$ and
\[
\zeta(v_iv_j) = \{v_i, v_j\} \cup \{v_l \mid [v_i, v_l] \ne 1 \text{ or } [v_j, v_l] \ne 1\}.
\]
It follows that either
\begin{itemize}
\item $j = k$, or
\item $j \ne k$ and $[v_i, v_k] \ne 1$, or
\item $j \ne k$ and $[v_j, v_k] \ne 1$.
\end{itemize}
Conversely, suppose that $m = v_iv_jv_k \in LE_3$. In order to show that $m$ is a Lyndon element, we use the third criterion of \cref{theo:equivalenceLyndonElements} and case distinction.
\emph{Case 1: $i = j < k$, $[v_i, v_k] \ne 1$:} Then both $v_i$ and $v_i v_k$ are Lyndon elements, $v_i < v_i v_k$ and $init(v_iv_k) = v_i \in \zeta(v_i)$, hence $m = v_iv_iv_k$ is a Lyndon element as well.
\emph{Case 2: $i < j < k$, $[v_i, v_j] \ne 1 \ne [v_i, v_k]$:} In this case, $v_i$ and $v_jv_k$ are Lyndon elements, $v_i < v_jv_k$ and $init(v_jv_k) = v_j \in \zeta(v_i)$, as $[v_i, v_j] \ne 1$. Hence, $m = v_iv_jv_k$ is a Lyndon element as well.
\emph{Case 3: $i < j = k$, $[v_i, v_j] \ne 1$:} As $[v_i, v_j] \ne 1$ and $i < j$, $v_iv_j$ is a Lyndon element, as is $v_j$. Since $std(v_iv_j) = v_iv_j$, it follows that $v_iv_j < v_j$ in $M$. Lastly, note that $init(v_j) = v_j \in \zeta(v_iv_j)$, so $m = v_iv_jv_j$ is a Lyndon element.
\emph{Case 4: $i < j \ne k > i, [v_i, v_j] \ne 1$ and $([v_i, v_k] \ne 1$ or $[v_j, v_k] \ne 1)$:} Put $x = v_iv_j$ and $y = v_k$. As $i < j$ and $[v_i, v_j] \ne 1$, $x$ is a Lyndon element, and clearly, so is $y$. Note that $init(y) = v_k$ and $\zeta(x) = \{v_i, v_j\} \cup \{v_l \mid [v_i, v_l] \ne 1 \text{ or } [v_j, v_l] \ne 1\}$. As either $[v_i, v_k] \ne 1$ or $[v_j, v_k] \ne 1$, it follows that $init(y) \in \zeta(x)$ and therefore, $m$ is a Lyndon element.
\end{proof}
This explicit description of $LE_2$ and $LE_3$ will be particularly useful to determine linearly independent subsets of $\gamma_2(A_\Gamma) / \gamma_3(A_\Gamma)$ and $\gamma_3(A_\Gamma) / \gamma_4(A_\Gamma)$.
\medskip
Let $m$ be a Lyndon element of length at least $2$. By \cref{theo:equivalenceLyndonElements}, there exist Lyndon elements $x < y$ such that $m = xy$ and $init(y) \in \zeta(x)$. However, there can be multiple factorizations of this form. The factorization of $m$ where $y$ is minimal is called the \emph{standard factorization of $m$} and we write $S(m) = (x, y)$.
Using this standard factorization, we can define a bracketing procedure on the Lyndon elements: let $m$ be a Lyndon element. If $\len(m) = 1$, then $[m] = m$. If $\len(m) \geq 2$, write $(x, y) = S(m)$. Then the bracketing of $m$ is defined as $[[x], [y]]$, where $[x], [y]$ is the bracketing of $x$ and $y$, respectively. This bracketing can be interpreted in $A_{\Gamma}$ as the commutator bracket.
R.\ Wade \cite{Wade} then essentially proved the following:
\begin{theorem} \label{theo:RAAGsAreFGTorsionFreeLCS}
Let $\Gamma$ be a graph and $A_\Gamma$ its associated RAAG. Then the factors of the lower central series of $A_\Gamma$ are finitely generated and torsion-free. Moreover, after bracketing, the Lyndon elements of length $i$ form a $\mathbb{Z}$-basis of $L_i(A_\Gamma)$ for all $i$.
\end{theorem}
First of all, this theorem proves that we may use \cref{theo:usingL(G)ToEstablishRinf} for RAAGs. Secondly, combining this theorem with \cref{prop:LyndonElementsLength<=3}, we find explicit bases of $L_{2}(A_{\Gamma})$ and $L_{3}(A_{\Gamma})$. For $L_{2}(A_{\Gamma})$, this is immediate:
\begin{prop} \label{prop:basisL2(AGamma)}
The set
\[
\{[v_{i}, v_{j}]\gamma_{3}(A_{\Gamma}) \mid 1 \leq i < j \leq n, [v_{i}, v_{j}] \ne 1\}
\]
is a basis of $L_{2}(A_{\Gamma})$.
\end{prop}
Turning the Lyndon elements of length $3$ into an explicit basis of $L_{3}(A_{\Gamma})$ gives a quite long and ugly set. However, in the proof of \cref{theo:transvectionFreeAutomorphism}, we only need a linearly independent subset of this basis.
\begin{prop} \label{prop:linearlyIndependentSubsetL3(AGamma)}
Let $1 \leq i < j \leq n$ be indices such that $[v_{i}, v_{j}] \ne 1$. Then the bracketing of the Lyndon elements $v_{i}v_{i}v_{j}$ and $v_{i}v_{j}v_{j}$ is given by $[v_{i}, [v_{i}, v_{j}]]$ and $[[v_{i}, v_{j}], v_{j}]$, respectively.
In particular, the set
\begin{align*}
&\{ [v_{i}, v_{j}, v_{i}]\gamma_{4}(A_{\Gamma}) \mid 1 \leq i < j \leq n, [v_{i}, v_{j}] \ne 1\} \, \cup \\
&\{ [v_{i}, v_{j}, v_{j}]\gamma_{4}(A_{\Gamma}) \mid 1 \leq i < j \leq n, [v_{i}, v_{j}] \ne 1\}
\end{align*}
forms a linearly independent subset of $L_{3}(A_{\Gamma})$.
\end{prop}
\begin{proof}
\cref{prop:LyndonElementsLength<=3} implies that $v_{i}v_{i}v_{j}$ and $v_{i}v_{j}v_{j}$ are indeed Lyndon elements. Now, as neither $v_{i}v_{i}$ nor $v_{j}v_{j}$ is a Lyndon element, the standard factorizations are given by
\[
S(v_{i}v_{i}v_{j}) = (v_{i}, v_{i}v_{j}), \quad S(v_{i}v_{j}v_{j}) = (v_{i}v_{j}, v_{j}),
\]
hence the bracketing procedure gives $[v_{i}, [v_{i}, v_{j}]]$ and $[[v_{i}, v_{j}], v_{j}]$.
The claim regarding the linearly independent subset follows from \cref{theo:RAAGsAreFGTorsionFreeLCS} and the fact that
\[
[v_{i}, [v_{i}, v_{j}]]\gamma_{4}(A_{\Gamma}) = -[v_{i}, v_{j}, v_{i}]\gamma_{4}(A_{\Gamma}). \qedhere
\]
\end{proof}
\subsubsection{Proof of \cref{theo:transvectionFreeAutomorphism}}
We now have all the background needed to give a proof of \cref{theo:transvectionFreeAutomorphism}, which we restate here.
\transvectionFreeAutomorphism*
The main idea of the proof is to use \cref{theo:usingL(G)ToEstablishRinf}, i.e.\ find for a given automorphism $\phi \in \Autnottrans(A_{\Gamma})$, an index $i$ such that the induced automorphisms $\phi_i$ on $L_i(A_\Gamma)$ has eigenvalue $1$. More precisely, the proof consists of the following steps:
\begin{enumerate}[Step 1.]
\item We argue that it is sufficient to only consider the automorphisms in $\Autnottrans(A_{\Gamma})$ generated by graph automorphisms and inversions, by showing that partial conjugations descend to the trivial automorphism of $L(A_\Gamma)$.
\item Denote by $\Autip(A_{\Gamma})$ the subgroup of $\Autnottrans(A_{\Gamma})$ generated by all graph automorphisms and inversions. We show that every element $\psi$ in $\Autip(A_{\Gamma})$ is conjugate to an element $\psi'$ of a certain `nice' form. By \cref{prop:conjugateEndomorphisms}, $R(\psi) = R(\psi')$.
\item Finally, we show that each such `nice' $\psi'$ has infinite Reidemeister number, by finding an eigenvalue $1$ in either $L_1(A_\Gamma)$, $L_2(A_\Gamma)$ or $L_3(A_\Gamma)$.
\end{enumerate}
Throughout this section, we fix a non-complete graph $\Gamma$ on $n$ vertices and its associated RAAG $A_\Gamma$. We start with the first of two technical lemmata, which can be proven by straightforward induction.
\begin{lemma} \label{lem:congruenceModuloLCS}
Let $G$ be a group.
\begin{enumerate}[(i)]
\item If $x, y, z \in G$ are such that $x \equiv y \bmod \gamma_i(G)$, then $[x, z] \equiv [y, z] \bmod \gamma_{i + 1}(G)$. \label{item:equalModGammai}
\item \label{item:conjugationModGammai}For every $k \geq 1$, $x_1, \ldots, x_k \in G$ and $w_1, \ldots, w_k \in G$, we have that
\[
[x_1^{w_1}, \ldots, x_k^{w_k}] \equiv [x_1, \ldots, x_k] \bmod \gamma_{k + 1}(G).
\]
\item \label{item:exponentiationModGammai}For every $k \geq 1$, $x_1, \ldots, x_k \in G$ and $n_1, \ldots, n_k \in \mathbb{Z}$, we have that
\[
[x_1^{n_1}, \ldots, x_k^{n_k}] \equiv [x_1, \ldots, x_k]^{n_1 \ldots n_k} \bmod \gamma_{k + 1}(G).
\]
\end{enumerate}
\end{lemma}
\begin{cor} \label{cor:conjugationIsTrivialInLiering}
For every partial conjugation $\phi \in \Aut(A_\Gamma)$ and every $i \geq 1$, we have that the induced automorphism $\phi_i$ on $\gamma_i(A_\Gamma) / \gamma_{i + 1}(A_\Gamma)$ is the identity map.
Moreover, the induced automorphism $\phi_*$ on $L(A_\Gamma)$ is also the identity map.
\end{cor}
\begin{proof}
As $\gamma_i(A_\Gamma) / \gamma_{i + 1}(A_\Gamma)$ is generated by the cosets of the $i$-fold commutators of the generators $a_1, \ldots, a_n$ of $A_\Gamma$, it is sufficient to show that $\phi_i$ is the identity map on these commutators. But since $\phi$ is given by conjugation for each generator, \cref{lem:congruenceModuloLCS} implies that
\[
\phi([a_{i_1}, \ldots, a_{i_i}]) = [\phi(a_{i_1}), \ldots, \phi(a_{i_i})] \equiv [a_{i_1}, \ldots, a_{i_i}] \bmod \gamma_{i + 1}(A_\Gamma).
\]
Hence, $\phi_i$ is the identity map. The `moreover' statement is immediate.
\end{proof}
The following lemma gives a more concrete description of a given element in $\Autip(A_{\Gamma})$.
\begin{lemma} \label{lem:descriptionAutomorphismsTildeA}
For $\phi \in \Autip(A_{\Gamma})$, there exist a permutation $\sigma \in S_n$ and integers $e_1, \ldots, e_n \in \{-1, 1\}$ such that
\[
\phi(a_i) = a_{\sigma(i)}^{e_i}
\]
for all $i \in \{1, \ldots, n\}$ and such that the map $a_i \mapsto a_{\sigma(i)}$ is a graph automorphism of $\Gamma$.
\end{lemma}
\begin{proof}
The result trivially holds for the identity map, graph automorphisms and inversions, and it easily seen that the given form is preserved under composition. Hence, the result follows.
\end{proof}
The following lemma concerns the determinant of matrices of a specific form which will occur frequently in the proof of \cref{theo:transvectionFreeAutomorphism}.
\begin{lemma} \label{lem:specialDeterminant}
Let $k \geq 1$ be an integer and suppose $e_1, \ldots, e_k$ are all either equal to $1$ or $-1$. Define
\[
P(e_1, \ldots, e_k) := \begin{pmatrix}
0 & 0 & \ldots & 0 & e_k \\
e_1 & 0 & \ldots & 0 & 0 \\
0 & e_2 & \ldots & 0 & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \ldots & 0 & 0 \\
0 & 0 & \ldots & e_{k - 1} & 0.
\end{pmatrix}
\]
Note that $P(e_1) = (e_1)$. Then
\[
\det(I_k - P(e_1, \ldots, e_k)) = 1 - \prod_{i = 1}^k e_i.
\]
In particular, $P(e_1, \ldots, e_k)$ has eigenvalue $1$ if and only if an even number of the $e_i$ equal $-1$.
\end{lemma}
\begin{proof}
For $k = 1$, it is clear that $\det(I_1 - P(e_1)) = 1 - e_1$. For $k \geq 2$, expanding the determinant $\det(I_k - P(e_1, \ldots, e_k))$ along the first row yields the desired result.
\end{proof}
The last lemma (and the second technical one) makes more concrete what we meant by automorphisms in $\Autip(A_{\Gamma})$ of a `nice' form.
\begin{lemma} \label{lem:conjugationInAtilde}
Let $\phi \in \Autip(A_{\Gamma})$. Then there is a graph automorphism $\psi$ and an automorphism $\iota$, which is the composition of (possibly zero) inversions, such that $\phi$ and $\psi \circ \iota$ are conjugate.
Moreover, if $\sigma \in S_n$ is such that $\psi(a_i) = a_{\sigma(i)}$ and $\sigma = c_1 \circ \ldots \circ c_k$ is the disjoint cycle decomposition of $\sigma$, we can choose $\iota$ such that each cycle $c_j = (c_{j, 1} \ldots c_{j, n_{j}})$ contains at most one number $i_j$ with $\iota(a_{i_j}) = \inv{a}_{i_j}$. (Here, we also write cycles of length $1$, e.g. the identity permutation is written as $(1)(2) \ldots (n)$).
\end{lemma}
\begin{proof}
By \cref{lem:descriptionAutomorphismsTildeA}, we know that there is a permutation $\sigma \in S_n$ and elements $e_1, \ldots, e_n \in \{-1, 1\}$ such that $\phi(a_i) = a_{\sigma(i)}^{e_i}$ and such that $\psi: \Gamma \to \Gamma: a_i \mapsto a_{\sigma(i)}$ is a well-defined graph automorphism of $\Gamma$. Therefore, $\psi$ induces a graph automorphism of $A_\Gamma$, which we still denote by $\psi$.
Write $\sigma = c_1 \circ \ldots \circ c_k$ in disjoint cycle decomposition. After renumbering the generators, we can assume that there are integers $1 = i_1 < i_2 < \ldots < i_k$ such that $c_j = (i_j \ldots i_{j + 1} - 1)$, where we put $i_{k + 1} - 1 = n$. If we consider $\phi_1: A_\Gamma / \gamma_2(A_\Gamma) \to A_\Gamma / \gamma_2(A_\Gamma)$, then the matrix of $\phi_1$ with respect to the basis $a_1, \ldots, a_n$ is of the form
\[
M := \begin{pmatrix}
P(e_{i_1}, \ldots, e_{i_2 - 1}) & 0 & 0 & \ldots & 0 \\
0 & P(e_{i_2}, \ldots, e_{i_3 - 1}) & 0 & \ldots & 0 \\
0 & 0 & P(e_{i_3}, \ldots, e_{i_4 - 1}) & \ldots & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \ldots & P(e_{i_k}, \ldots, e_n)
\end{pmatrix}
\]
with each $P(e_{i_j}, \ldots, e_{i_{j + 1} - 1})$ as in \cref{lem:specialDeterminant}.
If we can find a diagonal matrix $D$ with only $\pm 1$ on the diagonal such that $DMD$ is of the form
\[
\begin{pmatrix}
P(\pm1, 1, \ldots, 1) & 0 & 0 & \ldots & 0 \\
0 & P(\pm1, 1, \ldots, 1) & 0 & \ldots & 0 \\
0 & 0 & P(\pm1, 1, \ldots, 1) & \ldots & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \ldots & P(\pm1, 1, \ldots, 1)
\end{pmatrix}
\]
with each block of the same dimension as in $M$, we are done. Indeed, the matrix $D$ will then be the matrix of $\jmath_1: A_\Gamma / \gamma_2(A_\Gamma) \to A_\Gamma / \gamma_2(A_\Gamma)$ w.r.t.\ the basis $a_1, \ldots, a_n$, with $\jmath \in \Aut(A_\Gamma)$ a composition of inversions and $DMD$ corresponds to the automorphism $\psi \circ \iota$, where $\iota$ maps $a_{i_j}$ to $a^{\pm 1}_{i_j}$ for all $1 \leq j \leq k$ according to the sign in the $j$-th block of $DMD$ and leaves all other generators fixed. Then $\psi \circ \iota = \jmath \circ \phi \circ \jmath$ and $\iota$ will satisfy the `moreover'-part of the lemma, since each $P$-block contains at most one $-1$.
It is sufficient to find such a diagonal matrix $D$ for each $P$-block and then putting all these blocks into one matrix. For ease of notation, we put $m = i_2 - 1$ and consider the block $P(e_1, \ldots, e_m)$. If $m = 1$, there is nothing to prove, so assume $m \geq 2$. Denote by $D_i$ the diagonal matrix with $1$'s on the diagonal except for the $i$-th position, there we put $-1$. Note that any product of the $D_i$'s is a diagonal matrix with only $\pm 1$ on the diagonal. It is not hard to see that
\[
D_iP(e_1, \ldots, e_m)D_i = P(e_1, \ldots, e_{i - 2}, -e_{i - 1}, -e_i, e_{i + 1}, \ldots, e_m),
\]
where $e_{0} = e_{m}$. Note that the parity of the number of $-1$'s is left unchanged. By starting with the $-1$ with the highest index and moving to $e_1$, we can clear out all $-1$'s, except for one if there were an odd number of them to begin with. Wherever this last $-1$ is situated, we can move it to the first position by conjugating with suitable $D_i$'s. In the end, we end up with $P(\pm 1, 1, \ldots, 1)$ and we are done.
\end{proof}
At last, we can give the proof of \cref{theo:transvectionFreeAutomorphism}.
\begin{proof}[Proof of \cref{theo:transvectionFreeAutomorphism}]
Let $\phi \in A$. Since we will work with the induced morphism $\phi_*$ on $L(A_\Gamma)$, we may assume by \cref{cor:conjugationIsTrivialInLiering} that $\phi \in \Autip(A_{\Gamma})$. By \cref{lem:conjugationInAtilde}, $\phi$ is conjugate to $\psi \circ \iota$ with $\psi$ a graph automorphism and $\iota$ a composition of (possible zero) inversions satisfying the `moreover'-part of the statement of \cref{lem:conjugationInAtilde}. As $\phi$ and $\psi \circ \iota$ are conjugate, $R(\phi) = R(\psi \circ \iota)$ by \cref{prop:conjugateEndomorphisms}. So, we can assume that $\phi = \psi \circ \iota$.
Let $\sigma \in S_n$ be the permutation associated to $\psi$ and write $\sigma = c_1 \circ \ldots \circ c_k$ in disjoint cycle decomposition. Denote by $\phi_i$ the induced automorphism on $L_i(A_\Gamma)$. By \cref{theo:RAAGsAreFGTorsionFreeLCS}, we can apply \cref{theo:usingL(G)ToEstablishRinf} to find that $R(\phi) = \infty$ if $\phi_i$ has eigenvalue $1$ for some $i$. We will distinguish several cases, and as each case will end with the sentence `$\phi_i$ has eigenvalue $1$', we will not mention \cref{theo:usingL(G)ToEstablishRinf} each time.
\medskip
\emph{Case 1:} Suppose there is a cycle, say, $c_1$, such that $\iota$ does not invert any of the generators $a_i$ with $i \in c_1$. After renumbering, we can assume that $c_1 = (1 \ldots m)$ for some $m \geq 1$. Then, on $L_1(A_\Gamma)$, we have that
\[
\phi_1(a_1\ldots a_m\gamma_2(A_\Gamma)) = a_2a_3\ldots a_ma_1\gamma_2(A_\Gamma),
\]
so $\phi_1$ has a fixed point, i.e.\ $1$ is an eigenvalue of $\phi_1$, hence $R(\phi) = \infty$.
From now on, assume that each cycle contains an index $i_j$ such that $\iota(a_{i_j}) = \inv{a}_{i_j}$.
\medskip
\emph{Case 2:} $k = n$, i.e.\ each cycle in the decomposition of $\sigma$ has length $1$. Then $\phi(a_i) = \inv{a}_i$ for all $i$. As $\Gamma$ is not complete, there are $a_i \ne a_j$ with $a_ia_j \notin E$. Hence, $[a_i, a_j]\gamma_3(A_\Gamma)$ is non-trivial and
\[
\phi_2([a_i, a_j]\gamma_3(A_\Gamma)) = [\inv{a}_i, \inv{a}_j]\gamma_3(A_\Gamma) = [a_i, a_j]\gamma_3(A_\Gamma)
\]
by \cref{lem:congruenceModuloLCS}\eqref{item:exponentiationModGammai}. Then $\phi_2$ has eigenvalue $1$, and thus $R(\phi) = \infty$.
From now on, assume that $k < n$.
\medskip
\emph{Case 3:} There is a cycle, say, $c_1$ containing indices $i < j$ with $a_ia_j \notin E$. Again, we can assume that $c_1 = (1 \ldots m)$. As $\sigma$ induces a graph automorphism, we have that $a_{i + l}a_{j + l}$ is not an edge of $E$ either for all $1 \leq l \leq m$ (here, we work with indices modulo $m$ where we use \(1\) up to \(m\) as representatives, rather than \(0\) up to \(m - 1\)). After renumbering the generators, we can assume that $a_1$ is mapped onto $\inv{a}_2$. Consider then the set
\[
B := \{[a_{i + l}, a_{j + l}]\gamma_3(A_\Gamma) \mid 0 \leq l \leq m - 1\}
\]
Note that $B$ does not necessarily form a linearly independent set: if $i \equiv j + l \bmod m$ and $j \equiv i + l \bmod m$ for some $1 \leq l \leq m - 1$, then $[a_i, a_j] = - [a_{i + l}, a_{j + l}]$. These two congruences can be simultaneously fulfilled if and only if $2(j - i) \equiv 0 \bmod m$. If this condition is not satisfied, then $B$ is indeed a linearly independent set: each element in $B$ can be rewritten (up to sign) such that the first index is strictly less than the second. No two elements will have the same indices, and by \cref{prop:basisL2(AGamma)}, $B$ is a subset of a basis of $L_2(A_{\Gamma})$. We then can proceed as follows: first remark that
\[
\phi_2([a_{i + l}, a_{j + l}]\gamma_3(A_\Gamma)) = \begin{cases}
-[a_{i + l + 1}, a_{j + l + 1}]\gamma_3(A_\Gamma) & \parbox[t]{.25\textwidth}{if $i + l \equiv 1 \bmod m$ or $j + l \equiv 1 \bmod m$} \\[1em]
[a_{i + l + 1}, a_{j + l + 1}]\gamma_3(A_\Gamma) & \mbox{otherwise}.
\end{cases}
\]
Note that $i + l$ and $j + l$ cannot both be congruent to $1$ modulo $m$ simultaneously, as otherwise $i \equiv j \bmod m$, which is impossible as $1 \leq i < j \leq m$ (and thus also $m \geq 2$). As all elements in $B$ are distinct, there will be precisely two elements $[a_{i + l}, a_{j + l}]\gamma_3(A_\Gamma)$ that are mapped to $-[a_{i + l + 1}, a_{j + l + 1}]\gamma_3(A_\Gamma)$. Moreover, $\phi_2(\Span_\mathbb{Z}(B)) = \Span_\mathbb{Z}(B) =: V$, hence we can consider the matrix of $\phi_2$ restricted to $V$ with respect to this basis and find a matrix $P(e_1, \ldots, e_m)$ where precisely two of the $e_i$ are equal to $-1$ and the rest equals $1$. Hence, \cref{lem:specialDeterminant} implies that $P(e_1, \ldots, e_m)$ has eigenvalue $1$, so $\phi_2$ restricted to $V$ does too. This implies that $\phi_2$ has eigenvalue $1$, and consequently, $R(\phi) = \infty$.
\medskip
Now, suppose that $2(j - i) \equiv 0 \bmod m$. If $m = 2$, then $B$ contains $[a_1, a_2]\gamma_3(A_\Gamma)$ and in that case,
\[
\phi_2([a_1, a_2]\gamma_3(A_\Gamma)) = -[a_2, a_1]\gamma_3(A_\Gamma) = [a_1, a_2]\gamma_3(A_\Gamma).
\]
Consequently, $\phi_2$ has eigenvalue $1$ and therefore $R(\phi) = \infty$.
Finally, if $m > 2$ and $2(j - i) \equiv 0 \bmod m$, note that $0 < j - i < m$, hence $m$ is even and $j - i = m / 2$. The set
\begin{align*}
\tilde{B} &:= \{[a_{1 + l}, a_{m / 2 + 1 + l}, a_{1 + l}]\gamma_4(A_\Gamma) \mid 0 \leq l \leq m / 2 - 1\} \\
& \cup \{[a_{1 + l}, a_{m / 2 + 1 + l}, a_{m / 2 + 1 + l}]\gamma_{4}(A_{\Gamma}) \mid 0 \leq l \leq m / 2 - 1\}
\end{align*}
will be linearly independent, as all elements in $\tilde{B}$ are distinct and $\tilde{B}$ forms a subset of a linearly independent set of $L_3(A_\Gamma)$, by \cref{prop:linearlyIndependentSubsetL3(AGamma)}. Also note that
\begin{align*}
\phi_3([a_{1 + l}, a_{m / 2 + 1 + l}, a_{1 + l}]\gamma_4(A_\Gamma)) &= [a_{2 + l}, a_{m / 2 + 2 + l}, a_{2 + l}]\gamma_4(A_\Gamma) \\
\phi_3([a_{1 + l}, a_{m / 2 + 1 + l}, a_{m / 2 + 1 + l}]\gamma_4(A_\Gamma)) &= [a_{2 + l}, a_{m / 2 + 2 + l}, a_{m / 2 + 2 + l}]\gamma_4(A_\Gamma)
\end{align*}
for all $0 \leq l \leq m/2 - 2$ and that
\begin{align*}
\phi_{3}([a_{m / 2}, a_{m}, a_{m / 2}]\gamma_{4}(A_{\Gamma})) &= [a_{m / 2 + 1}, a_{1}, a_{m / 2 + 1}]\gamma_{4}(A_{\Gamma}) \\
&= -[a_{1}, a_{m / 2 + 1}, a_{m / 2 + 1}]\gamma_{4}(A_{\Gamma}), \\
\phi_{3}([a_{m / 2}, a_{m}, a_{m}]\gamma_{4}(A_{\Gamma})) &= [a_{m / 2 + 1}, a_{1}, a_{1}]\gamma_{4}(A_{\Gamma}) \\
&= -[a_{1}, a_{m / 2 + 1}, a_{1}]\gamma_{4}(A_{\Gamma})
\end{align*} Hence, putting $W = \Span_\mathbb{Z}(\tilde{B})$, we find that $\phi_3(W) = W$ and that the matrix of $\phi_3$ restricted to $W$ with respect to $\tilde{B}$ is of the form $P(e_{1}, \ldots, e_{2m})$ as in \cref{lem:specialDeterminant}, where precisely two of the $e_{i}$'s are $-1$. This matrix has eigenvalue $1$, hence so does $\phi_3$. We conclude that $R(\phi) = \infty$.
\medskip
From now on, assume that each cycle $c_j$ satisfies the following property: if $c_j = (i_1 \ldots i_j)$, the induced subgraph $\Gamma(\{a_{i_1}, \ldots a_{i_j}\})$ is complete. As $\Gamma$ itself is not complete, there are cycles, say $c_1 = (1 \ldots m)$ and $c_2 = (m + 1 \ldots m + l)$, such that there are $i \in \{1, \ldots, m\}$, $j \in \{1, \ldots, l\}$ with $a_ia_{m + j} \notin E$. For notational convenience, we put $b_\imath= a_{\imath + m}$ for $\imath \in \{1, \ldots, l\}$. Hence, $a_i b_j \notin E$. It follows that also $a_1 b_{j - i + 1} \notin E$, as $\sigma^{1 - i}$ maps the non-edge $a_i b_j$ to the non-edge $a_{i + 1 - i} b_{j + 1 - i}$. Again, after renumbering, we can assume that $j - i + 1 = 1$. However, then it does not necessarily hold that $b_1 \mapsto \inv{b}_2$. By use of conjugation, we can obtain this nonetheless: suppose $b_\alpha \mapsto \inv{b}_{\alpha + 1}$ with $\alpha \ne 1$. Putting
\(
\jmath = \iota_2 \circ \iota_3 \circ \ldots \circ \iota_\alpha,
\)
where $\iota_k$ is the inversion $b_k$ to $\inv{b}_k$, a similar argument as in \cref{lem:conjugationInAtilde} yields that $\jmath \circ \phi \circ \jmath$ maps $b_1$ to $\inv{b}_2$, $b_\beta$ to $b_{\beta + 1}$ for $\beta \ne 1$ and coincides for the rest with $\phi$. Thus, we can assume that $a_1b_1 \notin E$.
Put $K = \lcm(l, m)$ and
\[
B = \{[a_1, b_1]\gamma_3(A_\Gamma), \ldots, [a_K, b_K]\gamma_3(A_\Gamma) \}.
\]
Again, we consider the indices modulo $m$ (for $a_i$) and $l$ (for $b_j$), respectively. Then $B$ is a linearly independent subset of $L_2(A_\Gamma)$ and $V := \Span_\mathbb{Z}(B)$ satisfies $\phi_2(V) = V$. We count the number of elements in $B$ that are mapped to minus the next generator in $B$. For $1 \leq i \leq K$, we have that
\[
\phi_2([a_i, b_i]\gamma_3(A_\Gamma)) = \begin{cases}
-[a_{i + 1}, b_{i + 1}]\gamma_3(A_\Gamma) & \parbox[t]{.4\textwidth}{if $i \equiv 1 \bmod m$ or $i \equiv 1 \bmod l$, but not both} \\[1em]
[a_{i + 1}, b_{i + 1}]\gamma_3(A_\Gamma) & \mbox{otherwise}.
\end{cases}
\]
Note that $i \equiv 1 \bmod m$ and $i \equiv 1 \bmod l$ if and only if $i \equiv 1 \bmod K$, hence if and only if $i = 1$. The number of indices in $\{1, \ldots, K\}$ that are congruent to $1$ modulo $m$ is $K / m$, and similarly there are $K / l$ indices congruent $1$ modulo $l$. As we have counted $1$ twice, there are $K (\inv{m} + \inv{l}) - 1$ indices $i$ that are congruent to $1$ modulo $m$ or $l$. It follows that there are $K' := K (\inv{m} + \inv{l}) - 2$ elements in $B$ such that $\phi_2([a_i, b_i]\gamma_3(A_\Gamma)) = -[a_{i + 1}, b_{i + 1}]\gamma_3(A_\Gamma)$.
\emph{Case 4:} Suppose $K'$ is even. Then the matrix of $\phi_2$ restricted to $V$ with respect to the basis $B$ will be of the form $P(e_1, \ldots, e_{K'})$ with an even number of $e_i$ equal to $-1$. Applying \cref{lem:specialDeterminant} gives that $P(e_1, \ldots, e_{K'})$ and hence $\phi_2$ has eigenvalue $1$, so $R(\phi) = \infty$.
\emph{Case 5:} $K'$ is odd. Then either $K / l$ or $K / m$ is even, say, $K / m$ (the case $K / l$ even is analogous). The set
\[
B' := \{[a_1, b_1, b_1]\gamma_4(A_\Gamma), \ldots, [a_K, b_K, b_K]\gamma_4(A_\Gamma) \}.
\]
is a linearly independent subset of $L_3(A_\Gamma)$, by \cref{prop:linearlyIndependentSubsetL3(AGamma)}. Moreover, by \cref{lem:congruenceModuloLCS}\eqref{item:exponentiationModGammai}
\[
\phi_3([a_i, b_i, b_i]\gamma_4(A_\Gamma)) = \begin{cases}
-[a_{i + 1}, b_{i + 1}, b_{i + 1}]\gamma_4(A_\Gamma) & \mbox{ if $i \equiv 1 \bmod m$} \\
[a_{i + 1}, b_{i + 1}, b_{i + 1}]\gamma_4(A_\Gamma) & \mbox{ otherwise}.
\end{cases}
\]
Hence, precisely $K / m$ elements of $B'$ are mapped to minus the next element and $V' := \Span_\mathbb{Z}(B')$ satisfies $\phi_3(V') = V'$. The matrix of $\phi_3$ restricted to $V'$ with respect to $B'$ will then again be of the form as in \cref{lem:specialDeterminant} with an even number of $e_i$'s equal to $-1$, hence $\phi_3$ has eigenvalue $1$ and we can conclude that $R(\phi) = \infty$.
\end{proof}
\begin{defin}
Let $\Gamma(V, E)$ be a graph. We call $\Gamma$ \emph{transvection-free} if $\Gamma$ is not $K^{1}$ and $\Vtext{\overline{\tau}} = V$.
\end{defin}
\begin{remark}
For $K^{1}$, the sole vertex $v$ is, strictly speaking, transvection-free. Since $A_{K^{1}} = \mathbb{Z}$ is abelian, we exclude $K^{1}$ from the transvection-free graphs.
\end{remark}
The following important corollary is immediate.
\begin{cor} \label{cor:transvectionFreeGraphsRinf}
Let $\Gamma$ be a transvection-free graph. Then $A_\Gamma \in R_\infty$.
\end{cor}
\begin{proof}
If $A_\Gamma$ is non-abelian and does not admit any transvections, then $\Autnottrans(A_{\Gamma}) = \Aut(A_\Gamma)$ and \cref{theo:transvectionFreeAutomorphism} implies that every $\phi \in \Aut(A_\Gamma) = \Autnottrans(A_{\Gamma})$ has infinite Reidemeister number.
\end{proof}
Of course, a theorem whose proof deserves a separate section should be of significant value. There are indeed a certain amount of transvection-free graphs, which we will discuss in the following section, but first we would like to make the following remark which explains to some extent the length of the proof: \cref{theo:transvectionFreeAutomorphism} holds for \emph{every} non-abelian RAAG. The only assumption regarding $\Gamma$ we needed in the proof was the fact that there are two vertices which are not connected by an edge. As only the complete graph does not meet this requirement, we consider a huge amount of very different graphs in the proof, which explains the number of cases.
The condition that $\Gamma$ is non-complete is also crucial: for $\Gamma = K^{n}$, we have $A_{\Gamma} = \mathbb{Z}^{n}$ and the automorphism $-{\Id_{\mathbb{Z}^{n}}}$ is a composition of graph automorphisms and inversions, but an easy calculation gives $R(-{\Id_{\mathbb{Z}^{n}}}) = 2^{n}$.
\subsubsection{Examples of transvection-free graphs}
To illustrate the significance of \cref{theo:transvectionFreeAutomorphism} and \cref{cor:transvectionFreeGraphsRinf}, we give an example of a family of transvection-free graphs. Later, in the section on regular graphs, we will see that most of the strongly regular graphs are also transvection-free, providing yet another example.
\begin{example}[Cycle graphs] \label{ex:cycleGraphs}
Let $n \geq 3$ be an integer. The \emph{cycle graph $C_n$} is the graph with vertex set $\{v_0, \ldots, v_{n - 1}\}$ and edge set \(\{v_i v_{i + 1} \mid i \in \{0, \ldots, n - 1\}\},\)
where the indices are viewed modulo $n$. Note that $C_n$ is connected and regular for each $n \geq 3$, and non-complete for $n \geq 4$.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\GraphInit[vstyle = Simple]
\tikzset{VertexStyle/.style = {shape = circle,fill = black,minimum size = 4pt,inner sep=0pt}}
\SetUpEdge[lw = 0.4pt]
\grEmptyCycle[RA = 1.5, prefix = a, rotation = 10]{9}
\EdgeInGraphLoop{a}{9}
\end{tikzpicture}
\caption{Cycle graph $C_9$}
\end{figure}
We claim that $C_n$ is transvection-free if and only if $n \geq 5$. Suppose $n \geq 5$. For $i \in \{0, \ldots, n - 1\}$, we have that $lk(v_i) = \{v_{i - 1}, v_{i + 1}\}$ and $st(v_i) = \{v_{i - 1}, v_i, v_{i + 1}\}$. If $lk(v_i) \subseteq st(v_j)$, then either $v_i = v_j$, or $v_{i - 1} = v_{j + 1}$ and $v_{i + 1} = v_{j - 1}$. The latter case implies that $i - 1 \equiv j + 1 \bmod n$ and $i + 1 \equiv j - 1 \bmod n$. Hence, $-2 \equiv i - j \equiv 2 \bmod n$. As $n \geq 5$, this is impossible, hence $v_i = v_j$. We conclude that $C_n$ is indeed transvection-free. Consequently, $A_{C_n}$ has the $R_\infty$-property for $n \geq 5$.
Conversely, if $n = 4$, note that $lk(v_1) = \{v_0, v_2\} = lk(v_3)$, so $C_4$ is not transvection-free. As $C_3 = K^3$, it is clear that also $C_3$ is not transvection-free.
\end{example}
\subsubsection{Properties of transvection-free graphs}
We end this section with some properties of transvection-free graphs.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. The \emph{complement graph $\cl{\Gamma}$} is the graph with vertex set $V$ and edge set
\(
\cl{E} = \{vw \mid v \ne w, vw \notin E\}.
\)
We also call $\cl{\Gamma}$ the \emph{complement of $\Gamma$}.
\end{defin}
For example, the complement of the complete graph $K^{n}$ is the edgeless graph $\cl{K^{n}}$, which also explains the notation. Moreover, it is not hard to see that
\(
\cl{\Gamma_{1} * \Gamma_{2}} = \cl{\Gamma_{1}} \sqcup \cl{\Gamma_{2}}
\)
and that $\cl{\cl{\Gamma}} = \Gamma$.
The next result follows directly from the definitions.
\begin{lemma} \label{lem:linksStarsComplement}
Let $\Gamma$ be a graph and $v \in V$. Denote by $lk_{\Gamma}(v)$ the link of $v$ seen as vertex of $\Gamma$ and similarly for $st_{\Gamma}(v)$. Then $lk_{\cl{\Gamma}}(v) = V \setminus st_{\Gamma}(v)$ and $st_{\Gamma}(v) = V \setminus lk_{\cl{\Gamma}}(v)$.
\end{lemma}
\begin{defin}
Let $\Gamma(V, E)$ be a graph. A vertex $v \in V$ is called \emph{isolated} if $lk(v) = \emptyset$.
\end{defin}
\begin{prop} \label{prop:transvectionfreeGraphsUnderGraphsOperations}
Let $\Gamma_{1}, \Gamma_{2}$ be transvection-free graphs. Then $\cl{\Gamma}_{1}$, $\Gamma_{1} \sqcup \Gamma_{2}$ and $\Gamma_{1} * \Gamma_{2}$ are all transvection-free as well.
\end{prop}
\begin{proof}
A transvection-free graph cannot contain an isolated vertex $v$, since otherwise $lk(v) = \emptyset$ is contained in the star of every other vertex. Hence, the link of every vertex in $\Gamma_{1}$ and $\Gamma_{2}$ is non-empty, implying that any pair of dominating vertices in $\Gamma_{1} \sqcup \Gamma_{2}$ arises in $\Gamma_{1}$ or $\Gamma_{2}$. Consequently, $\Gamma_{1} \sqcup \Gamma_{2}$ is transvection-free.
\medskip
For the complement of $\Gamma := \Gamma_{1}$: note that $w \in V$ dominates $v \in V$ seen as vertices in $\cl{\Gamma}$ if and only if $lk_{\cl{\Gamma}}(v) \subseteq st_{\cl{\Gamma}}(w)$. By \cref{lem:linksStarsComplement}, this is equivalent to $lk_{\Gamma}(w) \subseteq st_{\Gamma}(v)$, i.e.\ $v$ dominates $w$ seen as vertices in $\Gamma$. As $\Gamma$ is transvection-free, it follows that $\cl{\Gamma}$ is transvection-free as well.
\medskip
Finally, consider the graph $\Gamma := \Gamma_{1} * \Gamma_{2}$. Suppose $v, w \in V_{1}$ are vertices with $v \leq w$ (in $\Gamma$). Since $V_{2} \subseteq lk(v), V_{2} \subseteq st(w)$, we have that $lk_{\Gamma}(v) = V_{2} \cup lk_{\Gamma_{1}}(v)$ and similarly for $st_{\Gamma}(w)$. This implies that $lk_{\Gamma_{1}}(v) \subseteq st_{\Gamma_{1}}(w)$, which contradicts the transvection-freeness of $\Gamma_{1}$. A similar argument holds if $v, w \in V_{2}$.
If $v \in V_{1}$ and $w \in V_{2}$ are such that $lk_{\Gamma}(v) \subseteq st_{\Gamma}(w)$, then $V_{2} \subseteq lk_{\Gamma}(v) \subseteq st_{\Gamma}(w)$, hence, $st_{\Gamma_{2}}(w) = V_{2}$, which contradicts the fact that $\Gamma_{2}$ is transvection-free. We conclude that $\Gamma_{1} * \Gamma_{2}$ is transvection-free as well.
\end{proof}
\section{Disconnected graphs} \label{sec:freeProductsRAAGs}
\addtocounter{subsection}{1}
From this section onwards, we discuss right-angled Artin groups associated to the three types of graphs arising from the Simplification Lemma. We start with disconnected graphs, which correspond to free products of RAAGs.
\begin{defin}
A group \(G\) is called \emph{freely indecomposable} if \(G = G_{1} * G_{2}\) implies \(G_{1} = 1\) or \(G_{2} = 1\), i.e.\ \(G\) cannot be written as a non-trivial free product.
\end{defin}
Clearly, abelian groups are freely indecomposable. Recently, D.\ Gon\c{c}alves, P.\ Sankaran and P.\ Wong proved the following \cite{GoncalvesSankaranWong}:
\begin{theorem} \label{theo:RinfFreeProducts}
Let \(n \geq 2\) and suppose \(G = G_{1}* \ldots* G_{n}\), where each \(G_{i}\) is freely indecomposable, and both \(G_{1}\) and \(G_{2}\) have a proper characteristic subgroup of finite index. Then \(G \in R_\infty\).
\end{theorem}
E.\ Green proved in her PhD-thesis \cite[Lemma 4.7]{Green} that a RAAG is freely indecomposable if and only if its defining graph is connected. Next, we show that each RAAG admits a proper characteristic subgroup of finite index.
\begin{lemma} \label{lem:RAAGsAdmitCharacteristicSubgroup}
Let \(\Gamma\) be a graph and \(A_{\Gamma}\) its associated RAAG. Then \(A_{\Gamma}\) has a proper characteristic subgroup of finite index.
\end{lemma}
\begin{proof}
Suppose \(\Gamma\) has \(n\) vertices. Let \(p: A_{\Gamma} \to \frac{A_{\Gamma}}{[A_{\Gamma}, A_{\Gamma}]}\) be the natural projection. By \cref{cor:abelianizationIsFreeAbelian}, the abelianisation is isomorphic to \(\mathbb{Z}^{n}\). Note that \(2 \mathbb{Z}^{n}\) is a proper characteristic subgroup of finite index in \(\mathbb{Z}^{n}\). Then \(\inv{p}(2 \mathbb{Z}^{n})\) will be the desired subgroup of \(A_{\Gamma}\).
\end{proof}
We can therefore apply \cref{theo:RinfFreeProducts} to find the following result.
\begin{theorem} \label{theo:RinfDisconnectedGraphs}
Let \(\Gamma\) be a disconnected graph. Then \(A_{\Gamma}\) has the \(R_\infty\)-property.
\end{theorem}
\begin{proof}
Write \(\Gamma = \bigsqcup_{i = 1}^{n} \Gamma_{i}\), where each \(\Gamma_{i}\) is connected and where \(n \geq 2\). By \cref{prop:directAndFreeProductRAAGs}, \(A_{\Gamma} \cong \Ast_{i = 1}^{n} A_{\Gamma_{i}}\). By \cite[Lemma 4.7]{Green}, each \(A_{\Gamma_{i}}\) is freely indecomposable and by \cref{lem:RAAGsAdmitCharacteristicSubgroup}, each \(A_{\Gamma_{i}}\) admits a proper finite index characteristic subgroup. Consequently, the conditions for \cref{theo:RinfFreeProducts} are fulfilled and we can conclude that \(A_{\Gamma} \in R_\infty\).
\end{proof}
\begin{remark}
In \cite[Chapter 5]{SendenThesis}, \cref{theo:RinfDisconnectedGraphs} was proved using the associated Lie ring of the RAAG.
\end{remark}
\section{Regular graphs} \label{sec:regularGraphs}
We start this section with some results regarding direct products of RAAGs, which will then be applied to RAAGs associated to regular graphs. In particular, we discuss strongly regular graphs, which provide examples of both transvection-free graphs and direct products of RAAGs.
\subsection{Direct products of RAAGs}
For groups $G_{1}, \ldots, G_{n}$, the automorphism group of the direct product of the $G_{i}$'s always contains $\Aut(G_{1}) \times \ldots \times \Aut(G_{n})$, but the full automorphism group $\Aut(G_{1} \times \ldots \times G_{n})$ can be much bigger and much more difficult to describe. For RAAGs, however, a general description in terms of the factors is possible. We start with some general theory regarding automorphism groups of direct products before heading to direct products of RAAGs.
\subsubsection{Subgroups of \texorpdfstring{$\Aut(G_{1} \times \ldots \times G_{n})$}{}} \label{subsec:subgroupsAutDirectProduct}
\begin{prop} \label{prop:ReidemeisterNumberDirectProductAutomorphismGroups}
Let $G_{1}, \ldots, G_{n}$ be groups and suppose (at least) one of them has the $R_\infty$-property. Then for each element $\phi$ of $\Aut(G_{1}) \times \ldots \times \Aut(G_{n}) \leq \Aut(G_{1} \times \ldots \times G_{n})$, we have $R(\phi) = \infty$.
\end{prop}
\begin{proof}
Write $\phi = (\phi_{1}, \ldots, \phi_{n})$. It is clear that, for $(g_{1}, \ldots, g_{n}), (h_{1}, \ldots, h_{n}) \in G_{1} \times \ldots \times G_{n}$, we have
\[
(g_{1}, \ldots, g_{n}) \conj{\phi} (h_{1}, \ldots, h_{n}) \iff \forall i \in \{1, \ldots, n\}: g_{i} \conj{\phi_{i}} h_{i}.
\]
Thus,
\[
R(\phi) = \prod_{i = 1}^{n} R(\phi_{i}),
\]
from which the result immediately follows.
\end{proof}
For an \(n\)-fold direct product of a group with itself, we can consider a bigger subgroup of the automorphism group than merely the direct product: there is an injective group homomorphism
\[
\psi: \Aut(G) \wr S_{n} \to \Aut(G^{n}): (\phi, \sigma) = (\phi_{1}, \ldots, \phi_{n}, \sigma) \mapsto \psi(\phi, \sigma)
\]
where
\[
\psi(\phi, \sigma)(g_{1}, \ldots, g_{n}) = (\phi_{1}(g_{\inv{\sigma}(1)}), \ldots, \phi_{n}(g_{\inv{\sigma}(n)})).
\]
Here, \(\Aut(G) \wr S_{n}\) is the wreath product, i.e.\ the semidirect product \(\Aut(G)^{n} \rtimes S_{n}\), where the action is given by
\[
\sigma \cdot (\phi_{1}, \ldots, \phi_{n}) = (\phi_{\inv{\sigma}(1)}, \ldots, \phi_{\inv{\sigma}(n)}).
\]
We will write elements in \(\Aut(G) \wr S_{n}\) simply as \((\phi, \sigma)\).
\begin{prop} \label{prop:ReidemeisterNumberWreathProduct}
Let $G$ be a group having the $R_\infty$-property and $n \geq 1$ an integer. Then for each $\chi \in \Aut(G) \wr S_{n} \leq \Aut(G^{n})$, we have $R(\chi) = \infty$.
\end{prop}
\begin{proof}
Write $\chi = (\phi, \sigma)$. Suppose $(g, 1, \ldots, 1) \conj{\chi} (h, 1, \ldots, 1)$. Then there are $x_{1}, \ldots, x_{n} \in G$ with
\begin{align}
g &= x_{1}h\inv{\phi_{1}(x_{\inv{\sigma}(1)})} \nonumber\\
1 &= x_{i} \inv{\phi_{i}(x_{\inv{\sigma}(i)})} \quad \text{ $i \geq 2$}. \label{eq:xi=phi(xi)}
\end{align}
First, suppose that $\inv{\sigma}(1) = 1$. Then $g \conj{\phi_{1}} h$, hence the map
\[
F: \{[(g, 1, \ldots, 1)]_{\chi} \mid g \in G\} \to \mathcal{R}[\phi_{1}]: [(g, 1, \ldots, 1)]_{\chi} \mapsto [g]_{\phi_{1}}
\]
is well-defined and surjective. The domain of $F$ is a subset of $\mathcal{R}[\chi]$, hence $R(\chi) \geq R(\phi_{1}) = \infty$, since $\phi_{1} \in \Aut(G)$ and $G \in R_\infty$.
If $\inv{\sigma}(1) \ne 1$, let $m > 0$ be the smallest integer such that $\sigma^{-m}(1) = 1$. Repeatedly using \eqref{eq:xi=phi(xi)} gives
\[
g = x_{1}h\inv{\phi_{1}(x_{\inv{\sigma}(1)})} = x_{1}h\inv{\phi_{1}(\phi_{\inv{\sigma}(1)}(x_{\sigma^{-2}(1)}))} = \ldots = x_{1}h\inv{\tilde{\phi}(x_{1})}
\]
where
\[
\tilde{\phi} = \phi_{1} \circ \phi_{\inv{\sigma}(1)} \circ \ldots \circ \phi_{\sigma^{1 - m}(1)}.
\]
Note that $\tilde{\phi}$ only depends on $\chi$. Thus, $g \conj{\tilde{\phi}} h$ and the map
\[
F: \{[(g, 1, \ldots, 1)]_{\chi} \mid g \in G\} \to \mathcal{R}[\tilde{\phi}]: [(g, 1, \ldots, 1)]_{\chi} \mapsto [g]_{\tilde{\phi}}
\]
is well-defined and surjective. Similarly as before, we find that $R(\chi) = \infty$.
\end{proof}
\begin{cor} \label{cor:wreathProductEqualAutomorphismGroupRinf}
If a group \(G\) has the \(R_\infty\)-property and \(n \geq 1\) is an integer such that \(\Aut(G^{n}) = \Aut(G) \wr S_{n}\), then \(G^{n} \in R_\infty\).
\end{cor}
\subsubsection{Automorphism group of direct product of RAAGs}
The description of the automorphism group of a direct product of RAAGs is due to N.~Fullarton \cite{Fullarton}, and G.~Giovanni and N.~Wahl \cite{Gandini}, whose results we present in this section, together with its implications for the $R_\infty$-property for RAAGs. Recall that a direct product on group theoretical level corresponds to a simplicial join on graph theoretical level.
\begin{prop}[{\cite[Proposition 3.1]{Gandini}}] \label{prop:primeDecompositionRAAGs}
Let $\Gamma$ be a graph. Then $A_{\Gamma}$ admits a unique maximal decomposition as
\[
A_{\Gamma} = A_{\Gamma_{1}} \times \ldots \times A_{\Gamma_{k}}
\]
for induced subgraphs $\Gamma_{1}, \ldots, \Gamma_{k}$. This decomposition is unique up to isomorphism and permutation of the factors.
\end{prop}
\begin{lemma}[{\cite[Proposition 2.2]{CharneyVogtmann}}] \label{lem:centreOfRAAG}
Let $\Gamma$ be a graph and $A_{\Gamma}$ its associated RAAG. Then the centre of $A_{\Gamma}$ is given by
\(
Z(A_{\Gamma}) = \gen{\{v \in V \mid \deg(v) = |V| - 1\}}.
\)
\end{lemma}
\begin{prop}[{\cite[Proposition 3.3]{Gandini}}] \label{prop:furtherDecompositionAutomorphismGroupRAAG}
Let $\Gamma$ be a graph with maximal decomposition
\begin{equation} \label{eq:simplicialDecompositionGamma}
\Gamma = K^{d} * \Ast_{j = 1}^{k} (*_{i_{j}} \Gamma_{j})
\end{equation}
where all $\Gamma_{j}$ are non-isomorphic and non-complete graphs.
Then
\[
\Aut(A_{\Gamma}) \cong \mathbb{Z}^{d |V'|} \rtimes (\GL_{d}(\mathbb{Z}) \times (\Aut(A_{\Gamma_{1}}) \wr S_{i_{1}}) \times \ldots \times (\Aut(A_{\Gamma_{k}}) \wr S_{i_{k}})).
\]
\end{prop}
Using this description and the results from \cref{subsec:subgroupsAutDirectProduct}, we can describe what happens on the level of Reidemeister numbers.
\begin{theorem} \label{theo:directProductsRAAGsRinf}
Let $\Gamma$ be a graph with maximal decomposition
\begin{equation*}
\Gamma = K^{d} * \Ast_{j = 1}^{k} (*_{i_{j}} \Gamma_{j})
\end{equation*}
where all $\Gamma_{j}$ are non-isomorphic and non-complete graphs. If any graph $\Gamma_{j}$ is such that $A_{\Gamma_{j}} \in R_\infty$, then $A_{\Gamma} \in R_\infty$.
\end{theorem}
\begin{proof}
Since $Z(A_{\Gamma}) = A_{K^{d}}$ by \cref{lem:centreOfRAAG} and the centre of a group is characteristic, we have by \cref{prop:eliminatingGenerators} the isomorphism
\[
\frac{A_{\Gamma}}{A_{K^{d}}} \cong A_{\Gamma'}
\]
where
\begin{equation} \label{eq:decompositionGamma'}
\Gamma' = \Ast_{j = 1}^{k} (*_{i_{j}} \Gamma_{j}).
\end{equation}
The maximality of the decomposition of $\Gamma$ implies that \eqref{eq:decompositionGamma'} is the maximal decomposition of $\Gamma'$. Hence, \cref{prop:furtherDecompositionAutomorphismGroupRAAG} implies that
\[
\Aut(A_{\Gamma'}) \cong \Times_{j = 1}^{k} (\Aut(A_{\Gamma_{j}}) \wr S_{i_{j}}).
\]
Suppose that $A_{\Gamma_{j}} \in R_\infty$ for some $1 \leq j \leq k$. Then the isomorphism above combined with \cref{prop:ReidemeisterNumberDirectProductAutomorphismGroups,prop:ReidemeisterNumberWreathProduct} implies that $A_{\Gamma'} \in R_\infty$ as well. Since $A_{\Gamma'}$ is a characteristic quotient of $A_{\Gamma}$, we consequently have that $A_{\Gamma} \in R_\infty$.
\end{proof}
\begin{example} \label{ex:finiteProductsStartingWithZ}
As an application of the previous theorem, consider the smallest family $\mathcal{G}$ of groups satisfying the following two properties: $\mathcal{G}$ contains $\mathbb{Z}$ and $\mathcal{G}$ is closed under taking finite free products and finite direct products. We claim that any non-abelian group in $\mathcal{G}$ has the $R_\infty$-property. Note that $\mathcal{G}$ contains only RAAGs.
Let $G \in \mathcal{G}$ be a non-abelian group and write $G = A_{\Gamma}$ for some (non-complete) graph $\Gamma$. The group $G$ splits as either a free or direct product of two groups in $\mathcal{G}$. In the former case, $\Gamma$ is disconnected, hence $G \in R_\infty$ by \cref{theo:RinfDisconnectedGraphs}. In the latter case, write $\Gamma = \Gamma_{1} * \ldots * \Gamma_{k}$ in maximal decomposition, with $k \geq 2$. Since $G$ is non-abelian, one factor, say, $\Gamma_{1}$, is non-complete. As the decomposition of $\Gamma$ is maximal and $A_{\Gamma_{1}} \in \mathcal{G}$, we must have that $\Gamma_{1}$ is disconnected. Consequently, $A_{\Gamma_{1}} \in R_\infty$ by \cref{theo:RinfDisconnectedGraphs}. We can then apply \cref{theo:directProductsRAAGsRinf} to conclude.
In particular, any (non-abelian) finite direct product of free groups has the $R_\infty$-property.
\end{example}
We would like to remark that \cref{theo:directProductsRAAGsRinf} does \emph{not} state that any direct product of RAAGs has the $R_\infty$-property if (at least) one factor has it. It does state so for certain types of RAAGs, namely those for which their corresponding graph does not split as a simplicial join. Nonetheless, the result is strong enough for our purposes.
\subsection{Regular graphs}
We formally state the definition of a regular graph.
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $k \geq 0$. We call $\Gamma$ \emph{$k$-regular} if $\deg(v) = k$ for all $v \in V$. We call $\Gamma$ \emph{regular} if there is some $k \geq 0$ such that $\Gamma$ is $k$-regular.
\end{defin}
For example, the complete graph $K^n$ is $(n - 1)$-regular, whereas all cycle graphs on at least $3$ vertices are $2$-regular. Clearly, there is (up to isomorphism) only one $0$-regular graph on $n$ vertices, namely $\cl{K^{n}}$. We start with some very basic properties of regular graphs. For a proof, we refer the reader to \cite[Chapter 2]{ChartrandZhang}.
\begin{lemma}[Handshaking Lemma] \label{lem:handshakingLemma}
For a graph $\Gamma$ the following equality holds:
\[
2|E| = \sum_{v \in V} \deg(v).
\]
\end{lemma}
\begin{lemma} \label{lem:basicPropertiesRegularGraphs}
Let $\Gamma$ be a $k$-regular graph on $n$ vertices. Then
\begin{enumerate}[(i)]
\item $kn \equiv 0 \bmod 2$. \label{item:knequiv0regularGraph}
\item $\cl{\Gamma}$ is $(n - k - 1)$-regular. \label{item:complementRegularGraph}
\end{enumerate}
\end{lemma}
With this lemma, we classify all $k$-regular graphs on $n$ vertices with $k \in \{1, 2, n - 2, n - 3\}$.
\begin{prop} \label{prop:1regularGraphs}
Suppose $\Gamma$ is a $1$-regular graph on $n$ vertices. Then $\Gamma$ is the disjoint union of $n / 2$ copies of $K^{2}$.
\end{prop}
\begin{proof}
Note that $n$ is even, since $n\cdot 1 \equiv 0 \bmod 2$ by \cref{lem:basicPropertiesRegularGraphs}\eqref{item:knequiv0regularGraph}. Write $\Gamma = \Gamma_{1} \sqcup \ldots \sqcup \Gamma_{k}$ with each $\Gamma_{i}$ connected. Then each $\Gamma_{i}$ is a connected $1$-regular graph, so $\Gamma_{i}$ contains at least $2$ vertices, say $v$ and $w$, that are connected by an edge. However, $\Gamma_{i}$ cannot contain more than $2$ vertices, since no additional vertex can be adjacent to either $v$ or $w$. Hence, $\Gamma_{i}$ is (isomorphic to) $K^{2}$, implying that $\Gamma$ is a disjoint union of $n / 2$ copies of $K^{2}$.
\end{proof}
Taking complements, we obtain all $(n - 2)$-regular graphs on $n$ vertices.
\begin{cor} \label{cor:n-2RegularGraphs}
Suppose $\Gamma$ is an $(n - 2)$-regular graph on $n$ vertices. Then $\Gamma$ is the simplicial join of $n / 2$ copies of $\cl{K^{2}}$.
\end{cor}
\begin{cor} \label{cor:1/n-2RegularGraphsRinf}
Suppose $\Gamma$ is a non-complete graph on $n$ vertices that is either $1$-regular or $(n - 2)$-regular. Then $A_{\Gamma} \in R_\infty$.
\end{cor}
\begin{proof}
If $\Gamma$ is $1$-regular and non-complete, it is disconnected by \cref{prop:1regularGraphs}, hence $A_{\Gamma} \in R_\infty$ by \cref{theo:RinfDisconnectedGraphs}. If $\Gamma$ is $(n - 2)$-regular, it is a simplicial join of $n / 2$ copies of $\cl{K^{2}}$. Then \cref{ex:finiteProductsStartingWithZ} implies that $A_{\Gamma} \in R_\infty$.
\end{proof}
\begin{prop} \label{prop:2regularGraphs}
Suppose $\Gamma$ is a $2$-regular graph. Then $\Gamma$ is the disjoint union of cycle graphs.
\end{prop}
\begin{proof}
It is sufficient to prove that every \emph{connected} $2$-regular graph is a cycle graph. So, suppose that $\Gamma$ is connected. Denote by $v_{1}, \ldots, v_{n}$ the vertices of $\Gamma$. Consider a path of maximal length $k$ in $\Gamma$. Without loss of generality, this is the path $v_{1}, v_{2}, \ldots, v_{k}$. For each $j \in \{2, \ldots, k - 1\}$, the link of $v_{j}$ is given by $lk(v_{j}) = \{v_{j - 1}, v_{j + 1}\}$. Since $\deg(v_{k}) = 2$, there is a vertex besides $v_{k - 1}$ adjacent to $v_{k}$, say $v_{i}$. Note that $i \in \{1, k + 1, \ldots, n\}$. If $i \geq k + 1$, then the path was not maximal. Hence, $i = 1$ and since $\Gamma$ is connected, $k$ must be equal to $n$; otherwise, there would not exist a path from $v_{k + 1}$ to $v_{k}$. We conclude that $\Gamma$ is indeed a cycle graph.
\end{proof}
Again, taking complements gives us all $(n - 3)$-regular graphs on $n$ vertices.
\begin{cor} \label{cor:n-3RegularGraphs}
Suppose $\Gamma$ is an $(n - 3)$-regular graph on $n$ vertices. Then $\Gamma$ is the simplicial join of complements of cycle graphs.
\end{cor}
\begin{cor} \label{cor:2/n-3RegularGraphsRinf}
Suppose $\Gamma$ is a non-complete graph on $n$ vertices that is either $2$-regular or $(n - 3)$-regular. Then $A_{\Gamma} \in R_\infty$.
\end{cor}
\begin{proof}
If $\Gamma$ is $2$-regular, then \cref{prop:2regularGraphs} implies $\Gamma$ is either disconnected or a cycle graph. In the former case, $A_{\Gamma} \in R_\infty$. In the latter, $n \geq 4$ since $\Gamma$ is non-complete. If $n = 4$, then $\Gamma = \cl{K^{2}} * \cl{K^{2}}$. Consequently, \cref{ex:finiteProductsStartingWithZ} implies that $A_{\Gamma} \in R_\infty$. If $n \geq 5$, then $\Gamma$ is transvection-free by \cref{ex:cycleGraphs}, so \cref{cor:transvectionFreeGraphsRinf} implies that $A_{\Gamma} \in R_\infty$.
\medskip
Now suppose that $\Gamma$ is $(n - 3)$-regular. Then $\Gamma = \cl{C}_{n_{1}} * \ldots * \cl{C}_{n_{k}}$ for some $k \geq 1$ and $n_{i} \geq 3$. Note that the complement of the cycle graph $C_{3}$ is $\cl{K^{3}}$, the complement of $C_{4}$ is the disjoint union $K^{2} \sqcup K^{2}$, whereas the complement of the cycle graph $C_{n}$ for $n \geq 5$ is transvection-free by \cref{ex:cycleGraphs} and \cref{prop:transvectionfreeGraphsUnderGraphsOperations}. Consequently, $A_{\cl{C}_{m}} \in R_\infty$ for all $m \geq 3$ by \cref{theo:RinfDisconnectedGraphs,cor:transvectionFreeGraphsRinf}. Moreover, as each cycle graph is connected, their complements do not split as simplicial joins. Hence, the maximal decomposition of $\Gamma$ is given by $\cl{C}_{n_{1}} * \ldots * \cl{C}_{n_{k}}$. Applying \cref{theo:directProductsRAAGsRinf} then gives the result.
\end{proof}
\subsection{Strongly regular graphs}
We follow \cite[Chapter 10]{GodsilRoyle} for the definition of strongly regular graphs. As the name suggests, these graphs have strong regularity conditions.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. We say that $\Gamma$ is \emph{strongly regular with parameters $n$, $k$, $\lambda$ and $\mu$} if \(\Gamma\) is a \(k\)-regular graph on \(n\) vertices such that
\begin{itemize}
\item every two adjacent vertices have $\lambda$ neighbours in common, i.e. if $vw \in E$, then $|lk(v) \cap lk(w)| = \lambda$,
\item every two non-adjacent vertices have $\mu$ neighbours in common, i.e. if $vw \notin E$ and $v \ne w$, then $|lk(v) \cap lk(w)| = \mu$,
\item $1 \leq k < n - 1$.
\end{itemize}
We also say that $\Gamma$ is an \emph{$srg(n, k, \lambda, \mu)$}.
\end{defin}
\begin{remark}
The only $0$-regular graph on $n$ vertices is the edgeless graph $\overline{K^n}$. If we would consider this graph to be strongly regular, then $\lambda$ would be undefined. On the other hand, the only $(n - 1)$-regular graph on $n$ vertices is the complete graph $K^n$, for which the parameter $\mu$ is undefined. This explains why we require that $1 \leq k < n - 1$. The corresponding RAAGs of the aforementioned graphs are the free group and the free abelian group, respectively, and for both we already determined the Reidemeister spectrum. Hence, it is no loss to exclude these graphs from the strongly regular ones.
\end{remark}
Note that we thus can assume that $n \geq 2$ and in that case, $\lambda \leq k - 1$ and $\mu \leq k$. For an example of a strongly regular graph, we first need a definition.
\begin{defin}
Let $\Gamma(V, E)$ be a graph and $p \geq 2$. We say that $\Gamma$ is \emph{$p$-partite} if $V$ admits a partition $V_1, \ldots, V_p$ such that no vertices in $V_i$ are connected by an edge. If $\Gamma$ is $p$-partite for some $p$, we say that $\Gamma$ is \emph{multipartite}.
\end{defin}
\begin{example}
Let $p \geq 2$ and $n_1, \ldots, n_p$ be strictly positive integers. The \emph{complete $p$-partite graph of order $n_1, \ldots, n_p$} is the graph
\[
K(n_1, \ldots, n_p) := \overline{K^{n_1}} * \overline{K^{n_2}} *\ldots * \overline{K^{n_p}}.
\]
If $n_1 = n_2 = \ldots = n_p =: n$, we also write $K^p_n$. If $n = 1$, then $K^p_1$ is the complete graph on $p$ vertices, hence the notation coincides.
\medskip
Now, we claim that for $n \geq 2$, $K_n^p$ is an $srg(np, n(p - 1), n(p - 2), n(p - 1))$. Clearly, the number of vertices is $np$. If we denote by $V_1, \ldots, V_p$ the partition of the vertex set $V$, then $|V_i| = n$ for all $i$ and each $v \in V_i$ is connected with every vertex in $V \setminus V_i$. Hence, $\deg(v) = (p - 1)n$, so $K_n^p$ is $(p - 1)n$-regular.
Two vertices that are not adjacent lie in the same $V_i$, hence they have $n(p - 1)$ common neighbours. If two vertices $v, w$ are adjacent, then $v \in V_i$ and $w \in V_j$ with $i \ne j$. A vertex $v' \in V$ is a neighbour of $v$ if and only if $v' \notin V_i$ and it is a neighbour of $w$ if and only if $v' \notin V_j$. Hence, the common neighbours of $v$ and $w$ are precisely all vertices in $V \setminus (V_i \cup V_j)$, which contains $n(p - 2)$ elements.
\end{example}
The regularity conditions for a strongly regular graph $\Gamma$ restrict the possibilities for characteristic vertex-subgroups: as $\Gamma$ is regular, $\Vtext{max} = V$, and $\Vtext{\overline{\tau}}$ will be either empty or the whole of $V$. In order to prove this last statement, it will be more convenient to work with the complement of $\Vtext{\overline{\tau}}$, which we now give a name.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. We call the complement of $\Vtext{\overline{\tau}}$ the set of all \emph{transvection-admitting vertices} and we denote it by $\Vtext{\tau}$. Note that
\[
\Vtext{\tau} = \{v \in V \mid \exists w \ne v \in V: v \leq w\}.
\]
\end{defin}
\begin{lemma} \label{lem:symmetryVtransRegular}
Let $\Gamma$ be a $k$-regular graph and $v, w \in V$ two vertices. Then
\begin{enumerate}[(i)]
\item $lk(v) \subseteq lk(w) \iff lk(w) \subseteq lk(v) \iff |lk(v) \cap lk(w)| = k$.
\item $st(v) \subseteq st(w) \iff st(w) \subseteq st(v) \iff |st(v) \cap st(w)| = k + 1$.
\end{enumerate}
\end{lemma}
\begin{proof}
Both statements follow directly from the fact that $|lk(v)| = |lk(w)| = k$ and $|st(v)| = |st(w)| = k + 1$.
\end{proof}
\begin{prop} \label{prop:transvectionfreeVerticesSRG}
Let $\Gamma$ be an $srg(n, k, \lambda, \mu)$. Then
\begin{equation} \label{eq:VtransInSRG}
\Vtext{\tau} = \begin{cases}
\emptyset & \mbox{ if $\lambda < k - 1$ and $\mu < k$} \\
V & \mbox{ otherwise}.
\end{cases}
\end{equation}
\end{prop}
\begin{proof}
As mentioned before, we can assume that $n \geq 2$. Let $v, w \in V$ be distinct vertices. We look for equivalent conditions for $v \leq w$. We distinguish two cases.
\emph{Case 1: $vw \notin E$}. Then $v \leq w$ is equivalent with $lk(v) \subseteq lk(w)$ and hence with $|lk(v) \cap lk(w)| = k$, by the previous lemma. As $\Gamma$ is strongly regular, $|lk(v) \cap lk(w)| = \mu$. Hence, $lk(v) \subseteq lk(w)$ if and only if $k = \mu$.
\emph{Case 2: $vw \in E$}. In this case, $v \leq w$ is equivalent with $st(v) \subseteq st(w)$ and hence with $|st(v) \cap st(w)| = k + 1$. Note that
\[
|st(v) \cap st(w)| = |\{v, w\} \cup (lk(v) \cap lk(w))| = 2 + \lambda
\]
Therefore, $st(v) \subseteq st(w)$ if and only if $\lambda = k - 1$.
\medskip
With this information, we can prove \eqref{eq:VtransInSRG}: if $\lambda < k - 1$ and $\mu < k$, then $v \leq w$ can never happen, hence $\Vtext{\tau} = \emptyset$.
If $\mu = k$, recall that $\Gamma$ is a $k$-regular non-complete graph. Hence, for every vertex $v$ there is a non-adjacent vertex $w$. Then $lk(v) \subseteq lk(w)$ as $\mu = k$. Consequently, $\Vtext{\tau} = V$.
If $\lambda = k - 1$, take a vertex $v$ and one of its neighbours $w$ (which exists, as $k \geq 1$). As $\lambda = k - 1$, $st(v) \subseteq st(w)$. This holds for all $v \in V$, so $\Vtext{\tau} = V$.
\end{proof}
\begin{cor} \label{cor:almostAllSRGhaveRinf}
If $\Gamma$ is an $srg(n, k, \lambda, \mu)$ with $\lambda < k - 1$ and $\mu < k$, then $A_\Gamma \in R_\infty$.
\end{cor}
\begin{proof}
From the previous proposition, it follows that $\Vtext{\tau} = \emptyset$, hence $\Gamma$ is transvection-free. Applying \cref{cor:transvectionFreeGraphsRinf} now yields the result.
\end{proof}
A natural question is to ask which strongly regular graphs have either $\lambda = k - 1$ or $\mu = k$. To answer this question, we first point out a relation between those two kinds of graphs.
\begin{lemma}[{\cite[p.~218]{GodsilRoyle}}] \label{lem:complementSRGisSRG}
Let $\Gamma$ be an $srg(n, k, \lambda, \mu)$. Then $\cl{\Gamma}$ is an $srg(n, n - k - 1, n - 2k - 2 + \mu, n - 2k + \lambda)$.
\end{lemma}
\begin{lemma}[{\cite[p.~219]{GodsilRoyle}}] \label{lem:relationParametersSRG}
Let $\Gamma$ be an $srg(n, k, \lambda, \mu)$. Then
\[
(n - k - 1) \mu = k (k - \lambda - 1).
\]
In particular, $\mu = k$ if and only if $\lambda = 2k - n$, and $\lambda = k - 1$ if and only if $\mu = 0$.
\end{lemma}
Hence, if $\Gamma$ is an $srg(n, k, k - 1, 0)$, then $\cl{\Gamma}$ is an $srg(n, k', \lambda', \mu')$ with
\begin{align*}
k' &= n - k - 1 \\
\lambda' &= n - 2k - 2 = 2k' - n \\
\mu' &= n - k - 1 = k'.
\end{align*}
So, we only have to determine the strongly regular graphs with $\mu = k$ in order to determine all srg's with either $\mu = k$ or $\lambda = k - 1$.
\begin{prop}[{\cite[Lemma 10.1.1]{GodsilRoyle}}] \label{prop:SRGwithlambda = k - 1}
Let \(\Gamma\) be an \(srg(n, k, \lambda, \mu)\). If \(\lambda = k - 1\), then \(\Gamma\) is the disjoint union of \(\frac{n}{k + 1}\) copies of \(K^{k + 1}\).
\end{prop}
\begin{cor} \label{cor:SRGwithmu = k}
Let $\Gamma$ be an $srg(n, k, \lambda, \mu)$. If $\mu = k$, then $\Gamma$ is the complete multipartite graph $K^{n / (n - k)}_{n - k}$.
\end{cor}
\begin{proof}
By \cref{lem:complementSRGisSRG} and \cref{prop:SRGwithlambda = k - 1}, \(\cl{\Gamma}\) is the disjoint union of \(\frac{n}{n - k}\) copies of \(K^{n - k}\). Taking complements again implies that \(\Gamma\) is the simplicial join of \(\frac{n}{n - k}\) copies of \(\cl{K^{n - k}}\), i.e.\ \(\Gamma\) is the complete multipartite graph \(K^{n / (n - k)}_{n - k}\).
\end{proof}
The RAAGs associated to srg's with $\lambda = k - 1$ and $\mu = k$ are
\[
\Ast_{i = 1}^{\frac{n}{k + 1}} \mathbb{Z}^{k + 1} \quad \text{ and } \quad \Times_{i = 1}^{\frac{n}{n - k}} F_{n - k},
\]
respectively. The first group has the $R_\infty$-property by \cref{theo:RinfDisconnectedGraphs}. For the second, we invoke \cref{ex:finiteProductsStartingWithZ}. Combining with \cref{cor:almostAllSRGhaveRinf}, we find the following result.
\begin{theorem} \label{theo:SRGsHaveRinf}
Let $\Gamma$ be a strongly regular graph. Then $A_{\Gamma} \in R_\infty$.
\end{theorem}
\section{Max-by-abelian graphs} \label{sec:MBA}
Max-by-abelian graphs possess more structure than arbitrary (connected) graphs, but this structure is much less apparent than the structure of regular graphs, for instance. We therefore start by discussing general graph theoretical properties of max-by-abelian graphs, and we also provide some examples. Thereafter, we move on to their associated RAAGs and use the aforementioned structural properties to prove the $R_\infty$-property in certain cases.
\subsection{Examples and general properties}
We extend the definition of a max-by-abelian graph with some parameters.
\begin{defin}
Let $\Gamma(V, E)$ be a graph. We say that $\Gamma$ is \emph{$(n, k, d)$-max-by-abelian} (or $(n, k, d)$-MBA) if
\begin{itemize}
\item $\Gamma$ is non-regular and connected,
\item $\Gamma(V \setminus \Vtext{max})$ is complete,
\item $|V| = n$, $|\Vtext{max}| = k$ and $\Delta(\Gamma) = d$.
\end{itemize}
\end{defin}
In particular, $k \leq n - 1$ and $d \leq n - 2$, since $\Gamma$ is non-regular (and thus a fortiori non-complete). \Cref{fig:example543MBAgraph} shows an example of a $(5, 4, 3)$-MBA graph with $\Vtext{max} = \{v_{1}, v_{2}, v_{3}, v_{4}\}$.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\GraphInit[vstyle = Simple]
\SetVertexMath
\tikzset{VertexStyle/.style = {shape = circle,fill = black,minimum size = 4pt,inner sep=0pt}}
\SetUpEdge[lw = 0.4pt]
\grEmptyCycle[RA=1.5, prefix=v, rotation = 90]{5}%
\tikzset{AssignStyle/.append style = {above= 2pt}}
\AssignVertexLabel[color = black, size = \small]{v}{$v_{0}$, , , ,}
\tikzset{AssignStyle/.append style = {left = 1pt}}
\AssignVertexLabel[color = black, size = \small]{v}{, $v_{1}$, $v_{2}$, , }
\tikzset{AssignStyle/.append style = {right = 1pt}}
\AssignVertexLabel[color = black, size = \small]{v}{, , , $v_{3}$, $v_{4}$}
\EdgeInGraphLoop{v}{5}
\EdgeFromOneToSel{v}{v}{1}{3}
\EdgeFromOneToSel{v}{v}{2}{4}
\end{tikzpicture}
\caption{A $(5, 4, 3)$-max-by-abelian graph}
\label{fig:example543MBAgraph}
\end{figure}
\begin{example} \label{ex:simplicialJoinDisjointUnionCompleteGraphs}
Let $n, n_{1}, n_{2}$ be strictly positive integers with $n \leq n_{1}, n \leq n_{2}$ and $n^{2} < n_{1} n_{2}$. We claim that $\Gamma := (K^{n} \sqcup K^{n_{1}}) * (K^{n} \sqcup K^{n_{2}})$ is an MBA-graph. It is clearly connected. Denote the vertices of the first copy of $K^{n}$ by $a_{i}$, of the second copy by $c_{i}$, the vertices of $K^{n_{1}}$ by $b_{i}$ and those of $K^{n_{2}}$ by $d_{i}$. Then
\begin{align*}
\deg(a_{i}) &= n - 1 + n + n_{2} = 2n + n_{2} - 1 \\
\deg(b_{i}) &= n_{1} - 1 + n + n_{2} = n + n_{1} + n_{2} - 1 \\
\deg(c_{i}) &= n - 1 + n + n_{1} = 2n + n_{1} - 1 \\
\deg(d_{i}) &= n_{2} - 1 + n + n_{1} = n + n_{1} + n_{2} - 1
\end{align*}
Since at least one of the inequalities $n < n_{1}$ and $n < n_{2}$ holds, $\Gamma$ is not regular. We may assume that $n_{1} \leq n_{2}$. If $n < n_{1}$, then
\[
\Vtext{max} = \{b_{i} \mid 1 \leq i \leq n_{1}\} \cup \{d_{i} \mid 1 \leq i \leq n_{2}\}
\]
and $\Gamma(V \setminus \Vtext{max}) = K^{n} * K^{n} = K^{2n}$ is complete. In this case, $\Gamma$ is an $(2n + n_{1} + n_{2}, n_{1} + n_{2}, n + n_{1} + n_{2} - 1)$-MBA graph.
If $n = n_{1}$, then $n < n_{2}$ and
\(
\Vtext{max} = V \setminus \{c_{i} \mid 1 \leq i \leq n\}.
\)
In this case, $\Gamma(V \setminus \Vtext{max}) = K^{n}$, so $\Gamma$ is an $(2n + n_{1} + n_{2}, n + n_{1} + n_{2}, n + n_{1} + n_{2} - 1)$-MBA graph.
Note that the corresponding RAAG is given by $(\mathbb{Z}^{n} * \mathbb{Z}^{n_{1}}) \times (\mathbb{Z}^{n} * \mathbb{Z}^{n_{2}})$, which has the $R_\infty$-property by \cref{theo:RinfDisconnectedGraphs,theo:directProductsRAAGsRinf}.
\end{example}
The first properties we derive are inequalities providing relations amongst or restrictions on the parameters $n, k$ and $d$.
\begin{prop}
Let $\Gamma$ be an $(n, k, d)$-MBA graph. Then
\begin{equation} \label{eq:boundsEdgesMBA}
\binom{n}{2} - k(n - d - 1) \leq |E| \leq \frac{n(d - 1) + k}{2}.
\end{equation}
\end{prop}
\begin{proof}
The inequalities follow from a count of the minimal resp.\ maximal number of edges in $\Gamma$. First, we prove the upper bound. Every vertex in $\Vtext{max}$ has degree $d$, so every vertex in $V \setminus \Vtext{max}$ has at most degree $d - 1$. We thus have that
\begin{equation*}
|E| = \frac{1}{2} \sum_{v \in V} \deg(v) \leq \frac{k d + (n - k)(d - 1)}{2} = \frac{n(d - 1) + k}{2}.
\end{equation*}
The lower bound is more involved. Since the induced subgraph $\Gamma(V \setminus \Vtext{max})$ is the complete graph on $n - k$ vertices, $\Gamma$ has at least $\binom{n - k}{2}$ edges. We now count how many edges we deleted going from $\Gamma$ to $\Gamma(V \setminus \Vtext{max})$. Put $\Vtext{max} = \{v_{1}, \ldots, v_{k}\}$. Deleting $v_{1}$ results in deleting $d$ edges, one for every neighbour of $v_{1}$. Deleting $v_{i}$ after deleting $v_{1}$ up to $v_{i - 1}$ results in deleting at least $d - i + 1$ edges. Indeed, the only edges with endpoint $v_{i}$ that could already have been deleted are those with $v_{j}$ with $1 \leq j \leq i - 1$ as (other) endpoint. There are at most $i - 1$ such vertices and since $\deg(v_{i}) = d$, there are at least $d - i + 1$ edges with one endpoint equal to $v_{i}$ and one in $V \setminus \Vtext{max}$. Hence, deleting $v_{i}$ after deleting $v_{1}$ up to $v_{i - 1}$ results in deleting at least $d - i + 1$ edges. It follows that
\begin{align*}
|E| &\geq \binom{n - k}{2} + \sum_{i = 1}^{k} (d - i + 1) \\
&= \binom{n - k}{2} + k(d + 1) - \frac{k(k + 1)}{2} \\
&= \frac{1}{2} ((n - k)(n - k - 1) + 2kd + 2k - k^{2} - k) \\
&= \frac{n^{2} - 2nk + k^{2} - n + k + 2kd - k^{2} + k}{2} \\
&= \frac{n^{2} - n}{2} + kd + k - nk \\
&= \binom{n}{2} - k(n - d - 1). \qedhere
\end{align*}
\end{proof}
\begin{cor} \label{cor:boundsOnkAndd}
Let $\Gamma$ be an $(n, k, d)$-MBA graph. Then $n < 2k$ and ${d \leq n - \frac{k}{2k - n}}$.
\end{cor}
\begin{proof}
We transform the inequality in \eqref{eq:boundsEdgesMBA}:
\begin{alignat}{2}
&\quad& \binom{n}{2} - k(n - d - 1) & \leq \frac{n(d - 1) + k}{2} \nonumber \\
\iff && n^{2} - n - 2k(n - d - 1) & \leq nd - n + k \nonumber \\
\iff && n^{2} - 2kn + 2kd + 2k & \leq nd + k \label{eq:intermediateInequalityBoundd}\\
\iff && n^{2} - nd & \leq k (2n - 2d - 1). \nonumber
\end{alignat}
By dividing both sides of the last inequality by $2n - 2d - 1 \ne 0$, we have
\[
\frac{n}{2} < \frac{n(n - d)}{2(n - d) - 1} \leq k,
\]
where the strict inequality is equivalent with $2n(n - d) - n < 2n(n - d)$.
Continuing from \eqref{eq:intermediateInequalityBoundd}, we find
\[
d(2k - n) \leq 2kn - k - n^{2}
\]
and hence
\[
d \leq \frac{2kn - n^{2} - k}{2k - n} = n - \frac{k}{2k - n}. \qedhere
\]
\end{proof}
Note that since $k$ is an integer, the condition $2k > n$ is equivalent to $k \geq \ceil{\frac{n + 1}{2}}$.
\begin{lemma} \label{lem:degreePlusVmaxAtLeastVPlus1}
Let $\Gamma$ be an $(n, k, d)$-MBA graph. Then $k + d \geq n + 1$.
\end{lemma}
\begin{proof}
Since $\Gamma(V \setminus \Vtext{max})$ is isomorphic to $K^{n - k}$, the degree of a vertex in $V \setminus \Vtext{max}$ is at least $n - k - 1$. Since $\Gamma$ is connected, there is a $v \in V \setminus \Vtext{max}$ adjacent to a vertex $w \in \Vtext{max}$. For said vertex $v$, we have $d > \deg(v) \geq n - k$. Hence, $n - k \leq d - 1$, and consequently $n + 1 \leq k + d$.
\end{proof}
All results combined give some non-existence results of certain max-by-abelian graphs.
\begin{cor} \label{cor:MBAAtLeast5Vertices}
Let $\Gamma$ be an $(n, k, d)$-MBA graph. Then $n \geq 5$.
\end{cor}
\begin{proof}
Combining $k \leq n - 1$ with \cref{lem:degreePlusVmaxAtLeastVPlus1}, we obtain
\[
2 = n + 1 - (n - 1) \leq n + 1 - k \leq d \leq n - 2,
\]
hence $n \geq 4$. If $n = 4$, then $d = 2$, and $\ceil{\frac{n + 1}{2}} \leq k \leq n - 1$ implies $k = 3$. The sole vertex $v$ of $\Gamma$ not in $\Vtext{max}$ then has degree $1$, since $\Gamma$ is connected. The sum of the degrees is then $k \cdot d + 1 = 7$. This, however, contradicts the Handshaking Lemma. We conclude that $n \geq 5$.
\end{proof}
\begin{prop} \label{prop:MBAMinimalkImpliesNeven}
Let $\Gamma$ be an $\left(n, \ceil{\frac{n + 1}{2}}, d\right)$-MBA graph. Then $n$ is even.
\end{prop}
\begin{proof}
Suppose $n$ is odd. Then $k = \ceil{\frac{n + 1}{2}} = \frac{n + 1}{2}$. Applying \cref{cor:boundsOnkAndd} gives
\[
d \leq n - \frac{(n + 1) / 2}{2(n + 1) / 2 - n} = n - \frac{n + 1}{2} = \frac{n - 1}{2}
\]
and thus $k + d \leq n$, whereas \cref{lem:degreePlusVmaxAtLeastVPlus1} states that $k + d \geq n + 1$, a contradiction. Hence, $n$ is even.
\end{proof}
\subsection{\texorpdfstring{$R_\infty$-property for large values of $k$}{}}
Next, we focus on the associated RAAG of a max-by-abelian graph $\Gamma$. For large values of $k$, i.e.\ $k$ close to the upper bound of $n - 1$, the graph is structured enough to find characteristic vertex-subgroups resulting in a non-trivial and non-abelian quotient.
We start with $(n, n - 1, d)$-MBA graphs.
\begin{prop} \label{prop:intersectionLinkNonMaximalDegreeCharacteristic}
Let $\Gamma$ be a graph. Put
\[
\tilde{V} = \bigcap_{v \notin \Vtext{max}} lk(v).
\]
Then $N(\tilde{V})$ is characteristic in $A_{\Gamma}$.
\end{prop}
\begin{proof}
If $\tilde{V}$ is empty, the claim is trivial. Since $v \notin lk(v)$ for any $v \in V$, we have that $(V \setminus \Vtext{max}) \cap \tilde{V} = \emptyset$, hence $\tilde{V} \subseteq \Vtext{max}$.
Let $v \in \tilde{V}$. By \cref{theo:classificationCharacteristicVertexSubgroups}, it is sufficient to prove that $\Vtext{char}(v) \subseteq \tilde{V}$. Put $V_{0} = \{v\}$ and let $V_{i}, V_{\omega}$ be as in the construction of $\Vtext{char}(v)$, i.e.\
\[
V_{i} = \{w \in V \mid \exists v' \in V_{i - 1}: v' \leq w\}
\]
and $V_{\omega} = \bigcup_{n \in \mathbb{N}} V_{n}$. We prove by induction that $V_{i} \subseteq \tilde{V}$. For $i = 0$, this is trivial. Suppose that $V_{i} \subseteq \tilde{V}$ and that $w \in V$ dominates $v' \in V_{i}$. As $v' \in \tilde{V} \subseteq \Vtext{max}$ by the induction hypothesis, also $w \in \Vtext{max}$ and \(v'' \in lk(v') \subseteq st(w)\) for all $v'' \in V \setminus \Vtext{max}$.
Since $w \in \Vtext{max}$, this implies that $w \in \bigcap_{v'' \notin \Vtext{max}} lk(v'') = \tilde{V}$. Consequently, $V_{i + 1} \subseteq \tilde{V}$. We conclude that $V_{\omega} \subseteq \tilde{V}$.
Now, for $\phi \in \Aut(\Gamma)$, note that $\phi(\Vtext{max}) = \Vtext{max}$, hence $\phi(V \setminus \Vtext{max}) = V \setminus \Vtext{max}$ and thus $\phi(\tilde{V}) = \tilde{V}$. In particular is
\(
\phi(V_{\omega}) \subseteq \phi(\tilde{V}) = \tilde{V},
\)
therefore, $\Vtext{char}(v) \subseteq \tilde{V}$.
\end{proof}
\begin{prop} \label{prop:nn-1dMBAHaveRinf}
Let $\Gamma$ be an $(n, n - 1, d)$-MBA graph. Then $A_{\Gamma} \in R_\infty$.
\end{prop}
\begin{proof}
Let $v$ be the sole vertex in $V \setminus \Vtext{max}$. By \cref{prop:intersectionLinkNonMaximalDegreeCharacteristic}, $N(lk(v))$ is characteristic in $A_{\Gamma}$. Note that $lk(v) \ne \Vtext{max}$, since
\[
|lk(v)| \leq d - 1 \leq n - 3 < n - 1 = |\Vtext{max}|.
\]
Hence, $\tilde{\Gamma} := \Gamma(V \setminus lk(v))$ is equal to $\Gamma(\{v\}) \sqcup \Gamma(\Vtext{max} \setminus lk(v))$. \cref{theo:RinfDisconnectedGraphs} implies that $A_{\tilde{\Gamma}}$ has the $R_\infty$-property. Since $A_{\tilde{\Gamma}} \cong \frac{A_{\Gamma}}{N(lk(v))}$ (by \cref{prop:eliminatingGenerators}) is a characteristic quotient of $A_{\Gamma}$, also $A_{\Gamma}$ has the $R_\infty$-property.
\end{proof}
Now we move on to $(n, n - 2, d)$-MBA graphs.
\begin{prop} \label{prop:intersectionVmaxUnionLinksMBA}
Let $\Gamma$ be a graph. Put
\[
\tilde{V} = \Vtext{max} \cap \left(\bigcup_{v \notin \Vtext{max}} lk(v)\right).
\]
Then $N(\tilde{V})$ is characteristic in $A_{\Gamma}$.
\end{prop}
\begin{proof}
Let $v \in \tilde{V}$. Again, by \cref{theo:classificationCharacteristicVertexSubgroups}, it is sufficient to prove that $\Vtext{char}(v) \subseteq \tilde{V}$. Put $V_{0} = \{v\}$ and let $V_{i}, V_{\omega}$ be as usual. We prove by induction that $V_{i} \subseteq \tilde{V}$. For $i = 0$, this is trivial. Suppose $V_{i} \subseteq \tilde{V}$ and let $w \in V$ be a vertex dominating $v' \in V_{i}$. By the induction hypothesis, $v' \in \Vtext{max}$. Since $w$ dominates $v'$, $\deg(w) \geq \deg(v') = \Delta(\Gamma)$, so $w \in \Vtext{max}$.
Also by the induction hypothesis, there exists a $w' \notin \Vtext{max}$ such that $v' \in lk(w')$. Consequently is $w' \in lk(v') \subseteq st(w)$. As $w' \notin \Vtext{max}$, we have that $w' \ne w$, so $w' \in lk(w)$. Therefore, $w \in lk(w')$ implying that $w \in \tilde{V}$. We conclude that $V_{\omega} \subseteq \tilde{V}$.
Next, let $\phi \in \Aut(\Gamma)$. As $\phi(\Vtext{max}) = \Vtext{max}$, we have that $\phi(V \setminus \Vtext{max}) = V \setminus \Vtext{max}$ and hence
\[
\phi(\tilde{V}) = \Vtext{max} \cap \left(\bigcup_{\phi(v) \notin \Vtext{max}} lk(\phi(v))\right) = \Vtext{max} \cap \left(\bigcup_{v \notin \Vtext{max}} lk(v)\right) = \tilde{V}.
\]
Consequently, $\phi(V_{\omega}) \subseteq \tilde{V}$ and thus $\Vtext{char}(v) \subseteq \tilde{V}$.
\end{proof}
Before heading to the next result, we state three commutator identities, each of which can be easily checked by working out both sides of the equality.
\begin{lemma} \label{lem:commutatorIdentities}
Let $G$ be a group, and $a, b, c \in G$. Then the following identities hold:
\begin{enumerate}[(i)]
\item \([ab, c] = [a, c]^{b}[b, c]\)
\item $[a^b, c] = [c, b]^{a^b} [a, c]^b [b, c]$
\item $[a^b, c] = [b, a] [a, c] [a, b]^c$
\end{enumerate}
\end{lemma}
\begin{prop} \label{prop:simplifyingn-2MBAGraphs}
Let $\Gamma$ be an $(n, n - 2, d)$-MBA graph. Write $\{v_{1}, v_{2}\} = V \setminus \Vtext{max}$. Suppose that $V$ is the disjoint union of $lk(v_{1})$ and $lk(v_{2})$. Then
\begin{enumerate}[(i)]
\item $\Gamma(lk(v_{i}))$ is disconnected for $i = 1, 2$.
\item The normal subgroup
\(
N := \normcl{\{[v, w] \mid v \in lk(v_{1}), w \in lk(v_{2})\}}_{A_{\Gamma}}
\)
is characteristic in $A_{\Gamma}$.
\item With $N$ as above and $\tilde{\Gamma} := \Gamma(lk(v_{1})) * \Gamma(lk(v_{2}))$, we have
\[
\frac{A_{\Gamma}}{N} \cong A_{\tilde{\Gamma}}.
\]
\end{enumerate}
\end{prop}
\begin{proof}
We give the proof for $i = 1$, the case $i = 2$ is analogous. Since $v_{2} \notin lk(v_{2})$ and $lk(v_{1}) \cup lk(v_{2}) = V$, we have $v_{2} \in lk(v_{1})$. Let $w \in lk(v_{1}) \setminus \{v_{2}\}$ be a vertex. Because $lk(v_{1})$ and $lk(v_{2})$ are disjoint, $w$ does not lie in $lk(v_{2})$. This means that $v_{2}$ is not connected to any other vertex in $lk(v_{1})$, implying that $v_{2}$ is an isolated vertex in $\Gamma(lk(v_{1}))$. Hence, $\Gamma(lk(v_{1}))$ is disconnected.
\medskip
Let $\phi \in \Aut(A_{\Gamma})$ be an automorphism of basic type. If $\phi = \imath_{a}$ is an inversion, then for vertices $b \ne c$ different from $a$ we have
\[
\imath_{a}([a, b]) = [\inv{a}, b] = [b, a]^{\inv{a}}
\]
and $\imath_{a}([b, c]) = [b, c]$.
Hence, if $[a, b] \in N$, then $\imath_{a}([a, b]) \in N$ as well.
Next, suppose that $\phi$ is a graph automorphism. As $\phi(\Vtext{max}) = \Vtext{max}$, we have $\phi(\{v_{1}, v_{2}\}) = \{v_{1}, v_{2}\}$. So either $\phi(v_{i}) = v_{i}$ or $\phi(v_{i}) = v_{3 - i}$ for $i = 1, 2$. Let $v \in lk(v_{1})$ and $w \in lk(v_{2})$. In the first case, $\phi(v) \in lk(v_{1})$ and $\phi(w) \in lk(v_{2})$, so $\phi([v, w]) = [\phi(v), \phi(w)] \in N$. In the second case, $\phi(v) \in lk(v_{2})$ and $\phi(w) \in lk(v_{1})$, and hence also $\phi([v, w]) = [\phi(v), \phi(w)] \in N$.
Suppose $\phi = \tau_{ab}$ is a transvection with $a \in lk(v_{1})$; the case $a \in lk(v_{2})$ is analogous and since $lk(v_{1}) \cup lk(v_{2}) = V$, this covers all cases. Let $w \in lk(v_{2})$. We have to show that $\tau_{ab}([a, w]) = [ab, w]$ lies in $N$. As $a \in lk(v_{1})$, we have $v_{1} \in lk(a) \subseteq st(b)$, where the latter inclusion follows from $a \leq b$.
First, suppose that $b \ne v_{1}$. Then $v_{1} \in lk(b)$, so $b \in lk(v_{1})$, implying that $[b, w] \in N$. Thus,
\[
\tau_{ab}([a, w]) = [ab, w] = [a, w]^{b}[b, w] \in N.
\]
Now, if $b = v_{1}$, then $\deg(a) \leq \deg(v_{1}) < \Delta(\Gamma)$, implying that $a = v_{2}$, as $V \setminus \Vtext{max} = \{v_{1}, v_{2}\}$. Hence, $lk(v_{2}) \subseteq st(v_{1})$. As $lk(v_{1})$ and $lk(v_{2})$ are disjoint, this implies that $lk(v_{2}) = \{v_{1}\}$. Consequently, $lk(v_{1}) = V \setminus \{v_{1}\}$ and thus $\deg(v_{1}) = n - 1$, which contradicts $v_{1} \notin \Vtext{max}$.
Finally, suppose $\phi = \gamma_{b, C}$ is a partial conjugation. Let $v \in lk(v_{1})$ and $w \in lk(v_{2})$. If $v$ and $w$ are both (not) conjugated by $b$ via $\phi$, we are done. So, suppose $\phi(v) = v^{b}$ and $\phi(w) = w$. By \cref{lem:commutatorIdentities}, we have
\begin{align}
[v^{b}, w] &= [w, b]^{v^{b}} [v, w]^{b} [b, w] \label{eq:partialConjugationMBAn-2Case1}\\
&= [b, v][v, w][v, b]^{w} \label{eq:partialConjugationMBAn-2Case2}.
\end{align}
If $b \in lk(v_{1})$, then \eqref{eq:partialConjugationMBAn-2Case1} implies that $[v^{b}, w] \in N$. If $b \in lk(v_{2})$, then the result follows from \eqref{eq:partialConjugationMBAn-2Case2}. This finishes the proof that $N$ is characteristic in $A_{\Gamma}$.
The isomorphism is a direct application of \cref{prop:addingEdges}.
\end{proof}
\begin{cor}
Let $\Gamma$ be an $(n, n - 2, d)$-MBA graph satisfying the same conditions as in \cref{prop:simplifyingn-2MBAGraphs}. Then $A_{\Gamma} \in R_\infty$.
\end{cor}
\begin{proof}
We use the same notation as before. Since $\Gamma(lk(v_{1}))$ is disconnected, \cref{theo:RinfDisconnectedGraphs} implies that $A_{\Gamma(lk(v_{1}))} \in R_\infty$. Putting $\tilde{\Gamma} = \Gamma(lk(v_{1})) * \Gamma(lk(v_{2}))$, we note that this is the maximal decomposition of $\tilde{\Gamma}$, since $\Gamma(lk(v_{1}))$ and $\Gamma(lk(v_{2}))$ are both disconnected graphs. Therefore, \cref{theo:directProductsRAAGsRinf} implies that $A_{\tilde{\Gamma}} \in R_\infty$ as well. Finally, since $A_{\tilde{\Gamma}}$ is a characteristic quotient of $A_{\Gamma}$, we conclude that $A_{\Gamma}$ has the $R_\infty$-property.
\end{proof}
\begin{example}
We provide two examples of $(n, n - 2, d)$-MBA graphs: one satisfying the conditions of \cref{prop:simplifyingn-2MBAGraphs} and one not.
Consider the graph ${\Gamma_{1} := (K^{1} \sqcup K^{2}) * (K^{1} \sqcup K^{2})}$ (see \Cref{fig:MBAgraph6vertices}; one copy of $K^{1} \sqcup K^{2}$ has solid vertices, the other one hollow). From \cref{ex:simplicialJoinDisjointUnionCompleteGraphs}, we know that $\Gamma_{1}$ is a $(6, 4, 4)$-MBA graph.
\begin{figure}[h]
\centering
\begin{subfigure}[h]{0.4\textwidth}
\centering
\begin{tikzpicture}
\GraphInit[vstyle = Simple]
\SetVertexMath
\SetUpVertex[LineWidth = 0.5pt]
\tikzset{VertexStyle/.append style = {shape = circle,fill = white,minimum size = 4pt,inner sep = 0pt}}
\SetUpEdge[lw = 0.4pt]
\grEmptyCycle[RA = 1.5, prefix = a]{3}
\tikzset{VertexStyle/.append style = {fill = black}}
\grEmptyCycle[RA = 1.5, prefix = b, rotation = 180]{3}
\EdgeFromOneToAll{a}{b}{0}{3}
\EdgeFromOneToAll{a}{b}{1}{3}
\EdgeFromOneToAll{a}{b}{2}{3}
\EdgeFromOneToSel{a}{a}{1}{2}
\EdgeFromOneToSel{b}{b}{1}{2}
\tikzset{AssignStyle/.append style = {right = 1pt}}
\AssignVertexLabel[color = black, size = \small]{a}{$w_{0}$, , }
\tikzset{AssignStyle/.append style = {left = 1pt}}
\AssignVertexLabel[color = black, size = \small]{b}{$w_{1}$, ,}
\tikzset{AssignStyle/.append style = {above = 1pt}}
\AssignVertexLabel[color = black, size = \small]{a}{, $v_{2}$, }
\AssignVertexLabel[color = black, size = \small]{b}{, ,$v_{1}$ }
\tikzset{AssignStyle/.append style = {below = 1pt}}
\AssignVertexLabel[color = black, size = \small]{b}{, $v_{0}$, }
\AssignVertexLabel[color = black, size = \small]{a}{, , $v_{3}$}
\end{tikzpicture}
\caption{$(K^{1} \sqcup K^{2}) * (K^{1} \sqcup K^{2})$}
\label{fig:MBAgraph6vertices}
\end{subfigure}
\begin{subfigure}[h]{0.4\textwidth}
\centering
\begin{tikzpicture}
\GraphInit[vstyle = Simple]
\SetVertexMath
\tikzset{VertexStyle/.style = {shape = circle,fill = black,minimum size = 4pt,inner sep=0pt}}
\SetUpEdge[lw = 0.4pt]
\grComplete[y = 2, RA = 1, prefix = w, rotation = 0]{2}
\grEmptyCycle[RA = 1.5, prefix=v, rotation = 90]{5}%
\EdgeFromOneToSel{w}{v}{0}{0,4}
\EdgeFromOneToSel{w}{v}{1}{0,1}
\tikzset{AssignStyle/.append style = {left = 1pt}}
\AssignVertexLabel[color = black, size = \small]{v}{, $v_{1}$, $v_{2}$, ,}
\tikzset{AssignStyle/.append style = {right = 1pt}}
\AssignVertexLabel[color = black, size = \small]{v}{, , , $v_{3}$ ,$v_{4}$}
\tikzset{AssignStyle/.append style = {below right = .5pt}}
\AssignVertexLabel[color = black, size = \small]{v}{$v_{0}$, , , ,}
\tikzset{AssignStyle/.append style = {left = 1pt}}
\AssignVertexLabel[color = black, size = \small]{w}{, $w_{1}$}
\tikzset{AssignStyle/.append style = {right = 1pt}}
\AssignVertexLabel[color = black, size = \small]{w}{$w_{0}$, }
\EdgeInGraphSeq{v}{1}{3}
\EdgeInGraphMod{v}{5}{2}
\end{tikzpicture}
\caption{$(7, 5, 4)$-MBA graph $\Gamma_{2}$}
\label{fig:MBAgraph7vertices}
\end{subfigure}
\caption{Further examples of MBA-graphs}
\end{figure}
The vertices in $V \setminus \Vtext{max}$ are $w_{0}$ and $w_{1}$, for which we have
\[
lk(w_{0}) = \{w_{1}, v_{0}, v_{1}\}, \quad lk(w_{1}) = \{w_{0}, v_{2}, v_{3}\},
\]
showing that $\Gamma_{1}$ satisfies the conditions of \cref{prop:simplifyingn-2MBAGraphs}.
On the other hand, consider $\Gamma_{2}$ as in \Cref{fig:MBAgraph7vertices}. Then $V \setminus \Vtext{max} = \{w_{0}, w_{1}\}$, but $v_{0}$ is adjacent to both $w_{0}$ and $w_{1}$. Moreover, $v_{2}$ lies in neither $lk(w_{0})$ nor $lk(w_{1})$.
\end{example}
For $(n, n - 2, d)$-MBA graphs not satisfying the conditions of \cref{prop:simplifyingn-2MBAGraphs}, we can nonetheless find a suitable characteristic vertex subgroup.
\begin{lemma} \label{lem:MBAIntersectionsUnionsLinks}
Let $\Gamma$ be a max-by-abelian graph. Then
\begin{equation} \label{eq:MBAintersectionLinksSubsetVmax}
\bigcap_{v \notin \Vtext{max}} lk(v) \subsetneq \Vtext{max}
\end{equation}
and
\begin{equation} \label{eq:MBAunionIntersectsVmax}
\Vtext{max} \cap \left(\bigcup_{v \notin \Vtext{max}} lk(v)\right) \ne \emptyset.
\end{equation}
\end{lemma}
\begin{proof}
In \cref{prop:intersectionLinkNonMaximalDegreeCharacteristic}, we have proven that the inclusion of \eqref{eq:MBAintersectionLinksSubsetVmax} holds, albeit not necessarily strict. Suppose equality holds. Then for ${v \in V \setminus \Vtext{max}}$, we have $\Vtext{max} \subseteq lk(v)$. As $\Gamma(V \setminus \Vtext{max})$ is complete, we also have
\[
V \setminus (\Vtext{max} \cup \{v\}) \subseteq lk(v).
\]
Consequently, $V \setminus \{v\} \subseteq lk(v)$, contradicting $v \notin \Vtext{max}$. Hence, the inclusion in \eqref{eq:MBAintersectionLinksSubsetVmax} is strict.
For \eqref{eq:MBAunionIntersectsVmax}, note that since $\Gamma$ is connected, there is a $v \notin \Vtext{max}$ that is adjacent to a vertex in $\Vtext{max}$. Hence, the intersection in \eqref{eq:MBAunionIntersectsVmax} is non-empty.
\end{proof}
\begin{prop} \label{prop:characteristicQuotientn-2MBAGraph}
Let $\Gamma$ be an $(n, n - 2, d)$-MBA graph and put $V \setminus \Vtext{max} = \{v_{1}, v_{2}\}$.
\begin{enumerate}[(i)]
\item If $lk(v_{1}) \cap lk(v_{2}) \ne \emptyset$, then $\Gamma\big(V \setminus (lk(v_{1}) \cap lk(v_{2}))\big)$ is non-complete. \label{item:nonDisjointLinks}
\item If $lk(v_{1}) \cup lk(v_{2}) \ne V$, then $\Gamma\big(V \setminus (\Vtext{max} \cap (lk(v_{1}) \cup lk(v_{2})))\big)$ is non-complete. \label{item:nonCoveringLinks}
\end{enumerate}
In either case, $A_{\Gamma}$ has a non-abelian RAAG $A_{\Gamma'}$ as characteristic quotient with $|V'| < |V|$.
\end{prop}
\begin{proof}
Suppose $lk(v_{1}) \cap lk(v_{2}) \ne \emptyset$. Note that $v_{1}, v_{2} \in V \setminus (lk(v_{1}) \cap lk(v_{2}))$ and because $lk(v_{1}) \cap lk(v_{2}) \subsetneq \Vtext{max}$ by \cref{lem:MBAIntersectionsUnionsLinks}, there is a vertex $v \in \Vtext{max} \setminus (lk(v_{1}) \cap lk(v_{2}))$. Then by definition of $v$, either $vv_{1} \notin E$ or $vv_{2} \notin E$, showing that $\Gamma(V \setminus (lk(v_{1}) \cap lk(v_{2}))$ is non-complete.
On the other hand, suppose $lk(v_{1}) \cup lk(v_{2}) \ne V$. As both $v_{1}$ and $v_{2}$ lie in $lk(v_{1}) \cup lk(v_{2})$, there is a $v \in \Vtext{max}$ such that $v \notin lk(v_{1}) \cup lk(v_{2})$. Then $v, v_{1}$ and $v_{2}$ all lie in $V \setminus (\Vtext{max} \cap (lk(v_{1}) \cup lk(v_{2})))$ and by definition of $v$ is $vv_{1} \notin E$. Hence, $\Gamma(V \setminus (\Vtext{max} \cap (lk(v_{1}) \cup lk(v_{2}))))$ is non-complete.
\medskip
The claim regarding the characteristic quotients follows for \eqref{item:nonDisjointLinks} from \cref{lem:MBAIntersectionsUnionsLinks} and \cref{prop:intersectionLinkNonMaximalDegreeCharacteristic}. For \eqref{item:nonCoveringLinks}, it follows from \cref{lem:MBAIntersectionsUnionsLinks} and \cref{prop:intersectionVmaxUnionLinksMBA}.
\end{proof}
\section{Partial answer to conjecture}
\addtocounter{subsection}{1}
Combining all the obtained results, we are able to partially answer the conjecture.
\begin{theorem}
Let \(\Gamma\) be a non-complete graph on at most \(7\) vertices. Then \(A_{\Gamma}\) has the \(R_\infty\)-property.
\end{theorem}
The proof boils down to applying the Simplification Lemma and showing that the results from \cref{sec:freeProductsRAAGs,sec:regularGraphs,sec:MBA} cover all remaining cases when \(\Gamma\) has at most \(7\) vertices, see \cite[p.~155]{SendenThesis} for the details.
\rmfamily
\printbibliography[heading=bibintoc]
\end{document}
| 2105ca7e872460fa498c1a552f3c6a3d355a0255 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Today, algorithmic systems driven by large amounts of data
are increasingly being used in all aspects of life.
Often, such systems are being used to assist, or, even replace
human decision-making.
This increased dependence on algorithms has given rise to the field of algorithmic fairness, where the goal is to ensure that algorithms do not exhibit biases towards specific individuals, or groups of users (see e.g., \cite{fairness-study} for a survey).
We also live in a connected world where networks, be it, social, communication,
interaction, or cooperation networks, play a central role.
However, surprisingly, fairness in networks
has received less attention.
Link analysis algorithms, such as Pagerank~\cite{pagerank}, take a graph as input and use the structure of the graph to determine the relative importance of its nodes.
The output of the algorithms is a numerical weight for each node that reflects its importance. The weights are
used to produce a ranking of the nodes, but also as input features in a variety of
machine learning algorithms including classification~\cite{spam-classifier}, and
search result ranking~\cite{pagerank}.
Pagerank performs a random walk on the input graph, and weights the nodes according to the stationary probability distribution of this walk.
At each step, the random walk restarts with probability $\gamma$, where
the restart node is selected according to a``jump'' distribution vector $\mathbf{v}$.
Since its introduction
in the Google search engine, Pagerank has been the
cornerstone algorithm in several applications (see, e.g., \cite{pagerank-survey}).
Previous research on the fairness of centrality measures has considered only degrees and found biases that arise as a network
evolves
\cite{glass-ceiling,glass-ceiling-recommend}, or has studied general notions of fairness in graphs based on the premise that similar nodes should get similar outputs \cite{inform}.
In this work, we focus on the fairness of the Pagerank algorithm.
As in previous research, we view fairness as lack of discrimination against a
protected group defined by the value of a sensitive attribute, such as, gender, or race \cite{fairness-study}.
We operationalize this view by saying that a link analysis algorithm is \textit{$\phi$-fair}, if the fraction of the total weight allocated to the members of the protected group is $\phi$.
The value of $\phi$ is a parameter that can be used to implement different fairness policies. For example, by setting $\phi$ equal to the ratio of the protected nodes in the graph, we ask that the protected nodes have a share in the weights proportional to their share in the population, a property also known as demographic parity
\cite{fairness-awarness}.
We also consider \emph{targeted} fairness, where we focus on a specific subset of nodes to which we want to allocate weights in a fair manner.
We revisit Pagerank through the lens of our fairness definitions,
and we consider the problem of defining families of Pagerank algorithms that are fair.
We also define the \emph{utility loss} of a fair algorithm as the difference between its output and the output of the original Pagerank algorithm, and we pose the problem of achieving fairness while minimizing utility.
We consider two approaches for achieving fairness.
The first family of algorithms we consider is the \emph{fairness-sensitive} Pagerank family which exploits the jump vector $\mathbf{v}$.
There has been a lot of work on modifying the jump vector to obtain variants of Pagerank biased towards a specific set of nodes.
The topic-sensitive Pagerank algorithm \cite{topic-sensitive} is such an example, where the probability is assigned to nodes of a specific topic.
In this paper, we take the novel approach of using the jump vector to achieve $\phi$-fairness.
We determine the conditions under which this is feasible
and formulate the problem of finding the jump vector that achieves $\phi$-fairness while minimizing utility loss
as a convex optimization problem.
Our second family of algorithms takes a microscopic view of fairness by looking at the behavior of each individual node in the graph.
Implicitly, a link analysis algorithm assumes that links in the graph correspond to endorsements between the nodes. Therefore, we can view each node, as an agent that \emph{endorses} (or \emph{votes for}) the nodes that it links to. Pagerank defines a process that takes these individual actions of the nodes and transforms them into a global weighting of the nodes.
We thus introduce, the \textit{locally fair PageRank algorithms}, where each individual node acts fairly by distributing its own pagerank to the protected and non-protected groups according to the fairness ratio $\phi$.
Local fairness defines a dynamic process that can be viewed as a \textit{fair random walk}, where \emph{at each step} of the random walk (not only at convergence), the probability of being at a node of the protected group is $\phi$.
In our first locally fair PageRank algorithm, termed the \textit{neighborhood locally fair} Pagerank algorithm, each node distributes its pagerank fairly among its immediate neighbors, allocating a fraction $\phi$ to the neighbors in the protected group, and $1-\phi$ to the neighbors in the non-protected group.
The \textit{residual-based locally fair} Pagerank algorithms generalizes this idea.
Consider a node $i$ that has less neighbors in the protected group than $\phi$. The node distributes an equal portion of its pagerank to each of its neighbors and a residual portion $\delta(i)$
to members in the protected group but not necessarily in its own neighborhood.
The residual is allocated based on \textit{a residual redistribution policy}, which allows us to control the fairness policy.
In this paper, we exploit a residual redistribution policy that minimizes the utility loss.
We then define a stronger fairness requirement, termed universal personalized fairness, that asks that the derived personalized pageranks of all nodes are fair. We prove that the locally fair algorithms achieve also universal personalized fairness.
Surprisingly, the locally fair algorithms are the \emph{only} family of algorithms with this property. Thus, we show that an algorithm is locally fair, if and only if, it is universally personalized fair.
We use real and synthetic datasets to study the conditions under which {{Pagerank}} and personalized {{Pagerank}} are fair.
We also evaluate both quantitatively and qualitatively the output of our fairness-aware algorithms.
In summary, in this paper, we make the following contributions:
\begin{itemize}
\item We initiate a study of fairness for the {{Pagerank}} algorithm.
\item We propose the fairness-sensitive Pagerank family that modifies the jump vector so as to achieve fairness, and the locally fair Pagerank family that guarantees that individually each node behaves in a fair manner.
\item We prove that local fairness implies universal personalized fairness and also that this is the only family of algorithms with this property, establishing an equivalence between local fairness and universal personalized fairness.
\item We perform experiments on several datasets to study the conditions under which Pagerank unfairness emerges and evaluate the utility loss for enforcing fairness.
\end{itemize}
The remainder of this paper is structured as follows.
In Section \ref{sec:definitions}, we provide definitions of fairness and we formulate our problems. In Sections \ref{sec:fairness-sensitive} and
\ref{sec:local-fair}, we introduce the fairness sensitive and
the locally fair families of {{Pagerank}} algorithms.
In Section \ref{sec:universal}, we discuss personalized fairness and we show an equivalence between local and universal personalized fairness. The results of our experimental evaluation are presented in Section \ref{sec:experiments}. Finally, we present related research in Section \ref{sec:related-work} and our conclusions in Section \ref{sec:conclusions}.
\section{Definitions}
\label{sec:definitions}
In this section, we first present background material and then we define Pagerank fairness.
\subsection{Preliminaries}
The {{Pagerank}} algorithm~\cite{pagerank} pioneered link analysis for weighting and ranking the nodes of a graph. It was popularized by its application in the Google search engine, but it has found a wide range of applications in different settings~\cite{pagerank-survey}. The algorithm takes as input a graph $G = (V,E)$, and produces a scoring vector, that assigns a weight to each node $v \in V$ in the graph. The scoring vector is the stationary distribution of a random walk on the graph $G$.
The {{Pagerank}} random walk is a random walk with restarts. It is parameterized by the value $\gamma$, which is the probability that the random walk will restart at any step.
The node from which the random walk restarts is selected according to the jump vector $\mathbf{v}$, which defines a distribution over the nodes in the graph.
Typically, the jump probability is set to $\gamma = 0.15$, and the jump vector is set to the uniform vector.
The ``organic'' part of the random walk is governed by the transition matrix $\mathbf{P}$, which defines the transition probability $P[i,j]$ between any two nodes $i$ and $j$. The transition matrix is typically defined as the normalized adjacency matrix of graph $G$. Special care is required for the sink nodes in the graph, that is, nodes with no outgoing edges. In our work, we adopt the convention that, when at a sink node, the random walk performs a jump to a node chosen uniformly at random~\cite{pagerank-survey}. That is, the corresponding zero-rows in the matrix $\mathbf{P}$ are replaced by the uniform vector.
The Pagerank vector $\mathbf{p}$ satisfies the equation:
\begin{equation}
\mathbf{p}^T= (1-\gamma) \mathbf{p}^T \mathbf{P} + \gamma \, \mathbf{v}^T
\label{eqn:PR}
\end{equation}
It can be computed either by solving the above equation, or by iteratively applying it to any initial probability vector.
The Pagerank algorithm is fully defined by the three parameters we described above: the transition matrix $\mathbf{P}$, the restart probability $\gamma$, and the restart (or jump) vector $\mathbf{v}$. Different settings for these parameters result in different algorithms.
Given a graph $G = (V,E)$, let $\mathcal{PR}(G)$ denote the family of all possible Pagerank algorithms on graph $G$. Each algorithm in $\mathcal{PR}(G)$ corresponds to a triplet $(\mathbf{P}(G), \gamma, \mathbf{v}(G))$ for the parameters of the random walk. This is a very general family that essentially includes all possible random walks defined over the nodes of graph $G$.
We will refer to the algorithm that uses the typical settings as the \emph{original} Pagerank algorithm, $\textsc{OPR}$, and use $\mathbf{p}_O$ to denote its pagerank vector.
Several variations of the original Pagerank algorithm have been proposed, that modify the above parameters to achieve different goals~\cite{pagerank-survey}.
We are interested in defining \emph{fair} Pagerank algorithms.
\subsection{Fair Pagerank}
We focus on graphs where a set of nodes defines a protected group based on the value of some protected attribute.
For example, in the case of social, collaboration, or citation networks where each node is a person, protected attributes may correspond to gender, age, race, or religion. In the following for simplicity, we assume binary such attributes, but our algorithms can be extended for the general case.
We consider two types of nodes, red and blue nodes, and the corresponding groups denoted $R$ and $B$ respectively. Group $R$ is the protected group.
We denote with $r = \frac{|R|}{n}$, and $b = \frac{|B|}{n}$, the fraction of nodes that belong to the red and blue group respectively.
Let $\textsc{PR}\in\mathcal{PR}(G)$ be a Pagerank algorithm on graph $G$. We will use $\textsc{PR}(u)$ to denote the pagerank mass that $\textsc{PR}$ assigns to node $u$, and, abusing the notation, $\textsc{PR}(R)$ to denote the total pagerank mass that $\textsc{PR}$ assigns to the nodes in the red group (for the blue group, $\textsc{PR}(B) = 1-\textsc{PR}(R))$.
We will say that $\textsc{PR}$ is \emph{fair}, if it assigns weights to each group according to a specified ratio $\phi$.
\begin{definition} [Pagerank Fairness]
Given a graph $G = (V,E)$ containing the protected group $R\subseteq V$, and a value $\phi \in (0,1)$, a Pagerank algorithm $\mathrm{PR} \in \mathcal{PR}(G)$ is $\phi$-fair on graph $G$, if $\mathrm{PR}(R) = \phi$.
\end{definition}
The ratio $\phi$ may be specified so as to implement specific affirmative action policies,
or other fairness enhancing interventions.
For example, $\phi$ may be set in accordance to the 80-percent rule advocated by the US
Equal Employment Opportunity Commission (EEOC), or some other formulation of disparate impact \cite{disparate-impact}.
Setting $\phi$ = $r$, we ask for a fair Pagerank algorithm that assigns weights proportionally to the sizes of the two groups.
In this case, fairness is analogous to demographic parity, i.e., the requirement that the demographics of those receiving a
positive outcome are identical to the demographics of the
population as a whole \cite{fairness-awarness}.
Our goal is to define fair Pagerank algorithms. We say that a \emph{family} of Pagerank algorithms $\textsl{FPR} \subseteq \mathcal{PR}(G)$ is $\phi$-fair if all the Pagerank algorithms in the family are $\phi$-fair. The first problem we consider is to find such families of algorithms.
\begin{problem}
Given a graph $G = (V,E)$ containing a protected group of nodes $R \subseteq V$, and a value $\phi \in (0,1)$, define a family of algorithms $\textsl{FPR} \subseteq \mathcal{PR}(G)$ that is $\phi$-fair.
\end{problem}
We can achieve fairness by modifying the parameters of the Pagerank algorithm.
For the following, we assume the jump probability $\gamma$ to be fixed, and we only consider modifications to the transition matrix $\mathbf{P}$ and the jump vector $\mathbf{v}$. A $\phi$-fair family of algorithms is defined by a specific process, parameterized by $\phi$, for defining $\mathbf{P}$ and $\mathbf{v}$.
A fair Pagerank algorithm will clearly output a different weight vector than the original Pagerank algorithm.
We assume that the weights of the original Pagerank algorithm carry some \emph{utility}, and use these weights to measure the \emph{utility loss} for achieving fairness. Concretely, if $\mathbf{f}$ is the output of a fair Pagerank algorithm $\textsc{FPR}$ and $\mathbf{p}_O$ is the output of the original Pagerank algorithm $\textsc{OPR}$, we define the utility loss of $\textsc{FPR}$ as: $L(\textsc{FPR}) = L(\mathbf{f},\mathbf{p}_O) = \|\mathbf{f}-\mathbf{p}_O\|^2$.
The second problem we consider is finding a fair algorithm that minimizes the utility loss.
\begin{problem}
Given a $\phi$-fair family of algorithms $\textsl{FPR} \subset \mathcal{PR}(G)$, find an algorithm $\mathrm{PR}$ $\in \textsl{FPR}$ that minimizes the utility loss $L(\mathrm{PR})$.
\end{problem}
Finally, we introduce an extension of the fairness definition, termed \emph{targeted fairness}, that asks for a fair distribution of
weights among a specific set of nodes $S$ that is given as input. The subset $S$ contains a protected group of nodes $S_R$.
Targeted fairness asks that a fraction $\phi$ of the pagerank mass that $\textsc{PR}$ assigns to $S$ goes to the protected group $S_R$.
For example, assume a co-authorship graph $G$, where $S$ is the set of authors of a particular male-dominated field. We want to allocate to the female authors $S_R$ in this field a fraction $\phi$ of the total pagerank allocated to $S$.
\begin{definition} [Targeted Fairness]
Given a graph $G = (V,E)$, a subset of nodes $S\subseteq V$ containing a protected group $S_R\subseteq S$, and a value $\phi \in (0,1)$, a Pagerank algorithm $\mathrm{PR}\in \mathcal{PR}(G)$, is targeted $\phi$-fair on the subset $S$ of $G$, if $\mathrm{PR}(S_R) = \phi \mathrm{PR}(S)$.
\end{definition}
The two problems we defined above can also be defined for targeted fairness.
\section{Fairness Sensitive PageRank}
\label{sec:fairness-sensitive}
Our first family of algorithms achieves fairness by keeping the transition matrix $\mathbf{P}$ fixed and changing the jump vector $\mathbf{v}$ so as to meet the fairness criterion. We denote this family of algorithms as the \emph{Fairness-Sensitive Pagerank} (\textsl{FSPR}) algorithms.
\subsection{The \textsl{FSPR} Algorithm}
First, we note that that pagerank vector $\mathbf{p}$ can be written as a linear function of the jump vector $\mathbf{v}$.
Solving Equation (\ref{eqn:PR}) for $\mathbf{p}$,
we have that $\mathbf{p}^T = \mathbf{v}^T\mathbf{Q}$, where
\[
\mathbf{Q} = \gamma \left[\mathbf{I} - (1-\gamma)\mathbf{P} \right]^{-1}
\]
Note that if we set $\mathbf{v} = \mathbf{e}_j$, the vector with $\mathbf{e}_j[j] = 1$ and zero elsewhere, then $\mathbf{p}^T = \mathbf{Q}_{j}^T$, the $j$-th row of matrix $\mathbf{Q}$. Therefore, the row vector $\mathbf{Q}_{j}^T$ corresponds to the personalized pagerank vector of node $j$. The pagerank vector $\mathbf{p}$ is a linear combination of the personalized pagerank vectors of all nodes, as defined by the jump vector.
Let $\mathbf{p}[R]$ denote the pagerank mass that is allocated to the nodes of the protected category.
We have that
$$
\mathbf{p}[R] = \sum_{i\in R} \left(\mathbf{v}^T \mathbf{Q} \right) [i]
= \mathbf{v}^T \left( \sum_{i\in R} \mathbf{Q}_{i} \right)
= \mathbf{v}^T \mathbf{Q}_R
$$
where $\mathbf{Q}_{i}$ is the $i$-th column of matrix $\mathbf{Q}$, and $\mathbf{Q}_R$ is the vector that is the sum of the columns of $\mathbf{Q}$ in the set $R$. $\mathbf{Q}_R[j]$ is the total personalized pagerank that node $j$ allocates to the red group.
For the algorithm to be fair, we need $\mathbf{p}[R] = \phi$.
Thus, our goal is to find a jump vector $\mathbf{v}$ such that $\mathbf{v}^T \mathbf{Q}_R = \phi$. Does such a vector always exist? We prove the following:
\begin{lemma}
\label{lemma:solution-feasibiliy}
Given the vector $\mathbf{Q}_R$, there exists a vector $\mathbf{v}$ such that $\mathbf{v}^T \mathbf{Q}_R = \phi$, if and only if, there exist nodes $i,j$ such that $\mathbf{Q}_R[i] \leq \phi$ and $\mathbf{Q}_R[j] \geq \phi$
\end{lemma}
\begin{proof}
We have
$
\mathbf{p}[R]= \mathbf{v}^T \mathbf{Q}_R = \sum_{j = 1}^N \mathbf{v}[j]\mathbf{Q}_R[j]
$,
with $0 \leq \mathbf{v}[j] \leq 1$. If $\mathbf{v}^T \mathbf{Q}_R = \phi$, there must exist $i,j$ with $\mathbf{Q}_R[i] \leq \phi$, and $\mathbf{Q}_R[j] \geq \phi$.
Conversely, if there exists two such entries $i,j$, then we can find values $\pi_i$ and $\pi_j$, such that
$\pi_i \mathbf{Q}_R[i] + \pi_j \mathbf{Q}_R[j] = \phi$ and $\pi_i + \pi_j = 1$.
\end{proof}
\noindent {\bf Complexity.} Note that there is a very efficient way to compute the personalized pagerank, $\mathbf{Q}_R[j]$, that node $j$ allocates to the red group. We can add a red and a blue absorbing node in the random walk and estimate the probability for each node $j$ to be absorbed by the corresponding absorbing node. This can be done in the time required for running Pagerank. Thus, it is possible to compute the $\mathbf{Q}_R$ vector without doing matrix inversion to compute $\mathbf{Q}$.
\subsection{Minimizing Utility Loss}
\label{sec:optimization-fair-sensitive}
An implication of Lemma~\ref{lemma:solution-feasibiliy} is that, in most cases, there are multiple jump vectors that give a fair pagerank vector. We are interested in the solution that minimizes the utility loss.
To solve this problem we exploit the fact that the utility loss function $L(\mathbf{p}_\mathbf{v},\mathbf{p}_O) = \|\mathbf{p}_\mathbf{v} - \mathbf{p}_O\|^2$, where $\mathbf{p}_\mathbf{v}$ is the fair pagerank vector and $\mathbf{p}_O$ the original vector, is convex. We also can express the fairness requirement as a linear constraint. Thus, we define the following convex optimization problem.
\begin{equation*}
\begin{aligned}
& \underset{\mathbf{x}}{\text{minimize}} & & \|\mathbf{x}^T \mathbf{Q} - \mathbf{p}_O\|^2 \\
& \text{subject to} & & \mathbf{x}^T \mathbf{Q}_R = \phi\\
& & & \sum_{i = 1}^{n} \mathbf{x}[i] = 1 \\
& & & 0 \leq \mathbf{x}[i] \leq 1, \; i = 1, \ldots, n \\
\end{aligned}
\end{equation*}
This problem can be solved using standard convex optimization solvers.
In our work, we used the CVXOPT software package\footnote{https://cvxopt.org/}.
The complexity of the algorithm is dominated by the computation of matrix $\mathbf{Q}$ which requires a matrix inversion. We can speed up this process by exploiting the fact that the rows of $\mathbf{Q}$ are personalized pagerank vectors, which can be computed (in parallel) by performing multiple random walks. We can improve performance further using approximate computation, e.g., \cite{approximatepr}.
\subsection{Targeted Fairness \textsl{FSPR} Algorithm}
We can formulate a similar convex optimization problem for the targeted fairness problem.
Let $\mathbf{Q}_S = \sum_{i\in S} \mathbf{Q}_i$ be the sum of columns of $\mathbf{Q}$ for the nodes in $S$, and $\mathbf{Q}_{S_R} = \sum_{i\in S_R} \mathbf{Q}_i$be the sum of columns of $\mathbf{Q}$ for the nodes in $S_R$. We define a convex optimization problem that is exactly the same as in Section~\ref{sec:optimization-fair-sensitive}, except for the fact that we replace the constraint $\mathbf{x}^T \mathbf{Q}_R = \phi$ with the constraint $\mathbf{x}^T \mathbf{Q}_{S_R} = \phi \mathbf{x}^T \mathbf{Q}_S $.
\section{Locally Fair PageRank}
\label{sec:local-fair}
Our second family of fair Pagerank algorithms, termed \emph{Locally Fair Pagerank}
(\textsl{LFPR}), takes a microscopic view of fairness,
by asking that \textit{each individual node} acts fairly, i.e., each node distributes its own pagerank to red and blue nodes fairly.
In random walk terms, local fairness defines a dynamic process
that can be viewed as a random walk that is
fair at each step, and not just at convergence.
The \textsl{LFPR} contains all Pagerank algorithms, where all rows of the transition matrix $\mathbf{P}$ are $\phi$-fair vectors. That is, for every node $i \in V$, $\sum_{j \in R} P[i,j] = \phi$. Also, the jump vector $\mathbf{v}$ is $\phi$-fair: $\sum_{j \in R} \mathbf{v}[j] = \phi$.
We now consider specific algorithms from the family of locally fair algorithms.
\subsection{The Neighborhood \textsl{LFPR} Algorithm}
We first consider a node that treats its neighbors fairly by allocating a fraction $\phi$ of its pagerank to its red neighbors
and the remaining $1 - \phi$ fraction to its blue neighbors.
In random walk terms, at each node the probability of transitioning to a red neighbor is $\phi$ and the probability of transitioning to a blue neighbor $1-\phi$.
Formally, we define the \textit{neighborhood locally fair pagerank} ({{\sc LFPR$_N$}}) as follows.
Let $out_R(i)$ and $out_B(i)$ be the number of outgoing edges from node $i$ to red nodes and blue nodes respectively. We define $\mathbf{P}_R$ as the stochastic transition matrix that handles transitions to red nodes, or random jumps to red nodes
if such links do not exist:
\[
\mathbf{P}_R[i,j] = \left \{
\begin{tabular}{cl}
$\frac{1}{out_R(i)}$, & if $(i, j)$ $\in$ E and $j$ $\in$ $R$ \\
$\frac{1}{|R|}$, & if $out_R(i) = 0$ and $j$ $\in$ $R$ \\
0, & otherwise \\
\end{tabular}
\right.
\]
The transition matrix
$\mathbf{P}_B$ for the blue nodes is defined similarly. The transition matrix $\mathbf{P}_N$ of the {{\sc LFPR$_N$}} algorithm is defined as:
\begin{center}
$\mathbf{P}_N = \phi \, \mathbf{P}_R + (1 - \phi) \, \mathbf{P}_B
$
\end{center}
We also define a $\phi$-fair jump vector $\mathbf{v}_N$ with $\mathbf{v}_N[i]$ = $\frac{\phi}{|R|}$, if $i \in R$, and
$\mathbf{v}_N[i]$ = $\frac{1-\phi}{|B|}$, if $i \in B$.
The neighborhood locally-fair pagerank vector $\mathbf{p}_N$ is defined as:
\[
\mathbf{p}_N^T = (1 - \gamma) \mathbf{p}_N^T \mathbf{P}_N + \gamma \, \mathbf{v}_N^T
\]
\subsection{The Residual-based \textsl{LFPR} Algorithms}
We consider an alternative fair behavior for individual nodes.
Similarly to the {{\sc LFPR$_N$}} algorithm, each node $i$ acts fairly by respecting the $\phi$ ratio when distributing its own pagerank to red and blue nodes. However, now node $i$
treats its neighbors the same, independently of their color and assigns to each of them the same portion of its pagerank. When a node is in a ``biased'' neighborhood, i.e., the ratio of its red neighbors is different than
$\phi$, to be fair, node $i$ distributes only a fraction of its pagerank to its neigbors, and the remaining portion of its pagerank
to nodes in the underrepresented group. We call the remaining portion \emph{residual} and denote it by $\delta(i)$.
How $\delta(i)$ is distributed to the underrepresented group is determined by a \textit{residual policy}.
Intuitively, this corresponds to a fair random walker that upon arriving at a node $i$, with probability 1-$\delta(i)$ follows one of $i$'s outlinks and with probability $\delta(i)$
jumps to one or more node belonging to the group that is locally underrepresented.
We now describe the algorithm formally.
We divide the (non sink) nodes in $V$ into two sets, $L_R$ and $L_B$, based on the fraction of their red and blue neighbors.
Set $L_R$ includes
all nodes $i$ such that
$out_R(i)/out(i) < \phi$, where $out(i)$ the out-degree of node $i$,
that is, the nodes for which the ratio of red nodes in their neighborhoods is smaller than the required $\phi$ ratio. These nodes have a residual that needs to be distributed to red nodes.
Analogously, $L_B$ includes
all nodes $i$ such that
$out_R(i)/out(i) \geq \phi$,
that have a residual to be distributed to blue nodes.
Consider a node $i$ in $L_R$.
To compute $\delta_R(i)$ note that each neighbor of $i$ gets a fraction $\rho_R(i) = \frac{1-\delta_R(i)}{out(i)}$ of $i$'s pagerank. The residual $\delta_R(i)$ of $i$'s pagerank goes to red nodes. In order for node $i$ to be $\phi$-fair, we have:
\begin{equation}
\frac{1-\delta_R(i)}{out(i)}out_R(i) + \delta_R(i)= \phi \label{eq:excess1}
\end{equation}
Solving for $\delta_R(i)$ and using the fact that $out(i) = out_R(i)+out_B(i)$ we get $\delta_R(i) = \phi - \frac{(1-\phi)\,out_R(i)}{out_B(i)}$, and $\rho_R(i) = \frac{1-\phi}{out_B(i)}$.
\iffalse
Consider a node $i$ in $L_R$.
Each neighbor of $i$ gets the same portion of $i$'s pagerank, let $\rho_R(i)$ be this portion. To attain the $\phi$ ratio, the residual $\delta_R(i)$ of $i$'s pagerank goes to the red nodes. Portions $\rho_R(i)$ and $\delta_R(i)$ must be such that:
\begin{equation}
(1- \phi) \,\, (out_R(i) \, \rho_R(i) + \delta_R(i)) = \phi \,\, (out_B(i) \, \rho_R(i)) \label{eq:excess1}
\end{equation}
\begin{equation}
out_R(i) \,\rho_R(i) + out_B(i) \, \rho_R(i) + \delta_R(i) = 1 \label{eq:excess2}
\end{equation}
From Equations (\ref{eq:excess1}) and (\ref{eq:excess2}), we get $\rho_R(i)$ = $\frac{1-\phi}{out_B(i)}$ and the residual is $\delta_R(i) = \phi - \frac{(1-\phi)\,out_R(i)}{out_B(i)}$.
\fi
Analogously, for a node $i$ in $L_B$, we get a residual $\delta_B(i) = (1 -\phi) - \frac{\phi\,out_B(i)}{out_R(i)}$ that goes to the blue nodes, and $\rho_B(i)$ = $\frac{\phi}{out_R(i)}$.
For a sink node $i$, we assume that $i$ belongs to both $L_R$ and $L_B$ with residual $\delta_R(i)$ = $\phi$ and $\delta_B(i)$ = $1- \phi$.
\vspace*{0.1in}
\noindent \textit{Example.}
Consider a node $i$ with 5 out-neighbors, 1 red and 4 blue, and let $\phi$ be $0.5$.
This is a ``blue-biased''node, that is, $i$ $\in$ $L_R$.
With the residual algorithm, each of $i$'s neighbors gets
$\rho_R(i)$ = $0.5/4 = 1/8$ portion of $i$'s pagerank, resulting in red neighbors getting $1/8$ and blue neighbors $4/8$ of i's pagerank. The residual $\delta_B(i)$ = $3/8$ goes to nodes in the red group so as to attain the $\phi$ ratio and make $i$ fair.
In terms of the random walker interpretation, a random walker that arrives at $i$, with probability $5/8$ chooses one of $i$'s outlinks uniformly at random and
with probability $3/8$ jumps to nodes in the red group. \hfill$\qed$.
Formally, we define $\mathbf{P}_L$
as follows:
\[
\mathbf{P}_L[i, j] = \left \{
\begin{tabular}{cl}
$\frac{1-\phi}{out_B(i)}$, & if $(i, j)$ $\in$ $E$ and $i \in L_R$\\
$\frac{\phi}{out_R(i)}$, & if $(i, j)$ $\in$ $E$ and $i \in L_B$ \\
$0$ & otherwise
\end{tabular}
\right.
\]
\iffals
Let $\mathbf{\delta}_R$ be the vector carrying the red residual, that is,
$\mathbf{\delta}_R[i] = \phi - \frac{(1-\phi)\,out_R(i)}{out_B(i)}$, if $i \in L_R$ and 0 otherwise.
Similarly, let $\mathbf{\delta}_B$ be the vector carrying the blue residual, that is,
$\delta_B(i) = (1 -\phi) - \frac{\phi\,out_B(i)}{out_R(i)}$, if $i \notin L_B$ and 0 otherwise.
We have a total red residual $\Delta_R = \mathbf{p}_L^T \, \mathbf{\delta}_R $ and a total blue residual
$\Delta_B = \mathbf{p}_L^T \, \mathbf{\delta}_B$, where $\mathbf{p}_L$ is the locally fair pagerank vector.
\f
$\mathbf{P}_L$ handles the transitions of nodes to their neighborhood.
To express the residual distribution policy, we introduce matrices $\mathbf{X}$ and $\mathbf{Y}$, that
capture the policy for distributing the residual to red and blue nodes respectively.
Specifically, $\mathbf{X}[i, j]$ denotes the portion of the $\delta_R(i)$ of node $i$ $\in$ $L_R$ that goes to node $j$ $\in$ $R$, and
$\mathbf{Y}[i, j]$ the portion of the $\delta_B(i)$ of node $i$ $\in$ $L_B$ that goes to node $j$ $\in$ $B$.
The locally-fair pagerank vector $\mathbf{p}_L$ is defined as:
\[
\mathbf{p}_L^T = (1 - \gamma) \mathbf{p}_L^T \, (\mathbf{P}_L + \mathbf{X} + \mathbf{Y}) + \gamma \, \mathbf{v}_N^T
\]
\noindent \textbf{Residual Distribution Policies.} The $\mathbf{X}$ and $\mathbf{Y}$ allocation matrices allow us the flexibility to
specify appropriate policies for distributing the residual.
For example,
the {{\sc LFPR$_N$}} algorithm is a special case of the residual-based algorithm, with
\vspace*{-0.05in}
\[
\mathbf{X}_N[i, j] = \left \{
\begin{tabular}{cl}
$\frac{\delta_R(i)}{out_R(i)}$, & if $i \in L_R$, $(i,j) \in E$, and $j \in R$ \\
$\frac{\delta_R(i)}{|R|}$, & if $i \in L_R$, $out_R(i) = 0$, and $j \in R$ \\
0 & otherwise \\
\end{tabular}
\right.
\]
and $\mathbf{Y}_N[i, j]$ defined analogously.
We also consider residual policies where all nodes follow the same policy in distributing their residual, that is, each red node gets the same portion of the red residuals
and each blue node the same portion and the blue residuals. In this case, the residual policy is expressed through two (column) vectors $\mathbf{x}$ and $\mathbf{y}$, with
$\mathbf{x}[j]$ = 0 if $j \in B$ and $\mathbf{y}[j]$ = 0, $j \in R$.
Each node $i \in L_R$ sends a fraction $\delta_R(i)\mathbf{x}[j]$ of its pagerank to red node $j$, while each node $i \in L_B$ sends a fraction $\delta_B(i)\mathbf{y}[j]$ of its pagerank to blue node $j$.
Let $\mathbf{\delta}_R$ be the vector carrying the red residual, and
$\mathbf{\delta}_B$ the vector carrying the blue residual.
We have:
\begin{equation*}
\mathbf{p}_L^T = (1 - \gamma)\mathbf{p}_L^T\, (\mathbf{P}_L + \mathbf{\delta}_R \, \mathbf{x}^T + \mathbf{\delta}_B \, \mathbf{y}^T) + \gamma \, \mathbf{v}_N^T.
\end{equation*}
We define two locally fair {{Pagerank}} algorithms based on two intuitive policies of distributing the residual:
\vspace{0.05in}
\noindent
The \textit{Uniform Locally Fair {{Pagerank}}} ({{\sc LFPR$_U$}}) algorithm distributes the residual uniformly. Specifically,
we define the vector $\mathbf{x}$, as $\mathbf{x}[i]$ = $\frac{1}{|R|}$ for $i \in R$,
and the vector $\mathbf{y}$, as $\mathbf{y}[i]$ = $\frac{1}{|B|}$, for $i \in B$.
\vspace{0.05in}
\noindent
The \textit{Proportional Locally Fair {{Pagerank}}} ({{\sc LFPR$_P$}}) algorithm distributes the residual proportionally to the original pagerank weights $\mathbf{p}_O$
Specifically,
we define the vector $\mathbf{x}$, as
$\mathbf{x}[i]$ = $\frac{ \mathbf{p}_O[i] }{\sum_{i \in R}\mathbf{p}_O[i]}$, for $i \in R$,
and the vector $\mathbf{y}$, as $\mathbf{y}[i]$ =
$\frac{ \mathbf{p}_O[i] }{\sum_{i \in B}\mathbf{p}_O[i]}$, for $i \in B$.
\subsection{Fairness of the \textsl{LFPR} Algorithms}
Although each node acts independently of the other nodes in the network,
this microscopic view of fairness
results in a macroscopic view of fairness. Specifically, we prove the following theorem.
\begin{theorem}
The locally fair {{Pagerank}} algorithms are $\phi$-fair. \label{theorem:local}
\end{theorem}
\iffals
\begin{proof}
We must show that $\sum_{v \in R} \mathbf{p}_N(u) = \phi$.
Since each node in the graph gives a portion $\phi$ of its pagerank to red nodes, we have
\[\sum_{v \in R} \mathbf{p}_N(u) = \sum_{v \in V} \phi \, \mathbf{p}_N(u)
\]
which proves the theorem.
\end{proof}
\fi
\begin{proof}
Let $\mathbf{e}_R$ denote the vector with 1's at the red nodes and zero at the blue nodes. The amount of pagerank that vector $\mathbf{p}_L$ gives to the red nodes can be expressed as $\mathbf{p}_L^T\mathbf{e}_R$. Let $\mathbf{P}_D$ = $\mathbf{P}_L + \mathbf{X} + \mathbf{Y}$, we have:
\[
\mathbf{p}_L^T\mathbf{e}_R= (1-\gamma)\mathbf{p}_L^T \mathbf{P}_D\mathbf{e}_R + \gamma \mathbf{v}_N^T\mathbf{e}_R
\]
By design we have that $\mathbf{v}_N^T\mathbf{e}_R = \phi$. For transition matrix $\mathbf{P}_D$, due to the local fairness property, we know that each row has probability $\phi$ of transitioning to a red node. Therefore, $\mathbf{P}_D\mathbf{e}_R = \phi\mathbf{e}$, where $\mathbf{e}$ is the vector of all ones. Note that since $\mathbf{p}_L$ is a distribution we have that $\mathbf{p}_L^T\mathbf{e} = 1$. We thus have:
$\mathbf{p}_L^T\mathbf{e}_R= (1-\gamma)\phi + \gamma\phi = \phi$.
\end{proof}
\subsection{Minimizing Utility Loss}
We now consider how to distribute the residual so as to minimize the utility loss of the locally fair Pagerank. We denote this algorithm as {{\sc LFPR$_O$}}. To this end,
we compute the $\mathbf{x}$ and $\mathbf{y}$ residual distribution vectors by formulating an optimization problem.
We can write the vector $\mathbf{p}_L$ as a function of the vectors $\mathbf{x}$ and $\mathbf{y}$ as follows:
\[
\mathbf{p}_L^T(\mathbf{x},\mathbf{y}) = \gamma \, \mathbf{v}_N^T \left[\mathbf{I} - (1-\gamma)(\mathbf{P}_L + \delta_R\, \mathbf{x}^T + \delta_B\, \mathbf{y}^T) \right]^{-1}
\]
We can now define the optimization problem of finding the vectors $\mathbf{x}$ and $\mathbf{y}$ that minimize the loss function $L(\mathbf{p}_L, \mathbf{p}_O) = \|\mathbf{p}_L(\mathbf{x},\mathbf{y}) - \mathbf{p}_O\|^2$ subject to the constraint that the vectors $\mathbf{x}$ and $\mathbf{y}$ define a distribution over the nodes in $R$ and $B$ respectively.
Since our problem is not convex, we implement a Stochastic Random Search algorithm for solving it, that looks at randomly selected directions for the gradient, and selects the one that causes the largest decrease.
We enforce the distribution constraints
by adding a penalty term
$\lambda$ $\left ( (\sum_{i = 1}^{n} \mathbf{x}_i - 1)^2 + (\sum_{i = 1}^{n} \mathbf{y}_i - 1)^2\right )$.
We enforce the positivity constraints
through proper bracketing at the line-search step.
The complexity of the algorithm is $O\left ( I\cdot K\cdot T_{PR}\right )$, where $I$ is the number of iterations, $K$ the number of random directions that are examined, and $T_{PR}$ the cost of running Pagerank. In our experiments, $I$ and $K$ are in the order of a few hundreds.
\subsection{Targeted Fairness \textsl{LFPR} Algorithms}
We can apply the locally fair algorithms
to the targeted fairness problem.
Let $S_R$ and $S_B$ be the red and blue nodes in the set $S$ respectively, and let $I_S$ be the set of in-neighbors of $S$.
The idea is that the nodes in $I_S$ should distribute their pagerank to $S_R$ and $S_B$ fairly,
such that the
portion that goes to nodes in $S_R$ is a $\phi$ fraction of the total pagerank that goes to the set $S$.
We can implement the same redistribution policies as in the case of the neighborhood local and the residual-based local fair algorithms.
We also need the (global) jump vector $\mathbf{v}$ to obey the $\phi$ ratio for the nodes in $S$. We can achieve this by redistributing the probability $|S|/n$ of the jump vector according to the $\phi$ ratio.
\section{Personalized Fairness}
\label{sec:universal}
A special case of the {{Pagerank}} algorithm is when the restart vector is defined to be the unit vector $\mathbf{e}_i$ that puts all the mass at a single node $i$. For any {{Pagerank}} algorithm $\textsc{PR} \in \mathcal{PR}$, we can define the corresponding personalized algorithm $\textsc{PR}_i$ by setting $\mathbf{v} = \mathbf{e}_i$.
The output of the algorithm $\textsc{PR}_i$ is a probability vector, where $\textsc{PR}_i(u)$ is the probability that a random walk that always restarts at node $i$ is at $u$ after infinite number of steps. We say that node $i$ allocates this probability to node $u$. Personalized random walks have found several applications in network analysis~\cite{pagerank-survey}. For example, the probability $\textsc{PR}_i(u)$ can be seen as a measure of proximity between node $i$ and node $u$, and it has been used for recommendations.
For a personalized Pagerank algorithm $\textsc{PR}_i$, we define $\textsc{PR}_i(R)$ and $\textsc{PR}_i(B)$ to be the probability that node $i$ allocates to the red and blue groups respectively. Recall that if $\mathbf{v}$ is the jump vector, then $\textsc{PR}(R) = \sum_{i\in V} \mathbf{v}[i] \textsc{PR}_i(R)$. We can think of the probability $\textsc{PR}_i(R)$, as a measure of how a specific node $i$ ``views'' the red group, while $\textsc{PR}(R)$ captures the value that the network places on the red nodes on aggregate.
We could define fairness for the $\textsc{PR}_i$ algorithm using the standard fairness definition. However, note that since the random walk jumps with probability $\gamma$ to node $i$ at every step, this immediately adds probability $\gamma$ to the category of node $i$. This probability is due to the random jump and not due to the ''organic'' random walk, and the structure of the graph. We thus subtract this probability, and we define the vector $\overline{\textsc{PR}_i}$, where $\overline{\textsc{PR}_i}(i) = \textsc{PR}_i(i)-\gamma$, and $\overline{\textsc{PR}_i}(u) = \textsc{PR}_i(u)$, for $u \neq i$. Another way to think of this is that an amount $\gamma$ of probability is reserved for restarting, and the remaining $1-\gamma$ is allocated through the random walk. This is the probability mass that we want to allocate fairly. We say that the \emph{personalized} Pagerank algorithm $\textsc{PR}_i$ is $\phi$-fair if $\overline{\textsc{PR}_i}(R) = \phi(1-\gamma)$.
Intuitively, fairness of $\textsc{PR}_i$ implies that node $i$ treats the red and blue groups fairly. For example, if we think of $\overline{\PR_i}(R)$ as a proximity measure, and $\phi = 0.5$, $\phi$-fairness implies that node $i$ is equally close to the red and blue groups. Given that this probability is often used for recommendations, this has also implications to the fairness of the recommendations.
\begin{figure}[]
\includegraphics[width = \columnwidth]{books_hist_sep2.pdf}
\caption{Histogram of personalized red ratio values.}
\label{books-histogram}
\end{figure}
Note that a Pagerank algorithm $\textsc{PR}$ may be fair, while the corresponding personalized Pagerank algorithms are not. In Figure~\ref{books-histogram}, we consider the original Pagerank algorithm $\textsc{OPR}$, and we show the histogram of the $\overline{\textsc{OPR}_i}(R)$ values for the \textsc{books} dataset (described in Section~\ref{sec:experiments}), for the red and blue nodes, in red and blue respectively. For this dataset, we have that $r = 0.47$ and $\textsc{OPR}(R)$ is 0.46. That is, the original Pagerank algorithm is almost fair for $\phi = r$. However, we observe that the distribution of the $\overline{\textsc{OPR}_i}(R)$ values is highly polarized.
The values for the red nodes (shown in red) are concentrated in the interval $[0.8,1]$, while the values for the blue nodes (in blue) are concentrated in the interval $[0,0.2]$. Thus, although the network as a whole has a fair view of the two groups, the individual nodes are highly biased in favor of their own group.
Motivated by this observation, we consider a stronger definition of fairness, where given an algorithm $\textsc{PR}$ we require that \emph{all} derivative Personalized Pagerank versions of this algorithm are fair. That is, it is not sufficient that the algorithm treats the red group fairly on aggregate, but we require that each individual node is also fair.
\begin{definition}[Universal Personalized Fairness]
Given a graph $G = (V,E)$ containing the protected group $R\subseteq V$, and a value $\phi \in (0,1)$, a Pagerank algorithm $\textsc{PR} \in \mathcal{PR}(G)$ is universally personalized $\phi$-fair on graph $G$, if for every node $i \in V$, the personalized Pagerank algorithm $\textsc{PR}_i$ is personalized $\phi$-fair.
\end{definition}
Since we want all personalized algorithms to be fair, universal personalized fairness (universal fairness for short) is a property of the transition matrix $\mathbf{P}$ of the Pagerank algorithm. Since the \textsl{FSPR} family does not change the matrix $\mathbf{P}$, it is not universally fair. We will show that the locally fair algorithms are universally fair. Furthermore, we can prove that universally fair algorithms are locally fair. Therefore, universal fairness is equivalent to local fairness.
\begin{theorem}
A Pagerank algorithm $\textsc{PR}$ is universally personalized $\phi$-fair if and only if it is locally fair.
\end{theorem}
\begin{proof}
We first prove that if an algorithm $\textsc{PR}$ is locally fair then it is personalized fair. The proof is similar to that of Theorem~\ref{theorem:local}.
Let $\mathbf{p}_i$ denote the personalized pagerank vector of algorithm $\textsc{PR}_i$. We know that
$
\mathbf{p}_i^T = (1-\gamma)\mathbf{p}_i^T\mathbf{P} + \gamma \mathbf{e}_i^T
$
where $\mathbf{e}_i$ is the vector with 1 at the position $i$, and 0 everywhere else.
The amount of probability that $\textsc{PR}_i$ allocates to the red category can be computed as $\textsc{PR}_i(R) = \mathbf{p}_i^T \mathbf{e}_R$, where $\mathbf{e}_R$ is the vector with 1 at the positions of all red nodes and 0 everywhere else. Multiplying the equation for $\mathbf{p}_i^T$ with $\mathbf{e}_R$ we have:
\[
\textsc{PR}_i(R) = (1-\gamma)\mathbf{p}_i^T\mathbf{P}\mathbf{e}_R + \gamma \mathbf{e}_i^T\mathbf{e}_R
\]
Since $\textsc{PR}$ is fair, for every row of the transition matrix $\mathbf{P}$, the probability of transitioning to a red node is $\phi$. Therefore we have that $\mathbf{P} \mathbf{e}_R = \phi\mathbf{e}$, where $\mathbf{e}$ is the vector of all 1's. Also, since $\mathbf{p}_i$ defines a distribution $\mathbf{p}_i^T\mathbf{e} = 1$. Therefore $(1-\gamma)\mathbf{p}_i^T\mathbf{P}\mathbf{e}_R = \phi(1-\gamma)$. We have:
\[
\textsc{PR}_i(R) = \phi(1-\gamma) + \gamma\mathbf{e}_i^T \mathbf{e}_R
\]
The value of the second term $\gamma\mathbf{e}_i^T \mathbf{e}_R$ depends on whether the node $i$ is red or blue. If $i$ is blue, $\gamma\mathbf{e}_i^T \mathbf{e}_R = 0$, and we have $\textsc{PR}_i(R) = \phi(1-\gamma)$. If $i$ is red, $\gamma\mathbf{e}_i^T \mathbf{e}_R = \gamma$, and thus $\textsc{PR}_i(R) = \phi(1-\gamma) + \gamma$, which proves our claim.
For the opposite direction we make use of the fact that the pagerank vector can be written as $\mathbf{p}^T = \mathbf{v}^T\mathbf{Q}$, where $\mathbf{v}$ is the jump vector and $\mathbf{Q} = \gamma\left[\mathbf{I} - (1-\gamma)\mathbf{P}\right]^{-1}$.
%
%
From Section~\ref{sec:fairness-sensitive} we know that
the $i$-th row of matrix $\mathbf{Q}$ is equal to the personalized pagerank vector $\mathbf{p}_i$. The product $\mathbf{r} = \mathbf{Q}\mathbf{e}_R$ is a vector where $\mathbf{r}[i] = \textsc{PR}_i(R)$. We have assumed that the $\textsc{PR}$ algorithm is universally personalized $\phi$-fair. Therefore $\mathbf{r}[i] = \phi(1-\gamma) + \gamma$ if $i$ is red, and $\mathbf{r}[i] = \phi(1-\gamma)$ if $i$ is blue. That is, $\mathbf{r} = \phi(1-\gamma)\mathbf{e} + \gamma\mathbf{e}_R$. Using the fact that $\mathbf{r} = \mathbf{Q}\mathbf{e}_R$, and that $\mathbf{Q}\mathbf{e} = \mathbf{e}$, since $\mathbf{Q}$ is stochastic, we have the following derivations:
\begin{align}
\mathbf{Q}\mathbf{e}_R & = \phi(1-\gamma)\mathbf{Q}\mathbf{e} + \gamma\mathbf{e}_R \nonumber\\
\mathbf{Q}^{-1}\mathbf{Q}\mathbf{e}_R & = \phi(1-\gamma)\mathbf{Q}^{-1}\mathbf{Q}\mathbf{e} + \gamma\mathbf{Q}^{-1}\mathbf{e}_R \nonumber\\
\mathbf{e}_R & = \phi(1-\gamma)\mathbf{e}
+ \gamma\frac1\gamma \left(\mathbf{I} - (1-\gamma)\mathbf{P} \right)\mathbf{e}_R \nonumber\\
\mathbf{e}_R & = \phi(1-\gamma)\mathbf{e} + \mathbf{e}_R - (1-\gamma) \mathbf{P} \mathbf{e}_R \nonumber\\
\mathbf{P} \mathbf{e}_R & = \phi \mathbf{e} \nonumber
\end{align}
The last equation means that the probability of transitioning from any node in the graph to a red node is $\phi$, which proves our claim.
\end{proof}
The theorem holds also when we consider targeted fairness. We can prove that an algorithm is universally personalized targeted fair, if and only if it is locally fair. We omit the details of the proof due to lack of space.
\section{Experimental Evaluation}
\label{sec:experiments}
Our goal is to evaluate Pagerank fairness in different kinds of networks, identify the conditions under which Pagerank unfairness emerges and evaluate the effect of the proposed fair Pagerank algorithms. Previous research has shown that homophily and size imbalance may lead to degree unfairness \cite{glass-ceiling,glass-ceiling-recommend,xyz}. Is this the case for Pagerank unfairness?
Specifically, we address the following three research questions:
\vspace*{0.03in}
\noindent \textbf{RQ1:} Under which conditions are the original Pagerank and personalized Pagerank algorithms fair?
\noindent \textbf{RQ2:} What is the utility loss incurred by the proposed fair Pagerank algorithms in different networks?
\noindent \textbf{RQ3:} What are the qualitative characteristics of the proposed fair Pagerank algorithms?
\begin{figure}[]
\centering
{\includegraphics[width = 0.4\textwidth]{real_violinplots.pdf}}
\caption{Distribution of the red personalized pagerank of the red and blue nodes for the real datasets, $\phi$-fairness when the red pagerank is $r$ (\sc{books} $r$ = 0.47, \sc{blogs} $r$ = 0.48, \sc{dblp} $r$ = 0.17, and \sc{twitter} $r$ = 0.61).}
\label{fig:real-violin-plots}
\end{figure}
\vspace*{0.03in}
\noindent \textbf{Datasets.}
We use both real and synthetic datasets. Our real datasets are the following:
\begin{itemize}
\item {\sc{books}}: A network of books about US politics where edges between books represented co-purchasing\footnote{\url{http://www-personal.umich.edu/~mejn/netdata/}}.
\item {\sc{blogs}}: A directed network of hyperlinks between weblogs on US politic \cite{blogs-dataset}.
\item {\sc{dblp}}: An author collaboration network constructed from DBLP including a subset of data mining and database conferences.
\item {\sc{twitter}}: A political retweet graph from \cite{nr}.
\end{itemize}
The characteristics of the real datasets are summarized in Table \ref{table:real}.
To infer the gender in {\sc{dblp}}, we used the python \textit{gender guesser} package\footnote{\url{https://pypi.org/project/gender-guesser/}}.
Regarding $homophily$,
we measure for each group, the percentage of the edges that are \textit{cross-edges} that is they point to nodes of the other group divided by the expected
number of such edges. We denote these quantities as $\textit{cross}_R$ and $\textit{cross}_B$.
Values significantly smaller than 1 indicate that the corresponding group exhibits homophily \cite{network-book}.
Synthetic networks are generated using the biased preferential attachment model introduced in \cite{glass-ceiling}. The graph evolves over time as follows.
Let $G_t = (V_t, E_t)$ and $d_t(v)$ denote the graph and the degree of node $v$ at time $t$, respectively.
The process starts with an arbitrary connected graph $G_0$, with $n_0 \, r$ red and $n_0 \,(1 - r)$ blue nodes.
At time step $t + 1$, $t > 0$, a new node $v$ enters the graph. The color of $v$ is red with probability $r$ and blue with probability $1-r$. Node $v$ chooses to connect with an existing node $u$ with probability
$\frac{d_t(u)}{\sum_{w \in G_{t} d_t(w)}}$. If the color of the chosen node $u$ is the same with the color of the new node $v$, then an edge between them is inserted with probability $\alpha$; otherwise an edge is inserted with probability
$1-\alpha$. If no edge is inserted, the
process of selecting a neighbor for node $v$ is repeated until an edge is created.
\begin{figure*}[]
\centering
\subfigure[{symmetric, $r$ = 0.3}]{
{\includegraphics[width = 0.22\textwidth]{sym_varyHom_size_03_utility_plot.pdf}}
}
\subfigure[{asymmetric, $r$ = 0.3}]{
{\includegraphics[width=0.22\textwidth]{asym_varyHom_size_03_utility_plot4.pdf}}
}
\subfigure[{symmetric, $r$ = 0.5}]{
{\includegraphics[width = 0.22\textwidth]{sym_varyHom_size_05_utility_plot.pdf}}
}
\subfigure[{asymmetric, $r$ = 0.5}]{
{\includegraphics[width=0.22\textwidth]{asym_varyHom_size_05_utility_plot1.pdf}}
}
\caption{Utility loss for synthetic networks, $\phi$ = 0.5.}
\label{fig:synth-utility-homo}
\end{figure*}
\begin{figure}[]
\centering
\subfigure[{$r$ = 0.3}]{
{\includegraphics[width = 0.22\textwidth]{sym_varyPhi_size_03_utility_plot.pdf}}
}
\subfigure[{$r$ = 0.5}]{
{\includegraphics[width = 0.22\textwidth]{sym_varyPhi_size_05_utility_plot4.pdf}}
}
\caption{Utility loss for the synthetic datasets, $\alpha$ = 0.5.}
\label{fig:synth-utility-phi}
\end{figure}
\begin{figure*}[]
\centering
\subfigure[{\sc{books}}]{
{\includegraphics[width = 0.22\textwidth]{books_utility_plot.pdf}}
}
\subfigure[{\sc{blogs}}]{
{\includegraphics[width = 0.22\textwidth]{blogs_utility_plot.pdf}}
}
\subfigure[{\sc{dblp}}]{
{\includegraphics[width = 0.22\textwidth]{dblp_course_utility_plot.pdf}}
}
\subfigure[{\sc{twitter}}]{
{\includegraphics[width = 0.22\textwidth]{twitter_utility_plot2.pdf}}
}
\caption{Utility loss for the real networks, (original red pagerank, \sc{books}: 0.48, \sc{blogs}: 0.33, \sc{dblp}: 0.16, and \sc{twitter}: 0.57).}
\label{fig:real-utility}
\end{figure*}
Parameter $r$ controls the group size imbalance.
Parameter $\alpha$ controls the level of homophily: $\alpha < 0.5$ corresponds to heterophily,
$\alpha = 0.5$ to neutrality and $\alpha > 0.5$ to homophily.
We consider: (a) a symmetric case, where $\alpha$ is the same for both groups and (b)
an asymmetric case, where we set $\alpha$ = 0.5 for the blue group, making it neutral, and vary $\alpha$ for the red group.
The datasets and code are available in GitHub\footnote{\url{https://github.com/SotirisTsioutsiouliklis/FairLaR}}.
\subsection{When is Pagerank Fair?}
We study the conditions under which the original Pagerank and personalized Pagerank
algorithms are fair. We assume that the algorithms are fair, if they respect demographic parity, that is, if each group gets pagerank equal to its ratio in the overall population ($\phi$ = $r$). For brevity, we shall call the (personalized) pagerank allocated to red nodes
\textit{red (personalized) pagerank} and the (personalized) pagerank allocated to blue nodes \textit{blue (personalized) pagerank} .
First, we study homophily and size imbalance using synthetic datasets.
In Figure \ref{fig:local-synthetic-all}(a), we report the red pagerank
for the symmetric and in Figure \ref{fig:local-synthetic-all}(b) for the asymmetric case.
Fairness corresponds to the identity line (red pagerank = $r$).
Values above the line indicate unfairness towards the blue group, while values below the line unfairness towards the red group.
We also plot the distribution of the red personalized pagerank
in Figures \ref{fig:sym-violin-plots} and \ref{fig:asym-violin-plots}
for the symmetric and asymmetric case respectively.
To test whether the red personalized pagerank of a node depends on its color,
we plot two distributions, one for the red personalized pagerank of the red nodes and one for the red personalized pagerank of the blue nodes.
Distributions are plotted in the form of violin plots. Personalized pagerank fairness corresponds to the case in which the two distributions overlap, with their mean on value $r$.
Note that when a group is homophilic, it is expected that a large part of its personalized pagerannk goes to nodes of its own color, e.g., red homophilic nodes have large red personalized pageranks.
\noindent \textbf {Symmetric case} (Figures \ref{fig:local-synthetic-all}(a) and \ref{fig:sym-violin-plots}): When both groups are homophilic ($\alpha$ = 0.7, 0.9), the nodes tend to form two clusters, one with red nodes and one with blue nodes
sparsely connected to each other.
This leads to an almost fair pagerank (with a very slight unfairness towards the minority group), but highly unfair personalized pageranks.
On the contrary, when there is
heterophily ($\alpha = 0.1, 0.3$),
there are no clusters, nodes tend to connect with nodes of the opposite color, and the larger group favors the smaller one. In this case, the pagerank and personalized pageranks of both the blue and the red nodes are all unfair towards the majority group. This is especially evident when the imbalance in size is large (small $r$).
\noindent \textbf {Asymmetric case} (Figures \ref{fig:local-synthetic-all}(b) and \ref{fig:asym-violin-plots}):
When the red nodes are homophilic ($\alpha$ = 0.7, 0.9), the red group keeps the pagerank to itself. As a result both pagerank and personalized pageranks are unfair towards the blue nodes, especially for larger $r$.
When the red nodes are heterophilic ($\alpha$ = 0.1, 0.3), the red group favors the blue group, and as a result, both the pagerank and the personalized pagerank are unfair towards the red nodes, especially for larger $r$.
Thus, independently of the size $r$,
pagerank is unfair to the blue (the neutral) group in case of homophily, and unfair to the red group in the case of heterophily.
\noindent \textbf {Universal fairness:} The only case when both pagerank and personalized pageranks are all fair is in a neutral network ($\alpha = 0.5$) with same-size groups ($r$ = 0.5) (middle violin plots in Figures
\ref{fig:sym-violin-plots}(c) and \ref{fig:asym-violin-plots}(c)).
\noindent \textbf {Real datasets:}
For the real datasets, we report the red pagerank in Table \ref{table:real} ($p_R$ value) and plot the distributions of the red personalized pagerank of the blue and red nodes in Figure \ref{fig:real-violin-plots}.
For {\sc books}, there is no size imbalance and there is strong symmetric homophily leading to fair pagerank and highly unfair personalized pageranks.
For {\sc blogs}, there is no size imbalance, but the blue group is slightly more homophilic, which leads both to unfairness in pagerank for the red group, and unfairness of the personalized pagerank of the blue nodes towards the red ones.
For {\sc dblp}, we have large size imbalance with the red group being the minority but almost neutral behavior in terms of homophily, leading to an almost fair pagepank, and the majority of personalized pageranks being fair.
Finally, for {\sc twitter}, the red group is larger than the blue group but less homophilic which leads to a slight unfairness towards the red group for both pagerank and personalized pageranks.
\subsection{What is the Utility Loss for Fairness?}
We now look into the utility loss
for achieving fairness.
We can view utility loss for each network as a measure of the cost we have to pay to achieve $\phi$-fairness for this network.
First,
to assess the utility loss of our algorithms in absolute terms
we compute a lower bound for the utility loss.
\noindent \textbf{Lower Bound.}
We compute a lower bound on the utility loss, by constructing the probability vector $\mathbf{w}$ that is $\phi$-fair, and it has the minimum utility loss compared to the original pagerank vector $\mathbf{p}_O$.
Note that vector $\mathbf{w}$ is not necessarily attainable by any Pagerank algorithm in $\mathcal{PR}$.
To compute $\mathbf{w}$, we start with $\mathbf{p}_O$ and we redistribute the probability between the two groups to make it fair. Let $\mathbf{p}_O(R)$ be the probability assigned to the red group by $\mathbf{p}$. Without loss of generality, assume that $\mathbf{p}_O(R)< \phi$, and let $\Delta = \phi-\mathbf{p}_O(R)$. To make the vector fair, we need to remove $\Delta$ probability mass from the nodes in $B$, and redistribute it to the nodes in $R$. It is easy to show that to minimize the loss, the optimal redistribution will remove uniformly $\Delta/|B|$ probability from all nodes in $B$, and add uniformly $\Delta/|R|$ to all nodes in $R$.
This follows from the fact that among all distribution vectors the one with the smallest length is the uniform one. However, this process does not guarantee that the resulting vector will not have negative entries, since some blue nodes may have probability less than $\Delta/|B|$. Let $\beta$ be the smallest non-zero such probability of any blue node. Our algorithm transfers $\beta$ probability from all the non-zero blue nodes to the red nodes, and then recursively applies the same procedure for the residual amount of probability that has not been transferred. It is not hard to show that this process will produce a fair vector with the minimum utility loss with respect to $\mathbf{p}_O$.
Figures \ref{fig:synth-utility-homo} and \ref{fig:synth-utility-phi} report the utility loss for selected synthetic networks
and Figure \ref{fig:real-utility} for different values of
$\phi$ for the real datasets. $LB$ is the
lower bound on utility loss.
\noindent \textbf{Effect of $\phi$:} In all cases, loss increases as the requested $\phi$ deviates from the red pagerank originally assigned to the red group
(Figures \ref{fig:synth-utility-phi} and \ref{fig:real-utility}).
\noindent \textbf{\textsl{FSPR}:} \, In some cases {{\sc FSPR}} incurs high utility loss and, occasionally, it is even unable to find an appropriate solution.
{{\sc FSPR}} achieves fairness by changing the jump vector of the Pagerank algorithm.
The overall pagerank vector is a linear combination of the personalized pagerank vectors, with the jump vector providing the coefficients of this linear combination.
{{\sc FSPR}} increases the jump probability for the vectors that favor the group it wants to promote and takes away probability from the vectors that favor the other group. However, when there are few, or no appropriate such vectors, {{\sc FSPR}} is forced to make extreme choices (assign a lot of probability to a few vectors) thus incurring high utility loss, or it is unable to find any solution.
There are several such examples in our datasets.
For the {\sc dblp} dataset (Figure \ref{fig:real-utility}(c)), for small values of $\phi$ ($\phi \leq 0.3$), the utility loss of {{\sc FSPR}} is close to the optimal, but for $\phi \geq 0.4$, it skyrockets. Looking at Figure \ref{fig:real-violin-plots}, we see that there are very few personalized pagerank vectors with red pagerank larger than 0.4. As a result, for $\phi \geq 0.4$, {{\sc FSPR}} is forced to allocate all the probability of the jump vector to these nodes, leading to high utility loss.
It is also interesting to observe the effect of homophily, or lack of, on the utility loss of {{\sc FSPR}}. In Figure \ref{fig:synth-utility-homo}(a) and (b), utility loss peaks when the network is neutral ($\alpha = 0.5$). In this case, there is no personalized pagerank vector that strongly favors one group over the other.
\noindent \textbf{\textsl{LFPR}:} \, Overall, for the locally fair family of algorithms, the utility loss function is smoother, avoiding high peaks.
The locally fair algorithms are directly affected by homophily, since this determines the composition of the neighborhoods of the nodes. As we deviate from neutrality, the loss increases (Figure \ref{fig:synth-utility-homo}).
This holds especially for the {{\sc LFPR$_N$}} algorithm.
This can be seen very clearly in {\sc books} (Figure \ref{fig:real-utility}(a)), where {{\sc FSPR}} almost achieves the lower bound, while {{\sc LFPR$_N$}} incurs high utility loss because of {\sc books} being very homophilic.
The utility loss of {{\sc LFPR$_U$}} and {{\sc LFPR$_P$}} follows in general the same trend as {{\sc LFPR$_N$}}.
Finally, {{\sc LFPR$_O$}} redistributes any residual Pagerank so that the
utility loss is optimized and in many cases its utility loss is close to the lower bound ($LB$).
\noindent \textbf{Summary:}
{{\sc FSPR}} works well when there are enough unfair nodes (i.e., nodes with unfair personalized pageranks), as it can distribute the jump probability to them to achieve fairness.
On the contrary, locally fair algorithms have high utility loss when there are many unfair nodes. {{\sc LFPR$_N$}} is the most sensitive to homophily.
The utility loss of {{\sc LFPR$_N$}} can be seen as a measure of local unfairness.
Overall, the locally fair algorithms are more stable than {{\sc FSPR}}. {{\sc LFPR$_O$}} works very well in terms of utility loss and in many cases it achieves utility close to the lower bound.
\subsection{Qualitative Evaluation}
In this section, we provide some qualitative experiments, to better understand the properties of the fair algorithms.
\begin{figure}[t]
\centering
\subfigure[Original {{Pagerank}}]{
{\includegraphics[width = 0.175\textwidth]{books_05_pagerank.png}}
}
\subfigure[{{\sc LFPR$_N$}}]{
{\includegraphics[width = 0.175\textwidth]{books_05_lfprNei.png}}
}
\subfigure[{{\sc FSPR}}]{
{\includegraphics[width = 0.175\textwidth]{books_05_sensitive.png}}
}
\subfigure[Jump vector for {{\sc FSPR}}]{
{\includegraphics[width = 0.175\textwidth]{books_05_jumpV.png}}
}
\caption{Visualization of the {\sc book} dataset.}
\label{fig:vis-books}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[Original {{Pagerank}}]{
{\includegraphics[width = 0.22\textwidth]{dblp-course-05-pagerank.pdf}}
}
\subfigure[{{\sc LFPR$_N$}}]{
{\includegraphics[width = 0.22\textwidth]{dblp-course-05-lfprNei.pdf}}
}
\subfigure[{{\sc FSPR}}]{
{\includegraphics[width = 0.22\textwidth]{dblp-course-05-sensitive.pdf}}
}
\subfigure[Jump vector for {{\sc FSPR}}]{
{\includegraphics[width = 0.22\textwidth]{dblp-course-05-jumpV.pdf}}
}
\caption{Visualization of the {\sc dblp} dataset.}
\label{fig:vis-dblp}
\end{figure}
\noindent \textbf{Visualization.}
In Figures \ref{fig:vis-books} and \ref{fig:vis-dblp}, we visualize the results of the algorithms for the {\sc books} and the {\sc dblp} dataset respectively, for $\phi$ = 0.5. Red nodes are colored red, and blue nodes are colored blue. Their size depends on the value of the quantity we visualize. We visualize the pagerank values for the original Pagerank, {{\sc FSPR}} and {{\sc LFPR$_N$}} algorithms, and the jump vector probabilities for {{\sc FSPR}}.
For {\sc books}, where the original red pagerank is close to $\phi$,
{{\sc FSPR}} is very similar to the original Pagerank algorithm.
{\sc books} is homophilic and the jump vector assigns rather uniform weights
to almost all nodes.
On the other hand, {{\sc LFPR$_N$}} promotes heavily nodes connecting the two opposite groups, i.e., nodes that are minorities in their neighborhoods.
We observe a different result in {\sc dblp}, where $\phi$ is much larger than
the original red pagerank.
{{\sc LFPR$_N$}} distributes weights broadly in the red community, while {{\sc FSPR}} is much more aggressive. This is especially evident in the jump vector which promotes a few nodes in the periphery.
\begin{table}[ht]
\centering
\caption{Top-10 authors with $\phi$ = 0.3; the number in parenthesis is the position of the author in the original {{Pagerank}} ($\textsc{OPR}$) (female authors in bold).
}
\footnotesize{
\begin{tabular}{l l l}
\toprule
$\textsc{OPR}$ & {\sc FSPR} & {\sc LFPR$_N$} \\
\midrule
C. Faloutsos & C. Faloutsos (1) & C. Faloutsos (1) \\
G. Weikum & G. Weikum (2) & G. Weikum (2) \\
P. S. Yu & P. S. Yu (3) & \textbf{J. Widom} (38) \\
M. Stonebraker & M. Stonebraker (4) & M. Stonebraker (4) \\
M. J. Franklin & M. J. Franklin (5) & P. S. Yu (3)\\
H. Garcia-Molina & H. Garcia-Molina (6) & \textbf{S. T. Dumais} (28) \\
D. Kossmann & D. Kossmann (7) & \textbf{M. Lalmas} (27) \\
W. Lehner & \textbf{E. A. Rundensteiner} (22) & P. Serdyukov (17) \\
M. J. Carey & R. Agrawal (11) & \textbf{E. Bertino} (25) \\
M. de Rijke & W. Lehner (8) & \textbf{E. A. Rundensteiner} (22) \\
\bottomrule
\end{tabular}
\label{table:case-study}
}
\end{table}
\noindent
\textbf{Anecdotal Examples.}
We will use the {\sc dblp} dataset for a qualitative evaluation of the results of the algorithms. Recall that for this dataset, women are the minority with $r$ = 0.17, and the original red pagerank is equal to 0.16.
To achieve a fairer representation of women, we apply the {{\sc LFPR$_N$}} and {{\sc FSPR}} algorithms with $\phi$ = 0.3. In Table \ref{table:case-study}, we present the first 10 authors for each algorithm.
In the original Pagerank algorithm $\textsc{OPR}$, there is no female author in the top-10 (the first one appears in position 22). {{\sc FSPR}} achieves fairness but also minimizes utility loss, so the result is fair but also close to that of $\textsc{OPR}$. {{\sc LFPR$_N$}} asks that all authors
have $\phi$-fair inter-gender collaborations resulting
in a large number of female authors appearing in the top-10 positions.
\begin{table}[ht]
\centering
\caption{Top-3 female authors by conference, $\phi$ = 0.3.}
\footnotesize {
\begin{tabular}{lll}
\toprule
&SIGIR&SIGMOD\\
\midrule
\multirow{3}{*}{{\sc FSPR}}&
Mounia Lalmas & Elke A. Rundensteiner \\
&Susan T. Dumais & Elisa Bertino \\
&Juliana Freire & Tova Milo \\
\midrule
\multirow{3}{*}{{\sc LFPR$_N$}}&
Susan T. Dumais & Jennifer Widom \\
& Mounia Lalmas & Elke A. Rundensteiner \\
& Emine Yilmaz & Fatma Ozcan \\
\bottomrule
\end{tabular}
}
\label{table:case-study-conf}
\end{table}
We also use {\sc dblp} to study the targeted fair Pagerank algorithms.
In this case, we want to enforce fair behavior towards authors in specific conferences. We consider two such conferences, SIGIR and SIGMOD, and select $S$ to include authors of each one of them. In Table \ref{table:case-study-conf}, we show the top-3 women authors for each conference according to our algorithms. We observe that the algorithms produce different results depending on the conference, each time promoting women that are authorities in their respective fields, such as, Suzan Dumais when $S$ is SIGIR, and Jennifer Widom when $S$ is SIGMOD.
\section{Related Work}
\label{sec:related-work}
\noindent \textbf{Algorithmic fairness.} Recently, there has been increasing interest in algorithmic fairness, especially in the context of machine learning.
Fairness is regarded as the lack of discrimination on the basis of some protective attribute.
Various definition of fairness having proposed especially for classification \cite{fairness-awarness,fairness-study,fairness-measures,bias-fairness-survey}. We use a group-fairness definition, based on parity.
Approaches to handing fairness can be classified as \textit{pre-processing}, that modify the input data, \textit{in-processing}, that modify the algorithm and \textit{post-processing} ones, that modify the output. We are mostly interested in in-processing techniques.
There is also prior work on fairness in ranking~\cite{fai*r,Julia17,Biega18,pair-wise}.
All of these works consider ranking as an ordered list of items, and use different rules for defining and enforcing fairness that consider different prefixes of the ranking~\cite{fai*r,Julia17}, pair-wise orderings~\cite{pair-wise}, or exposure and presentation bias~\cite{exposure-ranking,Biega18}.
Our goal in this paper is not to propose a new definition of ranking fairness, but rather to initiate a study of fairness in link analysis.
A distinguishing aspect of our approach is that we
take into account the actual Pagerank weights of the nodes, not just their ranking.
Furthermore, our focus in this paper, is to design in-processing algorithms that incorporate fairness in the inner working of the Pagerank algorithm. We present a post-processing approach as a means to estimate a lower bound on the utility loss. None of the previous approaches considers ranking in networks, so the proposed approaches are novel.
\noindent \textbf{Fairness in networks.}
There has been some recent work on network fairness in the context of graph embeddings \cite{fair-embedding,filter-bubbles,fair-walk}.
The work in \cite{fair-embedding} follows an in-processing approach
that extends the learning function with regulatory fairness enforcing terms, while the work in \cite{filter-bubbles} follows a post-processing approach so as to promote link recommendations to nodes belonging
to specific groups. Both works are not related to our approach.
The work in \cite{fair-walk} extends the node2vec graph embedding method by modifying the
random walks used in node2vec with fair walks, where nodes are partitioned into groups and each group is given the same probability of being selected when a node makes a transition. The random walk
introduced in \cite{fair-walk} has some similarity with the
random walk interpretation of {{\sc LFPR$_N$}}. It would
be interesting to see, whether our extended residual-based algorithms could be utilized also in the context of graph embeddings, besides its use in link analysis.
There are also previous studies on the effect of
homophily, preferential attachment and imbalances in group sizes.
It was shown that the combination of these three factors leads to
uneven degree distributions between groups \cite{glass-ceiling}.
Recent work shows that this phenomenon is exaggerated by many link recommendation algorithms \cite{glass-ceiling-recommend}. Evidence of inequality between degree distribution of minorities and majorities was also found in many real networks \cite{homophily-ranking}.
Our work extends this line of research by looking at Pagerank values instead of degrees.
Along this lines, recent work studies in depth how homophily and size imbalance can affect the visibility
that different groups get in link recommendations, i.e, how often nodes in each group get recommended \cite{xyz}.
Very recent work also looks at graph mining algorithms in general from the perspective of individual fairness, where the goal is to produce a similar output for similar nodes \cite{inform}.
Finally, there is previous work on diversity in network ranking.
The goal is to find important nodes that also maximally cover the nodes in the network~\cite{grasshoper,divrank}.
Our problem is fundamentally different, since we look for scoring functions that follow a parity constraint.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we initiate a study of fairness for Pagerank.
We provide definitions of fairness, and we propose two approaches for achieving fairness: one that modifies the jump vector, and one that imposes a fair behavior per node. We prove that the latter is equivalent to a stronger notion of fairness that also guarantees personalized Pagerank fairness. We also consider the problem of attaining fairness while minimizing the utility loss of Pagerank. Our experiments demonstrate the behavior of our different algorithms.
There are many direction for future work. First, we would like to study the role of
$\gamma$, i.e., the jump probability. Then, it would be interesting to explore other notions of Pagerank fairness, besides $\phi$-fairness, for instance ones based on
rank-aware fairness \cite{edbt-tut}. Furthermore, we plan to explore further applications of the theory behind our fair Pagerank algorithms to derive novel notions of fair random walks. | 273298219efd07f335ff99867a7f4bcabbb67a3f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$ and let $1 < p < \infty$.
We consider the existence of positive solutions to quasilinear elliptic equations of the type
\begin{equation}\label{eqn:p-laplace}
\begin{cases}
\displaystyle
- \Delta_{p, w} u = \sigma u^{q}
&
\text{in $\Omega$},
\\
u = 0
&
\text{on $\partial \Omega$,}
\end{cases}
\end{equation}
in the sub-natural growth case $0 < q < p - 1$,
where
$\Delta_{p, w}$ is a weighted $(p, w)$-Laplacian,
$w$ is a $p$-admissible weight on $\mathbb{R}^{n}$ (see Section \ref{sec:preliminaries} below)
and $\sigma$ is a nonnegative (locally finite) Radon measure on $\Omega$.
For the standard theory of sublinear equations,
we refer to \cite{MR0181881, MR679140, MR820658, MR1141779} and the references therein.
In the classical existence results of weak solutions, the boundedness of coefficients was assumed.
Boccardo and Orsina \cite{MR1272564} removed this assumption
and pointed out that if the integrability of the coefficients is low,
the solutions do not necessarily belong to the class of weak solutions.
Therefore, we interpret this equation in the sense of $p$-superharmonic functions, or locally renormalized solutions.
For details on such generalized solutions, especially on the relation between the two concepts, see \cite{MR1955596, MR2859927, MR3676369} and references therein.
Hereafter, we use the framework for $p$-superharmonic functions.
The measure-valued coefficient equation \eqref{eqn:p-laplace} is relevant to the following $L^{p}$-$L^{1 + q}$ trace inequality:
\begin{equation}\label{eqn:trace_ineq}
\| f \|_{L^{1 + q}(\Omega; \sigma)} \le C_{T} \| \nabla f \|_{L^{p}(\Omega; w)}, \quad \forall f \in C_{c}^{\infty}(\Omega).
\end{equation}
Maz'ya and Netrusov \cite{MR1313906} gave a capacitary condition that characterizes \eqref{eqn:trace_ineq}.
Cascante, Ortega and Verbitsky \cite{MR1734322, MR1747901} studied non-capacitary characterizations for inequalities of the type \eqref{eqn:trace_ineq}.
For example, if $\Omega = \mathbb{R}^{n}$ and $w = 1$, then the best constant $C_{T}$ in \eqref{eqn:trace_ineq} satisfies
\[
\frac{1}{c} \, C_{T}^{ \frac{(1 + q)p}{p - 1 - q} }
\le
\int_{\mathbb{R}^{n}}
\left( {\bf{W}}_{1, p} \sigma \right)^{\frac{(1 + q)(p - 1)}{p - 1 - q}}
\, d \sigma
\le
c \, C_{T}^{ \frac{(1 + q)p}{p - 1 - q} },
\]
where $c = c(n, p, q)$ and ${\bf{W}}_{1, p} \sigma$ is the \textit{Wolff potential} of $\sigma$ (see \cite{MR0409858, MR727526}).
Recently, Verbitsky and his colleagues studied the problem of
existence of solutions to elliptic equations related to \eqref{eqn:p-laplace}
and presented some criteria
(see \cite{MR3567503, MR3311903, MR3556326, MR3938014, MR4105916, MR3642745, MR3724493, MR3985926, MR3881877, MR3792109, VERBITSKY2019111516}).
In their study, they treated the cases of $\Omega = \mathbb{R}^{n}$, or $p = 2$.
For $\Omega = \mathbb{R}^{n}$, Wolff potentials are suitable potentials for the problem.
Every $p$-superharmonic function known to be locally estimated by Wolff potentials (see \cite{MR1205885,MR1264000}),
and these estimates are some of the key pieces of proof.
In contrast, we can directly use Green potentials for $p = 2$ or, more generally, for linear equations.
However, Wolff potentials are not sufficient for estimating the boundary behavior of solutions.
Furthermore, Radon measures satisfying \eqref{eqn:trace_ineq} may not have compact support in $\Omega$ and are not even finite in general (see Section \ref{sec:example2}).
Consequently, the complete criteria for the existence of solutions to quasilinear equations in bounded domains have not yet been obtained.
The purpose of this paper is to extend Verbitsky's theory to (weighted) quasilinear equations in bounded domains.
Our basic idea is to replace Wolff or Green potentials with
the minimal positive $p$-superharmonic solution $u$ to $- \Delta_{p, w} u = \sigma$ (see Definition \ref{def:potential} for the precise meaning).
The function $\mathcal{W}_{p, w} \sigma = u$ has no explicit integral representation; however, some required estimates can be obtained by directly using the properties of weak solutions.
To realize this approach, we investigate various properties of $p$-superharmonic functions, especially the solvability of Dirichlet problems in the case of infinite measure data.
Note that this existence problem has been stated as an open problem in \cite[Problem 2]{MR1990293}.
Let $\mathcal{M}^{+}_{0}(\Omega)$ be the set of all nonnegative Radon measures on $\Omega$
that are absolutely continuous with respect to the $(p, w)$-capacity.
Note that $\sigma$ must belong to $\mathcal{M}^{+}_{0}(\Omega)$ if \eqref{eqn:trace_ineq} holds.
Our main result is as follows.
\begin{theorem}\label{thm:main_theorem}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
Let $1 < p < \infty$ and $0 < q < p - 1$.
Suppose that $\sigma \in \mathcal{M}^{+}_{0}(\Omega) \setminus \{ 0 \}$.
Fix $0 < \gamma < \infty$.
Then the following statements are equivalent:
\begin{enumerate}[label=(\arabic*)]
\item\label{enum:01@main_theorem}
There exists a nontrivial nonnegative $(p, w)$-superharmonic supersolution $v$ to $- \Delta_{p, w} v = \sigma v^{q}$ in $\Omega$
satisfying $\| v \|_{L^{\gamma + q}(\sigma)} \le C_{1} < \infty$.
\item\label{enum:02@main_theorem}
The measure $\sigma$ satisfies
\begin{equation}\label{eqn:energy_cond}
\left(
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} } \, d \sigma
\right)^{\frac{1}{\gamma + q}}
\le C_{2} < \infty.
\end{equation}
\item\label{enum:03@main_theorem}
The following weighted norm inequality holds:
\begin{equation}\label{eqn:weighted_norm_ineq}
\| \mathcal{W}_{p, w}( |f| \sigma ) \|_{L^{\gamma + q}( \sigma )}
\le
C_{3} \| f \|_{L^{ \frac{\gamma + q}{q} }( \sigma )}^{\frac{1}{p - 1}},
\quad
\forall f \in L^{ \frac{\gamma + q}{q} }( \sigma ).
\end{equation}
\end{enumerate}
Moreover, if $C_{i}$ ($i = 1, 2 ,3$) are the best constants in the above statements, then
\[
C_{1}
\le
C_{3}^{\frac{p - 1}{p - 1 - q}}
\le
c_{E}^{\frac{1}{p - 1 - q}} \, C_{2}
\le
\frac{ c_{E}^{\frac{1}{p - 1 - q}} }{ c_{V} } \, C_{1},
\]
where
\begin{equation}\label{eqn:opt-const}
c_{E}
:=
\left( \frac{p - 1 + \gamma}{p} \right)^{p} \frac{1}{\gamma},
\quad
c_{V}
:=
\left( \frac{p - 1 - q}{p - 1} \right)^{ \frac{p - 1}{p - 1 - q} }.
\end{equation}
In addition, if one of the above statements holds,
then there exists a minimal positive $(p, w)$-superharmonic solution $u$ to $- \Delta_{p, w} u = \sigma u^{q}$ in $\Omega$
such that $\| u \|_{L^{\gamma + q}(\sigma)} \le C_{1}$.
\end{theorem}
The boundary behavior of $u$ is not discussed in Theorem \ref{thm:main_theorem}; however, under appropriate assumptions on $\sigma$,
we can prove that $u$ vanishes on $\partial \Omega$ (see Proposition \ref{prop:energy_class} and Corollary \ref{cor:energy_esti}).
One such condition is as follows.
\begin{theorem}\label{thm:main_theorem_FE}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
Let $1 < p < \infty$ and $0 < q < p - 1$.
Suppose that $\sigma \in \mathcal{M}^{+}_{0}(\Omega) \setminus \{ 0 \}$.
Then there exists a unique positive weak solution $u \in H_{0}^{1, p}(\Omega; w)$ to \eqref{eqn:p-laplace}
if and only if
\begin{equation}\label{eqn:energy_cond@FE}
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma
<
\infty.
\end{equation}
Moreover,
\[
c_{V}^{1 + q}
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma
\le
\| \nabla u \|_{L^{p}(w)}^{p}
\le
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma.
\]
\end{theorem}
We also give the Cascante-Ortega-Verbitsky type of theorem below.
\begin{theorem}\label{thm:embedding}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
Let $1 < p < \infty$ and let $-1 < q < p - 1$.
Suppose that $C_{T}$ is the best constant of \eqref{eqn:trace_ineq}.
Then
\[
C_{T}^{ \frac{(1 + q)p}{p - 1 - q} }
\le
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma
\le
\left( \frac{1}{1 + q} \right)^{ \frac{1 + q}{p - 1 - q} } \frac{1}{c_{V}^{1 + q}} \,
C_{T}^{ \frac{(1 + q)p}{p - 1 - q} }.
\]
In particular, if \eqref{eqn:trace_ineq} holds with some $-1 < q < p - 1$,
then the equation $- \Delta_{p, w} u = \sigma$ in $\Omega$ has
a minimal positive $(p, w)$-superharmonic solution.
\end{theorem}
\begin{remark}
The constants in Theorems \ref{thm:main_theorem}-\ref{thm:embedding}
do not depend on $n$ or the data of $w$.
Due to the qualitative arguments in the proof, $w$ must be $p$-admissible; however, its quantitative properties, especially the Sobolev-type inequalities are not used.
\end{remark}
In particular, if \eqref{eqn:trace_ineq} holds with $q > 0$, then there exists a unique positive weak solution $u$ to \eqref{eqn:p-laplace} such that
\[
\frac{1}{c(p, q)} C_{T}^{ \frac{1 + q}{p - 1 - q} } \le\| \nabla u \|_{L^{p}(w)} \le c(p, q) C_{T}^{ \frac{1 + q}{p - 1 - q} }.
\]
Examples of concrete sufficient conditions for \eqref{eqn:energy_cond}
will be discussed at the end of the paper.
One is a Lorentz scale refinement of \cite[Theorem 5.5]{MR1272564},
and the other is quasilinear ordinary differential equations with nonintegrable Hardy-type coefficients.
\subsection*{Organization of the paper}
Section \ref{sec:preliminaries} presents various facts in nonlinear potential theory and introduces classes of smooth measures.
Section \ref{sec:potentials} discusses minimal $p$-superharmonic solutions to Dirichlet problems.
Section \ref{sec:generalized_energy} defines the generalized energy of $p$-superharmonic functions
and investigates its properties including Theorem \ref{thm:embedding}.
Section \ref{sec:proof_of_main_theorem} provides the proof of Theorem \ref{thm:main_theorem}
by using results in Sections \ref{sec:potentials} and \ref{sec:generalized_energy}.
Sections \ref{sec:example1} and \ref{sec:example2} discuss two applications of Theorems \ref{thm:main_theorem}-\ref{thm:embedding}.
\subsection*{Acknowledgments}
The author would like to thank Professor Verbitsky for suggesting references to added to an earlier version of the manuscript.
This work was supported by JSPS KAKENHI Grant Number JP18J00965.
\subsection*{Notation}
We use the following notation.
Let $\Omega$ be a domain (connected open subset) in $\mathbb{R}^{n}$.
\begin{itemize}
\item
$\mathbf{1}_{E}(x) :=$ the indicator function of a set $E$.
\item
$C_{c}^{\infty}(\Omega) :=$
the set of all infinitely differentiable functions with compact support in $\Omega$.
\item
$\mathcal{M}^{+}(\Omega) :=$ the set of all nonnegative Radon measures on $\Omega$.
\item
$L^{p}(\mu) :=$ the $L^{p}$ space with respect to $\mu \in \mathcal{M}^{+}(\Omega)$.
\end{itemize}
For a ball $B = B(x, R)$ and $\lambda > 0$, $\lambda B := B(x, \lambda R)$.
For measures $\mu$ and $\nu$, we denote $\nu \le \mu$ if $\mu - \nu$ is a nonnegative measure.
For a sequence of extended real valued functions $\{ f_{j} \}_{j = 1}^{\infty}$,
we denote $f_{j} \uparrow f$
if $f_{j + 1} \ge f_{j}$ for all $j \ge 1$ and $\lim_{j \to \infty} f_{j} = f$.
Moreover, $c$ and $C$ denote various constants with and without indices.
\section{Preliminaries}\label{sec:preliminaries}
We first recall the basic properties of $p$-admissible weights from \cite[Chapter A.2]{MR2867756}, \cite[Chapter 20]{MR2305115} and references therein.
Throughout the paper, $1 < p < \infty$ is a fixed constant.
A Lebesgue measurable function $w$ on $\mathbb{R}^{n}$
to be said as \textit{weight} on $\mathbb{R}^{n}$ if $w \in L^{1}_{\mathrm{loc}}(\mathbb{R}^{n}; dx)$ and $w(x) > 0$ $dx$-a.e.
We write $w(E) = \int_{E} w \, dx$ for a Lebesgue measurable set $E \subset \mathbb{R}^{n}$.
We always assume that $w$ is \textit{$p$-admissible},
that is, positive constants $C_{D}$, $C_{P}$ and $\lambda \ge 1$ exist, such that
\[
w(2B) \le C_{D} w(B)
\]
and
\[
\fint_{B} |f - f_{B}| \, dw \le C_{P} \, \mathrm{diam}(B) \left( \fint_{\lambda B} |\nabla f|^{p} \, dw \right)^{\frac{1}{p}},
\quad
\forall f \in C_{c}^{\infty}(\mathbb{R}^{n}),
\]
where $B$ is an arbitrary ball in $\mathbb{R}^{n}$,
$\fint_{B} = w(B)^{-1} \int_{B}$
and
$f_{B} = \fint_{B} f \, dw$.
One of the important properties of $p$-admissible weights is the Sobolev inequality (\cite{MR1098839, MR1150597}).
In particular, the following form of the Poincar\'{e} inequality holds:
\begin{equation*}\label{eqn:poincare}
\int_{B} |f|^{p} \, dw
\le
C \, \mathrm{diam}(B)^{p} \int_{B} |\nabla f|^{p} \, dw,
\quad
\forall f \in C_{c}^{\infty}(B),
\end{equation*}
where $C$ is a constant depending only on $p$, $C_{D}$, $C_{P}$ and $\lambda$.
Next, we recall basics of nonlinear potential theory from \cite[Chapters 1-10 and 21]{MR2305115}.
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
The weighted Sobolev space $H^{1, p}(\Omega; w)$ is the closure of $C^{\infty}(\Omega)$
with respect to the norm
\[
\| u \|_{H^{1, p}(\Omega; w)}
:=
\left(
\int_{\Omega} |u|^{p} + |\nabla u|^{p} \, d w
\right)^{\frac{1}{p}},
\]
where $\nabla u$ is the gradient of $u$.
The corresponding local space $H^{1, p}_{\mathrm{loc}}(\Omega; w)$ is defined in the usual manner.
We denote the closure of $C_{c}^{\infty}(\Omega)$ in $H^{1, p}(\Omega; w)$ by $H_{0}^{1, p}(\Omega; w)$.
Since $\Omega$ is bounded, we can take $\| \nabla \cdot \|_{L^{p}(\Omega; w)}$
as the norm of $H_{0}^{1, p}(\Omega; w)$ by the Poincar\'{e} inequality.
For $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$,
we define the weighted $p$-Laplace operator $\Delta_{p, w}$ by
\[
\langle - \Delta_{p, w} u, \varphi \rangle
=
\int_{\Omega} |\nabla u|^{p - 2} \nabla u \cdot \nabla \varphi \, dw,
\quad
\forall \varphi \in C_{c}^{\infty}(\Omega).
\]
A function $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$ is called a \textit{supersolution} to
\begin{equation}\label{eqn:p-harmonic}
- \Delta_{p, w} u = 0 \quad \text{in} \ \Omega
\end{equation}
if $\langle - \Delta_{p, w} u, \varphi \rangle \geq 0$ for all nonnegative $\varphi \in C_{c}^{\infty}(\Omega)$.
If $\mu$ is an element of the dual of $H_{0}^{1, p}(\Omega; w)$,
then the Dirichlet problem
\begin{equation}\label{eqn:variational_problem}
\begin{cases}
\langle - \Delta_{p, w} u, \varphi \rangle = \langle \mu, \varphi \rangle
\quad
\forall \varphi \in H_{0}^{1, p}(\Omega; w)
\\
u \in H_{0}^{1, p}(\Omega; w)
\end{cases}
\end{equation}
has a unique weak solution $u$.
Let $\Omega \subset \mathbb{R}^{n}$ be open, and let $K \subset \Omega$ be compact.
The \textit{variational $(p, w)$-capacity} $\mathrm{cap}_{p, w}(K, \Omega)$
of the condenser $(K, \Omega)$ is defined by
\[
\mathrm{cap}_{p, w}(K, \Omega)
:=
\inf \left\{
\| \nabla u \|_{L^{p}(\Omega; w)}^{p} \colon u \geq 1 \ \text{on} \ K, \ u \in C_{c}^{\infty}(\Omega)
\right\}.
\]
Moreover, for $E \subset \Omega$, we define
\[
\mathrm{cap}_{p, w}(E, \Omega)
:=
\inf_{\substack{ E \subset U \subset \Omega \\ U \colon \text{open} }}
\sup_{K \subset U \colon \text{compact}}
\mathrm{cap}_{p , w}(K, \Omega).
\]
Since $\Omega \subset \mathbb{R}^{n}$ is bounded,
$\mathrm{cap}_{p, w}(E, \Omega) = 0$ if and only if $C_{p, w}(E) = 0$,
where $C_{p, w}(\cdot)$ is the \textit{(Sobolev) capacity} of $E$.
We say that a property holds \textit{quasieverywhere} (q.e.)
if it holds except on a set of $(p, w)$-capacity zero.
An extended real valued function $u$ on $\Omega$ is called as \textit{quasicontinuous}
if for every $\epsilon > 0$ there exists an open set $G$ such that
$C_{p, w}(G) < \epsilon$ and $u|_{\Omega \setminus G}$ is continuous.
Every $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$ has a quasicontinuous representative $\tilde{u}$
such that $u = \tilde{u}$ a.e.
A function $u \colon \Omega \to ( - \infty, \infty]$ is called \textit{$(p, w)$-superharmonic} if
$u$ is lower semicontinuous in $\Omega$, is not identically infinite,
and satisfies the comparison principle on each subdomain $D \Subset \Omega$;
if $h \in C(\overline{D})$ is a continuous weak solution to $- \Delta_{p, w} u = 0$ in $D$,
and if $u \geq h$ on $\partial D$, then $u \geq h$ in $D$.
If $u$ is a bounded $(p, w)$-superharmonic function,
then $u$ belongs to $H^{1, p}_{\mathrm{loc}}(\Omega; w)$ and is a supersolution to \eqref{eqn:p-harmonic}.
Conversely, if $u$ is a supersolution to \eqref{eqn:p-harmonic},
then its \textit{lsc-regularization}
\[
u^{*}(x)
=
\lim_{r \to 0} \essinf_{B(x, r)} u
\]
is $(p, w)$-superharmonic in $\Omega$.
If $u$ and $v$ are $(p, w)$-superharmonic in $\Omega$ and $u(x) \leq v(x)$ for a.e. $x \in \Omega$,
then $u(x) \leq v(x)$ for all $x \in \Omega$.
Every $(p, w)$-superharmonic function is known to be quasicontinuous.
In particular, the set $\{ u = \infty \}$ has zero $(p, w)$-capacity whenever $u$ is $(p, w)$-superharmonic.
Assume that $u$ is a $(p, w)$-superharmonic function in $\Omega$.
Its truncation $u_{k} = \min\{ u, k \}$ will then become a supersolution to \eqref{eqn:p-harmonic} for all $k > 0$.
Moreover, there exists a Radon measure $\mu[u]$ such that
\[
\lim_{k \to \infty}
\int_{\Omega} |\nabla u_{k}|^{p - 2} \nabla u_{k} \cdot \nabla \varphi \, dw
=
\int_{\Omega} \varphi \, d \mu[u],
\quad
\forall \varphi \in C_{c}^{\infty}(\Omega).
\]
The measure $\mu[u]$ is called the \textit{Riesz measure} of $u$.
By definition, if $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$, then $\mu[u] = - \Delta_{p, w} u$ in the sense of distribution.
As in Section \ref{sec:introduction},
we denote by $\mathcal{M}^{+}_{0}(\Omega)$ the set of all Radon measures $\mu$
that are absolutely continuous with respect to the $(p, w)$-capacity, i.e.,
$\mu(E) = 0$ whenever $E$ has zero $(p, w)$-capacity.
If $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$ is $(p, w)$-superharmonic in $\Omega$,
then the Riesz measure of $u$ belongs to $\mathcal{M}^{+}_{0}(\Omega)$.
It is known that if $\mu \in \mathcal{M}^{+}_{0}(\Omega)$ is finite,
then the integral $\int_{\Omega} f \, d \mu$ is well-defined for any quasicontinuous function $f$ on $\Omega$.
If $u \in H^{1, p}(\Omega; w)$ is a supersolution to \eqref{eqn:p-harmonic},
then the Riesz measure of $u^{*}$ satisfies
\begin{equation}\label{eqn:finite_energy}
\int_{\Omega} \varphi \, d \mu
\le
C \| \varphi \|_{H^{1, p}(\Omega; w)},
\quad \forall \varphi \in C_{c}^{\infty}(\Omega), \ \varphi \ge 0.
\end{equation}
If $\mu$ is also finite, then we can replace $C_{c}^{\infty}(\Omega)$ with $H_{0}^{1, p}(\Omega; w)$ up to taking a quasicontinuous representative,
i.e., the dual action of $\mu$ has the integral representative.
Conversely, if a finite measure $\mu$ satisfies \eqref{eqn:finite_energy},
then there exists a unique weak solution $u \in H_{0}^{1, p}(\Omega; w)$ to
\[
\int_{\Omega} |\nabla u|^{p - 2} \nabla u \cdot \nabla \varphi \, dw
=
\int_{\Omega} \tilde{ \varphi } \, d \mu
\quad
\forall \varphi \in H_{0}^{1, p}(\Omega; w).
\]
The following weak continuity result was given by Trudinger and Wang \cite{MR1890997}.
\begin{theorem}[{\cite[Theorem 3.1]{MR1890997}}]\label{thm:TW}
Suppose that $\{ u_{k} \}_{k = 1}^{\infty}$ is a sequence of nonnegative $(p, w)$-superharmonic functions in $\Omega$.
Assume that $u_{k} \to u$ a.e. in $\Omega$ and that $u$ is $(p, w)$-superharmonic in $\Omega$.
Let $\mu[u_{k}]$ and $\mu[u]$ be the Riesz measures of $u_{k}$ and $u$, respectively.
Then $\mu[u_{k}]$ converges to $\mu[u]$ weakly, that is
\[
\int_{\Omega} \varphi \, d \mu[u_{k}] \to \int_{\Omega} \varphi \, d \mu[u],
\quad
\forall \varphi \in C_{c}^{\infty}(\Omega).
\]
\end{theorem}
The following Harnack-type convergence theorem follows from combining Theorem \ref{thm:TW} and \cite[Lemma 7.3]{MR2305115}:
If $\{ u_{k} \}_{k = 1}^{\infty}$ is a nondecreasing sequence of $(p, w)$-superharmonic functions in $\Omega$
and if $u := \lim_{k \to \infty} u_{k} \not \equiv \infty$,
then $u$ is $(p, w)$-superharmonic in $\Omega$ and $\mu[u_{k}]$ converges to $\mu[u]$ weakly.
The following form of Wolff potential estimate was first established by
Kilpel\"{a}inen and Mal\'{y} \cite{MR1205885,MR1264000}.
The extension to weighted equations is due to Mikkonenn \cite{MR1386213}.
See also \cite{MR1890997,MR3842214} for other proofs.
\begin{theorem}[{\cite[Theorem 3.1]{MR1386213}}]\label{thm:wolff}
Suppose that $u$ is a nonnegative $(p, w)$-superharmonic function in $B(x, 2R)$.
Let $\mu$ be the Riesz measure of $u$.
Then
\[
\frac{1}{C} {\bf{W}}_{1, p, w}^{R} \mu(x)
\leq
u(x)
\leq
C \left(
\inf_{B(x, R)} u
+
{\bf{W}}_{1, p, w}^{2R} \mu(x)
\right),
\]
where $C \ge 1$ is a constant depending only on $p$, $C_{D}$, $C_{P}$ and $\lambda$, and
${\bf{W}}_{1, p, w}^{R} \mu$ is the \textit{truncated Wolff potential} of $\mu$, which is defined by
\[
{\bf{W}}_{1, p, w}^{R} \mu (x)
:=
\int_{0}^{R}
\left(
r^{p} \frac{ \mu(B(x, r))}{ w(B(x, r)) }
\right)^{\frac{1}{p - 1}}
\frac{dr}{r}.
\]
\end{theorem}
\section{Minimal $p$-superharmonic solutions}\label{sec:potentials}
The following comparison principle improves \cite[Lemma 5.2]{MR3567503}.
We do not assume the finiteness of the Riesz measures here.
\begin{theorem}\label{thm:comparison_principle}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
Let $u$ and $v$ be nonnegative $(p, w)$-superharmonic functions in $\Omega$
with the Riesz measures $\mu$ and $\nu$, respectively.
Assume also that $\mu \le \nu$ and $u \in H_{0}^{1, p}(\Omega; w)$.
Then $u(x) \leq v(x)$ for all $x \in \Omega$.
\end{theorem}
\begin{proof}
For each $k \in \mathbb{N}$, set $v_{k} = \min\{ v, k \}$. Let $\nu_{k}$ be the Riesz measure of $v_{k}$.
By the chain rule of Sobolev functions,
\[
\Psi^{k}_{\epsilon}(v)
:=
\frac{ (k - v)_{+} }{ (k - v)_{+} + \epsilon }
\in H^{1, p}_{\mathrm{loc}}(\Omega; w) \cap L^{\infty}(\Omega),
\quad
\forall \epsilon > 0.
\]
Thus, for any $l \ge k$,
\[
\begin{split}
\int_{\Omega} \varphi \, \Psi^{k}_{\epsilon}(v) \, d \nu_{l}
& =
\int_{\Omega} |\nabla v_{l}|^{p - 2} \nabla v_{l} \cdot \nabla ( \varphi \, \Psi^{k}_{\epsilon}(v) ) \, dw
\\
& \le
\int_{\Omega} |\nabla v_{l}|^{p - 2} \nabla v_{l} \cdot \nabla \varphi \, \Psi^{k}_{\epsilon}(v) \, dw,
\quad
\forall \varphi \in C_{c}^{\infty}(\Omega), \ \varphi \ge 0.
\end{split}
\]
Passing to the limits $\epsilon \to 0$ and $l \to \infty$, we see that $\mathbf{1}_{ \{ v < k \} } \nu \le \nu_{k}$.
Let $\mu_{k} = \mathbf{1}_{ \{ v < k \} } \mu$,
and let $\{ \Omega_{k} \}_{k = 1}^{\infty}$ be a sequence of open sets
such that $\bigcup_{k = 1}^{\infty} \Omega_{k} = \Omega$ and $\Omega_{k} \Subset \Omega_{k + 1}$ for all $k \ge 1$.
Since $\mu_{k} \le \mu$,
there exists a weak solution $u_{k} \in H_{0}^{1, p}(\Omega_{k}; w)$ to
$- \Delta_{p, w} u_{k} = \mu_{k}$ in $\Omega_{k}$.
Let us denote by $u_{k}$ the zero extension of the lsc-regularization of $u_{k}$ again.
By the comparison principle for weak solutions,
$u_{k}(x) \le v_{k}(x) \le v(x)$ for a.e. $x \in \Omega_{k}$.
Accordingly, $u_{k}(x) \le v(x)$ for all $x \in \Omega_{k}$ since $u$ and $v$ are $(p, w)$-superharmonic in $\Omega_{k}$.
Similarly, $u_{k} \le u_{k + 1} \le u$ in $\Omega_{k}$.
Let $u' = \lim_{k \to \infty} u_{k}$, and let $\mu'$ be the Riesz measure of $u'$.
Testing the equation of $u_{k}$ with $u_{k}$, we have
\[
\int_{\Omega} |\nabla u_{k}|^{p} \, d w
=
\int_{\Omega_{k}} u_{k} \, d \mu_{k}
\le
\int_{\Omega_{k}} u_{k} \, d \mu
\le
\| \nabla u \|_{L^{p}(\Omega; w)}^{p - 1}
\| \nabla u_{k} \|_{L^{p}(\Omega; w)}.
\]
Therefore $u' \in H_{0}^{1, p}(\Omega; w)$.
By Theorem \ref{thm:TW}, $\{ \mu_{k} \}_{k = 1}^{\infty}$ converges to $\mu'$ weakly.
Meanwhile, since $\mu( \{ v = \infty \} ) = 0$, $\{ \mu_{k} \}_{k = 1}^{\infty}$ converges to $\mu$ weakly.
Thus $\mu = \mu'$.
By the uniqueness of weak solutions, $u = u' \le v$ in $\Omega$.
\end{proof}
Following \cite[Chapter 2]{MR2778606},
we introduce classes of smooth measures.
\begin{definition}
Let ${S}_{0}$ be the set of all Radon measures satisfying \eqref{eqn:finite_energy}.
For $\mu \in {S}_{0}$,
we denote the lsc-regularization of the solution to \eqref{eqn:variational_problem} by $\mathcal{W}_{p, w}^{0} \mu$.
Furthermore, we define a subset ${S}_{00}$ of ${S}_{0}$ as
\[
{S}_{00}
:=
\left\{
\mu \in {S}_{0} \colon \sup_{\Omega} \mathcal{W}_{p, w}^{0} \mu < \infty \ \text{and} \ \mu(\Omega) < \infty
\right\}.
\]
\end{definition}
\begin{remark}\label{rem:wmp}
Assume that $\mu \in S_{0}$ is finite. Set $u = \mathcal{W}_{p, w}^{0} \mu$.
Then $\sup_{\Omega} u = \| u \|_{L^{\infty}(\mu)}$.
In fact, clearly $\sup_{\Omega} u \ge \| u \|_{L^{\infty}(\mu)}$.
To prove the converse inequality, set $k = \| u \|_{L^{\infty}(\mu)}$.
Then testing the equation of $u$ with $(u - k)_{+}$, we find that
\[
\int_{ \{ u > k \} } |\nabla u|^{p} \, dw
=
\int_{\Omega} (u - k)_{+} \, d \mu
=
0.
\]
Therefore $u(x) \le k$ for a.e. $x \in \Omega$.
Since $u$ is $(p, w)$-superharmonic in $\Omega$,
$u(x) \le k$ for all $x \in \Omega$.
\end{remark}
\begin{remark}
If $\mu \in {S}_{00}$ and $f \in L^{\infty}(\mu)$, then $|f| \mu \in {S}_{00}$.
\end{remark}
\begin{lemma}\label{lem:additivity_s_00}
Let $\mu, \nu \in {S}_{00}$.
Assume also that $\spt(\mu + \nu) \Subset \Omega$.
Then $\mu + \nu \in {S}_{00}$.
\end{lemma}
\begin{proof}
Clearly $\mu + \nu$ is finite and belongs to ${S}_{0}$.
Let $u = \mathcal{W}_{p, w}^{0}(\mu + \nu)$.
By Remark \ref{rem:wmp}, $\sup_{\Omega} u = \| u \|_{L^{\infty}(\mu + \nu)}$.
Let $R = \mathrm{dist}(\spt (\mu + \nu), \partial \Omega) / 4$ and fix $x \in \spt (\mu + \nu)$.
Then, by the latter inequality in Theorem \ref{thm:wolff}, we have
\[
u(x)
\le
C \left( \inf_{B(x, R)} u + {\bf{W}}_{1, p, w}^{2R} (\mu + \nu)(x) \right).
\]
Using a simple calculation and the former inequality in Theorem \ref{thm:wolff}, we also obtain
\[
\begin{split}
{\bf{W}}_{1, p, w}^{2R} (\mu + \nu)(x)
& \le
\max\{ 2^{ \frac{2 - p}{p - 1}}, 1 \} \left( {\bf{W}}_{1, p, w}^{2R}\mu(x) + {\bf{W}}_{1, p, w}^{2R}\nu(x) \right)
\\
& \le
C \left( \mathcal{W}_{p, w}^{0} \mu(x) + \mathcal{W}_{p, w}^{0} \nu(x) \right).
\end{split}
\]
Furthermore, testing the equation of $u$ with $\min \{ u, \inf_{B(x, R)} u \}$
and using the Poincar\'{e} inequality, we find that
\[
\inf_{B(x, R)} u
\le
\left(
\frac{ (\mu + \nu)(\Omega) }{ \mathrm{cap}_{p, w}( \overline{ B(x, R) }, \Omega) }
\right)^{\frac{1}{p - 1}}
\le
C
\left(
\mathrm{diam}(\Omega)^{p} \frac{ (\mu + \nu)(\Omega) }{ w(B(x, R)) }
\right)^{\frac{1}{p - 1}}.
\]
The right-hand side is continuous with respect to $x$.
Hence it is bounded on $\spt( \mu + \nu )$.
Combining these estimates, we obtain the desired boundedness.
\end{proof}
The following characterization for $\mathcal{M}^{+}_{0}(\Omega)$ is
the nonlinear counterpart of \cite[Theorem 2.2.4]{MR2778606} (see also \cite[Corollary 3.19]{MR0386032}).
For related characterizations for $\mathcal{M}_{0}^{+}(\Omega)$,
see also \cite{MR740203} and \cite[Proposition 1.2.7]{MR3676369}.
\begin{theorem}\label{thm:approximation}
Let $\mu \in \mathcal{M}^{+}(\Omega)$.
Then, $\mu \in \mathcal{M}^{+}_{0}(\Omega)$ if and only if
there exists an increasing sequence of compact sets $\{ F_{k } \}_{k = 1}^{\infty}$ such that
$\mu_{k} := \mathbf{1}_{F_{k}} \mu \in {S}_{00}$ for all $k \ge 1$
and
$\mu\left( \Omega \setminus \bigcup_{k = 1}^{\infty} F_{k} \right) = 0$.
\end{theorem}
\begin{proof}
The ``if'' part is readily obtained from the definition of ${S}_{0}$. Let us prove the ``only if'' part.
Consider a sequence of open sets $\{ \Omega_{j} \}_{j = 1}^{\infty}$ such that
$\bigcup_{j = 1}^{\infty} \Omega_{j} = \Omega$ and $\Omega_{j} \Subset \Omega_{j + 1}$ for all $j \ge 1$.
For each $j \ge 1$, $\mathbf{1}_{ \overline{\Omega_{j}} } \mu$ is finite.
Thus, by \cite[Theorem 6.6]{MR1386213},
there exists a $(p, w)$-superharmonic function $u_{j}$ satisfying
\[
\begin{cases}
- \Delta_{p, w} u_{j} = \mathbf{1}_{ \overline{\Omega_{j}} } \mu & \text{in $\Omega$,}
\\
\min\{ u_{j}, k \} \in H_{0}^{1, p}(\Omega; w) & \text{for all $k \ge 1$.}
\end{cases}
\]
Let $F_{k, j} = \{ u_{j} \le k \}$ and let $\mu_{j, k} = \mathbf{1}_{ F_{k, j} \cap \overline{\Omega_{j}} } \mu$.
Using the method in Theorem \ref{thm:comparison_principle},
we see that $\mu_{j, k} \le \mathbf{1}_{ \{ u _{j} < k + 1\} } \mathbf{1}_{ \overline{\Omega_{j}} } \mu \le \mu[ \min\{ u_{j}, k + 1 \} ]$.
Since $\min \{ u_{j}, k + 1 \} \in H_{0}^{1, p}(\Omega; w)$, this implies that $\mu_{j, k} \in {S}_{0}$.
By Remark \ref{rem:wmp},
\[
\sup_{\Omega} \mathcal{W}_{p, w}^{0} \mu_{j, k}
=
\| \mathcal{W}_{p, w}^{0} \mu_{j, k} \|_{L^{\infty}( \mu_{j, k} )}
\le
\| \min \{ u_{j}, k + 1 \} \|_{L^{\infty}( \mu_{j, k} )}
\le
k.
\]
Hence $\mu_{j, k} \in {S}_{00}$.
Let
\[
F_{k} = \bigcup_{1 \le j \le k} \left( F_{k, j} \cap \overline{ \Omega_{j} } \right).
\]
Clearly, $F_{k} \subset F_{k + 1}$ and $F_{k} \subset \overline{ \Omega_{k} }$.
Fix $j \ge 1$. Since $\mu \in \mathcal{M}^{+}_{0}(\Omega)$, we have $\mu(\{ u_{j} = \infty \}) = 0$.
Therefore,
\[
\mathbf{1}_{ \overline{ \Omega_{j} } }
=
\lim_{k \to \infty} \mathbf{1}_{ F_{k, j} \cap \overline{ \Omega_{j} } }
\le
\lim_{k \to \infty} \mathbf{1}_{ F_{k} }
\le
\mathbf{1}_{\Omega}
\quad
\text{ $\mu$-a.e. }
\]
Passing to the limit $j \to \infty$, we see that $\lim_{k \to \infty} \mathbf{1}_{ F_{k} } = \mathbf{1}_{\Omega}$ $\mu$-a.e.
It remains to be shown that $\mathbf{1}_{F_{k}} \mu \in {S}_{00}$.
By the comparison principle for weak solutions,
\[
\sup_{\Omega} \mathcal{W}_{p, w}^{0} \left( \mathbf{1}_{F_{k}} \mu \right)
\le
\sup_{\Omega} \mathcal{W}_{p, w}^{0} \left( \sum_{1\le j \le k} \mu_{j, k} \right).
\]
Iterating Lemma \ref{lem:additivity_s_00} $(k - 1)$ times, we find that the right-hand side is finite.
\end{proof}
Let us now consider the following Dirichlet problem:
\begin{equation}\label{eqn:poisson}
\begin{cases}
- \Delta_{p, w} u = \mu & \text{in $\Omega$},
\\
u = 0 & \text{on $\partial \Omega$.}
\end{cases}
\end{equation}
We say that a function $u$ is a \textit{$(p, w)$-superharmonic solution (supersolution)}
to $- \Delta_{p, w} u = \mu$ in $\Omega$,
if $u$ is a $(p, w)$-superharmonic function in $\Omega$ and $\mu[u] = \mu$ ($\mu[u] \ge \mu$),
where $\mu[u]$ is the Riesz measure of $u$.
We say that a nontrivial nonnegative solution $u$ is \textit{minimal}
if $v \geq u$ in $\Omega$ whenever
$v$ is a nontrivial nonnegative supersolution to the same equation.
\begin{definition}\label{def:potential}
For $\mu \in \mathcal{M}^{+}_{0}(\Omega)$,
we define
\[
\mathcal{W}_{p, w} \mu(x)
:=
\sup \left\{ \mathcal{W}_{p, w}^{0} \nu(x) \colon \nu \in {S}_{00} \ \text{and} \ \nu \le \mu \right\}.
\]
\end{definition}
\begin{remark}
Note that $u := \mathcal{W}_{p, w} \mu$ may be identically infinite.
If $\mu$ is finite, then $u$ is finite q.e. in $\Omega$ and satisfies \eqref{eqn:poisson}
in the sense of entropy or renormalized solutions
(see \cite{MR1205885, MR1409661,MR1402674, MR1386213, MR1760541}).
In general, even if $u \not \equiv \infty$ is a $p$-superharmonic solution to $- \Delta_{p, w} u = \mu$ in $\Omega$,
it does not satisfy the Dirichlet boundary condition in the sense of renormalized solutions.
The author is not aware of the renowned name for this class of solutions.
Sufficient conditions for $u \not \equiv \infty$ will be discussed in the next section.
\end{remark}
\begin{remark}
Clearly, $\mathcal{W}_{p, w}( a \mu) = a^{\frac{1}{p - 1}} \mathcal{W}_{p, w} \mu$ for any constant $a \ge 0$, and
\[
\mu \le \nu
\Rightarrow
\mathcal{W}_{p, w} \mu(x) \le \mathcal{W}_{p, w} \nu(x),
\quad
\forall x \in \Omega.
\]
Furthermore, Theorem \ref{thm:comparison_principle} is still valid when $u = \mathcal{W}_{p, w} \mu$.
\end{remark}
\begin{proposition}\label{prop:potentials}
Let $\mu \in \mathcal{M}^{+}_{0}(\Omega)$.
The following statements hold.
\begin{enumerate}[label=(\roman*)]
\item\label{cond:01_potentials}
Assume that $\mathcal{W}_{p, w} \mu \not \equiv \infty$.
Then $u = \mathcal{W}_{p, w} \mu$ is the minimal nonnegative $(p, w)$-superharmonic solution to
$- \Delta_{p, w} u = \mu$ in $\Omega$.
\item\label{cond:02_potentials}
Let $\{ f_{j} \}_{j = 1}^{\infty} \subset L^{1}_{\mathrm{loc}}(\mu)$ be a nondecreasing sequence of functions.
Assume that $f_{j} \uparrow f$ $\mu$-a.e.
Let $u_{j} = \mathcal{W}_{p, w} ( f_{j} \mu)$ and let $u = \lim_{j \to \infty} u_{j}$.
Assume also that $u \not \equiv \infty$.
Then $u = \mathcal{W}_{p, w} ( f \mu)$.
\item\label{cond:03_potentials}
If $\mu \in {S}_{0}$, then $\mathcal{W}_{p, w} \mu = \mathcal{W}_{p, w}^{0} \mu$.
\end{enumerate}
\end{proposition}
\begin{proof}
\ref{cond:01_potentials}
Let
\[
u'(x) = \lim_{k \to \infty} \mathcal{W}_{p, w}^{0} \mu_{k}(x),
\]
where $\{ \mu_{k} \}_{k = 1}^{\infty}$ is a sequence of Radon measures in Theorem \ref{thm:approximation}.
Since $u' \le u \not \equiv \infty$, it follows from Theorem \ref{thm:TW} that $u'$ is
a nonnegative $(p, w)$-superharmonic solution to $- \Delta_{p, w} u = \mu$ in $\Omega$.
By Theorem \ref{thm:comparison_principle}, if $\nu \le \mu$ and $\nu \in {S}_{00}$, then $\mathcal{W}_{p, w}^{0} \nu \le u'$ in $\Omega$; therefore $u = u'$.
Using the same argument again, we see that $u$ is minimal.
\ref{cond:02_potentials}
Let $\omega$ be the Riesz measure of $u$.
By Theorem \ref{thm:TW}, $f_{k} \mu$ converges to $\omega$ weakly.
By the monotone convergence theorem, $\omega = f \mu$ and $f \mu \in \mathcal{M}_{0}^{+}(\Omega)$.
For each $j \ge 1$, $u_{j} \le \mathcal{W}_{p, w} (f \mu)$, and hence $u \le \mathcal{W}_{p, w} (f \mu)$.
By the minimality of $\mathcal{W}_{p, w}(f \mu)$, $u = \mathcal{W}_{p, w}(f \mu)$.
\ref{cond:03_potentials}
Set $u_{k} = \mathcal{W}_{p, w}^{0} \mu_{k}$.
Since $\mu_{k} \in {S}_{00}$ and $\mu \in {S}_{0}$, we have
\[
\int_{\Omega} |\nabla u_{k}|^{p} \, dw
=
\int_{\Omega} u_{k} \, d \mu_{k}
\le
\int_{\Omega} u_{k} \, d \mu
\le
C \| \nabla u_{k} \|_{ L^{p}(\Omega; w) }.
\]
Passing to the limit $k \to \infty$, we see that $u = \lim_{k \to \infty} u$ belongs to $H_{0}^{1, p}(\Omega; w)$.
From Theorem \ref{thm:TW} and the uniqueness of weak solutions, the assertion follows.
\end{proof}
\begin{lemma}\label{lem:wmp}
Let $\mu \in \mathcal{M}^{+}_{0}(\Omega)$, and let $u = \mathcal{W}_{p, w} \mu$.
Then, $\sup_{\Omega} u = \| u \|_{L^{\infty}(\mu)}$.
\end{lemma}
\begin{proof}
Take $\{ \mu_{k} \}_{k = 1}^{\infty}$ by using Theorem \ref{thm:approximation}.
Set $u_{k} = \mathcal{W}_{p, w}^{0} \mu_{k}$.
By Remark \ref{rem:wmp}, for any $x \in \Omega$,
\[
u(x)
=
\lim_{k \to \infty} u_{k}(x)
\le
\lim_{k \to \infty} \sup_{\Omega} u_{k}
=
\lim_{k \to \infty} \| u_{k} \|_{L^{\infty}(\mu_{k})}
\le
\| u \|_{L^{\infty}(\mu)}.
\]
The converse inequality is clear.
\end{proof}
\section{Generalized energy}\label{sec:generalized_energy}
Let us define generalized $p$-energy by using minimal $p$-superharmonic solutions.
For $p = 2$ or $\Omega = \mathbb{R}^{n}$, we refer to \cite{MR3881877,HARA2020111847}.
\begin{definition}\label{def:generalized_energy}
For $0 \le \gamma < \infty$, set
${S}^{\gamma} := \left\{ \mu \in \mathcal{M}^{+}_{0}(\Omega) \colon \mathcal{E}_{\gamma}( \mu ) < \infty \right\}$,
where
\[
\mathcal{E}_{\gamma}(\mu)
:=
\int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu.
\]
Furthermore, let
$
{S}^{\infty}
:=
\left\{
\mu \in \mathcal{M}^{+}_{0}(\Omega) \colon \| \mathcal{W}_{p, w} \mu \|_{L^{\infty}(\mu)} < \infty
\right\}
$
for $\gamma = \infty$.
\end{definition}
By definition, if $\mu \in {S}^{\gamma}$, then the Dirichlet problem \eqref{eqn:poisson} has
a minimal nonnegative $(p, w)$-superharmonic solution $u = \mathcal{W}_{p, w} \mu$.
We can verify ${S}^{1} = {S}_{0}$ by modifying the proof of Theorem \ref{thm:embedding} below.
For any $0 \le \gamma \le \infty$,
\[
{S}_{00} = {S}^{0} \cap {S}^{\infty} \subset {S}^{\gamma} \subset \mathcal{M}^{+}_{0}(\Omega).
\]
Assume that $0 < \gamma < \infty$ and $\mu \in {S}_{00}$.
Let $u = \mathcal{W}_{p, w}^{0} \mu$ and let $v = u^{\frac{p - 1 + \gamma}{p}}$.
Then, for any $\epsilon > 0$,
\[
\int_{\Omega} (u^{\gamma} - \epsilon)_{+} \, d \mu
=
\gamma \int_{ \{ u^{\gamma} > \epsilon \} } |\nabla u|^{p} u^{\gamma - 1} \, dw
=
\frac{1}{c_{E}} \int_{ \Omega } |\nabla v_{\epsilon}|^{p} \, dw,
\]
where $v_{\epsilon} = (v - \epsilon^{ \frac{p - 1 + \gamma}{\gamma p} })_{+}$ and
$c_{E}$ is the constant in \eqref{eqn:opt-const}.
Thus, by the monotone convergence theorem,
\begin{equation}\label{eqn:integrating_by_parts}
\int_{\Omega} (\mathcal{W}_{p, w}^{0} \mu)^{\gamma} \, d \mu
=
\gamma \int_{\Omega} |\nabla u|^{p} u^{\gamma - 1} \, dw
=
\frac{1}{c_{E}} \int_{ \Omega } |\nabla v|^{p} \, dw.
\end{equation}
Moreover, since $\{ v_{\epsilon} \}$ is a bounded sequence in $H_{0}^{1, p}(\Omega; w)$,
$v \in H_{0}^{1, p}(\Omega; w)$.
\begin{proposition}\label{prop:energy_class}
Let $\mu \in \mathcal{M}^{+}_{0}(\Omega)$, and let $u = \mathcal{W}_{p, w} \mu$.
\begin{enumerate}[label=(\roman*)]
\item\label{statement:02@energy_class}
If $\mu \in {S}^{\gamma}$ with $0 \le \gamma \le 1$,
then $\min \{ u, l \} \in H_{0}^{1, p}(\Omega; w)$ for all $l > 0$.
\item\label{statement:03@energy_class}
If $\mu \in {S}^{\gamma}$ with $1 \le \gamma \le \infty$,
then $u$ belongs to $H^{1, p}_{\mathrm{loc}}(\Omega; w)$ and
satisfies $- \Delta_{p, w} u = \mu$ in the sense of weak solutions.
\item\label{statement:04@energy_class}
Assume that $\mu \in {S}^{\gamma_{0}} \cap {S}^{\gamma_{1}}$ with
$0 \le \gamma_{0} \le 1$ and $1 \le \gamma_{1} \le \infty$.
Then $u = \mathcal{W}_{p, w}^{0} \mu \in H_{0}^{1, p}(\Omega; w)$.
In particular, $u$ satisfies \eqref{eqn:poisson} in the sense of finite energy weak solutions.
\end{enumerate}
\end{proposition}
\begin{proof}
Let $\{ \mu_{k} \}_{k = 1}^{\infty}$ be a sequence of Radon measures in Theorem \ref{thm:approximation},
and let $u_{k} = \mathcal{W}_{p, w}^{0} \mu_{k}$.
\ref{statement:02@energy_class}
Testing the equation of $u_{k}$ with $\min \{ u_{k}, l \} \in H_{0}^{1, p}(\Omega; w)$, we get
\begin{equation}\label{eqn:grad_esti}
\int_{\Omega} |\nabla \min \{ u_{k}, l \}|^{p} dw
=
\int_{\Omega} \min \{ u_{k}, l \} \, d \mu_{k}
\le
l^{1 - \gamma}
\int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu.
\end{equation}
Letting $k \to \infty$ gives the desired boundary condition.
\ref{statement:03@energy_class}
Assume that $1 \le \gamma < \infty$.
Take a ball $B$ such that $B \Subset \Omega$.
Without loss of generality, we may assume that $u_{k_{0}} \neq 0$ for some $k_{0} \ge 1$.
By the strong minimum principle, $\inf_{ \overline{B} } u_{k_{0}} > 0$.
Let $k \geq k_{0}$.
Subsequently, \eqref{eqn:integrating_by_parts} gives
\[
\begin{split}
\int_{B} |\nabla u_{k}|^{p} \, dw
& \le
\frac{1}{\gamma} \left( \inf_{ \overline{B} } u_{k} \right)^{1 - \gamma}
\int_{\Omega} (\mathcal{W}_{p, w} \mu_{k})^{\gamma} \, d \mu_{k}
\\
& \le
\frac{1}{\gamma} \left( \inf_{ \overline{B} } u_{k_{0}} \right)^{1 - \gamma}
\int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu.
\end{split}
\]
On the other hand, by the Poincar\'{e} inequality,
\[
\begin{split}
\left( \int_{B} u_{k}^{p} \, d w \right)^{\frac{p - 1 + \gamma}{p}}
& \le
w(B)^{ \frac{\gamma - 1}{p} }
\int_{\Omega} u_{k}^{p - 1 + \gamma} \, d w
\le
C \int_{\Omega} |\nabla u_{k}^{ \frac{p - 1 + \gamma}{p} }|^{p} \, d w
\\
& =
c_{E} C \int_{\Omega} (\mathcal{W}_{p, w} \mu_{k})^{\gamma} \, d \mu_{k}
\le
c_{E} C \int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu.
\end{split}
\]
Hence, $\| u_{k} \|_{H^{1, p}(B; w)}$ is bounded.
Passing to the limit $k \to \infty$, we see that $u \in H^{1, p}(B; w)$.
Thus, $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$.
If $\mu \in {S}^{\infty}$, then $u$ is a bounded $(p, w)$-superharmonic function.
Therefore, $u \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$.
\ref{statement:04@energy_class}
By H\"{o}lder's inequality, ${S}^{\gamma_{0}} \cap {S}^{\gamma_{1}} \subset {S}^{1} = {S}_{0}$.
Hence, $u = \mathcal{W}_{p, w} \mu \in H_{0}^{1, p}(\Omega; w)$.
\end{proof}
As in the proof of \cite[Lemma 3.3]{HARA2020111847},
the Picone-type inequality in \cite{MR1618334,MR3273896} yields the following estimate.
\begin{lemma}\label{lem:trace_estimate}
Let $1 < p < \infty$, $0 < \gamma < \infty$ and $- \gamma < q < p - 1$.
Suppose that $\nu \in {S}_{00}$ and $v = \mathcal{W}_{p, w} \nu$.
Assume that $u \in H_{0}^{1, p}(\Omega; w) \cap L^{\infty}(\Omega)$ and $u \ge 0$ in $\Omega$.
Then,
\[
\begin{split}
\int_{\Omega} \tilde{u}^{\gamma + q} \, d \nu
& \le
\left(
\left( \frac{p - 1 + \gamma}{p} \right)^{p}
\int_{\Omega} |\nabla u|^{p} u^{\gamma - 1} \, d w
\right)^{ \frac{\gamma + q}{p - 1 + \gamma} }
\\
& \quad
\times
\left(
\int_{\Omega} v^{\frac{(\gamma + q)(p - 1)}{p - 1 - q}} \, d \nu
\right)^{ \frac{p - 1 - q}{p - 1 + \gamma} }.
\end{split}
\]
\end{lemma}
\begin{proof}[Proof of Theorem \ref{thm:embedding}]
We first prove the upper bound of $C_{T}$.
Take a sequence of measures $\{ \sigma_{k} \}_{k = 1}^{\infty} \subset {S}_{00}$ by using Theorem \ref{thm:approximation}.
Apply Lemma \ref{lem:trace_estimate} to $u_{k} = \min\{ |u|, k \}$ and $\sigma_{k}$; we get
\[
\begin{split}
\int_{\Omega} u_{k}^{1 + q} \, d \sigma_{k}
\le
\left( \int_{\Omega} |\nabla u_{k}|^{p} \, d w \right)^{ \frac{1 + q}{p} }
\left(
\int_{\Omega} (\mathcal{W}_{p, w} \sigma_{k})^{\frac{(1 + q)(p - 1)}{p - 1 - q}} \, d \sigma_{k}
\right)^{ \frac{p - 1 - q}{p} }.
\end{split}
\]
The desired estimate then follows from the monotone convergence theorem.
Let us prove the lower bound.
Take $\{ \sigma_{k} \}_{k = 1}^{\infty} \subset {S}_{00}$.
Let $u = \mathcal{W}_{p, w}^{0} \sigma_{k}$ and let $v = u^{\frac{p - 1}{p - 1 - q}}$.
Note that
\[
\nabla v = \frac{p - 1}{p - 1 - q} \nabla u u^{ \frac{q}{p - 1 - q} }
\quad
\text{a.e. in $\Omega$.}
\]
Thus, using \eqref{eqn:integrating_by_parts} with $\gamma = \frac{(1 + q)(p - 1) }{p - 1 - q}$, we get
\[
\begin{split}
\int_{\Omega} |\nabla v|^{p} \, d w
& =
\left( \frac{p - 1}{p - 1 - q} \right)^{p}
\int_{\Omega} |\nabla u|^{p} u^{ \frac{pq}{p - 1 - q} } \, dw
\\
& =
\frac{1}{1 + q}
\left( \frac{p - 1}{p - 1 - q} \right)^{p - 1}
\int_{\Omega} u^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma_{k}.
\end{split}
\]
By density, \eqref{eqn:trace_ineq} gives
\[
\begin{split}
&
\left(
\int_{\Omega} u^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma_{k}
\right)^{\frac{1}{1 + q}}
\le
\| v \|_{L^{1 + q}(\sigma)}
\le
C_{T} \| \nabla v \|_{L^{p}(w)}
\\
\quad & =
\left( \frac{1}{1 + q} \right)^{ \frac{1}{p} }
\left( \frac{p - 1}{p - 1 - q} \right)^{ \frac{p - 1}{p} }
C_{T}
\left(
\int_{\Omega} u^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma_{k}
\right)^{\frac{1}{p}}.
\end{split}
\]
Therefore,
\[
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma_{k} \right)^{ \frac{(1 + q)(p - 1)}{p - 1 - q} } \, d \sigma_{k}
\le
\left( \frac{1}{1 + q} \right)^{ \frac{1 + q}{p - 1 - q} } \frac{1}{c_{V}^{1 + q}} \,
C_{T}^{ \frac{(1 + q)p}{p - 1 - q} }.
\]
Passing to the limit $k \to \infty$, we arrive at the desired lower bound.
\end{proof}
\begin{corollary}\label{cor:energy_esti}
Let $\mu \in \mathcal{M}^{+}_{0}(\Omega)$ and let $0 < \gamma < \infty$.
Then, $\mu \in {S}^{\gamma}$ if and only if $v := ( \mathcal{W}_{p, w} \mu )^{ \frac{p - 1 + \gamma}{p} } \in H_{0}^{1, p}(\Omega; w)$.
Moreover,
\[
\int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu
\le
\int_{ \Omega } |\nabla v|^{p} \, dw
\le
c_{E} \int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu.
\]
\end{corollary}
\begin{proof}
Let $\{ \mu_{k} \}_{k = 1}^{\infty}$ be a sequence of Radon measures in Theorem \ref{thm:approximation}.
Set $u_{k} = \mathcal{W}_{p, w}^{0} \mu_{k}$ and $v_{k} = \left( u_{k} \right)^{ \frac{p - 1 + \gamma}{p} }$.
Assume that $\mu \in {S}^{\gamma}$.
Then, by \eqref{eqn:integrating_by_parts}, $\{ v_{k} \}_{k = 1}^{\infty}$ is a bounded sequence in $H_{0}^{1, p}(\Omega; w)$.
Since $v_{k} \uparrow v$, $v$ belongs to $H_{0}^{1, p}(\Omega; w)$ and satisfies the latter inequality.
Conversely, assume that $v \in H_{0}^{1, p}(\Omega; w)$.
Then, by Theorem \ref{thm:embedding},
\[
\int_{\Omega} u_{k}^{\gamma} \, d \mu_{k}
\le
\int_{\Omega} v^{1 + q} \, d \mu_{k}
\le
\left( \int_{\Omega} |\nabla v|^{p} \, dw \right)^{\frac{1 + q}{p}}
\left( \int_{\Omega} u_{k}^{\gamma} \, d \mu_{k} \right)^{ \frac{p - 1 - q}{p} },
\]
where $q = \frac{\gamma p}{p - 1 + \gamma} - 1$.
Thus using the monotone convergence theorem, we obtain
\[
\int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu
=
\lim_{k \to \infty} \int_{\Omega} u_{k}^{\gamma} \, d \mu_{k}
\le
\int_{\Omega} |\nabla v|^{p} \, dw.
\]
This completes the proof.
\end{proof}
We also obtain the following estimate using a similar approximation argument.
\begin{theorem}\label{thm:MEE}
Let $1 < p < \infty$, $0 < \gamma < \infty$ and $- \gamma < q < p - 1$.
Then, for any $\mu, \nu \in \mathcal{M}^{+}_{0}(\Omega)$,
\begin{equation}\label{eqn:01@MEE}
\int_{\Omega} ( \mathcal{W}_{p, w} \mu )^{\gamma + q} \, d \nu
\le
\left(
c_{E} \int_{\Omega} (\mathcal{W}_{p, w} \mu)^{\gamma} \, d \mu
\right)^{ \frac{\gamma + q}{p - 1 + \gamma} }
\mathcal{E}_{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }(\nu)^{\frac{p - 1 - q}{p - 1 + \gamma}}.
\end{equation}
In particular,
if $\mu \in {S}^{\gamma}$ and $\nu \in {S}^{\frac{ (\gamma + q)(p - 1) }{ p - 1 - q}}$,
then $\mathcal{W}_{p, w} \mu \in L^{\gamma + q}(\nu)$.
\end{theorem}
\begin{remark}
Note that $c_{E} = 1$ if $\gamma = 1$.
This sharp constant is achieved when $\mu = \nu$ is an equilibrium measure.
\end{remark}
\begin{remark}\label{rem:quasi-additivity}
For $0 < \gamma < \infty$, set $\trinorm{\mu}_{\gamma} = \mathcal{E}_{\gamma}(\mu)^{\frac{p - 1}{p - 1 + \gamma}}$.
Then by Theorem \ref{thm:MEE},
\[
\begin{split}
&
\int_{\Omega} \mathcal{W}_{p, w}(\mu + \nu)^{\gamma} d (\mu + \nu)
=
\int_{\Omega} \mathcal{W}_{p, w}(\mu + \nu)^{\gamma} d \mu
+
\int_{\Omega} \mathcal{W}_{p, w}(\mu + \nu)^{\gamma} d \nu
\\
& \quad
\le
c_{E}^{\gamma} \left( \int_{\Omega} \mathcal{W}_{p, w}(\mu + \nu)^{\gamma} d (\mu + \nu) \right)^{\frac{\gamma}{p - 1 + \gamma}}
\left( \trinorm{\mu}_{\gamma} + \trinorm{\nu}_{\gamma} \right),
\end{split}
\]
and hence
\[
\trinorm{\mu + \nu}_{\gamma} \le c_{E}^{\gamma} \left( \trinorm{\mu}_{\gamma} + \trinorm{\nu}_{\gamma} \right).
\]
In particular, ${S}^{\gamma}$ is a convex cone.
\end{remark}
Finally, we prove weighted norm inequalities.
\begin{theorem}\label{thm:WNI}
Let $1 < p < \infty$, $0 < q < p - 1$ and $0 < \gamma < \infty$.
Assume that $\sigma \in {S}^{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }$.
Then, for any $f \in L^{ \frac{\gamma + q}{q} }(\sigma)$, $|f| \sigma \in {S}^{\gamma}$.
Moreover,
\[
\mathcal{E}_{\gamma}( |f| \sigma )^{\frac{p - 1}{p - 1 + \gamma}}
\le
\left(
c_{E} \, \mathcal{E}_{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }(\sigma)^{\frac{p - 1 - q}{\gamma + q}}
\right)^{ \frac{\gamma}{p - 1 + \gamma} }
\| f \|_{L^{ \frac{\gamma + q}{q} }(\sigma)}
\]
and
\[
\| \mathcal{W}_{p, w} ( |f| \sigma ) \|_{L^{\gamma + q}( \sigma )}
\le
\left(
c_{E} \, \mathcal{E}_{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }(\sigma)^{\frac{p - 1 - q}{\gamma + q}}
\right)^{ \frac{1}{p - 1} }
\| f \|_{L^{ \frac{\gamma + q}{q} }(\sigma)}^{\frac{1}{p - 1}}.
\]
\end{theorem}
\begin{proof}
We may assume that $f \ge 0$ without loss of generality.
By Theorem \ref{thm:MEE},
\[
\begin{split}
\int_{\Omega} ( \mathcal{W}_{p, w}(f \sigma) )^{\gamma + q} \, d \sigma
& \le
\left(
c_{E}
\int_{\Omega} (\mathcal{W}_{p, w}(f \sigma))^{\gamma} f \, d \sigma
\right)^{ \frac{\gamma + q}{p - 1 + \gamma} }
\mathcal{E}_{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }(\sigma)^{\frac{p - 1 - q}{p - 1 + \gamma}}.
\end{split}
\]
Meanwhile, by H\"{o}lder's inequality,
\[
\int_{\Omega} (\mathcal{W}_{p, w}(f \sigma))^{\gamma} f \, d \sigma
\le
\left(
\int_{\Omega} (\mathcal{W}_{p, w}(f \sigma))^{\gamma + q} \, d \sigma
\right)^{\frac{\gamma}{\gamma + q}}
\| f \|_{L^{ \frac{\gamma + q}{q} }(\sigma)}.
\]
Combining the two inequalities, we obtain the desired estimates.
\end{proof}
\begin{remark}\label{rem:two-weight}
Under the same assumptions,
suppose also that
$\nu \in {S}^{ \frac{(\gamma + Q)(p - 1)}{p - 1 - Q} }$
with
$- \gamma < Q < p - 1$.
Then, the following two weight norm inequality holds:
\[
\| \mathcal{W}_{p, w}( |f| \sigma) \|_{L^{\gamma + Q}(\nu)}
\le
C \| f \|_{L^{ \frac{\gamma + q}{q} }(\sigma)}^{ \frac{1}{p - 1} },
\quad
\forall f \in L^{\frac{\gamma + q}{q}}(\sigma).
\]
In fact, by Theorem \ref{thm:MEE},
\[
\begin{split}
\int_{\Omega} ( \mathcal{W}_{p, w}(f \sigma) )^{\gamma + Q} \, d \nu
& \le
C
\left(
\int_{\Omega} (\mathcal{W}_{p, w}(f \sigma))^{\gamma} f \, d \sigma
\right)^{ \frac{\gamma + Q}{p - 1 + \gamma} }
\mathcal{E}_{ \frac{(\gamma + Q)(p - 1)}{p - 1 - Q} }(\nu)^{\frac{p - 1 - Q}{p - 1 + \gamma}}.
\end{split}
\]
The right-hand side is estimated by Theorem \ref{thm:WNI}.
\end{remark}
\section{Properties of solutions to \eqref{eqn:p-laplace}}\label{sec:proof_of_main_theorem}
First, we give the counterpart of \cite[Lemma 3.5]{MR3567503} or \cite[Remark 2.6]{MR4105916}.
\begin{lemma}\label{lem:iterated_ineq}
Let $\sigma \in \mathcal{M}_{0}(\Omega)$, and let $\beta \ge 1$.
Assume that $\left( \mathcal{W}_{p, w} \sigma \right)^{(\beta - 1)(p - 1)} \sigma \in \mathcal{M}_{0}(\Omega)$.
Then,
\[
\left( \mathcal{W}_{p, w} \sigma \right)^{\beta}(x)
\le
\beta \, \mathcal{W}_{p, w} \left( \left( \mathcal{W}_{p, w} \sigma \right)^{(\beta - 1)(p - 1)} \sigma \right)(x),
\quad
\forall x \in \Omega.
\]
\end{lemma}
\begin{proof}
By Proposition \ref{prop:potentials},
we may assume that $\sigma \in {S}_{00}$ without loss of generality.
We use the argument in \cite[Lemma 4.4]{HARA2020111847}.
Let $u = \mathcal{W}_{p, w} \sigma$.
Since $u$ is bounded on $\Omega$,
$u^{\beta} \in H_{0}^{1, p}(\Omega; w)$
and
$u^{(\beta - 1)(p - 1)} \sigma \in {S}_{00}$.
Fix a nonnegative function $\varphi \in C_{c}^{\infty}(\Omega)$.
Testing the equation of $u$ with $\varphi (u^{(\beta - 1)(p - 1)} - \epsilon)_{+}$, we find that
\[
\begin{split}
\int_{\Omega} \varphi (u^{(\beta - 1)(p - 1)} - \epsilon)_{+} \, d \sigma
& =
\int_{\Omega}
|\nabla u|^{p - 2} \nabla u \cdot \nabla \left( \varphi (u^{(\beta - 1)(p - 1)} - \epsilon)_{+} \right)
\, dw
\\
& \ge
\beta^{1 - p}
\int_{ \{ u^{(\beta - 1)(p - 1)} > \epsilon \} }
|\nabla u^{\beta}|^{p - 2} \nabla u^{\beta}
\cdot \nabla \varphi
\, dw.
\end{split}
\]
Applying the dominated convergence theorem to the right-hand side, we obtain
\[
\begin{split}
\int_{\Omega} \varphi u^{(\beta - 1)(p - 1)} \, d \sigma
\ge
\beta^{1 - p}
\int_{\Omega}
|\nabla u^{\beta}|^{p - 2} \nabla u^{\beta}
\cdot \nabla \varphi
\, dw.
\end{split}
\]
By the comparison principle for weak solutions, this implies that
\[
u^{\beta}
\le
\mathcal{W}_{p, w}^{0}( \beta^{p - 1} u^{(\beta - 1)(p - 1)} \sigma)
\quad
\text{a.e. in $\Omega$.}
\]
Since $u$ is $(p, w)$-superharmonic in $\Omega$, the desired inequality holds.
\end{proof}
Next, we give the counterpart of \cite[Theorem 3.4]{MR3567503} or \cite[Theorem 1.3]{MR4105916}.
\begin{theorem}\label{thm:lower_bound}
Let $1 < p < \infty$ and $0 < q < p - 1$.
Let $\sigma \in \mathcal{M}^{+}_{0}(\Omega)$.
Let $v \in L^{q}_{\mathrm{loc}}(\sigma)$ be a nontrivial nonnegative $(p, w)$-superharmonic supersolution to
$- \Delta_{p, w} v = \sigma v^{q}$ in $\Omega$.
Then,
\[
v(x) \ge c_{V} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{p - 1}{p - 1 - q} }(x),
\quad
\forall x \in \Omega,
\]
where $c_{V}$ is the constant in \eqref{eqn:opt-const}.
\end{theorem}
\begin{proof}
For simplicity, we write $\mathcal{W}_{p, w} \mu$ as $\mathcal{W} \mu$.
Let $u = \mathcal{W}(v^{q} \sigma)$.
By Theorem \ref{thm:comparison_principle}, $v(x) \ge u(x)$ for all $x \in \Omega$.
Fix $a > 0$, and set $\sigma_{a} = \mathbf{1}_ { \{ x\in \Omega \colon u(x) > a \} } \sigma$.
Using Theorem \ref{thm:comparison_principle} again, we get
\[
u
\ge
\mathcal{W}(u^{q} \sigma)
\ge
\mathcal{W}(a^{q} \sigma_{a} )
=
a^{\frac{q}{p - 1}} \mathcal{W} \sigma_{a}.
\]
Continuing this argument $k$-times, we obtain
\begin{equation}\label{eqn:01@thm:lower_bound}
\begin{split}
u
& \ge
\mathcal{W}( \mathcal{W}( \cdots ( \mathcal{W}( u^{q} \sigma) )^{q} \cdots \sigma)^{q} \sigma)
\\
& \ge
\mathcal{W}( \mathcal{W}( \cdots ( \mathcal{W}( a^{q} \sigma_{a}) )^{q} \cdots \sigma )^{q} \sigma )
\\
& \ge
a^{ (\frac{q}{p - 1})^{k} }
\mathcal{W}( \mathcal{W}( \cdots ( \mathcal{W} \sigma_{a} )^{q} \cdots \sigma_{a} )^{q} \sigma_{a} ).
\end{split}
\end{equation}
Meanwhile, by Lemma \ref{lem:iterated_ineq},
\[
\left( \mathcal{W} \sigma_{a} \right)^{\beta_{i + 1}}
\le
\beta_{i + 1} \, \mathcal{W} \left( \left( \mathcal{W} \sigma_{a} \right)^{\beta_{i} q} \sigma_{a} \right)
\]
for each $i \ge 0$, where $\beta_{0} = 1$ and $\beta_{i + 1} = \beta_{i} \frac{q}{p - 1} + 1$.
Iterating this estimate $k$ times, we get
\begin{equation}\label{eqn:02@thm:lower_bound}
(\mathcal{W} \sigma_{a})^{\beta_{k}}
\le
\prod_{i = 1}^{k} \beta_{i}^{ \left( \frac{q}{p - 1} \right)^{k - i} }
\mathcal{W}( \mathcal{W}( \cdots ( \mathcal{W} \sigma_{a} )^{q} \cdots \sigma_{a} )^{q} \sigma_{a} ).
\end{equation}
By definition, $\beta_{k} = \sum_{i = 0}^{k} \left( \frac{q}{p - 1} \right)^{i}$.
Therefore $\beta_{k} \uparrow \frac{p - 1}{p - 1 - q} $ as $k \to \infty$ and
\begin{equation}\label{eqn:03@thm:lower_bound}
\prod_{i = 1}^{k} \beta_{i}^{ \left( \frac{q}{p - 1} \right)^{k - i} }
\le
\left( \frac{p - 1}{p - 1 - q} \right)^{ \sum_{i = 1}^{k} \left( \frac{q}{p - 1} \right)^{k - i} }
\le
\left( \frac{p - 1}{p - 1 - q} \right)^{ \frac{p - 1}{p - 1 - q} }.
\end{equation}
Combining \eqref{eqn:01@thm:lower_bound}, \eqref{eqn:02@thm:lower_bound} and \eqref{eqn:03@thm:lower_bound}
and letting $k \to \infty$, we obtain
\[
u
\ge
\left( \frac{p - 1 - q}{p - 1} \right)^{ \frac{p - 1}{p - 1 - q} }
(\mathcal{W} \sigma_{a})^{ \frac{p - 1}{p - 1 - q} }.
\]
Without loss of generality, we may assume that $\sigma_{a} \neq 0$ for small $a > 0$.
Then $u \ge a^{\frac{q}{p - 1} }\mathcal{W} \sigma_{a} > 0$ in $\Omega$ by the strong minimum principle.
Thus, taking the limit $a \to 0$, we arrive at the desired estimate.
\end{proof}
Finally, we prove Theorem \ref{thm:main_theorem} and its variants.
\begin{proof}[Proof of Theorem \ref{thm:main_theorem}]
\ref{enum:01@main_theorem} $\Rightarrow$ \ref{enum:02@main_theorem}:
Using Theorem \ref{thm:lower_bound}, we find that
\[
\begin{split}
\left(
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma \right)^{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} } \, d \sigma
\right)^{\frac{1}{\gamma + q}}
& \le
\frac{1}{c_{V}}
\| v \|_{L^{\gamma + q}(\sigma)}
\le
\frac{C_{1}}{c_{V}}.
\end{split}
\]
\ref{enum:02@main_theorem} $\Rightarrow$ \ref{enum:03@main_theorem}:
By Theorem \ref{thm:WNI}, \eqref{eqn:weighted_norm_ineq} holds with
\[
\begin{split}
C_{3}^{ \frac{p - 1}{p - 1 - q}}
\le
\left(
c_{E} \mathcal{E}_{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} }(\sigma)^{\frac{p - 1 - q}{\gamma + q}}
\right)^{ \frac{1}{p - 1 - q} }
\le
c_{E}^{ \frac{1}{p - 1 - q} } C_{2}.
\end{split}
\]
\ref{enum:03@main_theorem} $\Rightarrow$ \ref{enum:01@main_theorem}:
Take $\{ \mathbf{1}_{F_{k}} \sigma \}_{k = 1}^{\infty} \subset {S}_{00}$ by using Theorem \ref{thm:approximation}.
Applying \eqref{eqn:weighted_norm_ineq} to
$f = \left( \mathcal{W} \sigma_{k} \right)^{ \frac{q(p - 1)}{p - 1 - q} } \mathbf{1}_{F_{k}} \in L^{ \frac{\gamma + q}{q}}(\sigma)$
and using Lemma \ref{lem:iterated_ineq},
we obtain
\[
\begin{split}
\left(
\int_{\Omega} \left( \mathcal{W}_{p, w} \sigma_{k} \right)^{ \frac{(\gamma +q)(p - 1)}{p - 1 - q} } \mathbf{1}_{F_{k}} \, d \sigma
\right)^{\frac{1}{\gamma + q}}
\le
C.
\end{split}
\]
Therefore, by the monotone convergence theorem,
\[
u_{0} := c_{V} \left(\mathcal{W}_{p, w} \sigma \right)^{\frac{p - 1}{p - 1 - q}} \in L^{ \gamma + q }(\sigma),
\]
where $c_{V}$ is the constant in \eqref{eqn:opt-const}.
Define a sequence of $(p, w)$-superharmonic functions $\{ u_{i} \}_{i = 1}^{\infty}$ by
\[
u_{i + 1} := \mathcal{W}_{p, w} (u_{i}^{q} \sigma), \quad i = 1, 2, \dots.
\]
By \eqref{eqn:weighted_norm_ineq},
\[
\begin{split}
\| u_{i + 1} \|_{L^{\gamma + q}(\sigma)}
=
\| \mathcal{W}_{p, w} ( u_{i}^{q} \sigma) \|_{L^{\gamma + q}(\sigma)}
\le
C_{3} \| u_{i} \|_{L^{ \gamma + q }(\sigma)}^{ \frac{q}{p - 1} },
\end{split}
\]
and hence $\{ u_{i} \}_{i = 1}^{\infty} \subset L^{\gamma + q}(\sigma) \subset L^{q}_{\mathrm{loc}}(\sigma)$.
By Lemma \ref{lem:iterated_ineq}, $u_{0} \le u_{1}$.
Hence, by induction, $u_{i} \le u_{i + 1}$ for all $i \ge 1$.
Let $u = \lim_{i \to \infty} u_{i}$.
By the monotone convergence theorem,
\[
\| u \|_{L^{\gamma + q}(\sigma)}
=
\lim_{i \to \infty} \| u_{i + 1} \|_{L^{\gamma + q}(\sigma)}
\le
C_{3}^{\frac{p - 1}{p - 1 - q}}.
\]
By Proposition \ref{prop:potentials}, $u$ is $(p, w)$-superharmonic in $\Omega$ and $u = \mathcal{W}_{p, w} ( u^{q} \sigma)$.
Assume that $v$ is a nontrivial nonnegative $(p,w)$-superharmonic solution to
$- \Delta_{p, w} v = \sigma v^{q}$ in $\Omega$.
Then $u_{0} \le v$ by Theorem \ref{thm:lower_bound}, and hence, $u_{i} \le v$ for all $i \ge 1$ by induction.
Therefore $u \le v$.
\end{proof}
\begin{remark}
In Theorem \ref{thm:main_theorem},
the equivalence \ref{enum:01@main_theorem} $\Leftrightarrow$ \ref{enum:02@main_theorem}
still holds even if $q = 0$.
\end{remark}
\begin{theorem}\label{thm:main_theorem_infty}
Let $\Omega$ be a bounded domain in $\mathbb{R}^{n}$.
Let $1 < p < \infty$ and $0 < q < p - 1$.
Suppose that $\sigma \in \mathcal{M}^{+}_{0}(\Omega) \setminus \{ 0 \}$.
Then, the following statements are equivalent:
\begin{enumerate}[label=(\arabic*)]
\item\label{enum:01@main_theorem_infty}
There exists a bounded positive weak supersolution
$v \in H^{1, p}_{\mathrm{loc}}(\Omega; w)$ to $- \Delta_{p, w} v = \sigma v^{q}$ in $\Omega$
satisfying $\| v \|_{ L^{\infty}(\sigma) } \le C_{1} < \infty$.
\item\label{enum:02@main_theorem_infty}
$\| \mathcal{W}_{p, w} \sigma \|_{L^{\infty}(\sigma)}^{ \frac{p - 1}{p - 1 - q} } \le C_{2} < \infty$.
\item\label{enum:03@main_theorem_infty}
The following weighted norm inequality holds:
\[
\| \mathcal{W}_{p, w} ( |f| \sigma ) \|_{L^{\infty}( \sigma )}
\le
C_{3} \| f \|_{L^{ \infty }( \sigma )}^{\frac{1}{p - 1}},
\quad
\forall f \in L^{\infty}( \sigma ).
\]
\end{enumerate}
Moreover, if $C_{i}$ ($i = 1, 2 ,3$) are the best constants in the above statements, then
\[
C_{1} \le C_{3}^{ \frac{p - 1}{p - 1 - q} } \le C_{2} \le \frac{ C_{1} }{c_{V}}.
\]
In addition, if one of the above statements holds,
then there exists a minimal positive $(p, w)$-superharmonic solution $u$ to
$- \Delta_{p, w} u = \sigma u^{q}$ in $\Omega$
such that $\sup_{\Omega} u \le C_{1}$.
\end{theorem}
\begin{proof}
We modify the Proof of Theorem \ref{thm:main_theorem}.
Clearly, \ref{enum:01@main_theorem_infty} $\Rightarrow$ \ref{enum:02@main_theorem_infty} $\Rightarrow$ \ref{enum:03@main_theorem_infty}.
Assume that \ref{enum:03@main_theorem_infty} holds. Then, for all $i \ge 1$,
\[
\begin{split}
\| u_{i + 1} \|_{L^{\infty}(\sigma)}
=
\| \mathcal{W}_{p, w} ( u_{i}^{q} \sigma) \|_{L^{\infty}(\sigma)}
\le
C_{3} \| u_{i} \|_{L^{\infty}(\sigma)}^{ \frac{q}{p - 1} }.
\end{split}
\]
Thus,
\[
\| u \|_{L^{\infty}(u^{q} \sigma)}
\le
\| u \|_{L^{\infty}(\sigma)}
\le
C_{3}^{ \frac{p - 1}{p - 1 - q} }.
\]
By Lemma \ref{lem:wmp},
$\sup_{\Omega} u \le C_{3}^{ \frac{p - 1}{p - 1 - q} }$ and \ref{enum:01@main_theorem_infty} holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:main_theorem_FE}]
The case of $q = 0$ is \cite[Corollary 21.18]{MR2305115}, so we consider $q > 0$.
Assume that a positive finite energy solution $u$ exists.
We may assume that $u$ is quasicontinuous without loss of generality.
Then, we find that
\[
\begin{split}
\| \nabla u \|_{L^{p}(w)}^{p}
& =
\int_{\Omega} |\nabla u|^{p} \, d w
=
\int_{\Omega} u^{1 + q} \, d \sigma
=
\| u \|_{L^{1 + q}(\sigma)}^{1 + q}.
\end{split}
\]
Therefore, \eqref{eqn:energy_cond@FE} follows from Theorem \ref{thm:main_theorem}.
Conversely, assume that \eqref{eqn:energy_cond@FE} holds.
Then, Theorem \ref{thm:main_theorem} and Proposition \ref{prop:energy_class} give
a minimal positive finite energy weak solution $u \in H_{0}^{1, p}(\Omega; w)$ to $- \Delta_{p, w} u = \sigma u^{q}$ in $\Omega$.
The uniqueness of such a solution follows from a convexity argument as in \cite[Theorem 5.1]{MR3311903}.
\end{proof}
\begin{remark}
As the proof of \cite[Corollary 6.3]{HARA2020111847},
using Remark \ref{rem:quasi-additivity},
we can construct weak solutions to quasilinear equations
of the type
\[
\begin{cases}
\displaystyle
- \Delta_{p, w} u = \sum_{j = 1}^{J} \sigma_{j} u^{q_{j}}
& \text{in} \ \Omega,
\\
u = 0
& \text{on} \ \partial \Omega,
\end{cases}
\]
where $0 \le q_{j} < p - 1$ and $\sigma_{j} \in {S}^{ \frac{ (1 + q_{j})(p - 1) }{p - 1 - q_{j}} }$ for $j = 1, 2, \dots, J$.
\end{remark}
\section{Quasilinear PDE with $L^{s, t}$ coefficients}\label{sec:example1}
Let us assume that $\mu \in \mathcal{M}^{+}_{0}(\Omega)$ is finite.
Then as in \cite[Theorem 2.1]{MR2456885},
\begin{equation}\label{eqn:wolff_UB}
\mathcal{W}_{p, w} \mu(x)
\le
C {\bf{W}}_{1, p, w}^{ 2 \mathrm{diam}(\Omega) } \mu(x),
\quad
\forall x \in \Omega.
\end{equation}
Using this, we can estimate the generalized $p$-energy of $\mu$.
We now consider unweighted equations.
As the usual notation, we write $H_{0}^{1, p}(\Omega; 1)$ as $W_{0}^{1, p}(\Omega)$.
For a Lebesgue measurable function $f$ on $\Omega$, we define
\[
\| f \|_{L^{r, \rho}(\Omega)}
=
\begin{cases}
\displaystyle
\left( \int_{0}^{\infty} \left( t^{\frac{1}{r}} f^{*}(t) \right)^{\rho} \frac{dt}{t} \right)^{\frac{1}{\rho}}
& \text{if} \ \rho < \infty, \\
\displaystyle
\sup_{t > 0} t^{\frac{1}{r}} f^{*}(t)
& \text{if} \ \rho = \infty,
\end{cases}
\]
where $0 < r, \rho \leq \infty$ and
$f^{*}(t) = \inf \{ \alpha > 0 \colon | \{ x \in \Omega \colon |f(x)| > \alpha \} | \leq t \}$.
The space of all $f$ with $\| f \|_{L^{r, \rho}(\Omega)} < \infty$ is called the \textit{Lorentz space}.
For the basics of Lorentz spaces, we refer to \cite[Chapter 1]{MR2445437}.
\begin{corollary}\label{cor:BO}
Let $1 < p < n$, $0 \le q < p - 1$ and $0 < \gamma \le \infty$.
Set
\[
s = \frac{n(p - 1 + \gamma)}{n (p - 1 - q) + p(\gamma + q)}
\quad \text{and} \quad
t = \frac{p - 1 + \gamma}{p - 1 - q}.
\]
Let $\sigma = \theta \, dx$, where $\theta \neq 0$ is a nonnegative function in $L^{s, t}(\Omega)$.
Then there exists a minimal positive $(p, w)$-superharmonic solution $u$ to \eqref{eqn:p-laplace} such that
$\min\{ u, l \} \in W_{0}^{1, p}(\Omega)$ for all $l > 0$.
Moreover, $u \in L^{r, \rho}(\Omega)$ and
\begin{equation}\label{eqn:00_BO}
\| u \|_{L^{r, \rho}(\Omega)}
\le
C \| \theta \|_{L^{s, t}(\Omega)}^{\frac{1}{p - 1 - q}},
\end{equation}
where
\[
r = \frac{n (p - 1 + \gamma)}{n - p},
\quad
\rho = p - 1 + \gamma
\]
and $C$ is a positive constant depending only on $n$, $p$, $q$ and $\gamma$.
Assume also that $\gamma \ge 1$.
Then $u$ belongs to $W_{0}^{1, p}(\Omega)$
and satisfies \eqref{eqn:p-laplace} in the sense of weak solutions.
\end{corollary}
\begin{proof}
As in the proof of \cite[Corollary 3.6]{HARA2020111847},
using \eqref{eqn:wolff_UB} and the Havin-Maz'ya potential estimate (see \cite{MR727526}),
we find that
\[
\int_{\Omega} (\mathcal{W}_{p, 1} (\theta \, dx))^{ \frac{(\gamma + q)(p - 1)}{p - 1 - q} } \theta \, dx
\le
C \| \theta \|_{L^{s, t}(\Omega)}^{\frac{p - 1 + \gamma}{p - 1 - q}}.
\]
By Theorem \ref{thm:main_theorem}, there exists a minimal $p$-superharmonic solution $u$ such that
\begin{equation}\label{eqn:01_BO}
\mathcal{E}_{\gamma}( u^{q} \sigma )
=
\| u \|_{L^{\gamma + q}(\sigma)}^{\gamma + q}
\le
C \| \theta \|_{L^{s, t}(\Omega)}^{\frac{p - 1 + \gamma}{p - 1 - q}}.
\end{equation}
Furthermore, by Corollary \ref{cor:energy_esti}
and a sharp form of Sobolev inequality (see, e.g., \cite[p.234]{MR2777530}),
\begin{equation}\label{eqn:02_BO}
\| u \|_{L^{r, \rho}(\Omega)}^{p - 1 + \gamma}
=
\| v \|_{L^{p^{*}, p}(\Omega)}^{p}
\le
C \int_{\Omega} |\nabla v|^{p} \, dx
\le
C \mathcal{E}_{\gamma}( u^{q} \sigma ),
\end{equation}
where $v = u^{ \frac{p - 1 + \gamma}{p} }$.
Combining \eqref{eqn:01_BO} and \eqref{eqn:02_BO}, we obtain \eqref{eqn:00_BO}.
Since $\Omega$ is bounded, we have
\[
\begin{split}
\int_{\Omega} u^{q} \theta dx
& \le
\left(
\int_{\Omega}
u^{\gamma + q} \theta \,dx
\right)^{\frac{q}{\gamma + q}}
\left(
\int_{\Omega} \theta \, dx
\right)^{\frac{\gamma}{\gamma + q}}
\\
& \le
\left(
\int_{\Omega}
u^{\gamma + q} \theta \,dx
\right)^{\frac{q}{\gamma + q}}
\left(
C \| \theta \|_{L^{s, t}(\Omega)} |\Omega|^{\frac{s - 1}{s}}
\right)^{\frac{\gamma}{\gamma + q}} < \infty.
\end{split}
\]
Thus, $u^{q} \sigma = u^{q} \theta \, dx \in {S}^{0}$.
By Proposition \ref{prop:energy_class}, this implies that $\min\{ u, l \} \in W_{0}^{1, p}(\Omega)$ for all $l > 0$.
If $\gamma \ge 1$, then $u^{q} \sigma \in {S}^{0} \cap {S}^{\gamma} \subset {S}^{1}$, and hence $u \in W_{0}^{1, p}(\Omega)$.
\end{proof}
\begin{remark}
For $0 < \gamma < 1$,
using an interpolation argument (see, e.g., \cite[Lemma 4.2]{MR1354907}),
from \eqref{eqn:02_BO} and \eqref{eqn:grad_esti}, we can deduce a gradient estimate of $u$.
\end{remark}
\section{Quasilinear ODE with Hardy-type coefficients}\label{sec:example2}
Let us now consider the model ordinary differential equation
\begin{equation}\label{eqn:p-laplace_bdr}
\begin{cases}
- (w |u'|^{p - 2} u')' = \theta u^{q} \quad \text{in $(-1, 1)$,}
\\
u(-1) = u(1) = 0,
\end{cases}
\end{equation}
where
\begin{align*}
w(x) & = (1 - |x|)^{\beta}, \quad \beta \in (-1, p - 1),
\\
\theta(x) & = (1 - |x|)^{- \alpha}, \quad \alpha \in \mathbb{R},
\end{align*}
and $' = \frac{d}{dx}$.
The function $w$ can be regarded as a Muckenhoupt $A_{p}$-weight in $\mathbb{R}$.
Therefore it is also $p$-admissible (see \cite[Chapter 15]{MR2305115} \cite[Theorem 2]{MR2180887}).
Note that $\theta$ is not integrable on $(-1, 1)$ if $\alpha \ge 1$.
\begin{corollary}
Let $1 < p < \infty$ and $0 \le q < p - 1$.
Assume that $\alpha < p - \beta$.
Then there exists a bounded minimal positive weak solution $u \in H^{1, p}_{\mathrm{loc}}((-1, 1); w) \cap C([-1, 1])$ to \eqref{eqn:p-laplace_bdr}.
Moreover, there exists a positive finite energy weak solution $u \in H_{0}^{1, p}((-1, 1); w)$ to \eqref{eqn:p-laplace_bdr}
if and only if
\begin{equation}\label{eqn:bdr_singularity}
\alpha < 1 + (1 + q) \left( 1 - \frac{1}{p} \right) \left( 1 - \frac{\beta}{p - 1} \right).
\end{equation}
\end{corollary}
\begin{proof}
Let $v = (1 - |x|)^{A}$. Then
\[
- (w |v'|^{p - 2} v')'
=
c(p, \beta, A) (1 - |x|)^{(A - 1)(p - 1) + \beta - 1} + 2A^{p - 1} \delta_{0}
\]
in the sense of distribution, where
$c(p, \beta, A) = - A^{p - 1}\{ (A - 1)(p - 1) + \beta \}$
and
$\delta_{0}$ is the Dirac mass concentrated at $0$.
Hence taking $A \in (0, 1 - \frac{\beta}{p - 1})$ such that $A \le \frac{p - \alpha - \beta}{p - 1- q}$ and
choosing a large $C$, we can make $V(x) = C (1 - |x|)^{A}$ a bounded supersolution
to \eqref{eqn:p-laplace_bdr}.
Then Theorem \ref{thm:main_theorem_infty} gives a positive bounded weak solution $u \in H^{1, p}_{\mathrm{loc}}((-1, 1); w)$ to \eqref{eqn:p-laplace_bdr}.
The Sobolev embedding theorem provides $u \in C(-1, 1)$.
Furthermore, $u$ is continuous up to the boundary because $0 \le u(x) \le V(x)$ for all $x \in (-1, 1)$.
According to the Hardy-type inequality in \cite[Theorem 1.3.3]{MR2777530},
\eqref{eqn:bdr_singularity} is necessary and sufficient for the embedding
\[
\left( \int_{-1}^{1} |u|^{1 + q} \theta \, dx \right)^{\frac{1}{1 + q}}
\le
C \left( \int_{-1}^{1} |u'|^{p} w \, dx \right)^{\frac{1}{p}},
\quad
\forall u \in C_{c}^{\infty}(-1, 1).
\]
Thus Theorems \ref{thm:main_theorem_FE} and \ref{thm:embedding} give the desired assertion.
\end{proof}
\bibliographystyle{abbrv}
| 96d6ef9df6e196d616b2be8c0ce0e1497723b307 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\blfootnote{\IEEEauthorrefmark{1}Saeid Haghighatshoar is currently with the Swiss Center for Electronics and Microtechnology (CSEM), however his contribution to this work was made while he was with the CommIT group ([email protected]).}
Massive multiple-input multiple-output (MIMO) communication systems have recently received considerable attention by promising unprecedented data rates, network coverage and link reliability \cite{larsson2014massive,bjornson2016massive}. This technology makes use of a large number ($M\gg 1$) of antennas at the base station (BS), enabling the spatial multiplexing of several data streams to a number of users $K$ that share the same time-frequency transmission resource \cite{yang2013performance}. Massive MIMO can be implemented for both narrow-band and wide-band signaling and is compatible with new trends such as millimeter wave (mmWave) communications \cite{heath2016overview}.
In order to precode effectively the Downlink (DL) data streams to the users, massive MIMO requires some level of channel state information at the transmitter side (CSIT). This can be easily obtained in TDD systems owing to channel reciprocity and calibrated RF front-ends \cite{marzetta2010noncooperative,shepard2012argos}.
However, in the case where the computation of the precoder is too complex for real-time operations, or in the FDD case, where obtaining accurate and timely CSIT is very difficult,
several schemes have been proposed in order to exploit the statistical structure of the massive MIMO channels, and in particular to decompose them into ``virtual beam directions", such that
the precoding and channel estimation can be performed in a virtual beam domain of reduced dimension \cite{adhikary2013joint,haghighatshoar2016massive,haghighatshoar2018low,khalilsarai2018fdd}. In particular, in many typical propagation environments the number of scatterers is much less than the number of array elements, the BS is connected to a user through a narrow angular aperture. Therefore, by a careful choice of the beam-space, the BS can use a variety of sparse signal processing tools to efficiently perform its tasks \cite{haghighatshoar2017massive,khalilsarai2018achieve,nguyen2013compressive,wunder2018hierarchical}.
Defining the array response vector as ${{\mathbb f} a} (\hbox{\boldmath$\xi$}) = [ e^{j\tfrac{2\pi}{\lambda}\langle {{\mathbb f} r}_1 , \hbox{\boldmath$\xi$} \rangle},\ldots,e^{j\tfrac{2\pi}{\lambda}\langle {{\mathbb f} r}_M , \hbox{\boldmath$\xi$} \rangle}, ]^{\sf T}$, the MIMO channel covariance can be written as \cite{khalilsarai2018fdd}
\begin{equation}\label{eq:ch_cov_0}
\widetilde{\hbox{\boldmath$\Sigma$}} = {\mathbb E} \left[ {{\mathbb f} h} {{\mathbb f} h}^{{\sf H}} \right] = \int_{\hbox{\boldmath$\Xi$}} \gamma (\hbox{\boldmath$\xi$}) {{\mathbb f} a} (\hbox{\boldmath$\xi$}) {{\mathbb f} a} (\hbox{\boldmath$\xi$})^{{\sf H}} d\hbox{\boldmath$\xi$},
\end{equation}
where $\lambda$ is the wavelength, $\{{{\mathbb f} r}_m \}_{m=1}^M$ denotes antenna locations, $\hbox{\boldmath$\xi$} \in \hbox{\boldmath$\Xi$}$ is the angle of arrival (AoA) and $\gamma (\hbox{\boldmath$\xi$})$ is the angular power spread function.
The beam-space for a user is characterized by a set of vectors, called \textit{virtual beams} that span the column space of $\widetilde{\hbox{\boldmath$\Sigma$}}$. The ``best" choice for such vectors is given by the eigenvectors of $\widetilde{\hbox{\boldmath$\Sigma$}}$ denoted as columns of the matrix ${{\mathbb f} U} = [{{\mathbb f} u}_1,\ldots,{{\mathbb f} u}_M]$, which also comprise the basis for the Karhunen-L{\`o}eve (KL) expansion of ${{\mathbb f} h}$ as
$
{{\mathbb f} h} = \sum_{m=1}^{M} w_m {{\mathbb f} u}_m,
$
where $w_m \in {\mathbb C}, \, m=1,\ldots,M$ are uncorrelated random variables. The KL expansion provides the best $k$-term approximation for ${{\mathbb f} h}$ and has, exclusive to itself, the property of decorrelating ${{\mathbb f} h}$ into orthonormal vectors $\{ {{\mathbb f} u}_m \}_{m=1}^M$ with uncorrelated coefficients $\{w_m \}_{m=1}^M$ \cite{unser2014introduction}. For Gaussian channels, widely assumed in wireless multipath channel models, this property is even more pronounced as the coefficients become statistically independent. The variance of a coefficient $\sigma_m^2 ={\mathbb E} [ |w_m|^2 ] $ represents the channel ``power" along virtual beam ${{\mathbb f} u}_m$ and reveals the channel angular sparsity, such that if the power along a beam is less than a threshold ($\sigma_m^2 < \epsilon$), one can assume that the BS receives almost no energy from that beam.
While obtaining the set of virtual beams is easy for a single user, it is not straightforward when one deals with multiple users, since in general the users do not share the same set of eigenvectors. On the other hand, having a suitable set of common virtual beams (CVB) is highly desirable for DL channel training and multi-user precoding \cite{sayeed2013beamspace,khalilsarai2018fdd}. Most of the works in the literature \textit{assume} the existence of a common basis of orthogonal beam directions that (approximately) diagonalizes all user covariances and represents a sort of common eigenvectors set (CES) \cite{heath2016overview}. For the simplest case of a \textit{uniform linear array} (ULA), it is typically taken for granted that the CVB is given by a uniform sampling of the array response vector, which in turn is equivalent to the DFT basis. The implicit (and seldom mentioned) reason for this assumption comes from an asymptotic ($M\to \infty$) equivalence between Toeplitz and circulant matrices according to the Szeg{\"o} theorem (see \cite{adhikary2013joint} and references therein). Since the covariance of the channel associated with a ULA is Toeplitz, and since circulant matrices are diagonalized by the DFT matrix, it holds that DFT columns form an approximate set of eigenvectors for any ULA covariance. As this theoretic result holds for large ULAs, the choice of a suitable set of virtual beams for small or moderate-sized arrays is unclear. Besides, for array geometries other than the ULA, such as circular, planar and cylindrical arrays, no universal virtual beam design is known.
This paper provides a solution for the problem above. Given random channel realizations ${{\mathbb f} h}_k^{(1)},\, {{\mathbb f} h}_k^{(2)},\, \ldots, \, {{\mathbb f} h}_k^{(N)}$ of a set of users $k=1,\ldots,K$ and without knowing the array geometry we propose a rigorous method for obtaining an orthogonal CVB set. Note that this problem is more general compared to the case in which we have user covariances $\widetilde{\hbox{\boldmath$\Sigma$}}_k,\, k=1,\ldots, K$. Our method is based on the maximum-likelihood (ML) estimation of the \textit{postulated} CES given random channel realizations. We formulate the ML problem as an optimization over the unitary manifold and proposed a \textit{projected gradient descent} (PGD) method to solve it. We prove the convergence of this algorithm to a stationary point of the likelihood function with arbitrary initialization. We also show that, with jointly diagonalizable covariances, the CES coincides with the global maximizer of the likelihood function. Finally, we compare our method to the classic joint approximate diagonalization of eigenmatrices (JADE) algorithm \cite{cardoso1993blind} to show its superior performance.
\section{System Setup}
Consider the following scenario: we observe $N$ random realizations for each of the $K$ independent, stationary, zero-mean vector Gaussian processes as
\begin{equation}\label{eq:vector_observations}
{{\mathbb f} H}_k = \left[ {{\mathbb f} h}_k^{(1)},\ldots,{{\mathbb f} h}_k^{(N)} \right] \in {\mathbb C}^{M\times N},\, k=1,\ldots,K.
\end{equation}
The covariance matrix of process $k$ is denoted by $\widetilde{\hbox{\boldmath$\Sigma$}}_k = {\mathbb E} [{{\mathbb f} h}_k {{\mathbb f} h}_k^{{\sf H}} ]$. The random vector ${{\mathbb f} h}_k$ can represent, for instance, the channel vector of a user $k$ communicating with a base station (BS) equipped with $M$ antennas. The eigendecomposition of $\widetilde{\hbox{\boldmath$\Sigma$}}_k$ is given as
\begin{equation}\label{eq:eigdecomp_true}
\widetilde{\hbox{\boldmath$\Sigma$}}_k = {{\mathbb f} U}_k \widetilde{\hbox{\boldmath$\Lambda$}}_k {{\mathbb f} U}_k^{{\sf H}},
\end{equation}
where ${{\mathbb f} U}_k$ is the unitary matrix of eigenvectors (${{\mathbb f} U}_k^{{\sf H}} {{\mathbb f} U}_k=\mathbf{I}_M$) and $\widetilde{\hbox{\boldmath$\Lambda$}}_k$ is the diagonal matrix of eigenvalues.
We are interested in obtaining a (approximate) common eigenstructure\footnote{We use the terms ``common eigenstructure", ``common eigenvectors set", and ``common virtual beams set" interchangeably.} among all covariances $\{ \widetilde{\hbox{\boldmath$\Sigma$}}_k \}_{k=1}^K$ given random samples $\{{{\mathbb f} H}_k\}_{k=1}^K$. If the covariance matrices are jointly diagonalizable, i.e. if there exists a unitary matrix ${{\mathbb f} U}^c$ such that ${{\mathbb f} U}_1={{\mathbb f} U}_2=\ldots={{\mathbb f} U}_K={{\mathbb f} U}^c$, then it is desirable to obtain ${{\mathbb f} U}^c$ as the common eigenstructure. If the covariances are \textit{not} jointly diagonalizable, then we want to obtain a unitary matrix ${{\mathbb f} U}^\star$ as the common eigenstructure that best diagonalizes the covariances.
To have a systematic way of estimating the common eigenstructure, we first impose the joint diagonalizability criterion on the estimation model, in which each covariance is decomposed as
\begin{equation}\label{eq:decomp}
\hbox{\boldmath$\Sigma$}_k = {{\mathbb f} U} \hbox{\boldmath$\Lambda$}_k {{\mathbb f} U}^{{\sf H}},
\end{equation}
where ${{\mathbb f} U}=[{{\mathbb f} u}_1,\ldots,{{\mathbb f} u}_M]\in {\mathbb C}^{M\times M}$ is a unitary matrix (${{\mathbb f} U}^{{\sf H}} {{\mathbb f} U}=\mathbf{I}_M$) representing the to-be-estimated common eigenstructure (or CVB set) and, assuming non-singular covariances for simplicity, $\hbox{\boldmath$\Lambda$}_k = \text{diag}(\hbox{\boldmath$\lambda$}_k)$ is a diagonal matrix with positive diagonal elements given in the vector $\hbox{\boldmath$\lambda$}_k$ for $k=1,\ldots,K$. Note the difference between the true covariance $\widetilde{\hbox{\boldmath$\Sigma$}}_k$ in \eqref{eq:eigdecomp_true} and the hypothetical covariance $\hbox{\boldmath$\Sigma$}_k$ in \eqref{eq:decomp}. Then, given random samples as in \eqref{eq:vector_observations}, we solve a maximum-likelihood optimization to estimate ${{\mathbb f} U}$. This approach can be seen as a \textit{deliberately mismatched} ML estimation method: the covariances do \textit{not} generally share the same set of eigenvectors, but we impose this property by adopting the model in \eqref{eq:decomp}. Notice that by imposing such model mismatch, we look for the common unitary matrix ${{\mathbb f} U}$ that best fits the sample data ${{\mathbb f} H}_1,\ldots, {{\mathbb f} H}_K$, where ``best fit" is to be interpreted in the Maximum Likelihood sense.
Observing the random samples $\{ {{\mathbb f} H}_k \}_{k=1}^K$, one can write the likelihood function as a function of the hypothetical covariance matrices according to
\begin{equation}\label{eq:ML_func}
\begin{aligned}
& p\left(\{{{\mathbb f} H}_k \}_{k=1}^K |\{ \hbox{\boldmath$\Sigma$}_k\}_{k=1}^K \right) =\\
&\prod_{k=1}^{K} \frac{1}{(2\pi)^{\frac{M}{2}} {\mathop{\rm det}} (\hbox{\boldmath$\Sigma$}_k)^{\frac{N}{2}}}\exp \left( -\frac{1}{2N}\text{trace}({{\mathbb f} H}_k^{{\sf H}} \hbox{\boldmath$\Sigma$}_k^{-1} {{\mathbb f} H}_k)\right)=\\
& \prod_{k=1}^{K} \frac{1}{(2\pi)^{\frac{M}{2}} {\mathop{\rm det}} (\hbox{\boldmath$\Sigma$}_k)^{\frac{N}{2}}}\exp \left( -\frac{1}{2}\text{trace}( \hbox{\boldmath$\Sigma$}_k^{-1} \widehat{\hbox{\boldmath$\Sigma$}}_k )\right),
\end{aligned}
\end{equation}
where we have defined $\widehat{\hbox{\boldmath$\Sigma$}}_k:= \frac{1}{N} {{\mathbb f} H}_k {{\mathbb f} H}_k^{{\sf H}}$, as the sample covariance of the $k$-th process. Taking the $-\log(\cdot)$ of the likelihood function, scaling it, omitting constant terms, and replacing $\hbox{\boldmath$\Sigma$}_k$ with ${{\mathbb f} U} \hbox{\boldmath$\Lambda$}_k {{\mathbb f} U}^{{\sf H}}$ from \eqref{eq:decomp}, one can show that maximizing the likelihood function is equivalent to minimizing the following ML cost as a function of the parameters ${{\mathbb f} U}$ and $ \{ \hbox{\boldmath$\lambda$}_k \}_{k=1}^K$:
\begin{equation}\label{eq:cost_func}
\begin{aligned}
& {\cal C} \left({{\mathbb f} U} , \{ \hbox{\boldmath$\lambda$}_k \}_{k=1}^K \right) =\\
&\scalebox{0.92}{$\sum_{k=1}^{K} \log{\mathop{\rm det}}\, ({{\mathbb f} U} \text{diag}(\hbox{\boldmath$\lambda$}_k) {{\mathbb f} U}^{{\sf H}}) + \text{trace} \left( ({{\mathbb f} U} \text{diag}(\hbox{\boldmath$\lambda$}_k) {{\mathbb f} U}^{{\sf H}})^{-1} \widehat{\hbox{\boldmath$\Sigma$}}_k \right).$}
\end{aligned}
\end{equation}
Since
${\mathop{\rm det}}\, ({{\mathbb f} U} \text{diag}(\hbox{\boldmath$\lambda$}_k) {{\mathbb f} U}^{{\sf H}}) =\nolinebreak {\mathop{\rm det}}\, (\text{diag}(\hbox{\boldmath$\lambda$}_k) ),$
and \scalebox{0.95}{$\text{trace} \left( ({{\mathbb f} U} \text{diag}(\hbox{\boldmath$\lambda$}_k) {{\mathbb f} U}^{{\sf H}})^{-1} \widehat{\hbox{\boldmath$\Sigma$}}_k \right) = \nolinebreak \text{trace} \left( \text{diag}(\hbox{\boldmath$\lambda$}_k)^{-1} {{\mathbb f} U}^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k{{\mathbb f} U} \right),$} we have
\begin{equation}
{\cal C} \left({{\mathbb f} U} , \{ \hbox{\boldmath$\lambda$}_k \}_{k=1}^K \right) = \sum_{m,k} \log \hbox{\boldmath$\lambda$}_{k,m} + \frac{{{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m}{\hbox{\boldmath$\lambda$}_{k,m}},
\end{equation}
where $\hbox{\boldmath$\lambda$}_{k,m}> 0$ is the $m$-th element of $\hbox{\boldmath$\lambda$}_k$. The ML optimization problem then can be formulated as
\begin{equation}\label{eq:ML_formula_1}
\begin{aligned}
&\underset{\{ {{\mathbb f} u}_m \}_{m=1}^M ,\{\hbox{\boldmath$\lambda$}_k \}_{k=1}^K}{\text{minimize}} && \sum_{m,k} \log \hbox{\boldmath$\lambda$}_{k,m} + \frac{{{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m}{\hbox{\boldmath$\lambda$}_{k,m}}\\
&\hspace{6mm} \text{subject to} && {{\mathbb f} u}_m^{{\sf H}} {{\mathbb f} u}_n = \delta_{m,n}
\end{aligned}
\end{equation}
For simplicity we assume that all sample covariance matrices are non-singular, such that ${{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m>0$ for any ${{\mathbb f} u}_m$. As a result, it is easy to show that, for given $\{ {{\mathbb f} u}_m\}_{m=1}^M$, the function $g(x)=\log x + \frac{{{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m}{x},\, x>0$ achieves its minimum at $x={{\mathbb f} u}_m^{{\sf H}}\widehat{ \hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m$. Therefore, we take the minimization in \eqref{eq:ML_formula_1} first with respect to $\hbox{\boldmath$\lambda$}_{k,m}, \, k\in [K],\, m\in [M]$ (for an integer $n$, we define $[n]:=\{1,\ldots,n \}$) and transform \eqref{eq:ML_formula_1} to
\begin{equation}\label{eq:ML_formula_2}
\begin{aligned}
&\underset{\{ {{\mathbb f} u}_m \}_{m=1}^M }{\text{minimize}} && f({{\mathbb f} U})=\sum_{m,k} \log \left( {{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m\right) \\
& \text{subject to} && {{\mathbb f} u}_m^{{\sf H}} {{\mathbb f} u}_n = \delta_{m,n}
\end{aligned}\tag{$P_{ ML}$}
\end{equation}
This presents an optimization problem over the manifold of unitary matrices ${\cal U} = \{ {{\mathbb f} U} \in {\mathbb C}^{M\times M}:\, {{\mathbb f} U}^{{\sf H}} {{\mathbb f} U} =\mathbf{I}_M \}$. To solve \ref{eq:ML_formula_2}, we propose a gradient projection method and show that it converges to a stationary point of the cost function $f({{\mathbb f} U})$.
\subsection{Intermezzo: Jointly Diagonalizable Covariances and Global Optimality of the CES}
In the discussion above, we said that for jointly diagonalizable covariances, we with to obtain the CES ${{\mathbb f} U}^c$ as a result of our estimation method. In order to show that the ML problem \eqref{eq:ML_formula_2} gives a reasonable framework to satisfy this property, here we prove that for jointly diagonalizable covariances and by assuming sample covariances to have converged ($\widehat{ \hbox{\boldmath$\Sigma$}}_k = \hbox{\boldmath$\Sigma$}_k$), the CES ${{\mathbb f} U}^c$ is in fact a global minimizer of $P_{ML}$. Let the set of covariances $\hbox{\boldmath$\Sigma$}_k,~k\in [K]$ to be decomposed as
\begin{equation}\label{eq:commut_cov}
\hbox{\boldmath$\Sigma$}_k = {{\mathbb f} U}^c \hbox{\boldmath$\Lambda$}_k {{\mathbb f} U}^{c\, {{\sf H}}},
\end{equation}
where ${{\mathbb f} U}^c \in {\mathbb C}^{M\times M}$ denotes the CES. Assume $\widehat{ \hbox{\boldmath$\Sigma$}}_k = \hbox{\boldmath$\Sigma$}_k$, and consider the following definition.
\begin{definition}[Majorization]
For ${{\mathbb f} x} \in {\mathbb R}^{M}$, define ${{\mathbb f} x}^{\downarrow}$ as the vector containing the elements of ${{\mathbb f} x}$ in descending order. Let ${{\mathbb f} y} \in {\mathbb R}^M$ be another vector such that $\sum_{i=1}^{M}{{\mathbb f} x}_i =\sum_{i=1}^{M}{{\mathbb f} y}_i $. We say ${{\mathbb f} x}$ \textit{majorizes} ${{\mathbb f} y}$ (${{\mathbb f} x} \succ {{\mathbb f} y}$) iff
\[ \sum_{i=1}^{m}{{\mathbb f} x}^{\downarrow}_i \ge \sum_{i=1}^{m}{{\mathbb f} y}^{\downarrow}_i, \]
for all $m\in [M]$.
\end{definition}
We have the following theorem on global optimality of ${{\mathbb f} U}^c$.
\begin{theorem}\label{thm_global_opt}
Let $\hbox{\boldmath$\Sigma$}_k,\, k=1,\ldots, K$ be a set of jointly diagonalizable covariance matrices as in \eqref{eq:commut_cov} and consider the optimization problem
\begin{equation}\label{eq:opt_new}
\underset{{{\mathbb f} U}}{\text{minimize}}\, f({{\mathbb f} U})=\sum_{m,k} \log \left( {{\mathbb f} u}_m^{{\sf H}} \hbox{\boldmath$\Sigma$}_k {{\mathbb f} u}_m \right) ~\text{s.t.}~ {{\mathbb f} u}_m^{{\sf H}} {{\mathbb f} u}_n = \delta_{m,n}.
\end{equation}
Then ${{\mathbb f} U}^\star = {{\mathbb f} U}^c$ is a global solution of \eqref{eq:opt_new}.
\end{theorem}
\begin{proof}
For any unitary ${{\mathbb f} U}$, define the vector $\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U}) \in {\mathbb R}^M$ where $[\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U})]_{m} = {{\mathbb f} u}_m^{{\sf H}} \hbox{\boldmath$\Sigma$}_k {{\mathbb f} u}_m$. In particular $\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U}^c)$ is the vector of eigenvalues of $\hbox{\boldmath$\Sigma$}_k$. Using the properties of eigenvalue decomposition one can show $\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U}^c) \succ \hbox{\boldmath$\sigma$}_k ({{\mathbb f} U})$ for all ${{\mathbb f} U} \in {\cal U}$ and all $k\in [K]$.
In addition, the function $h({{\mathbb f} x}) = \sum_i \log ({{\mathbb f} x}_i)$ is Schur-concave \cite{peajcariaac1992convex} and therefore $\sum_m \log ([\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U}^c)]_m) \le \sum_m \log ([\hbox{\boldmath$\sigma$}_k ({{\mathbb f} U})]_m)$. Hence, $f ({{\mathbb f} U}^c)\le f ({{\mathbb f} U})$ for all ${{\mathbb f} U} \in {\cal U}$, proving ${{\mathbb f} U}^c$ to be the global minimizer of $f({{\mathbb f} U})$ over ${\cal U}$.
\end{proof}
\section{ML via Projected Gradient Descent}
The projected gradient descent method (PGD) is a well-known iterative optimization algorithm \cite{bertsekas2015convex}. Starting from an initial point ${{\mathbb f} U}^{(0)}$, this method consists of the two following steps per iteration:
\begin{equation}\label{eq:PGD_step1}
\widetilde{{{\mathbb f} U}}^{(t)} = {{\mathbb f} U}^{(t)} - \alpha_t \nabla f({{\mathbb f} U}^{(t)})\tag{Gradient Step}
\end{equation}
\begin{equation}\label{eq:PGD_step2}
{{\mathbb f} U}^{(t+1)} = {\cal P}_{{\cal U}} (\widetilde{{{\mathbb f} U}}^{(t)}) \tag{Projection Step}
\end{equation}
where $\alpha_t>0$ is a step size, $\nabla f({{\mathbb f} U}^{(t)})\in {\mathbb C}^{M\times M}$ is the gradient of $f$ at ${{\mathbb f} U}^{(t)}$ and ${\cal P}_{{\cal U}}:{\mathbb C}^{M\times M}\to {\cal U} $ denotes the orthogonal projection operator onto the set of unitary matrices. The explicit expression of this operator is given in \cite{manton2002optimization}; Nevertheless, we derive it here through the following lemma for the sake of completeness.
\begin{lemma}
Let ${{\mathbb f} V} \in {\mathbb C}^{M\times M}$ be a matrix with singular value decomposition ${{\mathbb f} V} = {{\mathbb f} S} {{\mathbb f} D} {{\mathbb f} T}^{{\sf H}}$ where ${{\mathbb f} S}$ and ${{\mathbb f} T}$ are unitary matrices of left and right eigenvectors and ${{\mathbb f} D}=\text{diag}({{\mathbb f} d})$ is non-negative diagonal. Then, the orthogonal projection of ${{\mathbb f} V}$ onto the set of unitary matrices is given by ${\cal P}_{{\cal U}}({{\mathbb f} V}) ={{\mathbb f} S} {{\mathbb f} T}^{{\sf H}}$.
\end{lemma}
\begin{proof}
The orthogonal projection of ${{\mathbb f} V}$ is given by the minimizer of $g({{\mathbb f} U})= \Vert {{\mathbb f} V} -{{\mathbb f} U}\Vert_F^2$ over the set of unitary matrices, where $\Vert\cdot \Vert_F^2$ denotes Frobenius norm and ${{\mathbb f} U}^{{\sf H}}{{\mathbb f} U}=\mathbf{I}_M$. We can write
\begin{equation}\label{eq:fro}
\begin{aligned}
g({{\mathbb f} U})= \Vert {{\mathbb f} V} -{{\mathbb f} U}\Vert_F^2 & = \Vert {{\mathbb f} U}\Vert_{\sf F}^2 + \Vert {{\mathbb f} V} \Vert_{\sf F}^2 - 2\field{R}\{ \langle{{\mathbb f} V} , {{\mathbb f} U} \rangle \}\\
& =M + \Vert {{\mathbb f} V} \Vert_{\sf F}^2 - 2\field{R}\{ \langle{{\mathbb f} V} , {{\mathbb f} U} \rangle \},
\end{aligned}
\end{equation}
where the inner product is defined as $\langle{{\mathbb f} V} , {{\mathbb f} U} \rangle = \text{trace} ({{\mathbb f} U}^{{\sf H}} {{\mathbb f} V}) $ and we used the fact that $ \text{trace} ({{\mathbb f} U} {{\mathbb f} U}^{{\sf H}})=\text{trace}(\mathbf{I}_M)=M$. According to Von Neumann's trace inequality we have
$| \langle{{\mathbb f} V} , {{\mathbb f} U} \rangle |=|\text{trace}\left( {{\mathbb f} U}^{{\sf H}}{{\mathbb f} V} \right)|\le \langle {{\mathbb f} s}_{{{\mathbb f} U}},{{\mathbb f} d} \rangle, $\cite{mirsky1975trace}, where ${{\mathbb f} s}$ denotes the singular values vector of ${{\mathbb f} U}$. In the special case where ${{\mathbb f} U}$ is unitary, we have ${{\mathbb f} s}_{{{\mathbb f} U}} = [1,\ldots,1]^{\sf T}$ and
$|\langle{{\mathbb f} V} , {{\mathbb f} U} \rangle |\le \langle {{\mathbb f} s}_{{{\mathbb f} U}},{{\mathbb f} d} \rangle = \sum_i {{\mathbb f} d}_i.$
Now, using \eqref{eq:fro} we have
$ g({{\mathbb f} U})\ge M +\Vert {{\mathbb f} V}\Vert_{\sf F}^2 -2\sum_i {{\mathbb f} d}_i,$
where the right hand side of the inequality is independent of ${{\mathbb f} U}$. We show that the lower bound on $g({{\mathbb f} U})$ is achieved by ${{\mathbb f} U}^\star = {{\mathbb f} S} {{\mathbb f} T}^{{\sf H}}$. This is seen by the fact that
\begin{equation}
\begin{aligned}
g({{\mathbb f} S} {{\mathbb f} T}^{{\sf H}} )&= M + \Vert {{\mathbb f} S} {{\mathbb f} D} {{\mathbb f} T}^{{\sf H}}\Vert_{\sf F}^2 - 2\field{R}\{ \langle{{\mathbb f} S} {{\mathbb f} D} {{\mathbb f} T}^{{\sf H}} , {{\mathbb f} S} {{\mathbb f} T}^{{\sf H}} \rangle \}\\
& = M + \Vert {{\mathbb f} S} {{\mathbb f} D} {{\mathbb f} T}^{{\sf H}}\Vert_{\sf F}^2 - 2\text{trace}({{\mathbb f} D})\\
&= M + \Vert{{\mathbb f} V} \Vert_{\sf F}^2 - 2\sum_i {{\mathbb f} d}_i,.
\end{aligned}
\end{equation}
This completes the proof.
\end{proof}
This lemma provides an explicit formula for the orthogonal projection of a matrix into the unitary set as
\begin{equation}\label{eq:orthogonal_proj}
{\cal P}_{{\cal U}}:{\mathbb C}^{M\times M}\to {\cal U},\, {{\mathbb f} V} = {{\mathbb f} S} {{\mathbb f} D} {{\mathbb f} T}^{{\sf H}}\to {{\mathbb f} S} {{\mathbb f} T}^{{\sf H}}.
\end{equation}
The following theorem presents the main result of this work.
\begin{theorem}\label{th:convergance}
Let ${{\mathbb f} U}^{(0)}\in {\cal U}$ be an initial point and consider the gradient projection update rule
\begin{equation}\label{eq:grad_proj}
{{\mathbb f} U}^{(t+1)} = {\cal P}_{{\cal U}} \left( {{\mathbb f} U}^{(t)} - \alpha_t \nabla f ({{\mathbb f} U}^{(t)}) \right),\,\, t=0,1,\ldots,
\end{equation}
with $\alpha_t \in (0,\frac{1}{L})$ for all $t$, where $L$ is the Lipschitz constant of $\nabla f({{\mathbb f} U})$. Then the sequence $\{ {{\mathbb f} U}^{(t)},\, t=0,1,\ldots \}$ converges to a stationary point of $f({{\mathbb f} U})$.
\end{theorem}
In order to prove Theorem \ref{th:convergance}, we need to first prove some useful properties of the ML optimization problem.
\subsection{Lipschitz Continuity of the Cost Gradient}
As a first step, we prove that the cost gradient $\nabla f({{\mathbb f} U})$ is Lipschitz continuous over ${\cal U}$. Note that the manifold ${\cal U}$ is a subset of the closed convex ball ${\cal B}$ (${\cal U} \subset {\cal B}$) where
${\cal B} = \{ {{\mathbb f} U}\, :\, \Vert {{\mathbb f} U} \Vert_{\sf F}\le \sqrt{M} \}.$
One can show that $f({{\mathbb f} U})$ has Lipschitz continuous gradient over ${\cal B}$, i.e. there exists a constant $L$, such that
$\Vert \nabla f({{\mathbb f} U} )- \nabla f({{\mathbb f} U}') \Vert_{\sf F} \le L \Vert {{\mathbb f} U} - {{\mathbb f} U}' \Vert_{\sf F},$
for all ${{\mathbb f} U},{{\mathbb f} U}' \in {\cal B}$. One way to prove this is by showing that the Hessian of $f({{\mathbb f} U})$ has bounded operator norm over ${\cal B}$.
Define the complex Hessian as the $M^2\times M^2$ square matrix $\nabla^2 f({{\mathbb f} U})$ whose elements are given as \cite{gunning2009analytic}
$
[\nabla^2 f({{\mathbb f} U})]_{m,n} = \frac{\partial^2\, f({{\mathbb f} U})}{\partial [\vec ({{\mathbb f} U})]_m \partial [\vec ({{\mathbb f} U})]_n^\ast },$
for $m,n\in [M^2]$, where
$\vec ({{\mathbb f} U}) = [{{\mathbb f} u}_1^{\sf T},\ldots,{{\mathbb f} u}_M^{\sf T}]^{\sf T}$ is the vectorized version of ${{\mathbb f} U}$. Simple calculations show that the Hessian is a block-diagonal matrix with its $m$-th diagonal block given as
\begin{align}\label{eq:Hess_f_0}
{{\mathbb f} D}_f^{(m)} = \sum_{k=1}^K
\frac{\widehat{\hbox{\boldmath$\Sigma$}}_k^{\sf T}}{{{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m} - \sum_{k=1}^K \frac{\left( \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m {{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k \right)^{\sf T}}{\left({{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m\right)^2},
\end{align}
\noindent so that we have $\nabla^2 f({{\mathbb f} U}) = \text{blkdiag}\left( {{\mathbb f} D}_f^{(1)},\ldots,{{\mathbb f} D}_f^{(M)} \right)$. Note that both terms on the right-hand-side of \eqref{eq:Hess_f_0} are PSD and therefore ${{\mathbb f} D}_f^{(m)}$ is the difference of two PSD matrices.
\begin{lemma}
The Hessian matrix $\nabla^2 f({{\mathbb f} U})$ is bounded in operator norm.
\end{lemma}
\begin{proof}
Define the operator norm of a matrix ${{\mathbb f} A}\in {\mathbb C}^{M^2\times M^2}$ as $\Vert {{\mathbb f} A} \Vert_{op} = \underset{\Vert{{\mathbb f} x} \Vert=1}{\sup} \frac{\Vert {{\mathbb f} A} {{\mathbb f} x} \Vert}{\Vert{{\mathbb f} x} \Vert}$, where $\Vert \cdot \Vert$ is the $\ell_2$ norm.
For a block-diagonal matrix such as $\nabla^2 f ({{\mathbb f} U}) $, the operator norm is equal to the maximum of the operator norms of each individual block, i.e. $
\Vert \nabla^2 f ({{\mathbb f} U}) \Vert_{op} = \underset{m}{\max}~ \Vert {{\mathbb f} D}_f^{(m)} \Vert_{op}$.
Using \eqref{eq:Hess_f_0}, the operator norm of block $m$ is bounded as
\scalebox{0.97}{$\Vert {{\mathbb f} D}_f^{(m)} \Vert_{op} \le \max \left\{ \sum_k \frac{\Vert \widehat{\hbox{\boldmath$\Sigma$}}_k \Vert_{op}}{{{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m}\,,\, \sum_k\frac{\Vert \widehat{\hbox{\boldmath$\Sigma$}}_k{{\mathbb f} u}_m \Vert^2}{({{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m)^2} \right\}$}
where we used the fact that ${{\mathbb f} D}_f^{(m)}$ is the difference of two PSD matrices and therefore its operator norm is bounded by the maximum of the operator norms of the two. Also, since the matrix $\widehat{\hbox{\boldmath$\Sigma$}}_k {{\mathbb f} u}_m {{\mathbb f} u}_m^{{\sf H}} \widehat{\hbox{\boldmath$\Sigma$}}_k$ is of rank one, its operator norm is equal to $\Vert \widehat{\hbox{\boldmath$\Sigma$}}_k{{\mathbb f} u}_m \Vert^2$. Finally, since sample covariances are assumed to be non-singular, both arguments in $\max\{\cdot\}$ are finite. Taking the maximum over all $M$ bounds also results in a finite value and the proof is complete.
\end{proof}
Next we show that the Lipschitz constant of $\nabla f ({{\mathbb f} U})$ is related to the operator norm of $\nabla^2 f({{\mathbb f} U})$.
\begin{lemma}
For a twice differentiable function $f({{\mathbb f} U})$ with Hessian bounded in operator norm as $\Vert \nabla^2 f({{\mathbb f} U}) \Vert_{op} \le L$ for all ${{\mathbb f} U}$, the gradient $\nabla f ({{\mathbb f} U})$ is Lipschitz continuous with Lipschitz constant $L$.
\end{lemma}
\begin{proof}
We show that the Lipschitz continuity condition $\Vert \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \Vert_{\sf F}\le L \Vert {{\mathbb f} U} -{{\mathbb f} U}' \Vert_{\sf F} $ holds for any ${{\mathbb f} U},\, {{\mathbb f} U}'$ via the following sequence of inequalities:
\begin{equation}\label{eq:Ineq_5}
\begin{aligned}
\Vert \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \Vert_{\sf F}
&\overset{(a)}{\le} \underset{\Vert {{\mathbb f} B} \Vert_{\sf F}=1}{\sup} \left| \left\langle {{\mathbb f} B} , \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \right\rangle \right|\\
& \hspace{-32mm}\scalebox{0.89}{$=\underset{\Vert {{\mathbb f} B} \Vert_{\sf F}=1}{\sup} \left| \left\langle \vec({{\mathbb f} B}) ,\int_{0}^1 \nabla^2 f \left({{\mathbb f} U}'+t({{\mathbb f} U} - {{\mathbb f} U}')\right)\text{vec}({{\mathbb f} U} - {{\mathbb f} U}')\, dt \right\rangle \right|$}
\\
& \hspace{-32mm}\scalebox{0.9}{$\le \underset{\Vert {{\mathbb f} B} \Vert_{\sf F}=1}{\sup}\int_{0}^1 \left| \left\langle \vec({{\mathbb f} B}) , \nabla^2 f \left({{\mathbb f} U}'+t({{\mathbb f} U} - {{\mathbb f} U}')\right)\text{vec}({{\mathbb f} U} - {{\mathbb f} U}')\, \right\rangle \right|dt$}\\
&
\hspace{-32mm}\scalebox{0.85}{$\overset{(b)}{\le}\underset{\Vert {{\mathbb f} B} \Vert_{\sf F}=1}{\sup}\underset{t\in [0,1]}{\sup} \Vert \nabla^2 f \left({{\mathbb f} U}'+t({{\mathbb f} U} - {{\mathbb f} U}')\right) \Vert_{op} \Vert \text{vec}({{\mathbb f} U} - {{\mathbb f} U}') \Vert \Vert \vec({{\mathbb f} B}) \Vert$}\\
&
\hspace{-32mm}\scalebox{0.81}{$=\underset{t\in [0,1]}{\sup} \Vert \nabla^2 f \left({{\mathbb f} U}'+t({{\mathbb f} U} - {{\mathbb f} U}')\right) \Vert_{op} \Vert \text{vec}({{\mathbb f} U} - {{\mathbb f} U}') \Vert $}\overset{(c)}{\le } L \Vert {{\mathbb f} U} - {{\mathbb f} U}' \Vert_{\sf F}. \\
\end{aligned}
\end{equation}
Inequality $(a)$ holds by taking into account the fact that for the particular value of ${{\mathbb f} B}$ as ${{\mathbb f} B}_0 = \frac{\nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') }{\Vert \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \Vert_{\sf F} }$ we have
$ \langle {{\mathbb f} B}_0 , \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \rangle = \Vert \nabla f({{\mathbb f} U})-\nabla f({{\mathbb f} U}') \Vert_{\sf F}.$ Inequality $(b)$ comes from an application of the Cauchy-Schwarz inequality and the definition of the operator norm. Finally, inequality $(c)$ holds due to the assumption on the boundedness of the Hessian operator norm, i.e. $\Vert\nabla^2 f({{\mathbb f} U})\Vert_{op}\le L$ for all ${{\mathbb f} U}$, and the proof is complete.
\end{proof}
The next lemma emerges as a consequence of the discussion above.
\begin{lemma}
For any pair of matrices ${{\mathbb f} U},{{\mathbb f} U}'\in {\cal B}$ we have
\begin{equation}\label{eq:descent}
f({{\mathbb f} U})\le f({{\mathbb f} U}') + \langle \nabla f({{\mathbb f} U}'),{{\mathbb f} U} - {{\mathbb f} U}'\rangle + \frac{L}{2}\Vert{{\mathbb f} U}-{{\mathbb f} U}' \Vert_{\sf F}^2,
\end{equation}
where $L$ is the gradient Lipschitz constant.
\end{lemma}
\begin{proof}
See \cite{bertsekas2015convex}, proposition 6.1.2.
\end{proof}
This lemma is used as a tool to prove the convergence of PGD to a stationary point, as outlined by Theorem \ref{th:convergance}.
\subsection{\textbf{Proof of Theorem \ref{th:convergance}}}
We start by replacing ${{\mathbb f} U}'$ with ${{\mathbb f} U}^{(t)}$ in \eqref{eq:descent} and defining the RHS of \eqref{eq:descent} as the proxy function
\begin{equation}\label{eq:proxy}
\scalebox{0.92}{$f_{\sf proxy}^{(t)} ({{\mathbb f} U})= f({{\mathbb f} U}^{(t)}) + \langle \nabla f({{\mathbb f} U}^{(t)}),{{\mathbb f} U} - {{\mathbb f} U}^{(t)}\rangle + \frac{L}{2}\Vert{{\mathbb f} U}-{{\mathbb f} U}^{(t)} \Vert_{\sf F}^2,$}
\end{equation}
at point ${{\mathbb f} U}^{(t)}$, such that we have
\begin{equation}\label{eq:proxy_bound}
f({{\mathbb f} U})\le f_{\sf proxy}^{(t)} ({{\mathbb f} U})
\end{equation}
for all ${{\mathbb f} U}\in {\cal U}$ and
$
f({{\mathbb f} U}^{(t)})= f_{\sf proxy}^{(t)} ({{\mathbb f} U}^{(t)}).
$
Now let us show that the point ${{\mathbb f} U}^{(t+1)}={\cal P}_{{\cal U}} \left( {{\mathbb f} U}^{(t)} - \alpha_t \nabla f ({{\mathbb f} U}^{(t)}) \right)$ is indeed a minimizer of $f_{\sf proxy}^{(t)}({{\mathbb f} U})$ over ${\cal U}$ with $\alpha_t=\frac{1}{L}$. To see this, note that we can expand $f_{\sf proxy}^{(t)}({{\mathbb f} U})$ as
$ f_{\sf proxy}^{(t)}({{\mathbb f} U}) = \langle \nabla f({{\mathbb f} U}^{(t)}),{{\mathbb f} U}\rangle-L \langle {{\mathbb f} U}^{(t)},{{\mathbb f} U} \rangle + {\sf const.}, $
for all unitary ${{\mathbb f} U}$. Then, minimizing $f_{\sf proxy}^{(t)}({{\mathbb f} U}) $ is equivalent to the maximization problem:
$\underset{{{\mathbb f} U}^{{\sf H}} {{\mathbb f} U} = \mathbf{I}_{M}}{\text{maximize}}\,\, \langle {{\mathbb f} U}^{(t)} -\frac{1}{L}\nabla f({{\mathbb f} U}^{(t)}),{{\mathbb f} U} \rangle$.
But the maximum of this objective is achieved at the point ${{\mathbb f} U}^\star = {{\mathbb f} S}_t {{\mathbb f} T}_t^{{\sf H}}$, where ${{\mathbb f} S}_t$ and ${{\mathbb f} T}_t$ are matrices of left and right eigenvectors in the SVD form $ {{\mathbb f} U}^{(t)} -\frac{1}{L}\nabla f({{\mathbb f} U}^{(t)}) = {{\mathbb f} S}_t{{\mathbb f} D}_t {{\mathbb f} T}_t^{{\sf H}}$. But this implies that ${{\mathbb f} U}^\star = {{\mathbb f} S}_t {{\mathbb f} T}_t^{{\sf H}} = {\cal P}_{{\cal U}} \left( {{\mathbb f} U}^{(t)} - \frac{1}{L}\nabla f ({{\mathbb f} U}^{(t)}) \right)={{\mathbb f} U}^{(t+1)}$ and ${{\mathbb f} U}^{(t+1)}$ is a minimizer of $f_{\sf proxy}^{(t)}({{\mathbb f} U})$. The chain of inequalities below immediately follows:
\begin{equation}\label{eq:ineqs}
\scalebox{0.92}{$f({{\mathbb f} U}^{(t+1)})\overset{(a)}{\le } f_{\sf proxy}^{(t)}({{\mathbb f} U}^{(t+1)})\overset{(b)}{\le } f_{\sf proxy}^{(t)}({{\mathbb f} U}^{(t)})\overset{(c)}{= } f({{\mathbb f} U}^{(t)}),$}
\end{equation}
where $(a)$ follows from \eqref{eq:proxy_bound}, $(b)$ follows from the fact that ${{\mathbb f} U}^{(t+1)}$ is aminimizer of $f_{\sf proxy}^{(t)}({{\mathbb f} U})$, and $(c)$ is a result of $f({{\mathbb f} U}^{(t)})= f_{\sf proxy}^{(t)} ({{\mathbb f} U}^{(t)})$. Therefore we have
$f({{\mathbb f} U}^{(t+1)}) \le f({{\mathbb f} U}^{(t)}),
$
for $t=0,1,\ldots$ and since $f({{\mathbb f} U})$ is bounded from below, the gradient projection sequence $\{ {{\mathbb f} U}^{(t)},\, t=0,1,\ldots \}$ converges to a stationary point of $f({{\mathbb f} U})$. $\hfill \QED$
\begin{figure}[t]
\centering
\includegraphics[ width=0.37\textwidth]{Cost_Global.pdf}
\caption{The average cost $f({{\mathbb f} U})$ as a function of the number of samples, for the solution of PGD ${{\mathbb f} U}^\star (N)$ and the global optimum ${{\mathbb f} U}^c$. We have set $M=16$ and $K=8$. }
\label{fig:cost_global}
\end{figure}
\section{Simulation Results}
To study the performance of our proposed method, in this section we provide some empirical results.
\subsection{Jointly Diagonalizable Covariances}
The case of jointly diagonalizable covariances is especially interesting, as we know by Theorem \ref{thm_global_opt} that the global optimum of the ML problem is given by the shared CES ${{\mathbb f} U}^c$ (see \eqref{eq:commut_cov}). In order to assess the performance of our method, we compare the ML cost as a function of the number of random samples per process $N$, to the cost at the global minimum. Consider a signal dimension (number of antennas) $M=16$ and a number of processes (number of users) $K=8$. We generate a random unitary matrix as the CES ${{\mathbb f} U}^c$ by calculating the eigenvectors of a random matrix of size $M\times M$ with i.i.d complex Gaussian elements. Also, for each process $k\in [K]$, we generate a random vector of eigenvalues $\hbox{\boldmath$\lambda$}_k$ with i.i.d, positive elements given as $\hbox{\boldmath$\lambda$}_{k,m}=|\rho_m|$ where $\rho_m \sim {\cal N} (0,1)$. Then we form the covariance matrix of user $k$ as $\widetilde{\hbox{\boldmath$\Sigma$}}={{\mathbb f} U}^c \text{diag}(\hbox{\boldmath$\lambda$}_k){{\mathbb f} U}^{c\, {{\sf H}}}$. We also normalize the covariances to have trace equal to one. This way we have randomly generated covariances with a shared CES. Now, having the covariances, we can generate random realizations for each process for different sample sizes $N$. We run a Monte Carlo simulation with $1000$ iterations, at each iteration generating covariances as stated above, then for each sample size $N$ we run our proposed PGD method which converges to a point ${{\mathbb f} U}^\star (N)$ (explicitly noting the dependence on $N$). Then, we average the cost function ${{\mathbb f} U}^\star (N)$ in \eqref{eq:ML_formula_2} over the Monte Carlo iterations for each value of $N$ and compare it to the average cost at the global optimum $f({{\mathbb f} U}^c )$. Note that the latter of course is not a function of $N$. The PGD method is initialized with a random unitary matrix. Theorem \ref{th:convergance} guarantees convergence to a stationary point when the step size is chosen as $\alpha_t\in (0,1/L)$. However, practically we can be more ambitious by choosing larger step sizes as $\alpha_t = \frac{\alpha_0}{t},~t=1,2,\ldots$ with $\alpha_0 =2$ and our simulation results show that even with this choice, PGD converges.
Fig. \ref{fig:cost_global} illustrates the result. We denote the solution of our method with ${{\mathbb f} U}^\star (N)$, to explicitly highlight its dependence on the number of samples $N$. The interesting fact about this result is that, as the number of samples gets larger, the PGD with random initialization always converges to the global optimum, as its cost value is the same as that in the optimal point ${{\mathbb f} U}^c$. This is an empirical evidence for the convergence of our proposed method to the global solution for jointly diagonalizable covariances.
\begin{figure}[t]
\centering
\includegraphics[ width=0.386\textwidth]{eta_func.pdf}
\caption{The average diagonalization metric as function of the sampling ratio $N/M$, for the solution of our proposed PGD method vs the JADE method. We set $M=16$ and $K=8$. }
\label{fig:eta_func}
\end{figure}
\subsection{Non-Jointly Diagonalizable Covariances}
As a different scenario, we
consider covariances that are not jointly diagonalizable. This is done by generating a different random unitary eigenvector matrix for each process separately as $\widetilde{\hbox{\boldmath$\Sigma$}}={{\mathbb f} U}_k \text{diag}(\hbox{\boldmath$\lambda$}_k){{\mathbb f} U}_k^{ {{\sf H}}}$. The eigenvalue vectors $\hbox{\boldmath$\lambda$}_k$ are generated as before and we normalize the covariances to have unit trace. So, in this case we do not have a CES and the PGD method yields a unitary matrix that approximately jointly diagonalizes the covariances. Since we do not have the global optimum in this case, we compare our method to the JADE algorithm, which is a classic Jacobian-based method for joint covariance diagonalization (we do not explain the details of this method here due to space limitations and refer the reader to \cite{cardoso1993blind} for a full account).
One way to measure the performance of joint diagonalization methods is by defining the metric $\eta : {\mathbb C}^{M\times M}\to [0,1]$:
\begin{equation}\label{eq:eta}
\eta ({{\mathbb f} U}) = 1-\frac{1}{K}\sum_{k=1}^K\frac{\Vert \text{diag}({{\mathbb f} U}^{{\sf H}} \hbox{\boldmath$\Sigma$}_k{{\mathbb f} U}) \Vert}{\Vert \hbox{\boldmath$\Sigma$}_k \Vert_{\sf F}},
\end{equation}
where $\text{diag}(\cdot)$ with a matrix argument as in \eqref{eq:eta} denotes the $M$-dim vector of the diagonal elements of its argument. The smaller the value of $\eta ({{\mathbb f} U})$ is, the better ${{\mathbb f} U}$ jointly diagonalizes the covariances. In the extreme case, If ${{\mathbb f} U}$ diagonalizes all covariance matrices, we have $\eta ({{\mathbb f} U})=0$.
The joint diagonalization metric is empirically averaged over $1000$ Monte Carlo simulations and for different sample sizes for the solutions of our proposed PGD method and the JADE method. Fig. \ref{fig:eta_func} illustrates the results. It clearly shows that for the ranges of sample sizes considered here, the proposed PGD method outperforms the classic JADE method, yielding smaller values of the diagonalization metric on average, and hence achieving a better joint diagonalization of the covariances.
\subsection{CES for ULA: PGD vs the Fourier Basis}
For a ULA it is usually taken for granted that the CES is given by the Fourier basis vectors. While this is true in an asymptotic sense thanks to the Szeg{\"o} theorem, it does not hold for small to moderate array sizes. We conclude our simulations by showing that, in fact the unitary basis produced by our proposed method better diagonalizes ULA covariances, compared to the DFT basis.
\begin{figure}[t]
\centering
\includegraphics[ width=0.4\textwidth]{eta_U_vs_Fourier.pdf}
\caption{The average diagonalization metric as function of the number of antennas $M$, for the solution of our proposed PGD method vs the Fourier basis. Here we have set $K=5$. }
\label{fig:U_vs_F}
\end{figure}
We consider $K=5$ $M$-dimensional ULA covariances, generated randomly according to \eqref{eq:ch_cov_0} (see \cite{khalilsarai2018fdd} on randomly generating ULA covariances). Since each covariance is associated with a user that occupies a limited angular range as seen from the BS, we generate the covariances such that each of them has an ``effective" rank of $r_k=\text{effrank}(\widetilde{\hbox{\boldmath$\Sigma$}}_k)=\lceil\frac{M}{2} \rceil$. The effective rank is equivalent to the number of significant covariance eigenvalues as well as channel angular sparsity.
We plot the expected joint diagonalization metric ${\mathbb E} [ \eta (\cdot) ]$ as a function of the number of antennas for the unitary matrix yielded by our method as well as for the Fourier basis ${{\mathbb f} F}$ where $[{{\mathbb f} F}]_{m,n}=\frac{1}{\sqrt{M}}e^{j2\pi \frac{(m-1)(n-1)}{M}},\, m,n\in [M]$. We assume that the sample covariance has converged, i.e. $\widehat{ \hbox{\boldmath$\Sigma$}}_k = \widehat{ \hbox{\boldmath$\Sigma$}}_k $ for all $k$. The expectation ${\mathbb E} [ \eta (\cdot) ]$ is taken over random covariance realizations and is calculated empirically over 100 Monte-Carlo loops. Fig. \ref{fig:U_vs_F} illustrates the result. As we can see, the basis given by the PGD method achieves better diagonalization (smaller $\eta$ values) than the Fourier basis, which shows that using our method is in fact preferable even for the diagonalization of ULA covariances. Also, as $M$ increases, the two bases have closer diagonalization metrics since we are approaching the asymptotic regime in which the Fourier basis approximately diagonalizes the Toeplitz ULA covariances.
\section{Conclusion}
We presented a framework for estimating the common eigenstructure for a set of covariance matrices. Our approach was based on maximizing the likelihood function of the postulated common eigenvectors set, given random vector realizations. We proposed the PGD method and proved its convergence to a stationary point of the likelihood function. Our empirical results illustrated that the proposed method converges to the global optimum when the covariances indeed share a common eigenvectors set. It outperforms the classic JADE method in all ranges of sample sizes. It also performs better than the Fourier basis in diagonalizing covariances generated by a ULA geometry in a multi-user MIMO setup.
{\small
\bibliographystyle{IEEEtran}
\subsection*{}%
\hbox{\hspace{.05\hsize}\defbox{\medskip#1\bigskip}}%
\subsection*{}}
\def{}
\def{\hbox{\it\tiny T}}{{\hbox{\it\tiny T}}}
\def{\text{diag}}{{\text{diag}}}
\def\phantom{-}{\phantom{-}}
\def\buildrel{\rm w}\over\longrightarrow{\buildrel{\rm w}\over\longrightarrow}
\def\buildrel{\rm v}\over\longrightarrow{\buildrel{\rm v}\over\longrightarrow}
\def\tvlimit#1{\buildrel{\scriptscriptstyle #1}\over\Longrightarrow}
\def\buildrel{\rm d}\over\longrightarrow{\buildrel{\rm d}\over\longrightarrow}
\def\goes#1{\buildrel{#1}\over\longrightarrow}
\def\wiggles#1{\buildrel{#1}\over\leadsto}
\def\mathsf{tr}{\mathsf{tr}}
\def{\rm rank\,}{{\rm rank\,}}
\def\hbox{\rm deg\thinspace}{\hbox{\rm deg\thinspace}}
\def{\rm sign}{{\rm sign}}
\def{\rm supp\,}{{\rm supp\,}}
\newsavebox{\junk}
\savebox{\junk}[1.6mm]{\hbox{$|\!|\!|$}}
\def{\usebox{\junk}}{{\usebox{\junk}}}
\def{\mathop{\rm det}}{{\mathop{\rm det}}}
\def\mathop{\rm lim\ sup}{\mathop{\rm lim\ sup}}
\def\mathop{\rm lim\ inf}{\mathop{\rm lim\ inf}}
\def\mathop{\rm arg\, min}{\mathop{\rm arg\, min}}
\def\mathop{\rm arg\, max}{\mathop{\rm arg\, max}}
\def\field{A}{{\sf A}}
\def{\sf K}{{\sf K}}
\def{\sf U}{{\sf U}}
\def{\sf V}{{\sf V}}
\def{\sf W}{{\sf W}}
\def{\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX}{{\sf X}}
\def{\mathchoice{\sf Y}\sfY{\hbox{\scriptsize\sf Y}}\smallsfY}{{\sf Y}}
\def{\sf Z}{{\sf Z}}
\def{\mathbb y}{{\cal B}({\mathchoice{\sf Y}\sfY{\hbox{\scriptsize\sf Y}}\smallsfY})}
\def\tinyplus{\vbox{\hrule width 3pt depth -1.5pt height 1.9pt
\vskip -1.5pt
\hbox{\hskip 1.3pt\vrule height 3pt depth 0pt width .4pt}}}
\def{\sf X^{\rm z}}{{\sf X^{\rm z}}}
\def{\cal B}(\state^\hbox{\rm\tiny z}){{\cal B}({\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX}^\hbox{\rm\tiny z})}
\def{\sf X^{{\rm z}_{ \tinyplus}}}{{\sf X^{{\rm z}_{ \tinyplus}}}}
\def({\sf X},{\cal B}({{\mathbb f} X}),{{\mathbb f} m}){({\sf X},{\cal B}({{\mathbb f} X}),{{\mathbb f} m})}
\def{\mathbb x}{{{\cal B}({\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX})}}
\def{\cal B}(\state^{{\rm z}_{\tinyplus}} )} \def\bxplus{{{\cal B}^+(\state)}{{\cal B}({\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX}^{{\rm z}_{\tinyplus}} )} \def\bxplus{{{\cal B}^+({\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX})}}
\newcommand{\field}[1]{\mathbb{#1}}
\def\field{R}_+{\field{R}_+}
\def\field{R}{\field{R}}
\def\field{A}{\field{A}}
\def\mbox{\rm{\large{1}}}{\mbox{\rm{\large{1}}}}
\def\mbox{\rm{\large{0}}}{\mbox{\rm{\large{0}}}}
\def\field{I}{\field{I}}
\def\field{T}{\field{T}}
\def\field{Z}{\field{Z}}
\def\field{Z}{\field{Z}}
\def\field{Z}_+{\field{Z}_+}
\def\field{Q}{\field{Q}}
\def\field{C}{\field{C}}
\def{\check{{\cal F}}}{{\check{{\cal F}}}}
\def{\check{{\mathchoice{\bfatom}{\bfatom}{\alpha}{\alpha}}}}{{\check{{\mathchoice{\bfatom}{\bfatom}{\alpha}{\alpha}}}}}
\def\check{{\sf E}}{\check{{\sf E}}}
\def\check{\Phi}{\check{\Phi}}
\def\check{\bfmath{\Phi}}{\check{\bfmath{\Phi}}}
\def\check{{\sf P}}{\check{{\sf P}}}
\def\cProb\!{\check{{\sf P}}\!}
\def\check{\Phi}{\check{\Phi}}
\def\check{\pi}{\check{\pi}}
\def\check{A}{\check{A}}
\def{\check{G}}{{\check{G}}}
\def{\check{P}}{{\check{P}}}
\def{\cal B}^+({\check{\state}}){{\cal B}^+({\check{\state}})}
\def{\check{\state}}{{\check{{\mathchoice{\sf X}\sfX{\hbox{\scriptsize\sf X}}\smallsfX}}}}
\def{\cal B}(\cstate){{\cal B}({\check{\state}})}
\def\bfmath{\Phi}^{\Lambda}{\bfmath{\Phi}^{\Lambda}}
\def{\mathbb A}{{\mathbb A}}
\def{\mathbb B}{{\mathbb B}}
\def{\mathbb C}{{\mathbb C}}
\def{\mathbb D}{{\mathbb D}}
\def{\mathbb E}{{\mathbb E}}
\def{\mathbb F}{{\mathbb F}}
\def{\mathbb G}{{\mathbb G}}
\def{\mathbb H}{{\mathbb H}}
\def{\mathbb I}{{\mathbb I}}
\def{\mathbb J}{{\mathbb J}}
\def{\mathbb K}{{\mathbb K}}
\def{\mathbb L}{{\mathbb L}}
\def{\mathbb M}{{\mathbb M}}
\def{\mathbb N}{{\mathbb N}}
\def{\mathbb O}{{\mathbb O}}
\def{\mathbb P}{{\mathbb P}}
\def{\mathbb Q}{{\mathbb Q}}
\def{\mathbb R}{{\mathbb R}}
\def{\mathbb S}{{\mathbb S}}
\def{\mathbb T}{{\mathbb T}}
\def{\mathbb U}{{\mathbb U}}
\def{\mathbb V}{{\mathbb V}}
\def{\mathbb W}{{\mathbb W}}
\def{\mathbb X}{{\mathbb X}}
\def{\mathbb Y}{{\mathbb Y}}
\def{\mathbb Z}{{\mathbb Z}}
\def{\mathbb a}{{\mathbb a}}
\def{\mathbb b}{{\mathbb b}}
\def{\mathbb c}{{\mathbb c}}
\def{\mathbb d}{{\mathbb d}}
\def{\mathbb e}{{\mathbb e}}
\def{\mathbb f}{{\mathbb f}}
\def{\mathbb g}{{\mathbb g}}
\def{\mathbb h}{{\mathbb h}}
\def{\mathbb i}{{\mathbb i}}
\def{\mathbb j}{{\mathbb j}}
\def{\mathbb k}{{\mathbb k}}
\def{\mathbb l}{{\mathbb l}}
\def{\mathbb m}{{\mathbb m}}
\def{\mathbb n}{{\mathbb n}}
\def{\mathbb o}{{\mathbb o}}
\def{\mathbb p}{{\mathbb p}}
\def{\mathbb q}{{\mathbb q}}
\def{\mathbb r}{{\mathbb r}}
\def{\mathbb s}{{\mathbb s}}
\def{\mathbb t}{{\mathbb t}}
\def{\mathbb u}{{\mathbb u}}
\def{\mathbb v}{{\mathbb v}}
\def{\mathbb w}{{\mathbb w}}
\def{\mathbb x}{{\mathbb x}}
\def{\mathbb y}{{\mathbb y}}
\def{\mathbb z}{{\mathbb z}}
\def{\bf A}{{{\mathbb f} A}}
\def{\bf B}{{{\mathbb f} B}}
\def{\bf C}{{{\mathbb f} C}}
\def{\bf D}{{{\mathbb f} D}}
\def{\bf E}{{{\mathbb f} E}}
\def{\bf F}{{{\mathbb f} F}}
\def{\bf G}{{{\mathbb f} G}}
\def{\bf H}{{{\mathbb f} H}}
\def{\bf I}{{{\mathbb f} I}}
\def{\bf J}{{{\mathbb f} J}}
\def{\bf K}{{{\mathbb f} K}}
\def{\bf L}{{{\mathbb f} L}}
\def{\bf M}{{{\mathbb f} M}}
\def{\bf N}{{{\mathbb f} N}}
\def{\bf O}{{{\mathbb f} O}}
\def{\bf P}{{{\mathbb f} P}}
\def{\bf Q}{{{\mathbb f} Q}}
\def{\bf R}{{{\mathbb f} R}}
\def{\bf S}{{{\mathbb f} S}}
\def{\bf T}{{{\mathbb f} T}}
\def{\bf U}{{{\mathbb f} U}}
\def{\bf V}{{{\mathbb f} V}}
\def{\bf W}{{{\mathbb f} W}}
\def{\bf X}{{{\mathbb f} X}}
\def{\bf Y}{{{\mathbb f} Y}}
\def{\bf Z}{{{\mathbb f} Z}}
\def{\bf a}{{{\mathbb f} a}}
\def{\bf b}{{{\mathbb f} b}}
\def{\bf c}{{{\mathbb f} c}}
\def{\bf d}{{{\mathbb f} d}}
\def{\bf e}{{{\mathbb f} e}}
\def{\bf f}{{{\mathbb f} f}}
\def{\bf g}{{{\mathbb f} g}}
\def{\bf h}{{{\mathbb f} h}}
\def{\bf i}{{{\mathbb f} i}}
\def{\bf j}{{{\mathbb f} j}}
\def{\bf k}{{{\mathbb f} k}}
\def{\bf l}{{{\mathbb f} l}}
\def{\bf m}{{{\mathbb f} m}}
\def{\bf n}{{{\mathbb f} n}}
\def{\bf o}{{{\mathbb f} o}}
\def{\bf p}{{{\mathbb f} p}}
\def{\bf q}{{{\mathbb f} q}}
\def{\bf r}{{{\mathbb f} r}}
\def{\bf s}{{{\mathbb f} s}}
\def{\bf t}{{{\mathbb f} t}}
\def{\bf u}{{{\mathbb f} u}}
\def{\bf v}{{{\mathbb f} v}}
\def{\bf w}{{{\mathbb f} w}}
\def{\bf x}{{{\mathbb f} x}}
\def{\bf y}{{{\mathbb f} y}}
\def{\bf z}{{{\mathbb f} z}}
\def{\mathfrak{A}}{{\mathfrak{A}}}
\def{\mathfrak{B}}{{\mathfrak{B}}}
\def{\mathfrak{C}}{{\mathfrak{C}}}
\def{\mathfrak{D}}{{\mathfrak{D}}}
\def{\mathfrak{E}}{{\mathfrak{E}}}
\def{\mathfrak{F}}{{\mathfrak{F}}}
\def{\mathfrak{G}}{{\mathfrak{G}}}
\def{\mathfrak{H}}{{\mathfrak{H}}}
\def{\mathfrak{I}}{{\mathfrak{I}}}
\def{\mathfrak{J}}{{\mathfrak{J}}}
\def{\mathfrak{K}}{{\mathfrak{K}}}
\def{\mathfrak{L}}{{\mathfrak{L}}}
\def{\mathfrak{M}}{{\mathfrak{M}}}
\def{\mathfrak{N}}{{\mathfrak{N}}}
\def{\mathfrak{O}}{{\mathfrak{O}}}
\def{\mathfrak{P}}{{\mathfrak{P}}}
\def{\mathfrak{Q}}{{\mathfrak{Q}}}
\def{\mathfrak{R}}{{\mathfrak{R}}}
\def{\mathfrak{S}}{{\mathfrak{S}}}
\def{\mathfrak{T}}{{\mathfrak{T}}}
\def{\mathfrak{U}}{{\mathfrak{U}}}
\def{\mathfrak{V}}{{\mathfrak{V}}}
\def{\mathfrak{W}}{{\mathfrak{W}}}
\def{\mathfrak{X}}{{\mathfrak{X}}}
\def{\mathfrak{Y}}{{\mathfrak{Y}}}
\def{\mathfrak{Z}}{{\mathfrak{Z}}}
\def{\mathfrak a}{{\mathfrak a}}
\def{\mathfrak b}{{\mathfrak b}}
\def{\mathfrak c}{{\mathfrak c}}
\def{\mathfrak d}{{\mathfrak d}}
\def{\mathfrak e}{{\mathfrak e}}
\def{\mathfrak f}{{\mathfrak f}}
\def{\mathfrak g}{{\mathfrak g}}
\def{\mathfrak h}{{\mathfrak h}}
\def{\mathfrak i}{{\mathfrak i}}
\def{\mathfrak j}{{\mathfrak j}}
\def{\mathfrak k}{{\mathfrak k}}
\def{\mathfrak l}{{\mathfrak l}}
\def{\mathfrak m}{{\mathfrak m}}
\def{\mathfrak n}{{\mathfrak n}}
\def{\mathfrak o}{{\mathfrak o}}
\def{\mathfrak p}{{\mathfrak p}}
\def{\mathfrak q}{{\mathfrak q}}
\def{\mathfrak r}{{\mathfrak r}}
\def{\mathfrak s}{{\mathfrak s}}
\def{\mathfrak t}{{\mathfrak t}}
\def{\mathfrak u}{{\mathfrak u}}
\def{\mathfrak v}{{\mathfrak v}}
\def{\mathfrak w}{{\mathfrak w}}
\def{\mathfrak x}{{\mathfrak x}}
\def{\mathfrak y}{{\mathfrak y}}
\def{\mathfrak z}{{\mathfrak z}}
\def{\mathscr{A}}{{\mathscr{A}}}
\def{\mathscr{B}}{{\mathscr{B}}}
\def{\mathscr{C}}{{\mathscr{C}}}
\def{\mathscr{D}}{{\mathscr{D}}}
\def{\mathscr{E}}{{\mathscr{E}}}
\def{\mathscr{F}}{{\mathscr{F}}}
\def{\mathscr{G}}{{\mathscr{G}}}
\def{\mathscr{H}}{{\mathscr{H}}}
\def{\mathscr{I}}{{\mathscr{I}}}
\def{\mathscr{J}}{{\mathscr{J}}}
\def{\mathscr{K}}{{\mathscr{K}}}
\def{\mathscr{L}}{{\mathscr{L}}}
\def{\mathscr{M}}{{\mathscr{M}}}
\def{\mathscr{N}}{{\mathscr{N}}}
\def{\mathscr{O}}{{\mathscr{O}}}
\def{\mathscr{P}}{{\mathscr{P}}}
\def{\mathscr{Q}}{{\mathscr{Q}}}
\def{\mathscr{R}}{{\mathscr{R}}}
\def{\mathscr{S}}{{\mathscr{S}}}
\def{\mathscr{T}}{{\mathscr{T}}}
\def{\mathscr{U}}{{\mathscr{U}}}
\def{\mathscr{V}}{{\mathscr{V}}}
\def{\mathscr{W}}{{\mathscr{W}}}
\def{\mathscr{X}}{{\mathscr{X}}}
\def{\mathscr{Y}}{{\mathscr{Y}}}
\def{\mathscr{Z}}{{\mathscr{Z}}}
\def{\mathscr a}{{\mathscr a}}
\def{\mathscr b}{{\mathscr b}}
\def{\mathscr c}{{\mathscr c}}
\def{\mathscr d}{{\mathscr d}}
\def{\mathscr e}{{\mathscr e}}
\def{\mathscr f}{{\mathscr f}}
\def{\mathscr g}{{\mathscr g}}
\def{\mathscr h}{{\mathscr h}}
\def{\mathscr i}{{\mathscr i}}
\def{\mathscr j}{{\mathscr j}}
\def{\mathscr k}{{\mathscr k}}
\def{\mathscr l}{{\mathscr l}}
\def{\mathscr m}{{\mathscr m}}
\def{\mathscr n}{{\mathscr n}}
\def{\mathscr o}{{\mathscr o}}
\def{\mathscr p}{{\mathscr p}}
\def{\mathscr q}{{\mathscr q}}
\def{\mathscr r}{{\mathscr r}}
\def{\mathscr s}{{\mathscr s}}
\def{\mathscr t}{{\mathscr t}}
\def{\mathscr u}{{\mathscr u}}
\def{\mathscr v}{{\mathscr v}}
\def{\mathscr w}{{\mathscr w}}
\def{\mathscr x}{{\mathscr x}}
\def{\mathscr y}{{\mathscr y}}
\def{\mathscr z}{{\mathscr z}}
\def{\mathtt{A}}{{\mathtt{A}}}
\def{\mathtt{B}}{{\mathtt{B}}}
\def{\mathtt{C}}{{\mathtt{C}}}
\def{\mathtt{D}}{{\mathtt{D}}}
\def{\mathtt{E}}{{\mathtt{E}}}
\def{\mathtt{F}}{{\mathtt{F}}}
\def{\mathtt{G}}{{\mathtt{G}}}
\def{\mathtt{H}}{{\mathtt{H}}}
\def{\mathtt{I}}{{\mathtt{I}}}
\def{\mathtt{J}}{{\mathtt{J}}}
\def{\mathtt{K}}{{\mathtt{K}}}
\def{\mathtt{L}}{{\mathtt{L}}}
\def{\mathtt{M}}{{\mathtt{M}}}
\def{\mathtt{N}}{{\mathtt{N}}}
\def{\mathtt{O}}{{\mathtt{O}}}
\def{\mathtt{P}}{{\mathtt{P}}}
\def{\mathtt{Q}}{{\mathtt{Q}}}
\def{\mathtt{R}}{{\mathtt{R}}}
\def{\mathtt{S}}{{\mathtt{S}}}
\def{\mathtt{T}}{{\mathtt{T}}}
\def{\mathtt{U}}{{\mathtt{U}}}
\def{\mathtt{V}}{{\mathtt{V}}}
\def{\mathtt{W}}{{\mathtt{W}}}
\def{\mathtt{X}}{{\mathtt{X}}}
\def{\mathtt{Y}}{{\mathtt{Y}}}
\def{\mathtt{Z}}{{\mathtt{Z}}}
\def{\mathtt a}{{\mathtt a}}
\def{\mathtt b}{{\mathtt b}}
\def{\mathtt c}{{\mathtt c}}
\def{\mathtt d}{{\mathtt d}}
\def{\mathtt e}{{\mathtt e}}
\def{\mathtt f}{{\mathtt f}}
\def{\mathtt g}{{\mathtt g}}
\def{\mathtt h}{{\mathtt h}}
\def{\mathtt i}{{\mathtt i}}
\def{\mathtt j}{{\mathtt j}}
\def{\mathtt k}{{\mathtt k}}
\def{\mathtt l}{{\mathtt l}}
\def{\mathtt m}{{\mathtt m}}
\def{\mathtt n}{{\mathtt n}}
\def{\mathtt o}{{\mathtt o}}
\def{\mathtt p}{{\mathtt p}}
\def{\mathtt q}{{\mathtt q}}
\def{\mathtt r}{{\mathtt r}}
\def{\mathtt s}{{\mathtt s}}
\def{\mathtt t}{{\mathtt t}}
\def{\mathtt u}{{\mathtt u}}
\def{\mathtt v}{{\mathtt v}}
\def{\mathtt w}{{\mathtt w}}
\def{\mathtt x}{{\mathtt x}}
\def{\mathtt y}{{\mathtt y}}
\def{\mathtt z}{{\mathtt z}}
\def{\sf A}{{\sf A}}
\def{\sf B}{{\sf B}}
\def{\sf C}{{\sf C}}
\def{\sf D}{{\sf D}}
\def{\sf E}{{\sf E}}
\def{\sf F}{{\sf F}}
\def{\sf G}{{\sf G}}
\def{\sf H}{{\sf H}}
\def{\sf I}{{\sf I}}
\def{\sf J}{{\sf J}}
\def{\sf K}{{\sf K}}
\def{\sf L}{{\sf L}}
\def{\sf M}{{\sf M}}
\def{\sf N}{{\sf N}}
\def{\sf O}{{\sf O}}
\def{\sf P}{{\sf P}}
\def{\sf Q}{{\sf Q}}
\def{\sf R}{{\sf R}}
\def{\sf S}{{\sf S}}
\def{\sf T}{{\sf T}}
\def{\sf U}{{\sf U}}
\def{\sf V}{{\sf V}}
\def{\sf W}{{\sf W}}
\def{\sf X}{{\sf X}}
\def{\sf Y}{{\sf Y}}
\def{\sf Z}{{\sf Z}}
\def{\sf a}{{\sf a}}
\def{\sf b}{{\sf b}}
\def{\sf c}{{\sf c}}
\def{\sf d}{{\sf d}}
\def{\sf e}{{\sf e}}
\def{\sf f}{{\sf f}}
\def{\sf g}{{\sf g}}
\def{\sf h}{{\sf h}}
\def{\sf i}{{\sf i}}
\def{\sf j}{{\sf j}}
\def{\sf k}{{\sf k}}
\def{\sf l}{{\sf l}}
\def{\sf m}{{\sf m}}
\def{\sf n}{{\sf n}}
\def{\sf o}{{\sf o}}
\def{\sf p}{{\sf p}}
\def{\sf q}{{\sf q}}
\def{\sf r}{{\sf r}}
\def{\sf s}{{\sf s}}
\def{\sf t}{{\sf t}}
\def{\sf u}{{\sf u}}
\def{\sf v}{{\sf v}}
\def{\sf w}{{\sf w}}
\def{\sf x}{{\sf x}}
\def{\sf y}{{\sf y}}
\def{\sf z}{{\sf z}}
\def\bfmath#1{{\mathchoice{\mbox{\boldmath$#1$}}%
{\mbox{\boldmath$#1$}}%
{\mbox{\boldmath$\scriptstyle#1$}}%
{\mbox{\boldmath$\scriptscriptstyle#1$}}}}
\def\bfmath{a}{\bfmath{a}}
\def\bfmath{b}{\bfmath{b}}
\def\bfmath{d}{\bfmath{d}}
\def\bfmath{e}{\bfmath{e}}
\def\bfmath{m}{\bfmath{m}}
\def\bfmath{q}{\bfmath{q}}
\def\bfmath{r}{\bfmath{r}}
\def\bfmath{v}{\bfmath{v}}
\def\bfmath{x}{\bfmath{x}}
\def\bfmath{y}{\bfmath{y}}
\def\bfmath{u}{\bfmath{u}}
\def\bfmath{w}{\bfmath{w}}
\def\bfmath{\widehat w}{\bfmath{{\hat w}}}
\def\bfmath{z}{\bfmath{z}}
\def\bfmath{A}{\bfmath{A}}
\def\bfmath{B}{\bfmath{B}}
\def\bfmath{C}{\bfmath{C}}
\def\bfmath{D}{\bfmath{D}}
\def\bfmath{E}{\bfmath{E}}
\def\bfmath{F}{\bfmath{F}}
\def\bfmath{G}{\bfmath{G}}
\def\bfmath{H}{\bfmath{H}}
\def\bfmath{I}{\bfmath{I}}
\def\bfmath{{\hat I}}{\bfmath{{\hat I}}}
\def\bfmath{\haL}{\bfmath{\haL}}
\def\bfmath{L}{\bfmath{L}}
\def\bfmath{M}{\bfmath{M}}
\def\bfmath{N}{\bfmath{N}}
\def\bfmath{Q}{\bfmath{Q}}
\def\bfmath{R}{\bfmath{R}}
\def\bfmath{\bar{R}}{\bfmath{\bar{R}}}
\def\bfmath{S}{\bfmath{S}}
\def\bfmath{T}{\bfmath{T}}
\def\bfmath{U}{\bfmath{U}}
\def\bfmath{\widehat N}{\bfmath{\widehat N}}
\def{\bfmath{$\widehat Q$}}{\bfmath{\widehat Q}}
\def\bfmath{X}{\bfmath{X}}
\def\tilde \bfmX{\bfmath{\widetilde{X}}}
\def\bfmath{\widetilde{Q}}{\bfmath{\widetilde{Q}}}
\def\tilde{\bfmath{q}}{\tilde{\bfmath{q}}}
\def\bfmath{Y}{\bfmath{Y}}
\def\bfmath{\widehat Y}{\bfmath{\widehat Y}}
\def\hbox to 0pt{$\widehat{\bfmY}$\hss}\widehat{\phantom{\raise 1.25pt\hbox{$\bfmY$}}}{\bfmath{\hhaY}}
\def\hbox to 0pt{$\widehat{\bfmY}$\hss}\widehat{\phantom{\raise 1.25pt\hbox{$\bfmY$}}}{\hbox to 0pt{$\widehat{\bfmath{Y}}$\hss}\widehat{\phantom{\raise 1.25pt\hbox{$\bfmath{Y}$}}}}
\def\bfmath{V}{\bfmath{V}}
\def\bfmath{W}{\bfmath{W}}
\def\bfmath{\widehat W}{\bfmath{{\widehat W}}}
\def\bfmath{Z}{\bfmath{Z}}
\def{\bfmath{\widehat Z}}{{\bfmath{\widehat Z}}}
\def\bfmath{\widehat w}{\bfmath{\widehat w}}
\def\bfmath{\widehat y}{\bfmath{\widehat y}}
\def\bfmath{\widehat z}{\bfmath{\widehat z}}
\def{\bfmath{$\widehat Q$}}{\bfmath{\widehat Q}}
\def\bfmath{\widehat S}{\bfmath{\widehat S}}
\def\bfmath{\widehat U}{\bfmath{\widehat U}}
\def\bfmath{\widehat W}{\bfmath{\widehat W}}
\def{\mbox{\boldmath$\alpha$}}{\bfmath{\theta}}
\def\bfmath{\Phi}{\bfmath{\Phi}}
\def\bfmath{\eta}{\bfmath{\eta}}
\def\bfmath{\pi}{\bfmath{\pi}}
\def\bfmath{\phi}{\bfmath{\phi}}
\def\bfmath{\psi}{\bfmath{\psi}}
\def\bfmath{\rho}{\bfmath{\rho}}
\def\bfmath{\zeta}{\bfmath{\zeta}}
\def\bfmath{\Upsilon}{\bfmath{\Upsilon}}
\def{\widehat{\Psi}}{{\widehat{\Psi}}}
\def{\widehat{\Gamma}}{{\widehat{\Gamma}}}
\def{\widehat{\Sigma}}{{\widehat{\Sigma}}}
\def{\widehat{\bfPhi}}{{\widehat{\bfmath{\Phi}}}}
\def{\widehat P}{{\widehat P}}
\def{\hat c}{{\hat c}}
\def\hat{z}{\hat{z}}
\def{\hat {\i}}{{\hat {\i}}}
\def{\hat f}{{\hat f}}
\def{\hat g}{{\hat g}}
\def{\hat m}{{\hat m}}
\def{\hat h}{{\hat h}}
\def{\hat p}{{\hat p}}
\def{\hat q}{{\hat q}}
\def{\bfmath{$\widehat Q$}}{{\bfmath{$\widehat Q$}}}
\def\hat v{\hat v}
\def{\hat w}{{\hat w}}
\def{\hat y}{{\hat y}}
\def{\widehat F}{{\widehat F}}
\def{\hat I}{{\hat I}}
\def{\hat J}{{\hat J}}
\def\hat{N}{\hat{N}}
\def{\widehat Q}{{\widehat Q}}
\def{\widehat T}{{\widehat T}}
\def{\widehat W}{{\widehat W}}
\def{\hat\eta}{{\hat\eta}}
\def{\hat\theta}{{\hat\theta}}
\def{\hat\psi}{{\hat\psi}}
\def{\hat\pi}{{\hat\pi}}
\def{\hat\nu}{{\hat\nu}}
\def{\hat\mu}{{\hat\mu}}
\def{\hat\alpha}{{\hat\alpha}}
\def{\hat\beta}{{\hat\beta}}
\def{\hat\gamma}{{\hat\gamma}}
\def{\hat\lambda}{{\hat\lambda}}
\def{\hat\Lambda}{{\hat\Lambda}}
\def{\rm A}{{\rm A}}
\def{\rm B}{{\rm B}}
\def{\rm C}{{\rm C}}
\def{\rm C}{{\rm C}}
\def{\rm E}{{\rm E}}
\def{\rm H}{{\rm H}}
\def{\rm I}{{\rm I}}
\def{\rm M}{{\rm M}}
\def{\rm P}{{\rm P}}
\def{\rm Q}{{\rm Q}}
\def{\rm d}{{\rm d}}
\def{\rm g}{{\rm g}}
\def{\rm k}{{\rm k}}
\def{\rm u}{{\rm u}}
\def{\rm x}{{\rm x}}
\def{\rm P}{{\rm P}}
\def{\rm E}{{\rm E}}
\def\til={{\widetilde =}}
\def{\widetilde \Phi}{{\widetilde \Phi}}
\def{\widetilde N}{{\widetilde N}}
\def{\widetilde P}{{\widetilde P}}
\def{\widetilde Q}{{\widetilde Q}}
\def\widetilde T{\widetilde T}
\def\widetilde W{\widetilde W}
\def{\tilde \mu}{{\tilde \mu}}
\def{\tilde \pi}{{\tilde \pi}}
\def{\bf\tilde X}{{{\mathbb f}\tilde X}}
\def{\bf \tilde m}{{{\mathbb f} \tilde m}}
\def{\tilde \theta}{{\tilde \theta}}
\def\tilde e{\tilde e}
\def\tilde \sigma{\tilde \sigma}
\def\tilde \tau{\tilde \tau}
\def{\cal A}{{\cal A}}
\def{\cal B}{{\cal B}}
\def{\cal C}{{\cal C}}
\def{\cal D}{{\cal D}}
\def{\cal E}{{\cal E}}
\def{\cal F}{{\cal F}}
\def{\cal G}{{\cal G}}
\def{\cal H}{{\cal H}}
\def{\cal J}{{\cal J}}
\def{\cal I}{{\cal I}}
\def{\cal K}{{\cal K}}
\def{\cal L}{{\cal L}}
\def{\cal M}{{\cal M}}
\def{\cal N}{{\cal N}}
\def{\cal O}{{\cal O}}
\def{\cal P}{{\cal P}}
\def{\cal Q}{{\cal Q}}
\def{\cal R}{{\cal R}}
\def{\cal S}{{\cal S}}
\def{\cal T}{{\cal T}}
\def{\cal U}{{\cal U}}
\def{\cal V}{{\cal V}}
\def{\cal W}{{\cal W}}
\def{\cal X}{{\cal X}}
\def{\cal Y}{{\cal Y}}
\def{\cal Z}{{\cal Z}}
\def{\mbox{\boldmath$\alpha$}}{{\mbox{\boldmath$\alpha$}}}
\def{\mathchoice{\bfatom}{\bfatom}{\alpha}{\alpha}}{{\mathchoice{{\mbox{\boldmath$\alpha$}}}{{\mbox{\boldmath$\alpha$}}}{\alpha}{\alpha}}}
\def\clG^+(\gamma){{\cal G}^+(\gamma)}
\def\clG^+(s){{\cal G}^+(s)}
\def{\mathbb{V}\sfa \sfr}{{\mathbb{V}{\sf a} {\sf r}}}
\defA_+{A_+}
\def\barAt#1{\overline{A_+(#1)}}
\def\head#1{\subsubsection*{#1}}
\def\atopss#1#2{\genfrac{}{}{0cm}{2}{#1}{#2}}
\def\atop#1#2{\genfrac{}{}{0pt}{}{#1}{#2}}
\def\FRAC#1#2#3{\genfrac{}{}{}{#1}{#2}{#3}}
\def\fraction#1#2{{\mathchoice{\FRAC{0}{#1}{#2}}%
{\FRAC{1}{#1}{#2}}%
{\FRAC{3}{#1}{#2}}%
{\FRAC{3}{#1}{#2}}}}
\def\ddt{{\mathchoice{\FRAC{1}{d}{dt}}%
{\FRAC{1}{d}{dt}}%
{\FRAC{3}{d}{dt}}%
{\FRAC{3}{d}{dt}}}}
\def\ddr{{\mathchoice{\FRAC{1}{d}{dr}}%
{\FRAC{1}{d}{dr}}%
{\FRAC{3}{d}{dr}}%
{\FRAC{3}{d}{dr}}}}
\def\ddr{{\mathchoice{\FRAC{1}{d}{dy}}%
{\FRAC{1}{d}{dy}}%
{\FRAC{3}{d}{dy}}%
{\FRAC{3}{d}{dy}}}}
\def\ddtp{{\mathchoice{\FRAC{1}{d^{\hbox to 2pt{\rm\tiny +\hss}}}{dt}}%
{\FRAC{1}{d^{\hbox to 2pt{\rm\tiny +\hss}}}{dt}}%
{\FRAC{3}{d^{\hbox to 2pt{\rm\tiny +\hss}}}{dt}}%
{\FRAC{3}{d^{\hbox to 2pt{\rm\tiny +\hss}}}{dt}}}}
\def\ddalpha{{\mathchoice{\FRAC{1}{d}{d\alpha}}%
{\FRAC{1}{d}{d\alpha}}%
{\FRAC{3}{d}{d\alpha}}%
{\FRAC{3}{d}{d\alpha}}}}
\def\half{{\mathchoice{\FRAC{1}{1}{2}}%
{\FRAC{1}{1}{2}}%
{\FRAC{3}{1}{2}}%
{\FRAC{3}{1}{2}}}}
\def\third{{\mathchoice{\FRAC{1}{1}{3}}%
{\FRAC{1}{1}{3}}%
{\FRAC{3}{1}{3}}%
{\FRAC{3}{1}{3}}}}
\def\fourth{{\mathchoice{\FRAC{1}{1}{4}}%
{\FRAC{1}{1}{4}}%
{\FRAC{3}{1}{4}}%
{\FRAC{3}{1}{4}}}}
\def\threefourth{{\mathchoice{\FRAC{1}{3}{4}}%
{\FRAC{1}{3}{4}}%
{\FRAC{3}{3}{4}}%
{\FRAC{3}{3}{4}}}}
\def\sixth{{\mathchoice{\FRAC{1}{1}{6}}%
{\FRAC{1}{1}{6}}%
{\FRAC{3}{1}{6}}%
{\FRAC{3}{1}{6}}}}
\def\eighth{{\mathchoice{\FRAC{1}{1}{8}}%
{\FRAC{1}{1}{8}}%
{\FRAC{3}{1}{8}}%
{\FRAC{3}{1}{8}}}}
\def\buildrel{\rm dist}\over ={\buildrel{\rm dist}\over =}
\def\mathbin{:=}{\mathbin{:=}}
\def\mu^{\hbox{\rm \tiny Leb}}{\mu^{\hbox{\rm \tiny Leb}}}
\def{\cal M}{{\cal M}}
\def{\sf P}{{\sf P}}
\def{\sf P\! }{{\sf P\! }}
\def{\sf E}{{\sf E}}
\def\taboo#1{{{}_{#1}P}}
\def{\rm i.o.}{{\rm i.o.}}
\def{\rm a.s.}{{\rm a.s.}}
\def{\rm vec\, }{{\rm vec\, }}
\def{\displaystyle\frac{1}{ N}}{{\displaystyle\frac{1}{ N}}}
\def\average#1,#2,{{1\over #2} \sum_{#1}^{#2}}
\def\hbox{ \rm \ enters\ }{\hbox{ \rm \ enters\ }}
\def\eye(#1){{{\mathbb f}(#1)}\quad}
\def\varepsilon{\varepsilon}
\def\,\cdot\,{\,\cdot\,}
\def{{\mathbb V}{\sf a}{\sf r}}{{{\mathbb V}{\sf a}{\sf r}}}
\newtheorem{theorem}{{{\mathbb f} Theorem}}
\newtheorem{proposition}{{{\mathbb f} Proposition}}
\newtheorem{corollary}{Corollary}
\newtheorem{definition}{{{\mathbb f} Definition}}
\newtheorem{remark}{{{\mathbb f} Remark}}
\newtheorem{lemma}{{{\mathbb f} Lemma}}
\def\Lemma#1{Lemma~\ref{t:#1}}
\def\Proposition#1{Proposition~\ref{t:#1}}
\def\Theorem#1{Theorem~\ref{t:#1}}
\def\Corollary#1{Corollary~\ref{t:#1}}
\def\Remark#1{Remark~\ref{t:#1}}
\def\Section#1{Section~\ref{#1}}
\def\Figure#1{Figure~\ref{f:#1}}
\def\Chapter#1{Chapter~\ref{c:#1}}
\def\Appendix#1{Appendix~\ref{c:#1}}
\def\eq#1/{(\ref{e:#1})}
\newcommand{\inp}[2]{{\langle #1, #2 \rangle}}
\newcommand{\inpr}[2]{{\langle #1, #2 \rangle}_{\mathbb R}}
\newcommand{\beqn}[1]{\notes{#1}%
\begin{eqnarray} \elabel{#1}}
\newcommand{\end{eqnarray} }{\end{eqnarray} }
\newcommand{\beq}[1]{\notes{#1}%
\begin{equation}\elabel{#1}}
\newcommand{\end{equation}}{\end{equation}}
\def\begin{description}{\begin{description}}
\def\end{description}{\end{description}}
\def{\overline {a}}{{\overline {a}}}
\def{\overline {c}}{{\overline {c}}}
\def{\overline {f}}{{\overline {f}}}
\def{\overline {g}}{{\overline {g}}}
\def{\overline {h}}{{\overline {h}}}
\def{\overline {l}}{{\overline {l}}}
\def{\overline {m}}{{\overline {m}}}
\def{\overline {n}}{{\overline {n}}}
\def{\overline {p}}{{\overline {p}}}
\newcommand{{\bar{x}}}{{\bar{x}}}
\def{\overline {y}}{{\overline {y}}}
\def{\overline {A}}{{\overline {A}}}
\def{\overline {B}}{{\overline {B}}}
\def{\overline {C}}{{\overline {C}}}
\def{\overline {E}}{{\overline {E}}}
\def{\overline {M}}{{\overline {M}}}
\def{\overline {P}}{{\overline {P}}}
\def{\overline {Q}}{{\overline {Q}}}
\def{\overline {T}}{{\overline {T}}}
\def{\underline{n}}{{\underline{n}}}
\def{\underline{\rho}}{{\underline{\rho}}}
\def{\overline{\atom}}{{\overline{{\mathchoice{\bfatom}{\bfatom}{\alpha}{\alpha}}}}}
\def{\overline{\rho}}{{\overline{\rho}}}
\def{\overline{\mu}}{{\overline{\mu}}}
\def{\overline{\nu}}{{\overline{\nu}}}
\def{\overline{\alpha}}{{\overline{\alpha}}}
\def{\overline{\beta}}{{\overline{\beta}}}
\def{\overline{\alpha}}{{\overline{\alpha}}}
\def{\overline{\eta}}{{\overline{\eta}}}
\def\overline{\bfPhi}{\overline{\bfmath{\Phi}}}
\def\overline{\Phi}{\overline{\Phi}}
\def\overline{\clF}{\overline{{\cal F}}}
\def\overline{\sigma}{\overline{\sigma}}
\def\overline{\Sigma}{\overline{\Sigma}}
\def\overline{\tau}{\overline{\tau}}
\def\hat{\sigma}{\hat{\sigma}}
\def\hat{\tau}{\hat{\tau}}
\def\hat{P}{\hat{P}}
\def\hat{\Phi}{\hat{\Phi}}
\def\hat{\bfPhi}{\hat{\bfmath{\Phi}}}
\def\hat{\cal F}{\hat{\cal F}}
\newcounter{rmnum}
\newenvironment{romannum}{\begin{list}{{\upshape (\roman{rmnum})}}{\usecounter{rmnum}
\setlength{\leftmargin}{14pt}
\setlength{\rightmargin}{8pt}
\setlength{\itemindent}{-1pt}
}}{\end{list}}
\newcounter{anum}
\newenvironment{alphanum}{\begin{list}{{\upshape (\alph{anum})}}{\usecounter{anum}
\setlength{\leftmargin}{14pt}
\setlength{\rightmargin}{8pt}
\setlength{\itemindent}{-1pt}
}}{\end{list}}
\newenvironment{assumptions}[1]{%
\setcounter{keep_counter}{\value{#1}}
\def#1{#1}
\begin{list}{\sf\, (#1\arabic{#1})\mindex{Assumption (#1\arabic{#1})}\, }{\usecounter{#1}%
\setlength{\rightmargin}{\leftmargin}}
\setcounter{#1}{\value{keep_counter}}}%
{\end{list}}
\def\ass(#1:#2){(#1\ref{#1:#2})}
\def\ritem#1{
\item[{\sf \ass(\current_model:#1)}]
}
\newenvironment{recall-ass}[1]{%
\begin{description}
\def\current_model{#1}}{
\end{description}
}
\def\Ebox#1#2{%
\begin{center}
\parbox{#1\hsize}{\epsfxsize=\hsize \epsfbox{#2}}
\end{center}}
\def\cite{meytwe92d}{\cite{meytwe92d}}
| 7afc65112c152c9eb779bf98fe5dd669f52b8196 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{intro}
Cellular communication relies upon binding of ligands to specific cell surface receptors. \textit{G-protein-coupled receptors} (GPCRs) are key players in initiating and regulating cellular processes as they mediate responses to hormones, neurotransmitters, metabolites, ions, fatty acids, pathogens, and physical stimuli, such as light, smell, taste, and mechanical stretch. Moreover they can be activated by synthetic agonists, inhibited by antagonists and inverse agonists, or affected by allosteric modulators. They actually represent the most important superfamily of clinical targets in disorders of neural, immune, cardiovascular, endocrine, respiratory system and cancer \cite{Cole:2011, Unal:2012, Allard:2012, Mary:2013}.
In particular, for disease like asthma, it is known to involve specific GPCRs, called $\beta_2$-\textit{adrenergic receptors}. This is because asthma entails a continuous use of $\beta_2$-agonists, which are located in human airway epithelial cells. Indeed, this results in loss of bronchoprotective effects and deterioration of asthma control.
The importance of such receptors in cancer relates to the fact that cAMP-based markers have shown to have a great potential for the early diagnosis of certain tumors. Groundwork for translation of the so called $\beta$-blockade as a novel adjuvant to existing therapeutic strategies in clinical oncology is on its way (see e.g. \cite{Cho:2000, Cvijic:2000, Liu:2010, Banerjee:2016, Choi:2018}).
GPCRs are major modulators of communication between the internal and external milieu of cells. These receptors are integral membrane proteins with an extracellular N-terminus and seven \textit{TransMembrane} helical \textit{domains} (TMs), from TM1 to TM7, connected by loop regions. Nowadays GPCR signaling is recognized to be more complex that was originally understood \cite{Kobilka:2007}. Briefly, their binding to very different kinds of extracellular stimuli induces TM domain \textit{conformational changes} and the structural remodelling of the protein. The latter allows for the coupling with cytoplasmic G proteins, which is followed by the activation of second messenger generating enzymes. The produced second messenger can activate many downstream signaling pathways inside the cell \cite{Lefkowitz:2007}. This activity of GPCRs is known to be detected through measurements of \textit{cyclyc Adenosyne Monophospate} -cAMP-, the detectable signature of the pathway arising in response to ligands, such as epinephrine. The production of cAMP in the cellular environment turns out to be modulated by several factors. In particular, the presence of \textit{Multidrug Resistance Proteins}-MRPs allows for the efflux of cAMP from the interior of the cell to the extracellular fluid, by so maintaining homeostatic intracellular concentrations. For this reason, MRPs are called transporters, as they contribute to keep a balance of cAMP inside the cell. A schematic of this process is depicted in Figure \ref{fig_cell_with_receptors_and_transporters}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.7 \columnwidth]{fig_cell_with_receptors}
\caption{Schematic representation of the ligand-binding between epinephrine (red dots on the top-left) and GPCRs (left, formed by seven cylindrical Transmembrane Domains-TMs) with production of intracellular and extracellular cyclic-Adenosine MonoPhosphate-cAMP- in the presence of transporters-MRPs (yellow unit to the right). Intermediate units are the G-protein (brown circle, near the GPCRS) and the Adenylate Cyclase (purple unit, across the membrane). The surface density of GPCRs bound to ligands is denoted by $\xi$, while the analog for MRPs is labelled as $\zeta$. The letters A, B, C and D display the order in which the chain of events occurs from the time in which ligand-binding takes place to the effects of the cell response through the production of intracellular and extracellular cAMP. The scheme highlights that a broad
spectrum inhibitor of cAMP phosphodiesterase, labelled as PDE, is accounted for \cite{Biondi:2010}. Helical domains forming the GPCR are displayed as cylinders in the figure: the TMs, where TM3, TM5 and TM6 have been reported in colors, to highlight their conformational changes. $\,$ The latter are displayed at the bottom of this figure. They consist in a rotation ($\omega$) of such domains about their axis, orthogonal to the (local) mid plane of the membrane (whose unit vector is denoted by $\bold{e}_3$), and on the shear ($\mu$) of TM6 approximately towards TM5, in the direction $\bold{s}$, whose components in the local mid-plane are also displayed in the figure (see e.g. \cite{Ghanouni:2001}).}
\label{fig_cell_with_receptors_and_transporters}
\end{figure}
Historically, \textit{adrenergic receptors} are some of the most studied GPCRs. Two main subfamilies, $\alpha$ and $\beta$, which differ in tissue localization, ligand specificity, G protein coupling, and downstream effector mechanisms, are present. Diverse diseases, such as asthma, hypertension, and heart failure, are related to genetic modifications of adrenergic receptors. In particular, $\beta$2-adrenergic receptors ($\beta$2ARs) are found in smooth muscles throughout the body, and $\beta$2AR agonists are common treatments for asthma and preterm labor \cite{Cherezov:2007}. In \cite{Biondi:2006} it has been demonstrated the presence of $\beta$-adrenergic receptors in HTR-8/SVneo cell line, a well characterized first-trimester human extravillous trophoblast-derived cell line. Such cells invade the maternal uterine stroma, the decidua and the deeper portion of the myometrium evoking profound changes within the uterine vessels. A limited trophoblast invasion of maternal vessels has been correlated to both preeclampsia and fetal growth restriction, whereas an excessive trophoblast invasion is associated with invasive mole, placenta accreta and choriocarcinoma \cite{Lunghi:2007}. Other studies showed the importance of GPCRs during pregrancy \cite{LeBouteiller:2006, Hanna:2006}.
Ghanouni and coworkers \cite{Ghanouni:2001} focused their studies on $\beta$2AR \textit{conformational changes} induced by agonist binding using quenching experiments. Their results can best be explained by either a rotation of TM6 and a tilting of TM6 toward TM5 during agonist-induced activation of the receptor. The structure, activation, and signaling of a GPCR is heavily influenced by the bilayer environment. This influence seems likely to involve either indirect bilayer effects, specific membrane-GPCR interactions, or a combination of both \cite{Drake:2006, Chini:2004, Oates:2011}.
It has been reported that specific properties of this environment result in co-localization with downstream signaling components providing a rationale for regulation and specificity in GPCR activation \cite{Ostrom:2004, Patel:2008}. Some subsets of GPCRs are \textit{preferentially segregated to discrete regions} of the membrane defined as \textit{lipid rafts}
\cite{Bray:1998, foster2003, Chini:2004, Ostrom:2004, Zhang:2005, Becher:2005, Barnett-Norris:2005, Gopalakrishnan:2005, Patel:2008, Watkins:2011, Villar:2016, Shukla:2016}. Despite lipid rafts are involved in a number of biologically relevant phenomena and their formation has been demonstrated to often serve as entry port for some viruses, including SARS-CoV ones \cite{lu2008lipid}, why is this the case and what is the physics at the basis of the associated remodelling of the lipid membrane
remain still open questions \cite{Gopalakrishnan:2005, Fallahi-Sichani:2009, Jacobson2007}. Lipid raft are highly dynamic thickened areas on the lipid membrane with different dimensions and a complex heterogeneous composition. This is due to the fact that such rafts host a variety of transmembrane proteins responsible of compartmentalizing cellular processes to isolated districts of the cell. These planar regions form localized microdomains on the cell membrane
and are hardly directly observable \textit{in vivo} \cite{risselada2008, DHARANI2015123}. \color{black} In particular, they appear as dynamic nanoscale assemblies, where a high glycosphingolipid and cholesterol content in the outer leaflet of the lipid bilayer is present. This composition gives them a gel-like liquid-ordered organisation in comparison with the surrounding phospholipid-rich disordered membrane \cite{Chini:2004, Lingwood:2010}.
Each lipid molecule, composed by a hydrophilic head and a hydrophobic tail, may exhibit two different shape conditions of the tail, either straightened and taller (also known as \textit{ordered state}, $L_o$), or curly and shortened (also known as \textit{disordered state}, $L_d$). Such states depend on several conditions, among which the temperature and chemical composition of the lipid mixture are the main factors (see e.g. \cite{Bermudez:2004, Das:2008, Iglic:2012, Sackmann:1995}).
The interplay between lipids and proteins is known to influence the overall behavior of the cell membrane. In particular,the adenosine A1, $\alpha$1-AR, $\beta$1-AR, $\beta$2-AR, AT1R, the endothelin (ETA-A and ET-B) receptors, and the M2-muscarinic receptors have all been localized to lipid rafts and/or caveolae \cite{Drake:2006}.
The association of $\beta$2AR with caveolin3 in myocytes membranes has been demonstrated, as well as the fact that the confinement of $\beta$2-AR to caveolae is of critical importance for regulation of the intrinsic contraction rate in these membrane preparations \cite{Xiang:2002}. How cholesterol modifies GPCR activity appears very much receptor-dependent, with regards to both upregulation and downregulation and of direct and indirect action \cite{Oates:2011, Hanson:2008}.
The mechanical and chemical response of these amazing structures depends on many factors, such as the shape of the membrane, the temperature of the environment, the osmotic pressure, the chemical composition of lipid mixture, etc. \cite{Hu:2012, Norouzi:2006, Agrawal:2008, Agrawal:2009, Walani:2015, Baumgart:2003, Baumgart:2005,Honerkamp:2008}.
Depending on the presence of embedded specialized proteins into the lipid membrane, several predicting models have been developed \cite{DeseriPZD:2015, Deseri:2013, Deseri:2008, Zurlo:2006, Canham:1970, Jenkins:1977, Agrawal:2009, Biscari:2002, Rangamani:2014, Walani:2015}. Most of them are based on the characterization of the energy, with the aim of predicting the ordered-disordered phenomena arising in transition regions for contrasting the external stimuli. Indeed, thickness reduction of the bilayer is observed by changing the mechanical and chemical conditions in the environment surrounding the cell. For this reason, several Authors adopted the change in thickness as order parameter for studying such kind of systems \cite{Goldstein:1989, Owicki:1979, Sackmann:1995}. Without explicitly doing so, a theory for the chemo-mechanical coupling of lipid bilayers accounting for curvature changes, viscosity, diffusion and $L_o-L_d$ phase transitions has recently been presented in \cite{Sauer:2017}. The $L_o-L_d$ phase transition has been modelled through a dimensionless order parameter related to the fraction of dioleoylphosphatidylcholine, known as DOPC. This is known to exhibit disordered features in agreement with the curliness of the tails mentioned above. Furthermore, binding and diffusion of certain proteins, epsin-1, is thoroughly analyzed. \color{black} As pointed out above, the physical behavior of such membranes is regulated by many factors enabling the occurrence of out-of-plane tractions, while in-plane shear stresses are not transmitted unless their viscosity is accounted for.
The increasing availability of imaging techniques led indeed to a striking increase of interest in the study of lipid membranes, often revealing examples of the complex features characterizing their behavior (see, {\it e.g.}, \cite{Baumgart:2003, Baumgart:2005}). Bilayer elasticity has been fruitfully exploited in the literature for the study of equilibrium shapes of red blood cells \cite{Jenkins:1977} and GUVs (Giant Unilamellar Vesicle) \cite{Canham:1970, Helfrich:1973}. The effects of embedded proteins or rod-like inclusions in the lipid membrane have been analyzed in \cite{Agrawal:2009, Walani:2015, Biscari:2002}, together with the analysis of buds formation \cite{Lipowsky:1992} with the coexistence of domains characterized by different bending rigidities \cite{Agrawal:2008, Baumgart:2005}.
Models for the exhibited order-disorder transition can be found in \cite{Akimov:2003, Chen:2001, Falkovitz:1982, Goldstein:1989, Iglic:2012, Jahnig:1981, Owicki:1978, Owicki:1979, Deserno:2014, Steigmann:2015} among others, while the effects of special molecules (like cholesterol) on such transition have also been studied in \cite{Komura:2004, Pan:2009, Rawicz:2000, Akimov:2003, Honerkamp:2008, Lipowsky:1992}.
Models of lipid membranes where the bending behavior, the order-disorder transition and the chemical composition have been consistently taken into account can be found in \cite{Deseri:2008, Zurlo:2006, Maleki:2013, Agrawal:2009, Walani:2015, Steigmann:2013, Steigmann:2014a, Rangamani:2014, Deserno:2014}. In \cite{Deseri:2008, Zurlo:2006}, the energetics regulating the behavior of such membranes was obtained through asymptotic dimension reduction (see also \cite{Deseri:2010}).
The major point in such papers is that the ``quasi-incompressibility'' of the environment, together with the influence of chemical composition, is enough to describe the membrane $L_o-L_d$ transition.
With the aim of understanding why GPCRs prefer to live on lipid rafts and, in turn, why the activation of transmembrane proteins involved in cell processes induce the formation of such thickened microdomains, the mechanobiology of the remodelling of the lipid bilayer is investigated by focusing on the interplay among membrane elasticity \color{black} and density changes of active species. In order to predict the dynamic remodelling of the cell membrane, following thermodynamic arguments, a chemo-mechanical coupling is found for the obtained chemical potentials of such species. These potentials regulate active species diffusion through mass balance which, in turn, involves the kinetics of binding through the occurence of interspecific terms. The potentials of the active species determine a corresponding coupling for the resulting energetics. Furthermore, it is found that the chemo-mechanical coupling found for such potentials is the specific work exterted by the lateral pressure (arising across the membrane thickness) against the volume changes around the interaction sites among the species domains and the surrounding lipids. It is worth noting that accounting for the aspects mentioned above enables mechanobiology to give a mechanically-based justification to the experimental findings of Nobel Prizes 2012 Kobilka \cite{Kobilka:2007} and Leifkovitz \cite{Lefkowitz:2007}. \color{black}
\section{Experimental measurements to trace protein fractions}
\label{ExpMeas}
\setcounter{equation}{0}
Experiments involving human trophoblast cells known to contain $\beta$-adrenergic receptors have been performed to the extent of detecting the activity of such GPCR in the human placenta (see e.g. \cite{Lunghi:2007}, \cite{Biondi:2010}).
The occurrence of a Multidrug Resistant Protein (MRP)-dependent cAMP efflux is also shown in human first-trimester placenta explants. Extracellular cAMP has been hypothesized to represent a source for adenosine formation that, in turn, could modulate cAMP-dependent responses in placental tissue. Evidence is provided that adenosine receptor subtypes are present and functional in \textit{Human TRophoblast} (HTR)-derived \textit{cells}. A role for cAMP egress mechanism in the fine modulation of the nucleotide homeostasis is then highly probable.
\subsection{Materials and Methods}
In the papers cited above, HTR-8/SVneo trophoblast cell line obtained from human first-trimester placenta ex-plant cultures and immortalized using SV40 large T antigen, donated by Dr. C.H. Graham, Queen$^{^,}$s University (Kingston, ON-Canada) to the Cell Biology Laboratory of the University of Ferrara, were cultured at 37$^\circ$C in a controlled atmosphere of $5\%$ $CO_2/95\%$ air in RPMI 1640 medium containing $10\%$ fetal bovine serum, 100 U/ml penicillin and 100 $\mu$g/ml streptomycin. These cells were grown to confluence (2-3 days) in a 24-well plate (250,000 cells per well). The medium was then removed and replaced by serum-free RPMI.
The incubation process took place in wells of capacity $C_w=500 \,\mu l$ of such RPMI. This was carried out in the presence of isobutylmethylxanthine, otherwise known as IBMX, a broad
spectrum inhibitor of cAMP phosphodiesterase (PDE) \cite{Biondi:2010}. This was performed in the presence and in the absence of a \textit{given amount of epinephrine} for the indicated times.
Media were then collected and immediately frozen at -70$^{\circ}$C until \textit{extracellular cAMP }(cAMP$_e$) levels were measured. Ice-cold 0.1 N HCl (0.25 ml) was added to the cells and, after centrifugation at 12,500 x g for 10 min, supernatants were neutralized adding 0.5 M Trizma base (0.05 ml) and utilized for measuring \textit{intracellular cAMP} (cAMP$_i$).
Finally, cAMP$_i$ and cAMP$_e$ were determined by standard methods (e.g. method of Brown \cite{Brown:1972}) and the nucleotide levels were expressed as $pmoles/10^6 cells/time$.
Tests with increasing concentration of epinephrine were first performed and cAMP measurements were recorded at a given time (10 minutes).
Epinephrine turned out to enhance intracellular cAMP in a dose-dependent fashion, reaching a plateau at around $10^{-4}\, M$ (Figure \ref{fig:experiment}\hyperref[{fig:experiment}]{A}).
Furthermore, the cAMP evolution over time was monitored in the absense and in the presence of $10^{-6}\, M$ concentration of epinephrine. Such a concentration was chosen because it triggered a cAMP$_i$ production close to the half of the maximum response (Figure \ref{fig:experiment}\hyperref[{fig:experiment}]{B}).
The cAMP levels were measured in cells incubated up to 60 minutes.
In basal conditions, cAMP concentrations remained almost constant at all tested times (around $6.0$ pmoles/$10^6$ cells, not shown). In the presence of epinephrine, cAMP$_i$ increased as a function of the incubation time up to 15 min (14-fold), thereafter a reduction of the nucleotide level was observed.
At the same time, cAMP$_e$ gradually increases in time, thereby almost reaching a plateaux, at least during the 60 min of observation (not shown).
Analogous experimental observations in \cite{Biondi:2006, Lunghi:2007, Biondi:2010}, although with different ligands, showed an analogous trend. For the present case, when a concentration $c = 1 \,\mu M$ of epinephrine was used, the total quantity $Q$ of ligand utilized in the experiment involving $10^6$ cells can be then computed as follows:
\begin{equation}
\label{eq:totalQ}
Q = 4 \times \, c \,C_w = 4 \times 1 \, \mu M \times 500 \, \mu L = 2000 \,pmol .
\end{equation}
\begin{figure*}[t]
\includegraphics[width=0.45 \columnwidth]{fig05_experiment_a_MOD} \quad
\includegraphics[width=0.45 \columnwidth]{fig05_experiment_b}
\caption{Experimental results: (a) value of measured intracellular cAMP after 10 minutes depending on epinephrine concentration, (b) time evolution of measured intracellular cAMP for a fixed value of epinephrine concentration ($10^{-6} M$).}
\label{fig:experiment}
\end{figure*}
The experimental diagrams coming from the measurements of cAMP production are displayed in Figure \ref{fig:experiment}.
\paragraph{cAMPi-to-$\xi$ and cAMPe-to-$\zeta$ relationships} The fields of active receptors density $\xi(x,t)$ and its products, namely the intracellular cAMP$_i$, are intimately linked by a conversion factor, namely $a_{\xi}$, as follows:
\begin{equation}
\label{eq:camptoxi}
cAMP_i = a_{\xi} \left(\xi - \bar{\xi} \right) \approx a_{\xi}\, \xi ,
\end{equation}
where $a_{\xi} = 10^4$ is estimated experimentally \cite{Biondi:2010} and $\bar{\xi}$ is the basal value of the density of active receptors. Here, the cAMP level is referred to a population of $10^6$ cells (and it is expressed in $pmol$). Taking into account a parameter $c_1 = 10^6$ for switching between the population and the single cell, and $c_2 = 10^{12}$ for converting $pmol$ to $mol$, the number of active receptors can be computed as follows:
\begin{equation}
\label{eq:cAMP-to-Xi}
\xi_{\#} = \frac{1}{c_1}\frac{1}{a_{\xi}} \left( \frac{cAMP_i}{c_2}\right) N_A ,
\end{equation}
where $N_A$ is the Avogadro number. An analogous relationship holds for computing the number of transporters:
\begin{equation}
\label{eq:camptozeta}
cAMP_e = a_{\zeta} \left(\zeta - \bar{\zeta} \right) \approx a_{\zeta} \zeta,
\end{equation}
where $\bar{\zeta}$ represents the basal value of the density of active transporters. In both the cases, direct proportionality between cAMP concentrations has been used to obtain an indirect estimation of the protein fractions predicted by the dynamics.
\section{The biomechanical model}
\label{sec:2}
The mechanobiology of lipid raft formation is investigated by accounting for deformation localization phenomena during binding of receptors, namely GPCRs compounds, with the incoming ligand. To this \color{black} aim, \color{black} the cell membrane is modelled as a thin hyperelastic soft body on which \textit{active receptors} with density $\xi$ and \textit{transporters} (i.e. MRPs-Multidrug Resistant Proteins) with density $\zeta$ mediate the communication between the extra-cellular and the intra-cellular environment. \color{black} Receptor-Ligand (RL) compounds and MRPs directly initiate different signaling pathways involved in different cell processes, regulated by the transport and compartmentation of chemicals such as cyclic-AMP. Measuring cAMP concentrations then allows to indirectly estimate the amount of surface proteins involved, as explained in Section \ref{ExpMeas}. There the experiments used to trace the levels of receptors and transporters are described. In the proposed model, the cell membrane is characterized according to the constitutive framework proposed by Deseri and Zurlo in \cite{Deseri:2008, Deseri:2013, Zurlo:2006}. Here, this is explicitly coupled with the dynamics of GPCRs and MRPs activation, apt to describe the growth of active domains on the membrane surface by means of flow rules that include the receptor-transporter interplay. This dynamics is modelled according to a theory of interspecific growth mechanics recently proposed by Fraldi and Carotenuto in \cite{fraldi2018cells, carotenuto2018} and applied to model the growth of solid tumors. Here, the same framework is adapted to model the kinetics and diffusion of transmembrane proteins triggered by ligand-binding and mediated by the membrane elasticity. \color{black} Upon utilizing this approach, the evolution equations for the species provide Lotka-Volterra-inspired interspecific terms aimed to model the response of GPCRs to \color{black} affine ligands and the growth of RL complex. The latter stimulate the activation of MRPs through the intracellular cAMP production. In turn, it is observed a \color{black} localization of islands of active receptors and transporters on thickened domains of the cell membrane, known as lipid rafts. Such remodelling phenomenon is then governed by kinetics, diffusion and energetics, where binding/unbinding events turn out to be strongly coupled to the mechanical response of the plasma membrane through the energetics. Indeed, the mechanical interaction between the membrane and the active proteins during their remodelling and the conformational changes of their domains can be traced through the growth of their corresponding fractions. \color{black} \\
\subsection{Membrane elasticity model for lipid bilayers}
\label{sectionB3}
In this Section the main results obtained in \cite{Deseri:2008, Deseri:2013, Zurlo:2006} are briefly recalled, and a schematic description of the approach followed in these papers is provided. The formulation of the model is based on the following assumptions:
\begin{itemize}
\item[(i)] the effects leading to a spontaneous or natural curvature of the bilayer are ignored, i.e. it is assumed that the natural configuration is flat;
\item[(ii)] the membrane kinematics is restricted to the class of normal preserving deformations, i.e. the normal vector $\mathbf{n}$ remains always normal to the mid-surface.
\end{itemize}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5 \columnwidth]{fig00_schemagen}
\caption{Schematic representation of the deformation of a prismatic, plate-like reference configuration $\mathcal{B}_0$ into the current configuration $\mathcal{B}$. The gray box depicts the space occupied by two lipid molecules, their volume being conserved during the deformation. Courtesy of \cite{Deseri:2013}.}
\label{figB:fig00_schemagen}
\end{figure}
Under these assumptions, a simplified version of the elastic energy related to the change of the membrane geometry has been obtained in \cite{Deseri:2008, Deseri:2013}, The deformation mapping is denoted by $\mathbf{y}=\textbf{\textit{f}}(\textit{x})$, while its gradient is $\textbf{F}=\nabla\mathbf{y}= \nabla_\textit{x}\,\textbf{\textit{f}}$. The stored Helmholtz free-energy can be then expressed as
\begin{equation}
\mathcal{E}(\textbf{\textit{f}})=\int_{V^0}\Psi(\textbf{\textit{f}})\,dV^0=\int_{\Omega} \int_{-h_0/2}^{h_0/2}\Psi(\textbf{\textit{f}})\,dz\,d\Omega,
\end{equation}
where $\Psi$ denotes the Hemholtz energy density per unit of referential volume. It is easy to recognize that the surface density energy has the form:
\begin{equation}
\label{eqB:en2d}
\psi(\ff) = \int_{-h_0/2}^{h_0/2}\Psi(\ff)\,dz.
\end{equation}
For a fixed temperature, the \textit{natural} configuration $\mathcal{B}_0$ of the membrane is assumed to coincide with a thin flat prismatic shape of homogeneous thickness $h_0$ in direction $\textbf{e}_3$, width $B$ in direction $\textbf{e}_2$ and length $L$ in direction $\textbf{e}_1$, with ordered phase $L_o$ (see Figure \ref{figB:fig00_schemagen}). The membrane geometry is split into two entities, the two-dimensional mid-plane and the thickness, hence, the material particles $\textbf{\textit{x}}\,\in\mathcal{B}_0$ are described by:
\begin{equation}
\textbf{\textit{x}}=\textbf{x}+z\textbf{e}_3,
\end{equation}
where $\textbf{x}=x\,\textbf{e}_1+y\,\textbf{e}_2$ denotes locations of a flat mid-surface $\Omega$, and $z$ spans the whole thickness. The reference membrane mid-surface $\theta$ corresponds to $z=0$, and its edges are defined by $x=\pm L/2$ and $y=\pm B/2$. Experimental evidences suggest that lipid membranes exhibit the so-called \textit{in-plane fluidity}, i.e. the absence of viscosity does not allow for sustaining shear stress in planes perpendicular to $\textbf{e}_3$. Based on this constitutive assumption, it is possibile to find a relationship for the energy density $\Psi$ depending on the three invariants of the deformation gradient, namely $\{\tilde{J}(\textbf{\textit{x}}),\,\det \textbf{F}(\textbf{\textit{x}}),\,\bar{\phi}(\textbf{\textit{x}})\}$.
In particular, the quantity $\tilde{J}(\textbf{\textit{x}}) := \sqrt{\det \, C_{\alpha \beta}}$ represents the areal stretch of planes perpendicular to the direction $\textbf{e}_3$, after setting $\textbf{C} = \textbf{F}^{T}\textbf{F }= C_{ij} \textbf{e}_i \otimes\textbf{e}_j$, which is the usual Cauchy-Green stretch tensor. The remaining invariants are the volume change, $\det\textbf{F}$, and the $\bar{\phi}(\textbf{\textit{x}}) = h(\textbf{\textit{x}})/h_0$ thickness stretch in direction $\textbf{e}_3$, respectively \cite{Deseri:2008, Deseri:2013}.
In \cite{Lipowsky:1992, Sackmann:1995} it is suggested that the volume of lipid membranes does not significantly change. Some authors showed also that the volume of biological membranes can be assumed constant at several values of temperature \cite{Goldstein:1989, Owicki:1978}. The following Ansatz (see Figure \ref{figB:fig00_schemagen}) is assumed for describing the geometrical changes of the membrane:
\begin{equation}
\label{eqB:def}
\textbf{\textit{f}}(\textbf{\textit{x}}) = \textbf{\textit{g}}(\textbf{x}) + z\phi(\textbf{x}) \,\textbf{n}(\textbf{x})
\end{equation}
where \color{black} $\textbf{\textit{g}}(\circ)$ maps the mid-plane $\Omega$ of the membrane from the natural configuration $\mathcal{B}_0$ to the current mid-surface of the membrane (i.e $\theta=\textbf{\textit{g}}(\Omega)$), $\textbf{n}$ denotes the outward normal to $\theta$, and $\phi(\textbf{x})=\bar{\phi}(\textbf{x})=h(\textbf{x})/h_0$ is the thickness stretch. The latter is defined as the ratio of the current thickness $h$ over the reference value $h_0$. It is worth noting that a Monge representation $\bar{\textbf{\textit{g}}}(x,y):=\textbf{\textit{g}}(\textbf{x})$ for $\theta$ arises upon writing the position of material particles $\textbf{x} \in \Omega$ in components with respect to the pair $\textbf{e}_1, \, \textbf{e}_2$ of orthonormal vectors chosen above. The assumption of the Ansatz \eqref{eqB:def} is then coupled with a quasi-incompressibility constraint in the following form:
\begin{equation}
\label{eqB:quasinc}
\det\textbf{F}(\textbf{x},0)=\tilde{J}(\textbf{x},0) \phi(\textbf{x}) = 1.
\end{equation}
It is worth noting that such quasi-incompressibility is a first-oder approximation of the full iscochoricity requirement. An explicit expansion of equation \eqref{eqB:en2d} in powers of the reference thickness $h_0$ (see \cite{Zurlo:2006, Deseri:2008, Deseri:2013}) can be done by taking into account the choice of the constraint \eqref{eqB:quasinc}. Therefore, a restriction $\psi$ of the energy density $\Psi$ to $\Omega$ is considered by taking into account the requirement \eqref{eqB:quasinc}:
\begin{equation}
\label{eq:psiJ}
\psi(J)=\Psi(\tilde{J},\det\textbf{F},\bar{\phi}){\Big |}_{z=0}=\Psi(J,1,J^{-1}) ,
\end{equation}
where
\begin{equation}
J \, \, :=\, \tilde{J}(\textbf{x},0).
\label{J_def}
\end{equation}
Note that
\begin{equation}
\label{eq:J}
J=\left(\frac{h}{h_0}\right)^{-1}.
\end{equation}
In-plane fluidity, bulk incompressibility, and a dimension reduction \color{black} yield the following expression for the energy density of the lipid membrane $w$ per unit area in the reference configuration
\cite{Deseri:2008, Deseri:2013, deseri:2012}
\begin{equation}
\label{eqB:en2dexp}
w = \varphi(J) + k(J) H^2 + k_G(J) K + \alpha(J) \, ||(\mathop{\rm grad}\nolimits_{\theta} \widehat{J} )|_m||^2 ,
\end{equation}
\color{black} where $J$ accounts for the areal stretch of the membrane (see e.g. \cite{Wada:2015} for the importance of this quantity in lipid membranes), $H$ and $K$ are the mean and Gaussian curvatures of the mid-surface $\theta$, respectively, \color{black} $k(J)=\frac{h_0^2}{6}\tau'(J)$ and $k_G(J)=\frac{h_0^2}{12J} \tau(J)$ are the corresponding bending rigidities, $\varphi(J):= h_0 \, \psi(J)$ represents the \textit{local energy density} per unit area\color{black}, whereas
\begin{equation}
\label{eqB:alpha}
\alpha(J)=\frac{h_0^2}{24}\frac{\tau(J)}{J^3},
\end{equation}
where
\begin{equation}
\label{eq:tau_def}
\tau(J) = \varphi'(J).
\end{equation}
The quantity $\alpha(J)$ is a higher order extensional modules and it is related to the nonlocal part of the energy density. In equation \eqref{eqB:en2dexp}, $\hat{J}$ is the spatial description of $J$ , defined by the composition $\hat{J} \circ g = J$,
$\mathop{\rm grad}\nolimits_{\theta}$ is the gradient with respect to points of the current mid-surface $\theta$, while $(\cdot)_{\scriptscriptstyle m}$ gives its material description.
The term \eqref{eqB:en2dexp} recovers exactly the Helfrich energy \cite{Helfrich:1973} if $J$ is constant. The main ingredient of the two-dimensional membrane model derived in \cite{Deseri:2008, Deseri:2013} is the surface Helmholtz energy $\varphi(J)$, which regulates the in-plane stretching behavior of the membrane and describes the phase transition phenomena taking place in lipid bilayers.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\columnwidth]{figB_phi_tau}
\caption{The stretching energy $\varphi(J)$ (top) adapted from \cite{Goldstein:1989} for a temperature $T\sim 30^{\circ}C$, with related local stress $\varphi'(J) = \tau(J)$ (bottom). The areal stretch $J_o=1$ corresponds to the unstressed reference configuration $\mathcal{B}_0$, while $\Sigma_M$ indicates the Maxwell stress.}
\label{figB:fig_dz_energy}
\end{figure}
A Landau expansion for $\varphi(J)$ can be constructed because of the lack of more precise information (see, {\it e.g.}, \cite{Falkovitz:1982, Goldstein:1989, Komura:2004, Owicki:1978, Owicki:1979}). This approach results to be useful as it allows for relating the coefficients of the polynomial expansions to measurable quantities, such as the transition temperature, the latent heat and the order parameter jump (see \cite{Goldstein:1989} and the treatise \cite{Sackmann:1995} for a detailed discussion).
Under the current assumptions on the natural configuration $\mathcal{B}_0$ the Landau expansion of stretching energy \cite{Zurlo:2006} takes the form:
\begin{equation}
\label{eqB:phi}
\varphi(J) = a_0 + a_1 J + a_2 J^2 + a_3 J^3 +a_4 J^4,
\end{equation}
where the parameters $a_i\,\,(i=0 \div 4)$ depend on temperature and chemical composition. Because of the lack of avalaible experimental data, these parameters have been calibrated in \cite{Zurlo:2006} by considering experimental estimates provided by \cite{Goldstein:1989, Komura:2004, Komura:2006}. For a temperature $T\sim 30^{\circ}C$ we have:
\begin{equation}
\label{eqB:ai}
\begin{array}{c}
a_0=2.03 , \quad a_1=-7.1, \quad a_2=9.23 \\
a_3=-5.3, \quad a_4=1.13
\end{array}
\end{equation}
expressed in $[J][m]^{-2}$. The related local constitutive stress $\Sigma$ \eqref{eq:tau_def} shows the typical S-shaped form, as shown in Figure \ref{figB:fig_dz_energy}. The locations where the maximum $\Sigma_{\max}$ and the minimum $\Sigma_{\min}$ of the local stress occur are denoted by $J_{\max}$ and $J_{\min}$, respectively, and highlighted with squares, while the value of the Maxwell stress (determined by the equal area rule) is drawn as a straight horizontal dark-grey line (see e.g. \cite{Coleman:1988}). Whenever a generic stress $\Sigma$ is considered, it is also possible to identify three intersections between such a stress and the local stress curve. In the case of the Maxwell stress, these three intersections are denoted with black circles, and the corresponding areal stretch values are called $J_*$, $J_m$ and $J^*$ from left to right. As an example, Table \ref{tabB:coleman} collects the numerical values of these quantities for $\varphi(J)$ obtained by using the coefficients in \eqref{eqB:ai}.
\begin{table}[htbp]
\centering
\caption{\textbf{Table}\ref{tabB:coleman}. Characteristic values of the membrane stretching energy at $T\sim 30^{\circ}$. Stress expressed as $[\Sigma] = [J/m^2] \, \times 10^{-3}$}
\label{tabB:coleman}
\begin{tabular}{cccccccc}
\hline
$\Sigma_{M} $ & $\Sigma_{\max} $ & $\Sigma_{\min} $ & $J_*$ & $J_m$ & $J^*$ &$J_{\max}$ & $J_{\min}$ \\
\hline
5.922 & 10.855 & 0.989 & 1.025& 1.167& 1.308& 1.085 & 1.248 \\
\hline
\end{tabular}
\end{table}
\subsubsection{Planar case}\label{sec:planarcase}
The study of the equilibrium for a planar lipid membrane described by the energy \eqref{eqB:en2dexp} permits to elucidate the emergence of thickness inhomogeneities in the membrane. Moreover, this simple energetics allows one to calculate the corresponding rigidities and the shape of the boundary layer between the ordered and disordered phases. By fixing temperature, energy density coefficients are fixed to those ones in Table \ref{tabB:coleman}.
Whenever no curvature changes are experienced by the lipid bilayer, in the plane strain approximation the elastic energy density in \eqref{eqB:en2dexp} takes the form:
\begin{equation}
\label{eqB:energy_gamma_1D}
\psi(J) = \varphi\left(J\right) - \frac{1}{2} \gamma(J) J_x^2.
\end{equation}
We note that the functions
\begin{equation}
\label{eqB:gamma}
\gamma(J) = -2\,\alpha(J).
\end{equation}
and $\tau(J)$, defined by \eqref{eq:tau_def}, can be interpreted as \textit{transition} and \textit{stress-like} functions, respectively, as the former drives the boundary layer wherever transitions between two phases occur. According to the geometry introduced above, the three-dimensional membrane deformation is further restricted with respect to equation \eqref{eqB:def} as follow:
\begin{equation}
\label{eqB:def2}
\textbf{\textit{f}}(\textbf{\textit{x}})=g(x)\textbf{e}_1+y\textbf{e}_2+z\phi(x)\textbf{e}_3 ,
\end{equation}
so that the width $B$ is kept constant and its gradient takes the following form
\begin{equation}
\label{eqB:F}
\textbf{F}=\nabla\textbf{\textit{f}}(\textbf{\textit{x}})=
\left[
\begin{array}{ccc}
g_{x} & 0 & 0 \\
0 & 1 & 0 \\
z\phi_{x} & 0 & \phi
\end{array}
\right],
\end{equation}
where the subscript $x$ denotes differentiation with respect to $x$. The displacement component along $\textbf{e}_1$ is $u(x)=g(x) - x$. After setting
\begin{equation}
\label{eqB:lambda}
\lambda(x)= g_{x}(x),
\end{equation}
where the function $g$ maps the mid plane of the membrane from $\mathcal{B}_0$ to $\mathcal{B}$, see Figure \ref{figB:fig00_schemagen}.
The plane strain approximation yields
\begin{equation}\label{ThicknessChange}
\phi=\lambda^{-1}=\frac{h}{h_0},
\end{equation}
so that the membrane deformation is completely determined by
\begin{equation}\label{J=lambda}
J = \lambda.
\end{equation}
\subsection{Membrane remodelling due to GPCR growth via densification}
\color{black} A thermodynamically consistent framework is provided in this section with the aim of evaluating the chemo-mechanical coupling between transmembrane proteins and membrane elasticity. To this end, a total Helmoltz free energy per unit volume is introduced, say $W^*$, accounting for both the response of the lipid bilayer and the energetics associated to the proteins changes. In particular, the rearrangement
of protein domains on the cell membrane is modelled as a process of growth via pure densification \cite{Lubarda}, i.e. by considering the effect of the variation of the density of activated proteins on the free-energy of the membrane. In fact, by considering a membrane element with mass $dm^0=\rho^0 dV^0$ in the inactive (virgin) configuration, as displayed in figure \ref{fig:Stress00}, the potential activation of previously dormant receptors on the membrane is likely assumed not to produce mass variation.
However, the submacroscopic changes following the formation of receptor-ligand binding induce remodelling and conformational changes of the bonded receptors across the membrane. This is because the variation of the density and of the conformations of those receptors determine the re-organization of the surrounding lipids (see the \hyperref[{SB2}]{Appendix C}).
By denoting with the superscript $a$ the active (reference) configuration in which protein activation occurs (see the inermediate global configuration of the membrane in figure \ref{fig:Stress00}), the remodelling $K_r$ affecting the mass element can be written as follows (see the \hyperref[{KrCalc}]{Appendix A}): \color{black}
\begin{equation}\label{RemodelingTerm}
\rho^0\, dV^0=\rho^a \,dV^a \rightarrow \,
K_r := \frac{dV^a}{dV^0}=\frac{\rho^0}{\rho^a}=\frac{1+\kappa_u(1+\small{\Sigma}_i\,n_i^0\,\Delta_A)(\rho_p/\rho_l -1)}{1+\kappa_u(1+\small{\Sigma}_i\,n_i\,\Delta_A)(\varrho_p/\varrho_l -1)}.
\end{equation}
This factor is calculated by considering that the density of the heterogeneous medium in a certain configuration is the sum of the true densities weighted by the respective fractions, i.e. $\rho=\varrho_l\phi_l+\varrho_i\,n_i$ (see \cite{fraldi2018cells}), $\{n_i\}|_{\{i=1, \, 2\}}=\{\zeta, \xi\}$ representing the relative fractions of protein species (summation over \textit{i}) and $\phi_l$ the complementary lipid fraction, the additional superscript $0$ being used for indicating the initial ones. The other constants read \color{black} as follows \color{black}:
\begin{itemize}
\item $\kappa_u=N_{max} A_u/A$, i.e. the total areal fraction of inactive proteins ($N_{max}=1.13\times 10^7$ is the total number of potentially activating proteins, estimated from the experimental data, $A_u$ is the area of the inactive domain (with diameter of $\approx 4 nm$ \cite{gurevich2018gpcrs}, $A$ is the surface of the cell for which a radius of $30 \mu m$ was considered)
\item $\Delta_A= (A_a/A_u-1)$ is the relative change in area of protein domains passing from inactive to active state, the sight of the latter exhibiting a diameter of $\approx 5 nm$ \cite{gurevich2018gpcrs}
\item the terms $\varrho_p$ and $\varrho_l$ indicate the specific true density of the transmembrane proteins and of the bilayer lipids, respectively, both assumed constants and calculated from their molecular weights and sizes. In particular, $\varrho_p=100 kDa/h_0 \pi (2.5 nm)^2$ and $\varrho_l=2\times(1 kDa/h_0 \pi (0.5 nm)^2)$ \cite{cevc1993, carpenter2016, mendelson2013, HUME2014363}.
\end{itemize}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\columnwidth]{RaftKinNew.png}
\caption{Schematic picture of protein activation and lipid membrane remodelling. A density change of active receptors drives the system towards an active configuration in which the lipid bilayer shrinks around active domains, this causing membrane deformation and thickening in the form of lipid rafts. Unlike usual description of finite deformations and multiphysics, all the three configurations are global. This is justified through a non-standard multiscale geometric framework provided by the theory of \textit{Structured Deformations} (see e.g. \cite{deseri:2003}, \cite{OwenMix:2008}, \cite{Deseri:2010}, \cite{deseri:2015}, \cite{deseri:2019}, \cite{palumbo:2018}). Here, each material particle of the body in the virgin configuration (displayed in the top-left side of this figure) gets first mapped into an intermediate region (center of the figure) through a pair of smooth bijections $(\bf{i}, \bf{K})$ (where $\bf{i}$ stands for the vector-valued identity and the tensor $\bf{K}$ accounts for all the submacroscopic changes). Indeed, it can be shown that $K_r = det \bf{K}$ and, hence, this invariant of $\bf{K}$ accounts for remodelling. Secondly, each neighborhood of each particle located in this intermediate global configuration gets classically deformed through a pair $(\bf{y}, \nabla\bf{y})$, where $\bf{y}$ is a smooth deformation field from the intermediate global configuration onto the current one, while the second item in the pair is the deformation gradient from the intermediate global configuration. Features of Structured Deformations in connection with the examined biological systems are discussed in more details in \hyperref[{SB2}]{Appendix C}. There, an alternative scheme of the kinematics is given in Fig.\ref{fig:StructuredDeformations}.
}
\label{fig:Stress00}
\end{figure}
\bigskip
\color{black} In the light of \eqref{RemodelingTerm}, the
energy stored through remodelling, stretching and species redistribution can be expressed in terms of the total (three-dimensional) Helmoltz free energy $W^*$ of the membrane in the natural configuration (i.e. the reference configuration in fig. \ref{fig:Stress00}) as:
\begin{equation}\label{StretchingEnergy}
\int_{V^a} W^* dV^a=\int_{V^0} K_r\,W^* dV^0=\int_{V^0} W^*_0 dV^0
\end{equation}
where $W^*_0$ is instead the three-dimensional energy density relative to the
virgin configuration. In analogy with equation \eqref{eqB:en2d}, because of the thinnes of the cell membrane relative to its in-plane sizes, a dimension reduction is performed to obtain a corresponding effective energy density $W_0$ per unit area in the virgin configuration:
\begin{equation}\label{position}
W_0=K_r\,W(\lambda,\nabla\lambda,n_i)=\int_{-h_0/2}^{h_0/2}W^*_0 dz=h_0\,K_r\,W^*(\lambda,\nabla\lambda,n_i).
\end{equation}
\subsubsection{Mass conservation equations and interspecific terms}
The one-dimensional balance of mass for each constituents can be written as
\begin{equation}\label{massbalance}
\dot{n}_i+M_{i,x}=\Gamma_i
\end{equation}
where the dot notation denotes the material time derivative, $M_i$ describes the material species
flux and $\Gamma_i$ is the rate term describing the evolutionary behavior of each species. This is done by taking into account intrinsic rates as well as mutual interactions with other fields characterizing the system by means of suitable coupling terms. The present model mainly focuses on the active receptors and transport proteins on the cell membrane. In particular, in the presence of a specific
ligand, binding and activation GPCRs kindles cytoplasmatic signaling pathways. The latter allow for both regulating some intracellular processes and communicating with the extra-cellular environment by expelling molecules
through the activation of specific transport proteins, such as MRPs.
This coupled mechanism can be followed in practice by measuring the levels of intracellular and extracellular concentrations of cAMP, involved in many physiological processes of the cell. In particular, intracellular cAMP production follows the response of the GPCRs to certain ligands, such as epinephrine. The level of cytoplasmatic cAMP is modulated by MRPs that permit the efflux of cAMP from the interior of the cell to the extracellular fluid. In this way, MRPs control cell homeostasis and determine an agonist interaction with membrane receptors.
In order to trace the interplay of the transmembrane proteins, we introduce mass balances for two selected species, $n_1=\xi$, representing the G proteins, and $n_2=\zeta$, namely the transport proteins. The main physiological aspects associated to their activity have been described by means of the following interspecific (Volterra-Lotka like) equations:
\begin{align}
&\dot{\xi}+M_{\xi,x} = \alpha_\xi \, \xi -\delta_\xi \, \xi -\beta_{\xi\zeta} \, \zeta \, \xi \label{massbalance2.1}\\
&\dot{\zeta}+M_{\zeta,x} = \beta_{\zeta\xi} \, \xi \, \zeta -\delta_\zeta \, \zeta . \label{massbalance2.2}
\end{align}
\paragraph{Physical meaning of the species rates parameters.} In equation \eqref{massbalance2.1}, the term $\alpha_\xi$ denotes \textit{intrinsic activation in response to a given ligand concentration}
(such as the epinephrine), modelled by means of a specific uptake function. In particular, it is likely assumed that the ligand precipitation rate obeys a generic Gamma distribution with probability density function $\Upsilon(t)=a\,t\,e^{-bt}$. The coefficients $a$ and $b$ are calibrated by means of experiments. By observing the time at which $\Upsilon$ is maximum -- i.e. $\dot{\Upsilon}(t_m)=0$ -- it readily follows that $b=t_m^{-1}$. Also, the total quantity of ligand precipitated over the experiment time $T$ then will be
\begin{equation}\label{Q}
Q=\int_0^T \Upsilon(t) dt= ab^{-2} (1- e^{-bT}(1+bT))
\end{equation}
that allows to estimate the parameter $a={Q}/(t_m^2-e^{-\frac{T}{t_m}}t_m\left(t_m+T\right))$. Given this, the coefficient $\alpha_\xi$ results the relative uptake function multiplied by a suitable binding constant:
\begin{equation}\label{alphaxi}
\alpha_\xi=k_b \, Q^{-1}\Upsilon(T).
\end{equation}
The coefficient $\delta_\xi$ denotes an \textit{intrinsic deactivation constant}, while the \textit{interspecific term} $\beta_{\xi \zeta}$ is introduced in the light that the efflux of extracellular cyclic Adenosyne Monophospate, $cAMP_e$ (whose amount is proportional to the active MRP on lipidic membrane through \eqref{eq:camptozeta}), is assumed to reduce the activity of receptor-ligand compounds by establishing chemical equilibrium between the extra- and the intra-cellular space. Dually, in equation \eqref{massbalance2.2}, the rate of activation of transport proteins is stimulated by the level of intracellular cyclic Adenosyne Monophospate, $cAMP_i$ (which is proportional to the activated G-proteins at time $t$ through \eqref{eq:camptoxi}). This is accounted for through the \textit{interspecific positive term} $\beta_{\zeta\xi}$ governing the growth of activating MRPs. Model parameters have been therefore estimated through the experimental evaluation of intracellular and extracellular cAMP concentrations (see \hyperref[{ExpMeas}]{Section 2}).
\subsection{The Helmoltz free-energy of the membrane proteins}
\label{sec:3P}
The
energy $W^*$ accounts for the hyperelastic contribution of the lipid membrane and for the Helmoltz free energy of the active proteins. The latter interact with the membrane by exchanging mechanical forces and showing remodelling changes during their activity. We start from the isothermal dissipation inequality. Here, we single out the contribution of mass transport of the
species over the generic elementary volume, say $M_i$ for the $i^{th}$ species, through the action of the correspoinding (thermodynamically conjugate) chemical potential $\mu_i^*$, as well as the presence of thermodynamic driving forces $F_r^*$ expending power against the rate of change of remodelling (see e.g. \cite{Lubarda, BMMB}). Hence, for an incompressible material and under isothermal assumptions, combined energy-entropy equations yield the following form of the second law of thermodynamics:
\begin{equation}\label{dissip}
\int_{V^0}{P^*\dot{\lambda}}\,dV^0 - \int_{V^0}\left(\mu^*_i\,M_i\right)_{x}\,dV^0+\int_{V^0}F_r^* \dot{K}_r\,dV^0 \geq \frac{d}{dt}\int_{V^0}W^*_0 \, dV^0
\end{equation}
\color{black} where repeated index $i$ means summation over that index. By appealing to relation \eqref{StretchingEnergy}, in agreement with the dimension reduction performed in \eqref{position}, quantities $X$ per unit area can be introduced by considering that $\int_{V^0}X^*dV^0=\int_\Omega X\, d\Omega$, with $X=h_0 X^*$ for any field $X^*$ per unit volume in the virgin configuration.
\noindent Provided that
\begin{equation}
\label{energy-reduced-reference}
W=W(\lambda,\lambda_x,n_i)
\end{equation}
from \eqref{position}, one also obtains the dimensionally reduced version of \eqref{dissip}:
\begin{equation}\label{dissip2}
\resizebox{.9 \textwidth}{!} {$
\int_{\Omega}{P\dot{\lambda}}\,dV^0 - \int_{\Omega}\left(\mu_i\,M_i\right)_{x}\,d\Omega+\int_{\Omega}F_r \dot{K}_r\,d\Omega \geq \int_{\Omega}\left[K_r\left(\frac{\partial W}{\partial \lambda}\dot{\lambda}+\frac{\partial W}{\partial \lambda_x}\dot{\lambda}_x +\frac{\partial W}{\partial n_i}\dot{n}_i\right)+W\,\dot{K}_r\right] \, d\Omega .
$}
\end{equation}
By substituting the mass balance \eqref{massbalance} in the second integral of the left-hand side and by integrating by parts the second term on the right-hand side, equation \eqref{dissip2} can be rearranged as follows:
\begin{equation}\label{dissip3}
\resizebox{.9 \textwidth}{!} {$
\int_{\Omega}{\left[P\dot{\lambda}-\mu_{i,x}M_i +\mu_i\,\dot{n}_i -\mu_i\,\Gamma_i +F_r\dot{K}_r\right]}\,d\Omega
\geq \left[\frac{\partial W_0}{\partial \lambda_x}\dot{\lambda}\right]_{\partial \Omega}+\int_{\Omega}\left\{K_r\left[\left(\frac{\partial W}{\partial \lambda}-\frac{\partial}{\partial x}\left(\frac{\partial W}{\partial \lambda_x}\right)\right)\dot{\lambda}+\frac{\partial W}{\partial n_i}\dot{n}_i\right]+W\,\dot{K}_r\right\} \, dV^0
$}
\end{equation}
Provided the condition
\begin{equation}\label{conditionstress}
\left[\frac{\partial W}{\partial \lambda_x}\dot{\lambda}\right]_{\partial \Omega}= 0
\end{equation}
is satisfied, grouping and localization of the terms in \eqref{dissip3} yields the following inequality
\begin{align}\label{localdissip}
\left[P-K_r\left(\frac{\partial W}{\partial \lambda}-\frac{\partial}{\partial x}\frac{\partial W}{\partial \lambda_x}\right)\right]\dot{\lambda} +\left[\mu_i - K_r \frac{\partial W}{\partial n_i}\right]\dot{n}_i + \left[F_r-W\right]\dot{K}_r -\mu_{i,x}M_i -\mu_i\,\Gamma_i \, \geq \, 0.
\end{align}
Following the standard Coleman and Noll's procedure, it is then possible to obtain the following constitutive relations for the Piola-Kirchhoff stress, the chemical potential and the remodelling driving force, which turns out to be entirely identifiable with the mechanical energy involved in the process:
\begin{align}\label{ConstEqual}
&P=K_r\left(\frac{\partial W}{\partial \lambda}-\frac{\partial}{\partial x}\frac{\partial W}{\partial \lambda_x}\right), \nonumber\\
&\mu_i=K_r\frac{\partial W}{\partial n_i},\\
&F_r=W. \nonumber
\end{align}
Inequality \eqref{localdissip2} thus reduces to:
\begin{equation}\label{localdissip2}
-\mu_{i,x}M_i -\mu_i\,\Gamma_i\geq 0
\end{equation}
that can be fulfilled by expressing proteins movement term as proportional to the gradient of the chemical potential, i.e. $M_i =-L_i\,n_i\,\mu_{i,x}$, where $L_i\geq0$ are suitable mobilities.
A reduced dissipation inequality in the form $\mu_i\,\Gamma_i \leq 0$ remains to be fulfilled in order to ensure thermodynamic compatibility. In otehr words, the evolution of the system is thermodynamically consistent when the chemical potential becomes negative in the presence of increasing binding proteins on the membrane. An example of a similar instance was encountered in \cite{Gao:2005b} in which an \textit{ad hoc} flow rule respecting such condition was introduced.
\subsection{Constitutive equations for the membrane}
\label{sec:3M}
In presence of the uniaxial kinematics discussed in Section \ref{sec:planarcase}, the problem is entirely governed by a one-dimensional energetics.
In agreement with \eqref{position}, a complete representation formula for the Helmoltz free energy density introduced in \eqref{energy-reduced-reference} is sought. Indeed $W=W(\lambda,\lambda_x, n_i)$ represents the energy per unit area in the active (reference) configuration
for which no complete expression is yet available. With the aim of providing an explicit representation for
$W$,
an additive decompositon is assumed between the (hyperelastic)energy of the membrane and a second potential associated to the energetics of the transmembrane proteins.
This assumption is stated in the following form:
\begin{equation}\label{W01D}
W=W_{hyp}(\lambda,\lambda_x)-W_{n_i}(\phi,n_i)=W_{hyp}(\lambda,\lambda_x)-W_{n_i}(\lambda^{-1},n_i).
\end{equation}
it is worth notning that the second term account for the possibility
that activation and conformational changes of GPCRs may affect the membrane thickness, measured through $\lambda^{-1}$. The total differential of $W$ thus gives:
\begin{equation}\label{W01Dvar}
dW=\frac{\partial W}{\partial \lambda} d\lambda+\frac{\partial W}{\partial \lambda_x} d\lambda_x + \frac{\partial W}{\partial n_i} dn_i= \tilde{P}\,d\lambda + \tilde{\mu}_i\, dn_i.
\end{equation}
Because of relations \eqref{ConstEqual}, the following equations for the stress and for the chemical potentials relative to the active (reference) configurations follow:
\begin{align}
&\tilde{P}=K_r^{-1}P=\frac{\partial W}{\partial \lambda}-\frac{\partial}{\partial x}\frac{\partial W}{\partial \lambda_x}\label{costPmu1}\\
&\tilde{\mu}_i=K_r^{-1}\mu_i= \frac{\partial W}{\partial n_i}=-\frac{\partial W_{n_i}}{\partial n_i}.\label{costPmu2}
\end{align}
A Legendre transformation of \eqref{W01Dvar} can now be performed by subtracting the variation $d(n_i\tilde{\mu}_i)=\tilde{\mu}_i\, dn_i+ n_i\,d\tilde{\mu}_i$, by obtaining the coupled potential \color{black}
\begin{equation}\label{W01Dvar2}
d\hat{W}=d(W_{hyp}(\lambda,\lambda_x)-\hat{W}_{\mu_i}(\lambda^{-1},\mu_i))= \tilde{P}\,d\lambda - n_i\,d\tilde{\mu}_i
\end{equation}
from which it follows that
\begin{equation}\label{niW}
n_i=-\frac{\partial \hat{W}}{\partial \tilde{\mu}_i}=\frac{\partial \hat{W}_{\mu_i}}{\partial \tilde{\mu}_i}.
\end{equation}
Because $W_{hyp}$ is independent of the chemical potentials \color{black} and since $\hat{W}_{\mu_i}$ depends on the variables $\lambda^{-1}$ and $\tilde{\mu}_i$, the latter relation also implies that \color{black}
\begin{align}\label{dniW}
dn_i&=-d\left(\frac{\partial \hat{W}}{\partial \tilde{\mu}_i}\right)=\frac{1}{\lambda^{-2}}\frac{\partial\tilde{P}}{\partial \tilde{\mu}_i}\,d\lambda^{-1} - \frac{\partial^2\hat{W}}{\partial \tilde{\mu}_i^2}\,d\tilde{\mu}_i= - c_{\lambda_i}\,d\lambda^{-1} + c_{\mu_i}d\tilde{\mu}_i.
\end{align}
\color{black}
\noindent Herein, the coefficient $c_{\lambda_i}$ and $c_{\mu_i}$ relate the variation of the $i^{th}$ species density $n_i$ to the change of the membrane stretch and of the associated chemical potential $\tilde{\mu}_i$.
It is assumed that the terms $c_{\lambda_i}$ and $c_{\mu_i}$ can be modelled as first order variation coefficients that depend on the current amount of active proteins, i.e. $c_{\lambda_i}\approx n_i \overline{c}_{\lambda_i}$ and, in the light of the antagonism \eqref{localdissip2}, $c_{\mu_i}\approx -n_i \overline{c}_{\mu_i}$. Under this constitutive assumption, equation \eqref{dniW} rewrites as:
\begin{equation}\label{dniW2}
\frac{dn_i}{n_i}= - \overline{c}_{\lambda_i}\,d\lambda^{-1} - \overline{c}_{\mu_i}d\tilde{\mu}_i
\end{equation}
Therefore, \color{black}
integration of Equation \eqref{dniW2} leads to the following relation:\color{black}
\begin{equation}\label{niRel}
-\log \frac{n_i}{n_i^0}\simeq \, \overline{c}_{\lambda_i}\,\left(\lambda^{-1}-1\right) +\overline{c}_{\mu_i}\left(\tilde{\mu}_i-\tilde{\mu}_i^0\right)
\end{equation}
which allows for evaluating the chemical potential of the species as follows
\begin{equation}\label{ChemPot}
\tilde{\mu}_i=\tilde{\mu}_i^0 - \omega_i \, \left(\lambda^{-1}-1\right) -\eta_i \, \log \frac{n_i}{n_i^0},
\end{equation}
\color{black} after setting
\begin{align}
&\omega_i := \overline{c}_{\lambda_i}/\overline{c}_{\mu_i},\notag\\
\label{Chemical_Potential_Coefficients}\\
&\eta_i := 1/\overline{c}_{\mu_i}. \notag
\end{align}
Equation \eqref{costPmu2} in combination with \eqref{ChemPot} then allows for finding the following (normalized) expression for the potential $W_{n_i}$:
\begin{equation}\label{Wnidef}
W_{n_i}(\lambda^{-1},n_i)=-\int_{n_i^0}^{n_i}\tilde{\mu}_i\,dn_i= \eta_i \, n_i \, \log \frac{n_i}{n_i^0}+(n_i-n_i^0)\left(\omega_i\left(\lambda^{-1}-1\right) -\tilde{\mu}_{i}^0 -\eta_i \right).
\end{equation}
This relation exhibits a Boltzmann-type entropic contribution related to the chemical species coupled with a mechanical term. In \hyperref[{lateralpress}]{Appendix B} it has been shown that the latter represents the specific work done by the lateral pressure between the protein domains and the lipids in the cell membrane during conformational changes.
Representation formula \eqref{Wnidef} for $W_{hyp}$ permits to obtain the following relationships for the stress and the chemical potentials:
\begin{align}
&P=K_r\left(\frac{\partial W_{hyp}}{\partial \lambda}-\frac{\partial}{\partial x}\frac{\partial W_{hyp}}{\partial \lambda_x}-\frac{\partial W_{n_i}}{\partial \lambda}\right)=\notag\\
&\quad=K_r\left(\frac{\partial W_{hyp}}{\partial \lambda}-\frac{\partial}{\partial x}\frac{\partial W_{hyp}}{\partial \lambda_x}-\omega_i \frac{(n_i^0-n_i)}{\lambda^2}\right)=P_{eff}-P_{n}, \label{Pmudef1}\\
&\mu_i = K_r \tilde{\mu}_i=K_r\left(\tilde{\mu}_i^0 - \omega_i \, \left(\lambda^{-1}-1\right) -\eta_i \, \log \frac{n_i}{n_i^0}\right). \label{Pmudef2}
\end{align}
\color{black}
In equation \eqref{Pmudef1}, $P$ can be interpreted as the membrane net stress, while $P_{eff}$ and $P_n$ respectively represent the effective (hyperelastic) response of membrane and the \textit{chemical stress} due to protein specific work. \color{black} Indeed, the assumed correlation between the potential associated to transmembrane proteins activation and lipid bilayer thickening has been introduced in equation \eqref{W01D}. It can be noted that each term $\omega_i (n_i^0-n_i)/ \lambda^2$ forming $P_n$ relates to the change of the corresponding specific work exerted by the lateral pressure arising in species-lipid interactions. In \hyperref[{lateralpress}]{Appendix C} this is explicitly worked out for the active receptors through a bottom-up approach that considers the surface interactions between receptors and lipids, thereby giving a mechobiological explanation to recently observed experimental findings.\color{black} The effective response of the lipid membrane is instead fully identifiable with the hyperelastic law \eqref{eqB:energy_gamma_1D} for $J=\lambda$, here reported for convenience:
\begin{equation}\label{DZpot}
W_{hyp}(\lambda) := \psi(\lambda)=\varphi(\lambda) -\frac{1}{2}\gamma(\lambda)(\lambda_x)^2
\end{equation}
\subsection{Mechanical equilibrium of the biological cell membrane}
\label{sectionD0}
In this Section, the use of total free-energy \eqref{W01D} expressed by means of \eqref{Wnidef} and \eqref{DZpot} as a simplified energy governing the behaviour of biological membranes allows for searching the set of mechanical equilibria. To this end, the variational derivative of the resulting Gibbs free energy with respect to the in-plane displacement $u$ will be computed in order to isolate configurations corresponding to its stationary points. The Gibbs free energy \color{black}, here denoted by $\mathcal{E}$ \color{black} is obtained as \color{black} the difference between the Helmholtz one $\mathcal{H}$, defined by \eqref{eq:energy_basic}$_2$ below,\color{black} minus the work $\mathcal{L}$ done by the external load, i.e
\begin{equation}
\label{eq:energy_basic}
\mathcal{E} :=\mathcal{H} - \mathcal{L}, \qquad \mathcal{H}:=\int_\Omega W_0 d\Omega
\end{equation}
The work $\mathcal{L}$ results from the presence of two possible mechanical agents, a traction $\Sigma(t)$ acting against the displacement, and a hyperstress $H(t)$ acting against the gradient of the displacement at the boundary, i.e.
\begin{equation}
\mathcal{L} = \left[\Sigma\,u + H\, u_x\right]_{\partial\Omega},
\end{equation}
\color{black} Because $u(x,t)=g(x,t)-x$, since \eqref{eqB:lambda} holds, a compatibility relation between the stretch and the derivative of the displacement is obtained $\lambda= 1 + u_x$. Henceforth, upon considering \color{black} a perturbation $\delta u$,
the first variation of \eqref{eq:energy_basic} with respect to the displacement can be written
with the help of relation \eqref{dissip3}:
\begin{align}\label{first_variation}
\delta\mathcal{E}&= \left[\frac{\partial W_0}{\partial \lambda_x}\,\delta u_x\right]_{\partial \Omega} +\int_\Omega P\,\delta u_x\, d\Omega - \left[\Sigma\,u + H\, u_x\right]_{\partial\Omega}\notag\\
&= \left[\left(\frac{\partial W_0}{\partial \lambda_x}-H\right)\,\delta u_x\right]_{\partial \Omega}+\left[\left(P-\Sigma\right)\delta u\right] -\int_\Omega P_x \,\delta u\, d\Omega.
\end{align}
Stationary condition of the energy \eqref{first_variation} supplies the following equations governing the equilibrium of the membrane:
\begin{equation}\label{equilibrium}
\begin{cases}
P_x =0, \quad x\in\Omega\\
\text{either}\,\, P=\Sigma\,\,\text{or}\,\, \delta u=0, \quad x\in \partial \Omega\\
\text{either}\,\, \gamma\,u_{xx}=H\,\,\text{or}\,\, \delta u_x=0, \quad x\in \partial \Omega
\end{cases}
\end{equation}
in which equation \eqref{DZpot} and compatibility relation have been employed, while \color{black} the stress $P$ relative to the virgin configuration reads as follows:
\begin{equation}\label{StressDisp}
P=K_r\left[\tau(\lambda) + \frac{1}{2} \gamma'(\lambda) (u_{xx})^2+\gamma(\lambda) u_{xxx} - \omega_\xi \frac{\xi-\xi^0}{(1+u_x)^2} - \omega_\zeta \frac{\zeta-\zeta^0}{(1+u_x)^2}\right].
\end{equation}
Here \color{black} $\tau$ and $\gamma$ have been defined through \eqref{eq:tau_def} and\eqref{eqB:gamma} respectively.
\subsection{Coupling of the equations}
For the sake of illustration, prediction the formation of lipid rafts on a cell membrane is obtained in the case of plane strain. Coupling between the balance of linear momentum \eqref{equilibrium}$_1$ with the evolution laws provided by the mass balance of transmembrane proteins, \eqref{massbalance2.1} and \eqref{massbalance2.2}, yields the following set of governing equations:
\begin{equation}\label{Alleqs}
\begin{cases}
P=\Sigma\\
\dot{\xi}-\left(L_\xi\,\xi\,\mu_{\xi,x}\right)_x = \alpha_\xi \, \xi -\delta_\xi \, \xi -\beta_{\xi\zeta} \zeta \, \xi \\
\dot{\zeta}-\left(L_\zeta\,\zeta\,\mu_{\zeta,x}\right)_x= \beta_{\zeta\xi} \xi \, \zeta -\delta_\zeta \, \zeta .
\end{cases}
\end{equation}
Here, because of symmetry assumptions, $x\in\left[0, L\right]$ with $L$ set to $L=1000 h_0$ , while the time $t\in\left[0^+,t_{max}\right]$, $t_{max}$ being set to 1 hour on the basis the experiments duration (see Sect. appendix \ref{ExpMeas}). The unknowns of the problem are the displacement $u(x,t)$ and the protein fractions $\xi(x,t)$ and $\zeta(x,t)$, subjected to the boundary conditions $u(0,t)=0$, $u_{xx}(0,t)=u_{xx}(L,t)=0$, $\xi_x(0,t)=\xi_x(L,t)=0$ and $\zeta_x(0,t)=\zeta_x(L,t)=0$ in absence of hyperstress $H$ and protein species leakage at the boundaries. Initial conditions provide initial null displacement $u(x,0)=0$, while suitable initial distribution have been assigned to proteins. In particular, $\zeta(x,0)=\zeta^0$ is assumed constant, while, in order to take into account localization of lipid rafts on the membrane in response to prescribed ligand distributions, an initially non-homogeneous spatial profile of inactive GPCRs has been introduced. To do this, the initial condition $\xi(x,0)=\xi^0(x)$ is expressed by means of a suitable generating function:
\begin{equation}\label{Xi0}
\xi^0(x)=\xi_o+ \left(\xi_a-\xi_o\right) \sum_{m=0}^{{m}_{raft}}\Pi(x), \qquad \Pi(x)=\frac{E_{-}(x)}{1+E_{-}(x)}-\frac{E_{+}(x)}{1+E_{+}(x)}
\end{equation}
where $E_{\pm}(x)=\exp\left[-c_a\left(x-x_a \pm \Delta_a/2\right)\right]$, $x_a$ and $\Delta_a$ governing the positioning and the amplitude of the activating raft. In particular, the formation of a single raft region centred at the origin with amplitude $\Delta_a=400h_0$ and multiple (quasi-periodic, with ${m}_{raft}=4$) rafts regions with amplitude $\Delta_a=100h_0$ have been simulated in the sequel. Moreover, a stress-free (i.e. $\Sigma=0$) and a stress-prescribed case have been both considered, by imposing that the nominal traction equals the Maxwell stress of the membrane through the expression $\Sigma(t)=\Sigma_M(1-e^{-t})$, consistently with the initial undeformed conditions and with characteristic values of internal cell pressures \cite{cellpress}. Equations \eqref{Alleqs} have been numerically solved with the aid of the software Mathematica\textsuperscript{\textregistered} \cite{MathematicaProgram} through the method of lines. Spatial coordinates and model parameters, reported in \hyperref[{tab}]{Table} \ref{tab}, have been normalized with respect to the reference thickness $h_0$.
\begin{table}[]\label{tab}
\centering
\caption{\textbf{Table} \ref{tab}. Coefficients used in numerical simulations (e.d.= available experimental data).}
\label{tab}
\begin{tabular}{cccc}
\hline
Coefficient & Value {[}Unit{]} & Range {[}Unit{]} & Reference \\
\hline
$k_b$ & $5.18$ &$3.89-5.7$ & \cite{Bridge2018d1, Li2017} - e.d. \\
$\delta_\xi$ & $1.1\times10^{-3} \, s^{-1}$ &$(0.9-1.65)\times10^{-3} \, s^{-1}$ & \cite{Bridge2018d1} \\
$\beta_{\xi\zeta}$ & $4\times10^{-6}\, s^{-1}$ &$(3-6)\times10^{-6} \, s^{-1}$ & \cite{Bridge2018d1, Rich2001} \\
$\beta_{\zeta\xi}$ & $3.3\times 10^{-3}\, s^{-1}$ &$(2.5-4.15)\times10^{-3} \, s^{-1}$ & \cite{Saucerman2013, Rich2001, Agarwal2014} - e.d. \\
$\delta_\zeta$ & $10^{-7}$ &$10^{-8}-10^{-6}$ & assumed \\
$\zeta^0$ & $10^{-2}$ &$ $ & - \\
$\xi_o$ & $10^{-4}$ &$ $ & - \\
$\xi_a$ & $10^{-2}$ &$ $ & - \\
$L_i$ & $1.87\times10^{-17}\,{m^4 J^{-1}s^{-1}}$ &$(10^{-20}-10^{-15})\,{m^4 J^{-1}s^{-1}}$ & \cite{gurevich2018gpcrs, Kim2016} \\
\hline
\end{tabular}
\end{table}
\section{Results and discussion}
The outcomes of the numerical simulations derived by the biomechanical model introduced above show the lipid bilayer undergoing dynamic remodelling triggered by the evolution of transmembrane protein activity. In particular, here strain localization within the cell membrane leading to the coexistence of thicker (ordered $L_o$-phases) and thinner lipid zones (disordered $L_d$-phases) is predicted.
As shown in Figure
\ref{fig:Phi}, \color{black} the densification of lipids across the regions in which active GPCRs tend to cluster
is accompanied by the progressive formation of lipid rafts. In agreement with previous findings (see e.g. \cite{niemela2007}), on such sites a difference of about $0.9$ nm between raft and non-raft membrane zones is detected. \color{black} In particular, in stress-free simulations, there is a pure chemically-induced thickening of the membrane that forms a single raft and multiple rafts according to the analysed cases (Figure \ref{fig:Phi}). Under the same initial conditions, rafts also occur in the tensed membrane, where a transverse contraction and initial longitudinal elongation additionally take place in the $L_d$-phase as elastic response of the applied traction.
In Figure \ref{fig:XiPhi}\hyperref[{fig:XiPhi}]{A} the time-evolution of both the type of transmembrane protein fractions \textcolor{black}{occurring on a raft site} is reported. The interplay between RL proteins $\xi$ and transporters $\zeta$ is studied as a result of the introduced interspecific dynamics. More specifically, an initial growth of the GPCRs binding with ligands first occurs. Upon comparing this finding with the experimentally determined intracellular cAMP (cAMP\textsubscript{i}) normalized concentration, an estimate of the binding Receptor-Ligand (RL) compounds is provided according to \eqref{eq:camptoxi}. This is because GPCRs activation directly triggers the cell response by initiating the cAMP signaling pathway. Homeostatic levels of cAMP\textsubscript{i} are restored thanks to the activation of the Multidrug Resistant transport Proteins (MRPs) driving the efflux of part of the produced cAMP towards the extracellular environment.
It is worth noting that a very good agreement is found between the predicted MRPs fraction and the (normalized) increase of cAMP\textsubscript{e}.
Furthermore, Figure \ref{fig:XiPhi}\hyperref[{fig:XiPhi}]{A} displays the simultaneous evolution of the ordered (thicker) phases. The latter evolve with similar trends in both the unconfined and the tensed membranes, exception made for the load-induced shrinkage in the initial stage, when chemical stress $P_{n} \, =\, K_r \,\omega_i \, (n_i^0-n_i)/\lambda^2$ (introduced in \eqref{Pmudef1}) is not yet acting. At the time at which such stress is maximum, i.e $t\simeq 642 s$ in the simulations, a clear spatial correspondence between active domains and raft formation can be recognized in all situations, see Figure \ref{fig:XiPhi}\hyperref[{fig:XiPhi}]{B}. \textit{This predicts that active GPCRs localize on lipid rafts} as an effect of the chemo-mechanical coupling between:
\begin{itemize}
\item the chemical pressure arising from the deformation of the lipid bilayer;\\
\item the configurational changes of TMs domains, to which a decrease chemical potential as in Figure \ref{fig:Mu} is associated according to the above discussed thermodynamic arguments.
\end{itemize}
Although the model and the associated numerics is carried out for a one-dimensional problem, the findings are fully consistent with the experimentally observed clustering of ligand-binding receptors on lipid rafts across the cell membrane (see e.g. \cite{Bray:1998, foster2003, Chini:2004, Ostrom:2004, Zhang:2005, Becher:2005, Watkins:2011, Villar:2016} among others).
\textcolor{black}{In order to investigate how model coefficients influence the formation of lipid rafts in high coupling with the dynamics of active receptors, the robustness of the model was tested by considering the variation of both interspecific and chemo-mechanical coefficients. In this regard, more than two thousands numerical simulations were performed by assigning values within the ranges reported in Table \ref{tab} and by then evaluating the associated membrane thickening. In particular, as also shown in Figure \ref{fig:sensan}, analyses revealed that lipid rafts formation is mainly affected by the level of intrinsic rate of activation in response to ligand concentration $\alpha_\xi$ (through the uptake coefficient $k_b$), and the chemo-mechanical coefficient $\omega_\xi$, representing the specific mechanical work that GPCRs exert on the lipid membrane by inducing its thickening. }
As displayed in Figure \ref{fig:StressSM}, the imposed constant reference traction induces a nonhomogeneous distribution of the Cauchy stress, which relaxes within the rafts (namely across the thickened regions), thereby exhibiting values lower than the tension experienced by the thinner (hence disordered) phases. The binding kinetics and protein activation are instead responsible of a progressive, chemically-induced, compression of the lipids in the raft region. \color{black} Such compression decreases over time during the unbinding phase (see Figure \ref{fig:StressSM}). \color{black}
In stress free conditions, the chemical stress exerted by the activating proteins is entirely converted into effective membrane compression (see Figure \ref{fig:Stress0}). From a mechanical point of view, the stress relaxation renders ordered islands more compliant to demand TMs to undergo conformational changes, and hence GPCRs location on lipid rafts also implies a reduced energy expenditure during the protein structural re-organization.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\columnwidth]{Fig1Phi}
\caption{Lipid raft formation in stress-free (upper row) and tensed (lower row) membranes. Both the case of a single raft and multiple raft are analysed, the membrane thickening following the activation of transmembrane proteins.}
\label{fig:Phi}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\columnwidth]{Fig2XiPhi}
\caption{\textbf{A}. Time evolution of active receptors (blue) and transporters (red) in comparison with the experimental cAMP measurements. It is shown that the raft dynamically follows the binding/unbinding kinetics. \textbf{B}. High degree of localization between lipid raft formation and active GPCRs in single rafts (top) and multiple raft domains (bottom).}
\label{fig:XiPhi}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\columnwidth]{plotsens2}
\caption{(Left) Variation of the maximum thickening of the lipid membrane as a function of the intrinsic uptake coefficient $k_b$ and of the chemo-mechanical work coefficient $\omega_\xi$. The red spot indicates the couple $\{\overline{ k }_b , \overline{\omega}_\xi \}$ used in the simulation. (Right) Evolution of the membrane stretch for different values of coefficients $k_b$ and $\omega_\xi$. The red curve highlights the one obtained for the adopted parameters.}
\label{fig:sensan}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\columnwidth]{ChemPot}
\caption{Evolution of chemical potentials associated to protein species during binding/unbinding kinetics.}
\label{fig:Mu}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\columnwidth]{Fig4_Stress_SM2}
\caption{Stress in tensed membranes. The imposed constant reference traction induces a nonhomogeneous distribution of the Cauchy stress (upper row), which relaxes in thickened raft regions. In the lower row, the chemical stress generated by the protein species is reported, causing a progressive compression of the lipids inhabiting the raft domains.}
\label{fig:StressSM}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\columnwidth]{Fig3_Stress_new}
\caption{Effective and chemical stress balancing in stress-free remodelling membranes.}
\label{fig:Stress0}
\end{figure}
\section*{Conclusions}
\label{sectionH0}
\color{black} Predictions and interpretation of the findings provided in this paper allow for explaining \textit{why active receptors} of the GPCRs family \textit{tend to cluster on lipid rafts}. Furthermore, \textit{active receptors are predicted to enhance rafts formation}.
Such predictions are shown to be in relation with: \color{black}
i) stress reduction arising within rafts because of the presence of active receptors and transporters that need local compliant zones to be able to perform conformational changes;
ii) decrease of chemical potential owing (1) binding of the GPRCs with the ligand, thereby favoring the formation of higher density regions of RL complex and (2) uptaking of intracellular cAMP from the MRPs transporters.
It is shown how the mechanobiology involved at the level of the cell membrane does entail strong coupling among (a) mechanical equilibrium governed by the (entropic, conformational, elastic) energy, (b) the kinetics of RL-binding and unbinding, regulated by the chemical potential and traced through the diffusion \color{black} of GPCRs and MRPs (related to cAMP$_i$ and cAMP$_e$ respectively), \color{black} and (c) the consistency requirement that the effective chemical potential of the RL-compound must agree both with the energetics and with the fact that lateral pressure arising between lipids surrounding \color{black} the domains of the active species \color{black} and confining their conformational changes. The strict correlation between how cAMP is triggered by RL-compounds \color{black} and modulated by MRPs and the tendency of GPCRs to locate on lipid rafts has been shown by the simulations.
\textcolor{black}{It is worth notning that on the basis of the boundary and limit conditions of interest and by considering the interspecific terms taken into account in the model, coalescence due to Cahn-Hilliard-like dynamics was not expected, this phenomenon being object of future investigations.} \color{black}
The entire analysis has be done with special regard to an existing set of experiments involving HTR-8/SVneo cells and $\beta$2AR. Diffusive phenomena have been considered and the arising chemo-mechanical coupling has been accounted for through the kinetics of ligand binding. The outcomes of the model highlight the mechanobiology of lipid membrane remodelling by providing a mechanical explanation for the observed correspondence between the membrane regions where distribution density of active receptors is higher and locations of lipid rafts. These theoretical predictions actually confirm experimental findings by Nobel Prize 2012 Kobilka \cite{Kobilka:2007} and Lefkovitz \cite{Lefkowitz:2007}. Importantly, the predictive potential of the obtained model can help to provide a key support to quantitative diagnostics, owing for evaluation of cell response to endogenous and exogenous ligands regulated by GPCRs under different operating conditions of the cells.
\section*{Acknowledgements}
L.D., N.M.P. and M.F. acknowledge the Italian Ministry of Education, University and Research (MIUR) under the (1) ARS01-01384-PROSCAN and (2) PRIN 2017 20177TTP3S grants.
N.M.P. and L.D. also acknowledge the partial support from (3) the Italian MIUR "Departments of Excellence" grant L. 232/2016.
N.M.P. and L.D. are supported by the European Commission (EC) under (4) the FET Proactive (Neurofibres) grant No. 732344.
A.R.C. acknowledges support from (5) PON-AIM1849854-1.
M. B. and K. D. acknowledge support from (6, 7, 8) NSF (1150002, 1635407 and 1635435), (9) ARO (W911NF-17-1-0084), (10) ONR Applied and Computational Analysis (N00014-18-1-2528), and (11) BSF (2018183).
The authors wish to acknowledge Prof. Carla Biondi, Prof. Maria Enrica Ferretti and Prof. Fortunato Vesce, from the University of Ferrara, Italy, for the endless conversations about the role of cAMP, its detection and its relationships with submacroscopic effects at the level of bound receptors, and their help to Dr. Laura Lunghi during her experimental work. Prof. Giuseppe Valacchi, from North Carolina State University, USA, and Ferrara, is also acknowledged.
The authors also gratefully acknowledge Prof. Davide Bigoni and Dr. Diego Misseroni, from the University of Trento (Italy), for the discussions about biomechanics of cancer in relation to the findings of this paper, as well as Prof. Anne M. Robertson, from the University of Pittsburgh), Prof. David R. Owen, from Carnegie Mellon University, and Dr. Giuseppe Zurlo, from NUI-Galway-Ireland, for their valuable discussions through the last few years about various biological and theoretical aspects of this paper.
\bigskip
\section*{Appendix A} \label{KrCalc}
\subsection*{Density ratio associated to remodelling}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
The \color{black} effective density $\rho(x,t)$ of the heterogeneous membrane at any time $t$ and at any point $x$ \color{black} can be written as the sum of the products between the true densities $\varrho_k$ and the volumetric fractions $\upsilon_k$ composing the body. \color{black} The membrane is
essentially inhabited by lipids and by transmembrane proteins. The latter share the same true density, although they can appear either in an active or inactive state. Furthermore, it is worth notning that in the intermediate active (reference) configuration displayed in figure \ref{fig:Stress00}) the thickness of the membrane $h_0$ is homogeneous. Henceforth, the following expression can been written for the density $\rho(x,t)$: \color{black}
\begin{align}\label{dens1}
\rho(x,t)&= \upsilon_k\,\varrho_k=\upsilon_{u}\,\varrho^p_{u}+\upsilon_{a}\,\varrho^p_{a}+\upsilon_{l}\,\varrho^{l}\notag\\
&=\left(N_{u}\,\frac{A_{u}}{A}+N_{a}\,\frac{A_{a}}{A}\right)\varrho^p+\left[1-\left(N_{u}\,\frac{A_{u}}{A}+N_{a}\,\frac{A_{a}}{A}\right)\right]\varrho^{l}\notag\\
&=\varrho^l\left[1+\left(N_{u}\,\frac{A_{u}}{A}+N_{a}\,\frac{A_{a}}{A}\right)\left(\frac{\varrho^p}{\varrho^l}-1\right)\right]
\end{align}
where $\varrho^p=dm^p/dV^p$ and $\varrho^{l}=dm^l/dV^l$ are the protein and lipid true densities, respectively, while $A_k/A$ is the areal fraction of the species. By introducing the relation $N_u+N_a=N_{max}$, \color{black} where $N_{max}$ denotes an estimate of the maximum reference value of total proteins present on a surface surface $A$ on the cell membrane, the expression \eqref{dens1} above can be rearranged as follows: \color{black}
\begin{align}\label{dens2}
\rho(x,t)&=\varrho^l\left[1+N_{max}\left(\frac{A_{u}}{A}+\frac{N_a}{N_{max}}\left(\frac{A_a-A_u}{A}\right)\right)\left(\frac{\varrho^p}{\varrho^l}-1\right)\right]\notag\\
&=\varrho^l\left[1+\frac{N_{max}A_u}{A}\left(1+\sum_i\,n_i\left(\frac{A_a}{A_u}-1\right)\right)\left(\frac{\varrho^p}{\varrho^l}-1\right)\right]
\end{align}
where $\sum_i\,n_i=N_a/N_{max}=\xi+\zeta$ represents the cumulative fraction of activated \color{black} receptor and transporters\color{black}. Therefore, the density ratio associated to remodelling, reported in equation \eqref{RemodelingTerm}, is finally obtained as $K_r={\rho(x,0)}/{\rho(x,t)}$. \color{black} It is worth notning that $K_r$ has a meaningful multiscale geometric interpretation through the \textit{Theory of Structured Deformations}, as highlighted in Fig.\ref{fig:Stress0} (see e.g. \cite{deseri:2003}, \cite{Deseri:2010}, \cite{deseri:2012}, \cite{deseri:2015}, \cite{deseri:2019}, \cite{palumbo:2018}), as discussed in details in \hyperref[{SB2}]{Appendix C}.\color{black}
\section*{Appendix B} \label{lateralpress}
\renewcommand{\theequation}{B.\arabic{equation}}
\setcounter{equation}{0}
\subsection*{The chemical potential is influenced by the lateral pressure experienced by the cell membrane}\label{lateral_pressure}
\renewcommand{\theequation}{B.\arabic{equation}}
\setcounter{equation}{0}
The mechanical term $\omega_i \, \left(\lambda^{-1}-1\right)$ appearing in the chemical potential $\tilde{\mu}_i=\tilde{\mu}_i^0 - \omega_i \, \left(\lambda^{-1}-1\right) -\eta_i \, \log \frac{n_i}{n_i^0}$, i.e. \eqref{ChemPot}, which is the parent of the chemo-mechanical energy \eqref{Wnidef}, allows for obtaining the chemical stress through derivation as in equations \eqref{Pmudef1} and \eqref{StressDisp}, i.e.
\begin{equation}\label{ChemStress}
\tilde{P}_n = (n_i^0-n_i) \frac{\partial \tilde{\mu}_i} {\partial \lambda} = - \frac{\partial W_{n_{i}}}{\partial \lambda} = \omega_i \frac{(n_i^0-n_i)}{\lambda^2}.
\end{equation}
As the first formula clearly displays, \color{black} the chemical stress is obtained through the change in chemical potential caused by local stretch changes, \color{black} $\lambda$ (or, equivalently, caused by the local changes of the membrane thickness as \eqref{ThicknessChange}, i.e. $\phi=\lambda^{-1}$, so that $d\lambda=- \phi^{-2} d\phi$). The chemical potential of the species is then key for understanding the mechanobiology involved in ligand-binding of the receptors analyzed in this paper. Because of its high importance, in this section we provide a more physical understanding of the origins of such a key quantity. Indeed, direct considerations at the microscale can be performed by taking into account the lateral pressure arising between \color{black} the active species domains and the adjacent lipids within the bilayer. \color{black} Lateral pressure in lipid membranes is known to be present in bilayers even without receptors. It is worth noting that lateral pressure is known to contribute to the stress across the thickness. In mechanical terms, such a pressure can be seen as the reaction to incompressibility, thereby not affecting the elastic part of the energy. When it comes to considering the membrane stress, formed by a hyperelastic part added on a reactive one, the lateral pressure is indeed such a term in the expression for the stress.
In turn, the lateral pressure changes in the presence of proteins (see e.g. \cite{Cantor:1999, Lee:2004, Marsh:2007, Ollila:2010}) and it influences their chemical potential. \color{black}
\color{black} For the sake of illustration, in the sequel we focus on the active receptors, namely the RL complex, as reported in \cite{Pollaci:2016}. \color{black} In Figure \ref{fig:LateralPressure} it is shown a schematic of a TM domain involved in the conformational changes and the lateral pressure RL profile $\hat{\pi}(\textbf{x}, z, t)$ (marked with the green line in Figure \ref{fig:LateralPressure}.b).
\begin{figure*}[th]
\centering
\includegraphics[width=0.75\textwidth]{figF_LateralPressure_mod.pdf}
\caption{Schematic representation of the lateral pressure between a TM domain and the lipid bilayer as in \cite{Pollaci:2016} (see also \cite{Marsh:2007}).}
\label{fig:LateralPressure}
\end{figure*}
Indeed, this lateral pressure exhibits higher magnitudes in the zones where the lipid head groups interact with the top and bottom sites of the TMs. Therefore $\hat{\pi}(\textbf{x}, z, t)$ performs work against the lateral variation of area, namely the local area change, $\hat{J}(\textbf{x}, z, t)$ at each level $z$ across the thickness.
We approximate the lateral pressure profile $\hat{\pi}(\textbf{x},z,t)=\hat{\pi}(\textbf{x},-z,t)$ with an even piecewise function \color{black} as displayed in red in Figure \ref{fig:LateralPressure}.b) and defined as follows:\color{black}
{\small
\renewcommand\arraystretch{1.5}
\begin{equation}
\label{eq:laterl_pressure_profile}
\hat{\pi}(\textbf{x},z,t) = \left\{
\begin{array}{ll}
-p & 0< z <\dfrac{h}{2} -\phi\\
\dfrac{q}{\phi/2} \left[ z - \left(\dfrac{h}{2} - \phi\right)\right] & \dfrac{h}{2} - \phi < z < \dfrac{h}{2} - \dfrac{\phi}{2}\\
q - \dfrac{q}{\phi/2} \left[ z - \left(\dfrac{h}{2} - \dfrac{\phi}{2} \right)\right]& \dfrac{h}{2} - \dfrac{\phi}{2} < z < \dfrac{h}{2}
\end{array}
\right.
\end{equation}
}
Here $h(\textbf{x},t)$ is the current thickness of the membrane at $\textbf{x}$ at the time $t$, $q$ denotes the value of repulsive pressure arising because of the contrast of the lipid headgroup against the receptor domain, and $p$ is the value of the attractive pressure along the hydrocarbon chain region arising to produce a self-balanced pressure profile.
Balance through the thickness implies that the following relationship must hold:
\begin{equation*}
2 p \left(\dfrac{h}{2}- \phi \right) = 4 \, \dfrac{q \, \phi/2}{2},
\end{equation*}
hence $p$ and $q$ turn out to relate as follows:
\begin{equation}
\label{p_q}
p = \frac{q}{2} \frac{\phi}{h - \phi} .
\end{equation}
The work $\tilde{w}_{\xi}$ done by the lateral pressure \eqref{eq:laterl_pressure_profile} against the change in lateral stretch $\hat{J}(\textbf{x},z,t)$ can be expressed as:
\begin{equation}
\label{W}
\tilde{w}_{\xi} = \pi\, d_s \,h \, \hat{w}_{\xi} ,
\end{equation}
where $\hat{w}_\xi$ denotes the work done per unit area and $\pi\, d_s \,h$ is the exhibited lateral surface of the TM.
If we set
\begin{equation*}
\theta = z - \left( \frac{h}{2} - \phi\right) \qquad \text{and} \qquad \theta' = z - \left( \frac{h}{2} - \frac{\phi}{2} \right),
\end{equation*}
the quantity $\hat{w}_\xi$ can be evaluated as follows:
{\small
\begin{equation}
\label{W*}
\begin{aligned}
\hat{w}_\xi & = 2 \int_0^{\frac{h}{2}} \hat{\pi}(\textbf{x},\,z,t) \, \hat{J}(\textbf{x},\,z,t) \, dz \\
&
\begin{aligned}
= 2 \Bigg( -p \int_{0}^{h/2 - \phi} \hat{J}(\textbf{x},\,z,t) \,d z & + q\frac{1}{\phi/2} \int_{0}^{\frac{\phi}{2}} \theta \, \hat{J}(\textbf{x},\,\theta+\frac{h}{2}-\phi,t) \, d\theta
+ q \frac{\phi}{2} +\\
& - q\frac{4}{\phi} \int_0^{\frac{\phi}{4}} \theta' \hat{J}(\textbf{x},\, \theta'+\frac{h}{2}-\frac{\phi}{2},t) \, d \theta'
\end{aligned}
\\
&\approx 2 \left[
q \frac{\phi}{2}
- p \int_{0}^{\frac{h}{2} - \phi} \left[ \hat{J}(\textbf{x},\,0,t) + z \, \hat{J}_{z}(\textbf{x},\,0,t) \right] \,d z
\right] \\
&= 2 \left[q \frac{\phi}{2} - p \left( \frac{h}{2} - \phi\right) \,J \right] \\
&= \, \frac{\gamma_{RL}}{d_s} \, \phi \left(1 - J \right)
\end{aligned}
\end{equation}
}
\noindent \color{black} The result arises \color{black} after utilizing \eqref{p_q} relating $p$ to $q$
and upon recognizing that the surface tension $\gamma_{RL}$ of the RL compound at the lipid headgroup- is related to the lateral pressure as
\begin{equation*}
q = \gamma_{{RL}}/d_s
\end{equation*}
where $d_s$ denotes the diameter of a TM domain.
Bearing in mind that $h=h_0/\lambda$, where $\lambda=J$, an approximaion for the work done by the lateral pressure is the computed through \eqref{W}, namely $\tilde{w}_{\xi} = \pi\, d_s \,h \, \hat{w}_{\xi}$, by virtue of \eqref{W*}, i.e.
\begin{equation}
\begin{aligned}
\tilde{w}_\xi & = \pi \, \phi \, h_0 \, \gamma_{RL} \, \left(\frac{1}{\lambda} \, - 1\right). \label{Work_lateral_pressure}
\end{aligned}
\end{equation}
With the identification
\begin{equation}\label{omega-xi}
\omega_\xi = \pi \, \phi \, h_0 \, \gamma_{RL}
\end{equation}
equation \eqref{Work_lateral_pressure} yields
\begin{equation}
\begin{aligned}
\tilde{w}_\xi & = \omega_\xi \, \left(\frac{1}{\lambda} \, - 1\right). \label{Work_lateral_pressure-1}
\end{aligned}
\end{equation}
namely \textit{the} (opposite of the) \textit{work per unit area and per unit receptor, reperesents the contribution to the chemical potential $\mu_\xi$ in \eqref{Pmudef2} due to mechanical actions associated with conformational changes, against which the lateral pressure performs work}.
The remaining terms in \eqref{ChemPot}, namely the entropic one, $\eta_\xi \, \log (\xi / \xi^0)$, and the reference value $\tilde{\mu}_\xi^0$ add together and sum up to $\tilde{w}_\xi$ to deliver the complete expression of the chemical potential for the RL compound. Obviously, for the MRPs, namely the transporters, the reasoning is analog, thereby retrieving relation \eqref{Pmudef2} for $n_2=\zeta$, i.e. the analog expression for the chemical potential for such species.
The chemical stress \eqref{ChemStress} can be revisited in the light of \eqref{Work_lateral_pressure-1}. In particular, for the RL species $n_1= \xi$ it is worth noting that $\partial \tilde{\mu}_\xi / \partial \lambda= - \partial \tilde{w}_\xi / \partial \lambda$ and hence the chemical stress associated with the RL compound reads as follows:
\begin{equation}\label{ChemStress-xi}
\tilde{P}_\xi = - (\xi^0-\xi) \, \frac{\partial \tilde{w}_\xi} {\partial \lambda}.
\end{equation}
It is then worth emphasizing that, \textit{in spite of the fact that the lateral stress profile between TMs and lipids is self-balanced across the thickness, the rate of change of its work} (done against the lateral variation of area) \textit{times the relative density of the RL compound actually produces a quota of the membrane stress}. Of course for the MRPs an analog reasoning follows.
\section*{Appendix C} \label{SB2}
\subsection*{Conformational changes, remodelling and associated energetics}
\renewcommand{\theequation}{C.\arabic{equation}}
\setcounter{equation}{0}
A schematic of the kinematics of the cell membrane in the presence of \textit{conformational changes} and, possibly, \textit{remodelling} of the lipid membrane is displayed in the sequel in Fig.\ref{fig:StructuredDeformations}. For the sake of illustration, this focuses on active receptors alone, although the treatment for active transporters (MRPs) can be done in the same way.
Geometrical changes occuring among the TransMembrane domains of TM3, 5 and 6 (see \ref{fig_cell_with_receptors_and_transporters}) are shown together with the deformation of the lipid bilayer.
As pointed out in Sect. 3.2, the
submacroscopic changes caused by the formation of the RL compound (i.e. the receptor-ligand binding) do cause a remodelling of the
cell membrane. This is due to a change of the density of active receptors thereby determining the re-organization of the surrounding
lipids.
\textit{Structured Deformations} (see e.g. \cite{deseri:2003}, \cite{Deseri:2010}, \cite{deseri:2012}, \cite{deseri:2015}, \cite{deseri:2019}, \cite{palumbo:2018}) is a multiscale geometric framework that permits to account for both conformational changes and remodelling arising at the submacroscopic level. Such a framework does allow for envisioning three full configurations for the membrane.
\begin{figure}[htb!]
\centering
\includegraphics[width=0.95\columnwidth]{fig_StructuredDeformations.pdf}
\caption{Conformational changes and membrane deformation.}
\label{fig:StructuredDeformations}
\end{figure}
In particular, on the top-left side of Fig.\ref{fig:StructuredDeformations}, a schematic of a piece of the membrane in its virgin configuration is sketched.
Besides the reference thickness for the membrane and for the TMs, $h_0$ and $h_0^r$ respectively, and their counterparts in the current configurations, $h$ and $h'$, the quantities highlighted there are the normal $\textbf{N}$ to the mid-plane of the lipid membrane in the virgin configuration, its counterpart $\textbf{n}$ in the deformed configuration, the reference value $\phi_{\sss{T}}$ of the diameter of each TM, and the available diameter $\phi_0$ for such movements to arise. The quantity $\rho_{\sss{T}}:=\phi_0/\phi_{\sss{T}}$ denotes the \textit{room space} available between the TMs for conformational changes in the virgin configuration, this indicates the \textit{degree of packing} of the TMs. The minimum value of the \textit{room space} $\rho_{\sss{T}}$ is estimated to be $1+2/\sqrt{3}$ (closest packing of the three domains involved in the movements), hence its value is taken to be between such a value and $3$, corresponding to the available room space of a TM in the middle of TM3, 5 and 6, as shown in the sequel in Fig.\ref{fig_room_space}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\columnwidth]{fig_room_space.pdf}
\caption{Room space for conformational changes in active receptors as in \cite{Pollaci:2016}: (a) schematic of TM3, 5 and 6, (b) minimum and regular room space among the domains involved in the conformational changes.}
\label{fig_room_space}
\end{figure}
It is worth notning that this schematic allows for distinguishing two different deformed configurations: the classically deformed one and the current configuration. The latter is represented on the top right side of Fig. \ref{fig:StructuredDeformations}. Mathematically, this is described through a given Structured Deformation, namely a pair $(\textbf{y}, \textbf{G})$, where $\textbf{y}$ is the deformation mapping from the virgin configuration, and $\textbf{G}$ is a tensor field which would be equal to the gradient of $\textbf{y}$ if no conformational changes would occur. The lower part of Fig. \ref{fig:StructuredDeformations} offers an explanation of such interpretation through the multiplicative decompostition $(\textbf{y},\textbf{G}) \,=\, (\textbf{i},\textbf{H}) \circ (\textbf{y},\textbf{F})$ of the pair $(\textbf{y},\textbf{G})$ introduced above. An intermediate classically deformed configuration is represented in the bottom of this figure where $(\textbf{y}, \textbf{F})$, with $\textbf{F} = \text{Grad}$ $\textbf{y}$, maps points in the virgin configuration in this intermediate one. This classical deformation is followed by a non-classical one. The latter is represented by a pair $(\textbf{i}, \textbf{H})$ and it is displayed in Fig. \ref{fig:StructuredDeformations}: here $\textbf{i}$ is the identity mapping (leaving then the macroscopic configuration untouched), while $\textbf{H}=\textbf{G}\textbf{F}^{-1}$ is a tensor field accounting for all the local conformational changes.
It is also worth noting that this representation is fully compatible with the schematic depicted in Fig.\ref{fig:Stress0} in the main text. As pointed out in \cite{deseri:2003}, a factorization of the structured deformation $(\textbf{y},\textbf{G}) \,=\, (\textbf{y},\textbf{F}) \circ (\textbf{i},\textbf{K}) $ such that a classical deformation follows a non-classical one where the macroscopic deformation of the body does not change, while conformational changes such that $\textbf{K}=\textbf{F}^{-1}\textbf{G}$ can occur at the submacroscopic level. Obviously there is a one-to-one relationship between the configurational changes described (through $\textbf{H}$) from the classically deformed configuration to the current configuration and their description starting from the virgin configuration (through $\textbf{K}$). Indeed, pure geometry is suggestive of the fact that the deformation of the membrane plays a key role (through $\textbf{F}$ $=$ Grad $\textbf{y}$) on carrying over the information related to the conformational changes, as $\textbf{K}=\textbf{F}^{-1} \textbf{H} \textbf{F}$.
It is worth noting that the definition of the remodelling factor introduced in \eqref{RemodelingTerm} can be exactly written as
\begin{equation}
K_r \, =\, det \textbf{K},
\label{remodelling-Structured Deformation}
\end{equation}
as the change in volume between the virgin configuration and the reference configuration is solely due to disarrangements which, in this case, are due to (submacroscopic) remodelling.
Furthermore, it is obvious that $det \textbf{K} \, = \, det \textbf{H}$. Henceforth, the evaluation of the remodelling factor can either be done upon following the scheme displayed in Fig.\ref{fig:Stress0} ($(\textbf{y},\textbf{F}) \circ (\textbf{i},\textbf{K})$, namely by letting disarrangements to act first, then followed by a classical deformation) or the one in Fig.\ref{fig:StructuredDeformations} ($(\textbf{i},\textbf{H}) \circ (\textbf{y},\textbf{F})$, i.e. a classical deformation first, then the conformational changes).
\subsection*{Conformational energy: the case of active receptors}
\label{sectionB2}
As is has been recalled in Fig.\ref{fig_cell_with_receptors_and_transporters}, conformational changes in active receptors are mainly characterized by a rotation $\omega$ of TM6 about its axis and by a translation $\mu h$ of the TM6 domain with respect to TM3 and TM5 (see e.g. \cite{Ghanouni:2001}).
Henceforth, the change in Helmholtz free energy per unit area and per unit receptor relative to the classically deformed configuration (as displayed in Fig. \ref{fig:StructuredDeformations}) due to such movements can be then interpreted as a result of two contributions, both arising in the current configuration (see \cite{Pollaci:2016}). For both of such terms, the change in entropy per unit area and per unit receptor is evaluated as in \cite{Pollaci:2016}. The change of entropy due to rotation is accounted for by generalizing the result obtained in \cite{Finkelstein:1989}, namely
\begin{equation}
\label{eq:rotational}
\varphi_{\scriptscriptstyle CR}^{(1)} = A \log \left(\frac{\omega}{2 \pi^{2/3}} \right) ,
\end{equation}
where $\omega$ represents the rotation of the TM domains involved in the conformational change about the normal $\textbf{e}_3$ to the mid-plane of the membrane in the current configuration, $A$ is a normalization constant and $A \cdot K_B T$ represents the conformational energy level per unit receptor per unit area at the reference angle $\omega^* = 2 \pi^{2/3} e$.
The local translation measure is relative to the free volume available for the conformational changes in the current configuration. This is calculated by multiplying the current value of the thickness $h$ of the cell membrane by the available area $\pi (d_0/2)^2 J$, where $d_0$ represents the referential diameter of the zone in which conformational changes of the TMs can occur. It is worth recalling that $J=h_0/h$ depends on the location $\textbf{x}=(x_1, x_2)$ of the mid plane of the bilayer. As recalled in Sect. \ref{sectionB3}, it has been shown that $J$ is the order parameter in lipid membranes (see e.g. \cite{Zurlo:2006, Deseri:2008, Deseri:2013}), allowing for discriminating whether or not lipids exhibit ordered ($J=1$ indicates the completely straight configuration of the lipid tails) or disordered phases ($J>1$, curlier configuration). Henceforth, the resulting change in entropy due to translational changes may be written as follows:
\begin{equation}
\label{eq:translational}
\varphi_{\scriptscriptstyle CR}^{(2)} = A \log \left(\frac{\mu \, h}{ \left(J \, (d_0/2)^2 \, h \right)^{1/3}} \right),
\end{equation}
where $\mu \,h$ represents a measure of the local translation of the TMs with respect to current location and $\left(J \,(d_0/2)^2 \, h\right)^{1/3}=\left((d_0/2)^2 \, h_0\right)^{1/3}$ is a reference measure of such translation. Upon introducing the ``room space" parameter
\begin{equation}
\rho_T=d_0/d_s,
\label{room_space}
\end{equation}
which measures how much room is available for the conformational changes of the TMs, as $d_s$ indicates the diameter of any of the TMs involved in such changes.
The total \textit{conformational energy density per unit area and per unit (distribution density of) receptors} is obtained as the sum of both contributions \eqref{eq:rotational} and \eqref{eq:translational} above, i.e.
\begin{equation}
\label{eq:ConformationEnergy}
\varphi_{\scriptscriptstyle CR} = \varphi_{\scriptscriptstyle CR}^{(1)} + \varphi_{\scriptscriptstyle CR}^{(2)} = A \log\left(\kappa \frac{\eta}{J}\right) ,
\end{equation}
where
\begin{equation}
\eta = \omega \, \mu
\label{eta}
\end{equation}
represents the \textit{conformational field} and $\kappa$ is a dimensionless geometric constant defined as follows:
\begin{equation}
\label{eq:K}
\kappa = \frac{1}{2^{}} \left( \frac{2\,h_0}{\pi \, d_s \, \rho_{\sss{T}}} \right)^{\frac{2}{3}} .
\end{equation}
Accordingly to equation \eqref{eq:K} and a range of geometric quantities collected from literature in \Tab{tabF:K_parameters}, a range for $\kappa$ can be then determined to be $\kappa_{\min} \leq \kappa \leq \kappa_{\max}$, where $\kappa_{\min} = 0.587$, $\kappa_{\max} = 1.634$ and $\kappa_m = (\kappa_{\min}+\kappa_{\max})/2 = 1.11$ will be used.
\begin{table}[htbp]
\centering
\begin{tabular}{ccc}
\toprule
\small{thickness $h_0$} & \small{diameter $d_s$} & \small{room space $\rho_{\sss{T}}$} \\
$[nm]$ & $[nm]$ & - \\
\midrule
$3\div 6$ & $0.3\div 0.5$ & $2.15 \div 3$ \\
\bottomrule
\end{tabular}
\caption{\textbf{Table} \ref{tabF:K_parameters}. Ranges of values of geometrical parameters assumed from \cite{Lee:2004, Hiden:2007, Pollaci:2016}, while the room space $\rho_{\sss{T}}$ is estimated from \Fig{fig_room_space}. \label{tabF:K_parameters}}
\end{table}
\noindent In \cite{Pollaci:2016} bounds for $\eta$ are found in in terms of the ratio $h/h_0$. Nonetheless, for a given set of geometrical parameters $h_0, d_s, \rho_T$, \color{black} it can be shown that $\eta^*=K^{-1}$ represents the lowest possible value for $\eta$. \color{black} Maximum admissible values of the conformational changes $\eta$ can also be estimated by considering extreme configurations of a single TM
. Indeed $\eta$ is composed by a rotation, $\omega$ in the range $(0, \pi]$ (i.e. the value $0$ cannot be achieved), and a shear, $\mu = \tan \, \alpha$ where $\alpha$ is an angle across thickness (see \Fig{fig:StructuredDeformations}). The angle $\alpha$ can conceivably achieve values up to $\pi/4$, due to the presence of the surrounding TMs. Then, the conformational field is found in the following interval: $\eta^* \, \leq \, \eta \, \leq \, \pi$.
\noindent The result \eqref{eq:ConformationEnergy} is a generalization of \cite{Finkelstein:1989} and \cite{Murray:2002}, as the entropic changes are measured while accounting for the deformation of the lipid membrane.
\subsection*{Conformational energy and work done by the lateral pressure: the case of active receptors}
\label{sectionB3}
The contribution to the total Helmholtz free energy $W$
of the energetics of active receptors, RL, and transporters, MRPs,
has been worked out in Sect. \ref{sec:3M}. In particular, equation \eqref{Wnidef} for the RL compounds specializes as follows
\begin{equation}\label{Wnidef-xi}
W_{\xi}(\lambda^{-1},\xi)= \eta_\xi \, \xi \, \log \frac{\xi}{\xi^0}+(\xi-\xi^0)\left(\omega_\xi\left(\lambda^{-1}-1\right) -\tilde{\mu}_{\xi}^0 -\eta_\xi \right).
\end{equation}
Equation \eqref{Pmudef2} particularized for such a species and integrated over the domain $\Omega_a$ in the intermediate configuration, can be considered. Its variational derivative with respect to $\xi$, the density of RL, leads the following expression for the chemical potential
\begin{equation}
\tilde{\mu}_\xi = \tilde{\mu}_\xi^0 - \omega_\xi \, \left(\lambda^{-1}-1\right) -\eta_\xi \, \log \frac{\xi}{\xi^0} .\label{Pmudef2-2}
\end{equation}
In \hyperref[{lateralpress}]{Appendix C} it has been obtained that the term $\omega_\xi \, \left(\lambda^{-1}-1\right)$ appearing both in \eqref{Wnidef-xi} and \eqref{Pmudef2-2} solely regards the work done by the lateral pressure against the local area variation arising across the membrane during conformational changes.
\noindent Mixing of the the RL (active receptors) throughout the membrane contributes to lower the energy of the system in a purely entropic way (see e.g. \cite{Rangamani:2014}). Upon identifying
\begin{equation}
\eta_\xi = K_B \, T,
\label{etacsi}
\end{equation}
where $K_B$ is the Boltzmann constant, $T$ is the absolute temperature of the bath in which the cells are embedded, i.e.
\begin{equation}
\label{eq:ligand-receptor-diffusion}
K_B T ~ \xi \left(-e_{RL} + \log \left(\frac{\xi}{\xi^0} \right) \right),
\end{equation}
where $e_{RL}$ is the specific activation energy for the complex receptor-ligand.
The corresponding energy density \eqref{eq:ConformationEnergy} per unit receptors, per unit area and per unit $K_B T$ due to conformational changes and ideal mixing allows for writing the quota of the total Helmholtz free energy associated with $\xi$ as follows:
\begin{equation}
\label{eq:energy-RL}
\begin{aligned}
\int_{\Omega_a} \xi &\left[-e_{\scriptscriptstyle RL} + \log\left(\frac{\xi}{\xi^0} \right) - \varphi_{\scriptscriptstyle CR} \right]\, d\Omega_a. \\
\end{aligned}
\end{equation}
The variational derivative of the overall energy with respect to $\xi$ yields the following expression of the corresponding chemical potential
\begin{equation}
\label{eq:chemical_potential_RL}
\tilde{\mu}_\xi = K_B T \left( -e_{RL} + 1 + \ln \left( \frac{\xi}{\xi^0} \right) - \left(\xi \,\varphi_{\scriptscriptstyle CR} \right)_{,\xi}\right).
\end{equation}
As in \cite{Pollaci:2016}, by taking \eqref{etacsi} into account, the comparison between \eqref{eq:chemical_potential_RL} and \eqref{Pmudef2-2} establishes the consistency of both expressions for the chemical potetial of the RL compound and leads to the following relation among the density of active receptors $\xi$, the conformational field $\eta$ and the membrane thickness change $h$:
\begin{equation}
\label{eq:X-1}
\left(\frac{\xi}{\xi^0} \right)^{-1} = 1 - \dfrac{C}{\dfrac{h}{h_0} -1 } \ln \left(K \dfrac{h}{h_0} \eta \right).
\end{equation}
where
\begin{equation}
\label{eq:C}
C = \dfrac{K_B T A}{\pi \phi_T h_0 \gamma_{RL}}.
\end{equation}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\columnwidth]{fig_raft_conformation.pdf}
\caption{Behavior of $\xi/\xi^0$ vs $h/h_0$ using $\eta$ as parameter. Here $\eta^* = 1.07$ represents a basal value and $\Delta \eta = 2.06$ (see \cite{Pollaci:2016}).}
\label{fig:raft_conformation}
\end{figure}
In \Fig{fig:raft_conformation}, slices obtained by fixing values of $\eta$ are shown, namely the quantity $X = \xi/\xi^0$ as a function of the relative thickness change $H = \lambda^{-1}= h/h_0$. This shows that relative thickness change $H$ increases with an increase of active receptor density $X$. This result, is found to be in agreement with the experimental findings of Lefkovitz \cite{Lefkowitz:2007} and Kobilka \cite{Kobilka:2007}. As in a purification process the addition of cholesterol to the membrane induces lipid ordering, this actually enhances the clustering of activated receptors on rafts. Our model is then consistent to this observation already at the constitutive level. Furthermore, in \cite{Pollaci:2016} it is shown that equation \eqref{eq:X-1} allows one to obtain the conformation in terms of relative distribution of active receptors $X$ and of the thinning field $H$, in the following form:
\begin{equation}
\label{eq:relationship_eta}
\eta = \frac{1}{H K} \exp \left( - \frac{(1-H)(1-X)}{C \, X}\right) .
\end{equation}
The relative receptor density $X$ is a positive definite quantity and it is increasing with $H$. In \cite{Pollaci:2016} it is shown that the simultaneous occurrence of both these conditions implies the occurrence of the following inequalities:
\begin{subequations}
\label{eq:eta_bounds}
\begin{align}
\eta < \frac{1}{K~H} \exp \left(\frac{1-H}{C} \right) = \eta_{\sss{UB}}, \\
\eta \ge \frac{1}{K~H} \exp \left(\frac{H-1}{H} \right) = \eta_{\sss{LB}},
\end{align}
\end{subequations}
for given values of $H$, $C$ and $K$. Further investigations (see Table \ref{tab:A} below) allow for estimating $C$ in the range $[0.05 \div 0.75]$, while the value taken for this parametric study is $C = 0.3$.
\begin{table}[htbp]
\caption{\textbf{Table} \ref{tab:A} Estimated values for $C$. For each choice of parameters, the first pair of rows is referred to $A =1$, while the second refers to $A = 5$. Values of $\gamma_{\sss{RL}}$ come from literature \cite{Ollila:2010, Marsh:2007, Pollaci:2016}.} \label{tab:A}
\centering
\small
\begin{tabular}{ccccc}
\toprule
$h_0 \, [nm]$ & $\phi_{\sss{T}} \, [nm]$ & $\gamma_{\sss{RL}} \, [mN/m]$ & $A/C$ & $C$ \\
\midrule
\multirow{2}{*}{3} & \multirow{2}{*}{0.5} & \multirow{2}{*}{10} & \multirow{2}{*}{11.01} & 0.091\\
& & & & 0.454\\
\midrule
\multirow{2}{*}{6} & \multirow{2}{*}{0.3} & \multirow{2}{*}{10} & \multirow{2}{*}{13.21} & 0.076\\
& & & & 0.379\\
\midrule
\multirow{2}{*}{3} & \multirow{2}{*}{0.5} & \multirow{2}{*}{35} & \multirow{2}{*}{38.52} & 0.026\\
& & & & 0.129\\
\midrule
\multirow{2}{*}{6} & \multirow{2}{*}{0.3} & \multirow{2}{*}{35} & \multirow{2}{*}{42.22} & 0.022\\
& & & & 0.108\\
\bottomrule
\end{tabular}
\end{table}
\newpage
\bibliographystyle{unsrt}
| 8a26d6de146f298c13ccbe14cf393a8d2ce2d624 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{introduction}
A citation network refers to a (un)weighted, directed graph with the directed edges representing the citations, the edge weights (if exist) representing the number of citations between two nodes (e.g., papers, journals, authors). The impact made by the work done by an entity/node is measured using the impact factor. The impact factor of a journal for a particular year is measured by the total number of citations received by the articles published in the journal during the preceding two years, divided by the total number of articles published in that journal during the preceding two years \cite{impactfactor}.
It is widely known and we shall also explain below that anomalous citations can purposefully manipulate impact factors. Since the Journal Impact Factor (JIF) is used to determine the importance of the contributions made by the journal in its own domain, it becomes essential to spot the anomalies in citation networks in order to understand if there is any misuse of citations to manipulate JIF and what are the reasons behind such misuse.
Anomalous citations can be divided into two major categories:
\begin{enumerate}
\item \textbf{Self-citations}:
They refer to the citations made by entities, i.e., journals or authors, to themselves, which can, in certain cases, lead to a disproportionate increase in JIF of the entity \cite{chakraborty2018good}.
\item \textbf{Citation Stack}:
It refers to the citation made by an entity, i.e., a journal or an author, to another entity, which can, in certain cases, be anomalously high and deviate from the usual behaviour of the entity \cite{chakraborty2018good}.
\end{enumerate}
\citet{fowler2007does} showed that the more one cites oneself (self-citation), the more others cite them. \citet{ioannidis2015generalized} classified the phenomenon of self-citation to different types -- direct, co-author, collaborative and coercive-induced self-citation. They further raised concerns over how inappropriate self-citations can affect impact factor, and suggested the usage of citation indices which are more foolproof to collusion using inappropriate self-citations. \citet{Bartneck2010DetectingHM} discussed how h-index can be inflated by authors by manipulating self-citation. \citet{SelfCitation}, in their study on analyzing the effect of excessive self-citation, argued for a new metric based on more transparency to help curb excessive self-citations, which may have a negative effect on our ability to truly analyze the impact and contributions of a scientific research in its research domain. However, limited work has been done in the field of analyzing citation stack.
In 2011, Thomson Reuters \cite{ThomsonReuter} (currently known as Clarivate Analytics) suspended 33 journals from its Journal Citation Report due to a very high rate of self-citations, which contributed to as much as 90\% of the JIF of those journals. Therefore, anomalous citation detection is an important task in bibliometrics.
\textbf{Our Contributions}: In this paper, we study a citation network at the journal level, define the notion of anomalies for scientific citation networks, and detect the anomalous journal pairs with focus on the anomalies from one entity to another (citation stack), along with a confidence score indicating the correctness of our prediction. We also categorize them into different types, and figure out possible reasons for such anomalies.
The major contributions of this paper are mentioned below:
\begin{itemize}
\item We present a novel approach to detect the anomalies in a journal-level scientific citation network, and compare the results with the existing graph anomaly detection algorithms.
\item Due to the lack of ground-truth for citation anomalies, we introduce JoCAD, a novel dataset which consists of synthetically injected citation anomalies (Section \ref{Datasets_JoCAD}), and use it to evaluate our methodology and compare the performance of our method with the state-of-the-art graph anomaly detection methods.
\item Our model is able to predict the anomalous citation pairs with 100\% precision and 86\% F1-score (Section \ref{experimental-results}).
\item We further categorize the anomalies detected into various types and dig out possible reasons to develop a deeper understanding of the root causes.
\item We also analyze our model on the Microsoft Academic Search dataset, a real-world citation dataset (Section \ref{Experiment_MAS}), labeled by human annotators (Section \ref{human-annotator}).
\item We interpret our results using a case study, wherein our results resemble the citations and Scimago Journal \& Country Rank (SJR) rating-change charts (Section \ref{case-study}).
\item We further design an interactive web portal - `Journal Citations Analysis Tool' which can process any citation network and show journal-level anomalous citations. It helps the users analyze the temporal citation pattern of a given journal (Section \ref{portal}).
\end{itemize}
\noindent\fbox{%
\parbox{0.78\columnwidth}{%
\textbf{Reproducibility:} The codes and the datasets are available at: \url{http://bit.ly/anomalydata}.}}
\section{Related Work}\label{relatedwork}
The methodology used in this paper is closely related to the techniques used in anomaly detection in complex networks, and journal-level study of citation networks. Both these topics are well-studied; however, to the best of our knowledge, anomaly detection in journal-level citation network has not been studied yet.
\subsection{Anomaly Detection in Complex Networks}
Anomaly detection in networks is a widely explored area. \citet{Noble} proposed a subdue based method for anomaly detection in graphs.
\citet{Hodge2004} wrote a survey of machine learning and statistical techniques for outlier detection, and categorized the outlier detection algorithms into three major types, namely, unsupervised clustering, supervised classification and semi-supervised recognition.
The anomalies themselves have been categorized into two broad types \cite{chen} -- \textit{white crow} anomalies, which can be easily identified as outliers, and \textit{in-disguise} anomalies, which try to hide themselves and are thus difficult to detect. \citet{moon} presented OutRank, an algorithm to detect white-crow anomalies by assigning a similarity score to objects. \citet{Eberle2007MiningFS} detected in-disguise anomalies, which they called ``structural anomalies".
We present two methods that cover both kinds of anomalies mentioned above. The box plot method in Section \ref{sec:method} detects white-crow anomalies as it identifies high number of citations between journals (which is an easily noticeable feature) after bucketing them according to their sizes. The time-series method in Section \ref{sec:method} detects in-disguise anomalies as it analyzes citations over the years and detects outliers from the usual behaviour (which can not be noticed from an overall study of the data).
Anomalous groups in a citation network can be seen as communities with high citation activity with each other.
\citet{Sun2005NeighborhoodFA} presented algorithms for community detection in bipartite graphs. They introduced two kinds of functions for the purpose of community detection and outlier detection.
Oddball algorithm was proposed for unsupervised anomaly detection in weighted graphs, wherein each node is treated as an ego, and the neighbourhood network around it is studied as the egonet \cite{oddball}.
\citet{Noble} detected anomalies in graph-based data by categorizing them in two ways - anomalous sub-structures and anomalous sub-graphs. They claimed that these anomalous sub-structures and sub-graphs are extremely rare as compared to ``usual" sub-structures and sub-graphs.
\citet{Chandola:2009:ADS:1541880.1541882} presented a survey which provided classified anomalies into three types:
\begin{itemize}
\item \textbf{Point Anomalies:}
It refers to the individual data points which are anomalous with respect to the entire data.
For example, consider the real life scenario of the price of a stationary item. The cost of the item suddenly rises and becomes twice the previous cost. The new cost is thus, an example of a point anomaly.
\item \textbf{Contextual Anomalies:}
If a data instance is anomalous only in a particular context, then it is a contextual anomaly. It is also known as a conditional anomaly. \citet{Huber} designed a method to process event streams from technical systems for timely detecting anomalies and thus helping prevent huge industrial losses.
\item \textbf{Collective Anomalies:}
If the existence of certain data points individually is not anomalous but their collective existence collectively is, then they are called collective anomalies.
\end{itemize}
In our study, we focus on the collective anomalies.
The standard anomaly detection algorithms for graphs could not be used for detecting pairwise anomalies in a citation network because the subtle signal strengths of the anomalies at the journal level in the citation network could not be detected by the pre-existing graph algorithms. An alternate way of modelling this problem can be using supervised machine learning techniques, which can be used to learn the suitable feature representation that can help in detecting anomalous behaviour among journals, authors, paper, etc. However, due to the lack of a ground-truth dataset of considerable size which has sufficient examples of anomalous behaviours of the above entities, machine learning algorithms could not be applied. We empirically show that existing graph anomaly detection methods could not work on the problem under consideration (Section \ref{experimental-results}).
This motivated us to study detection of pair-wise anomalies.
\subsection{Journal-level Study of Citation Networks}
\citet{Hummon1989ConnectivityIA} modelled the connectivity in citation networks as an exhaustive search problem and applied a depth-first-search variant to quantify the similarity in the network. They focused on the similarity of edges. \citet{Zhang} presented a clustering algorithm which calculated the similarity between adjacent nodes and formed clusters based on that score.
Study of citation network is a sufficiently explored topic. It dates back to 1965, when \citet{solla} studied the citation network and made inferences about the nature of references of a scientific paper with respect to its field and length. Various facets of citation networks have been studied, most widely explored being the network of papers and its outgoing citations. \citet{Osareh} studied citation network and its applications. \citet{Zhao:2008:CPC:1458082.1458125} studied the evolution of heterogeneous citation networks.
Researches have also focused on the citation interactions among authors using citation analysis. \citet{Prabha} carried out a pilot study on the reasons behind the citations in scientific papers. He found that many of the citations made by papers are not essential for the results produced by it. \citet{Cronin2002} studied paper reference patterns with respect to the authors of the paper.
They defined an author's \textit{citation identity} as the authors they cite in their work, and \textit{citation image} as the authors who cite them. They argued that the reference styles of authors are a form of watermark for their work.
One important application of anomaly detection in citation network is to find the possibility of collusion among entities. \citet{cartel} came up with an algorithm to detect citation cartels \cite{journal}. It was a community detection algorithm focusing on anomalous pair of citations. The authors also discussed the possibility of some anomalous pairs being cases of citation cartels. Claiming whether a citation is a result of possible collusion or not is a non-trivial task, and requires an in-depth examination of every citation. In this paper, we focus only on the anomalies in the citation graph, and we do not try to claim any intentions behind those anomalous citations.
\section{Datasets}\label{datasets}
For the purpose of this study, we consider two different datasets. One of the major challenges while devising the experiments was the lack of ground-truth for anomalous citations in the real-world dataset. We thus introduce a journal citation anomaly dataset (called JoCAD), a dataset which consists of synthetically injected anomalies (see Section \ref{Datasets_JoCAD}). We use this for the purpose of testing our model and reporting the final accuracy in terms of the F1-score (Section \ref{experimental-results}). We also use the Microsoft Academic Search Dataset, a real-world citation dataset, for carrying out a case-study on the anomalous pair of journals detected by our model (Section \ref{case-study}).
\subsection{Microsoft Academic Search Dataset}\label{Datasets_MAS}
The Microsoft Academic Search (MAS) Dataset \cite{chakraborty2018good,MAS} consists of 1.6 million publications and 1 million authors related to the Computer Science domain. It has various metadata information about a paper, including the year of publication, keywords of the paper, the references of the paper, the specific field (AI, algorithms, etc.) the paper is targeted towards, and the abstract of the paper. The dataset spans from 1970 to 2014. This dataset has been extensively used in the past for bibliographic analysis \cite{chakraborty2015categorization,chakraborty2018universal}.
We preprocessed the dataset according to the needs of our model. We removed the metadata which were not useful for the study. These included keywords, abstract, and the related fields of the paper. Table \ref{tab:MAS} shows statistics of the dataset.
\begin{table}[!h]
\begin{center}
\caption{ Microsoft Academic Search Dataset \cite{chakraborty2018good,MAS}}\label{tab:MAS}
\begin{tabular}{ c | c }
\hline
Entity & Number \\
\hline
No. of papers & 1.6 million \\
No. of citations & 6 million \\
No. of journals & 1,800 \\
No. of journal pairs & 8,500\\
\hline
\end{tabular}
\end{center}
\vspace{-5mm}
\end{table}
\subsection{JoCAD - Our Synthetically Injected Anomaly Dataset}\label{Datasets_JoCAD}
For anomaly detection in citation graphs, the ground-truth is not well-defined. Hence, we created a synthetic dataset, wherein we synthetically injected anomalies to serve as the ground truth in our further experiments. The synthetic data generation process is motivated by \citet{Hayat2017}.
Our synthetic dataset contains 100 journals, each having a journal index between 0 and 99, and containing citation information for the journals ranging over 20 years (2000-2020).
We assign the number of papers in the journal as a random number in a pre-decided range. Similarly, we assign the citation count between each pair of journals as a random number in a pre-decided range which we consider as the normal citation behaviour of the journal pairs.
For injecting anomalies, we intuitively list properties of an anomalous pair of journals. The properties consist of a sudden spike in the number of citations from one journal to another or from both journals to each other in some random year; or a gradual increase from one journal to another or from both journals to each other, which result in an anomalously high number of citations over all the years. Using these methods with reasonable randomness, we create a total of 110 anomalies. The final dataset consists of 110 anomalous pairs out of total 10,000 pairs of journals.
The types of anomalies we inject into the dataset can be broadly classified into five types:
\begin{enumerate}
\item If the number of citations from journal $J_i$ to journal $J_j$ in a particular year is significantly more than the normal citation behaviour of the journal pairs, i.e., there is a sudden spike in the number of citations from $J_i$ to $J_j$ in that year, then the citation from $J_i$ to $J_j$ is an instance of {\bf Type 1 anomaly}.
\item If the number of citations from $J_i$ to $J_j$ increases significantly over successive years, with the difference in the number of citations between any two years $Y_i$ and $Y_j$ being greater than the expected difference in the number of citations between any journal pair, then the citation from $J_i$ to $J_j$ is an instance of {\bf Type 2 anomaly}.
\item If the number of citations from $J_i$ to $J_j$ in the year $Y_k$ is significantly higher than the average number of citations between any journal pairs, and the number of citations from $J_j$ to $J_i$ in the year $Y_{k+1}$ is significantly higher than the average number of citations between any journal pairs, then both the citations from $J_i$ to $J_j$ and vice-versa are instances of {\bf Type 3 anomaly}.
\item If the number of citations from $J_i$ to $J_j$ in the year $Y_i$ is greater than or equal to double the number of citations received by $J_i$ from $J_j$ in the previous year $Y(i-1)$, then the citation from $J_i$ to $J_j$ is an instance of {\bf Type 4 anomaly}.
\item If the number of citations from $J_i$ to $J_j$ in the year $Y_i$ is greater than or equal to double the number of citations from $J_i$ to $J_j$ in the previous year $Y_{i-1}$, then the citation from $J_i$ to $J_j$ is an instance of {\bf Type 5 anomaly}.
\end{enumerate}
\section{Methodology}\label{methodology}
In this section, we first introduce the problem definition, followed by the proposed methodology.
\begin{figure*}[!]
\centering
\includegraphics[width=15 cm]{dia1.jpeg}
\caption{Flow diagram of our methodology.}
\label{fig:methodology}
\end{figure*}
\subsection{Problem Definition}
Given a set of papers $P$, where each paper has 4 attributes -- journal name, authors, citations, and the year of publication, as the input to the system, the task is to predict the journal-to-journal anomalies with a confidence score of the prediction made, classify the anomaly into its type and suggest possible reasons for the existence of the same.
\subsection{Proposed Methodology}\label{sec:method}
We use a two-fold approach (as shown in Figure \ref{fig:methodology}) to classify the citation anomalies --- box plot and time-series analysis.
\subsubsection{Box Plot Bucket Analysis:}
We start by clustering the journals into buckets based on similar behaviour in terms of the total number of papers published in the journal.
For each journal in our database, we define the usual behaviour of the given journal by analyzing the behaviour of all the journals lying in the same bucket as the given journal, i.e., all the journals having similar number of papers as that of the given journal, and then use the box plot method to predict the anomalies.
For each possible bucket, we pair each of the papers in the current bucket with each of the papers in the bucket of the given journal. We refer to the set of all these paper-paper pairs as a grid. We now detect the anomalies using box plot. To calculate the outlier behaviour, using a box plot, we calculate the first quartile and the third quartile\footnote{
The first quartile is the 25$th$ percentile of the data and the third quartile is the 75$th$ percentile of the data. The second quartile refers to the median of the distribution.} of the data points corresponding to the journal-journal citation counts in the selected grid.
We define anomalous journal-journal pairs as the data points which lie outside the area:
[(first-quartile - ($1.5\times IQR$)) , (third-quartile + ($1.5\times IQR$))], where interquartile range $IQR$ = (third-quartile - first-quartile).
The anomalies detected by the box plot method are static anomalies as these anomalies are based on the total number of citations between two journals over all the years.
\subsubsection{Time-series Analysis:}
Once we have the pairs declared anomalous by the box plot, we analyze each of them with time-series analysis to confirm if there exists some year during which there was a sudden change in their usual citation behaviour.
The anomalies detected by the time-series analysis method are dynamic anomalies as these anomalies are susceptible to change every year.
To define the ``usual behaviour", we use the empirical rule. The empirical rule, also called the ``68-95-99.7 rule'' \cite{pukelsheim1994three}, is a statistical rule that states the percentage of values of a Gaussian distribution that \textcolor{black}{lies} within 2, 4, and 6 standard deviations around the mean. According to the rule, 2 standard deviations around the mean cover 68.27\% of the values, 4 standard deviations around the mean cover 95.45\% of the values, and 6 standard deviations around the mean cover 99.73\% of the values.
To check the usual behaviour for a pair of journals for a given year (say $Y$), we list the number of citations from one journal to another till $Y$, and calculate mean and standard deviation from the distribution. Then, as stated by the empirical formula, 99.73\% of the data is expected to lie within 6 standard deviations around the mean. Hence, if the number of citations from one journal to another exceeds this range, it is suggested to be anomalous by the time-series analysis.
To understand the different types of unusual behaviour, we extend the existing notions of synchronous and dianchronous citations \cite{macro} and use synchronous citations to refer to the citations from a journal to another journal, and dianchronous citations to refer to the citations received by a journal.
For a given pair of journals, we check four types of behaviours for each year:
\begin{enumerate}
\item \textbf{One sided synchronous citations}:
This consists of detecting the unusual behaviour of outgoing citations from one journal to the other for each side, i.e., if the two journals are $J_1$ and $J_2$, the above test is performed separately for the outgoing citations from $J_1$ to $J_2$ and for the outgoing citations from $J_2$ to $J_1$, for each year as stated above.
\item \textbf{One sided dianchronous citations}:
This consists of detecting the unusual behaviour of incoming citations from one journal to the other for each side i.e., if two journals are $J_1$ and $J_2$, the above test is performed separately for the incoming citations from $J_1$ to $J_2$ and the incoming citations from $J_2$ to $J_1$, for each year as stated above.
\item \textbf{Double sided synchronous citations}:
This consists of detecting the unusual behaviour of outgoing citations from one journal to the other for both the journals simultaneously, and if both of the sides lie in the outlier range, this pair is said to have double sided synchronous behaviour.
\item \textbf{Double sided dianchronous citations}:
This consists of detecting the unusual behaviour of incoming citations from one journal to the other for both the journals simultaneously, and if both of the sides lie in the outlier range, this pair is said to have double sided synchronous behaviour.
\end{enumerate}
Checking both synchronous and dianchronous citation behaviours is essential because a particular number of citations from one journal to another can be in the normal range for the citing journal but not in the normal range for the cited journal, or vice-versa. This kind of case can happen in many instances. For example, one unpopular journal suddenly increases citations to a very popular journal and to an unusually high level; or one very popular journal suddenly increases citations to an unpopular journal. In both cases, only one side of the single side anomalies will be evident from the distribution based check described above.
\subsubsection{Reasons for Possible Anomalies:}
We categorize the detected anomalies hinting at five possible reasons for their occurrence, namely, many-many anomaly, many-one anomaly, one-many anomaly, one-one anomaly, and previous-author collaboration.
For the first four, we supplement a metric that captures the essence of the citation graph being highly crowded at both the ends (journals $J_1$ and $J_2$) for many-many anomaly, and so on.
To do this, we define the ``normal behaviour" of the citations {\em given by a paper to a journal} by first creating a Gaussian distribution with the citation count from a paper to a journal as the data points and then, using the empirical formula, we use $\mu \pm \sigma$ to define the normal behaviour of nearly 70\% of the data points.
Now consider all the papers in the citing (sender) journal, i.e., the journal $J_1$, and count the number of papers in the journal which contribute to $J_2$ (cited/receiver journal) that is at least the normal behaviour: $paper-to-journal\ count > \mu \pm \sigma$. Dividing this by the total number of papers in published in $J_1$, we get the percentage of papers (vertices in the citation network) in the $J_1$'s side which is crowded in terms of the citation count they produce for $J_2$, denoted as `sender-percentage'.
Now, we define the ``normal behaviour" of the citations received by a paper from a journal, by creating a Gaussian distribution with the citation count from a journal to a paper as the data points, and then, we use the empirical formula to define the normal behaviour of the data points.
Similar to the previous case, based on the above definition of the normal behaviour, we now consider all the papers in the receiver journal $J_2$, and count the number of papers in $J_2$ which are cited by $J_1$ (the sender journal) more than or equal to the normal behaviour: $journal-to-paper\ count > \mu \pm \sigma$. Dividing this by the total number of papers in $J_2$, we get the percentage of papers (vertices) in the receiver's side which is crowded in terms of the citation count they receive from $J_1$, denoted as `receiver-percentage'.
Now we are ready to categorize an anomaly as many-many, many-one, one-many, or one-one based on the following criteria:
\begin{enumerate}
\item \textbf{Many-many Anomaly:}
If the sender-percentage is $>75\%$ and the receiver-percentage is $>75\%$, then it is the case of a many-many anomaly since both the receiver's and the sender's sides are extremely crowded.
\item \textbf{Many-one Anomaly:}
If the sender-percentage is $>75\%$ and the receiver-percentage is $<25\%$, then it is the case of a many-one anomaly since the receiver's side is not crowded while the sender's side is extremely crowded.
\item \textbf{One-many Anomaly:}
If the sender-percentage is $<25\%$ and the receiver-percentage is $>75\%$, then it is the case of a one-many anomaly since the sender's side is not crowded while the receiver's side is extremely crowded.
\item \textbf{One-one Anomaly:}
If the sender-percentage is $<25\%$ and the receiver-percentage is $<25\%$, then it is the case of a one-one anomaly since both the sender's and receiver's side are not crowded.
\end{enumerate}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.1\columnwidth]{reasons.png}
\caption{Pictorial representation of four of the five possible reasons of anomalies. The fifth reason (effect of collaboration) is not shown here.}
\label{fig:reasons}
\end{figure*}
Figure \ref{fig:reasons} diagrammatically shows the four reasons listed above.
To explain the reasons for the anomalies, we also look into the previous collaboration of the authors of a publication and count the number of such occurrences between the journal-journal pairs.
In particular, we detect if at the time of publishing a paper in year $Y$ any of the authors of the paper has previously collaborated with any of the authors of the paper it is citing. We then increment a counter by $1$. We do so for paper-pairs consisting of every paper in the sender-journal $J_1$ and all the papers it cites in the receiver-journal $J_2$, and thus calculate the total number of \textbf{previous-author-collaborations}.
\subsection{Confidence Score}\label{confidence}
Our model finally assigns a confidence score to every anomaly detected, which states how confident the model is on the pair of journals being anomalous in a year \cite{ConfScore}. A confidence score is first assigned by box plot method to each anomaly, then by the time-series method, and at last both the scores are combined, giving the final confidence score.
We use the $tanh$ function\footnote{$tanh$ is a hyperbolic function which gives a value ranging from -1 to 1, and in case of only positive inputs, 0 to 1.} for calculating the individual confidence scores from the two methods -- box plot and time-series.
One intuition is that if our method states a pair to be anomalous, we are at least $50\%$ sure of it being anomalous. Following this intuition, we scale the confidence scores given by individual methods to be from $0.5-1$ instead of $0-1$. Another intuition is that in the time-series method, both-sided anomaly is a more unusual event than a single-sided anomaly. This is because a both-sided anomaly in a particular year means both the journals went beyond the normal behaviour and cited the other. We state the confidence of a both-sided anomaly to be at least $75\%$. Therefore, a both-sided anomaly is scaled between $0.75-1$ instead of $0-1$.
\begin{figure*}
\centering
\includegraphics[width=15 cm]{labelled_chart.png}
\caption{Histogram of confidence scores of anomalies in the MAS dataset.}
\label{fig:histogram}
\end{figure*}
For combining the scores, we choose to take the average of the two scores. This is done because both the methods have equal importance in the detection of the anomaly.
Figure \ref{fig:histogram} shows the histogram of the confidence scores of anomalies detected from the MAS dataset. We analyze the most anomalous pair with confidence score 0.93 and find that it is logical for the pair to be anomalous (see case studies in Section \ref{case-study}).
\section{Experiments}\label{experiments}
\subsection{Comparison of Time-series Analysis with SJR ratings}\label{experimentA}
To evaluate the performance of the time-series method for anomalous behaviour detection, we use the SJR ratings to check changes in ratings of journals stated anomalous in some particular years. Scimago Journal \& Country Rank, better known as SJR, is a journal rating system \cite{SJR}. It uses data similar to what is used by H-index, and gives year-wise ranks to journals. It is a well-recognised rating system for journals and conferences.
We scraped the SJR ratings over the years for all the journals present in the MAS dataset from the portal made for SJR ratings. The comparison of sudden increase in citations to each other for pairs of journals mostly resonates with the respective SJR rating changes of the citation-receiving journal in the next two years. The case-study of the most anomalous pair of journals also supports the same result (see Section \ref{case-study}).
\subsection{Experimental Results}\label{experimental-results}
For anomaly detection in citation graphs, there is no well-defined ground-truth or a proper evaluation metric. Hence, for the purpose of evaluating the performance of our method, we created a synthetic dataset wherein we synthetically injected anomalies~\ref{datasets}. They were treated as the ground-truth in our experiments, and we report our results based on them. We use precision, recall and F1 score for the final evaluation.
\textbf{Baseline Methods:} Due to the lack of any existing baseline for the given task, we decided to use the existing graph anomaly detection algorithms and each of the methods used by us, namely, the box plot method and the time-series method, to act as baselines for evaluating our proposed model.
We used K-means clustering, one of the popular graph anomaly detection algorithms, but it could only detect 38 anomalous pairs out of the total of 110. The detected pairs are the pairs with a large number of citations on both sides i.e., to and from one journal to another over the years. These are clearly visible by taking the summation of the citations on both sides, similar to the box plot method. However, K-means clustering algorithm is unable to detect other types of anomalous pairs, and thus has a lower recall and F1 score as compared to our method.
\begin{table}[h!]
\label{F1_scores}
\begin{center}
\caption{Results of anomaly detection methods and our models.} \label{tab:result}
\begin{tabular}{ c c c c }
\hline
Method & Precision (\%) & Recall (\%) & F1 Score (\%) \\
\hline
K-means & 100 & 34.5 & 67.2 \\
Box plot & 59.19 & 93.63 & 72.53 \\
Time-series & 86.73 & 77.27 & 81.73 \\
Our model & 100 & 75.45 & 86.01 \\
\hline
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:result} shows the comparative results of the baselines and our model. The F1 score of our model comes out to be $86.01\%$ which is higher than that of the baselines. Also, the precision of our model comes out to be $100\%$, which means that all the pairs stated anomalous with the agreement of both the methods are actually anomalous.
\subsection{Results on the MAS Dataset}\label{Experiment_MAS}
We ran the model on the real-world MAS dataset. The histogram of confidence scores of the anomalies found in the dataset is shown in Figure \ref{fig:histogram} (explained in Section \ref{confidence}). Out of total 8.5 thousand journal-pairs, total of 328 anomalies are found. The number of unique pairs of anomalous journals (irrespective of the year of anomaly) is 230, and the total number of journals found in any anomaly is 103. We observe that the occurrence of anomalies is rare (only $3\%$), and the occurrence of both-sided anomalies is extremely rare (only $4$ cases out of $8.5$ thousand pairs), as shown in Table \ref{tab:distribution}
\begin{table}[h!]
\begin{center}
\caption{Distribution of types of journal-pair anomalies found in the MAS dataset.}\label{tab:distribution}
\begin{tabular}{ c c c }
\hline
Type & Number & Percentage of total pairs \\
\hline
Single (One sided) & 324 & 3.85\% \\
Double (Both sided) & 4 & 0.04\%\\
Total & 328 & 3.90\% \\
\hline
\end{tabular}
\end{center}
\end{table}
We plot the average number of publications per year that are anomalous, i.e., the ratio of the number of anomalous publications in a year to the total number of publications in that year. The temporal frequency mapping of anomalies is shown in Figure \ref{fig:year}. We can infer that the frequency of the relative year-wise anomalous journals decreases over the years. This decrease might be due to a huge boost given to the academic research in the field of Computer Science, which led to an increase in the total number of journals. This might have decreased the ratio over time. Another reason could be more stringent rules and regulations imposed in the scientific community, which could have led to a decrease in anomalous activity amongst journals.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{size.png}
\caption{Average anomalous publications throughout the years.}
\label{fig:year}
\end{figure}
Next, we plot the trends with respect to the size of the anomalous journals, defined in terms of the number of papers published in it. Figure \ref{fig:size} shows the decline of anomalous activities as the size of a journal increases. This might be due to a bigger journal having a wider variety of papers, research domains, and authors. Another reason might be the journals which have gained recognition and prestige over time are likely to receive high number of novel and high-quality papers which should be published. Thus, prestigious and renowned journals may have a bigger journal size, which are less likely to be anomalous.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{time.png}
\caption{Fraction of anomalous papers published in a journal with the size of the journal. }
\label{fig:size}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=15 cm]{lab_case.png}
\caption{The number of citations from Symposium on Information theory to Transactions of Information Theory over years.}
\label{fig:case}
\end{figure*}
\subsection{Human Annotators}\label{human-annotator}
Two human annotators\footnote{The annotators were experts in data mining domain and their age ranges between 25-35 years.} were asked to independently label the anomalies detected by our method on the MAS dataset as anomalous or non-anomalous. The annotation was done on the basis of the number of citations between the journals over years, and relevant metrics such as the average number of citations over all the years between the journals and the popularity of both journals. They checked the popularity of journals based on factors like impact factor and cite score.
Each annotator annotated all 328 anomalous pairs. They placed 38 pairs and 25 pairs respectively in the non-anomalous category. The inter-annotator agreement was $0.78$, based on \textit{Cohen's kappa coefficient}.
\subsection{Case Study}\label{case-study}
We conduct a case study for the anomalous pair with the highest confidence in the MAS dataset.
The pair is the journal \textit{IEEE Transactions on Information Theory} and the conference \textit{IEEE International Symposium on Information Theory}, and the confidence score of being an anomaly is $93\%$. IEEE has a professional society, namely \textit{The Information Theory Society} \cite{itsoc}, under whose umbrella lies its main journal - \textit{IEEE Transactions on Information Theory} (IEEE TIT) and its flagship conference, \textit{IEEE International Symposium on Information Theory}. The symposium is a conference for researchers to meet and discuss work done in the field of information theory. The work can be previously published or unpublished.
As the work can be previously published, and the conference is a part of the same society as the journal, many authors of the published papers in the journal could attend the conference to display their work. Hence, it is natural for the two to have numerous citations to each other over the years. Because of this, there can be two reasons for a citation:
\begin{enumerate}
\item A paper has already been published in the journal and the authors were invited to the conference to present and discuss their work.
\item A paper was not accepted by the journal and the authors presented their work \textcolor{black}{at} the conference. Later, some other researcher used their work and cited them. The paper then got published in the journal.
\end{enumerate}
Some citations may not exist because of the journal and the conference being a part of the same society. There would be many instances where a paper not published in the journal but presented at the conference had cited a paper previously published in the journal.
\begin{figure*}[!t]
\centering
\includegraphics[width=15 cm]{Portal_final_27_04.png}
\caption{A snapshot of our portal. It shows the result of detecting multiple anomalous relationships for the journal IET software/IEEE Proceedings in the year 1989.}
\label{fig:Portal_2}
\end{figure*}
Charts in Figure \ref{fig:case} show the number of citations from Symposium on Information Theory to IEEE TIT, the change in SJR rating of IEEE TIT over the years, the number of citations from IEEE TIT to Symposium on Information Theory, and the change in SJR ratings of Symposium on Information Theory over the years.
A clear resemblance in citations and SJR rating-change charts \cite{scimagojr.com} for both the journal and the conference shows that the findings of time-series analyzes correspond to the changes in the SJR rating of IEEE TIT. As stated in the introduction, JIF depends upon the number of citations the journal had received in the two previous years. A sudden spike can be seen in the citations from the conference to the journal \textcolor{black}{in the year} 2003. At the same time, a spike can be seen in the SJR rating of the journal \textcolor{black}{in the year} 2004, which shows that the increased citation count has contributed towards \textcolor{black}{an} increase in the rating of the journal.
\section{A Web Portal}\label{portal}
We use the journal level citation network created using the MAS dataset to develop a portal which allows the user to find the anomalies corresponding to a journal for all the possible years. It provides an interactive graphical interface, which visually depicts the anomalies in a citation network.
The user can enter the details of the journal, based on which the portal provide the user with a year-wise analysis of the anomalies. A screenshot of the portal is presented in Figure \ref{fig:Portal_2}.
For all pairs of journals in our dataset which form an anomalous pair with the journal queried by the user, we display the year-wise anomalies in an anomaly graph. This graph depicts each of the journals as nodes, while the citations between them is represented by the edges. The size of a node is proportional to the number of papers in the journal, i.e., a larger size of the node indicates that the journal has published a higher number of papers. The size of an edge depicts the number of citations between the two journals, i.e., a thicker edge between two journals indicates a higher number of citations between them. The number on top of each node represents its journal index, which is mapped to the respective journal name in our dataset.\smallskip{}
\noindent\fbox{%
\parbox{0.98\columnwidth}{%
A beta version of the portal is live and can be accessed here: \url{https://journalcitationsanalysistool.herokuapp.com/}.}}
\section{Conclusion}\label{conclusions}
In this paper, we presented a novel model to detect the journal-level anomalous pairs. The model also gives the confidence score by analyzing both static and dynamic anomalies using box plot bucketing method and time-series analysis. We also curated JoCAD, a novel dataset, which consists of synthetically injected citation anomalies. We further ran our model on two datasets, namely, JoCAD and the Microsoft Academic Dataset, and used it to evaluate our methodology. We achieved 86\% F1 score on JoCAD in the comparison of our method with the standard graph anomaly detection methods.
We interpreted our results in a case study using the Microsoft Academic Dataset, wherein our results resemble the citations and Scimago Journal \& Country Rank (SJR) rating-change charts. We experimentally showed a high similarity between the time-series trends predicted by our method and the time-series trends as reflected by the SJR. We further designed an interactive web portal - `Journal Citations Analysis Tool` which given the citation network as an input, shows the journal-level anomalous citations, and helps users analyze the temporal citation pattern of a given journal.
Future work in this direction can include an in-depth study about the reasons for the anomaly detected as well as extend the work in developing methods for the detection of citation cartels. We acknowledge that while curating the dataset, JoCAD, we could have possibly introduced a human bias in it, which can be reduced in the future. We would like to explore the anomalous citation behavior in other domains as well. Other relevant future work can take into account the textual context of a citation made, and calculate its similarity with the content of the paper being cited. This will help us understand the relevance of the citation made and thus help predict anomalous citations better.
\section*{Acknowldgement}
T. Chakraborty would like to acknowledge the support of Ramanujan Fellowship, DST (ECR/2017/00l691), and the Infosys Centre of AI, IIIT-Delhi, India.
\bibliographystyle{ACM-Reference-Format} | d163606dcc2370cc8ef0666016007c1496c590d6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Constructing a latent space for human motion is an important problem as it has a wide range of applications such as motion recognition, prediction, interpolation, and synthesis.
Ideal motion spaces should be compact in the sense that random sampling in the space leads to plausible motions and comprehensive so as to generate a wide range of human motions. In addition, locally linear arrangement of the semantically related hidden vectors would benefit motion synthesis, e.g., by simple algebraic operations.
However, constructing a compact and versatile motion space and extracting valid motions from it remains a challenging problem because the body parts of human body are highly correlated in general actions and the joints are constrained to satisfy the bone lengths and the range of movement. The high dimensionality of the joint space adds additional difficulty to this problem.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\linewidth]{figures/interpolation.jpeg}
\caption{Examples of motion interpolation on the latent motion manifold generated by our method. The first and last columns are snapshots of two input motions, and the intermediate columns show the snapshots of four individual motions obtained by the linear interpolation on the motion manifold.}
\label{fig:interpolation}
\end{figure}
In this paper, we present a novel framework to construct a latent motion manifold and to produce various human motions from the motion manifold.
In order to embrace the temporal characteristic of human motion, our model is based on the sequence-to-sequence model.
The unsupervised sequence-to-sequence models have been shown to be effective by previous studies on motion prediction \cite{martinez2017human, pavllo2018quaternet}. %
Based on these studies, we develop several novel technical contributions to achieve a compact yet versatile latent motion manifold and a motion generation method as follows.
First, our model is characterized by the combination of one encoder and two decoders.
Given a motion manifold vector, one decoder learns to generate the joint rotation while the other learns to output joint rotation velocities. As will be discussed later, the joint rotation decoder has the advantage of reconstructing long term motions better.
In comparison, the joint velocity decoder has the advantage of improving the continuity of the motion.
By complementing each other, our two decoder model shows a higher reconstruction accuracy than that of the single decoder model.
Second, unlike previous studies that deal with only either joint angles or joint positions, by adding a forward kinematics (FK) layer \cite{villegas2018neural}, our joint angle-based human representation achieves the advantage of satisfying bone-length constraints and simplifying joint limit representation.
By additionally considering joint position computed by the FK layer while training, our method reduces the joint position error, which is visually more perceptible than the joint angle error.
Lastly, we introduce several loss functions, each of which contributes to enhancing the quality of the motion manifold in different aspects.
A reconstruction loss reduces the difference between the reconstructed motion and the input motion and thus allows the manifold to synthesize motion content and details observed in the training motion dataset.
A regularizer loss improves the distribution quality of the motion manifold and thus enables random sampling and interpolation on the manifold.
In addition, an adversarial loss increases the naturalness of the motions generated from the motion manifold.
In this paper we show that, based on these technical contributions, our method allows for various practical applications such as random generation of motions, motion interpolation, motion denoising and motion analogy as will be shown in Sec.~\ref{sec:result}. The capability of our method is demonstrated by the comparison with other approaches, such as the seq2seq model~\cite{martinez2017human} and the convolution model~\cite{holden2015learning, holden2016deep}.
The remaining part of this paper proceeds as follows: After reviewing previous studies related to our work in Sec.~\ref{sec:related}, we present our method and loss function in detail in Sec.~\ref{sec:method}.
Sections~\ref{sec:data} detail the data pre-processing and Sec.~\ref{sec:result} reports a number of experiments performed to verify the effectiveness of our method.
Section~\ref{sec:conclusion} discusses the limitations of our work, future research directions, and concludes the paper.
Our code and networks are available at {\small \url{https://github.com/DK-Jang/human_motion_manifold}}.
\section{Related work}
\label{sec:related}
Researcher have developed several methods to construct motion manifold to generate natural human motions, but compared with studies on manifold learning for other data such as image, research on motion data is scarce.
Linear methods such as PCA can model human motion in only a local region.
Chai \textit{et al}. \cite{chai2005performance} apply local PCA to produce a motion manifold that includes a certain range of human motion, and apply it for synthesizing movements from low dimensional inputs such as the position of end effectors.
Lawrence \cite{lawrence2004gaussian} use Gaussian Process Latent Variable Model (GPLVM) to find a low dimensional latent space for high dimensional motion data.
Taylor \textit{et al}. \cite{taylor2007modeling} propose a modified Restricted Boltzmann Machine that is able to deal with the temporal coherency of the motion data. Lee \textit{et al}. \cite{lee2010motion} propose motion fields method, a novel representation of motion data, which allows for creating human motion responsive to arbitrary external disturbances.
Recently, with the development of deep learning technology, a method of constructing a motion manifold by using Convolutional Neural Network (CNN)-based encoder was introduced by Holden \textit{et al}. \cite{holden2015learning, holden2016deep}.
Butepage \textit{et al}. \cite{butepage2017deep} compare a number of deep learning frameworks for modeling human motion data.
Our method for constructing motion manifold is based on previous studies on sequence learning for motion to predict the joint position sequences of a 3D human body given past motions.
Martinez \textit{et al}. \cite{martinez2017human} develop a novel sequence-to-sequence encoder-decoder model that predicts human motion given a short duration of past motion. The presented result is impressive but has a few limitations that sometimes implausible motions such as foot sliding are generated and the initial pose of the predicted motion is somewhat discontinuous from the input motion.
Pavllo \textit{et al}. \cite{pavllo2018quaternet} selectively use a joint rotation-based loss for short term prediction and a joint position-based loss for long term prediction. The latter includes forward kinematics to compute the joint positions. However, the basic sequence-to-sequence model can only predict short term motions and has limitations in predicting non-trivial, long term motions. In addition, a loss function that minimizes only the prediction error does not guarantee to construct compact and versatile motion manifold. Our method solves these problems by jointly considering joint rotation and position errors in the loss function and by adding regularization to the motion manifold.
In a broader perspective, our work is related with the studies on recognizing and generating human motion, which
remains a challenging research topic due to the high dimensionality and dynamic nature of the human motion.
Wu and Shao \cite{wu2014leveraging} propose a hierarchical dynamic framework that extracts top-level skeletal joint features and uses the learned representation to infer the probability of emissions to infer motion sequences. Du \textit{et al}. ~\cite{du2015hierarchical} and Wang \textit{et al}. ~\cite{wang2017modeling} use recurrent neural network (RNN) to model temporal motion sequences and propose hierarchical structure for action recognition.
With regard to motion synthesis, Mittelman \textit{et al}. \cite{mittelman2014structured} propose a new class of Recurrent Temporal Restricted Boltzmann Machine (RTRBM). The structured RTRBM explicitly graphs to model the dependency structure to improve the quality of motion synthesis.
Fragkiadaki \textit{et al}. \cite{fragkiadaki2015recurrent} propose the Encoder-Recurrent-Decoder (ERD) that combines representation learning with learning temporal dynamics for recognition and prediction of human body pose in videos and motion capture.
Jain \textit{et al}. \cite{jain2016structural} propose structural RNN for combining the power of high-level spatio-temporal graphs.
\section{Method}
\label{sec:method}
This section details our framework.
After defining notations used in this paper, we explain the structure of the network and the design of the loss function for training.
\subsection{Representation and notations}
\label{subsec:rep}
We denote the human motion set by $\mathcal{Q}$ and corresponding random variable by $\mQ$.
A motion with a time range of $[t,t+\Delta t -1]$ is written as $\mQ_{t:(t+\Delta t -1)} = [\vq_t, \ldots, \vq_{t+\Delta t -1}]$, where $\vq_t$ denotes the pose at time $t$. A pose is represented with a set of joint angles written in the exponential coordinates, i.e., $\vq_t = [q_{i,x}^t, q_{i,y}^t, q_{i,z}^t]_{i=1}^{n_{joint}}$ where $(q_{i,x}^t, q_{i,y}^t, q_{i,z}^t)$ are the three components of the exponential coordinates and $n_{joint}$ is the number of joints.
Therefore, the dimension of a human motion is $\mathcal{Q} \in \real^{\Delta t \times n_{joint} \times 3}$.
Lastly, $\vp_t$ is the pose represented with the joint positions at time $t$ corresponding to $\vq_t$, and $\mP_{t:(t+\Delta t -1)}=[\vp_t, \ldots, \vp_{t+\Delta t -1}]$.
$\mP$ is also a random variable of motion set $\mathcal{Q}$.
\subsection{Motion manifold with sequential networks}
\label{subsec:motionmanifold}
\begin{figure*}[th]
\centering
\includegraphics[width=0.99\textwidth]{figures/overview.jpeg}
\caption{Structure of our sequential networks for constructing the motion manifold.}
\label{fig:overview}
\end{figure*}
We construct a motion manifold in an end-to-end unsupervised way using a network of sequential networks, with an objective to minimize the difference between the ground truth motion space distribution and the reconstructed motion space distribution extracted from the latent motion manifold.
To this end, we develop a sequential model that consists of the RNN with Gated Recurrent Unit (GRU). Our model has a sequence-to-sequence structure \cite{martinez2017human}, which is often used in machine translation. This RNN structure is effective for maintaining the temporal coherency in motion, and {\color{green} it is trained to generate a fixed length of motion (150 frames) in our study.}
As shown in Fig. \ref{fig:overview}, our model includes the combination of one encoder and two decoders with a regularizer.
The encoder takes the source motion as an input and maps it to the latent motion space.
The regularizer encourages the encoded motion distribution to approximate some prior distribution.
The two decoders are designed to map the latent motion space to joint angles and joint velocities, respectively.
Details of our model are given next.
\subsubsection{Encoder}
\label{subsubsec:enc}
The encoder consists of a GRU and one linear layer, and Fig. \ref{fig:overview} shows the unrolled schematic diagram of the encoder.
The $\Delta t$ poses $[\vq_t, \ldots, \vq_{t+\Delta t -1}]$ of a motion are input to the GRU sequentially.
The GRU encodes the current frame while being conditioned by the previous frames with their hidden representation. Specifically, the pose $\vq_i$ in the $i$-th frame is encoded as follows:
\begin{equation} \label{eq:en}
\begin{aligned}
h_i^{Enc} = \text{GRU}_{W_{Enc}}(h_{i-1}^{Enc}, \, \vq_i),
\end{aligned}
\end{equation}
where $h_{i}$ is the hidden state at frame $i$, and $W_{Enc} \in \mathbb{R}^{\text{3}n_{\text{joint}} \times d_h}$ are the training parameters with $d_h$ being the hidden dimension of the GRU.
After the final pose of the input motion is read, one linear layer of parameter $W_{\text{c}} \in \mathbb{R}^{d_h \times d_m}$ receives $h_{t+\Delta t -1}$ and compresses it to produce the $d_m$-dimensional code $Z \in \mathcal{Z}$ where $\mathcal{Z}$ denotes the motion manifold.
It is worth mentioning that this compression brings the benefit of denoising input data.
Now the encoder mapping $Enc: \, \mathcal{Q} \rightarrow \mathcal{Z}$ is completed.
\subsubsection{Latent motion manifold with the Wasserstein regularizer}
\label{subsubsec:reg}
We adopt the Wasserstein regularizer for matching the distribution $E_Z:=\mathbb{E}_{P_{\mQ}}[E(Z \mid \mQ)]$ of the motion manifold to the desired prior distribution $P_Z$.
Unlike the variational auto-encoder \cite{rezende2014stochastic}, the sequential networks trained with the Wasserstein regularizer allows non-random encoders to deterministically map inputs to the latent codes, and thus it helps randomly sampled or interpolated points in the motion manifold correspond to plausible motions
Refer to \cite{tolstikhin2017wasserstein} for more details about the Wasserstein regularizer.
\subsubsection{Decoder with joint rotation and joint velocity}
\label{subsubsec:dec}
Our decoder model consists of two kinds: One decoder learns the joint rotation and the other learns joint rotational velocity as shown in Fig. \ref{fig:overview}.
Both decoders are based on the GRU while the connection structures of the two are different.
Unlike the rotation decoder, the velocity decoder adds a residual connection between the input and the output to construct joint rotation.
Each decoder then generates the reconstructed joint angle sequence in reverse temporal order as suggested by \cite{srivastava2015unsupervised}.
The decoders are trained simultaneously with backpropagation.
This dual decoder model is based on the idea of \cite{srivastava2015unsupervised}.
By combining the two decoders, we can alleviate the limitations of individual decoder models.
The rotation decoder shows strength when reconstructing long term motions because it learns joint angle itself.
Conversely, it may cause pose discontinuity between frames.
The velocity decoder has the advantage of reconstructing continuous human motion as it outputs difference between consecutive rotations, which is usually small and easier to learn. However, training velocities tends to be unstable in a long-term sequence because the longer the motion is, the more error is accumulated.
As our two decoders have contrasting strengths and weaknesses, when combined, they complement each other in synergy.
Unlike previous studies about motion prediction, recognition and manifold ~\cite{butepage2017deep,martinez2017human,holden2015learning,fragkiadaki2015recurrent,pavllo2018quaternet} in which either only the joint rotations or the joint positions are used, our model considers both the joint rotations and positions in the motion reconstruction loss term, $L_R$ (See Eq.~\ref{eq:motion_recon_loss}).
Loss with joint angles has the advantage of preventing errors such as inconsistent bone length or deviation from human motion range, and thus learning with joint angle loss can generate plausible motions.
However, rotation prediction is often paired with a loss that averages errors over joints by giving each joint the same weight.
The ignorance of varying influence of different joints on the reconstructed motion can yield large errors in the important joints and degrade the quality of the generated poses.
The joint position loss minimizes the averaged position errors over 3D points, which better reflects perceptual differences between poses.
To combine both joint rotations and positions in the motion reconstruction loss $L_R$, we add a forward kinematics (FK) layer that computes the joint positions from the joint rotations.
This allows for calculating the loss between the joint positions of the target motion and the reconstruction motion.
The FK module is valid for network training because its output is differentiable with respect to joint rotation.
Finally, our method reconstructs the motion in the reverse order of the input sequence.
Reversing the target sequence has an advantage in learning in that the first output frame of the decoder needs only to match the last frame input of the encoder, which allows for a continuous transition of hidden space vectors from the encoder to the decoders. {\color{green} Refer to \cite{srivastava2015unsupervised} for a theoretical background on this approach.}
Details of our decoder are explained next.
\paragraph*{Joint Rotation Decoder}
\label{par:dec_rot}
The unfolded schematic diagram of the joint rotation decoder is shown in the upper row in Fig. \ref{fig:overview}.
It first transforms an element of the motion manifold $z \in Z$ to a $d_h$-dimensional hidden space vector with a linear layer of parameter $W_e^r \in \mathbb{R}^{d_m \times d_h}$.
Then, conditioned by the hidden space vector representing the future frames, the GRU and a linear layer outputs the reconstructed pose $\widehat{\vq}_i^r$ at the $i$-th frame given its next pose $\widehat{\vq}_{i+1}^r$:
\begin{align}
h_i^{{Dec}^r} &= \text{GRU}_{W_{{Dec}^r}}(h_{i+1}^{{Dec}^r}, {\color{green} \widehat{\vq}_{i+1}^r} ), \label{eq:de_rot}
\\
\widehat{\vq}_{i}^r &= W_o^{r \, T} h_i^{{Dec}^r}, \label{eq:out_rot}
\end{align}
where $W_{{Dec}^r} \in \mathbb{R}^{\text{3}n_{\text{joint}} \times d_h }$ is learning parameter of the GRU and $W_o^r \in \mathbb{R}^{d_h \times \text{3}n_{joint}}$ is the parameter of the linear layer.
Note that, as mentioned earlier, the decoder uses the reversed input motion as the target motion, so the reconstruction is performed in the order of $\widehat{\mQ}_{(t+\Delta t -1):t} = [\widehatt{\vq}_{t+\Delta t -1}, \ldots, \widehat{\vq}_t]$.
Unlike the encoder, the decoder uses the reconstructed result of the previous frame as the input \cite{martinez2017human, li2017auto}.
This is equivalent to the noise scheduling \cite{bengio2015scheduled} without parameter tuning for long term reconstruction, and it also helps prevent the overfitting.
The initial input $\widehat{\vq}_{t+\Delta t}^r$ to the GRU is set zero because there is no reconstruction result of the previous frame.
The reconstructed joint rotations are used to calculate the angle loss with respect to the target motion, and are also used to calculate the position $\widehat{\vp}_{i}^r$ through the FK layer.
\begin{equation} \label{eq:out_rot_fk}
\begin{aligned}
\widehat{\vp}_{i}^r = \, \text{Forward Kinematics} \, (\widehat{\vq}_{i}^r)
\end{aligned}
\end{equation}
After the last pose $\widehat{\vq}_t$ is generated, the joint decoder mapping $Dec^r: \, \mathcal{Z} \rightarrow \mathcal{Q}$ is completed.
\paragraph*{Joint Velocity Decoder}
\label{par:dec_vel}
The joint velocity decoder has the similar structure to the joint rotation decoder.
The main difference is that it has a residual connection to generate $\widehat{\vq}_{i}^v$.
\begin{align}
h_i^{\text{dec}^v} &= \text{GRU}_{W_{\text{Dec}^v}}(h_{i+1}^{\text{Dec}^v}, \, {\color{green} \widehat{\vq}_{i+1}^v} ), \label{eq:de_vel} \\
\widehat{\vq}_{i}^v &= W_o^{v \, T} h_i^{\text{Dec}^v} + \, \widehat{\vq}_{i+1}^v, \label{eq:de_vel_output}\\
\widehat{\vp}_{i}^v &= \, \text{Forward Kinematics} \, (\widehat{\vq}_{i}^v), \label{eq:out_vel_fk}
\end{align}
where $W_{\text{Dec}^v} \in \mathbb{R}^{\text{3}n_{\text{joint}} \times d_h}$ and $W_o^{v}$ are the learning parameters.
This residual network learns the difference between the current frame pose $\widehat{\vq}_{i}^v$ and the previous frame pose $\widehat{\vq}_{i+1}^v$.
Therefore, the model predicts the angle difference or velocity and integrates it over time.
After the last pose is generated, the joint velocity decoder mapping $Dec^v: \, \mathcal{Z} \rightarrow \mathcal{Q}$ is completed.
\subsection{Training the motion manifold}
\label{subsec:train}
\begin{figure*}[th]
\centering
\includegraphics[width=0.85\textwidth]{figures/loss_overview.jpeg}
\caption{Each loss term is evaluated from the data processed in the network pipeline shown with black arrows. Red arrows indicate the data used for the individual loss terms.}
\label{fig:loss_overview}
\end{figure*}
We model a number of loss functions, each of which contributes to enhancing the quality of the motion generated from the motion manifold from different perspectives.
To reduce the reconstruction loss, we employ two kinds of loss functions: Motion reconstruction loss $L_R$ that encourages a motion to be reconstructed after going through the encoder and decoder, and manifold reconstruction loss $L_M$ that helps a latent vector be reconstructed after going through the decoder and encoder.
In addition, we include Wasserstein loss $L_W$ that penalizes the discrepancy between $P_Z$ and the distribution $E_Z$ induced by the encoder, and an adversarial loss $L_G$ to achieve more natural motions from the motion manifold.
Figure~\ref{fig:loss_overview} shows overview of our loss functions.
\paragraph*{Motion reconstruction loss}
\label{par:motion_recon_loss}
The motion reconstruction loss penalizes the difference between the motion and the reconstructed motion, which is obtained by encoding the motion followed by decoding it.
Specifically, we measure the discrepancy of both the joint rotation angle $\vq$ and the joint position $\vp$ as follows:
\begin{align}
L_R &= L_{ang} + w_p L_{pos} \label{eq:motion_recon_loss} \\
L_{ang} &= \sum_i^{n_{joint}} \parallel \widehat{\vq}_{i}^r - \vq_i \parallel + \parallel \widehat{\vq}_{i}^v - \vq_i \parallel \\
L_{pos} &= \sum_i^{n_{joint}} \parallel \widehat{\vp}_{i}^r - \vp_i \parallel + \parallel \widehat{\vp}_{i}^v - \vp_i \parallel,
\end{align}
where $\parallel \cdot \parallel$ is the Euclidean norm and $w_p$ (= 5 in our experiment) is the weight of the position error.
\paragraph*{Manifold reconstruction loss}
\label{par:manifold_recon_loss}
A latent code sampled from the latent distribution should be reconstructed after decoding and encoding.
Manifold reconstruction loss encourages this reciprocal mapping between the motions and the manifold space.
To this end, we apply $L_{1}$ loss similar to \cite{lee2018diverse}.
We draw a motion manifold vector $Z$ from the encoded motion sequences and reconstruct it with $\widehat{Z}^r = Enc(Dec^r(Z))$ and $\widehat{Z}^v = Enc(Dec^v(Z))$, where $Z = Enc(\mQ_{t:(t+\Delta t -1)})$.
\begin{equation} \label{eq:manifold_recon_loss}
L_M = \, \parallel \widehat{Z}^r - Z \parallel_1 \, + \parallel \widehat{Z}^v - Z \parallel_1
\end{equation}
\paragraph*{Wasserstein regularizer loss}
\label{par:wa_loss}
In order to make the manifold space have a particular desired prior distribution so that we can efficiently sample from the distribution, we use the Wasserstein regularizer that penalizes deviation of the distribution $E_Z$ of the latent manifold from the desired prior distribution $P_Z$.
\begin{equation} \label{eq:MMD}
L_W = \text{MMD}_k(P_Z,E_Z),
\end{equation}
where $P_Z(Z) = \mathcal{N}(Z; \, \mathbf{0}, \, \sigma_z^2 \cdot \mathbf{I_d})$ is modeled as the multivariate normal distribution with $\sigma_z^2$ being decided through validation. We use the maximum mean discrepancy $\text{MMD}_k$ to measure the divergence between two distributions with the inverse multi-quadratics kernel $k(x,y)=C/(C+\parallel x-y \parallel_2^2)$ with $C=2Z_{\text{dim}}\sigma_z^2$. We set $\sigma_z^2=1$ and the dimension of motion manifold space $Z_{\text{dim}}=64$.
\paragraph*{Adversarial loss}
\label{par:ad_loss}
Finally, we employ the least squares generative adversarial network (LSGAN) to match the distribution of generated motion to the real motion data distribution, i.e., to promote motions generated by our model to be indistinguishable from real motions.
\begin{equation} \label{eq:dis_loss}
\begin{aligned}
L_D =&\frac{1}{2} \sum_{\widehat{\mQ}_{t:(t+\Delta t -1)}} \left[D(\widehat{\mQ}_{t:(t+\Delta t -1)})-0 \right]^2 \,\, + \\ &\frac{1}{2} \sum_{\mQ_{t:(t+\Delta t -1)}} \left[D(\mQ_{t:(t+\Delta t -1)})-1 \right]^2
\end{aligned}
\end{equation}
\begin{equation} \label{eq:GAN_loss}
L_G =\frac{1}{2} \sum_{\widehat{\mQ}_{t:(t+\Delta t -1)}} \left[D(\widehat{\mQ}_{t:(t+\Delta t -1)})-1 \right]^2 \,\,\,\,\,\,
\end{equation}
where the discriminator $D$ tries to distinguish between the reconstructed motions and the real motions.
The discriminator is then used to help our decoder generate realistic motions.
\paragraph*{Total loss}
\label{par:total_loss}
We jointly train the encoder, joint rotation decoder, joint velocity decoder and discriminator to optimize the total objective function, which is a weighted sum of the reconstruction loss, Wasserstein regularizer loss and adversarial loss.
The total objective function of manifold network is:
\begin{equation} \label{eq:GAN_loss}
\begin{aligned}
\min_{Enc,\, Dec^r,\, Dec^v} & \,\, L(Enc, Dec^r, Dec^v) \\
& = L_R + \lambda_M \,\, L_M +\lambda_W \,\, L_W + \lambda_G \,\, L_G
\end{aligned}
\end{equation}
and the discriminator loss is:
\begin{equation} \label{eq:GAN_loss}
\min_{D} \,\, L(D) = \lambda_G \,\, L_D,
\end{equation}
where weighting parameters $\lambda_M$, $\lambda_W$ and $\lambda_G$ are $0.001$, $0.1$, and $0.001$ determined through validation.
\section{Data pre-processing}
\label{sec:data}
We tested our method with H3.6M dataset.
Every motion in the dataset has the same skeletal structure.
All the poses are represented with the position and orientation of the root and the joint rotations expressed with the exponential coordinates.
For the training, motion clips of 150 frames are randomly selected from the input motion sequence and used to learn a motion manifold.
The root position in the transverse plane is removed and other data are normalized for better performance.
We will explain how motion dataset is processed.
\paragraph*{H3.6M dataset}
H3.6M dataset \cite{h36m_pami} consists of 15 activities such as walking, smoking, discussion, taking pictures, and phoning performed by 7 subjects.
We reduce 32 joints in the original data to 17 joints by removing redundant joints as done by \cite{martinez2017human}, and configured all data to have a frame rate of 25 Hz.
Therefore, 150 frames motion applied to our model cover 6 seconds.
The activities of subject S5 were used as the test data and those of the remaining subjects S1, S6, S7, S8, S9 and S11 were used as the training data.
{\color{green} Some motion data contain noises such as joint popping, but was used without noise removal.}
\section{Experimental Results}
\label{sec:result}
We perform several experiments to evaluate the performance of our method.
First, we compare the reconstruction accuracy of the proposed model with its own variations with some components ablated as well as the sequence-to-sequence model proposed by \cite{martinez2017human}.
Next, we test random sampling, motion interpolation via motion manifold, and motion denoising, followed by an experiment for motion analogies. {\color{green} For these tests, we use the joint rotation decoder to generate motions.} We qualitatively compare the result of motion interpolation and motion analogies with that of \cite{holden2015learning} \footnote{\cite{holden2015learning} is not compared with ours with respect to the reconstruction quality as it deals only with joint positions and not joint angles.}. All experiments were conducted with test sets not included in the training set. The supplemental video shows the resulting motions from the experiments.
\subsection{Motion and manifold reconstruction}
\label{subsec:result_recon}
\begin{table*}[t]
\centering
\resizebox{0.9\textwidth}{!}{%
\begin{tabular}{@{}ccccccccccccc@{}}
\toprule
\multicolumn{2}{c}{} & \multicolumn{2}{c}{1.2s} & \multicolumn{2}{c}{2.4s} & \multicolumn{2}{c}{3.6s} & \multicolumn{2}{c}{4.8s} & \multicolumn{2}{c}{6.0s} & \\ \cmidrule(lr){3-12}
\multicolumn{2}{c}{\multirow{-2}{*}{Model}} & $E_r$ & $E_p$ & $E_r$ & $E_p$ & $E_r$ & $E_p$ & $E_r$ & $E_p$ & $E_r$ & $E_p$ & \multirow{-2}{*}{$E_z$} \\ \midrule
& rot & 0.889 & 0.957 & 0.971 & 0.978 & 0.990 & 1.040 & 1.097 & 1.078 & 1.195 & \multicolumn{1}{c|}{1.181} & 0.317 \\
\multirow{-2}{*}{{$\mathbf{S}$}} & vel & - & - & - & - & - & - & - & - & - & \multicolumn{1}{c|}{-} & - \\
& rot & \textbf{0.823} & 0.855 & {\ul 0.868} & 0.923 & 0.925 & 0.999 & 1.039 & 1.032 & 1.164 & \multicolumn{1}{c|}{1.167} & 0.264 \\
\multirow{-2}{*}{{$\mathbf{D}$}} & vel & {\ul 0.856} & 0.889 & \textbf{0.843} & 0.889 & {\ul 0.877} & 0.961 & 1.008 & 1.081 & 1.127 & \multicolumn{1}{c|}{1.212} & 0.259 \\
& rot & 1.020 & 0.561 & 1.099 & 0.682 & 1.110 & 0.706 & 1.195 & 0.761 & 1.261 & \multicolumn{1}{c|}{0.822} & 0.196 \\
\multirow{-2}{*}{{{$\mathbf{DK}$}}} & vel & 1.347 & 0.600 & 1.353 & 0.698 & 1.323 & 0.723 & 1.382 & 0.756 & 1.391 & \multicolumn{1}{c|}{{\color[HTML]{000000} 0.809}} & 0.288 \\
& rot & 0.986 & {\ul 0.549} & 1.077 & \textbf{0.657} & 1.094 & {\ul 0.679} & 1.180 & 0.726 & 1.251 & \multicolumn{1}{c|}{0.810} & 0.188 \\
\multirow{-2}{*}{{{$\mathbf{DKG}$}}} & vel & 1.343 & 0.589 & 1.345 & 0.682 & 1.332 & 0.702 & 1.405 & 0.765 & 1.415 & \multicolumn{1}{c|}{ 0.834} & 0.307 \\
& rot & 0.997 & \textbf{0.541} & 1.066 & {\ul 0.659} & 1.084 & \textbf{0.668} & 1.162 & \textbf{0.696} & 1.258 & \multicolumn{1}{c|}{\textbf{0.780}} & 0.182 \\
\multirow{-2}{*}{{{$\mathbf{DKGM}$}} (ours)} & vel & 1.356 & 0.590 & 1.381 & 0.673 & 1.338 & 0.694 & 1.400 & 0.735 & 1.406 & \multicolumn{1}{c|}{0.792} & 0.293 \\
& rot & 0.906 & 0.629 & 0.909 & 0.730 & 0.886 & 0.724 & {\ul 0.954} & 0.754 & {\ul 1.053} & \multicolumn{1}{c|}{{\ul 0.788}} & {\ul 0.164} \\
\multirow{-2}{*}{{{$\mathbf{DKGMZ}$}}} & vel & 0.877 & 0.635 & 0.883 & 0.703 & \textbf{0.848} & 0.689 & \textbf{0.916} & {\ul 0.706} & \textbf{1.030} & \multicolumn{1}{c|}{0.815} & \textbf{0.157} \\
& rot & - & - & - & - & - & - & - & - & - & \multicolumn{1}{c|}{-} & - \\
\multirow{-2}{*}{Seq2seq} & vel & 0.875 & 0.863 & 0.870 & 0.954 & 0.891 & 1.059 & 1.039 & 1.177 & 1.154 & \multicolumn{1}{c|}{1.258} & 0.216 \\ \bottomrule
\end{tabular}
}
\caption{{\color{green} Reconstruction errors of joint angles ($E_r$) and joint positions ($E_p$) at sample time frames, and the reconstruction error of the manifold vector ($E_z$). The error is measured with respect to the general actions (all the actions in the DB) in H3.6M dataset.}}
\label{table:H3.6M_general}
\end{table*}
We assess the {accuracy} of the reconstructed motion $\widehat{\mQ}$ with respect to the input motion $\mQ$,
as well as the {accuracy} of the reconstructed motion manifold vector $\widehat{\vz}$ with respect to the motion manifold vector $\vz$ obtained by encoding a motion. The results are provided in Table~\ref{table:H3.6M_general}.
Generally, the reconstruction {accuracy} and the data generation quality of a manifold conflict with each other to some degree. As our purpose is to achieve a motion manifold that supports not only the motion reconstruction but also motion generation, it is important to strike a balance among various performance measures, and our method should not be evaluated only by the reconstruction {accuracy}. This trade off will be discussed in Sec. \ref{subsec:tradeoff}.
The sequence-to-sequence model (Seq2seq) compared with ours is based on \cite{martinez2017human}. The only difference is that a fully connected layer of 64 dimension is implemented between the encoder and the decoder to construct a motion manifold.
{\color{green}
For ablation study, we prepare a set of variations of our model.
The most basic model, denoted {$\mathbf{S}$}, has only joint rotation decoder with reconstruction and Wasserstein regularizer losses, without the FK layer in the network. Next model {$\mathbf{D}$} is the dual decoder model by adding the velocity decoder. From the dual model, we make variations by incrementally accumulating FK layer ({{$\mathbf{DK}$}}), adversarial loss ({{$\mathbf{DKG}$}}), manifold reconstruction loss ({{$\mathbf{DKGM}$}}, our method).
The last variation {{$\mathbf{DKGMZ}$}} is made by concatenating the manifold vector to the decoder input, i.e., $\left[ \widehat{\vq}_{i+1}, \, Z \right]$ is used instead of $\widehat{\vq}_{i+1}$ in Eqs. \ref{eq:out_rot} and \ref{eq:de_vel}. The idea of this last variation is to prevent the decoder from forgetting the motion manifold vector.
}
All variations
have the same network weight dimensions and hyper-parameters as our model. Supplemental material includes details of implementing the compared models. All models are trained with datasets that include all action categories.
The {accuracy} of the motion reconstruction is evaluated for both the joint rotation decoder ($Dec^r$) and the joint velocity decoder ($Dec^v$).
Both the Euclidean distances of joint angle errors ($L_{ang}$, also denoted as $E_r$) and joint position errors ($L_{pos}$ or $E_p$) are {\color{green} used for each decoder for the reconstruction loss}.
As for the reconstruction quality of the motion manifold vector, we measure the $L_1$-norm ($E_z$) of the difference between the motion manifold vector $\vz$ obtained by encoding a motion sequence and the reconstructed vector $\widehat{\vz}^r$ obtained by sequentially decoding $\vz$ and encoding it.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{figures/reconstruction.jpeg}
\caption{Ground truth motions (green) and reconstruction results (coral) of our method from H3.6M dataset.}
\label{fig:recon}
\end{figure}
Table~\ref{table:H3.6M_general} shows the reconstruction errors of our method and others for the datasets containing all action categories (15 actions in H3.6M dataset).
The reported errors are the average of 30 motions randomly selected from a test dataset.
A total of 150 frames are divided into 5 intervals, and errors ($E_r$, $E_p$) are measured for each interval to investigate the temporal characteristic.
The lowest and the next lowest errors are marked in bold and with underline, respectively.
{\color{green}
We first compare with respect to $E_r$ and $E_p$ errors.
Comparing {$\mathbf{S}$} and {$\mathbf{D}$}, the latter has lower $E_r$ and $E_p$ errors, which suggests that the joint rotation and velocity decoders complement with each other to reduce the errors.
Comparing {$\mathbf{D}$} and {{$\mathbf{DK}$}}, the latter reduces $E_p$ error significantly while only mildly sacrificing $E_r$ error.
{{$\mathbf{DKG}$}} has lower $E_r$ and $E_p$ errors than {{$\mathbf{DK}$}}, but higher errors than {$\mathbf{D}$} and {$\mathbf{S}$}.
This shows that adversarial loss slightly reduces reconstruction error. However, it turns out that the adversarial loss helps reconstruct the original behaviors, as will be discussed in Sec.~\ref{subsec:ZAd}.
Examining the error of {{$\mathbf{DKGM}$}} and {{$\mathbf{DKGMZ}$}}, we can see that adding manifold reconstruction loss does not significantly affect the reconstruction errors while explicitly feeding the manifold vector to the decoder helps reduce the errors.
Next, we examine manifold reconstruction error, $E_z$ ({\color{green} = $L_M$}).
Comparing {$\mathbf{D}$} and {$\mathbf{S}$}, it is remarkable that {$\mathbf{D}$} reduces $E_z$ error even without any manifold-related loss term.
However, adding FK layer to reduce joint position error slightly increases $E_z$ for the velocity decoder while it is decreased for the rotation decoder. Comparing {{$\mathbf{DK}$}} and {{$\mathbf{DKG}$}}, we can see that adversarial loss has negligible effect to the manifold reconstruction error.
Subsequently, {{$\mathbf{DKGM}$}} reduces $E_z$ slightly by adding the manifold reconstruction error, and {{$\mathbf{DKGMZ}$}} achieves the lowest $E_z$ error by explicitly feeding the manifold vector to the decoder.
{\color{green} Seq2seq~\cite{martinez2017human} shows less $E_r$ than our model, but $E_p$ is higher. In addition, our model shows better $E_z$ errors with respect to rotation decoder.}
Figure~\ref{fig:recon} visualizes the reconstruction results with our model over time in comparison with the ground truth input motion.
}
\subsubsection{Trade off between joint angle, joint position and motion manifold}
\label{subsec:tradeoff}
\begin{figure*}[th]
\centering
\subfigure[$L_{ang}$ with respect to $\lambda_W$]
{\includegraphics[width=0.33\linewidth]{figures/MMD_joint_anlge.png}}
\hfill
\subfigure[$L_{pos}$ with respect to $\lambda_W$]
{\includegraphics[width=0.33\linewidth]{figures/MMD_joint_pos.png}}
\hfill
\subfigure[$L_{M}$ with respect to $\lambda_W$]
{\includegraphics[width=0.33\linewidth]{figures/MMD_manifold.png}}
\hfill
\caption{Reconstruction errors of joint angle, joint position and manifold according to training step while adjusting $\lambda_W$ for H3.6M dataset.}
\label{fig:tradeoff}
\end{figure*}
This experiment examines the effect of different settings of the weight $\lambda_W$ for the regularization on the reconstruction errors ($L_{ang}$ and $L_{pos}$) and on motion manifold ($L_M$) on the test set. We employed {$\mathbf{D}$} model for this experiment to exclude the effect of other loss terms.
Figure~\ref{fig:tradeoff} (a) and (b) show that the joint reconstruction errors decrease as $\lambda_W$ becomes smaller, which makes {$\mathbf{D}$} model closer to a pure autoencoder, sacrificing the ability to enforce a prior over the motion manifold space while obtaining better reconstruction loss. For the same reason, Fig.~\ref{fig:tradeoff} (c) shows that the motion manifold reconstruction error $L_M$ decreases as $\lambda_W$ becomes larger.
As our goal is to obtain an effective motion manifold that is able to generate realistic motions, it is important to find a suitable set of weight parameters that compromise among different qualities.
{\color{green}
\subsubsection{Adversarial loss and explicit feeding manifold vector}
\label{subsec:ZAd}
\begin{figure}[h]
\centering
\includegraphics[width=0.9\columnwidth]{figures/ablation_study.jpeg}
\caption{{\color{green} Reconstruction results of different loss combinations for a posing while walking motion. Supplementary video includes full motions.} }
\label{fig:ablation}
\end{figure}
Here we discuss the effects of adversarial loss (Sec.~\ref{par:ad_loss}) and explicitly feeding motion manifold vector to the decoders on motion quality.
First, Table~\ref{table:H3.6M_general} shows that {{$\mathbf{DKG}$}} decreases $E_r$ from {{$\mathbf{DK}$}} only slightly. However, Fig.~\ref{fig:ablation} shows that {{$\mathbf{DK}$}} cannot properly reconstruct the original motion,
reconstructing only posing motion from the original motion of posing with walking.
In contrast, {{$\mathbf{DKG}$}} improves the overall motion quality by better reconstructing the behaviors in the original motion.
Comparing our method ({{$\mathbf{DKGM}$}}) and {{$\mathbf{DKGMZ}$}}, the latter results in lower $E_r$ and $E_p$ than our method as shown in Table.~\ref{table:H3.6M_general}.
However, Fig.~\ref{fig:ablation} reveals that {{$\mathbf{DKGMZ}$}} fails to capture walking motion. We conjecture that directly feeding manifold vector to decoder reduces reconstruction loss by explicitly retaining the motion manifold vector, but tends to converge to mean pose.
In contrast, our method successfully reconstructs the original posing with walking behavior.
This observation suggests that, while the joint reconstruction error is an important indicator of motion quality, it may not appropriately assess the motion quality in terms of reconstructing the original behaviors.
}
\subsection{Random motion samples}
\label{subsec:sampling}
To verify whether the latent motion manifold can create meaningful motions, we randomly sampled $P_Z$ and decoded to obtain motions.
We extracted 30 random samples from the motion manifold learned with H3.6M dataset.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{figures/random_sampling_compare.jpeg}
\caption{Results of randomly sampling motions from the motion manifold $P_Z$.}
\label{fig:randomsampling}
\end{figure}
Figure \ref{fig:randomsampling} is the results of random sampling from $P_Z$, and one can see that our method can create various actions including sitting, crossing the legs, and resting on the wall. This result suggests that our motion manifold and decoder can create a wide range of plausible behaviors.
{\color{green}
To examine the importance of WAE, we experimented random sampling by replacing the WAE regularizer with a simple $L_2$-norm ${\lVert z \rVert}^2$ loss.
Sampled motions from this method, as shown in Fig. \ref{fig:randomsampling} (right), often show unnatural poses and extreme joint rotations.
This experiment shows that the WAE regularizer not only helps achieve the desired motion manifold distribution but also improves quality of motion sampling.
}
\subsection{Motion interpolation with latent motion manifold}
\label{subsec:interpolation}
We can interpolate two different motions by encoding them into the latent motion manifold and then performing linear interpolation between the encoded motion manifold vectors.
The resulting interpolated motion created by our method is not just frame-by-frame interpolation, but may contain meaningful transition between the input motions.
For example, interpolating sitting down motion and photo taking motion creates hand raising motion to prepare to take a picture from sitting posture.
When waiting and smoking motions are interpolated, an interesting motion that a character seems tired of waiting and starts to smoke is created.
The capability of creating such meaningful motions is due to the Wasserstein regularizer that shortens the distance between the encoded vectors by matching the motion manifold to the multivariate normal prior.
Figure \ref{fig:interpolation} and the supplemental video show the interpolated motions.
Figure~\ref{fig:inter_compare} compares our model with \cite{holden2015learning} with respect to interpolation. See supplementary material for the implementation of \cite{holden2015learning}.
For the interpolation from sitting to walking (top) and from sitting down to taking photo (bottom), our model shows a natural transition between two motions while \cite{holden2015learning} creates somewhat averaged motion between the two motions.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{figures/interpolation_compare.jpeg}
\caption{Interpolation from sitting to walking (top) and from sitting down to taking photo (bottom) made by our model (left) and \cite{holden2015learning} (right).}
\label{fig:inter_compare}
\end{figure}
\subsection{Denoising motion data}
\label{subsec:denoising}
Our motion model can denoise motion data by projecting it to the latent motion manifold and decoding the motion manifold vector to obtain a reconstructed motion.
{\color{green}
Since the motion manifold is constructed only from human motion capture data, any element in the manifold is likely to be decoded to natural motion. Therefore, denoising effect occurs when noisy motion data is projected to the motion manifold.}
We experiment on the denoising capability of our method in the similar manner as in \cite{holden2015learning}.
We generate noise corrupted motion by randomly setting joint angles to zero with a probability of 0.5, which makes half of the joint angle information meaningless.
Figure \ref{fig:noise} shows the denoised results which are quite similar to the ground truth motions.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{figures/noise.jpeg}
\caption{Denoising experiment. Three poses are shown from the noise corrupted motion (orange), denoised motion by our method (coral), and the ground truth motion (green). Two motions (top and bottom) are shown.}
\label{fig:noise}
\end{figure}
\subsection{Motion analogy}
\label{subsec:analogy}
\begin{figure}[h]
\centering
\subfigure[{Motion analogy among ``Walking with posing'', ``Walking'' and ``Sitting'' actions.}]
{\includegraphics[width=0.99\linewidth]{figures/analogy.jpeg}}
\hfill
\subfigure[{Motion analogy among ``Smoking with sitting'', ``Sitting'' and ``Walking'' actions.}]
{\includegraphics[width=0.99\linewidth]{figures/analogy2.jpeg}}
\hfill
\caption{{\color{green} Motion analogy experiments performing arithmetic operations in the motion manifold.}}
\label{fig:analogies}
\end{figure}
Through motion analogy, we can understand how our model organizes motion manifold to represent the feature of actions.
Details about analogy can be found in \cite{white2016sampling}.
We perform vector algebraic operations with the latent vectors encoded from different motions and explore how the model organizes the latent space to represent motions.
{\color{green} Figure \ref{fig:analogies} (a) shows that subtracting a motion manifold vector for ``sitting down'' motion from ``taking photo with sitting down'' motion creates a vector representing ``taking photo'' motion. The character is standing because a zero vector in our motion manifold corresponds to an idle standing motion. Subsequently, when an encoded ``walking'' motion manifold vector is added, the motion vector becomes a vector for ``taking photo with walking'' motion.
Figure \ref{fig:analogies} (b) shows a similar analogy among ``walking'', ``smoking with walking'', and ``sitting'' motions.}
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{figures/analogy_Holden.jpeg}
\caption{{\color{green} Motion analogy experiment with \cite{holden2015learning}.}}
\label{fig:anal_Holden}
\end{figure}
Figure \ref{fig:anal_Holden} shows the experiments of performing analogy with \cite{holden2015learning}.
{\color{green} Figure \ref{fig:anal_Holden} (top) is the result of taking photo (left) and taking photo with walking (right) that correspond to Fig. \ref{fig:analogies} (a), and Fig. \ref{fig:anal_Holden} (bottom) shows smoking and smoking with sitting to compare with Fig. \ref{fig:analogies} (b).} One can see that the motion manifold obtained with \cite{holden2015learning} does not support analogy on the motion manifold.
\section{Conclusion and future work}
\label{sec:conclusion}
In this paper, we presented a novel sequential network for constructing a latent motion manifold for modeling human motion.
The main contributions of our method are the combined decoder for the joint rotation and joint velocity, and considering both the joint rotations and positions by adding the FK layer in both decoders, which improve the reconstruction accuracy.
In addition, we composed a set of loss functions, each of which contribute to enhancing the quality of motions generated from the motion manifold space from different aspects.
The capabilities of our model have been examined through various experiments such as random sampling, motion interpolation, denoising, and motion analogy.
Our method has several limitations.
First, as a sequence-to-sequence framework, the performance of our model degrades if trained to produce motions longer than 10 seconds.
{\color{green}
The supplementary video shows randomly generated motions with our network being trained to learn 300 frames (approx. 13 seconds). Resulting motions tend to lose details.
}
This limitation may be alleviated by employing an attention mechanism ~\cite{luong2015effective, bahdanau2014neural}.
Second, the encoded motions tend to be smoothed in the process of matching the latent motion manifold to the prior distribution through the regularizer.
For example, motions that contain frequent hand shaking, such as ``walking with dog'' or ``discussion'' motions in H3.6M dataset, lose fine details when reconstructed. Overcoming these limitations will be important future work.
{\color{green} We only considered joint rotations in the encoder, but incorporating additional information, such as joint positions and velocities, may be beneficial to achieve better motion qualities.
In addition, in the process of learning a motion manifold, loss terms to check validity of motions, such as joint limit, velocity limit and foot sliding, are not needed as all input motion data are considered valid. However, when an actual motion is sampled from the manifold and applied to an environment, such criteria may need to be checked.
}
Most studies on motion space learning have focused on representing a wide range of motion categories with a compact representation. In fact, the range of motion categories is only one aspect of the variedness of human motions.
Even a single motion category such as walking exhibits widely different styles depending on gender, body scale, emotion, and personality. Developing a motion manifold that can generate stylistic variations of motion is another important future research direction.
\section*{Acknowledgement}
This work was supported by Giga Korea Project (GK17P0200)
and Basic Science Research Program (NRF-2020R1A2C2011541) funded
by Ministry of Science and ICT, Korea.
\bibliographystyle{eg-alpha-doi}
| 286a2cbda1a9fdd927c81bcb49e51e5d3636d979 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{}
\begin{abstract}
We consider the problem of optimising the achievable EPR-pair distribution rate between multiple source-destination pairs in a quantum internet, where the repeaters may perform a probabilistic bell-state measurement and we may impose a minimum end-to-end fidelity as a requirement. We construct an efficient linear programming formulation that computes the maximum total achievable entanglement distribution rate, satisfying the end-to-end fidelity constraint in polynomial time (in the number of nodes in the network). We also propose an efficient algorithm that takes the output of the linear programming solver as an input and runs in polynomial time (in the number of nodes) to produce the set of paths to be used to achieve the entanglement distribution rate. Moreover, we point out a practical entanglement generation protocol which can achieve those rates.
\end{abstract}
\maketitle
\section{Introduction}
The quantum internet will provide a facility for communicating qubits between quantum information processing devices \cite{Van14, LSWK04, Kim08,WEH18}. It will enable us to implement interesting applications such as quantum key distribution \cite{bb14,E91}, clock synchronisation \cite{kkbj14}, secure multi-party computation \cite{CGS02}, and others \cite{WEH18}. To enable a full quantum internet the network needs to be able to produce entanglement between any two end nodes connected to the network \cite{MSLM13,Cal17,PKTT19,CRDW19}.
In this paper, we consider the problem of optimising the achievable rates for distributing EPR-pairs among multiple source-destination pairs in a network of quantum repeaters while keeping a lower bound on the end-to-end fidelity as a requirement. We propose a polynomial time algorithm for solving this problem and we show that, for a particular entanglement distribution protocol, our solution is tight and achieves the optimal rate. Our algorithm is inspired by \emph{multi-commodity flow} optimisation which is a very well-studied subject and has been used in many optimisation problems, including classical internet routing \cite{hu63}. In the context of a classical internet, a \emph{flow} is the total number of data packets, transmitted between a source and a destination per unit time (\emph{rate}). In this context, a \emph{commodity} is a \emph{demand}, which consists of a source, destination and potentially other requirements like the desired end-to-end packet transmission rate, quality of service, etc. In a classical network, a source and a destination can be connected via multiple communication channels as well as a sequence of repeaters and each of the communication channels has a certain \emph{capacity} which upper bounds the amount of flow it can transmit. In this context, a flow must satisfy another restriction, called \emph{flow conservation}, which says that the amount of flow entering a node (\emph{inflow}), except the source and destination node, equals the amount of flow leaving the node \footnote{Assuming that the intermediary repeater nodes do not lose packets while processing them.} (\emph{outflow}). With these constraints, one of the goals of a multi-commodity flow optimisation problem is to maximise the total amount of flows (end-to-end packet transmission rates) in a network given a set of commodities (demands). There exist \emph{linear programming} formulations for solving this problem and if we allow the flows to be a fraction then this \emph{linear programming} (LP) can be solved in polynomial time (in the number of nodes) \cite{karm84}.
In a quantum internet, we abstract the entire network as a graph $G= (V,E, C)$, where $V$ represents the set of repeaters as well as the set of end nodes, and the set of edges, $E$, abstracts the physical communication links. Corresponding to each edge we define edge capacities $C: E\rightarrow \mathbb{R}^+$, which denotes the maximum elementary EPR-pair generation rate. We assume that the fidelity of all the EPR-pairs, generated between any two nodes $u,v \in V$ such that $(u,v) \in E$, is the same (say $F$). We refer to such EPR-pair as an \emph{elementary pair} and the physical communication link via which we create such an elementary pair is called an \emph{elementary link}. Flow in such a network is the EPR-pair generation rate between a source-destination pair. Depending on the applications, the end nodes may need to generate EPR-pairs with a certain fidelity. Keeping the analogy with the classical internet, here we refer to such requirement as a demand (commodity) and it consists of four items, a source $s \in V$, a destination $e \in V$, end-to-end desired entanglement distribution rate $r$ and an end-to-end fidelity requirement $F_{\text{end}}$. We denote the set of all such demands (commodities) as $D$. In this paper, we are interested in computing the maximum entanglement distribution rate (flow). Given a quantum network $G$ and a set of demands $D$, we investigate how to produce a set of paths $\mathcal{P}_i$ and an end-to-end entanglement generation rate $r_i$ (flow), corresponding to each demand $(s_i,e_i,F_i)$, such that the total entanglement generation rate $\sum_{i=1}^{|D|} r_i$ is maximised. In the rest of this paper, we refer to this maximisation problem as \emph{rate maximisation problem}. What is more, here we also investigate what type of practical entanglement distribution protocol achieves such rate.
In the case of the quantum internet, we can use an LP for maximising the total flow $\sum_{i=1}^{|D|} r_i$. However, the working principle of quantum repeaters is different, unlike classical networks, the repeaters extend the length of the shared EPR-pairs by performing \emph{entanglement swapping} operations \footnote{Entanglement swapping is an important tool for establishing entanglement over long-distances. If two quantum repeaters, $A$ and $B$ are both connected to an intermediary quantum repeater $r$, but not directly connected themselves by a physical quantum communication channel such as fiber, then $A$ and $B$ can nevertheless create entanglement between themselves with the help of $r$. First, $A$ and $B$ each individually create entanglement with $r$. This requires one qubit of quantum storage at $A$ and $B$ to hold their end of the entanglement, and two qubits of quantum storage at $r$. Repeater $r$ then performs an entanglement swap, destroying its own entanglement with $A$ and $B$, but instead creating entanglement between $A$ and $B$. This process can be understood as repeater $r$ teleporting its qubit entangled with $A$ onto repeater $B$ using the entanglement that it shares with $B$.} \cite{MATN15, teleport93, ZZHE93,GWZ08}. However, entanglement swapping operations might be probabilistic depending on the repeater technology used in the quantum internet. This implies that the usual flow-conservation property which we use in classical networks does not hold in the quantum networks, i.e., the sum of the inflow is not always equal to the sum of the outflow. Hence, the standard multi-commodity flow-based approach cannot be applied directly.
For example, if the repeaters are built using \emph{atomic ensemble} and \emph{linear optics} then they use a probabilistic Bell-state measurement (BSM) for the entanglement swap operation \cite{SSDG11, GMLC13, SSMS14}. Due to the probabilistic nature of the BSM, the entanglement generation rate decays exponentially with the number of swap operations.
Another difficulty for using the standard multi-commodity flow-based approach for solving our problem occurs due to the end-to-end fidelity requirement in the demand. In a quantum network, the fidelity of an EPR-pair drops with each entanglement swap operation. This implies that a longer path-length results in a lower end-to-end fidelity. One can enhance the end-to-end fidelity using \emph{entanglement distillation}. However, some repeater technologies are unable to perform such quantum operations (for instance, the atomic ensemble-based quantum repeaters). Hence, for such cases, one can achieve the end-to-end fidelity requirement only by increasing the fidelity $F$ of the elementary pairs and reducing the length of the discovered path. The first of these two options, the elementary pair fidelity, depends on the hardware parameters at fabrication. The second option is related to the path-length and it is under the control of the routing algorithm that determines the path from the source to the destination. For the routing algorithms, one possible way to guarantee the end-to-end fidelity is to put an upper bound on the discovered path-lengths. The standard multi-commodity flow-based LP-formulations does not take into account this path-length constraint. However, there exists one class of multi-commodity flow-based LP-formulation, called \emph{length-constrained multi-commodity flow} \cite{MM10}, which takes into account such constraints. In this paper, our proposed LP-formulation is inspired from the length-constrained multi-commodity flow problem and it takes into account the path-length constraint.
Given these differences, one might use the LP-formulation corresponding to the standard multi-commodity flow-based approach which we described before, but this would lead to a very loose upper bound on the achievable entanglement generation rate \cite{BAKE18,pir16,Pir19,Pir19bounds,AML16,AK17,RKBK18,BA17} in this setting.
In our setting, it is not clear whether one can still have an efficient LP-formulation for the flow maximisation problem in a quantum internet. In fact, recently in \cite{li20} the authors mention that multi-commodity flow optimisation-based routing in a quantum internet may, in general, be an NP-hard problem. In this paper, we show that for some classes of practical entanglement generation protocols, one can still have efficient LP-formulation which maximises the total flow for all the commodities in polynomial time (in the number of nodes).
The organisation of our paper is as follows: section \ref{sum_contr} provides a summary of our results. In section \ref{lp_form}, we give the exact LP-formulation for solving rate maximisation problem. In section \ref{meth} we prove that our LP formulation solve the desired rate maximisation problem. Later, in the same section we show how one can achieve the entanglement generation rates proposed by the LP-formulation using an entanglement distribution protocol. We also analyse the complexity of our proposed algorithms in section \ref{meth}. We conclude our paper in section \ref{concl}.
\section{Results}
\label{contr}
\subsection{Our contributions in a nutshell}
\label{sum_contr}
In this paper, all of our results are directed towards solving the rate maximisation problem in a quantum internet, where given a quantum network and a set of demands, the goal is to produce a set of paths such that the total end-to-end entanglement generation rate is maximised and in addition for each of the demands the end-to-end fidelity of the EPR-pairs satisfy a minimal requirement.
In this section, we summarise our contributions.
\begin{itemize}
\item In order to solve the maximisation problem, we propose an LP-formulation called \emph{edge-based formulation} where both the number of variables and the number of constraints as well as the algorithm for solving such LPs scale polynomially with the number of nodes in the graph. However, it is non-trivial to see whether this formulation provides a valid solution to the problem or not. In this paper, by showing the equivalence between the edge-based formulation and another intuitive LP-formulation, called \emph{path-based formulation}, we show that the edge-based formulation provides a valid solution.
\item A disadvantage of the solution of the edge-based formulation is that it only gives the total achievable rate, not the set of paths which the underlying entanglement distribution protocol would use to distribute the EPR-pairs to achieve such rate. In this paper, we provide an algorithm, called the \emph{path extraction algorithm}, which takes the solutions of the edge-based formulation and for each of the commodities it extracts the set of paths to be used and the corresponding entanglement distribution rate along that path. The worst case time complexity of this algorithm is $O(|V|^4|E||D|)$, where $|V|, |E|$ denote the total number of nodes and edges in the network graph $G$ and $|D|$ denotes the total number of demands. What is more, we point out that there exists a practical entanglement distribution protocol along a path, called the \emph{prepare and swap protocol}, which achieves the rates (asymptotically) proposed by path extraction algorithm.
\end{itemize}
\subsection{From the fidelity constraint to the path-length constraint}
\label{fid_to_len}
In a quantum network, the fidelity of the EPR-pairs drops with each entanglement swap operation. The fidelity of the output state after a successful swap operation depends on the fidelity of the two input states. If a mixed state $\rho$ has fidelity $F$, corresponding to an EPR-pair (say $|\Psi^+\rangle = \frac{1}{\sqrt{2}}(|00\rangle + |11\rangle)$) then the corresponding Werner state \cite{wern89} with parameter $W$ can be written as follows,
\begin{align*}
&\rho = W|\Psi^+\rangle \langle ^+\Psi| + \frac{1-W}{4}~\mathbb{I}_4,
\end{align*}
where $\mathbb{I}_4$ is the identity matrix of dimension $4$. The fidelity of this state is $\frac{1+3W}{4}$.
In this paper we assume that all the mixed entangled states in the network are Werner states. The main reason is that Werner states can be written as mixing with isotropic noise and hence form the worst case assumption. For the Werner states, if a node performs a noise-free entanglement swap operation between two EPR-pairs with fidelities $F$, then the fidelity of the resulting state is $\frac{1+3W^2}{4} $ which is equal to $1+ \frac{3}{4}\left(\frac{4F-1}{3}\right)^2$\cite{BDCZ98}.
Here, each demand $(s_i,e_i,F_i)$ (where $1\leq i\leq |D|=k$) has $F_i$ as the end-to-end fidelity requirement. We assume that the fidelity of each of the elementary pairs is lower bounded by a constant $F$. Note that, in our model, we do not consider entanglement distillation, so in order to have a feasible solution, here we always assume that the fidelity requirement of the $i$-th demand, $F_i$ is at most the fidelity of the elementary pair $F > 0.5$.
Corresponding to a demand $(s_i,e_i,F_i)$, if we start generating the EPR-pairs along a path $p = ((s_i,u_1), (u_1,u_2), \ldots , (u_{|p|-1},e_i))$, then the total number of required entanglement swap operations is $|p|-1$, where the path-length is $|p|$. As with each swap operation the fidelity drops exponentially, this implies the end-to-end fidelity will be $\frac{1+3W^{|p|}}{4} = 1+ \frac{3}{4}\left(\frac{4F-1}{3}\right)^{|p|}$. In order to satisfy the demand, $1+ \frac{3}{4}\left(\frac{4F-1}{3}\right)^{|p|}$ should be greater than $F_i$, i.e., $1+ \frac{3}{4}\left(\frac{4F-1}{3}\right)^{|p|} \geq F_i$. From this relation, we get the following constraint on the length of the path.
\begin{equation}
\label{path_length}
|p| \leq \left\lfloor\frac{\log \left(\frac{4F_i-1}{3}\right)}{\log \left(\frac{4F-1}{3}\right)}\right\rfloor.
\end{equation}
This implies, for the $i$-th demand all the paths should have length at most $\left\lfloor\frac{\log \left(\frac{4F_i-1}{3}\right)}{\log \left(\frac{4F-1}{3}\right)}\right\rfloor$. In the rest of the paper, for the $i$-th demand we assume,
\begin{equation}
\label{len_const}
l_i := \left\lfloor\frac{\log \left(\frac{4F_i-1}{3}\right)}{\log \left(\frac{4F-1}{3}\right)}\right\rfloor.
\end{equation}
Using this constraint on the number of intermediate repeaters, we can rewrite the demand set $D$ in following way,
\begin{equation}
\label{eq:dem_len}
D = \{(s_1,e_1,l_1), \ldots , (s_k,e_k,l_k)\}.
\end{equation}
\subsection{LP-formulation}
\label{lp_form}
In this section, we construct the LP-formulation for computing the maximum flow in a quantum network. For the simplicity, we consider the network $G=(V,E,C)$ as a directed graph and construct all the LP-formulations accordingly. Note that, one can easily extend our result to an undirected graph, just by converting each of the edges which connects two nodes $u,v$ in the undirected graph into two directed edges $(u,v)$ and $(v,u)$.
For the entanglement distribution rate, here we let the achievable rate between two end nodes $s_i,e_i \in V$ along a repeater chain (or a path) $p = ((s_i,u_1), (u_1,u_2), \ldots , (u_{|p|-1},e_i))$ be $r_p$ such that,
\begin{equation}
\label{eq:rate}
r_p \leq (q)^{|p|-1}\text{Min}\{C(s_i,u_1), \ldots ,C(u_{|p|-1},e_i)\},
\end{equation}
where $C(u,v)$ denotes the capacity of the edge $(u,v)\in E$ and $|p|$ is the length of the path and $q$ is the success probability of the BSM. Later, in section \ref{meth} we show that there exists a practical protocol called prepare and swap which achieves this rate requirement along a path. For an idea of such protocol we refer to the example of figure \ref{exam0}. In the next section, we give the LP-construction of the edge-based formulation.
\subsubsection{Edge-based formulation}
\label{lbf_net}
In this section, we give the edge-based formulation for solving the rate maximisation problem. In this formulation, we assign one variable to each of the edges of the network. As the total number of edges, $|E|$, in a graph of $|V|$ nodes scales quadratically with the number of nodes in the graph, the total number of variables is polynomial in $|V|$. This makes the edge-based formulation efficient. However, it is challenging to formulate the path-length constraint in this formulation. The main reason is that, an edge can be shared by multiple paths of different lengths and the variables of the edge-based formulation corresponding to that edge do not give any information about the length of the paths. In this paper, we borrow ideas from the length-constrained multi-commodity flow \cite{MM10} to handle this problem. In order to implement the length constraint we need to modify the network graph $G$ as well as the demand set $D$. In the next section, we show how to modify the network graph and the demand set.
\vspace{0.2in}
\textbf{Network modification.} To implement the length constraint in the edge-based formulation, we define an expanded graph $G' = (V',E', C')$ from $G = (V,E,C)$ such that it contains $l_{\mathrm{max}} +1$ copies of each of the nodes, where $l_{\mathrm{max}} = \max\{l_1, \ldots , l_k\}$ and for all $1\leq i \leq k$, $l_i$ denotes the length constraint of the $i$-th demand $(s_i,e_i,l_i)$. For a node $u \in V$, we denote the copies of $u$ as $u^0,u^1, \ldots , u^{l_{\mathrm{max}}}$. We denote $V^j$ as the collection of the $j$-th copy of all the nodes. This implies, $V' = \bigcup_{j=0}^{l_{\mathrm{max}}}V^j$. For each edge $(u,v) \in E$ and for each $0\leq j < l_{\mathrm{max}}$, we connect $u^j \in V'$ and $v^{j+1} \in V'$ with an edge, i.e., $(u^j , v^{j+1}) \in E'$. For each edge $(u^j,v^{j+1}) \in E'$ we define $C'(u^j,v^{j+1}) := C(u,v)$. In figure \ref{exam3} we give an example of this extension procedure corresponding to a network graph of figure \ref{exam_glob}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth, left]{examp_glob_mod.pdf}
\caption{Network Graph $G = (V,E, C)$, with demand $D = \{(s,e,2)\}$ and $q = \frac{1}{2}$, i.e., here the source, $s$ wants to share EPR-pairs with $e$. In this network, for each edge $(u,v)\in E$, the quantity, $C(u,v)$ denotes the EPR-pair generation rate corresponding to that edge. }
\label{exam_glob}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\textwidth, left]{examp_glob_fid_mod.pdf}
\caption{Extended Network Graph $G' = (V',E', C')$ of the original graph in figure \ref{exam_glob}. Here we are interested in finding the paths between $s$ and $e$ with path-length at most $2$. For the construction of $G'$, we create three copies $u^0,u^1,u^2$ of each of the nodes $u \in G$. There are five nodes $s,u,v,w,e$ in the graph of figure \ref{exam_glob}. This implies, in this modified graph we have $(s^0,s^1,s^2),(u^0,u^1,u^2), (v^0,v^1,v^2),(w^0,w^1,w^2),(e^0,e^1,e^2)$, fifteen nodes. In this original graph of figure \ref{exam_glob}, if $u$ is connected to $v$, then in this modified graph, we connect, $u^0$ with $v^1$ and $u^1$ with $v^2$. Note that, all the paths from $s^0$ to $e^1$ or $e^2$ has hop length at most $2$. In this modified graph the new demand set corresponding to the demand $D = \{(s,e)\}$ in the original graph $G$ in figure \ref{exam_glob} is $ D_{\mathrm{mod}} = \{(s^0,e^1),(s^0,e^2)\} $.}
\label{exam3}
\end{figure}
According to this construction, the length of all the paths in $G'$ from $s^0_i$ to $e^j_i$ is exactly $j$ and all the paths from $s^0_i$ to $e^1_i, \ldots , e^{l_i}_i$ have a path-length at most $l_i$. This implies, for the edge formulation, the $i$-th demand $(s_i,e_i,l_i)$ can be decomposed into $l_i$ demands $\{(s^0_i,e^1_i), \ldots , (s^0_i,e^{l_i}_i)\}$. This implies the total modified demand set would become,
\begin{equation}
\label{mod_dem}
D_{\mathrm{mod}} := \{D_1, \ldots , D_{k}\},
\end{equation}
where for each $1 \leq i \leq k$, $D_i = \{(s^0_i,e^1_i), \ldots , (s^0_i,e^{l_i}_i)\}$. Note that, the new demand set $D_{\mathrm{mod}}$ doesn't have any length or fidelity constraint.
\vspace{0.2in}
\textbf{Edge-based formulation.} Here, we give the exact LP-construction of the edge-based formulation. Note that, one can use a standard LP-solver (in this paper we use Python 3.7 pulp class \cite{LPPulp}) to solve this LP (see figure \ref{graph_exam_surf} for an example).
In this LP-construction, for the $i,j$-th demand $(s^0_i,e^j_i) \in D_i$ (where $D_i \in D_{\mathrm{mod}}$), we define one function $g_{ij}: E' \rightarrow \mathbb{R}^+$. The value of this function $g_{ij}(u,v)$, corresponding to an edge $(u,v) \in E'$ denotes the flow across that edge for the $i,j$-th demand. We give the edge-based formulation in table \ref{EBF}.
\begin{table}[ht]
\begin{mdframed}
\begin{align}
\nonumber
&\text{Maximize}\\ \label{edge_form_plen3}
&\sum_{i=1}^k\sum_{j=1}^{l_i} (q)^{j-1}\sum_{\substack{v^1:(s^0_i,v^1) \in E'}} g_{ij}(s^0_i,v^1). \\ \nonumber
&\text{Subject to:}\\ \nonumber
&\text{For all } 1 \leq i \leq k, 1 \leq j \leq l_{i}, 0 \leq t \leq l_{\mathrm{max}}-1, (u,v) \in E,\\ \label{const_edge31}
& g_{ij}(u^t,v^{t+1}) \geq 0, \text{ and} \\ \label{const_edge32}
& \sum_{i=1}^k \sum_{j=1}^{l_i} \sum_{t=0}^{l_{\mathrm{max}} -1}g_{ij}(u^t,v^{t+1}) \leq C(u,v).\\ \nonumber
&\text{For all } 1 \leq i \leq k, 1 \leq j \leq l_{i}, u',v',w' \in V' : v \neq s^0_i, v \neq e^{j}_i , \\ \label{const_edge33}
&\sum_{u': (u',v') \in E'} g_{ij}(u',v')= \sum_{w': (v',w') \in E'} g_{ij}(v',w').
\end{align}
\end{mdframed}
\caption{Edge-based formulation}
\label{EBF}
\end{table}
In the objective function equation \ref{edge_form_plen3} of the edge-based formulation, the sum $\sum_{\substack{v^1:(s^0_i,v^1) \in E'}}g_{ij}(s^0_i,v^1)$ denotes the entanglement distribution rate between $s^0_i,e^j_i$, for a fixed $i,j$, when $(q) = 1$. Note that, according to the construction of the graph $G'$, all the paths between $s^0_i,e^j_i$ have path-length exactly $j$. This implies that $(q)^{j-1}\sum_{\substack{v^1:(s^0_i,v^1) \in E'}}g_{ij}(s^0_i,v^1)$ denotes the entanglement distribution rate between $s^0_i,e^j_i$ and $\sum_{i=1}^k\sum_{j=1}^{l_i} (q)^{j-1}\sum_{\substack{v^1:(s^0_i,v^1) \in E'}} g_{ij}(s^0_i,v^1)$ denotes the total entanglement distribution rate for all the demands. The condition in equation \ref{const_edge32} represents the capacity constraint and condition equation \ref{const_edge33} denotes the flow conservation property.
Although this construction is efficient, it does not give any intuition whether it will solve the rate maximisation problem. In section \ref{meth} we give another intuitive LP-formulation, called \emph{path-based formulation} and we explain the equivalence between the path-based and the edge-based formulation. This proof guarantees that the solution of the edge-based formulation gives a solution for the rate maximisation problem.
\begin{figure}
\includegraphics[width=0.7\textwidth, left]{examp_0_mod.pdf}
\caption{Repeater chain network with three intermediate nodes and one source-destination pair $s,e$. Here, there is a demand to create EPR-pairs between source $s$ and destination $e$. The capacity of an edge $(u,v)$ is given by the function $C(u,v)$. All the repeaters use BSM for the entanglement swap. Here we assume that the success probability of the BSM is $q = \frac{1}{2}$.
In the prepare and swap protocol, all the intermediate repeaters perform the swap operation at the same time. This implies, the expected entanglement generation rate between $s$ and $e$ for this protocol would be $r_{s,e} = (q)^{4-1} \text{Min}\{C(s,u),C(u,v),C(v,w),C(w,e)\} =2(\frac{1}{2})^{3} = 0.25$. }
\label{exam0}
\end{figure}
\subsubsection{Path-extraction algorithm}
\label{sec:path_extrac}
The edge-based formulation, proposed in the last section is a compact LP construction and it can be solved in polynomial time. However, this solution only gives the total achievable rates for all the commodities, it does not give us any information about the set of paths corresponding to each commodity along which one should distribute the EPR-pairs to maximise the entanglement distribution rate. In this section, we give an efficient method for doing so in algorithm \ref{flow_dec3}, which takes the solution of the edge-based formulation and produces a set of paths as well as the achievable rates across each path for each of the demands. Later, in section \ref{meth} we show that the set of extracted paths satisfies the path-length constraint for each of the demands and if one uses the prepare and swap method for distributing entanglement across each of the paths then one can achieve the entanglement distribution rate suggested from this algorithm.
\begin{algorithm}
\begin{mdframed}
\vspace{0.1in}
\hspace*{\algorithmicindent} \textbf{Input: } The solution of the edge-based formulation, i.e., $\{\{g_{ij}(u',v')\}_{(u',v')\in E'}\}_{1\leq i\leq k, 1 \leq j \leq l_i}$. \\
\hspace*{\algorithmicindent} \textbf{Output: } Set of paths as well as the rate across each of the paths $\{\mathcal{P}_{i,j}\}_{1\leq i\leq k, 1 \leq j \leq l_i}$.
\begin{algorithmic}[1]
\FOR{($i=1$; $i \leq k$; $i++$)}
\FOR{($j=1$; $j \leq l_i$; $j++$)}
\State $m =0$.
\State $F_{i,j}=g_{ij}$.
\State $\mathcal{P}_{i,j} = \emptyset$.
\WHILE{$\sum_{v:(s^0_i,v^1)\in E'} F_{i,j,m}(s^0_i,v^1) >0$}
\State Find a path $p_{j,m}$ from $s^0_i$ to $e^j_i$ such that,
\State $\forall(u',v')\in p_{j,m}$, $F_{i,j,m}(u',v') >0$
\State $\tilde{r}_{p_{j,m}} = (q)^{j-1}\{\text{Min}_{(u',v')\in p_{j,m}}\{F_{i,j,m}(u',v')\}$
\State $\forall (u',v')\in p_{j,m}$,
\State $F_{i,j,m+1}(u',v') = F_{i,j,m}(u',v') - \frac{\tilde{r}_{p_{j,m}}}{(q)^{j-1}}$.
\State $\mathcal{P}_{i,j} = \mathcal{P}_{i,j} \cup (p_{j,m},\tilde{r}_{p_{j,m}})$.
\State $m=m+1$.
\ENDWHILE
\ENDFOR
\ENDFOR
\end{algorithmic}
\vspace{0.1in}
\end{mdframed}
\caption{Path Extraction and Rate Allocation Algorithm.}
\label{flow_dec3}
\end{algorithm}
\subsubsection{Example}
In this section, we give an example of our algorithms on a real world network topology $G = (V,E,C)$. In order to do so, we choose a SURFnet topology from the \emph{internet topology zoo} \cite{KNH11}. This is a publicly available example of a dutch classical telecommunication network, with $50$ nodes (see figure \ref{exam_surf}). In this network, we assume that each of the nodes in the network is an atomic ensemble and linear optics-based quantum repeater. These types of repeaters can generate elementary pairs almost deterministically \cite{GMLC13, SSMS14}, due to their multiplexing abilities. Here we assume that the elementary pair generation is a deterministic process. The elementary pair generation rate depends only on the entanglement source and its efficiency. Here, we choose the elementary pair generation rate uniformly randomly from $1$ to $400$ EPR-pairs per second. The success probability for the BSM, $q$ is considered to be $0.5$ for all the nodes. We also assume that the memory efficiency is one and all the memories are on-demand memories, i.e., they can retrieve the stored EPR-pairs whenever required \cite{GMLC13}.
\begin{figure}[h]
\includegraphics[width=0.48\textwidth, left]{surfnet.jpg}
\caption{Pictorial view of the dutch SURFnet network, taken from internet topology zoo \cite{KNH11}. In this paper, we suppose that this is a quantum network and all the nodes in this network are quantum repeater nodes and some of them are the end nodes. We also assume that the repeaters can generate EPR-pairs using the communication links, shown in this figure. We run our proposed edge-based formulation and path-extraction algorithm on this network topology for maximising the total end-to-end EPR-pairs generation rate. We refer to figure \ref{graph_exam_surf} for detailed description.}
\label{exam_surf}
\end{figure}
We additionally assume that the minimum storage time of all the memories is the maximum round trip communication time between any two nodes in the network that are directly connected by an optical fiber. In the SURFnet network, the maximum length of the optical fiber connecting any two nodes is $50$ km. Hence, the minimum storage time is $\frac{2\times 50000}{c} = 500 \mu s$, where $c$ is the speed of light in a telecommunication fiber, which is approximately $c \approx 2\times 10^8 $ meters per second. In this example, we consider the fidelity of all elementary pairs to be $F = 0.9925$. We generate the demands uniformly at random, i.e., we choose the sources and the destinations uniformly at random between $1$ and $50$. We also choose the end-to-end fidelity randomly from $0.93$ to $0.99$ for each of the demands. Substituting these fidelity constraints in equation \ref{len_const}, we obtain a maximum path-length $l_{\text{max}} = 8$. Here, we have generated only four demands and we assume that all the entanglement distribution tasks are performed in parallel. We optimise the total achievable rate using the LP solver available in the Python $3.7$ pulp class \cite{LPPulp}. The rates and the paths corresponding to the four demands are shown in figure \ref{graph_exam_surf}. An overview of the entire procedure is given in algorithm \ref{qpce} and the code of this implementation is publicly available in \cite{githubcode_rout20} and the data set can be found in \cite{data_set_routing_20}.
\begin{algorithm}
\begin{mdframed}
\vspace{0.1in}
\hspace*{\algorithmicindent} \textbf{Input: }Set of demands $D$, Network Graph $G = (V,E,C)$.\\
\hspace*{\algorithmicindent} \textbf{Output: }Set of paths $\mathcal{P}_i$ for the $i$-th demand and rate $r_p$, across each of the path $p\in \mathcal{P}_i$.
\begin{algorithmic}[1]
\State Convert the fidelity requirement $F_i$ of the $i$-th demand $(s_i,e_i,F_i)\in D$ into a path-length constraint $l_i$ (use equation \ref{len_const}).
\State Compute the modified demand set $D_{\mathrm{mod}}$ from $D$ according to the path-length constraint which we compute at the previous step (see subsection \ref{lbf_net} for details).
\State Compute the extended network $G'$ from $G$ using the procedure, described in subsection \ref{lbf_net}.
\State Implement the edge-based LP-formulation, proposed in table \ref{EBF} and compute the total maximum achievable rate $\sum_{i=1}^k r_i$ using the LP solver available in the Python $3.7$ pulp class \cite{LPPulp}.
\State For the $i$-th demand, extract the set of paths $\mathcal{P}_i$ and compute the required rate $r_p$ across each of the paths $p\in \mathcal{P}_i$, such that $r_i = \sum_{p \in \mathcal{P}_i}r_p$ using the algorithm \ref{flow_dec3}.
\end{algorithmic}
\vspace{0.1in}
\end{mdframed}
\caption{Method to solve the rate maximisation problem.}
\label{qpce}
\end{algorithm}
\begin{figure}
\includegraphics[width=0.65\textwidth, left]{example_surf.pdf}
\caption{This figure is the graphical representation of the SURFnet network of figure \ref{exam_surf}. We run our rate maximisation algorithm on this graph. Here, we consider the demand set as $D = \{(40,45,7), (21,13,6),(30,50,7),(38,15,8)\}$. The end-nodes are represented using square boxes. The capacity of each of the link is chosen uniformly at random between $[1,400]$-EPR-pairs per second. We assume that the success probability of the bell-state measurement is $q = 0.5$ and the fidelity of the each of the elementary pairs is $F=0.9925$ and the value of the corresponding Werner parameter $W= 0.99$. Given such a network and demand set, we run our edge-based formulation proposed in table \ref{EBF} and get a total achievable rate of $18.140625$ EPR-pairs per second. Next, we feed the solution to the path extraction algorithm, proposed in algorithm \ref{flow_dec3} and extract the paths for each of the demands. In this figure, all the thick edges, participate in the paths. For the demand, $(45,40,7)$ we get two paths, $((45,43),(43,47),(47,48),(48,9),(9,33),(33,40))$ and $((45, 44), (44, 47), (47, 48), (48, 9), (9, 39), (39, 40))$. Note that each of the paths has path-length $6$. Hence the end-to-end fidelity for this demand is $\frac{1+3(0.99)^6}{4} = 0.955$. Similarly, for the demand $(21,13,6)$ the extracted path is $((21, 27), (27, 28), (28, 25), (25, 20), (20, 31), (31, 13))$ and end-to-end fidelity is $\frac{1+3(0.99)^6}{4} = 0.955$. For the demand $(30,50,7)$, there is no path of length smaller than $8$. Hence, this demand can not be satisfied in this model. For the demand $(38,15,8)$, we extract four paths, $((38, 23), (23, 31), (31, 14), (14, 15))$, and $((38, 39), (39, 31), (31, 14), (14, 15))$, and $((38, 39), (39, 9),$ $(9, 48), (48, 46), (46, 15))$, $((38, 39), (39, 9), (9, 6), (6, 48),$ $(48, 47), (47, 43), (43, 15))$ and the end-to-end fidelity for the paths are $0.97$, $0.97$, $0.9625$ and $0.9475$. If we use the prepare and swap protocol for generating EPR-pairs, along each of the paths, then we can achieve the total entanglement distribution rate of $18.140625$ EPR-pairs per second.}
\label{graph_exam_surf}
\end{figure}
\section{Methods}
\label{meth}
In this section, we provide more details of the results, presented in the last section. First we propose an intuitive LP-construction, called path-based formulation for solving the rate maximisation problem. Later, we give an idea about how to prove the equivalence between both of the proposed LP-constructions. Next, we explain the prepare and swap protocol in detail and show that using this protocol one can achieve the optimised entanglement generation rates along the paths from algorithm \ref{flow_dec3}. We finish this section with the complexity analysis of the edge-based formulation and algorithm \ref{flow_dec3}.
\subsection{Path-based formulation}
\label{p_form_prep_swap}
In this formulation, for each path $p$ corresponding to each source-destination pair $s_i,e_i$, we define one variable $r_p$. This $r_p$ denotes the achievable rate between $s_i,e_i$ along the path $p$. From equation \ref{eq:rate}, we have, for all $(u,v) \in p$,
\begin{equation}
r_p \leq (q)^{|p|-1}C(u,v)
\end{equation}
Note that, for the $i$-th demand the total rate $r_i$ can be achieved via multiple paths. Let $\mathcal{P}_i$ be the set of all possible paths which connect $s_i$ and $e_i$, hence,
\begin{equation}
r_i =\sum_{p \in \mathcal{P}_{i}} r_p.
\end{equation}
From equation \ref{eq:dem_len}, we have that for all the paths $p \in \mathcal{P}_i$, $|p|\leq l_i$. We give the exact LP-formulation which takes into account all these constraints in table \ref{path_form_normal}.
\begin{table}[ht]
\begin{mdframed}
\begin{alignat}{3}
\nonumber
&\text{Maximize}~~~~~~~~ \\ \label{simp_opt_cr_swap_len1}
&\sum_{i=1}^k \sum_{p \in \mathcal{P}_i}r_p &\\ \nonumber
&\text{Subject to :}~~~~~~~\\ \label{eq:path_rate}
&\sum_{i=1}^k \sum_{\substack{p \in \mathcal{P}_i :\\ (u,v) \in p}}\frac{r_p}{(q)^{|p|-1}} \leq C(u,v), &\forall ~(u,v) \in E, \\ \nonumber
&r_p \geq 0, &\forall i \in \{1, \ldots ,k\},\forall ~p\in \mathcal{P}_i, \\ \nonumber
&|p| \leq l_i &\forall i \in \{1, \ldots ,k\},\forall ~p\in \mathcal{P}_i.
\end{alignat}
\vspace{0.1in}
\end{mdframed}
\caption{Path-Based Formulation}
\label{path_form_normal}
\end{table}
Note that, in this LP-formulation, as we introduce one variable corresponding to each path so the total number of variables is of the order $O(|V|!)$. This scaling stops us from using the path-based formulation for solving the maximum flow problem. However, this formulation helps us to prove that the edge-based formulation gives a solution to the rate maximisation problem. In the next section, we give an idea of the proof. The full details of the proof are given in the supplementary material.
\subsection{Equivalence between the two formulations}
\label{ref:equiv}
The idea of the proof of equivalence is that, first we try to construct a solution of the edge-based formulation from the solution of the path-based formulation and then we try to construct a solution of the path-based formulation from the edge-based formulation. If both of the constructions are successful then we can conclude that both of the formulations are equivalent.
In table \ref{path_form_normal} we provide the path-based formulation on the basis of the network graph $G = (V,E,C)$ and the demand set $D$, whereas in table \ref{EBF} we propose the edge-based formulation using the network graph $G' = (V',E',C')$ and the demand set $D_{\mathrm{mod}}$. In order to show the equivalence between both of the formulations, we first need to rewrite the path-based formulation using the network graph $G' = (V',E',C')$ and the demand set $D_{\mathrm{mod}}$. The next section focuses on this. After that, we focus on proving the equivalence.
\subsubsection{Path-based formulation on the modified network}
\label{p_form_prep_swap}
In this section, we construct the path-based formulation on the basis of the new demand set $D_{\mathrm{mod}}$, defined in equation \ref{mod_dem} and the modified network $G' = (V',E',C')$. In this demand set, we use the term $i,j$-th demand to denote the $j$-th source destination pair $(s^0_i,e^j_i)$ of the $i$-th demand $(s_i,e_i) \in D$. We denote the set of all possible paths for the $i,j$-th demand as $\mathcal{P}_{i,j}$ and we assign a variable $r_p$, corresponding to each path $p \in \mathcal{P}_{i,j}$. From equation \ref{eq:rate} we have, for all the edges $(u',v') \in E'$ in a path $p$,
\begin{equation}
r_p \leq (q)^{|p|-1}C'(u',v').
\end{equation}
Note that, from the construction of $G'$ and the new demand set $D_{\mathrm{mod}}$, the length of all the paths in $\mathcal{P}_{i,j}$ is $j$. This implies, if $\mathcal{P}_i$ denotes the set of all possible paths for the $i$-th demand $(s_i,e_i,l_i)\in D$, then $\mathcal{P}_i = \bigcup_{j=1}^{l_i} \mathcal{P}_{i,j}$. For a fixed source-destination pair $(s_i,e_i)$, if along a path $p$, the achievable rate is $r_p$ then,
\begin{equation}
r_i = \sum_{j=1}^{l_i} \sum_{p \in \mathcal{P}_{i,j}} r_p.
\end{equation}
We give the exact formulation in table \ref{path_form}.
\begin{table}[ht]
\begin{mdframed}
\begin{alignat}{3}
\nonumber
&\text{Maximize} \quad \quad \quad \quad \quad \quad \quad \quad\\ \label{opt_cr_swap_len1}
&\sum_{i=1}^k \sum_{j=1}^{l_i}\sum_{p \in \mathcal{P}_{i,j}}r_p \quad \quad \quad \quad \quad \quad \\ \nonumber
&\text{Subject to :}\quad \quad \quad \quad \quad \quad \quad \quad \\ \label{swp_const1}
&\sum_{i=1}^k \sum_{j=1}^{l_i} \sum_{t=0}^{l_{\mathrm{max}} -1}\hspace{-0.13in}\sum_{\substack{p \in \mathcal{P}_{i,j} :\\ (u^t,v^{t+1}) \in p}}\hspace{-0.13in}\frac{r_p}{(q)^{|p|-1}} \leq C(u,v), ~ &\forall ~(u,v) \in E,\\\ \nonumber
&\forall i \in \{1, \ldots ,k\}, \forall j \in \{1, \ldots, l_i\}, \forall ~p\in \mathcal{P}_{i,j}, &\\
&r_p \geq 0.\quad \quad \quad \quad \quad \quad \quad \quad&
\end{alignat}
\vspace{0.1in}
\end{mdframed}
\caption{Path-Based Formulation on the Modified Network}
\label{path_form}
\end{table}
\subsubsection{Path-based formulation to edge-based formulation}
\label{pe_prep_swap}
In this section we construct a solution of the edge-based formulation from the solution of the path-based formulation.
In the edge-based formulation, we construct a new demand set $D_{\mathrm{mod}}$, where the $i$-th demand $(s_i,e_i,l_i)$ in the original demand set $D$ is decomposed into $l_i$-demands $(s^0_i,e^1_i), \ldots , (s^0_i,e^{l_i}_i)$. Recall that, the quantity $l_i$ denotes the upper bound on the discovered path-length which reflects the lower bound on the required end-to-end fidelity of the EPR-pairs generated between $s_i$ and $e_i$ (See section \ref{fid_to_len} for the details). Here, each of the $(s^0_i,e^j_i)$ are the nodes in the modified graph $G'$. From the construction of $G'$, it is clear that all the paths from $s^0_i$ to $e^j_i$ have length $j$. If we have the solutions of the path formulation, proposed in table \ref{path_form}, then from there, for each edge $(u',v') \in E'$ and for the $i,j$-th demand $(s^0_i,e^j_i)$ we define the value of $\tilde{g}_{ij}(u,v)$ as,
\begin{equation}
\label{eq_gij_prep_swap}
\tilde{g}_{ij}(u',v') := \sum_{\substack{p \in P_{i,j}, \\ (u',v') \in p}} \frac{r_p}{(q)^{j-1}}.
\end{equation}
One can easily check that the definition of $\tilde{g}_{ij}$, defined in equation \ref{eq_gij_prep_swap} satisfies all the constraints of the edge-based formulation, proposed in the equations \ref{const_edge31}, \ref{const_edge32}, \ref{const_edge33}. Moreover, with this definition of the $\tilde{g}_{ij}$, the objective function (equation \ref{edge_form_plen3}) of the edge-based formulation becomes same as the objective function of the path-based formulation. This shows that, the optimal value of the edge-based formulation is at least as good as the solution of the path-based formulation.
\subsubsection{Edge-based formulation to path-based formulation}
\label{etp_prep_swap}
Here, we construct a solution of the path-based formulation from the solution of the edge-based formulation. We use the algorithm, proposed in algorithm \ref{flow_dec3} for extracting the paths and corresponding rate along that path. In the algorithm we compute the rate $\tilde{r}_{p_{j,m}}$ corresponding to a path $p_{j,m}$ for a demand $(s^0_i,e^j_i) \in D_{\mathrm{mod}}$ as follows,
\begin{equation}
\label{prep_swap_rate}
\tilde{r}_{p_{j,m}} := (q)^{j-1}\text{Min}_{(u',v')\in p_{j,m}}\{F_{i,j,m}(u',v')\},
\end{equation}
where the function $F_{i,j,m}(u',v')$ is related to $g_{i,j}(u',v')$ and $\tilde{r}_{p_{j,m}}$. Here, $F_{i,j,0} = g_{i,j}$ and for all $m \geq 0$ we compute $F_{i,j,m+1}$ as follows,
\begin{equation}
\label{prep_swap_update}
\forall (u,v) \in p_{j,m}~~F_{i,j,m+1}(u',v')= F_{i,j,m} (u',v')- \frac{\tilde{r}_{p_{j,m}}}{(q)^{j-1}}.
\end{equation}
We give a detailed proof of the fact that the paths as well as the allocated rate corresponding to each path, extracted from algorithm \ref{flow_dec3} corresponds to the feasible solution of the path-based formulation in the appendix \ref{app_etp_prep_swap}. Moreover, if we consider the equation \ref{prep_swap_rate} as the definition $\tilde{r}_{p_{j,m}}$ then the objective function of the edge-based formulation is same as the objective function of the edge-based formulation. This shows that this is a valid solution of the path-based formulation. In the last section we showed that the solution of the path-based formulation is at least as good as the solution of the edge-based formulation. Hence, the solutions of both of the formulations are equivalent.
\subsection{Prepare and swap protocol and the LP-formulations}
In this section, we explain the prepare and swap protocol and show that with this protocol one can achieve the entanglement distribution rate along a path, proposed by algorithm \ref{flow_dec3}. In the next section we explain the protocol for a repeater chain with a single demand. After that we extend the protocol for the case for multiple demands.
\subsubsection{Prepare and swap protocol for a repeater chain}
Suppose in a repeater chain $u_0=s, u_1, \ldots, u_n, u_{n+1} = e$, where for all $0 \leq i \leq n$, the nodes $u_{i},u_{i+1}$ are neighbours of each other and $s$ would like to share EPR-pairs with $e$. In this protocol, first, all the repeaters generate entanglement with its neighbours/neighbour in parallel and store the entangled links in the memory. Here we assume that entanglement generation across an elementary link is a deterministic event, i.e., the entanglement generation probability per each attempt is one. An intermediate node $u_i$ which resides between $u_{i-1}$ and $u_i$, performs the swap operation when both of the EPR-pairs between $u_{i-1},u_i$ and $u_i, u_{i+1}$ are ready. As we assume that each of the swap operations is probabilistic, so the entanglement generation rate with this protocol is lower. However, due to the independent swap operations, the protocol doesn't need a long storage time. This makes the protocol more practical.
We give an example of such an entanglement generation protocol on a repeater chain with three intermediate nodes and one source-destination pair in figure \ref{exam0}. In the next lemma, we derive an analytical expression of the end-to-end entanglement generation rate in a repeater chain network for the prepare and swap protocol. Note that, a variant of the proof of lemma \ref{egr_prep_swap} can be found in the literature \cite{SSMS14}. For completeness, in the supplementary material we include the proof of this lemma.
\begin{lemma}
\label{egr_prep_swap}
In a repeater chain network with $n+1$ repeaters $\{u_0, u_1, \ldots, u_{n}\}$, if the probability of generating an elementary pair per attempt is one, the probability of a successful BSM is $(q)$, the capacity of an elementary link $(u_i,u_{i+1})$ (for $0 \leq i \leq n-1$) is denoted by $C_i$ and if the repeaters follow the prepare and swap protocol for generating EPR-pairs, then the expected end-to-end entanglement generation rate $r_{u_0,u_{n}}$ is,
\begin{align}
\label{eg_rate_prep_mes}
r_{u_0,u_{n}} & = (q)^{n-1}\text{Min}\{C_0, \ldots ,C_{n-1}\}.
\end{align}
\end{lemma}
Notice that the EPR-pair generation rate for this protocol is exactly same as the EPR-pair generation rate, proposed in equation \ref{eq:rate}, which we use for the path-based formulation. In the previous sections we show that both of the path-based and the edge-based formulations are equivalent. This implies, the rates extracted from algorithm \ref{flow_dec3} can be achieved with this prepare and swap protocol.
\subsubsection{Prepare and swap protocol for an arbitrary network}
In a quantum network if there are multiple demands then one link might be shared between multiple paths. From algorithm \ref{flow_dec3} we get the set of paths and the desired EPR-pair generation rate $r_p$, across each of the paths $p$, passing through an edge $(u,v)$. In this scenario we use the prepare and swap protocol for each of the paths in a sequential manner or round-robin manner\footnote{One can use a more sophisticated scheduling algorithm for allocating the EPR-pair generation resources across an elementary link. The details of this scheduling algorithm are beyond the scope of this paper.}. From lemma \ref{egr_prep_swap} we get that, to achieve the rate $r_p$, we need to generate on average $\frac{r_p}{(q)^{|p|-1}}$ elementary EPR-pairs per second across each of the elementary links $(u,v)$ along the path $p$. Here we also assume that the elementary link generation is a deterministic event, i.e., the two nodes $u$ and $v$ connecting the elementary link $(u,v)$ can generate exactly $C(u,v)$ EPR-pairs per second. Hence, each of the paths uses the elementary link $(u,v)$ for $\frac{r_p}{(q)^{|p|-1}C(u,v)}$ seconds (on average) for generating the required elementary EPR-pairs\footnote{Note that, if $\frac{r_p}{(q)^{|p|-1}}$ is a rational number then one can always achieve this average rate. For an irrational value of $\frac{r_p}{(q)^{|p|-1}}$ we need to approximate it to a rational number.}. For generating $r_p$ EPR-pairs per second, all the nodes along the path $p$ need to use the elementary links for at most $\max_{(u,v)\in p}\left\{\frac{r_p}{(q)^{|p|-1}C(u,v)}\right\}$ seconds, which is equal to $\frac{r_p}{(q)^{|p|-1}}\frac{1}{\text{Min}_{(u,v)\in p}\{C(u,v)\}}$ seconds. This gives an upper bound on the average storage time of the quantum memory for generating the EPR-pairs along the path $p$. In figure \ref{exam1} we give an example of allocating EPR-pair generation resources for multiple demands.
\begin{figure}
\includegraphics[width=0.7\textwidth, left]{examp_1_mod.pdf}
\caption{Repeater network $G = (V,E,C)$ with the demand set $D = \{(s_1,e),(s_2,e)\}$. All the repeaters use BSM for the entanglement swap. Here we assume that the success probability of the BSM is $q = \frac{1}{2}$. From the algorithm \ref{flow_dec3} we get a path $p_1 = ((s_1,u),(u,v),(v,w),(w,e))$ for the demand $(s_1,e)$ and a path $p_2 = ((s_2,w),(w,e))$ for the demand $(s_2,e)$. The desired rate across $p_1$ is $r_{p_1} = 0.25$ EPR-pairs per second and the desired rate across $p_2$ is $r_{p_2} = 5$ EPR-pairs per second.
If we use the prepare and swap protocol along both of the paths separately then we get the desired EPR-pairs generation rate. However, here the elementary link $(w,e)$ is being shared by both of the paths. In this case, a simple scheduling technique would be to distribute the EPR-pairs sequentially. For example, at the beginning the elementary link $(w,e)$ generates the EPR-pairs for the demand $(s_1,e)$. For generating on average $0.25$ EPR-pairs across the path $p_1$, the elementary link $(w,e)$ has to generate $0.25 \times 2^{|p_1|-1} = 2$ EPR-pairs per second. As the capacity of the elementary link is $20$ EPR-pairs per second. Hence, it can generate $2$ EPR-pairs within $\frac{2}{20} = 0.1$ seconds (on average). Then it can start generating the EPR-pairs for the demand $(s_2,e)$. For this demand, the elementary link $(w,e)$ has to generate $5 \times 2^{|p_2|-1} = 10$ EPR-pairs per second. Hence, it will take on average $\frac{10}{20} = 0.5$ seconds to generate $10$ EPR-pairs.}
\label{exam1}
\end{figure}
\subsection{Complexity analysis}
\label{complex_seq_swap_wo_fid}
In this section, we analyse the complexity of the LP formulations as well as the path-extraction and rate allocation algorithm (algorithm \ref{flow_dec3}). The edge-based LP-formulation, proposed in this paper is based on the modified network graph $G'$ and modified demand set $D_{\mathrm{mod}}$ and the running time of the edge-based LP-formulation solver depends on the size of this network and modified demand set. In the next lemma we give an upper bound on the size of $G'$ and $D_{\mathrm{mod}}$.
\begin{lemma}
\label{lem:graph_size}
The edge-based formulation, proposed in equations \ref{edge_form_plen3} to \ref{const_edge33}, has at most $N = |D||E||V|$ variables and $M = |V|^2|E||D| + |V||E| + |V|^2|D|$ constraints, where $|V|, |E|, |D|$ denote the total number of repeater nodes, total number of edges in the network graph $G$ and size of the demand set $D$ respectively.
\end{lemma}
\begin{proof}
The edge-based formulation, is based on the modified network $G' = (V',E',C')$ which we construct from the actual internet network $G = (V,E,C)$. In $G'$ we create at most $l_{\text{max}}$ copies of each of the nodes and edges. This implies, $|V'| \leq l_{\mathrm{max}} |V|$ and $|E'| \leq l_{\mathrm{max}}|E|$. As, $l_{\mathrm{max}} = \max \{l_1, \ldots, l_{|D|}\}$, hence $l_{\max} \leq |V|$. This implies, $|V'| \leq |V|^2$ and $|E'| \leq |E||V|$. In the construction of the edge-based formulation, for each demand, we introduce one variable corresponding to each edge of $E'$. Hence, the total number of variables for this formulation is $N = |D||E'| = |D||E||V|$.
In the edge-based formulation, the constraint equation \ref{const_edge31} ($g_{ij}(u,v) \geq 0$) holds for all $1\leq i \leq |D|$, for all $1 \leq j \leq l_i$ and for all edge $(u,v) \in E'$. This implies, the total number of constraints corresponding to equation \ref{const_edge31} is $|V|^2|E||D|$. Similarly, the constraint equation \ref{const_edge32} holds for all the edges $|E'|$. This implies, there are at most $|V||E|$ constraints corresponding to equation \ref{const_edge32}. The constraint equation \ref{const_edge33} should be satisfied by all the nodes in $V'$ and for all $1\leq i \leq |D|$, for all $1 \leq j \leq l_i$. This implies, the total number of constraints corresponding to that equation is $|V|^3|D|$. Hence, the total number of constraints in the edge-based formulation is $M = |V|^2|E||D| + |V||E| + |V|^3|D|$.
\end{proof}
The previous lemma implies that the total number of variables and constraints in the LP-formulation corresponding to the maximum flow problem scales polynomially with the number of nodes in the network graph $G$. This implies, the time complexity of the LP solvers for this problem also scales polynomially with the number of nodes in the network graph $G = (V,E,C)$. Hence, one can compute the maximum achievable rate for the quantum internet in polynomial time. Now we focus on the complexity of algorithm \ref{flow_dec3}. The algorithm \ref{flow_dec3} uses the solution of the edge-based formulation for extracting the set of paths for each of the demands. In the next proposition, we show that the size of the set of extracted paths corresponding to each demand is upper bounded by $|V||E|$. Then we use this result for computing the running time of algorithm \ref{flow_dec3}.
\begin{proposition}
\label{prop_prep_cond2}
In algorithm \ref{flow_dec3},
\begin{equation}
|\mathcal{P}_{i,j}| \leq |E||V|.
\end{equation}
\end{proposition}
\begin{proof}
Due to the flow conservation property of the edge-based formulation, if for some neighbour of $s^0_i$, $F_{i,j,m}(s^0_i,v^1) > 0$ then there exist a path $p_{j,m}$ from $s^0_i$ to $e^j_i$ such that $F_{i,j,m}(u',v') >0$ for all $(u',v') \in p_{j,m}$. Note that, at each step $m$ of the algorithm \ref{flow_dec3} there exist at least one edge $(u',v') \in E'$ in the discovered path $p_{j,m}$, such that $F_{i,j,m+1}(u',v') =0$. As there are in total, $|E'|$ number of edges and the algorithm runs until $\sum_{v^1:(s^0_i,v^1) \in E'}F_{i,j,m+1}(s^0_i,v^1)=0$, so the maximum value of $m$ could not be larger than $|E'|$. From the construction of the modified network, $G'$ we have $|E'| \leq |V||E|$. This implies, $|\mathcal{P}_{i,j}| \leq |E'|\leq |E||V|$.
\end{proof}
In the next theorem we show that, running time of algorithm \ref{flow_dec3} is $O(|D||V|^4|E|)$.
\begin{theorem}
\label{thm_comp_rate_fid_prep}
The algorithm \ref{flow_dec3} takes the solution of the edge-based LP-formulation and extract the set of paths in $O(|D||V|^4|E|)$ time, where $|D|$ is the size of the demand set, $|V|, |E|$ denote the total number of nodes and edges in the network $G = (V,E,C)$.
\end{theorem}
\begin{proof}
In algorithm \ref{flow_dec3} we compute the paths based on the modified network $G' = (V',E',C')$, which we construct from the original network $G = (V,E,C)$. In this modified network $|V'| \leq |V|l_{\max} \leq |V|^2$.
In algorithm \ref{flow_dec3} at step $11$ we compute a path in the graph $G'$. Note that, in the worst case, it takes $O(|V'|) = O(|V|^2)$ time to find a path between a source-destination pair in a network. According to proposition \ref{prop_prep_cond2}, we have that for a fixed $1\leq i \leq |D|$ and a fixed $1 \leq j \leq l_i$ the total number of paths discovered by algorithm \ref{flow_dec3} is upper bounded by $O(|V||E|)$. Hence, for that $i,j$, the running time of algorithm \ref{flow_dec3} is $O(|V|^3|E|)$. As, $i \leq |D|$ and $j \leq l_i \leq l_{\max} \leq |V|$, so in the worst case scenario, the total running time of algorithm \ref{flow_dec3} is upper bounded by $O(|D||V|^4|E|)$. This concludes the proof.
\end{proof}
\section{Conclusion}
\label{concl}
In this paper, we use techniques from the length constrained multi-commodity flow theory for developing a polynomial-time algorithm for maximising achievable expected entanglement generation rate between multiple source-destination pairs in a quantum internet. Here, we have maximised the end-to-end entanglement distribution rate, satisfying a constraint on the end-to-end fidelity for each of the demand. We have shown that our LP-formulation provides a maximal solution if it exists. Our path extraction algorithm produces a set of paths and the achievable rates along each of the paths. The path-extraction algorithm has high running time as a function of the path-length. Here we consider the worst case scenarios, where we assume that the length of the discovered path scales with $|V|$. In practical scenarios, without distillation, the end-to-end fidelity of the distributed EPR-pairs would drop drastically with the path-length. Hence, it is fair to consider that the length of the allowed path increases slowly with the size of the network node set. This would make the path-extraction algorithm faster.
One can use any entanglement generation protocol for distributing EPR-pairs across the paths that are discovered by the path-extraction algorithm. However, our LP-formulation is inspired from the atomic ensemble and linear optics based quantum repeaters, where the storage time is very short and the entanglement swap operation is probabilistic in nature \cite{SSDG11, GMLC13, SSMS14}. Here, we have also pointed out that, there exists a practical protocol, called prepare and swap protocol, which can be implemented using atomic ensemble based repeaters and if one uses this protocol for distributing entanglement across each of the paths, then one can generate EPR-pairs with the rate proposed by our path-extraction algorithm.
In this paper, we focus on maximising the end-to-end entanglement generation rate. However, one can easily extend our results for other objective functions, like minimising the weighted sum of congestion at edges.
In future work, it would be interesting to include the more realistic parameters like the bounded storage capacity, time to perform the swap operation, etc., in our model and modify our current formulations to come up with more sophisticated routing algorithms.
The proposed LP-formulations give an optimal achievable EPR-pairs distribution rate with respect to prepare and swap protocol. This protocol is practical and require very less amount of quantum storage time. However, there exist more sophisticated protocol which can achieve higher EPR-pairs distribution rate but require higher quantum storage time. Another interesting future research direction would be to find out a protocol for distributing EPR-pairs along a chain which achieves the optimal EPR-pair generation rate and find an LP-formulation for such protocol.
\begin{acknowledgements}
We would like to acknowledge W. Kozlowski for many stimulating discussions. We would like to thank M. Skrzypczyk for giving useful feedback on the draft. This publication is supported by an ERC Starting grant and the QIA-project that has received funding from the European Union's Horizon 2020 research and innovation program under grant Agreement No. 820445.
\end{acknowledgements}
\bibliographystyle{apsrev4-1}
| 4819f0f6becad26712b2cd1d0deabd7ee705dd22 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}~\par
Clifford analysis is introduced as a generalization of complex function theory to the higher dimensional cases. Many researchers have successfully generalized the theory of one dimensional complex analysis to the higher dimension cases via Clifford analysis. For instance, the function theory of quaternion analysis is applied to study some boundary value problems in three or four dimensional spaces in \cite{Guer}. Many important properties and problems, such as integral formulas, series expansion, integral transforms and boundary value problems, have been generalized to higher dimensions as well in \cite{Guer,Guer1,Guer2}.
\par
The higher spin theory in Clifford analysis is the theory on functions taking values in irreducible representations of the spin group. These representation spaces are usually realized as spaces of homogeneous harmonic or monogenic (null solutions of the Dirac operator) polynomials. The study on this topic can be traced back to the work of Stein and Weiss given in \cite{SW}. Stein and Weiss introduced a technique to construct first order conformally invariant differential operators, named as Stein-Weiss gradients, with a certain type of projections. Bure\v{s} et. al. \cite{Bures} investigated a class of generalized Rarita-Schwinger operators acting on functions taking values in irreducible representations of the spin group with weight $k+1/2$ via Clifford analysis in 2002. These Rarita-Schwinger type operators were also studied by Dunkl et. al. \cite{Dunkl} with an analytic approach. In these two papers, one can notice that Rarita-Schwinger operators and the Dirac operator have very similar properties, such as Cauchy's Theorem, Cauchy's integral formula, Stokes' Theorem, etc. Therefore, Rarita-Schwinger operators are also considered as generalizations of the Dirac operator in the higher spin theory. In 2016, Eelbode et. al. \cite{Eelbode} and De Bie et. al. \cite{DeBie} introduced generalizations of the Laplace operator with respect to the conformal invariance property in the higher spin theory. These differential operators are named as bosonic Laplacians ( the higher spin Laplace operators) or the generalized Maxwell operators, which are special cases of bosonic Laplacians. It is reasonable to expect bosonic Laplacians also have similar properties as the Laplace operator has. In \cite{DR,DWR}, the authors discovered intertwining operators, a Borel-Pompeiu formula and a Green type integral formula for bosonic Laplacians. Recently, the authors \cite{DTR} also studied Dirichlet problems for bosonic Laplacians in the upper-half space and the unit ball. Further, many important properties, such as the mean-value property, Cauchy's estimates and Liouville's Theorem, for null solutions to bosonic Laplacians have been found in \cite{DTR}. Here, we continue our investigation on properties of bosonic Laplacians in Euclidean space.
\par
\textbf{Main results:} In this paper, we firstly look into the generalized Maxwell operators by generalizing classical Maxwell equations to the higher spin spaces in Section $2$. This provides us the motivation for studying the generalized Maxwell operators and bosonic Laplacians. Some preliminaries of Clifford analysis setting, Rarita-Schwinger operators and bosonic Laplacians will also be introduced here as well. In Section $3$, we use some properties of Rarita-Schwinger operators and their connections to bosonic Laplacians to solve Poisson's equation in the higher spin spaces. Section $4$ will be devoted to introducing Green's formulas in the higher spin spaces, which reveal that bosonic Laplacians are self-adjoint with respect to a given $L^2$ inner product. As in the harmonic anlysis, these formulas can possibly be applied in our future work on constructing Green's functions for solving certain boundary value problems.
\subsection*{Acknowledgements}
This paper is dedicated to Klaus G\"urlebeck on his 65th birthday. Chao Ding is supported by Czech Science Foundation, project GJ19-14413Y.
\section{Preliminaries}
\subsection{Notations}
Suppose that $\{e_1,\cdots,e_m\}$ is a standard orthonormal basis for the $m$-dimensional Euclidean space $\mathbb{R}^m$. The (real) Clifford algebra $\mathcal{C}l_m$ is generated by $\mathbb{R}^m$ with the relationship $e_ie_j+e_je_i=-2\delta_{i,j},\ 1\leq i,j\leq m.$ This implies that an element of the basis of the Clifford algebra can be written as $e_A=e_{j_1}\cdots e_{j_r},$ where $A=\{j_1, \cdots, j_r\}\subset \{1, 2, \cdots, m\}$ and $1\leq j_1< j_2 < \cdots < j_r \leq m.$ Hence any element $a\in \mathcal{C}l_m$ can be represented by $a=\sum_Aa_Ae_A$, where $a_A\in \mathbb{R}$. In particular, $\boldsymbol{e}_{\emptyset}=1$ and we call $a_{\emptyset}=Scr(a)$ the scalar part of $a$. The $m$-dimensional Euclidean space $\mathbb{R}^m$ is embedded into $\mathcal{C}l_m$ as follows.
\begin{eqnarray*}
\mathbb{R}^m&&\longrightarrow\quad \mathcal{C}l_m,\\
(x_1,\cdots,x_m)&&\ \mapsto\quad \sum_{j=1}^mx_je_j.
\end{eqnarray*}
Hence, for $\boldsymbol{x}\in\mathbb{R}^m$, one can easily see that $\|\boldsymbol{x}\|^2=\sum_{j=1}^mx_j^2=-\boldsymbol{x}^2$.
For $a=\sum_Aa_Ae_A\in\mathcal{C}l_m$, we define the reversion of $a$ as
\begin{eqnarray*}
\widetilde{a}=\sum_{A}(-1)^{|A|(|A|-1)/2}a_Ae_A,
\end{eqnarray*}
where $|A|$ is the cardinality of $A$. In particular, $\widetilde{e_{j_1}\cdots e_{j_r}}=e_{j_r}\cdots e_{j_1}$. Also $\widetilde{ab}=\widetilde{b}\widetilde{a}$ for $a, b\in\mathcal{C}l_m$.
\par
Now suppose $\boldsymbol{a}\in \mathbb{S}^{m-1}\subseteq \mathbb{R}^m$ and $\boldsymbol{x}\in\mathbb{R}^m$. If we consider $\boldsymbol{a}\boldsymbol{x}\boldsymbol{a}$, we may decompose
$$\boldsymbol{x}=\boldsymbol{x}_{\boldsymbol{a}\parallel}+\boldsymbol{x}_{\boldsymbol{a}\perp},$$
where $\boldsymbol{x}_{\boldsymbol{a}\parallel}$ is the projection of $\boldsymbol{x}$ onto $\boldsymbol{a}$ and $\boldsymbol{x}_{\boldsymbol{a}\perp}$ is the remainder part of $\boldsymbol{x}$ perpendicular to $\boldsymbol{a}$. Hence $\boldsymbol{x}_{\boldsymbol{a}\parallel}$ is a scalar multiple of $a$ and we have
$$\boldsymbol{a}\boldsymbol{x}\boldsymbol{a}=\boldsymbol{a}\boldsymbol{x}_{\boldsymbol{a}\parallel}\boldsymbol{a}+\boldsymbol{a}\boldsymbol{x}_{\boldsymbol{a}\perp}\boldsymbol{a}=-\boldsymbol{x}_{\boldsymbol{a}\parallel}+\boldsymbol{x}_{\boldsymbol{a}\perp}.$$
So the action $\boldsymbol{a}\boldsymbol{x}\boldsymbol{a}$ describes a reflection of $x$ in the direction of $\boldsymbol{a}$. More details can be found in, for instance, \cite{Del}.
\par
The classical Dirac operator is defined as $D_{\boldsymbol{x}}=\sum_{j=1}^m\partial_{x_j}e_j$, which factorizes the Laplace operator $\Delta_{\boldsymbol{x}}=-D_{\boldsymbol{x}}^2$. A $\mathcal{C}l_m$-valued function $f(\boldsymbol{x})$ defined on a domain $\Omega$ in $\mathbb{R}^m$ is called \emph{left monogenic} if it satisfies $D_{\boldsymbol{x}}f(\boldsymbol{x})=0$ in $\Omega$. Since multiplication of Clifford numbers is not commutative in general, there is a similar definition for right monogenic functions.
\subsection{Rarita-Schwinger type operators}
Let $\mathcal{H}_k(\mathcal{C}l_m)\ (\mathcal{M}_k(\mathcal{C}l_m))$ stand for the space of Clifford-valued harmonic (monogenic) polynomials homogeneous of degree $k$. Notice that if $h_k(\boldsymbol{u})\in\mathcal{H}_k(\mathcal{C}l_m)$, then $D_{\boldsymbol{u}}h_k(\boldsymbol{u})\in\mathcal{M}_{k-1}(\mathcal{C}l_m)$, but $D_{\boldsymbol{u}}\boldsymbol{u} p_{k-1}(\boldsymbol{u})=(-m-2k+2)p_{k-1}(\boldsymbol{u}),$ where $p_{k-1}(\boldsymbol{u})\in \mathcal{M}_{k-1}(\mathcal{C}l_m)$. Hence, we have
\begin{eqnarray}\label{Almansi}
\mathcal{H}_k(\mathcal{C}l_m)=\mathcal{M}_k(\mathcal{C}l_m)\oplus \boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m),\ h_k=p_k+\boldsymbol{u} p_{k-1}.
\end{eqnarray}
This is called an \emph{Almansi-Fischer decomposition} of $\mathcal{H}_k(\mathcal{C}l_m)$ \cite{Dunkl}. In this decomposition, we have $P_k^+$ and $P_k^-$ as the projection maps
\begin{eqnarray*}
&&P_k^+=1+\frac{\boldsymbol{u} D_{\boldsymbol{u}}}{m+2k-2}:\ \mathcal{H}_k(\mathcal{C}l_m)\longrightarrow \mathcal{M}_k(\mathcal{C}l_m), \label{Pk+}\\
&&P_k^-=I-P_k^+=\frac{-\boldsymbol{u} D_{\boldsymbol{u}}}{m+2k-2}:\ \mathcal{H}_k(\mathcal{C}l_m)\longrightarrow \boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m) \label{Pk-}.
\end{eqnarray*}
Suppose $\Omega$ is a domain in $\mathbb{R}^m$. Consider a differentiable function $f: \Omega\times \mathbb{R}^m\longrightarrow \mathcal{C}l_m$
such that, for each $\boldsymbol{x}\in \Omega$, $f(\boldsymbol{x},\boldsymbol{u})$ is a left monogenic polynomial homogeneous of degree $k$ in $\boldsymbol{u}$. Then \textbf{the Rarita-Schwinger operator} \cite{Bures,Dunkl} is defined by
$$R_k=P_k^+D_{\boldsymbol{x}}:\ C^{\infty}(\mathbb{R}^m,\mathcal{M}_k(\mathcal{C}l_m))\longrightarrow C^{\infty}(\mathbb{R}^m,\mathcal{M}_k(\mathcal{C}l_m)).$$
We also need the following three Rarita-Schwinger type operators.
\begin{eqnarray*}
&&\text{\textbf{The twistor operator:}}\\
&& T_k=P_k^+D_{\boldsymbol{x}}:\ C^{\infty}(\mathbb{R}^m,\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m))\longrightarrow C^{\infty}(\mathbb{R}^m,\mathcal{M}_k(\mathcal{C}l_m)),\\
&&\text{\textbf{The dual twistor operator:}}\\
&&T_k^*=P_k^-D_{\boldsymbol{x}}:\ C^{\infty}(\mathbb{R}^m,\mathcal{M}_k(\mathcal{C}l_m))\longrightarrow C^{\infty}(\mathbb{R}^m,\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m)),\\
&&\text{\textbf{The remaining operator:}}\\
&& Q_k=P_k^-D_{\boldsymbol{x}}:\ C^{\infty}(\mathbb{R}^m,\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m))\longrightarrow C^{\infty}(\mathbb{R}^m,\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m)).
\end{eqnarray*}
More details can be found in \cite{Bures,Dunkl}.
\subsection{The generalized Maxwell operators}
In \cite{Eelbode}, the authors constructed the generalized Maxwell operators as a type of second order conformally invariant differential operators on particular function spaces. It was also pointed out that these differential operators reduced to the classical source-free Maxwell equations in the Minkowski space. Here, we will show the details that how to obtain the generalized Maxwell operators from the classical source-free Maxwell equations. This gives us the motivation to study these particular type of second order differential operators.
\par
Recall that the classical source-free coupled Maxwell equations are given by
\begin{eqnarray*}
&&\nabla\cdot\boldsymbol{E}=0,\ \nabla\times\boldsymbol{E}=-\frac{\partial{\boldsymbol{B}}}{\partial t},\\
&&
\nabla\cdot\boldsymbol{B}=0,\ \nabla\times\boldsymbol{B}=\mu_0\epsilon_0\frac{\partial\boldsymbol{E}}{\partial t},
\end{eqnarray*}
Where $\boldsymbol{E}$ stands for the electric field, $\boldsymbol{B}$ stands for the magnetic field, $\boldsymbol{B}$ and $\boldsymbol{E}$ are both vector fields in $\mathbb{R}^3$. $\mu_0$ is the permeability of free space and $\epsilon_0$ is the permittivity of free space.
\par
Since $\nabla\cdot\boldsymbol{B}=0$, we can define $\boldsymbol{B}$ in terms of a vector potential $\boldsymbol{C}$ as $\boldsymbol{B}=\nabla\times\boldsymbol{C}$. From Maxwell Faraday's equation, one obtains that $\nabla\times(\boldsymbol{E}+\frac{\partial\boldsymbol{C}}{\partial t})=0$. This means that $\boldsymbol{E}+\frac{\partial\boldsymbol{C}}{\partial t}$ can be written as the gradient of some scalar function, namely, a scalar potential $\Phi$. Hence, one has
\begin{eqnarray}\label{Electric}
\boldsymbol{E}=-\nabla\Phi-\frac{\partial\boldsymbol{C}}{\partial t}
\end{eqnarray}
with a scalar potential $\Phi$.
\par
Plugging $\boldsymbol{B}=\nabla\times\boldsymbol{C}$ and $\boldsymbol{E}=-\nabla\Phi-\frac{\partial\boldsymbol{C}}{\partial t}$ into $\nabla\cdot\boldsymbol{E}=0,\ \nabla\times\boldsymbol{B}=\mu_0\epsilon_0\frac{\partial\boldsymbol{E}}{\partial t}$, one has
\begin{eqnarray}\label{newmax}
&&\nabla^2\Phi+\frac{\partial}{\partial t}(\nabla\cdot\boldsymbol{C})=0,\nonumber\\
&&\nabla^2\boldsymbol{C}-\mu_0\epsilon_0\frac{\partial^2\boldsymbol{C}}{\partial^2 t}-\nabla(\nabla\cdot\boldsymbol{C}+\mu_0\epsilon_0\frac{\partial\Phi}{\partial t})=0.
\end{eqnarray}
Next, we choose a set of potentials $(\boldsymbol{C},\Phi)$ to satisfy the Lorenz condition $\nabla\cdot \boldsymbol{C}+\mu_0\epsilon_0\frac{\partial\Phi}{\partial t}=0$, which uncouples the pair of equations given in (\ref{newmax}). Now one has
\begin{eqnarray*}
\nabla^2\Phi-\mu_0\epsilon_0\frac{\partial^2\Phi}{\partial t^2}=0,\ \nabla^2\boldsymbol{C}-\mu_0\epsilon_0\frac{\partial^2\boldsymbol{C}}{\partial t^2}=0.
\end{eqnarray*}
The potentials $\boldsymbol{C}$ and $\Phi$ form a 4-vector potential $\boldsymbol{C}^{\alpha}=(\Phi,\boldsymbol{C})$. Then the two equations above and the Lorenz condition becomes
\begin{eqnarray*}
\square C^{\alpha}=0,\ \partial_{\alpha}A^{\alpha}=0,
\end{eqnarray*}
where $\square=\nabla^2-\mu_0\epsilon_0\frac{\partial^2}{\partial t^2}$ is the d'Alembert operator. For $\boldsymbol{E}=-\nabla\Phi-\frac{\partial\boldsymbol{C}}{\partial t}$ and $\boldsymbol{B}=\nabla\times\boldsymbol{C}$, if the space coordinate is given by $(x,y,z)$, the $x$ components of $\boldsymbol{E}$ and $\boldsymbol{B}$ can be written explicitly as
\begin{eqnarray*}
&&E_x=-\mu_o\epsilon_0\frac{\partial C_x}{\partial t}-\frac{\partial \Phi}{\partial x}=-(\partial^0C^1-\partial^1C^0),\\
&&B_x=\frac{\partial C_z}{\partial y}-\frac{\partial C_y}{\partial z}=-(\partial^2C^3-\partial^3C^2),
\end{eqnarray*}
where $\partial^{\alpha}=(\frac{\partial}{\partial t},-\nabla)$. These equations imply that the electric and magnetic fields are the elements of a second rank, antisymmetric field-strength tensor $F^{\alpha\beta}=\partial^{\alpha}C^{\beta}-\partial^{\beta}C^{\alpha}$.
\par
This leads to a different formulation of Maxwell equations in terms of a differential 2-form $F^{\alpha\beta}$, known as the Maxwell-Faraday tensor, $\partial_{\alpha}F^{\alpha\beta}=0$, which can also be rewritten as
\begin{eqnarray*}
\square C^{\beta}-\partial^{\beta}\partial_{\alpha}C^{\alpha}=0,
\end{eqnarray*}
where $C^{\alpha}=(\Phi,\boldsymbol{C})$, $\partial^{\alpha}=(\partial_t,-\nabla)$ and $\partial_{\alpha}=(\partial_t,\nabla)$ and
\[
F^{\alpha\beta}
=
\begin{pmatrix}
0 & E_x & E_y & E_z \\
-E_x & 0 & -B_z & B_y \\
-E_y & B_z & 0 & -B_x \\
-E_z & -B_y & B_x & 0
\end{pmatrix}.
\]
The equation $\square C^{\beta}-\partial_{\beta}\partial^{\alpha}C^{\alpha}=0$ in the Minkowski space has a generalization to the $m$-dimensional Euclidean space given below.
\begin{eqnarray*}
\Delta_{\boldsymbol{x}}\boldsymbol{f_s}(\boldsymbol{x})-\frac{4}{m}\sum_{j=1}^{m}\partial_{x_s}\partial_{x_j}\boldsymbol{f_j}(\boldsymbol{x})=0,\ 1\leq s\leq m.
\end{eqnarray*}
where $f_s(\boldsymbol{x})$ is a vector valued function, with $\boldsymbol{x}\in\mathbb{R}^m$. The constant $4/m$ allows the generalized operator above to preserve the conformal invariance property of the Maxwell equations. Notice that the space of real-valued homogeneous of harmonic polynomials with degree-$1$ with respect to a variable $\boldsymbol{u}\in\mathbb{R}^m$ , denoted by $\mathcal{H}_1$ , is spanned by $\{u_1,\cdots,u_m\}$. The $m$ equations above can be replaced by one equation
\begin{eqnarray*}
\sum_{s=1}^{m}u_s\bigg(\Delta_{\boldsymbol{x}}\boldsymbol{f_s}(\boldsymbol{x})-\frac{4}{m}\sum_{j=1}^{m}\partial_{x_s}\partial_{x_j}\boldsymbol{f_j}(\boldsymbol{x})\bigg)=0.
\end{eqnarray*}
This equation can also be rewritten as
\begin{eqnarray*}
0&=&\sum_{s=1}^m\Delta_{\boldsymbol{x}}u_sf_s(\boldsymbol{x})-\frac{4}{m}\sum_{j,s=1}^mu_s\partial_{x_s}\partial_{x_j}f_j(\boldsymbol{x})\\
&=&\sum_{s=1}^m\Delta_{\boldsymbol{x}}u_sf_s(\boldsymbol{x})-\frac{4}{m}\sum_{j,s,k=1}^mu_s\partial_{x_s}\partial_{x_j}\partial_{u_j}u_kf_k(\boldsymbol{x}).
\end{eqnarray*}
Now, we consider $f(\boldsymbol{x},\boldsymbol{u}):=\sum_{s=1}^mu_sf_s(\boldsymbol{x})$ as a $\mathcal{H}_1(\mathbb{R})$-valued function on $\mathbb{R}^m$. With the help of the Dirac operator $\boldsymbol{D}_{\boldsymbol{x}}=\sum_{s=1}^m\boldsymbol{e}_s\partial_{x_s}$ in the $m$-dimensional Euclidean space. This equation can be written in a compact form
\begin{eqnarray*}
\left(\Delta_{\boldsymbol{x}}-\frac{4}{m}\langle \boldsymbol{u},\boldsymbol{D}_{\boldsymbol{x}}\rangle\langle \boldsymbol{D}_{\boldsymbol{u}},\boldsymbol{D}_{\boldsymbol{x}}\rangle\right)\boldsymbol{f}(\boldsymbol{x},\boldsymbol{u})=0,
\end{eqnarray*}
where $\langle\ ,\ \rangle$ is the standard inner product in Euclidean space.
The operator $\Delta_{\boldsymbol{x}}-\frac{4}{m}\langle \boldsymbol{u},\boldsymbol{D}_{\boldsymbol{x}}\rangle\langle \boldsymbol{D}_{\boldsymbol{u}},\boldsymbol{D}_{\boldsymbol{x}}\rangle$, denoted by $\mathcal{D}_1$ is called the generalized Maxwell operator, which is a second order conformally invariant differential operator in the higher spin spaces in Clifford analysis. It was firstly constructed by Eelbode et. al. in \cite{Eelbode}.
\par
More generally, if we consider a function $f(\boldsymbol{x},\boldsymbol{u})\in C^{\infty}(\mathbb{R}^m,\mathcal{H}_k(\mathbb{R}))$, i.e., for a fixed $\boldsymbol{x}\in\mathbb{R}^m$, $f(\boldsymbol{x},\boldsymbol{u})\in\mathcal{H}_k(\mathbb{R})$ with respect to $\boldsymbol{u}$. The second order conformally invariant differential operators, named as bosonic Laplacians (also known as the higher spin Laplace operators \cite{Eelbode}), are defined as
\begin{eqnarray}\label{Dtwo}
&&\mathcal{D}_k:\ C^{\infty}(\mathbb{R}^m,\mathcal{H}_k(\mathbb{R}))\longrightarrow C^{\infty}(\mathbb{R}^m,\mathcal{H}_k(\mathbb{R})),\nonumber\\
&&\mathcal{D}_k=\Delta_{\boldsymbol{x}}-\frac{4\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle}{m+2k-2}+\frac{4|u|^2\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle^2}{(m+2k-2)(m+2k-4)},
\end{eqnarray}
where $\langle\ ,\ \rangle$ is the standard inner product in $\mathbb{R}^m$. One can easily see that $\mathcal{D}_k$ reduces to the generalized Maxwell operator when $k=1$.
\section{Poisson's equation and representation formula}
In this section, we will review some properties of the Rarita-Schwinger type operators and bosonic Laplacians from \cite{DeBie,DR,Dunkl,Li}, which will be needed for solving a Poisson's equation in the higher spin spaces.
\par
Let $Z_k^1(\boldsymbol{u},\boldsymbol{v})$ be the reproducing kernel for $\mathcal{M}_k(\mathcal{C}l_m)$, which satisfies
\begin{eqnarray*}
f(\boldsymbol{v})=\int\displaylimits_{\mathbb{S}^{m-1}}Z_k^1(\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{u})dS(\boldsymbol{u}),\ for\ all\ f(\boldsymbol{v})\in\mathcal{M}_k(\mathcal{C}l_m).
\end{eqnarray*}
Then the fundamental solution for $R_k$ (Section $5.1$, \cite{Bures}) is
\begin{eqnarray*}
E_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=\displaystyle\frac{m+2k-2}{(m-2)\omega_{m}}\frac{\boldsymbol{y}-\boldsymbol{x}}{|\boldsymbol{y}-\boldsymbol{x}|^m}Z_k^1\bigg(\frac{(\boldsymbol{y}-\boldsymbol{x})\boldsymbol{u}(\boldsymbol{y}-\boldsymbol{x})}{|\boldsymbol{y}-\boldsymbol{x}|^2},\boldsymbol{v}\bigg).
\end{eqnarray*}
Similarly, we have the fundamental solution for $Q_k$ (Section $3$, \cite{Li}) as follows.
\begin{eqnarray*}
F_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=\displaystyle\frac{m+2k-2}{(2-m)\omega_{m}}\boldsymbol{u}\frac{\boldsymbol{y}-\boldsymbol{x}}{|\boldsymbol{y}-\boldsymbol{x}|^m}Z_{k-1}^1\bigg(\frac{(\boldsymbol{y}-\boldsymbol{x})\boldsymbol{u}(\boldsymbol{y}-\boldsymbol{x})}{|\boldsymbol{y}-\boldsymbol{x}|^2},\boldsymbol{v}\bigg)\boldsymbol{v}.
\end{eqnarray*}
We also have that
\begin{theorem}[Theorem 10, \cite{Dunkl}]
Let $f\in C_c^{\infty}(\mathbb{R}^m,\mathcal{M}_k(\mathcal{C}l_m))$, we have
\begin{eqnarray*}
R_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}E_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}=f(\boldsymbol{y},\boldsymbol{v}),
\end{eqnarray*}
\end{theorem}
and
\begin{theorem}[Theorem 6, \cite{Li}]
Let $f\in C_c^{\infty}(\mathbb{R}^m,\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m))$, we have
\begin{eqnarray*}
Q_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}F_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}=f(\boldsymbol{y},\boldsymbol{v}).
\end{eqnarray*}
\end{theorem}
Recall that (Proposition 1, \cite{DR}) the connection between bosonic Laplacians and the Rarita-Schwinger type operators is given by
\begin{align*}
&\mathcal{D}_k =-R_k^2P_k^++\frac{2R_kT_kP_k^-}{m+2k-4}-\frac{2Q_kT_k^*P_k^+}{m+2k-4}-\frac{(m+2k)Q_k^2P_k^-}{m+2k-4},\\
=&R_k\bigg[-R_kP_k^++\frac{2T_kP_k^-}{m+2k-4}\bigg]+Q_k\bigg[-\frac{2T_k^*P_k^+}{m+2k-4}-\frac{(m+2k)Q_kP_k^-}{m+2k-4}\bigg].
\end{align*}
We let $A_k=-R_kP_k^++\displaystyle\frac{2T_kP_k^-}{m+2k-4}$ and $B_k=-\displaystyle\frac{2T_k^*P_k^+}{m+2k-4}-\displaystyle\frac{(m+2k)Q_kP_k^-}{m+2k-4}$ for convenience, then we have
\begin{eqnarray}\label{connection}
\mathcal{D}_k=R_kA_k+Q_kB_k.
\end{eqnarray}
Now, we can solve a Poisson's equation for bosonic Laplacian $\mathcal{D}_k$ as follows.
\begin{theorem}[Solving Poisson's equation]\label{PoissonEqn}
Let $f\in C^2_c(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$, that is $f\in C^2(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$ and $f$ has compact support with respect to $\boldsymbol{x}$, and set
\begin{eqnarray}
\Phi(\boldsymbol{y},\boldsymbol{v})=\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}.
\end{eqnarray}
Then, we have
\begin{enumerate}
\item $\Phi\in C^2(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$,
\item $\mathcal{D}_k\Phi=f\ \text{in}\ \mathbb{R}^m\times\mathbb{B}^{m}$.
\end{enumerate}
\end{theorem}
\begin{proof}
1.Firstly, we notice that $Z_k\bigg(\displaystyle\frac{\boldsymbol{x}\boldsymbol{u}\boldsymbol{x}}{|\boldsymbol{x}|^2},\boldsymbol{v}\bigg)$ is a $k$-homogeneous harmonic polynomial with respect to $\boldsymbol{v}$, so $\Phi$ is harmonic with respect to $\boldsymbol{v}$. Further, we have
\begin{eqnarray*}
\Phi(\boldsymbol{y},\boldsymbol{v})&=&\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&=&\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},0,\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{y}-\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x},
\end{eqnarray*}
hence,
\begin{eqnarray*}
&&\frac{\Phi(\boldsymbol{y}+h\boldsymbol{e}_j,\boldsymbol{v})-\Phi(\boldsymbol{y},\boldsymbol{v})}{h}\\
&=&\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},0,\boldsymbol{u},\boldsymbol{v})\frac{f(\boldsymbol{y}-\boldsymbol{x}+h\boldsymbol{e}_j,\boldsymbol{u})-f(\boldsymbol{y}-\boldsymbol{x},\boldsymbol{u})}{h}dS(\boldsymbol{u})d\boldsymbol{x},
\end{eqnarray*}
where $h\neq 0$ and $\boldsymbol{e}_j=(0,\cdots,1,\cdots,0)$ with $1$ in the $j^{th}$ spot. Notice that $f\in C^2_c(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$, which tells us that
\begin{eqnarray*}
\frac{f(\boldsymbol{y}-\boldsymbol{x}+h\boldsymbol{e}_j,\boldsymbol{u})-f(\boldsymbol{y}-\boldsymbol{x},\boldsymbol{u})}{h}\rightarrow f_{y_j}(\boldsymbol{y}-\boldsymbol{x},\boldsymbol{u}),\ h\rightarrow 0
\end{eqnarray*}
uniformly. Hence, we have
\begin{eqnarray*}
\Phi_{y_j}(\boldsymbol{y},\boldsymbol{v})=\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},0,\boldsymbol{u},\boldsymbol{v})f_{y_j}(\boldsymbol{y}-\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}.
\end{eqnarray*}
A similar argument can be applied to the second derivatives, which implies $\Phi\in C^2(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$.
\par
2. We prove the second claim by applying the properties of the Rarita-Schwinger type operators and the connection between the Rarita-Schwinger type operators and bosonic Laplacians given in \eqref{connection}.
\par
Recall that $\mathcal{D}_k=R_kA_k+Q_kB_k$, and from Proposition 2 in \cite{DR}, we know that
\begin{eqnarray*}
&&R_kA_kH_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=R_kE_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=\delta(\boldsymbol{y}-\boldsymbol{x})Z_k^1(\boldsymbol{u},\boldsymbol{v}),\\
&&Q_kB_kH_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=R_kE_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=\delta(\boldsymbol{y}-\boldsymbol{x})\boldsymbol{v} Z_{k-1}^1(\boldsymbol{u},\boldsymbol{v})\boldsymbol{u},
\end{eqnarray*}
in the distributional sense. Further, it also tells us that
\begin{align*}
&\mathcal{D}_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
=&(R_kA_k+Q_kB_k)\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})(P_k^++P_k^-)f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
=&R_kA_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})P_k^+f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&+Q_kB_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})P_k^-f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}.
\end{align*}
The last equation comes from the fact that
\begin{eqnarray*}
R_kA_kH_k,P_k^+f\in\mathcal{M}_k(\mathcal{C}l_m);\ Q_kB_kH_k,P_k^-f\in\boldsymbol{u}\mathcal{M}_{k-1}(\mathcal{C}l_m),
\end{eqnarray*}
and two functions from different function spaces above are orthogonal to each other with respect to the integral over the unit sphere with respect to $\boldsymbol{u}$ (Lemma $5$ in \cite{Dunkl}). Further, \cite[Theorem 10]{Dunkl} shows that
\begin{eqnarray*}
&&R_kA_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})P_k^+f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&=&\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}R_kE_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})P_k^+f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}=P^+_kf(\boldsymbol{y},\boldsymbol{v}),
\end{eqnarray*}
and \cite[Theorem 6]{Li} gives us that
\begin{eqnarray*}
Q_kB_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})P_k^-f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}=P^-_kf(\boldsymbol{y},\boldsymbol{v}).
\end{eqnarray*}
Hence, one has
\begin{eqnarray*}
&&\mathcal{D}_k\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&=&P_k^+f(\boldsymbol{y},\boldsymbol{v})+P_k^-f(\boldsymbol{y},\boldsymbol{v})=f(\boldsymbol{y},\boldsymbol{v}),
\end{eqnarray*}
which completes the proof.
\end{proof}
With Theorem \ref{PoissonEqn} and the Liouville-type theorem given in \cite[Theorem 5.7]{DTR}, one can immediately have a representation formula as follows.
\begin{theorem}[Representation formula] Assume $m>4$ and $f\in C^2_c(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$. Then any bounded solution of $\mathcal{D}_k g=f$ in $\mathbb{R}^m\times\mathbb{B}^{m}$ has the form
\begin{eqnarray*}
g(\boldsymbol{y},\boldsymbol{v})=\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}+h(\boldsymbol{v}), \quad \boldsymbol{y}\in\mathbb{R}^m,\, \boldsymbol{v}\in\mathbb{B}^{m},
\end{eqnarray*}
where $h\in\mathcal{H}_k(\mathbb{R})$.
\end{theorem}
\begin{proof}
Notice that $H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})\rightarrow 0$ when $|\boldsymbol{x}-\boldsymbol{y}|\rightarrow \infty$ for $m>4$, hence,
\begin{eqnarray*}
\Phi(\boldsymbol{y},\boldsymbol{v}):=\int\displaylimits_{\mathbb{R}^m}\int\displaylimits_{\mathbb{S}^{m-1}}H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})f(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}
\end{eqnarray*}
is a bounded solution in $C^2(\mathbb{R}^m\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$ for $\mathcal{D}_k g=f$ in $\mathbb{R}^m\times\mathbb{B}^{m}$. If $\Phi'$ is another bounded solution in $\mathbb{R}^m\times\mathbb{B}^{m}$, then we have a bounded solution $\Phi-\Phi'$ for $\mathcal{D}_k g=0$ in $\mathbb{R}^m\times\mathbb{B}^{m}$. In accordance to the Liouville-type theorem, $\Phi-\Phi'=h(\boldsymbol{v})$ in $\mathbb{R}^m\times\mathbb{B}^{m}$, where $h(\boldsymbol{v})\in\mathcal{H}_k(\mathbb{R})$, which completes the proof.
\end{proof}
In particular, if we let $g(\boldsymbol{x},\boldsymbol{u})$ above be the fundamental solution of $\mathcal{D}_k$ given by (Theorem $5.2$, \cite{DeBie})
\begin{eqnarray*}
H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})=c_{m,k}|\boldsymbol{y}-\boldsymbol{x}|^{2-m}Z_k\bigg(\displaystyle\frac{(\boldsymbol{y}-\boldsymbol{x})\boldsymbol{u}(\boldsymbol{y}-\boldsymbol{x})}{|\boldsymbol{y}-\boldsymbol{x}|^2},v\bigg),
\end{eqnarray*}
and discuss the only singular point $\boldsymbol{y}$ with the technique applied in \cite[Theorem 6]{DR}, we can obtain a Green's integral formula as follows.
\begin{proposition}[Green's integral formula]
Let $\Omega\subset\mathbb{R}^m$ be an open domain and $f\in C^2(\Omega\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$. Then, we have
\begin{eqnarray*}
f(\boldsymbol{y},\boldsymbol{v})
&=&\int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}(Af)(\boldsymbol{x},\boldsymbol{u})H_k(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})\\
&&\quad\quad\quad-f(\boldsymbol{x},\boldsymbol{u})(AH_k)(\boldsymbol{x},\boldsymbol{y},\boldsymbol{u},\boldsymbol{v})dS(\boldsymbol{u})d\sigma(\boldsymbol{x}),
\end{eqnarray*}
where $A$ is the operator given in Theorem \ref{Green}.
\end{proposition}
\section{Green's formulas in the higher spin spaces}
It is well-known that Green's formulas are very important in many applications, especially in the study of boundary value problems. In this section, we provide Green's formulas for bosonic Laplacians acting on scalar-valued functions and Clifford-valued functions, respectively.
\subsection*{Green's formula: scalar-valued version}
\hfill \\
Let $\Omega\subset\mathbb{R}^m$ be an open domain, we consider scalar-valued function spaces $C^2(\Omega\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$ and define the following inner product
\begin{eqnarray}\label{innerproduct}
\langle f\ |\ g\rangle=\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x},\quad f,g\in C^2(\Omega\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R})).
\end{eqnarray}
\begin{theorem}[Green's formula: scalar-valued version]\label{Green}
Let $\Omega\subset\mathbb{R}^m$ be an open domain and $f,g\in C^2(\Omega\times\mathbb{B}^{m},\mathcal{H}_k(\mathbb{R}))$. Then, we have
\begin{eqnarray*}
&&\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}(\mathcal{D}_k f(\boldsymbol{x},\boldsymbol{u}))g(\boldsymbol{x},\boldsymbol{u})-f(\boldsymbol{x},\boldsymbol{u})(\mathcal{D}_k g(\boldsymbol{x},\boldsymbol{u}))dS(\boldsymbol{u})d\boldsymbol{x}\\
&=&\int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}(Af)(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})-f(\boldsymbol{x},\boldsymbol{u})(Ag)(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\sigma(\boldsymbol{x}),
\end{eqnarray*}
where $\sigma(x)$ is the area element on $\partial\Omega$ and $n_{\boldsymbol{x}}$ is the outward unit normal vector on $\partial\Omega$ and
\begin{eqnarray*}
A=\frac{\partial}{\partial n_{\boldsymbol{x}}}-\frac{4\langle\boldsymbol{u},n_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle}{m+2k-2}.
\end{eqnarray*}
\end{theorem}
\begin{proof}
The main tools used here are Stokes' Theorem and the orthogonality between homogeneous harmonic polynomials with respect to the $L^2$ integral inner product over the unit sphere. Firstly, let us look at
\begin{align}\label{ezero}
&\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}(\mathcal{D}_k f(\boldsymbol{x},\boldsymbol{u}))g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\bigg[\bigg(\Delta_{\boldsymbol{x}}-\frac{4\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle}{m+2k-2}+\frac{4|u|^2\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle^2}{(m+2k-2)(m+2k-4)}\bigg) f(\boldsymbol{x},\boldsymbol{u})\bigg]\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\bigg[\bigg(\Delta_{\boldsymbol{x}}-\frac{4\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle}{m+2k-2}\bigg) f(\boldsymbol{x},\boldsymbol{u})\bigg]g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}.
\end{align}
the last equation comes from the fact that
\begin{eqnarray*}
&&\int\displaylimits_{\mathbb{S}^{m-1}}|\boldsymbol{u}|^2\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle^2f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})\\
&=&\int\displaylimits_{\mathbb{S}^{m-1}}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle^2f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})=0,
\end{eqnarray*}
where $\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle^2f\in\mathcal{H}_{k-2}$, $g\in\mathcal{H}_k$, hence they are orthogonal to each other with respect to the integral over the unit sphere. On the one hand,
by Green's formula with respect to $\boldsymbol{x}$, we have
\begin{align}\label{eone}
& \int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\Delta_{\boldsymbol{x}}f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}=\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}f(\boldsymbol{x},\boldsymbol{u})\Delta_{\boldsymbol{x}}g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
&+ \int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\frac{\partial f(\boldsymbol{x},\boldsymbol{u})}{\partial n_x}g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\sigma(\boldsymbol{x})
- \int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}} f(\boldsymbol{x},\boldsymbol{u})\frac{\partial g(\boldsymbol{x},\boldsymbol{u})}{\partial n_x}dS(\boldsymbol{u})d\sigma(\boldsymbol{x}).
\end{align}
On the other hand, we have
\begin{eqnarray*}
&& \int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&=&-\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})\langle\boldsymbol{u},D_{\boldsymbol{x}}\rangle g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&&+\int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle\boldsymbol{u},n_x\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\sigma(\boldsymbol{x}).
\end{eqnarray*}
Noticing that $\boldsymbol{u}$ is the unit normal vector on the unit sphere, we apply Stokes' Theorem with respect to $\boldsymbol{u}$ to the first integral above, which is equal to
\begin{eqnarray*}
&&-\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{B}^{m}}\sum_{j=1}^m\partial_{u_j}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})\partial_{x_j} g(\boldsymbol{x},\boldsymbol{u})d\boldsymbol{u} d\boldsymbol{x}\\
&-&\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{B}^{m}}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle g(\boldsymbol{x},\boldsymbol{u})d\boldsymbol{u} d\boldsymbol{x}\\
&=&-\int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{B}^{m}}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle g(\boldsymbol{x},\boldsymbol{u})d\boldsymbol{u} d\boldsymbol{x}.
\end{eqnarray*}
where the equation above comes from the fact that $\partial_{u_j}\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})\in\mathcal{H}_{k-2}$ and $\partial_{x_j} g(\boldsymbol{x},\boldsymbol{u})\in\mathcal{H}_k$, which implies that they are orthogonal to each other. Now, we apply Stokes' Theorem to $\boldsymbol{x}$ and $\boldsymbol{u}$ separately and also use the orthogonality between $\mathcal{H}_k$ and $\mathcal{H}_{k-2}$, we eventually obtain
\begin{eqnarray}\label{etwo}
&& \int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
&=& \int\displaylimits_{\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle \boldsymbol{u},D_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
&+& \int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}}\langle \boldsymbol{u},n_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\sigma(\boldsymbol{x})\nonumber\\
&-&\int\displaylimits_{\partial\Omega}\int\displaylimits_{\mathbb{S}^{m-1}} f(\boldsymbol{x},\boldsymbol{u})\langle \boldsymbol{u},n_{\boldsymbol{x}}\rangle\langle D_{\boldsymbol{u}},D_{\boldsymbol{x}}\rangle g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\sigma(\boldsymbol{x}).
\end{eqnarray}
Now, we plug \eqref{eone} and \eqref{etwo} into \eqref{ezero}, which completes the proof.
\end{proof}
\begin{remark}
The Green's formula above shows that bosonic Laplacians $\mathcal{D}_k$ are self-adjoint with respect to the inner product $\langle\ |\ \rangle$ given in \eqref{innerproduct} for functions vanishing on the boundary of the domain with respect to $\boldsymbol{x}$.
\end{remark}
\subsection*{Green's formula: Clifford-valued version}
\hfill\\
The difficulty in the Clifford-valued case is caused by the fact that multiplication of Clifford numbers is not commutative in general. In our case, we need to use the connection between Rarita-Schwinger type operators and bosonic Laplacians given in \eqref{connection} and Stokes' Theorems for Rairta-Schwinger type operators given in \cite[Theorem 4,6,8,10]{DR1}.
\begin{theorem}[Green's formula: Clifford-valued version]
Let $\Omega\subset\mathbb{R}^m$ be a bounded domain, we assume that $f,g\in C^2(\Omega\times\mathbb{B}^{m},\mathcal{H}_k(\mathcal{C}l_m))$, then we have
\begin{align*}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}\mathcal{D}_k f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}} f(\boldsymbol{x},\boldsymbol{u})\mathcal{D}_k g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+d\sigma_{\boldsymbol{x}}\bigg(-R_kP_k^++\frac{2Q_kP_k^-}{m+2k-4}\bigg)gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-d\sigma_{\boldsymbol{x}}\bigg(-\frac{2R_kP_k^+}{m+2k-4}+\frac{(m+2k)Q_kP_k^-}{m+2k-4}\bigg)gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}f\bigg(P_{k,r}^+R_{k,r}+\frac{2P_{k,r}^-T_{k,r}}{m+2k-4}\bigg)d\sigma_{\boldsymbol{x}}P_k^+gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}f\bigg(-\frac{2P_{k,r}^+T_{k,r}^*}{m+2k-4}-\frac{(m+2k)P_{k,r}^-Q_{k,r}}{m+2k-4}\bigg)d\sigma_{\boldsymbol{x}}P_k^-gdS(\boldsymbol{u}),
\end{align*}
where $P_{k,r}$ stands for the projection operator $P_k$ acting from the right hand side.
\end{theorem}
\begin{proof}
Recall that
\begin{align*}
\mathcal{D}_k &=-R_k^2P_k^++\frac{2R_kT_kP_k^-}{m+2k-4}-\frac{2Q_kT_k^*P_k^+}{m+2k-4}-\frac{(m+2k)Q_k^2P_k^-}{m+2k-4}\\
&=-R_k^2P_k^++\frac{2T_k^*R_kP_k^+}{m+2k-4}-\frac{2T_kQ_kP_k^-}{m+2k-4}-\frac{(m+2k)Q_k^2P_k^-}{m+2k-4}.
\end{align*}
Then, we will calculate
\begin{eqnarray}\label{id0}
\int_{\Omega}\int_{\mathbb{S}^{m-1}}\mathcal{D}_k f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}
\end{eqnarray}
as the sum of four integrals regarding the four terms in the first expression of $\mathcal{D}_k$ above. In the calculation below, we write the functions $f$ without the variables for convenience. Firstly, we apply \cite[Lemma 5]{Dunkl} and Stokes' Theorem for $R_k$ \cite[Theorem 4]{DR1} to obtain
\begin{align}\label{id1}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_{k,r}^2gdS(\boldsymbol{u})d\boldsymbol{x}
=\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_{k,r}^2P_k^+gdS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&-\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_{k,r}R_kP_k^+gdS(\boldsymbol{u})d\boldsymbol{x}+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_{k,r}d\sigma_{\boldsymbol{x}}P_k^+gdS(\boldsymbol{u})\nonumber\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_k^2P_k^+gdS(\boldsymbol{u})d\boldsymbol{x}-\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+d\sigma_{\boldsymbol{x}}R_kP_k^+gdS(\boldsymbol{u})\nonumber\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+R_{k,r}d\sigma_{\boldsymbol{x}}P_k^+gdS(\boldsymbol{u}).
\end{align}
With the help of Stokes' Theorem for $R_k$ \cite[Theorem 4]{DR1} and Stokes' Theorem for $T_k$ \cite[Theorem 10]{DR1}, we can also have
\begin{align}\label{id2}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-T_{k,r}R_{k,r}gdS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^- T_k^*R_kP_k^+gdS(\boldsymbol{u})d\boldsymbol{x}-\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-d\sigma_{\boldsymbol{x}}R_kP_k^+gdS(\boldsymbol{u})\nonumber\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-T_{k,r}d\sigma_{\boldsymbol{x}}P_k^+gdS(\boldsymbol{u}).
\end{align}
Similarly, one can obtain
\begin{align}\label{id3}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+T_{k,r}^*Q_{k,r}gdS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+T_kQ_kP_k^-gdS(\boldsymbol{u})d\boldsymbol{x}-\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+d\sigma_{\boldsymbol{x}}Q_kP_k^-gdS(\boldsymbol{u})\nonumber\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+T_{k,r}^*d\sigma_{\boldsymbol{x}}P_k^-gdS(\boldsymbol{u}),
\end{align}
and
\begin{align}\label{id4}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-Q_{k,r}^2gdS(\boldsymbol{u})d\boldsymbol{x}\nonumber\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-Q_k^2P_k^-gdS(\boldsymbol{u})d\boldsymbol{x}-\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-d\sigma_{\boldsymbol{x}}Q_kP_k^-gdS(\boldsymbol{u})\nonumber\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-Q_{k,r}d\sigma_{\boldsymbol{x}}P_k^-gdS(\boldsymbol{u}).
\end{align}
Now, we plug \eqref{id1}-\eqref{id4} into \eqref{id0}, we obtain
\begin{align*}
&\int_{\Omega}\int_{\mathbb{S}^{m-1}}\mathcal{D}_k f(\boldsymbol{x},\boldsymbol{u})g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
=&\int_{\Omega}\int_{\mathbb{S}^{m-1}} f(\boldsymbol{x},\boldsymbol{u})\mathcal{D}_k g(\boldsymbol{x},\boldsymbol{u})dS(\boldsymbol{u})d\boldsymbol{x}\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^+d\sigma_{\boldsymbol{x}}\bigg(-R_kP_k^++\frac{2Q_kP_k^-}{m+2k-4}\bigg)gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}fP_{k,r}^-d\sigma_{\boldsymbol{x}}\bigg(-\frac{2R_kP_k^+}{m+2k-4}+\frac{(m+2k)Q_kP_k^-}{m+2k-4}\bigg)gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}f\bigg(P_{k,r}^+R_{k,r}+\frac{2P_{k,r}^-T_{k,r}}{m+2k-4}\bigg)d\sigma_{\boldsymbol{x}}P_k^+gdS(\boldsymbol{u})\\
&+\int_{\partial\Omega}\int_{\mathbb{S}^{m-1}}f\bigg(-\frac{2P_{k,r}^+T_{k,r}^*}{m+2k-4}-\frac{(m+2k)P_{k,r}^-Q_{k,r}}{m+2k-4}\bigg)d\sigma_{\boldsymbol{x}}P_k^-gdS(\boldsymbol{u}),
\end{align*}
as desired. We remind the reader that $\mathcal{D}_k$ on the right hand side above is obtained from the second expression of $\mathcal{D}_k$ given in the very beginning of the proof.
\end{proof}
\begin{remark}
If we replace $g$ by the fundamental solution of $\mathcal{D}_k$ in the previous theorem, one can have the Borel-Pompeiu formula obtained in \cite{DR} with a standard argument at the singular point.
\end{remark}
| 889bc7ce1ae59fcacf22aa3bb2fd70398c014634 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Boolean automata networks (BANs) are studied for their capacity to succintly
expose the complexity that comes with the composition of simple entities into
a network. They belong to a wide family of systems which include
cellular automata and neural networks, and can be described as cellular
automata with arbitrary functions and on arbitrary graph structures.
Understanding and predicting the dynamics of computing with BANs has been a
focus of the scientific community which studies them, in particular since
their applications include the modelling of gene regulatory
networks~\cite{J-Kauffman1969,J-Thomas1973,J-Mendoza1998,J-Davidich2008,J-Demongeot2010}.
In those applications, fixed points of a BAN are often viewed as
cellular types and limit cycles as biological
rhythms~\cite{J-Kauffman1969,J-Thomas1973}.
It follows that most biological studies
relying on BANs require the complete computation of their dynamics to propose
conclusions.
The complete computation of the dynamics of BANs is an exponentially costly
process. Indeed, for $n$ the size of a BAN, the size of its dynamics
is precisely $2^n$. The dynamics of a BAN is usually partitionned in two sorts
of configurations: the recurring ones that are parts of attractors and
either belong to a limit cycle or are fixed points; the others that evolve
towards these attractors and belong to their attraction basins.
The questions of characterising, computing or counting those attractors from a
simple description of the
network have been
explored~\cite{C-Orponen1989,J-Aracena2008,J-Goles2008,J-Demongeot2012,C-Noual2012,J-Aracena2017},
and have been shown to be difficult
problems~\cite{C-Orponen1989,C-Orponen1992,C-Bridoux2019,U-Bridoux2020,C-Nous2020}.
In this paper, we propose a new method for computing the
attractors of a BAN under the parallel update schedule. For any input network,
this method generates another network which is possibly smaller and which is
guaranteed to possess attractors isomorphic to those of the input network.
Computing
the dynamics of this smaller network therefore takes as much time as needed
to compute the dynamics of the input networks, divided by some power of two.
This method uses tools and results developed in previous works by the
authors~\cite{C-Perrot2018,C-Perrot2020}.
These works involve adding inputs to BANs, in a generalisation called modules
that proposes in some cases the study of the computationnal capabilities of the
network as the computation in terms of the inputs.
In particular, a result states that two networks that have equivalent such
computations share isomorphic attractors.
Section~\ref{s:Definitions} starts by exposing all the definitions
needed to read this paper.
Section~\ref{s:PartA} explores the question of obtaining an acyclic module
(AM) from
a BAN. Section~\ref{s:PartB} explains how to extract so called output functions
from a module. Section~\ref{s:PartC} details how to generate a minimal module
from a set of output functions. Finally Section~\ref{s:PartD} shows the final
step of the method, which implies constructing a BAN out of an AM
and computing its dynamics. Each section explores complexity results
of the different parts of the process, and details examples along the way.
An illustrative outline of the paper can be found in Figure~\ref{f:Plan}.
\input{figures/plan}
\section{Definitions} \label{s:Definitions}
\subsection{Boolean functions}
In this paper, we consider a Boolean function as any function
$f : \mathbb{B}^A \to \mathbb{B}$, for $A$ a finite set. An affectation $x$ of $f$ is a vector
in $\mathbb{B}^A$.
When considered as the input or output of a complexity problem, we encode
Boolean functions as Boolean circuits.
A \emph{Boolean circuit} of $f$ is an acyclic digraph in which nodes without
incoming edges are
labelled by an element in $A$, and every other node by a Boolean gate in
$\{\wedge, \vee, \neg\}$, with a special node marked as the output of the circuit.
The evaluation $f(x)$ is
computed by mapping $x$ to the input nodes of the circuit, and propagating
the evaluation along the circuit using the gates until the output node is
reached.
\subsection{Boolean automata networks and acyclic modules}
\subsubsection{Boolean automata networks}
BANs are composed of a set $S$ of automata.
Each automaton in $S$, or node, is at any time in a state in $\mathbb{B}$.
Gathering those isolated states into a vector of dimension $|S|$ provides us
with a configuration of the network.
More formally, a \emph{configuration} of $S$ over $\mathbb{B}$ is a vector in
$\mathbb{B}^S$.
The state of every automaton is bound to evolve as a function of the
configuration of the entire network. Each node has a unique function, called
a local function, that is predefined and does not change over time.
A \emph{local function} is thus a function $f$ defined as $f: \mathbb{B}^S \to
\mathbb{B}$.
Formally, a BAN $F$ is a set that assigns a local
function $f_s$ over $S$ for every $s \in S$.
\begin{example}
\label{ex1-BAN-def}
Let $S_A = \{a, b, c, d\}$. Let $F_A$ be the BAN defined by
$f_a(x) = x_d$, $f_b(x) = f_c(x) = x_a$, and
$f_d(x) = \neg x_b \vee \neg x_c$.
The interaction digraph of this BAN is depicted in Figure~\ref{f:ex:BAN}
(left panel).
\end{example}
\begin{example}
\label{ex2-BAN-def}
Let $S_B = \{St, Sl, Sk, Pp, Ru, S9, C, C25, M, C^*\}$. Let $F_B$ be the BAN
defined by
$f_{St}(x) = \neg x_{St}$,
$f_{Sl}(x) = \neg x_{Sl} \vee x_{C^*}$,
$f_{Sk}(x) = x_{St} \vee \neg x_{Sk}$,
$f_{Pp}(x) = x_{Sl} \vee \neg x_{Pp}$,
$f_{Ru}(x) = f_{S9}(x) = \neg x_{Sk} \vee x_{Pp} \vee \neg x_C \vee \neg x_{C^*}$,
$f_C(x) = \neg x_{Ru} \vee \neg x_{S9} \vee \neg x_{Sl}$,
$f_{C25}(x) = \neg x_{Pp} \vee x_C$,
$f_{M}(x) = x_{Pp} \vee \neg x_C$,
and $f_{C^*}(x) = \neg x_{Ru} \vee \neg x_{S9} \vee x_{C25} \vee \neg x_M$.
The interaction digraph of this BAN is depicted in Figure~\ref{f:ex:BAN}
(right panel).
\end{example}
\input{figures/exampleBAN}
In the scope of this paper, BANs (and modules) are udpated according to
the parallel update schedule.
Formally, for $F$ a BAN and $x$ a configuration of $F$, the update of $x$ under
$F$ is denoted by configuration $F(x)$, and defined as for all $s$ in $S$,
$F(x)_s = f_s(x)$.
\begin{example}
\label{ex1-BAN-update}
Consider $F_A$ of Example~\ref{ex1-BAN-def}, and $x \in \mathbb{B}^{S_A}$ such that
$x = 1001$.
We observe that $F_A(x) = 1111$. Configurations $1000$ and $0111$ are recurring
and form a limit cycle of size $2$,
as well as configurations $0000$, $0001$, $1001$, $1111$, $1110$ and $0110$
that form a limit cycle of size $6$.
\end{example}
\subsubsection{Interaction digraph}
BANs are usually represented by the influence that automata hold on each other.
As such the visual representation of a BAN is a digraph, called an
interaction digraph, whose nodes are the automata of the network,
and arcs are the influences that link the different automata.
Formally, $s$ \emph{influences} $s'$ if and only if there exist two
configurations $x,x'$ such that $f_{s'}(x) \neq f_{s'}(x')$
and for all $r$ in $S$, $r \neq s$ if and only if $x_r = x'_r$.
\subsubsection{Dynamics}
Finally, we define the \emph{dynamics} of a BAN $F$ as the digraph with $\mathbb{B}^S$ as its
set of vertices. There exists an edge from $x$ to $y$ if and only if $F(x) = y$.
Computing the dynamics of a BAN from the description of its local function
is an exponential process.
See~\cite{B-Robert1986} for a more throughout introduction to BANs and
related subjects.
\subsubsection{Modules}
Modules were first introduced in~\cite{C-Perrot2018}.
A module $M$ is a BAN with added inputs. It is defined on two sets: $S$
a set of automata, and $I$ a set of inputs, with $S \cap I = \emptyset$.
Similarly to standard BANs, we can define configurations as vectors in $\mathbb{B}^S$,
and we define input configurations as vectors in $\mathbb{B}^I$.
A local function of a module updates itself based on a configuration $x$
and an input configuration $i$, concatenated into one configuration.
Formally, a local function is defined from
$\mathbb{B}^{S \cup I}$ to $\mathbb{B}$. The module $M$ defines a local function for every node
$s$ in $S$.
\begin{example}
Let $M_e$ be the module defined on $S_e = \{p, q, r\}$ and $I = \{\alpha, \beta\}$,
such that $f_p(x) = x_\alpha$,
$f_q(x) = \neg x_p$,
and $f_r(x) = x_q \vee \neg x_\beta$.
\end{example}
We represent modules with an interaction digraph, in the same way as for BANs.
The interaction digraph of a module has added arrows that represent the
influence of the inputs over the nodes; for every node $s$ and every input
$\alpha$, the node $s$ of the interaction digraph has an ingoing arrow labelled
$\alpha$ if and only if $\alpha$ influences $s$, that is, there exists two
input configurations $i,i'$ such that for all $\beta$ in $I$,
$\beta \neq \alpha$ if and only if
$i_\beta = {i'}_\beta$, and $x$ a configuration such that
$f_s(x\cdot i) \neq f_s(x\cdot i')$, where $\cdot$ denotes the concatenation
operator.
A module is \emph{acyclic}
if and only if its interaction digraph is cycle-free.
\subsubsection{Recursive wirings}
A recursive wiring over a module $M$ is defined by a partial function
$\omega : I \not\to S$. The result of such a wiring is denoted
$\circlearrowright_\omega M$, a module defined over sets $S$ and
$I \setminus \mathrm{dom}(\omega)$, in which the local function of node $s$ is denoted
$f'_s$ and defined as
\begin{equation*}
\forall x \in \mathbb{B}^{S \cup I},\ f'_s(x) = f_s( x \circ \hat \omega ),
\text{ with } \hat\omega(i) = \begin{array}\{{ll}.
\omega(i) & \text{if } i \in \mathrm{dom}(\omega)\\
i & \text{if } i \in I \setminus \mathrm{dom}(\omega)
\end{array}\text{.}
\end{equation*}
\subsubsection{Output functions}
Output functions were first introduced in~\cite{C-Perrot2020} and present another
way of computing the evolution of an acyclic module. In the Boolean case, those
functions are defined on $\mathbb{B}^{I \times \{1, \ldots, D\}} \to \mathbb{B}$,
for I the input set of the module, and D some integer.
We interpret an input in $\mathbb{B}^{I \times \{1, \ldots, D\}}$ as an evaluation over
$\mathbb{B}$ of a set of variables ${I \times \{1, \ldots, D\}}$,
and for $\alpha \in I$ and $d \leq D$, we denote this variable by $\alpha_d$.
In the context of an acyclic module $M$, $\alpha_d$
is refering to the evaluation of the
input $\alpha$ on the $d$th update of the module. A vector
$j \in \mathbb{B}^{I \times \{1, \ldots, D\}}$ simply describes an evaluation of all
the inputs of the network over $D$ iterations. With such a vector, and
$x \in \mathbb{B}^S$, it is easy to see that the acyclic module $M$ can be updated
$k$ times in a row, for any $k \leq D$.
The result of this update is denoted by $M(x, j_{[1, \ldots, k]})$.
The \emph{delay} of an output function $O$ is the maximal value in the set of
all the $d \in \mathbb{N}$ for which there exists $\alpha \in I$ such that variable
$\alpha_d$ has an influence on the computation of $O$.
Finally,
for $M$ an acyclic module defined on the sets $S$ and $I$,
for $D$ a large enough integer, for $x \in \mathbb{B}^S$ and
$j \in \mathbb{B}^{I \times \{1, \ldots, D\}}$ some vectors, and for $s$ a node in $S$,
we define the output function of $s$, denoted $O_s$, as the output function
with minimal delay $d$ such that $O_s(j) = M(x, j_{[1, \ldots, k]})_s$.
Such a function always exists and is always unique.
\subsection{Promise problems}
In this paper, we make the hypothesis that every module passed as an input
of a complexity problem follows the property that each of its local functions
has only \emph{essential} variables. That is, a variable is included in the
circuit encoding that function if and only if the automaton or input represented
by that variable has an influence on said function.
This hypothesis will be implemented throughout this paper by the use of
promise problems~\cite{O-Goldreich2006}, which include a decision method which
can dismiss instances of the problem without that method's complexity cost
being included in the complexity of the problem.
This approach is motivated by the fact that obfuscating the relation
between automatons by building redundant variables in a circuit
increases the complexity of most considered problems. We justify our
decision in two points: first, the approach of this paper is one of providing
and studying an applicable method in a context where misleading inputs in
local functions are unlikely. Second, despite the inclusion of these
promises
we find high complexity issues in our pipeline, and as such we consider that
it helps understanding the precise issues that
prevent our method from being efficient.
\section{From BANs to AMs} \label{s:PartA}
The first step of our process is to unfold a BAN into an AM. This simply
requires the removal of any cycle in the interaction digraph of the
BAN, and their replacement by inputs.
In the scope of this paper, the number of inputs generated is required to be
minimal. This is justified by the fact that the complexity of most of the
problems addressed in the pipeline highly depends on the number of inputs
of the considered AM.
\begin{fpproblem}
\fpproblemtitle{Acyclic Unfolding Functional Problem}
\fpprobleminput{A Boolean automata network $F$, an integer $k$.}
\fpproblempromise{The encoding of the local functions of $F$ only has
essential variables.}
\fpproblemoutput{An acyclic module $M$ with at most $k$ inputs and a
recursive wiring $\omega$ such that $\circlearrowright_\omega M = F$.}
\end{fpproblem}
\begin{theorem}
\label{th-unfold-up}
The Acyclic Unfolding Functional Problem is in $\text{FNP}$.
\end{theorem}
\begin{proof}
The promise of this problem allows us to compute the interaction digraph of
$F$ in polynomial time.
Consider the following simple non-deterministic algorithm: first guess a module
$M$ and a wiring $\omega$; then check that the number of inputs in $M$ is no more than $k$
and that $\circlearrowright_\omega M$ syntactically equals $F$.
This algorithm operates in polynomial non-deterministic time since the
recursive wiring is a simple substitution of variables,
and thanks to the fact that one only needs to compare
$\circlearrowright_\omega M$ and $F$ at a syntactical level.
Indeed, if any solution exists, then a solution exists with the
same number of nodes, the same inputs, the same wirings, and such that
the substitution operated by $\omega$ on $M$ leads to the local functions
of $F$ written identically:
because all local functions are equal on a semantic level,
this is always possible,
by starting from the local functions of $F$ and operating variable
substitutions that are then reversed by the recursive wiring $\omega$
(remark that this last ``reversed'' construction is not required to be
computable in polynomial time).
\end{proof}
\begin{theorem}
\label{th-unfold-down}
The Acyclic Unfolding Functional Problem is $\text{NP}$-hard.
\end{theorem}
\begin{proof}
Let us provide a reduction from the Feedback Vertex Set problem. We provide
$f$ a function that for any instance $(G, k)$ of the Feedback Vertex Set problem,
provides an instance $(F, k)$ of the Optimal Acyclic Unfolding Problem
where $S = V(G)$ and $f_s$ is a OR function of exactly every node $s'$ such
that $(s', s) \in A(G)$. This construction is explicitly designed so that
the interaction digraph of $F$ is isomorphic to $G$. Clearly $f$ is computable
in polynomial time.
We also provide $g$ a function that for $(G, k)$ an instance of the
Feedback Vertex Set problem, and $M$ a solution to the
Optimal Acyclic Unfolding Problem, checks if $S = V(G)$, and then
deduces the solution for $(G, k)$ the following way: $s$ is part of
the feedback vertex set if and only if its influence has been replaced by an
input
in $M$. This means that the
variable $x_s$ has been replaced in every local function in $M$ by the same
input variable.
Finally $g$ checks that the size of the obtained set is not greater
than $k$. It is clear that $g$ is polynomial.
From the definition of $f$ and $g$, it follows that
the Feedback Arc Set problem reduces in polynomial time to
the Optimal Acyclic Unfolding Problem, which implies the result.
\end{proof}
\begin{example}
\label{ex1-AM}
Consider $S_A$ and $F_A$ of Example~\ref{ex1-BAN-def}. Let us define
$I_A = \{\alpha\}$. Let $M_A$ be the acyclic module that defines
$f'_a(x) = x_{\alpha}$,
$f'_b(x) = f'_c(x) = x_a$,
and $f'_d(x) = \neg x_b \vee \neg x_c$. The module $M_A$ is a valid answer to
the instance $F_A, k = 1$ of the Acyclic Unfolding Functional Problem.
The interaction digraph of this module is represented in Figure~\ref{f:ex:MOD}
(left panel).
\end{example}
\begin{example}
\label{ex2-AM}
Consider $S_B$ and $F_B$ of Example~\ref{ex2-BAN-def}. Let us define
$I_B = \{\alpha_{St}, \alpha_{Sl}, \alpha_{Sk},$
$\alpha_{Pp}, \alpha_{C}, $ $\alpha_{C^*}\}$.
Let $M_B$ be the acyclic module that defines
$f'_{St}(x) = \neg x_{\alpha_{St}}$,
$f'_{Sl}(x) = \neg x_{\alpha_{Sl}} \vee x_{\alpha_{C^*}}$,
$f'_{Sk}(x) = x_{\alpha_{St}} \vee \neg x_{\alpha_{Sk}}$,
$f'_{Pp}(x) = x_{\alpha_{Sl}} \vee \neg x_{\alpha_{Pp}}$,
$f'_{Ru}(x) = f_{S9}(x) = \neg x_{\alpha_{Sk}} \vee x_{\alpha_{Pp}}
\vee \neg x_{\alpha_C} \vee \neg x_{\alpha_{C^*}}$,
$f'_C(x) = \neg x_{Ru} \vee \neg x_{S9} \vee \neg x_{\alpha_{Sl}}$,
$f'_{C25}(x) = \neg x_{\alpha_{Pp}} \vee x_{\alpha_C}$,
$f'_{M}(x) = x_{\alpha_{Pp}} \vee \neg x_{\alpha_C}$,
and $f'_{C^*}(x) = \neg x_{Ru} \vee \neg x_{S9} \vee x_{C25} \vee \neg x_M$.
The module $M_B$ is a valid answer to
the instance $F_B, k = 6$ of the Acyclic Unfolding Functional Problem.
The interaction digraph of this module is represented in
Figure~\ref{f:ex:MOD} (right panel).
\end{example}
\input{figures/exampleMOD}
\section{Output functions} \label{s:PartB}
Output functions were first introduced in~\cite{C-Perrot2020}. They
are a way to characterise the asymptotic behaviour of an AM as a
set of Boolean functions that are computed from the local functions of the AM.
Computing the output functions of an AM is a crucial step in the pipeline
proposed in this work.
\begin{fpproblem}
\fpproblemtitle{Output Circuit Computation Problem}
\fpprobleminput{An acyclic module $M$, and $X \subseteq S$ a set of output nodes.}
\fpproblempromise{The encoding of the local functions of $M$ only has
essential variables.}
\fpproblemoutput
{An output function for each node in $X$, encoded as a Boolean circuit.}
\end{fpproblem}
\begin{theorem}
The Output Circuit Computation Problem is in FP.
\end{theorem}
\begin{proof}
The promise of this problem allows us to deduce
the interaction digraph of $M$ in polynomial time.
To compute $X$, we provide an algorithm to compute the output function circuit
of any node $s \in S$ in polynomial time.
The algorithm first constructs a list of requirements.
This list is initialy $R_0 = \{(s, 0)\}$, which can be interpreted to say that
we require the construction of the output
function of $s$ with added delay $0$.
We construct the next list the following way:
$(t', d') \in R_{k+1}$ if and only if there exists some $(t, d) \in R_{k}$ such
that $t'$ influences $t$ in $M$.
The total list $R$ is simply defined as $R = \bigcup_{i \in \mathbb{N}} R_i$.
\begin{claim}
$R$ is computable in polynomial time in the size of $M$.
\end{claim}
To see that this is true, consider that for $D$ the maximal depth of $M$,
the maximal $d$ such that $(t, d) \in R_k$ for any $k$ is $D$. Indeed since
the interaction digraph of $M$ is acyclic, the maximal delay value can only
be obtained by following the longest path in $M$. As such we can conclude that
the size of any $R_k$ is bounded by $D \times n$, for $n$ the size of the
network $M$. Finally, by a similar argument,
consider that the list $R$ converges after a maximum of $D$ steps. This implies
that the list $R$ is computed after $D$ steps of a $D \times n$ costly
process, and $R$ can therefore be computed in polynomial time.
We can construct the Boolean circuit from $R$ in the following way: for
every pair $(t, d) \in R$, take an instance of the Boolean circuit which encodes
the local function of $t$. Combine all of these instances the following way:
any input variable in $I$ is replaced by its delayed counterpart with delay
$1 + d$. For example, if a variable $\alpha$ appeared in the local function
of node $t$, substitute it by the variable $\alpha_{d + 1}$. Then, for any
gate displaying an input variable $t' \in S$, replace it with the same gate,
which rather than taking the value of variable $t'$, takes the value of the
output of the circuit that computes the local function of the node $t'$ with
added delay $d + 1$. By definition of $R$ this circuit will always be in $R$.
The obtained circuit computes the output function of the node $s$.
Repeat this process for every $s \in X$.
\end{proof}
\begin{example}
\label{ex1-OUT}
Consider $M_A$ of Example~\ref{ex1-AM}. Let
$X_A = \{d\}$ be an instance of the
Output Circuit Computation Problem. The circuit
$O_d = \neg \alpha_3$ is a
valid answer to that instance.
\end{example}
\begin{example}
\label{ex2-OUT}
Consider $M_B$ of Example~\ref{ex2-AM}. Let
$X_B = \{St, Sk, Sl, Pp, C, C^*\}$ be an instance of the
Output Circuit Computation Problem. The circuits
$O_{St} = \neg \alpha_{St, 1}$,
$O_{Sl} = \neg \alpha_{Sl, 1} \vee \alpha_{C^*, 1}$,
$O_{Sk} = \alpha_{St, 1} \vee \neg \alpha_{Sk, 1}$,
$O_{Pp} = \alpha_{Sl, 1} \vee \neg \alpha_{Pp, 1}$,
$O_C = ( \alpha_{Sk, 2} \wedge \neg \alpha_{Pp, 2}
\wedge \alpha_{C, 2} \wedge \alpha_{C^*, 2} ) \vee \neg \alpha_{Sl, 1}$
and $O_{C^*} = \alpha_{C, 2} \vee \neg \alpha_{Pp, 2}$ taken altogether are a
valid answer to that instance.
\end{example}
\section{Optimal acyclic module synthesis} \label{s:PartC}
\subsection{Module Synthesis}
This part of the process takes in a set of output functions and generates a
module that realizes these functions with an hopefully minimal number of nodes.
In this part the actual optimisation of the pipeline, if any, can
be directly observed. It is also the part of the pipeline which is the most
computationnaly costly.
\begin{fproblem}
\fproblemtitle{Module Synthesis Problem}
\fprobleminput{A set $I$ of input labels, a finite set of output functions $O$,
encoded as Boolean circuits,
defined on those labels, and $k$ an integer.}
\fproblemoutput{An acyclic module $M$ with at most $k$ nodes such that every
function in $O$ is the output function of at least one node in $M$.}
\end{fproblem}
\subsection{Complexity results}
\begin{theorem}
\label{th-syn-down}
The Module Synthesis Problem is coNP-hard.
\end{theorem}
\begin{proof}
Consider $f$ an instance of the Tautology problem, with $I$ the set of propositional
variables contained in $f$. We define $f'$ as the output function defined on the
labels $I$ such that $f'$ is obtained from $f$ by substituting all variables
$\alpha \in I$ by their equivalent of delay 1, $\alpha_1$. Let us also define
$f_1$ as the constant output function of delay $0$ which value is always $1$.
We compose an instance of the Module Synthesis Problem with $I$ the set of input
labels, $O = \{f', f_1\}$ and $k = 0$. This instance has a solution if and only
there exists an acyclic module with only one node such that the output function
of this node is equivalent to all the output functions in $O$. This implies that,
if the problem has a solution, $f'$ is equivalent to $f_1$, which proves that
$f'$ and $f$ are tautologies.
Therefore computing the output of the Module Synthesis
Problem requires solving a coNP-hard decision problem.
\end{proof}
\begin{theorem}
The Module Synthesis Problem is in $\text{FNP}^\text{coNP}$.
\end{theorem}
\begin{proof}
Consider the following algorithm. First, guess an acyclic module $M$, with
size $k$. Compute every output function of the network, which is in FP. Then
simply check that every function in $O$ is equivalent to at least one
output function in $M$, which requires at most $|M| \times |O|$ calls to
a $\text{coNP}$\ oracle.
\end{proof}
\subsection{Refining the complexity bounds}
It is unclear whether the synthesis problem can be proven to be in $\text{FcoNP}$\ or
to be $\text{NP}^\text{coNP}$-hard. An attempt has been made to prove the former by using a
greedy algorithm which would fuse nodes in an acyclic module, starting
from a trivially large enough module. This method requires solving the
following problem:
\begin{dproblem}
\problemtitle{Module Local Fusion Problem}
\probleminput{An acyclic module $M$ defined on sets $S$ and $I$,
and $a, b$ two different nodes.}
\problemquestion{Is there a local function $f_c$ such that there exists
some acyclic module $M'$ defined on the node set
$S \cup \{c\} \setminus \{a, b\}$ and input set $I$, such that $f_c \equiv f'_c$
and $O_s \equiv O'_s$ for $s \in S \setminus \{a, b\}$?}
\end{dproblem}
This problem formalises the idea of replacing two nodes by one in an acyclic
module, such that every other output function in the module is conserved.
Assuming the removed nodes are not considered outputs of the network is
an important step of any greedy algorithm that would try to optimise the size
of an acyclic module.
It is rather simple to prove that this problem is coNP-hard, since its computation
requires checking the equivalence of multiple pairs of Boolean circuits. It
is also rather easy to see that it is in $\text{NP}^\text{coNP}$, as one can guess $f_c$ and
$M'$ in polynomial time and verify the solution using a polynomial amount of
calls to a $\text{coNP}$\ oracle, one for every equivalence check.
The function $f_c$ could be composed as a binary function of the results
of $f_a$ and $f_b$, as it is intuitive to suppose that assuming such a fusion
is possible, then every node influenced by $a$ or $b$ should be computable
from such a composition of $a$ and $b$. The issue however is that this process
requires modifying every node influenced by $a$ or $b$ such that their output
functions match the output functions in $M$.
It is unclear that there should exist a method in $\text{coNP}$\ to ensure this
modification such that the output functions are conserved.
This leads us to
believe that a greedy algorithm wouldn't prove the Optimal Module Synthesis
Problem to be in $\text{FcoNP}$.
Similarly, it is interesting to consider the open question of whether or not the
Module Synthesis Problem can be proven $\text{NP}^\text{coNP}$-hard. This implies to prove,
between other things, that the problem is $\text{NP}$-hard. This is, to us, another
open problem as the Module Synthesis Problem does not seem equiped to compute
the satisfaction of a Boolean formula or circuit.
\subsection{Similarities to other optimisation problems}
This open question bears strong ressemblance to another open problem that
concerns Boolean circuits. The Circuit Minimisation Problem is a problem that
asks to provide a Boolean circuit below a given size such that it computes
a Boolean function given as a truth table as the input of the
problem~\cite{C-Kabanets2000}. The problem is trivially in $\text{NP}$\ but it is not
known whether the problem is in $\text{P}$ or $\text{NP}$-hard, as both possibilities
imply
proving other results which seem beyond the currently known techniques.
The same problem has been found to be $\text{NP}$-complete in both restricted (DNFs) and
generalised (unrestricted Boolean circuits)
variations of the Boolean circuit model~\cite{C-Ilango2020}.
There are strong similarities between acyclic modules and Boolean circuits.
Both are defined on acyclic digraphs, have inputs and outputs,
and compute Boolean functions. It is important to note that this analogy is
misleading when talking about the optimisation of their size. Optimising
a Boolean circuit requires the optimisation of a Boolean function in terms
of the number of gates that computes it. Optimising an acyclic module, however,
requires the optimisation of a network of functions with respect to a notion
of delay of the inputs,
whereas in this case one node may contain an arbitrary Boolean function.
As such these problems seem too independent
to provide any reduction between them.
\subsection{Examples}
\begin{example}
\label{ex1-SYN}
Consider the output function $O_d$ defined in Example~\ref{ex1-OUT}.
Let us define $M'_A$ as the module defined on $S'_A = \{a, b, d\}$ and
$I_A = \{\alpha\}$, such that
$f''_a = x_{\alpha}$,
$f''_b = x_a$ and $f_d = \neg x_b$.
The module $M'_A$ is a valid answer to the instance $I_A, \{O_d\}, k = 3$ of
the Module Synthesis Problem.
The interaction digraph of this module is depicted in Figure~\ref{f:ex:MODSYN}
(left panel).
\end{example}
\begin{example}
\label{ex2-SYN}
Consider the output functions $O_B = \{O_{St}, O_{Sl}, O_{Sk}, O_{Pp}, O_C, O_{C^*}\}$
defined in Example~\ref{ex2-OUT}.
Let us define $M'_B$ as the module defined on
$S'_B = \{St, Sl, Sk,$ $ Pp, Ru, C25, C^*\}$ and
$I_B = \{\alpha_{St}, \alpha_{Sl}, \alpha_{Sk}, \alpha_{Pp},
\alpha_C, \alpha_{C^*}\}$, such that
$f''_{St}(x) = \neg x_{\alpha_{St}}$,
$f''_{Sl}(x) = \neg x_{\alpha_{Sl}} \vee x_{\alpha_{C^*}}$,
$f''_{Sk}(x) = x_{\alpha_{St}} \vee \neg x_{\alpha_{Sk}}$,
$f''_{Pp}(x) = x_{\alpha_{Sl}} \vee \neg x_{\alpha_{Pp}}$,
$f''_{Ru}(x) = \neg x_{\alpha_{Sk}} \vee x_{\alpha_{Pp}}
\vee \neg x_{\alpha_C} \vee \neg x_{\alpha_{C^*}}$,
$f''_C(x) = \neg x_{Ru} \vee \neg x_{\alpha_{Sl}}$,
$f''_{C25}(x) = \neg x_{\alpha_{Pp}} \vee x_{\alpha_C}$,
and $f'_{C^*}(x) = x_{C25}$.
The module $M'_B$ is a valid answer to the instance $I_B, O_B, k = 8$ of
the Module Synthesis Problem.
The interaction digraph of this module is depicted in Figure~\ref{f:ex:MODSYN}
(right panel).
\end{example}
\input{figures/exampleMODSYN}
\section{Final wiring and analysis} \label{s:PartD}
The final step in the pipeline is simply to wire the module obtained in
Section~\ref{s:PartC} so that the obtained networks hold isomorphic attractors
to the input network. This is ensured by application of the following
result..
\begin{theorem}[\cite{C-Perrot2020}]
\label{th-limit-from-output}
Let $M$ and $M'$ be two acyclic modules, with $T$ and $T'$ subsets of their nodes
such that $|T| = |T'|$. If there exists $g$ a bijection from $I$ to $I'$ and
$h$ a bijection from $T$ to $T'$ such that for
every $s \in T$, $O_s$ and $O'_{h(s)}$ have same delay, and for every input
sequence $j$ with length the delay of $O_s$,
\[O_s(j) = O'_{h(s)}(j \circ g^{-1})\]
then for any function $\omega : I \to T$, the networks
$\circlearrowright_{\omega} M$ and $\circlearrowright_{h \circ \omega \circ g^{-1}} M'$ have
isomorphic attractors (up to the renaming of automata given by $h$).
\end{theorem}
Applying this theorem to the current problem is simple: the module $M$ is the
module obtained in Section~\ref{s:PartA}, and the module $M'$ is the
module obtained in Section~\ref{s:PartC}. The set $T$ is the set of nodes
which are substituted by new inputs in the process described in
Section~\ref{s:PartA}.
The set $T'$ is the set of nodes in $M'$ which are considered as the output
of the module, for example when the module $M'$ is obtained as the result
of the application of the functional problem defined in Section~\ref{s:PartC}.
As modules $M$ and $M'$ are defined over the same set of inputs, the
bijection $g$ is the identity. The bijection $h$ is directly constructed so
that for all $s \in T$, $h(s)$ in $M'$ has an equivalent output function as
$s$ in $M$, which is always possible thanks to the careful structure of our
pipeline. It follows quite clearly that for any $s \in T$, and for
any input sequence $j$, $O_s(j) = O'_{h(s)}(j \circ g^{-1})$ holds, and
the theorem applies.
\begin{example}
\label{ex1-FIN}
Consider $M'_A$ of Example~\ref{ex1-SYN}. Let $\omega_A(\alpha) = d$. The AN
$\circlearrowright_{\omega_A} M'_A$
is defined over $S'_A = \{a, b, d\}$ such that
$f'''_a(x) = x_d$,
$f'''_b(x) = x_a$,
$f'''_d(x) = \neg x_b$.
The interaction digraph of this module is depicted in Figure~\ref{f:ex:FIN}
(left panel).
\end{example}
\begin{example}
\label{ex2-FIN}
Consider $M'_B$ of Example~\ref{ex2-SYN}. Let $\omega_B(\alpha_s) = s$,
for all $s \in X_B$.
The AN $\circlearrowright_{\omega_B} M'_B$
is defined over $S'_B = \{St, Sl, Sk, Pp, Ru, C25, C^*\}$ such that
$f'''_{St}(x) = \neg x_{St}$,
$f'''_{Sl}(x) = \neg x_{Sl} \vee x_{C^*}$,
$f'''_{Sk}(x) = x_{St} \vee \neg x_{Sk}$,
$f'''_{Pp}(x) = x_{Sl} \vee \neg x_{Pp}$,
$f'''_{Ru}(x) = \neg x_{Sk} \vee x_{Pp}
\vee \neg x_C \vee \neg x_{C^*}$,
$f'''_C(x) = \neg x_{Ru} \vee \neg x_{Sl}$,
$f'''_{C25}(x) = \neg x_{Pp} \vee x_C$,
and $f'_{C^*}(x) = x_{C25}$.
The interaction digraph of this module is depicted in Figure~\ref{f:ex:FIN}
(right panel).
\end{example}
This allows us to compute the attractors of any BAN by computing the dynamics
of another BAN with possibly less nodes, thus dividing the number of computed
configurations by some power of two. Examples throughout this paper showcase
the application of the pipeline over two initial examples.
Examples~\ref{ex1-BAN-def}, \ref{ex1-AM}, \ref{ex1-OUT}, \ref{ex1-SYN} and
\ref{ex1-FIN} show the optimisation of a simple four nodes network into a three
nodes equivalent network.
The optimisation proceeds here by \lq compacting\rq\ two
trivially equivalent nodes, $b$ and $c$, into one. The resulting BAN has
dynamics $2^1$ times smaller than the initial network, with isomorphic attractors.
Examples~\ref{ex2-BAN-def}, \ref{ex2-AM}, \ref{ex2-OUT}, \ref{ex2-SYN} and
\ref{ex1-FIN} show the optimisation of a larger, more intricate network which
is drawn from a model predicting the cell cycle sequence of fission
yeast~\cite{J-Davidich2008}. This practical example, processed through our
pipeline, reduces from $10$ nodes to $8$. This implies a reduction in dynamics
size of $2^2$, while keeping isomorphic attractors.
Both sets of examples are illustrated throughout the paper in
Figures~\ref{f:ex:BAN}, \ref{f:ex:MOD}, \ref{f:ex:MODSYN} and \ref{f:ex:FIN}.
\input{figures/exampleFIN}
\section{Conclusion}
The present paper showcases an innovative way of reducing the cost of computing
the attractors of Boolean automata networks.
The method provides better optimisation
on networks showing structural redundancies, which are removed by the pipeline.
The limitations of this method
are still significant; it requires solving a problem that is at
least $\text{coNP}$-hard, and believed to be $\text{FNP}^\text{coNP}$-complete.
As it presently stands, this method is not as much a convincing practical tool as it is
a good argument in favor of the powerfulness of acyclic modules, their
output functions, and the approaches they allow together towards the
computation of BAN dynamics.
Future perspectives include finding better complexity bounds to the
Module Synthesis Problem, and generalising the formalism of output functions
and the optimisation pipeline to
different update schedules distinct from parallel.
\bibliographystyle{plain}
{\small{ | 9a1c89ca531a182b0990a8797f0843c88634236b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
One of the most interesting predictions of general relativity is
the existence of null circular orbits outside compact astrophysical objects,
such as black holes and horizonless ultra-compact stars \cite{c1,c2}.
These null circular geodesics provide the way that massless
particles (photons, gravitons) can orbit a central
compact objects, which gives important information about the
highly curved spacetime. At present, the null circular orbits have been widely studied
\cite{ad1}-\cite{r2}.
Recently, it was proved that the null circular orbit
can be used to describe the distribution of exterior matter fields
outside black holes. In spherically symmetric hairy black hole spacetimes,
it was found that the effective radius of hairs must
extend beyond the null circular orbit \cite{s1,s2}.
The authors in \cite{s1} have proposed a nice heuristic physical picture
that the null circular orbit divides matter fields
into two parts (fields below the orbit tending to be swallowed
by black holes and fields above the orbit tending to be radiated away to infinity)
and the self-interaction between these two parts bind
together fields in different regions leading to
the existence of black hole hairs.
And it was further conjectured that the region above the null circular orbit
contains at least half of the total hair mass \cite{s2,s3,s4,s5}.
The main goal of this paper is to study properties of
horizonless ultra-compact spherical stars with null circular orbits.
We shall analytically prove a no short hair theorem that
the effective radius of matter fields must
extend beyond the null circular orbit.
\section{Investigations on effective radius of matter fields}
We consider static horizonless stars with null circular orbits.
In Schwarzschild coordinates, these ultra-compact spherically
symmetric objects take the following form \cite{s1,s2}
\begin{eqnarray}\label{AdSBH}
ds^{2}&=&-g(r)e^{-2\chi(r)}dt^{2}+\frac{dr^{2}}{g(r)}+r^{2}(d\theta^2+sin^{2}\theta d\phi^{2}).
\end{eqnarray}
The metric functions $\chi(r)$ and $g(r)$ only
depend on the Schwarzschild radial coordinate r.
Asymptotical flatness of the spacetime at the infinity
requires that $g(r\rightarrow \infty)=1$ and $\chi(r\rightarrow \infty)=0$.
Near the origin, regularity conditions of the gravity are \cite{b1,b2}
\begin{eqnarray}\label{AdSBH}
g(r\rightarrow 0)=1+O(r^2)~~~~~~and~~~~~~\chi(0)<\infty.
\end{eqnarray}
We take $\rho=-T^{t}_{t}$, $p=T^{r}_{r}$ and
$p_{T}=T^{\theta}_{\theta}=T^{\phi}_{\phi}$,
where $\rho$, $p$ and $p_{T}$ are interpreted as
the energy density, the radial pressure and
the tangential pressure respectively \cite{s2}.
Einstein differential equations $G^{\mu}_{\nu}=8\pi T^{\mu}_{\nu}$ yield
the metric equations
\begin{eqnarray}\label{BHg}
g'=-8\pi r \rho+\frac{1-g}{r},
\end{eqnarray}
\begin{eqnarray}\label{BHg}
\chi'=\frac{-4\pi r (\rho+p)}{g}.
\end{eqnarray}
We label $m(r)$ as the gravitational mass contained within a sphere of radial radius r.
It can be expressed by the integral relation
\begin{eqnarray}\label{AdSBH}
m(r)=\int_{0}^{r}4\pi r'^{2}\rho(r')dr'.
\end{eqnarray}
Considering relations (3) and (5), one
can express the metric function $g(r)$ in
the form \cite{b2}
\begin{eqnarray}\label{BHg}
g=1-\frac{2m(r)}{r}.
\end{eqnarray}
Following approaches in \cite{s2}, we obtain equations of null circular orbits.
As the metric (1) is independent of the time $t$ and angular coordinates $\phi$,
the geodesic trajectories are characterized by conserved energy
E and conserved angular momentum L. And the null circular orbits are
determined by an effective potential
\begin{eqnarray}\label{BHg}
V_{r}=(1-e^{2\chi})E^{2}+g\frac{L^2}{r^2}
\end{eqnarray}
with the characteristic relations
\begin{eqnarray}\label{BHg}
V_{r}=E^{2}~~~~~~and~~~~~~V_{r}'=0.
\end{eqnarray}
Substituting Einstein equations (3) and (4) into
(7) and (8), we deduce the null circular orbit equation
\cite{b1}
\begin{eqnarray}\label{BHg}
N(r_{\gamma})=3g(r_{\gamma})-1-8\pi (r_{\gamma})^2p(r_{\gamma})=0,
\end{eqnarray}
where we have introduced a new function $N(r)=3g(r)-1-8\pi (r)^2p(r)$.
The discrete roots of (9) correspond to the radii of the null circular orbit.
With the relation (2) and the regular condition $p(0)<\infty$, one obtains
\begin{eqnarray}\label{BHg}
N(0)=2.
\end{eqnarray}
We label $r_{\gamma}^{in}$ as the radius of the innermost null circular orbit.
That is to say $r_{\gamma}^{in}$ is the smallest positive root of $N(r)=0$.
Also considering relation (10), one deduces that
\begin{eqnarray}\label{BHg}
N(r)\geqslant0~~~~for~~~~r\in [0,r_{\gamma}^{in}].
\end{eqnarray}
The conservation equation $T^{\mu}_{\nu;\mu}=0$ has only one nontrivial component
\begin{eqnarray}\label{BHg}
T^{\mu}_{r;\mu}=0.
\end{eqnarray}
Substituting equations (3) and (4) into (12),
one gets equation of the pressure
\begin{eqnarray}\label{BHg}
p'(r)=\frac{1}{2rg}[(3g-1-8\pi r^2p)(\rho+p)+2gT-8gp]
\end{eqnarray}
with $T=-\rho+p+2p_{T}$ standing for the trace of the energy momentum tensor.
With a new pressure function $P(r)=r^4p$,
the equation (13) can be expressed as
\begin{eqnarray}\label{BHg}
P'(r)=\frac{r}{2g}[N(\rho+p)+2gT],
\end{eqnarray}
where $N=3g-1-8\pi r^2p$.
The energy condition usually plays a
central role in determining the spacetime geometry.
We assume the dominant energy condition
\begin{eqnarray}\label{BHg}
\rho\geqslant |p|,~|p_{T}|\geqslant 0,
\end{eqnarray}
which means that the energy density $\rho$ is non-negative and
bounds the pressures \cite{s1,s2}. We also assume the
non-negative trace of the energy momentum tensor
expressed as
\begin{eqnarray}\label{BHg}
T=-\rho+p+2p_{\tau}\geqslant 0.
\end{eqnarray}
In this work, we take the non-negative trace conditions \cite{b1,b2,b3,b4}.
A well known example for such horizonless configurations
is the gravitating Einstein-Yang-Mills solitons,
which are characterized by the identity $T=0$ \cite{T}.
In contrast, the usual non-positive trace condition was
imposed in the black hole spacetime \cite{s1,s2}.
From (11) and (14-16), one deduces that
\begin{eqnarray}\label{BHg}
P'(r)\geqslant 0~~~~for~~~~r\in [0,r_{\gamma}^{in}].
\end{eqnarray}
We impose the condition that $\rho$ goes to zero
faster than $r^{-4}$ \cite{s1,s2}. Also considering (15)
and $P(r)=r^4p(r)$, there is the asymptotical behavior
\begin{eqnarray}\label{BHg}
P(r\rightarrow \infty)=0.
\end{eqnarray}
Near the origin, the pressure function $P(r)$ has the asymptotical behavior
\begin{eqnarray}\label{BHg}
P(r\rightarrow 0)=0.
\end{eqnarray}
Relations (18) and (19) imply that $|P(r)|$ must have a local maximum value
at some extremum point $r_{0}$. We can define $r_{m}=r_{0}$ as the effective radii
of matter fields \cite{s1,s2}. According to (17), $P(r)$ is an increasing function of r before it
reaches the innermost null circular orbit.
So the effective radii have a lower bound
\begin{eqnarray}\label{BHg}
r_{m}\geqslant r_{\gamma}^{in},
\end{eqnarray}
which is the same as the no short hair theorem of black holes \cite{s1,s2}.
\section{Conclusions}
We investigated distributions of matter fields
in the background of horizonless ultra-compact spherically symmetric stars.
We assumed the dominant energy condition and the non-negative trace condition.
We defined an effective matter field radius at an extremum point
where the pressure function $|P(r)|$ possesses a local maximum value.
Using analytical methods, we obtained a lower bound
on the effective matter field radius expressed as
$r_{m}\geqslant r_{\gamma}^{in}$ with
$r_{m}$ as the effective matter field radius
and $r_{\gamma}^{in}$ corresponding to the innermost null circular orbit radius.
So we found a no short hair behavior that the
effective matter field radius must extend beyond the
innermost null circular orbit.
\begin{acknowledgments}
This work was supported by the Shandong Provincial
Natural Science Foundation of China under Grant
No. ZR2018QA008. This work was also supported by
a grant from Qufu Normal University of China under
Grant No. xkjjc201906.
\end{acknowledgments}
| cc712b65ed1a1e8bf798894c3057a7bc87899a15 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Visual Simultaneous Localisation and Mapping (VSLAM) and the closely related Visual Odometry (VO) are established topics in the robotics community \citep{2012_Kaess_IJRR,2015_Leutenegger_ijrr,2016_Faessler_JFR,2017_ForsterScaramuzza_tro,2017_ForsterDelaert_tro}.
They are key components of almost all aerial robotic systems \citep{2018_Delmerico_icra} and are used in a host of other robotic applications \citep{2008_Bonin-Font_JIRS} including autonomous driving and underwater robotics.
VSLAM is used to refer to the case of the general SLAM problem where the available measurements are the bearings of landmarks such as provided by image features obtained using a monocular camera.
Visual Odometry is a variant of the VSLAM problem where the solution is optimised for local consistency of the localisation of the system and updates landmark states only for currently visible landmarks.
While the VO community has focused on embedded systems applications, and places a premium on algorithms with low computational and memory requirements \citep{2018_Delmerico_icra}, the VSLAM community has placed a premium on large scale map building, loop closure and accuracy \citep{2016_Stachniss_Handbook,2016_Cadena_TRO}.
As a consequence, many state-of-the-art VO systems use filter based formulations \citep{2007_Mourikis_icra,2015_Bloesch_iros,2017_ForsterScaramuzza_tro,2014_Lynen_iros} in contrast to the full trajectory smoothing and graph based optimization formulation accepted as the community standard for SLAM problems \citep{2016_Cadena_TRO}.
Well engineered trajectory smoothing algorithms using short sliding-windows are still highly competitive algorithms for VO
\citep{2012_Kaess_IJRR,2015_Leutenegger_ijrr,2017_Qin_arxive,2017_ForsterDelaert_tro}.
The non-linear observer community has become interested in the visual SLAM and VO problem in the last few years.
Work by Guerrerio \textit{et al.}~ \citep{2013_Guerreiro_TRO} and Louren{\c{c}}o \textit{et al.}~ \citep{2016_LouGueBatOliSil} propose a non-linear observer for the ``robo-centric'' SLAM problem.
Recent work by Barrau \textit{et al.}~ \citep{2016_Barrau_arxive,2017_Barrau_tac} introduce a symmetry group $\mathbf{SE}_{n+1}(3)$ for the SLAM problem and use this to derive an Invariant Kalman Filter algorithm that overcomes consistency issues that have plagued the EKF algorithms from the classical SLAM era \citep{2011_Dissanayake_ICIIS,2017_Zhang_ral}.
Parallel work by Mahony \textit{et al.}~
\citep{2017_Mahony_cdc} show that this symmetry acts transitively on a principle fibre bundle $\mathcal{M}_n(3)$ which forms a natural geometric state-space for the SLAM problem, overcoming the gauge uncertainty present in the usual pose-map state representation.
However, the symmetry induced by the group $\mathbf{SE}_{n+1}(3)$ applies only to the SLAM configuration state and is not compatible with bearing measurements.
As a consequence, applying the $\mathbf{SE}_{n+1}(3)$ symmetry to the visual SLAM problem still requires linearisation of the output map.
A new symmetry for the VSLAM problem was proposed in \citep{2019_vangoor_cdc_vslam} along with a non-linear observer.
However, in this prior work the observer is closely based on \citep{2016_Hamel_cdc} and is derived in local coordinates and then lifted onto the symmetry group.
It is of interest to consider the case where the observer is designed explicitly using the symmetry structure.
In this paper, we present a highly robust, simple, and computationally cheap non-linear observer for the visual SLAM problem based on the new symmetry of the SLAM configuration space, first presented in \citep{2019_vangoor_cdc_vslam}.
The symmetry is associated with the novel $\mathbf{VSLAM}_n(3)$ Lie-group, which acts on the raw pose-map configuration coordinates and is compatible with the SLAM configuration manifold $\mathcal{M}_n(3)$ \citep{2017_Mahony_cdc}.
The symmetry group structure introduced allows direct application of previous work by Mahony \textit{et al.}~ \citep{RM_2013_Mahony_nolcos} in development of non-linear observers to yield a novel observer for continuous-time VSLAM.
In the design of this observer, it is assumed measurements of the inverse depths of landmarks are available in addition to the bearing measurements.
In practice, the inverse depth may be measured by using optical flow, triangulation, or depth cameras.
The resulting algorithm is fully non-linear; no linearisation is required of the system or output maps.
The approach has the advantage that constant gains can be used in the filter (no Riccati gains need be computed on-line) leading to lower computation and memory requirements.
This additionally leads to a reduction in the number of parameters that need to be tuned in comparison with a standard EKF, making the proposed filter simpler to use in practice.
The inherent symmetry of the approach ensures high levels of robustness and Theorem \ref{th:observer} proves almost global asymptotic and local exponential stability of the error coordinates.
The convergence properties of the filter are demonstrated through a simulation experiment.
Additional simulation experiments compare an EKF with our observer.
These show that our observer achieves comparable mean RMSE to the EKF and has fewer outliers, and operates with a computational complexity that is only linear in the number of landmarks compared to quadratic complexity for the EKF.
\section{Notation}\label{sec:notation}
The special orthogonal group is the set of rotation matrices and is denoted $\mathbf{SO}(3)$ with Lie algebra $\mathfrak{so}(3)$.
The special Euclidean group is the set of rigid body transformations and is denoted $\mathbf{SE}(3)$ with Lie algebra $\mathfrak{se}(3)$.
The group of positive real numbers equipped with multiplication is denoted $\mathbf{MR}$ with Lie algebra $\mathfrak{mr}$.
We use the notation $R_P \in \mathbf{SO}(3)$ and $x_P \in \mathbb{R}^3$ to denote the rotation and translation components of a rigid-body transformation $P \in \mathbf{SE}(3)$ and write
\begin{equation}
P = \begin{pmatrix}
R_P & x_P \\
0 & 1 \end{pmatrix}.
\label{eq:matrix_pose_SE3}
\end{equation}
The pose of a vehicle moving in Euclidean space is written $P \in \mathbf{SE}(3)$.
The kinematics of such a pose frame are written as
\begin{gather}
\dot{P} = PU, \quad \dot{R}_P = R_P \Omega_U^\times, \quad \dot{x}_P = R_P V_U \label{eq:rigid_kinematics_coords}
\end{gather}
where $\Omega_U=(\Omega_1,\Omega_2,\Omega_3)^\top$ and $V_U$ are the body-fixed rotational and translational velocity vectors, respectively, and
\begin{equation}
U = (\Omega_U^\times,V_u)^{\wedge}
:=
\begin{pmatrix}
\Omega_U^\times & V_U \\ 0 & 0
\end{pmatrix}, \quad \Omega^\times_U=
\begin{pmatrix} 0 & -\Omega_3 & \Omega_2 \\
\Omega_3 & 0 & -\Omega_1 \\
-\Omega_2 & \Omega_1 & 0
\end{pmatrix}.
\label{eq:wedge}
\end{equation}
One has that $\Omega^\times_U w = \Omega_U \times w$ for any $w \in \mathbb{R}^3$ where $\times$ refers to the vector (cross) product.
For a unit vector $y \in \mathrm{S}^2 \subset \mathbb{R}^3$, the projector is given by
\begin{align} \label{eq:projector}
\Pi_y := I_3 - yy^\top,
\end{align}
and has the property $\Pi_y = - y^\times y^\times$.
\section{Problem formulation}\label{sec:problem_formulation}
The \emph{total space} coordinates for the SLAM problem are defined with respect to an unknown fixed but arbitrary reference $\mbox{$\{0\}$}$.
Let $P \in \mathbf{SE}(3)$ represent the body-fixed frame coordinates of the robot with respect to this reference frame.
Let
\[
p_i \in \mathbb{R}^3, \quad\quad i = 1, \ldots, n,
\]
be sparse points in the environment expressed with respect to the reference frame $\mbox{$\{0\}$}$.
The \emph{total space} of the SLAM problem is the product space
\begin{align}
\mathcal{T}_n(3) = \mathbf{SE}(3) \times \mathbb{R}^3 \times \cdots \times \mathbb{R}^3
\label{eq:tot_T}
\end{align}
made up of these raw coordinates $\Xi := (P, p_1, \ldots, p_n)$.
The bearing of a point $p_i$ co-located with the robot pose centre $x_P$ is undefined, so the VSLAM problem can only be considered on the \emph{reduced total space}
\begin{align}
\mr{\mathcal{T}}_n(3) = \left\{ (P, p_i) \in \mathcal{T}_n(3) \; \vline \; (\forall i) \; p_i \neq x_P \right\}.
\label{eq:reduced_total_space}
\end{align}
Moreover, since all the measurements of the VSLAM problem considered are made in the body-fixed frame the solution is only well defined up to an $\mathbf{SE}(3)$ gauge transformation \citep{2001_Kanatani_TIT}.
This property can be expressed as an invariance of the problem formulation and leads to the quotient structure of the SLAM manifold proposed in \citep{2017_Mahony_cdc}.
To keep the derivation simple and more accessible, in the present paper we will define the group actions and derive the observer on the reduced total space.
The measurements considered are spherical coordinates $y_i$ of body-fixed frame observations of points in the environment, which in practice may be obtained from a calibrated monocular camera.
Additionally, in this analysis, we assume inverse depth estimates $z_i$ are also available.
That is, for a given robot pose $P$ and environment point $p_i$,
\begin{subequations}\label{eq:h_output}
\begin{align}
y_i & := \frac{R_P^\top(p_i-x_P)}{|p_i-x_P|}, \label{eq:h_output_y} \\
z_i & := |p_i - x_P|^{-1}. \label{eq:h_output_z}
\end{align}
\end{subequations}
The combined output space is $\mathcal{N}_n(3) = (S^2 \times \mathbb{R}^+) \times \cdots \times (S^2 \times \mathbb{R}^+)$ and we write $(y, z) = h(\Xi) = (h_1(\Xi),\dots,h_n(\Xi)) = ((y_1, z_1), \dots, (y_n, z_n))$, where appropriate.
Let $U \in \mathfrak{se}(3)$ denote the velocity of the robot.
Assume that the environment points $p_i$ being observed are static, and thus do not have a velocity.
The tangent space of $\mathcal{T}_n(3)$ at a point $\Xi = (P, p_1, \ldots, p_n)$ can be identified with the matrix subspace
\[
(PU, u_1,...,u_n), \quad U \in \mathfrak{se}(3), \quad u_1,...,u_n \in \mathbb{R}^3.
\]
The system kinematics can then be written as
\begin{equation}
\frac{\td}{\td t} ( P , p_1, \ldots, p_n ) = ( P U, 0, \ldots, 0 ).
\label{eq:tot_kinematics}
\end{equation}
We assume that the robot velocity $U \in \mathfrak{se}(3)$ is measured.
We will also measure the optical flow of each landmark
\[
\phi_i \in T_{y_i} \mathrm{S}^2 \subset \mathbb{R}^3
\]
by numerically differentiating the coordinates of $y_i$ between consecutive measurements.
Here, we express $\phi_i \in T_{y_i} \mathrm{S}^2$ using the coordinates obtained by embedding $\mathrm{S}^2 \subset \mathbb{R}^3$.
Define a measurement velocity space
\begin{align}
\mathbb{V} = \mathfrak{se}(3) \times \mathbb{R}^3 \times \cdots \times \mathbb{R}^3
\label{eq:vecV}
\end{align}
with elements $(U, \phi_1, \ldots, \phi_n)$.
\section{Symmetry of the VSLAM problem}\label{sec:symmetry}
The symmetry group of the VSLAM for $n$ landmarks in Euclidean 3-space with separate bearing and range measurements problem is the \emph{visual SLAM group} $\mathbf{VSLAM}_n(3)$ first described in \citep{2019_vangoor_cdc_vslam}.
However, in this paper the VSLAM group and its actions are presented in a different form to \citep{2019_vangoor_cdc_vslam}.
In the following, we will write
$(Q_A,a)_1$ instead of the more formal $({\left(Q_A\right)}_1,a_1)$ and sometimes write $(Q_A,a)_i$ to represent the tuple $(Q_A,a)_1,\dots,(Q_A,a)_n$.
Similarly, we will sometimes write $(P,p_i)$ instead of
$(P,p_1,\dots,p_n)$.
The VSLAM group \citep{2019_vangoor_cdc_vslam} may be written
\begin{multline*}
\mathbf{VSLAM}_n(3) = \{ (A,(Q_A,a)_1, \ldots,(Q_A,a)_n) \;|\; \\
A \in \mathbf{SE}(3),\; {\left(Q_A\right)}_i \in \mathbf{SO}(3),\; a_i\in \mathbf{MR}, \; i = 1, \ldots, n\}.
\end{multline*}
\begin{lemma}
The set $\mathbf{VSLAM}_n(3)$ is a Lie group, defined as
$$ \mathbf{VSLAM}_n(3) := \mathbf{SE}(3) \times (\mathbf{SO}(3)\times \mathbf{MR})^n.$$
\end{lemma}
The visual SLAM group acts as a symmetry group on the reduced total space $\mr{\mathcal{T}}_n(3)$.
\begin{lemma}\label{lem:Upsilon_action}
The mapping $\Upsilon : \mathbf{VSLAM}_n(3) \times \mr{\mathcal{T}}_n(3) \rightarrow \mr{\mathcal{T}}_n(3)$ defined by
\begin{align}
\Upsilon((A,&(Q_A,a)_i),(P,p_i)) \notag \\
&= (PA,(a^{-1}R_{PA}Q_A^\top R_P^\top(p-x_P) +x_{PA})_i), \label{eq:Upsilon}
\end{align}
is a right group action of $\mathbf{VSLAM}_n(3)$ on $\mr{\mathcal{T}}_n(3)$.
\end{lemma}
The group action for the robot pose is rigid-body transformation.
The group action for environment points is considerably more subtle and can be understood conceptually as a sequence of operations: firstly, the reference frame coordinates of an environment point are written in the body-fixed frame, this point is then rotated by $Q_A^\top$ and then scaled by $a^{-1}$,
before these body-fixed frame coordinates are rewritten in the inertial frame using the \emph{new body-fixed frame} reference.
\begin{figure}[ht]
\begin{centering}
\includegraphics[width = 0.8\linewidth]{figures/group_action_updated.eps}
\caption{Group action $\Upsilon((A,(Q_A,a)_i),(P,p_1, \ldots, p_n))$.
The pose $P \mapsto P A$, that is the tip point of the pose is updated by the correction $A \in \mathbf{SE}(3)$.
The body fixed-frame environment points $R_A^\top (p_i - x_P)$ are rotated by $Q_A^\top$ and scaled by $a^{-1}$ in the body-fixed frame before \emph{transforming with the robot pose} to a new point $p_i'$ which is rewritten in the inertial frame.
}
\label{fig:symmetry_action}
\end{centering}
\end{figure}
A key property of the proposed structure is that there is a compatible group operation on the output $\mathcal{N}_n(3)$ of the system.
\begin{lemma}\label{lem:rho}
The action $\rho : \mathbf{VSLAM}_n(3) \times \mathcal{N}_n(3) \rightarrow \mathcal{N}_n(3)$ defined by
\begin{align}\label{eq:rho}
\rho((A,(Q_A,a)_i),(y, z)_i)& := ((Q_A^\top y, az)_i)
\end{align}
is a transitive right action on $\mathcal{N}_n(3)$.
Furthermore, one has
\begin{align*}
\rho((A,(Q_A,a)_i), h(\Xi)) & = h(\Upsilon((A, (Q_A,a)_i),\Xi))
\end{align*}
where $h$ is given by \eqref{eq:h_output}.
That is, $h$ is equivariant with respect to the actions $\Upsilon$ and $\rho$.
\end{lemma}
\subsection{Lift of the SLAM kinematics}
A key aspect of the proposed approach is that the symmetry group $\mathbf{VSLAM}_n(3)$ and the reduced total space $\mr{\mathcal{T}}_n(3)$ are quite different spaces.
The difference is particularly clear in studying the structure of the lifted kinematics on the $\mathbf{VSLAM}_n(3)$ group.
The Lie-algebra of $\mathbf{VSLAM}_n(3)$ is
\begin{equation*}
\mathfrak{vslam}_n(3) = \mathfrak{se}(3) \times (\mathfrak{so}(3) \times \mathfrak{mr}) \times \dots \times (\mathfrak{so}(3) \times \mathfrak{mr}).
\end{equation*}
We write $(U,(W,w)_i) \in \mathfrak{vslam}_n(3)$ with $U \in \mathfrak{se}(3)$, $W_i \in \mathfrak{so}(3)$ and $w_i \in \mathfrak{r}$, where
\begin{align*}
U=\begin{pmatrix} \Omega_U^\times & V_U \\ 0 & 0 \end{pmatrix}.
\end{align*}
In order to implement an observer on the VSLAM group, it is necessary to lift the velocity measurements $(U,\phi_i) \in \mathbb{V}$ \eqref{eq:vecV} to elements $(U,(W,u)_i) \in \mathfrak{vslam}_n(3)$ such that the resulting group velocity is compatible with the system kinematics.
That is, an algebraic map $\lambda : \mr{\mathcal{T}}_n(3) \times \mathbb{V} \rightarrow \mathfrak{vslam}_n(3)$ is required such that
\begin{align} \label{eq:lift-condition}
\mathrm{d} \Upsilon_{(P,p_i)} \lambda ((P,p_i),(U,\phi_i)) = (PU, 0, \ldots, 0).
\end{align}
\begin{proposition}
The map $\lambda :\mr{\mathcal{T}}_n(3) \times \mathbb{V} \rightarrow \mathfrak{vslam}_n(3)$ defined by
\begin{align} \label{eq:lambda}
\lambda((P,p_i),(U,\dot{y}_i)) := \left(U ,\left((\phi \times y)^\times, z y^\top V_U\right)_i \right),
\end{align}
where $(y,z)_i = h((P,p_i))$ satisfies the lift condition \eqref{eq:lift-condition}.
\end{proposition}
\begin{proof}
Under the static landmark assumption $p_i = 0$, the optic flow $\phi_i = \dot{y_i}$ is given by
\begin{align} \label{eq:optical-flow}
\phi_i &= -\Omega_U^\times y_i - z_i (I-y_i y_i^\top) V_U.
\end{align}
Let $(\lambda_A, (\lambda_Q^\times, \lambda_a)_i) := \lambda((P,p_i),(U,\phi_i))$.
Evaluating the left-hand side of \eqref{eq:lift-condition}, one has
\begin{align*}
\mathrm{d} &\Upsilon_{(P,p_i)} \lambda ((P,p_i),(U,\phi_i)) \\
&= \mathrm{D} \Upsilon_{(P,p_i)}(\id) [ (\lambda_A, (\lambda_Q^\times, \lambda_a)_i) ], \\
&= (P\lambda_A, (- \lambda_a (p-x_P) + R_P \Omega_{\lambda_A}^\times R_P^\top (p - x_P) \\
&\hspace{2cm} - R_P \lambda_Q^\times R_P^\top (p - x_P) + R_P V_U )_i).
\end{align*}
This expression may be written in terms of $(y,z)_i$ as follows.
\begin{align*}
\mathrm{d} &\Upsilon_{(P,p_i)} \lambda ((P,p_i),(U,\phi_i)) \\
&= (P\lambda_P, (- \lambda_a R_P z^{-1} y + R_P (\Omega_{\lambda_A}^\times - \lambda_Q^\times) z^{-1} y + R_P V_U )_i).
\end{align*}
Multiply the landmark velocity terms by $R_P^\top$ and substitute in the values for $\lambda$ to obtain
\begin{align*}
- \lambda_a & z^{-1} y + (\Omega_{\lambda_A}^\times - \lambda_Q^\times) z^{-1} y + V_U \\
&= -yy^\top V + z^{-1} (\Omega_U - (\phi \times y)^\times) y + V_U, \\
&= z^{-1} \Omega_U^\times y - z^{-1} y^\times y^\times \phi + (I-yy^\top) V , \\
&= z^{-1} \Omega_U^\times y - z^{-1} (\Omega_U^\times y_i + z_i (I-y_i y_i^\top) V_U) + (I-yy^\top) V , \\
&= 0.
\end{align*}
Hence $\mathrm{d} \Upsilon_{(P,p_i)} \lambda ((P,p_i),(U,\phi_i)) = (P\lambda_A, R_P 0) = (PU, 0)$ as required.
\end{proof}
The lifted velocity $(U,(W,u)_i) = \lambda((P,p_i), (U, \phi_i))$ induces kinematics on the symmetry group that project down to the state space trajectory.
Since the group action is not free, the stabiliser of $\Upsilon$ is non-trivial, and there are directions in $\mathfrak{vslam}_n(3)$, in particular $y_i^\times \in \mathfrak{so}(3)$, that are not constrained by the lift requirement \eqref{eq:lift-condition}.
The lift in direction $y_i^\times \in \mathfrak{so}(3)$ is chosen to be zero without loss of generality.
The lift $\lambda$ will enable us to go on and apply the observer design methodology developed in \citep{RM_2013_Mahony_nolcos}.
\section{Observer design}\label{sec:observer}
We approach the observer design by considering the lifted kinematics of the system on the symmetry group and designing the observer on $\mathbf{VSLAM}_n(3)$.
Let $\Xi(t) = (P(t),p_i(t)) \in \mathcal{T}_n(3) $ be the `true' configuration of the SLAM problem, noting that $\Xi(t)$ is defined relative to some arbitrary reference $\mbox{$\{0\}$}$.
Let $X(t) = (A(t),(Q(t) ,a(t))_i) \in \mathbf{VSLAM}_n(3)$ and define the \emph{lifted kinematics} \citep{RM_2013_Mahony_nolcos}
\begin{align}
\frac{\td}{\td t} X(t)&= X \lambda(\Upsilon(X, \Xi(0)), (U, \phi_i)) \notag \\
\quad X(0) &= \id.
\label{eq:ddtX}
\end{align}
Equation \eqref{eq:ddtX} evolves on the VSLAM group where $(U, \phi_i) \in \mathbb{V}$ are the measured velocities and $\lambda$ is the lift function \eqref{eq:lambda}.
Choose an arbitrary origin configuration
\[
\mr{\Xi}= (\mr{P}, \mr{p}_i) \in \mathcal{T}_n(3).
\]
If the initial condition $X(0) \in \mathbf{VSLAM}_n(3)$ of the lifted kinematics satisfies $\Upsilon(X(0),\mr{\Xi}) = \Xi(0)$ then \eqref{eq:ddtX} induces a trajectory that satisfies
\[
\Xi(t) = \Upsilon( X(t) , \mr{\Xi}) \in \mathcal{T}_n(3)
\]
for all time \citep{RM_2013_Mahony_nolcos}.
Let the observer be defined as
\[
\hat{X} = (\hat{A},(\hat{Q},\hat{a})_i) \in \mathbf{VSLAM}_n(3).
\]
The lifted kinematics \eqref{eq:ddtX} provide the internal model for the observer design.
That is, the kinematics of the observer are given by
\begin{align}
\frac{\td}{\td t} \hat{X}(t)&= \hat{X} \lambda(\Upsilon(X, \mr{\Xi}), (U, \phi_i)) - \Delta \hat{X} \notag, \\
\hat{X}(0) &= \id,
\label{eq:ObserXi}
\end{align}
where $\Delta \in \mathfrak{vslam}_n(3)$ is an innovation term to be assigned.
Note that $\lambda(\Upsilon(X, \mr{\Xi}), (U, \phi_i))$ is shown in \eqref{eq:lambda} to depend only on the measured quantities $y_i, z_i, U, \phi_i$, and therefore can be implemented in the observer kinematics \eqref{eq:ObserXi}.
The configuration estimate generated by the observer is given by
\[
\hat{\Xi} = (\hat{P}, \hat{p}_i)
= \Upsilon( \hat{X}, \mr{\Xi}) \in \mr{\mathcal{T}}_n(3)
\]
given the reference $\mr{\Xi} \in \mathcal{T}_n(3)$.
\begin{theorem}\label{th:observer}
Consider the kinematics \eqref{eq:ddtX} evolving on $\mathbf{VSLAM}_n(3)$ along with bounded outputs ${\bf y}=\{(y,z)_i\}=h(\Xi(t)) \in \mathcal{N}_n(3)$ given by \eqref{eq:h_output}.
Fix an arbitrary origin configuration $\mr{\Xi}= (\mr{P}, \mr{p}_i)$ and define the output error ${\bf e}=\{(e_y,e_z)_i\}$ as
\begin{align}
{\bf e}=\rho(\hat{X}^{-1},\{(y,z)_i\})=\rho(E^{-1}, \{(\mr{y},\mr{z})_i\}) \in \mathcal{N}_n(3). \label{e}
\end{align}
where $\{(\mr{y},\mr{z})_i\} = h(\mr{\Xi})$ and $E=\hat{X} X^{-1}$.
Consider the observer defined in \eqref{eq:ObserXi} and choose the
innovation term $\Delta := (\Delta_A, (\Delta_Q, \Delta_a)^i)$ as follows:
\begin{align}
\Delta_Q^i & := -k_{Q_i}( e_{y_i} \times \mr{y}_i)^\times \label{eq:Delta_Q}\\
\Delta_a^i & := -k_{a_i} \frac{e_{z_i} - \mr{z}_i }{e_{z_i}} \label{eq:Delta_a} \\
\Delta_A &:= -k_{A} \Ad_{\hat{A}} \left( \begin{matrix}
\Omega_\Delta^\times & V_\Delta \\ 0 & 0
\end{matrix} \right), \label{eq:Delta_A} \\
\begin{pmatrix} \Omega_{\Delta} \\ V_\Delta \end{pmatrix}
&:= \left( \sum_{i=1}^n \begin{pmatrix} \Pi_{\hat{y}_i} & \hat{z}_i \hat{y}_i^\times \\ -\hat{z}_i \hat{y}_i^\times & \hat{z}_i^2 \Pi_{\hat{y}_i} \end{pmatrix} \right)^{-1}
\left( \sum_{i=1}^n \begin{pmatrix}
-\hat{y}_i^\times \phi_i \\ -\hat{z}_i \phi_i
\end{pmatrix} \right) - \begin{pmatrix}
\Omega_U \\ V_U
\end{pmatrix}, \notag
\end{align}
where the gains $k_{a_i}, k_{Q_i}$ and $k_{A}$ are positive scalars (for $i = 1, \ldots, n$), and the matrix inverse in the definition of $\Delta_{A}$ is assumed to be well-defined
Then the configuration estimate $\hat{\Xi}(t) = \Upsilon(\hat{X}(t),\mr{\Xi})$ converges almost globally asymptotically and locally exponentially to the true state $\Xi(t)=\Upsilon(X(t),\mr{\Xi})$ up to a possibly time-varying element in $\mathbf{SE}(3)$.
\end{theorem}
\begin{proof}
Let $X(t) \in \mathbf{VSLAM}_n(3)$ satisfy the lifted kinematics \eqref{eq:ddtX} with $\Upsilon(X(0),\mr{\Xi}) = \Xi(0)$.
It follows that $\Xi(t)=\Upsilon(X(t),\mr{\Xi})$ \citep{RM_2013_Mahony_nolcos}.
Define
\begin{equation}
E=\hat{X} X^{-1} =: (\tilde{A}, (\tilde{Q},\tilde{a})_i) \in \mathbf{VSLAM}_n(3),
\label{eq:E}
\end{equation}
with $\tilde{A}:=\hat{A}A^{-1}$, $(\tilde{Q},\tilde{a})_i=\left( \hat{Q} Q^\top , \hat{a} a^{-1} \right)_i$.
Using \eqref{eq:ddtX} and \eqref{eq:ObserXi}, it is straightforward to verify that
\begin{equation}
\dot E = (-\Delta_A \tilde{A}, (-\Delta_Q \tilde{Q}, -\Delta_a \tilde{a})_i).
\label{dotE}
\end{equation}
Using the fact that $E^{-1} = (\tilde{A}^{-1},(\tilde{Q}^\top, \tilde{a}^{-1})_i)$ then each element of equation \eqref{e} becomes
\begin{align}
(e_{y_i}, e_{z_i})&=\left (\tilde{Q} \mr{y}_i,\tilde{a}^{-1} \mr{z}_i \right)\label{eq:e}.
\end{align}
Based on \eqref{dotE}, the error kinematics satisfy
\begin{align} \label{eq:error_kinematics}
(\dot{e}_{y_i}, \dot{e}_{z_i})&= \left (-\Delta_Q^i e_{y_i}, \Delta_a^i e_{z_i}\right ).
\end{align}
We first prove almost-global asymptotic and local exponential stability of the equilibrium $(e_{y_i}, e_{z_i})=(\mr{y}_i, \mr{z}_i)$ for the error kinematics \eqref{eq:error_kinematics}.
Consider the following candidate (positive definite) Lyapunov function $\Lyap: \mathcal{N}_n(3) \to \mathbb{R}^+$,
\begin{equation}
{\cal L}=\frac{1}{2}\sum_{i=1}^{n} \left (\left|e_{y_i}-\mr{y}_i \right|^2+\left (e_{z_i}-\mr{z}_i\right)^2 \right).
\label{eq:lyap}
\end{equation}
Differentiating ${\cal L}$ and using \eqref{eq:Delta_Q} and \eqref{eq:Delta_a}, one gets:
\begin{align*}
\dot{\cal L} &=\sum_{i=1}^{n} \left (\left (e_{y_i}-\mr{y}_i \right )^\top \dot{e}_{y_i}+\left (e_{z_i}-\mr{z}_i\right)\dot{e}_{z_i} \right ) , \\
&= \sum_{i=1}^{n} \left( - \left( e_{y_i}-\mr{y}_i \right)^\top \Delta_Q^i e_{y_i} + \left( e_{z_i}-\mr{z}_i \right) \Delta_a^i e_{z_i} \right ) , \\
&=-\sum_{i=1}^{n} \left ( k_{y_i} \left | e_{y_i} \times \mr{y}_i \right |^2 + k_{z_i} \left (e_{z_i}-\mr{z}_i\right)^2 \right ).
\end{align*}
The time derivative of the Lyapunov function is negative definite and equal to zero when $e_{y_i}=\pm\mr{y}_i$ and $e_{z_i}=\mr{z}_i$.
Direct application of Lyapunov's theorem ensures that the equilibrium $(e_{y_i},e_{z_i})=(\mr{y}_i, \mr{z}_i)$ is almost-globally \emph{asymptotically} stable\footnote{It is straightforward to verify that the equilibrium point $e_{y_i}=-\mr{y}_i$ is unstable.}.
To prove local exponential stability of the observer it suffices to split the Lyapunov function into two parts
\[
{\cal L}= {\cal L}_y +{\cal L}_z
\]
${\cal L}_z =\frac{1}{2}\sum_{i=1}^{n} \left (e_{z_i}-\mr{z}_i\right)^2$
and verify that $\dot{\cal L}_z \leq -2 \min(k_{z_i}) {\cal L}_z$, with ${\cal L}_z$ converging exponentially to zero.
Consider ${\cal L}_y=\sum_{i=1}^{n} {\cal L}_{y_i}$ with ${\cal L}_{y_i}=\frac{1}{2} \left|e_{y_i}-\mr{y}_i \right|^2$.
If there exists a positive number $\epsilon$ such that ${\cal L}_{y_i} \leq 2 - \epsilon$, that is $e_{y_i}(0)$ is not in the opposite direction of $\mr{y}(0)$, for all $i=\{1,\ldots,n\}$, then
\[
\dot{\cal L}_{y_i} \leq -2 \min(k_{y_i}) \epsilon {\cal L}_{y_i}
\]
This demonstrates local exponential stability (in a large domain) of the equilibrium
$(e_{y_i},e_{z_i})=(\mr{y}_i, \mr{z}_i)$
In the limit, at the stable equilibrium point $(e_{y_i},e_{z_i})=(\mr{y}_i, \mr{z}_i)$), \eqref{eq:e} implies that
\begin{align*}
(\hat{y},\hat{z})_i =
\left(Q_{\hat{A}}^\top \mr{y},\hat{a} \mr{z} \right)_i =
\left(Q_{\hat{A}}^\top e_{y},\hat{a} e_{z} \right)_i =
(y, z)_i,
\end{align*}
for all $i=1,\ldots,n$.
This in turn implies
\begin{align*}
\rho(\hat{X}, (\mr{y},\mr{z})_i) & = h(\Upsilon(\hat{X}, \mr{\Xi})) = h ( \Upsilon(X, \mr{\Xi})) \\
& = \rho(X, (\mr{y}, \mr{z})_i)
\end{align*}
Regarding just the central equality, and noting that $h$ only preserves the relative pose on $\mr{\mathcal{T}}_n(3)$, it follows that
\[
\Xi = \Upsilon(X, \mr{\Xi}) = \Upsilon(\hat{X}, \mr{\Xi}) \cong \hat{\Xi}
\]
The symbol $\cong$ indicates that $\Xi = \hat{\Xi}$ up to the possibly time-varying gauge transformation $\tilde{A}$.
That is
$$(\hat{P},\hat{p}_i) = (\tilde{A}^{-1} P, \tilde{A}^{-1} (p_i)) = (\tilde{A}^{-1} P, R_{\tilde{A}}^\top(p_i - x_{\tilde{A}})).$$
This concludes the proof of the almost-global asymptotic and local exponential stability.
\end{proof}
\begin{remark}
Observe that the output error ${\bf e}$ is independent of the $\mathbf{SE}(3)$ innovation $\Delta_A$ and the primary stability analysis in Theorem \ref{th:observer} is undertaken on the output space, not the state-space.
This is a key property of the $\mathbf{VSLAM}_n(3)$ symmetry and is intrinsic to the invariance of the underlying SLAM problem discussed in Section \S \ref{sec:problem_formulation}.
The particular choice of innovation $\Delta_A$ in \eqref{eq:Delta_A} minimizes the least squares drift in the visual odometry error as observed from direct measurements of landmark coordinates and optical flow.
This is only one of a family of possible choices (for example, \cite{2017_Mahony_cdc}), however, further analysis of this question is beyond the scope of the present paper.
\end{remark}
\section{Simulation Results}
The first simulation experiment was conducted to verify the observer design in Theorem \ref{th:observer}.
A robot is simulated to move in a circle with velocity $V_U = (0.1,0,0)$ m/s, $\Omega_U = (0,0,0.02\pi)$ rad/s on the ground, with 10 landmarks uniformly distributed in a 0.5-1 m band around the robot's path.
The reference configuration $\mr{\xi}$ of the observer is randomly set, and the observer $\hat{X}$ is initialised to the identity group element.
All landmarks are assumed to be measured at all times, and no noise is added to the system.
The gains of the observer are set to $k_{Q_i} = 0.05$, $k_{a_i} = 0.02$, $k_{A} = 0.03$ for all $i=1,...,10$.
The observer equations are implemented with Euler integration using a time step of 0.5 s.
Figure \ref{fig:simple_sim} shows the evolution of the Lyapunov function \eqref{eq:lyap} components for each landmark over 100 s.
The bearing storage refers to the component $l_y^i := \frac{1}{2}\vert e_{y_i} - \mr{y}_i \vert^2$ and the inverse depth storage refers to the component $l_z^i := \frac{1}{2}\vert e_{z_i} - \mr{z}_i \vert^2$ for each landmark index $i$.
The top two plots show the value of these functions for each landmark, and the bottom two plots show the log value for each landmark.
The plots clearly show the almost-global asymptotic and local exponential convergence of the observer's error system.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.85\linewidth]{figures/simple_sim_results.eps}
\caption{The evolution of the individual components of the Lyapunov function \eqref{eq:lyap} for 10 landmarks over time.}
\label{fig:simple_sim}
\end{figure}
Additional simulations were carried out to compare the non-linear observer proposed in Theorem \ref{th:observer} with an Extended Kalman Filter (EKF).
A robot is simulated to move in a circle with velocity $V_U = (0.1,0,0)$ m/s, $\Omega_U = (0,0,0.02\pi)$ rad/s on the ground, with $n$ landmarks uniformly distributed in a 0.5-1 m band around the robot's path.
The robot is modelled to have a sensor range of 1 m.
The reference configuration $\mr{\xi}$ is initialised without any landmarks, and the observer group element is initialised to identity.
When landmarks are first seen, their inertial frame position is computed using the observer's current position estimate, and the reference configuration is augmented with this value.
When landmarks are not within the sensing range, the observer equations cannot be used, and the current observer estimate of the landmark position is fixed until the landmark is next seen.
All noise added to the input velocities and output measurements is drawn from zero-mean Gaussian distributions.
The linear velocity noise has variance $0.2$, the angular velocity noise has variance $0.1$, the optical flow noise has variance $0.02$, the bearing measurement noise has variance $0.01$, and the inverse depth measurement noise has variance $0.4$.
The EKF is implemented with the system equation $\eqref{eq:tot_kinematics}$, and the measurement equations $\eqref{eq:h_output}$.
The gains of the observer are set to $k_{Q_i} = 0.25$, $k_{a_i} = 0.1$, $k_{A} = 0.1$ for all $i=1,...,n$, and the observer equations were implemented using Euler integration with a time step of $0.5$ s.
\begin{figure}[!htb]
\centering
\begin{subfigure}[t]{0.45\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/sim_rmse_comparison.eps}
\caption{Boxplot of the RMSE of EKF and our observer on 50 landmarks over 500 trials.}
\label{fig:sim_rmse_comparison}
\end{subfigure}
\hspace{0.05\linewidth}
\begin{subfigure}[t]{0.45\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/sim_time_comparison.eps}
\caption{Mean computation time of EKF and our observer on 10-400 landmarks over 500 trials.}
\label{fig:sim_time_comparison}
\end{subfigure}
\caption{Results of the simulation experiments comparing an EKF with the proposed observer.}
\end{figure}
Figure \ref{fig:sim_rmse_comparison} compares the statistics of the RMSE of the EKF and our observer for $n=50$ landmarks after 100 s over 500 trials.
While the EKF has a slightly lower mean RMSE, there are also more outliers due to linearisations errors.
Figure \ref{fig:sim_time_comparison} shows the mean computation time of the EKF and our observer for an increasing value of $n$ between 10 and 400 landmarks over 500 trials per number of landmarks.
While the processing time depends on the implementation of the EKF and of our observer, the figure clearly illustrates the quadratic complexity of the EKF and the linear complexity of our observer.
\section{Conclusion}
\label{sec:conclusion}
This paper proposes a new symmetry for visual SLAM and VIO problems.
This geometry is exploited to develop a visual SLAM observer and provide an almost global asymptotic and local exponential stability proof.
The authors believe that the inherent simplicity and robustness of the proposed approach makes it useful as a tool for embedded robotics applications.
\section*{Acknowledgement}
This research was supported by the Australian Research Council
through the ``Australian Centre of Excellence for Robotic Vision'' CE140100016.
\bibliographystyle{plainnat}
| f914526ef479bf51a80b6d50819ed198674c478e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and results}\label{intro}
Hankel operators on Bergman spaces on bounded pseudoconvex domains with
symbols that are continuous on the closure of the domain are compact when
the $\overline{\partial}$--Neumann operator on the domain is compact
(\cite{FuStraube01, HaslingerBook, StraubeBook}). It is natural to ask what
happens when the $\overline{\partial}$--Neumann operator is not compact. Must
there necessarily be noncompact Hankel operators (with, say, symbol continuous
on the closure of the domain)? The answer is known only in cases where
compactness (or lack thereof) of the $\overline{\partial}$--Neumann operator
is understood : when the domain is convex, bounded, in $\mathbb{C}^{n}$
(\cite{FuStraube01}, Remark 2), or when it is a smooth bounded pseudoconvex
Hartogs domain in $\mathbb{C}^{2}$ (\cite{SahutogluZeytuncu17}). In these cases,
compactness of all Hankel operators with symbols that are smooth on the closure
implies compactness of the $\overline{\partial}$--Neumann operator. In general,
the question is open\footnote{For the case $q=0$ and $N_{1}$. The answer is
known to be affirmative for Hankel operators on $(0,q)$--forms with $1\leq q\leq (n-1)$
(\cite{CelikSahutoglu14}); the relevant $\overline{\partial}$--Neumann operator
is then $N_{q+1}$. Here, pseudoconvexity is important; see
\cite{CelikSahutoglu12}.}; having an answer would be very interesting.
Alternatively, one can consider a particular situation where the
$\overline{\partial}$--Neumann operator is not compact and ask for necessary
and/or sufficient conditions on a symbol that will imply compactness of
the associated Hankel operator. Given that two symbols whose difference extends
continuously to the boundary as zero yield Hankel operators that agree modulo a
compact operator, one would in particular like to understand how the interaction
of the symbol with the boundary affects compactness of the associated Hankel
operator. It is this question that we are interested in in the current paper.
Such a study was initiated in \cite{CuckovicSahutoglu09}. For symbols smooth on
the closure of a smooth bounded pseudoconvex domain in $\mathbb{C}^{n}$, the
authors show, under the condition that the rank of the Levi form is at least
$(n-2)$, that compactness of the Hankel operator requires that the symbol is
holomorphic along analytic discs in the boundary. When the domain is also
convex, they can dispense with the condition on the Levi form.\footnote{In both
these cases, compactness of the $\overline{\partial}$--Neumann operator was known
to fail when there are discs in the boundary; \cite[Theorem 1]{SahutogluStraube06},
\cite[Theorem 1.1]{FuStraube98}.} Moreover, for (smooth) convex domains in
$\mathbb{C}^{2}$, holomorphy of the symbol along analytic discs in the boundary
is also sufficient for compactness of the Hankel operator. Further contributions
are in \cite{Le10,CuckovicSahutoglu17,ClosSahutoglu18,ClosAccepted};
we refer the reader to the introduction in \cite{ClosCelikSahutoglu18} for a
summary. The latter authors significantly reduce the regularity requirements
on both the domain (Lipschitz in $\mathbb{C}^{2}$ or convex in $\mathbb{C}^{n}$)
and the symbol (in $C(\overline{\Omega})$) that is required to infer holomorphicity
of the symbol along analytic discs in the boundary from compactness of the
Hankel operator. Because compactness of a Hankel operator localizes
(\cite{ClosCelikSahutoglu18}; see also \cite{Sahutoglu12}), the latter result carries
over to locally convexifiable domains. In \cite{CelikSahutogluStraube20}, the
authors, among other things, extend the result from \cite{ClosCelikSahutoglu18}
on convex domains in $\C^n$ to Hankel operators on ($\overline{\partial}$--closed)
$(0,q)$--forms with $0\leq q\leq (n-1)$, but with symbol assumed in $C^{1}$ of
the closure. When the (convex) domain is smooth and satisfies so called maximal
estimates\footnote{This condition is equivalent to a comparable eigenvalues
condition for the Levi form of the boundary, see the discussion in
\cite{CelikSahutogluStraube20} and their references.}, holomorphicity of the
symbol along $(n-1)$--dimensional analytic (equivalently: affine) polydiscs in
the boundary suffices for compactness of the associated Hankel
operator.\footnote{As noted in \cite{CelikSahutogluStraube20}, these two results
combined imply that a convex domain that satisfies maximal estimates for
$(0,q)$--forms does not have any analytic varieties of dimension $\geq q$ in its
boundary except ones in top dimension $(n-1)$ (and their subvarieties). It would
be desirable to have a direct proof for this fact.} Finally we
mention the recent \cite{ChengJinWang}, where the authors consider Hankel
operators with form symbols (replacing multiplication with the wedge product)
and prove many of the results discussed above for this situation.
Our first result (Theorem \ref{ThmNonCompact}) reduces the
regularity of the symbol in Theorem 1 in \cite{CelikSahutogluStraube20} to
$C(\overline{\Omega})$. It also corrects a geometric issue in the proof of
Theorem 2 in \cite{ClosCelikSahutoglu18} (where the result is given for $q=1$
only).
Since necessary and sufficient conditions for compactness of the
$\overline{\partial}$--Neumann operator are understood on convex domains
(\cite{FuStraube98, FuStraube01, StraubeBook}), it is reasonable to expect that
when the domain is convex, one should likewise obtain simple necessary and
sufficient conditions on the symbol that will guarantee compactness of the
Hankel operator. That is, one expects the converse of Theorem
\ref{ThmNonCompact} to hold: when the symbol is holomorphic along
$q$--dimensional varieties in the boundary, $H^{q-1}_{\phi}$ should be compact.
As mentioned above, this implication is known for $C^{1}$ convex domains in
$\mathbb{C}^{2}$ and symbol in $C^{1}(\overline{\Omega})$
(\cite[Theorem 3]{CuckovicSahutoglu09},
\cite[Theorem (\v{C}u\c{c}kovi\'{c}--\c{S}ahuto\u{g}lu)]{ClosCelikSahutoglu18},
as well as for some classes of Reinhardt domains in $\mathbb{C}^{2}$,
with symbol only assumed in $C(\overline{\Omega})$
(\cite[Theorem 1]{ClosSahutoglu18}, \cite[Theorem 1]{ClosAccepted}).
Theorem \ref{converse} says that the implication is true for convex domains in
higher dimensions as well, when there are only `finitely many' varieties in
the boundary. We expect this additional assumption to be an artifact of our
current proof. On the other hand, our result appears to be the first instance of a
converse to Theorem \ref{ThmNonCompact} in dimension greater than two.
Before stating our results precisely, we recall the notation form
\cite{CelikSahutogluStraube20} (which is fairly standard). For a
bounded domain $\Omega\subset\mathbb{C}^{n}$, denote by $K^{2}_{(0,q)}(\Omega)$
the space of square integrable $\overline{\partial}$--closed $(0,q)$--forms, and
by $A^{2}_{(0,q)}(\Omega)$ the subspace of forms with holomorphic coefficients.
$P_{q}: L^{2}_{(0,q)}(\Omega)\rightarrow K^{2}_{(0,q)}(\Omega)$ is the
orthogonal projection, the Bergman projection on $(0,q)$--forms. For a symbol
$\phi\in L^{\infty}(\Omega)$, the associated Hankel operator is
$H^{q}_{\phi}:K^{2}_{(0,q)}(\Omega)\rightarrow L^{2}_{(0,q)}(\Omega)$,
\begin{align*}
H^{q}_{\phi}f = \phi f - P_{q}(\phi f),
\end{align*}
with $\|H^{q}_{\phi}\|\leq \|\phi\|_{L^{\infty}(\Omega)}$. $H^{q}_{\phi}$ equals
the commutator $[\phi,P_{q}]$ (since $P_{q}f = f$), so that statements about
(compactness of) Hankel operators may also be viewed as statements about
commutators between the Bergman projection and multiplication operators.
When the symbol $\phi$ is in $C^{1}(\overline{\Omega})$, Kohn's formula,
$P_{q} = \overline{\partial}^{*}N_{q+1}\overline{\partial}$, implies
\begin{align}\label{Kohn}
H^{q}_{\phi}f = \overline{\partial}^{*}N_{q+1}(\overline{\partial}\phi\wedge f);
\end{align}
here, $N_{q+1}$ is the $\overline{\partial}$--Neumann operator on $(0,q+1)$--forms.
We can now state our first result.
\begin{theorem}\label{ThmNonCompact}
Let $\Omega$ be a bounded convex domain in $\C^n$, $n\geq 2$, and
$\phi\in C(\overline{\Omega})$. Let $\psi:\mathbb{D}^q\to b\D$ be a
holomorphic embedding for some $q$ with $1\leq q\leq n-1$.
If $H^{q-1}_{\phi}$ is compact on $A^2_{(0,q-1)}(\D)$ (a fortiori if it is
compact on $K^{2}_{(0,q-1)}(\Omega)$), then $\phi\circ \psi$ is holomorphic.
\end{theorem}
Before stating Theorem \ref{converse}, we first recall the basic facts about
analytic varieties in the boundaries of convex domains from
\cite[Section 2 and Proposition 3.2]{FuStraube98} and
\cite[Lemma 2]{CuckovicSahutoglu09}. Suppose
$\psi:\mathbb{D}^{q}\rightarrow b\Omega$ is a holomorphic embedding, where
$\Omega\subset\mathbb{C}^{n}$ is convex. Then the convex hull of the image
$\psi(\mathbb{D}^{q})$ is contained in the intersection of a complex hyperplane
$H$ through $P\in \psi(\mathbb{D}^{q})$ with $\overline{\Omega}$
(\cite[Section 2]{FuStraube98}). So it suffices to consider affine varieties in the
boundary of this kind. Among these varieties through a boundary point $P$,
there is a unique one, denoted by $V_{P}$, that has $P$ as a relative interior
point and whose dimension $m$ is maximal. Then $0\leq m\leq (n-1)$, and
the relative closure $\overline{V_{P}}$ is the intersection of an $m$--dimensional
affine subspace through $P$ (contained in $H$) with $\overline{\Omega}$.
Our second result is as follows.
\begin{theorem}\label{converse}
Let $\D$ be a bounded convex domain in $\C^n, \phi\in C(\overline{\Omega}),$ and
$1\leq q\leq n$. Assume that the boundary of $\D$ contains at most finitely many
disjoint varieties $\{\overline{V_{P_{k}}}\}$ of dimension $q$ or higher as above,
$k=1, \ldots, N$. Furthermore, assume that $\phi\circ\psi$ is holomorphic for every
embedding $\psi:\mathbb{D}^q\to b\Omega$. Then $H^{q-1}_{\phi}$ is compact
on $K^2_{(0,q-1)}(\D)$.
\end{theorem}
Note that the condition on holomorphicity of $\phi\circ\psi$ just says
that $\phi|_{V_{P_{K}}}$ is holomorphic on all $V_{P_{k}}$ of dimension $q$
or higher. We also point out that this assumption on $\phi$ does not
imply that the tangential component of $\overline{\partial}\phi$
(say when $\Omega$ is smooth) vanishes on the $V_{P_{k}}$ of dimension
$q$ or higher; the components transverse to the varieties need not be zero.
If $\Omega$ is assumed $C^{1}$, the relative closures of distinct varieties are
automatically disjoint: if $Q\in
\overline{V_{P_{k_{1}}}}\cap\overline{V_{P_{k_{2}}}}$, then both $V_{P{k_{1}}}$
and $V_{P{k_{2}}}$ have to be contained in the complex tangent space to
$b\Omega$ at $Q$, and considering the convex hull as above produces a variety
which contains them both. However, without a regularity assumption on $b\Omega$,
the supporting complex hyperplane is not unique, and two distinct varieties may
share a boundary point (as on the boundary of $\mathbb{D}\times\mathbb{D}$).
When $q=n$, there are no $q$--dimensional varieties in the boundary, and the
theorem says that $H^{n-1}_{\phi}$ is always (when $\phi\in
C(\overline{\Omega})$) compact. But this is clear from \eqref{Kohn}, at least
when $\phi\in C^{1}(\overline{\Omega})$, because $N_{n}$, and hence
$\overline{\partial}^{*}N_{n}$, is compact (\cite[Theorem 1.1]{FuStraube98}).
When $\phi$ is merely in $C(\overline{\Omega})$, it can be approximated
uniformly on $\overline{\Omega}$ by smooth functions; the corresponding
(compact) Hankel operators converge in norm to $H_{\phi}^{n-1}$.
As pointed out in \cite[Remark 1]{ClosCelikSahutoglu18}, Theorem \ref{converse}
fails on general domains, that is, without the assumption of convexity or a
related condition. This failure is related to the subtleties surrounding
compactness in the $\overline{\partial}$--Neumann problem on general domains
(but absent in the case of convex domains). Namely, there are smooth bounded
pseudoconvex complete Hartogs domains in $\mathbb{C}^{2}$ without discs in the
boundary and noncompact Hankel operators on them whose symbols are smooth
on the closure. These symbols trivially satisfy the assumption on holomorphicity
along analytic discs. The domains were originally constructed in \cite{M98}
as examples of smooth bounded pseudoconvex complete Hartogs domains
without discs in the boundary whose $\overline{\partial}$--Neumann operator
$N_{1}$ is nevertheless not compact
(see also \cite[Theorem 10]{FuStraube01}, \cite[Theorem 4.25]{StraubeBook}).
But on these domains, noncompactness of $N_{1}$ implies that there are
symbols smooth on the closure so that the associated Hankel operator is not
compact (\cite[Theorem 1]{SahutogluZeytuncu17}).
\section{Proofs}\label{proofs}
The following simple lemma formulates the lack of holomorphicity of a function
without relying on differentiability, in contrast to
\cite{CelikSahutogluStraube20}; this ultimately allows the regularity of the
symbol to be lowered from $C^{1}(\overline{\Omega})$ to $C(\overline{\Omega})$
in Theorem \ref{ThmNonCompact}.
\begin{lemma}\label{LemWeakHolo}
Let $\Omega$ be a domain in $\C^n$ and $\phi\in C(\Omega)$ that is not
holomorphic.
Then there exist $h\in C^{\infty}_0(\Omega)$ and $1\leq j\leq n$ such that
$ \int_{\Omega}\phi\overline{h_{z_j}} \neq 0.$
\end{lemma}
\begin{proof}
Assume that $\int_{\Omega}\phi\overline{h_{z_{j}}} = 0$ for all $j$ and
$h\in C^{\infty}_0(\Omega)$. This means that $\overline{\partial}\phi=0$ as a
distribution
on $\Omega$. But $\overline{\partial}$ is elliptic in the interior, so that
$\phi$ is an ordinary holomorphic function\footnote{Alternatively,
$\overline{\partial}\phi=0$ implies $\Delta\phi = 0$, so by Weyl's Lemma, $\phi$
is a $C^{\infty}$ function (and therefore holomorphic).}, a
contradiction.
\end{proof}
\begin{proof}[Proof of Theorem \ref{ThmNonCompact}]
We start the proof of Theorem \ref{ThmNonCompact} as in \cite[Proof of Theorem
1]{CelikSahutogluStraube20}.
After dilation, translation and rotation if necessary, we assume that
$(2\mathbb{D})^q\times\{0\}\subset b\D$ and, seeking to derive a contradiction,
that the function
$\phi(z_1,\ldots,z_q,0,\ldots,0)$ is not holomorphic on $\mathbb{D}^q$.
Let us denote
$\D_0=\{(z_{q+1},\ldots,z_n)\in \C^{n-q}:(0,\ldots,0,z_{q+1},\ldots,z_n)\in
\D\}$,
the slice transversal to $\mathbb{D}^q$ through the origin. Then
convexity of $\D$ implies that $\mathbb{D}^q\times \beta \D_0\subset \D$
for any $0<\beta\leq 1/2$ (\cite[page 636, proof of $(1)\Rightarrow
(2)$]{FuStraube98}; \cite[Proof of Theorem 2]{ClosCelikSahutoglu18}).
Lemma \ref{LemWeakHolo} implies that there exist
$h\in C^{\infty}_0(\mathbb{D}^q)$ and $\delta>0$ such that
\begin{align*}
\left| \int_{\mathbb{D}^q} \phi(z_1,\ldots,z_q,0,\ldots,0)
\frac{\partial \overline{h(z_1,\ldots,z_q)}}{\partial
\zb_1}dV(z_1,\ldots,z_q)\right|
\geq 2\delta.
\end{align*}
Furthermore, continuity of $\phi$ implies that there exists $0<\beta<1/2$ such
that
\begin{align*}
\left| \int_{\mathbb{D}^q} \phi(z_1,\ldots,z_n)
\frac{\partial \overline{h(z_1,\ldots,z_q)}}{\partial
\zb_1}dV(z_1,\ldots,z_q)\right|
\geq \delta
\end{align*}
for all $(z_{q+1},\ldots,z_n)\in \beta \D_0$. As in
\cite[Proof of Theorem 2]{ClosCelikSahutoglu18} we use their Lemma 5
to produce a bounded sequence
$\{f_j\}_{j=1}^{\infty}=\{f_j(z_{n})\}_{j=1}^{\infty}\subset A^2(\D)$
such that $\|f_j\|_{L^2(\beta\D_0)}=1$ and $f_{j}\rightarrow 0$ weakly in
$A^{2}_{(0,0)}(\Omega)$\footnote{Alternatively, one can modify the argument
below slightly and use the construction in
\cite[Proof of $(1)\Rightarrow (2)$ in Theorem 1.1]{FuStraube98}.}.
We define $\alpha_j=f_jd\zb_2\wedge \cdots\wedge d\zb_q.$ Then the sequence
$\{\alpha_{j}\}_{j=1}^{\infty}$ also has (the analogues of) these two
properties.
We have
\begin{align}
\nonumber \delta^2=& \delta^2\|f_j\|^2_{L^2(\beta \D_0)}\\
\nonumber \leq & \int_{\beta\D_0}\left|\int_{\mathbb{D}^q}\phi(z_1,\ldots,z_n)
\frac{\partial \overline{h(z_1,\ldots,z_q)}}{\partial
\zb_1}dV(z_1,\ldots,z_q)\right|^2
f_j(z_n)\overline{f_j(z_n)}dV(z_{q+1},\ldots,z_n) \\
\nonumber =& \int_{\beta\D_0}\left|\int_{\mathbb{D}^q}\phi(z_1,\ldots,z_n)
f_j(z_n)
\frac{\partial \overline{h(z_1,\ldots,z_q)}}{\partial \zb_1}dV(z_1,\ldots,z_q)
\right|^2 dV(z_{q+1},\ldots,z_n) \\
\label{Eqn1} =& \int_{\beta\D_0}\left|\Big( \phi \alpha_j,
\frac{\partial h}{\partial z_1}d\zb_2\wedge \cdots \wedge
d\zb_q\Big)_{\mathbb{D}^q}
\right|^2 dV(z_{q+1},\ldots,z_n)\\
\nonumber =& \int_{\beta\D_0}\left|\Big( \phi \alpha_j,
\dbar^*_{z_1\cdots z_q}(hd\zb_1\wedge \cdots \wedge d\zb_q)\Big)_{\mathbb{D}^q}
\right|^2 dV(z_{q+1},\ldots,z_n).
\end{align}
In the last two lines $(\cdot,\cdot)$ denotes the standard inner product between
forms on the fibers $\mathbb{D}^{q}\times\{(z_{q+1}, \ldots, z_{n})\}$, and
$\dbar^*_{z_1\cdots z_q}$ is the adjoint of $\overline{\partial}_{z_1\cdots
z_q}$, the $\overline{\partial}$ on each fiber (note that $hd\zb_1\wedge \cdots
\wedge d\zb_q \in dom(\dbar^*_{z_1\cdots z_q})$ since $h\in
C^{\infty}_{0}(\mathbb{D}^{q})$).
In the last equality above we used the fact that all the additional terms
in $\dbar^*_{z_1\cdots z_q}(hd\zb_1\wedge \cdots \wedge d\zb_q)$ involve
$d\zb_1$ and are thus (even pointwise) orthogonal to $\phi\alpha_{j}$.
Next, note that
\begin{align*}
\dbar_{z_1\cdots z_q} (P_{q-1}\phi\alpha_j|_{\mathbb{D}^q})
=(\dbar_{z_1\cdots z_n} P_{q-1}\phi\alpha_j)|_{\mathbb{D}^q} =0,
\end{align*}
where $|_{\mathbb{D}^q}$ denotes the restriction (pull back) of
forms to each fiber $\mathbb{D}^q\times \{(z_{q+1}, \ldots, z_{n})\}$.
Therefore, since $\phi\alpha_{j} = \phi\alpha_{j}|_{\mathbb{D}^q}$ on
$\mathbb{D}^{q}\times \beta\Omega_{0}$,
\begin{align*}
\Big( \phi \alpha_j,
\dbar^*_{z_1\cdots z_q}hd\zb_1\wedge \cdots \wedge d\zb_q\Big)_{\mathbb{D}^q}
=& \Big( (\phi\alpha_j-P_{q-1}(\phi\alpha_j))|_{\mathbb{D}^q},
\dbar^*_{z_1\cdots z_q}hd\zb_1\wedge \cdots \wedge d\zb_q\Big)_{\mathbb{D}^q} \\
=&\Big( (H^{q-1}_{\phi}\alpha_j)|_{\mathbb{D}^q},
\dbar^*_{z_1\cdots z_q}hd\zb_1\wedge \cdots \wedge
d\zb_q\Big)_{\mathbb{D}^q}
\end{align*}
and consequently
\begin{align*}
\left|\Big( \phi \alpha_j,
\dbar^*_{z_1\cdots z_n}hd\zb_1\wedge \cdots \wedge d\zb_q
\Big)_{\mathbb{D}^q}\right|
\leq
\left\|(H^{q-1}_{\phi}\alpha_j)|_{\mathbb{D}^q}\right\|_{L^2(\mathbb{D}^q)}
\left\| \dbar^*_{z_1\cdots z_q}
hd\zb_1\wedge \cdots \wedge d\zb_q \right\|_{L^2(\mathbb{D}^q)}.
\end{align*}
Combining this estimate with \eqref{Eqn1} and observing that
$|_{\mathbb{D}^{q}}$ decreases norms (pointwise on each fiber $\mathbb{D}^{q}$:
the omitted terms containing $d\overline{z}_{m}$ with $m>q$ are orthogonal to
$(H^{q-1}_{\phi}\alpha_{j})|_{\mathbb{D}^{q}}$) and that
$\mathbb{D}^{q}\times\beta\Omega_{0} \subseteq \Omega$ gives
\begin{align*}
\delta \lesssim \left\|(H^{q-1}_{\phi}\alpha_j)\right\|_{L^2(\D)}
\left\| \dbar^*_{z_1\cdots z_q}
hd\zb_1\wedge \cdots \wedge d\zb_q \right\|_{L^2(\mathbb{D}^q)}.
\end{align*}
Therefore, $\{H^{q-1}_{\phi}\alpha_j\}_{j=1}^{\infty}$ does not converge to 0 in
$L^2_{(0,q-1)}(\D)$
as $j\to \infty$ (since the second factor on the right hand side is nonzero).
On the other hand, the sequence $\{\alpha_j\}_{j=1}^{\infty}$ converges to zero
weakly
in $A^2_{(0,q-1)}(\D)$, and therefore in $K^2_{(0,q-1)}(\D)$, as $j\to \infty$,
and compactness of $H_{\phi}^{q-1}$ then forces convergence of
$\{H^{q-1}_{\phi}\alpha_j\}_{j=1}^{\infty}$ to zero in $L^2_{(0,q-1)}(\D)$. This
contradiction completes the proof of Theorem \ref{ThmNonCompact}.
\end{proof}
The above proof of Theorem \ref{ThmNonCompact} uses ideas from
\cite{FuStraube98, CuckovicSahutoglu09, ClosCelikSahutoglu18,
CelikSahutogluStraube20}; in turn, these ideas can be traced back at least to
\cite{Catlin81, DiederichPflug81}.
In the proof of Theorem \ref{converse}, we will repeatedly use the following
sufficient condition for compactness of an operator $T:X\rightarrow Y$, where
$X$ and $Y$ are Hilbert spaces (see e.g. \cite[Lemma 4.3(ii)]{StraubeBook}): for
all $\varepsilon>0$ there are a Hilbert space $Z_{\varepsilon}$, a linear
compact operator $S_{\varepsilon}:X\rightarrow Z_{\varepsilon}$, and a constant
$C_{\varepsilon}$ such that
\begin{align}\label{compcond}
\|Tx\|_{Y} \leq \varepsilon\|x\|_{X} +
C_{\varepsilon}\|S_{\varepsilon}x\|_{Z_{\varepsilon}}.
\end{align}
In addition, we need the following sufficient conditions for compactness of a
Hankel operator on $\Omega$ (notation as in the theorem).
\begin{lemma}\label{preplemma}
Suppose that $\Omega$ is as in Theorem \ref{converse} and
\begin{itemize}
\item[(i)] $\phi\in C^{1}(\overline{\Omega})$ and $\overline{\partial}\phi$
vanishes on $\cup_{j=1}^{N}\overline{V_{P_{j}}}$, or
\item[(ii)] $\phi\in C(\overline{\Omega})$ vanishes on
$\cup_{j=1}^{N}\overline{V_{P_{j}}}$.
\end{itemize}
Then $H^{q-1}_{\phi}$ is compact on $K^{2}_{(0,q-1)}(\Omega)$.
\end{lemma}
Note that the condition in $(i)$ is stronger than saying that $\phi$ is
holomorphic along the $V_{P_{j}}$: $\overline{\partial}\phi = 0$ also in the
directions transverse to $V_{P_{j}}$.
\begin{proof}
We start with $(i)$\footnote{When $\Omega$ is smooth, $(i)$ is a
special case of Theorem 1.3 in \cite{ChengJinWang}.}. We want to estimate
$\|H^{q-1}_{\phi}f\|^{2}$ in such a way that we can use \eqref{compcond}. So fix
$\varepsilon>0$. In view of \eqref{Kohn}, we have
\begin{align}\label{practical}
\|H^{q-1}_{\phi}f\|^{2} =
\langle\overline{\partial}^{*}N_{q}(\overline{\partial}\phi\wedge f),
\overline{\partial}^{*}N_{q}(\overline{\partial}\phi\wedge f)\rangle
= \langle N_{q}(\overline{\partial}\phi\wedge f), \overline{\partial}\phi\wedge
f\rangle
\end{align}
for $f\in K^{2}_{(0,q-1)}(\Omega)$.
The (absolute value of the) inner product on the right hand side is
\begin{align}\label{innerprod}
\sum_{j=1}^{n}\;\sideset{}{'}\sum_{|K|=q-1} \int_{\Omega}
(N_{q}(\overline{\partial}\phi\wedge f))_{jK}
\overline{(\partial\phi/\partial\overline{z_{j}})f_{K}}dV
\lesssim \sum_{j=1}^{n}\|(\partial\overline{\phi}/\partial
z_{j})N_{q}(\overline{\partial}\phi\wedge f)\|\,\|f\|,
\end{align}
with the usual notation $f=\sideset{}{'}\sum_{|K|=q-1}f_{K}d\overline{z_{K}}$,
the $'$ indicating summation over increasing multi--indices, and
$jK=(j,k_{1},\ldots,k_{q-1})$. Using the inequality $2ab\leq (a^{2}/\varepsilon
+ \varepsilon b^{2})$, where $a,b>0$, gives
\begin{align}\label{eq13}
2\|(\partial\overline{\phi}/\partial z_{j})N_{q}(\overline{\partial}\phi\wedge
f)\|\,\|f\|\leq \frac{1}{\varepsilon}\|(\partial\overline{\phi}/\partial
z_{j})N_{q}(\overline{\partial}\phi\wedge f)\|^{2} + \varepsilon \|f\|^{2}.
\end{align}
Because $(\partial\overline{\phi}/\partial z_{j})$ vanishes on
$\cup_{j=1}^{N}\overline{V_{P_{j}}}$, it is a compactness multiplier on
$(0,q)$--forms (\cite[Proposition 1, Theorem 3]{CelikStraube09}): for all
$\varepsilon^{\prime}>0$, there is a constant $C_{\varepsilon^{\prime},j}$ such
that
\begin{align}\label{compmultiplier}
\|(\partial\overline{\phi}/\partial z_{j})u\|^{2} \leq
\varepsilon^{\prime}(\|\overline{\partial}u\|^{2} +
\|\overline{\partial}^{*}u\|^{2}) + C_{\varepsilon^{\prime},j}\|u\|_{-1}^{2},
\end{align}
$u\in dom(\overline{\partial})\cap dom(\overline{\partial}^{*})\subset
L^{2}_{(0,q)}(\Omega)$.
We apply \eqref{compmultiplier} to the form
$u=N_{q}(\overline{\partial}\phi\wedge f)$ on the right hand side of
\eqref{eq13}, with $\varepsilon^{\prime}=\varepsilon^{2}$, to obtain that the
right hand side of \eqref{eq13} is dominated by
\begin{align}\label{hankelest}
\varepsilon\|\overline{\partial}^{*}N_{q}(\overline{\partial}\phi\wedge f)\|^{2}
+ C_{\varepsilon,j}\|N_{q}(\overline{\partial}\phi\wedge f)\|_{-1}^{2} +
\varepsilon\|f\|^{2} \\
\lesssim \varepsilon\|f\|^{2} +
C_{\varepsilon,j}\|N_{q}(\overline{\partial}\phi\wedge f)\|_{-1}^{2};
\end{align}
the constants involved (other than $C_{\varepsilon,j}$) are independent of
$\varepsilon$ and $f$ (but not of $\phi$). We have used here that
$N_{q}(\overline{\partial}\phi\wedge f)\in dom(\overline{\partial})\cap
dom(\overline{\partial}^{*})$ and that
$\overline{\partial}N_{q}(\overline{\partial}\phi\wedge f)=0$ (since
$\overline{\partial}f=0$). Combining \eqref{practical} --\eqref{eq13} and
\eqref{hankelest} and summing over $j=1,\ldots,n$ results in the estimate we are
looking for:
\begin{align*}
\|H^{q-1}_{\phi}f\|^{2} \lesssim \varepsilon\|f\|^{2} +
C_{\varepsilon}\|N_{q}(\overline{\partial}\phi\wedge f)\|_{-1}^{2}\;;\;
f\in K^{2}_{(0,q-1)}(\Omega),
\end{align*}
with $C_{\varepsilon} = max\{C_{\varepsilon,j}\,|\,j=1,\ldots,n\}$.
The operator $f\rightarrow N_{q}(\overline{\partial}\phi\wedge f)$ is continuous
from $K^{2}_{(0,q-1)}(\Omega)$ to $L^{2}_{(0,q)}(\Omega)$, hence is compact to
$W^{-1}_{(0,q)}(\Omega)$ (see e.g. (\cite[Theorem 1.4.3.2]{GrisvardBook}:
${W^{1}_{0}}(\Omega)\hookrightarrow L^{2}(\Omega)$ is compact; by duality,
so is $L^{2}(\Omega)\hookrightarrow W^{-1}(\Omega)$). Therefore
\eqref{compcond} is satisfied for $T=H^{q-1}_{\phi}$, with
$X=K^{2}_{(0,q-1)}(\Omega)$, $Y=L^{2}_{(0,q-1)}(\Omega)$,
$Z_{\varepsilon}=W^{-1}_{(0,q)}(\Omega)$, and
$S_{\varepsilon}=N_{q}(\overline{\partial}\phi\wedge f)\,:
\,K^{2}_{(0,q-1)}(\Omega)\rightarrow W^{-1}_{(0,q)}(\Omega)$, and
$H^{q-1}_{\phi}$ is compact. (In this particular case, $Z_{\varepsilon}$ and
$S_{\varepsilon}$ are independent of $\varepsilon$.)
$(ii)$ is an easy consequence of $(i)$. Any symbol as in $(ii)$ can be
approximated uniformly on $\overline{\Omega}$ by symbols in
$C^{1}(\overline{\Omega})$ and vanishing in a neighborhood of the compact set
$\cup_{j=1}^{N}\overline{V_{P_{j}}}$. These symbols satisfy the assumption in
$(i)$, and the corresponding (compact) Hankel operators converge to
$H^{q-1}_{\phi}$ in norm\footnote{Because $\|H^{q-1}_{\phi} -
H^{q-1}_{\phi_{j}}\| = \|H^{q-1}_{\phi - \phi_{j}}\| \leq \|\phi -
\phi_{j}\|_{L^{\infty}(\Omega)}$.}. Therefore, $H^{q-1}_{\phi}$ is compact as
well.\footnote{By \cite[Proposition 1, Theorem 3]{CelikStraube09}, $\phi$ is a
compactness multiplier for $(0,q)$--forms. When $\Omega$ is smooth (and $q=1$),
one can use \cite[Theorem 1]{CelikZeytuncu16}: symbols which are compactness
multipliers on a bounded smooth pseudoconvex domain produce compact Hankel
operators.}
\end{proof}
We are now ready to prove Theorem \ref{converse}.
\begin{proof}[Proof of Theorem \ref{converse}]
For the moment, fix $j$, $1\leq j\leq N$. $\overline{V_{P_{j}}}$ is a convex
subset of an affine subspace, and $\phi$ is holomorphic on $V_{P_{j}}$ and
continuous on the closure. Via dilatation, we can approximate $\phi$ uniformly
of $\overline{V_{P_{j}}}$, first by functions holomorphic in a relative
neighborhood of $\overline{V_{P_{j}}}$, and then (by trivial extension)
holomorphic in a neighborhood in $\mathbb{C}^{n}$ of $\overline{V_{P_{j}}}$.
To prove compactness of $H^{q-1}_{\phi}$, we use the sufficient condition in
\eqref{compcond}. So fix $\varepsilon>0$. For $j=1,\ldots,N$, choose open sets
$U_{j}$ in $\mathbb{C}^{n}$ that contain $\overline{V_{P_{j}}}$ and are pairwise
disjoint, and cutoff functions $\chi_{j}\in C^{\infty}_{0}(U_{j})$,
$0\leq\chi_{j}\leq 1$, which are identically equal to one in a neighborhood of
$\overline{V_{P_{j}}}$. In light of the previous paragraph, we can do that in
such a way that there exists a holomorphic function
$h_{j}$ on $U_{j}$ that first approximates $\phi$ to within $\varepsilon$ on
$\overline{V_{P_{j}}}$, and then by continuity to within $2\varepsilon$ on
$U_{j}\cap\overline{\Omega}$ (shrinking $U_{j}$ if necessary). Also set
$\chi_{0} =(1-\sum_{j=1}^{N}\chi_{j})\in C(\overline{\Omega})$. With this setup,
we split $H^{q-1}_{\phi}$ as follows:
\begin{align}\label{split}
H^{q-1}_{\phi} = H^{q-1}_{(\chi_{0}\phi)} + H^{q-1}_{\sum_{j=1}^{N}\chi_{j}\phi}
= H^{q-1}_{(\chi_{0}\phi)} + H^{q-1}_{\sum_{j=1}^{N}\chi_{j}h_{j}}
+ H^{q-1}_{\sum_{j=1}^{N}\chi_{j}(\phi-h_{j})}\;.
\end{align}
$\chi_{0}\phi$ vanishes on $\cup_{j=1}^{N}\overline{V_{P_{j}}}$, and
Lemma \ref{preplemma}, part $(ii)$, implies that $H^{q-1}_{(\chi_{0}\phi)}$
is compact. It remains to consider the remaining terms on the right hand
side of \eqref{split}.
We have $\sum_{j=1}^{N}\chi_{j}h_{j}\in C^{\infty}(\overline{\Omega})$ and
$\overline{\partial}(\sum_{j=1}^{N}\chi_{j}h_{j}) = \sum_{j=1}^{N}h_{j}\overline{\partial}\chi_{j}$.
This symbol therefore satisfies the assumptions in part $(i)$ of Lemma \ref{preplemma},
and the corresponding Hankel operator is compact. Because
$\|\chi_{j}(\phi - h_{j})\|_{L^{\infty}(\Omega)} \leq 2\varepsilon$, and the $\chi_{j}$
have disjoint supports, the norm of the last operator in \eqref{split} is
bounded by $2\varepsilon$, and we conclude
\begin{align}\label{chi-k}
\|H^{q-1}_{\phi}f\| \leq 2\varepsilon\|f\| + \|H^{q-1}_{(\chi_{0}\phi + \sum_{j=1}^{N}\chi_{j}h_{j})}f\|\;.
\end{align}
Estimate \eqref{chi-k} is \eqref{compcond} (after rescaling $\varepsilon$) with
$X=K^{2}_{(0,q-1}(\Omega)$, $Y=L^{2}_{(0,q-1}(\Omega)$,
$Z_{\varepsilon}=L^{2}_{(0,q-1)}(\Omega)$, and the compact operator
$S_{\varepsilon}=H^{q-1}_{(\chi_{0}\phi + \sum_{j=1}^{N}\chi_{j}h_{j})}
= H^{q-1}_{\chi_{0}\phi} + H^{q-1}_{\sum_{j=1}^{N}\chi_{j}h_{j}}$ (in contrast to
$Z_{\varepsilon}$, $S_{\varepsilon}$ does depend on $\varepsilon$, via $\chi_{j}$
and $h_{j}$). As $\varepsilon > 0$ was arbitrary, we conclude that $H^{q-1}_{\phi}$
is compact and complete the proof of Theorem \ref{converse}.
\end{proof}
| 8d076fe8c26d92cd80a3122afa2e1d1bbcecd453 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{O}{ver} the last decade, we have witnessed the evolution of vehicular networking from 'Vehicular Ad-Hoc Networks' (VANETs), enabling spontaneous direct communications between vehicles, to 'Connected Vehicles' generalizing information exchange among vehicles and infrastructure. Vehicular networking has been developed as an enabler of innovative applications intended to improve traffic safety, reduce congestion and even provide infotainment on-board. Early applications were designed to only provide information to drivers, delegating any decision making to them.
However, in recent ambitious applications, such as automated driving or platooning, simple information treatment and forwarding mechanisms are not sufficient anymore. Instead, decision-making is based on models of the environment built and improved from increasingly large and diverse sets of input information. Models are designed to learn from experience rather than react to static input signals. In this context, the current standards of vehicular networking, designed for the exchange of static information alone, are no longer optimal. As the individual piece of information becomes meaningless compared with the model that was extracted from it, the scope of vehicular networks is gradually shifting to the creation and exchange of \emph{knowledge} models, as opposed to the sum of their training input.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{img/data_info_knowledge.pdf}
\caption{The Relationship between Data, Information and Knowledge}
\label{fig:data_info_knowledge}
\end{figure}
The existing literature in the information science domain covers conceptual definition of data, information and knowledge \cite{zins2007}. In this paper, for the sake of clarity, we make similar distinctions among these categories. As shown in Figure~\ref{fig:data_info_knowledge}, the most fundamental element is data, that we define as an atomic value with a unit e.g. (5km). Next, information is built by aggregating pieces of data that describe a situation, e.g. (15:00, 30kph) a vehicle's speed at a given time. On top of information lies knowledge, which describes general patterns and relationships obtained through the analysis of sets of information. For example, clustering or classification algorithms can be used to extract hidden relationships within a set. Considering vehicles' pollutant emissions and speed through a day, a piece of extracted knowledge could be an estimate of the extra emissions linked to peak time.
Various techniques, such as Artificial Intelligence (AI), Machine Learning (ML) or Formal Language (FL) have been used to extract knowledge in vehicular contexts through the analysis of various sources of information~\cite{yeLiang2018}. For example, Ruta~{\em et al.}~\cite{ruta2018} composed {\em knowledge} to recognize a high level context of driving by using sensor information from multiple cars in a common geographical area. Qi, Wang~{\em et al.}~\cite{qi2019} applied both learning algorithms and edge computing units offloading to provide optimal caching of high level connected driving services to vehicles, including image auto annotation or locally relevant recommendations yielding. Khan~{\em et al.}~\cite{khan2019} applied deep learning to learn the transmission patterns of neighboring vehicles and paved the way for fewer packet collisions.
Regardless of the technique, extracting {\em knowledge} from information is a complex and expensive process, and the generated {\em knowledge} may be beneficial to other vehicles. So far each vehicle remains autonomous for its {\em knowledge} building, which requires highly specialized algorithms and a large amount of input information, potentially sourced from multiple different vehicles. This can be seen as a significant overhead considering that {\em knowledge} can be shared and not individually recreated. As a reaction, research has recently been focused on defining a knowledge-centric approach to networking, where information would not be the main focus anymore. Instead, knowledge would be created by nodes in the network, and directly stored and shared among them.
Wu~{\em et al.}~\cite{wu2019} described the concept of a knowledge-centric networking framework, separated into the three building blocks of knowledge creation, composition and distribution. A literature survey on means of creating and distributing knowledge is performed. However, the concept of knowledge remains abstract and its implementation or format is left for future work.
The structure and understanding of {\em knowledge} needs to be harmonized. Knowledge must move, follow and adapt in symbiosis with information as well as other pieces of knowledge.
Finally, existing information networking mechanisms such as Information-Centric Networking (ICN) lack semantic annotations describing such context, which prevails them to uniquely locate and acquire the required knowledge~\cite{yao2019}. The naming, storage and dissemination of {\em knowledge} thus must integrate semantic annotations, which in turn need to be harmonized.
In this paper we aim to present Vehicular Knowledge Networking (VKN), a knowledge-centric networking framework applied to vehicular networks. We suggest a common architecture for knowledge description, needed for subsequent storage, composition and exchange with other connected vehicles. VKN is a framework that makes performance improvements in other applications possible. In a conceptual passenger comfort-based rerouting application, we show a potential load and delay reduction in vehicular networks through cooperative knowledge building and sharing.
The rest of this article is organized as follows: Section~\ref{sec:infovsknowledge} introduces information treatment standards and defines the scope of \emph{knowledge} in vehicular networks. Based on this understanding, Section~\ref{sec:vkn} describes a structure for knowledge description, storage and distribution. In Section~\ref{sec:application}, an application of the concept shows potential load and delay improvement for the network. Section~\ref{sec:opportunities} finally points out the potential research applicability behind VKN, while Section~\ref{sec:conclusion} summarizes the article.
\section{Vehicular Information and Vehicular Knowledge}
\label{sec:infovsknowledge}
In this section, we first describe the current forms of information in vehicular networks, as well as various standards for information storing and sharing. Then, we build on this understanding to define a format for vehicular knowledge representation.
\subsection{Information in Vehicular Networks}
Nodes of the vehicular network may exchange diverse types of information, including but not limited to:
\begin{itemize}
\item Safety notifications, e.g. accidents or road condition.
\item Vehicle state information and sensor measurements.
\item Navigation information, e.g. maps, road or parking data.
\item Information on topics such as weather or traffic flow.
\item Road-related information, e.g. gas station opening times.
\item Multimedia contents for user infotainment.
\end{itemize}
Various standards to represent and store information on-board vehicles have been developed in the past years.
In ETSI standards, the storage of content in connected vehicles is performed inside the \emph{Local Dynamic Map (LDM)} information base~\cite{ETSIstandard_LDM}, itself divided into four layers:
\begin{enumerate}
\item Permanent static data, i.e. map data.
\item Temporary static data, i.e. roadside infrastructure.
\item Temporary dynamic data, e.g. roadblock, signal phase.
\item Highly Dynamic data, e.g. vehicles, pedestrians.
\end{enumerate}
Although the LDM provides a standard approach for storing information, it does not provide any standard for information naming. Generally, any information may be stored in the LDM as long as it is labeled with a space-time area of relevance. This can lead to a lack of interoperability between the contents generated by different providers. To tackle this issue, semantic standards have been developed in order to provide nodes with a "common language" and avoid redundancy. For example, Syzdykbayev {\em et al.}~\cite{syzdykbayev2019} defined an ontology for interactions between the different actors of vehicular networks and the infrastructure. Moreover, the Vehicle Signal Specification (VSS), later extended to the VSSo ontology~\cite{vsso2018}, provides a standard way to address the state of the inner components of vehicles, e.g. steering wheels or windows opening.
Finally, after it has been created and stored, information is spread within the vehicular network. In most vehicular applications, nodes care more about the information itself than about the host it originated from. Routing algorithms have thus been developed that focus on the information being shared rather than on its actual host and location in the network. Information-Centric Networking (ICN) is a networking paradigm that may be suitable for some vehicular applications. Rather than sending a request to a specific host, using ICN, a vehicle disseminates a request to fetch a specific information identified by a unique content name.
\subsection{Vehicular Knowledge Representation}
In next generation vehicular networks, we envision a shift away from the information-based architecture to the benefit of knowledge. As with information, mechanisms must be developed to create, compose, store and distribute knowledge, as introduced in the survey by Wu {\em et al.}~\cite{wu2019}. In order to go further and based on definitions of knowledge, we propose a formal structure for knowledge representation. In turn, it paves the way for the development of knowledge storage and distribution mechanisms.
We understand knowledge as a piece of abstract content obtained from the analysis of a larger set of information~\cite{zins2007}.
Knowledge can be extracted from information using machine learning algorithms, divided into three classes. Supervised learning is a technique applied to classification or regression. A model is trained based on a number of samples of the form: (information, class) for classification -- or (information, value) for regression. This step, called model training, allows it to grasp knowledge as the relationship between the information and its associated class i.e. the function that takes information as an input and returns its estimated class.
Unsupervised learning works with a set of information in which it is able to find clusters of similar items. It creates knowledge from a set of information by exposing relationships among information items and sorting them into different clusters. Reinforcement learning, finally, can be used by an agent to learn the optimal behavior to adopt in a context of interaction with an environment to maximize a user-defined reward.
Once trained, a machine learning model acts as a piece of knowledge, able to return synthetic knowledge from input information. The knowledge that is extracted through learning techniques can be further leveraged through knowledge composition methods where existing knowledge is further analyzed/collated to produce new knowledge. For instance, if a user wants to avoid traffic congestion, the system needs to first detect the congested zones and decompose the necessary factors including current location, destination, the estimated route/arrival time based on the current traffic. In this case, in addition to knowledge creation, the knowledge composition also collates some information and/or other knowledge, such as closed roads and/or construction zones, so as not to exacerbate the congestion.
Thus, the word \emph{knowledge} can be used to refer to both: (i)
the trained model or algorithm able to synthesize sets of information into a piece of knowledge, and (ii) the generated piece of knowledge, that further abstracts the available input information. As a consequence, we believe it is necessary to take both aspects into account as we suggest a formalization of knowledge definition, storage and distribution in vehicular networks.
\begin{figure}[H]
\centering
\includegraphics[width=\columnwidth]{img/arch_format_knowledge.pdf}
\caption{The Vehicular Knowledge Ecosystem}
\label{fig:archi_vehic_knowledge}
\end{figure}
Figure~\ref{fig:archi_vehic_knowledge} shows a general picture of when and how knowledge is handled in vehicular networks for safety and driving-related applications. To the left of the figure, information is meant to be stored in the LDM, and has a time and area of validity. For the sake of interoperability with other vehicles, we consider it to be named and structured following well-known constraints, e.g. following the VSS specification.
A \emph{knowledge model}, typically implemented as a trained machine learning algorithm, takes as input several pieces of information and produces a certain piece of knowledge as an output. In Figure~\ref{fig:archi_vehic_knowledge}, we distinguish two aspects of a knowledge model: a semantic description and a bytecode. The semantic description is used to describe the necessary input, the produced output and the potential preconditions necessary to the application of the model. We define the \emph{bytecode} of a model as the executable file that produces an output from a well-formed set of input information. By executing the bytecode of a model, one can produce a \emph{knowledge sample} that, just like the information used to produce it, has a time and area of validity. Another pertinent possible output for vehicles is an actuation signal, that may be triggered in reaction to the analysed input information.
In practice, the generated \emph{knowledge sample} is structured similarly to information, the difference being that it is abstract and obtained through analysis. As a consequence, it can be fed as input to another knowledge model, generating new composed knowledge.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.75\textwidth]{img/arch_applied_comfort.pdf}
\caption{Knowledge Models for Passenger Comfort Estimation.}
\label{fig:archi_applied_comfort}
\end{figure*}
Figure~\ref{fig:archi_applied_comfort} shows an application of this definition of knowledge. It is related to the estimation of passenger comfort on-board a connected vehicle. We will come back to the \texttt{model.env\_comfort} model in Section~\ref{sec:application}, as we describe an application of VKN.
\section{Aspects of Vehicular Knowledge Networking}
\label{sec:vkn}
\subsection{Knowledge Description \& Storage}
The creation of knowledge in vehicular networks takes place at two levels:
\begin{itemize}
\item Using machine learning algorithms, automakers and organizations train and provide models capable of generating a piece of abstract knowledge from a set of input (information or \emph{knowledge samples}).
\item The models are applied to real input, generating new \emph{knowledge samples}.
\end{itemize}
Knowledge models themselves are made of two separable components: (i) A description of the input and output parameters of the model, and (ii) the actual bytecode of the model, a program that will generate output \emph{knowledge samples} given well-formed input. Traditionally in vehicular networks, the two components are not separated. All aspects of the creation of a knowledge model, including training information gathering, model definition and training, are typically planned and organized by a single entity. The obtained model is then downloaded exclusively to the vehicle fleet of the training entity so that \emph{knowledge samples} can be locally generated.
In some cases, keeping models proprietary can be an intentional choice by automakers in order to protect their competitiveness in the market. However, in other cases, a lack of cooperation in knowledge model building may lead to an inefficient use of resources. On the one hand, similar knowledge models are likely to be independently trained by competing entities, leading to redundant computations. On the other hand, no common format to describe the input, output, and preconditions of a model are provided by training entities. This prevents the cooperative use of knowledge models by a larger number of nodes.
To tackle these issues, the semantic description of a model may be physically separated from its actual execution place, as shown in Figure~\ref{fig:knowledge_separation}. The knowledge model description formally states the input, output and preconditions to the application of a given model. It is a lightweight content to be shared with multiple nodes. The knowledge model's bytecode is the implemented program whose input and output match those of the semantic description. It performs the knowledge samples creation. Even if a vehicle is not in possession of a model's bytecode, it may request knowledge creation from another node using the constraints detailed in the model description.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{img/kdesc_vs_bytecode.pdf}
\caption{Separation of Model Description and Bytecode}
\label{fig:knowledge_separation}
\end{figure}
As a requirement for model description, the inputs and outputs of a model should be named according to standard semantics specifications such as VSS~\cite{vsso2018}. By consulting the specification for the name associated with each input or output, nodes are able to deduce the format of information and \emph{knowledge samples} required to apply the model. Moreover, preconditions to the model application may be set, e.g. conditions on the input information's age.
A possible candidate model description language matching these requirements is OWL-S, as described by Martin~\emph{et al.}~\cite{owls}. It was originally developed for automatic Web Services discovery, composition and invocation. The \emph{process model} standard of OWL-S provides a means of description for the set of input, output and preconditions of a model. Figure~\ref{fig:owls_example} provides an example of a OWL-S description of the \texttt{model.env\_comfort} model introduced in the bottom left corner of Figure~\ref{fig:archi_applied_comfort}.
\begin{figure}
\centering
\hspace*{0.05in}
\begin{tabular}{c}
\begin{lstlisting}[language=Xml]
<AtomicProcess ID="model.env_comfort">
<hasInput resource="#traffic" />
<hasInput resource="#visibility" />
<hasInput resource="#twoWheelers" />
</AtomicProcess>
<Input ID="traffic">
<parameterType resource="#Road.Traffic" />
</Input>
<Input ID="visibility">
<parameterType resource="#Road.Visibility" />
</Input>
<Input ID="twoWheelers">
<parameterType resource="#TwoWheelers.Concentration" />
</Input>
<Output ID="comfort">
<parameterType resource="#Road.ComfortLevel" />
</Output>
\end{lstlisting}
\end{tabular}
\caption{Comfort Model Description Structure in OWL-S}
\label{fig:owls_example}
\end{figure}
As part of VKN, we suggest separating the storage of models' descriptions and bytecodes. On-board each connected vehicle, as shown in Figure~\ref{fig:knowledge_storage}, we suggest adding a \emph{knowledge} layer on top of the ITS facilities layer containing the LDM:
\begin{itemize}
\item In a Knowledge Base (KB), a list of known knowledge model descriptions are stored.
\item A local storage in the knowledge layer may store knowledge model bytecodes. It is responsible for the actual execution of the knowledge creation process. The stored bytecodes are independent from the model descriptions stored in the KB.
\item The generated \emph{knowledge samples} are stored in the LDM along with traditional information.
\end{itemize}
The KB is responsible for knowledge creation, and knowledge retrieval when the knowledge is remotely created. When creating knowledge, the KB looks up in the LDM for the necessary input items. Then, if the required model bytecode is stored locally, it is executed immediately. The KB also communicates with other nodes' KBs to reach for non locally-stored bytecodes. This allows for a flexible framework, allowing to expand or limit the creation and distribution of knowledge within a certain group of vehicles as needed. The modalities for such remote knowledge creation will be further developed in the knowledge distribution section.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{img/knowledge_storage.pdf}
\caption{On-Board Storage of Knowledge and Information}
\label{fig:knowledge_storage}
\end{figure}
\subsection{Knowledge Distribution}
As we separate knowledge models' bytecodes and their description, a structural need appears for the distribution of knowledge. As shown in Figure~\ref{fig:knowledge_layer_networking}, nodes of the vehicular network are inter-connected and knowledge creation may be the product of a cooperation between multiple nodes.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{img/knowledge_networking_toyota.pdf}
\caption{Architecture of Vehicular Knowledge Networking}
\label{fig:knowledge_layer_networking}
\end{figure}
The input information or \emph{knowledge samples} required for knowledge creation through a knowledge model are spread across the vehicular network. As a consequence, centralizing all the input content required for knowledge creation in a single node can be inefficient. VKN makes knowledge models mutually understandable. It brings the opportunity not only to share models themselves, but also to outsource knowledge creation to remote nodes fitted with pertinent input. By cooperatively creating knowledge among well-chosen nodes, the amount of potentially heavyweight model inputs transmitted within the vehicular network can be reduced, saving time and load.
In order to achieve knowledge distribution, a \emph{Vehicular Knowledge Querying Language} (VKQL) must be defined to:
\begin{enumerate}
\item Lookup and retrieve the available knowledge models descriptions, their required input, output and preconditions.
\item Spread knowledge model bytecodes efficiently.
\item Delegate knowledge creation to remote nodes.
\end{enumerate}
An example of the delegation of knowledge creation to remote nodes and its potential benefits is developed in Section~\ref{sec:application}, through the example of passenger comfort-based rerouting.
Finally, a balance must be found between widespread and too scarce distribution of bytecodes. The former would increase load in the network and negate the point of separating descriptions and bytecodes, while the latter would make access to bytecodes too difficult, increasing delay. Mechanisms to cache knowledge close to the vehicles could make use of technologies such as edge computing units~\cite{mach2017} or vehicular micro-clouds~\cite{VC5}.
\section{Application: Passenger Comfort-based rerouting}
\label{sec:application}
As an application of VKN and to show potential benefits for vehicular networks, we investigate on the use case of \emph{passenger comfort-based rerouting}.
The scenario is the following: a Highly-Automated Vehicle (HAV) $V_{ego}$ seeks to provide the most \emph{comfortable} driving itinerary to its passengers.
\subsection{Comfort Knowledge Model}
Under certain driving situations, as surveyed by Xin~\emph{et al.}~\cite{xin2019}, a HAV's autonomous driving ability can be challenged. In somes cases, it can lead to a take over by a human driver, which can be regarded as a factor of discomfort for passengers.
Based on those factors and as an example, we defined a simplified knowledge model \texttt{model.env\_comfort} to determine the level of \emph{passenger comfort} in a given area, as introduced in the bottom left corner of Figure~\ref{fig:archi_applied_comfort}. The model reads a set of area-related input with semantically-defined names:
\begin{enumerate}
\item The current traffic conditions,\\$tr \in $ \texttt{Road.Traffic} $\equiv$ [\texttt{FLUID}, \texttt{CONGESTED}].
\item The visibility in the area, $v \in $ \texttt{Road.Visibility} $\equiv$ [\texttt{CLEAR}, \texttt{OBSTRUCTED}].
\item The concentration of two-wheelers in the surroundings, $c_{tw} \in $ \texttt{TwoWheelers.Concentration} $\equiv$ [\texttt{HIGH}, \texttt{MEDIUM}, \texttt{LOW}].
\end{enumerate}
Based on this input, the model returns a discrete qualification of the level of comfort associated with driving in the area, as $cft \in $ \texttt{Road.ComfortLevel} $\equiv$ [\texttt{GOOD}, \texttt{FAIR}, \texttt{POOR}].
To demonstrate what a simplified version of a knowledge bytecode would be, we provide a simple pseudocode implementation of the model in Algorithm~\ref{algo:the_algo}.
\begin{algorithm}[H]
\small
\caption{Simplified algorithm to compute comfort from environmental parameters}
\label{algo:the_algo}
\hspace*{\algorithmicindent} \textbf{Input} \\
\hspace*{\algorithmicindent} $c_{tw} \in $ [\texttt{HIGH}, \texttt{MEDIUM}, \texttt{LOW}] \\
\hspace*{\algorithmicindent} $v \in $ [\texttt{CLEAR}, \texttt{OBSTRUCTED}] \\
\hspace*{\algorithmicindent} $tr \in $ [\texttt{FLUID}, \texttt{CONGESTED}] \\
\hspace*{\algorithmicindent} \textbf{Output} \\
\hspace*{\algorithmicindent} $cft \in $ [\texttt{GOOD}, \texttt{FAIR}, \texttt{POOR}]
\begin{algorithmic}[1]
\IF{$c_{tw}$ = \texttt{LOW} \AND $v =$ \texttt{CLEAR} \AND $tr$ = \texttt{FLUID}}
\STATE $cft \leftarrow$ \texttt{GOOD}
\ELSIF{$c_{tw}$ = \texttt{HIGH}}
\STATE $cft \leftarrow$ \texttt{POOR}
\ELSE
\STATE $cft \leftarrow$ \texttt{FAIR}
\ENDIF
\end{algorithmic}
\end{algorithm}
\subsection{Comfort Retrieval Scenario}
The vehicle $V_{ego}$ has three itinerary options to reach its destination: crossing remote area $A$, $B$, or $C$. In order to compute the most comfortable itinerary for its passengers and take a decision on which route to take, $V_{ego}$ wishes to obtain a \texttt{Road.ComfortLevel} knowledge sample for each of these areas. Figure~\ref{fig:glob_scenario} illustrates the problem of the scenario.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{img/reroutescenario.pdf}
\caption{Illustration of the Comfort-based Rerouting Problem.}
\label{fig:glob_scenario}
\end{figure}
Once $V_{ego}$ has obtained the comfort level on each of the areas $A$, $B$, and $C$, it is able to take an educated decision on what itinerary to follow. As an illustration, we describe the process of retrieving the driving comfort level from Area $A$. The problem is as follows: $V_{ego}$ is currently driving $N \geq 5km$ away from Area $A$, and wishes to obtain the driving comfort level in $A$ prior to crossing it, in order to potentially reroute beforehand.
In a traditional, information-centric approach lacking standards for knowledge exchange, a typical solution to this problem would be as follows:
\begin{enumerate}
\setcounter{enumi}{-1}
\item The comfort knowledge model is downloaded to $V_{ego}$.
\item $V_{ego}$ transmits an information request via ICN to the target area.
\item The information acquisition is sensed by a node in the target area and retrieved to $V_{ego}$.
\item The remote comfort level is computed on-board $V_{ego}$.
\end{enumerate}
The input information is typically heavier than the produced knowledge. As a consequence, this approach increases load and delay in vehicular networks compared with a direct transfer of the knowledge.
In a novel solution, making use of VKN, we suggest to outsource the production of knowledge to a node in the target area. As shown in Figure~\ref{fig:scenario}, instead of transmitting a request for input information to nodes in the target area, $V_{ego}$ directly transmits a request for the comfort level knowledge in the area. The knowledge creation request is then forwarded to a node in the target area in possession of the required \texttt{model.env\_comfort} model's bytecode. In turn, the remote node locally computes the comfort knowledge using input information stored in its LDM and obtained through sensors. After the knowledge has been produced, it is directly returned to $V_{ego}$. Knowledge is typically much lighter than the input it was extracted from. In turn, this allows to save load and delay in vehicular networks.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{img/summary_scenario.pdf}
\caption{Summary of the Comfort Level Retrieval in a Remote Area using VKN}
\label{fig:scenario}
\end{figure}
\section{Research Applicability}
\label{sec:opportunities}
We described an application of cooperative \emph{knowledge samples} creation. By remotely running a knowledge model's bytecode in the area where its input information is sourced, unnecessary transfers of information are avoided to the benefit of knowledge.
Similarly, mechanisms have been defined in the literature to train \emph{knowledge models} themselves while avoiding the transmission of training information for privacy and efficiency concerns. Federated Machine Learning (FML) is an open research topic in which multiple nodes cooperatively train and update a shared model without exchanging actual training information~\cite{fml2019}. A centralized service provider selects a set of client nodes to cooperatively train a model. A current state of the model to be trained is downloaded to each client. Then, each client locally performs a few training steps of the model with local information before sending the updated model state back to the server. The server then aggregates the updated client model states to compute a global update of the model, taking advantage of the local training of each client.
An issue of this approach is that it is not trivial to ensure that all local nodes are interested in training and using the same model. Before being able to start the training, local nodes should be able to determine who among their neighbors is in possession of what type of model and has access to what kind of information.
We believe this to be a pertinent use case for VKN, as a protocol of discovery of the knowledge already possessed by one's neighbor is needed to initialize a session of FML.
\section{Conclusion}
\label{sec:conclusion}
Vehicular networks have been extensively studied in the past years. Several standards have been developed to store and share information. However, challenges remain to transition from an information-centric networking model to a model where common standards for knowledge characterization, description, storage and sharing allow nodes in vehicular networks to take full advantage of data-driven AI techniques. In this paper, using a common definition of knowledge, we determined under what forms it exists in vehicular networks, allowing us to concretely propose a structure for knowledge description, storage and sharing. Through a passenger comfort-based rerouting application, we exemplified the concept and showed potential performance improvements. Finally, we note the potential benefits of Vehicular Knowledge Networking for the open topic of Federated Machine Learning. Future work will focus on implementing, simulating and measuring the benefits of using VKN, with regards to traditional information-centric models.
\balance
\printbibliography
\end{document} | 755cef4a6c5ce02400dc8764beba86618b7e6ab0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In practice, one often encounters heterogeneous data, whose distribution is not constant, but depends on certain covariates. For example, data can be collected from several different sources, its distribution might differ across certain subpopulations or it could even change with time, etc. Inferring valid conclusions about a certain target of interest from such data can be very challenging as many different aspects of the distribution could potentially change. As an example, in medical studies, the effectiveness of a certain treatment might not be constant throughout the population but depend on certain patient characteristics such as age, race, gender, or medical history. Another issue could be that different patient groups were not equally likely to receive the same treatment in the observed data.
Obviously, pooling all available data together can result in invalid conclusions. On the other hand, if for a given test point of interest one only considers similar training data points, i.e.\ a small homogeneous subpopulation, one may end up with too few samples for accurate statistical estimation. In this paper, we propose a method based on the Random Forest algorithm \citep{breiman2001random} which in a data-adaptive way determines for any given test point which training data points are relevant for it. This in turn can be used for drawing valid conclusions or for accurately estimating any quantity of interest.
Let $\bold{Y}=(Y_1, Y_2, \ldots, Y_d) \in \mathbb{R}^{d}$ be a multivariate random variable representing the data of interest, but whose joint distribution is heterogeneous and depends on some subset of a potentially large number of covariates $\bold{X}=(X_1, X_2, \ldots, X_p) \in \mathbb{R}^{p}$. Throughout the paper, vector quantities are denoted in bold. We aim to estimate a certain target object $\tau(\bold{x})$ that depends on the conditional distribution $\P(\bold{Y} \mid \bold{X} \myeq \bold{x}) = \P(\bold{Y} \mid X_1 \myeq x_1, \ldots, X_p \myeq x_p)$, where $\bold{x} = (x_1, \ldots, x_p)$ is an arbitrary point in $\mathbb{R}^{p}$. The estimation target $\tau(\bold{x})$ can range from simple quantities, such as the conditional expectations $\E[f(\bold{Y}) \mid \bold{X}]$ \citep{breiman2001random} or quantiles $Q_\alpha[f(\bold{Y}) \mid \bold{X}]$ \citep{meinshausen2006quantile} for some function $f:\R^d \to \R$, to some more complicated aspects of the conditional distribution $\P(\bold{Y} \mid \bold{X} \myeq \bold{x})$, such as conditional copulas or conditional independence measures. Given the observed data $\{(\bold{x}_i,\bold{y}_i)\}_{i=1}^n$, the most straightforward way of estimating $\tau(\bold{x})$ nonparametrically would be to consider only the data points in some neighborhood $\mathcal{N}_{\bold{x}}$ around $\bold{x}$, e.g.\ by considering the $k$ nearest neighbors according to some metric. However, such methods typically suffer from the curse of dimensionality even when $p$ is only moderately large: for a reasonably small neighborhood, such that the distribution $\P(\bold{Y}\mid\bold{X}\in \mathcal{N}_\bold{x})$ is close to the distribution $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$, the number of training data points contained in $\mathcal{N}_{\bold{x}}$ will be very small, thus making the accurate estimation of the target $\tau(\bold{x})$ difficult. The same phenomenon occurs with other methods which locally weight the training observations such as kernel methods \citep{silverman1986density}, local MLE \citep{fan1998local} or weighted regression \citep{cleveland1979robust} even for the relatively simple problem of estimating the conditional mean $\E[\bold{Y}\mid\bold{X}\myeq\bold{x}]$ for fairly small $p$. For that reason, more importance should be given to the training data points $(\bold{x}_i, \bold{y}_i)$ for which the response distribution $\P(\bold{Y}\mid\bold{X}\myeq\bold{x}_i)$ at point $\bold{x}_i$ is similar to the target distribution $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$, even if $\bold{x}_i$ is not necessarily close to $\bold{x}$ in every component.
In this paper, we propose the Distributional Random Forest (DRF) algorithm which estimates the multivariate conditional distribution $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$ in a locally adaptive fashion. This is done by repeatedly dividing the data points in the spirit of the Random Forest algorithm \citep{breiman2001random}: at each step, we split the data points into two groups based on some feature $X_j$ in such a way that the distribution of $\bold{Y}$ for which $X_j \leq l$, for some level $l$, differs the most compared to the distribution of $\bold{Y}$ when $X_j > l$, according to some distributional metric. One can use any multivariate two-sample test statistic, provided it can detect a wide variety of distributional changes. As the default choice, we propose a criterion based on the Maximal Mean Discrepancy (MMD) statistic \citep{gretton2007kernel} with many interesting properties.
This splitting procedure partitions the data such that the distribution of the multivariate response $\bold{Y}$ in the resulting leaf nodes is as homogeneous as possible, thus defining neighborhoods of relevant training data points for every $\bold{x}$. Repeating this many times with randomization induces a weighting function $w_{\bold{x}}(\bold{x}_i)$ as in \citet{lin2002random, lin2006random}, described in detail in Section \ref{method_description}, which quantifies the relevance of each training data point $\bold{x}_i$ for a given test point $\bold{x}$. The conditional distribution is then estimated by an empirical distribution determined by these weights \citep{meinshausen2006quantile}. This construction is data-adaptive as it assigns more weight to the training points $\bold{x}_i$ that are closer to the test point $\bold{x}$ in the components which are more relevant for the distribution of $\bold{Y}$.
Our forest construction does not depend on the estimation target $\tau(\bold{x})$, but it rather estimates the conditional distribution $\P(\bold{Y} \mid \bold{\mathbf{X}=\mathbf{x}})$ directly and the induced forest weights can be used to estimate $\tau(\bold{x})$ in a second step. This approach has several advantages. First, only one DRF fit is required to obtain estimates of many different targets, which has a big computational advantage. Furthermore, since those estimates are obtained from the same forest fit, they are mutually compatible. For example, if the conditional correlation matrix $\{\Cor(Y_i,\, Y_j \mid \bold{X}\myeq\bold{x})\}_{i,j=1}^d$ were estimated componentwise, the resulting matrix might not be positive semidefinite, and as another example, the CDF estimates $\hat{\P}(\bold{Y} \leq \bold{y} \mid \bold{X}\myeq\bold{x})$ might not be monotone in $\bold{y}$, see Figure \ref{comparison}. Finally, it can be extremely difficult to tailor forest construction to more complex targets $\tau(\bold{x})$.
The induced weighting function can thus be used not only for obtaining simple distributional aspects such as, for example, the conditional quantiles, conditional correlations, or joint conditional probability statements, but also to obtain more complex objectives, such as conditional independence tests \citep{zhang2012kernel}, heterogeneous regression (see also Section
\ref{sec: causality} for more details) \citep{kunzel2019metalearners, wager2018estimation} or semiparametric estimation by fitting a parametric model for $\bold{Y}$, having nonparametrically adjusted for $\bold{X}$ \citep{bickel1993efficient}. Representation of the conditional distribution via the weighting function has a great potential for applications in causality such as causal effect estimation or as a way of implementing do-calculus \citep{pearl2009causality} for finite samples, as we discuss in Section \ref{sec: causality}.
Therefore, DRF is used in two steps: in the first step, we obtain the weighting function $w_\bold{x}(\cdot)$ describing the conditional distribution $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$ in a target- and model-free way, which is then used as an input for the second step. Even if the method used in the second step does not directly support weighting of the training data points, one can easily resample the data set with sampling probabilities equal to $\{w_\bold{x}(\bold{x}_i)\}_{i=1}^n$. The two-step approach is visualized in the following diagram:
\begin{center}
\hspace{2cm}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=2em, column sep=12em,minimum width=2em]
{
\P(\bold{Y}\mid\bold{X}\myeq\bold{x}) & \hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x}) \\
\tau(\P) & \tau(\hat{\P}) \\};
\path[-stealth]
(m-1-1) edge node [above] {1) get $w_\bold{x}(\cdot)$ with DRF} (m-1-2)
(m-1-1) edge [dashed] node [left] {objective} (m-2-1)
(m-1-2) edge node [right]{2) compute from $w_\bold{x}(\cdot)$} (m-2-2)
(m-2-2) edge [dashed] node [above] {induced estimator} (m-2-1);
\end{tikzpicture}
\end{center}
\subsection{Related work and our contribution}
Several adaptations of the Random Forest algorithm have been proposed for targets beyond the original one of the univariate conditional mean $\E[Y \mid \bold{\mathbf{X}\myeq\mathbf{x}}]$: for survival analysis \citep{hothorn2006survival}, conditional quantiles \citep{meinshausen2006quantile}, density estimation \citep{pospisil2018rfcde}, CDF estimation \citep{hothorn2017transformation} or heterogeneous treatment effects \citep{wager2018estimation}. Almost all such methods use the weights induced by the forest, as described in Section \ref{method_description}, rather than averaging the estimates obtained per tree. This view of Random Forests as a powerful adaptive nearest neighbor method is well known and dates back to \citet{lin2002random, lin2006random}. It was first used for targets beyond the conditional mean in \citet{meinshausen2006quantile}, where the original forest construction with univariate $Y$ was used \citep{breiman2001random}. However, the univariate response setting considered there severely restricts the number of interesting targets $\tau(\bold{x})$ and our DRF can thus be viewed as an important generalization of this approach to the multivariate setting.
In order to be able to perform certain tasks or to achieve a better accuracy, many forest-based methods adapt the forest construction by using a custom splitting criterion tailored to their specific target, instead of relying on the standard CART criterion.
In \citet{zeileis2008model} and \citet{hothorn2017transformation}, a parametric model for the response $\bold{Y} \mid \bold{\mathbf{X}\myeq\mathbf{x}} \sim f(\theta(\bold{x}), \cdot)$ is assumed and recursive splitting is performed based on a permutation test which uses the user-provided score functions. Similarly, \citet{athey2019generalized} estimate certain univariate targets for which there exist corresponding score functions defining the local estimating equations. The data is split so that the estimates of the target in resulting child nodes differ the most. This is different, though, to the target-free splitting criterion of DRF, which splits so that the distribution of $\bold{Y}$ in child nodes is as different as possible.
Since the splitting step is extensively used in the algorithm, its complexity is crucial for the overall computational efficiency of the method, and one often needs to resort to approximating the splitting criterion \citep{pospisil2018rfcde, athey2019generalized} to obtain good computational run time. We propose a splitting criterion based on a fast random approximation of the MMD statistic \citep{gretton2012kernel, zhao2015fastmmd}, which is commonly used in practice for two-sample testing as it is able to detect any change in the multivariate distribution of $\bold{Y}$ with good power \citep{gretton2007kernel}. DRF with the MMD splitting criterion also has interesting theoretical properties as shown in Section \ref{sec: theory} below.
The multivariate response case has not received much attention in the Random Forest literature. Most of the existing forest-based methods focus on either a univariate response $Y$ or on a certain univariate target $\tau(\bold{x})$. One interesting line of work considers density estimation \citep{pospisil2018rfcde} and uses aggregation of the CART criteria for different response transformations. Another approach \citep{kocev2007ensembles, segal2011multivariate} is based on aggregating standard univariate CART splitting criteria for $Y_1, \ldots, Y_d$ and targets only the conditional mean of the responses, a task which could also be solved by separate regression fits for each $Y_i$. In order to capture any change in the distribution of the multivariate response $\bold{Y}$, one needs to not only consider the marginal distributions for each component $Y_i$, but also to determine whether their dependence structure changes, see e.g.\ Figure \ref{fig: conditional-independence}.
There are not many methods that nonparametrically estimate the joint multivariate conditional distribution $\P(\bold{Y} \mid \bold{X}\myeq\bold{x})$ in the statistics or machine learning literature. Other than a few simple classical methods such as $k$-nearest neighbors and kernel regression, there are methods based on normalizing flows such as Inverse Autoregressive Flow \citep{kingma2016improving} or Masked Autoregressive Flow \citep{papamakarios2017masked} and also conditional variants of several popular generative models such as Conditional Generative Adversarial Networks \citep{mirza2014conditional} or Conditional Variational Autoencoder \citep{sohn2015learning}. The focus of these methods is more on the settings with large $d$ and small $p$, such as image or text generation. The comparison of DRF with the competing methods for distributional estimation can be found in Section \ref{sec: benchmarks}.
Our contribution, resulting in the proposal of the Distributional Random Forest (DRF), can be summarized as follows: First, we introduce the idea of forest construction based on sequential multivariate two-sample test statistics. It does not depend on a particular estimation target and is completely nonparametric, which makes its implementation and usage very simple and universal. Not only does it not require additional user input such as the log-likelihoods or score functions, but it can be used even for complicated targets for which there is no obvious forest construction. Furthermore, it has a computational advantage as only a single forest fit is needed for producing estimates of many different targets that are additionally compatible with each other.
Second, we propose an MMD-based splitting criterion with good statistical and computational properties, for which we also derive interesting theoretical results in Section \ref{sec: theory}. It underpins our implementation, which we provide as \texttt{R} and \texttt{Python} packages \texttt{drf}.
Finally, we show on a broad range of examples in Section \ref{sec: applications} how many different statistical estimation problems, some of which not being easily tractable by existing forest-based methods, can be cast to our framework, thus illustrating the usefulness and versatility of DRF.
\section{Method} \label{method_description}
In this section we describe the details of the Distributional Random Forest (DRF) algorithm. We closely follow the implementations of the \texttt{grf} \citep{athey2019generalized} and \texttt{ranger} \citep{wright2015ranger} \texttt{R}-packages. A detailed description of the method and its implementation and the corresponding pseudocode can be found in the Appendix \ref{appendix: impdetails}.
\subsection{Forest Building}
The trees are grown recursively in a model-free and target-free way as follows:
For every parent node $P$, we determine how to best split it into two child nodes of the form $C_L = \{X_j \leq l\}$ and $C_R = \{X_j > l\}$, where the variable $X_j$ is one of the randomly chosen splitting candidates and $l$ denotes its level based on which we perform the splitting. The split is chosen such that we maximize a certain (multivariate) two-sample test statistic
\begin{equation} \label{eq: splitcrit}
\mathcal{D}\left(\{\bold{y}_i\mid \bold{x}_i \in C_L\} \,,\, \{\bold{y}_i \mid \bold{x}_i \in C_R\}\right),
\end{equation}
which measures the difference of the empirical distributions of the data $\bold{Y}$ in the two resulting child nodes $C_L$ and $C_R$. Therefore, in each step we select the candidate predictor $X_j$ which seems to affect the distribution of $\bold{Y}$ the most, as measured by the metric $\mathcal{D}(\cdot,\cdot)$. Intuitively, in this way we ensure that the distribution of the data points in every leaf of the resulting tree is as homogeneous as possible, which helps mitigate the bias caused by pooling the heterogeneous data together. A related idea can be found in GRF \citep{athey2019generalized}, where one attempts to split the data so that the resulting estimates $\hat{\tau}_L$ and $\hat{\tau}_R$, obtained respectively from data points in $C_L$ and $C_R$, differ the most:
\begin{equation}\label{eq: GRF split}
\frac{n_Ln_R}{n_P^2} \left(\hat{\tau}_L - \hat{\tau}_R\right)^2,
\end{equation}
where we write $n_P = |\{i \mid \bold{x}_i \in P\}|$ and $n_L, n_R$ are defined analogously.
One could construct the forest using any metric $\mathcal{D}(\cdot,\cdot)$ for empirical distributions. However, in order to have a good accuracy of the overall method, the corresponding two-sample test using $\mathcal{D}(\cdot,\cdot)$ needs to have a good power for detecting any kind of change in distribution, which is a difficult task in general, especially for multivariate data \citep{bai1996effect, szekely2004testing}. Another very important aspect of the choice of distributional metric $\mathcal{D}(\cdot, \cdot)$ is the computational efficiency; one needs to be able to sequentially compute the values of $\mathcal{D}\left(\{\bold{y}_i\mid \bold{x}_i \in C_L\} \,,\, \{\bold{y}_i \mid \bold{x}_i \in C_R\}\right)$ for every possible split very fast for the overall algorithm to be computationally feasible, even for moderately large datasets. Below, we propose a splitting criterion based on the MMD two-sample test statistic \citep{gretton2007kernel} which has both good statistical and computational properties.
In contrast to other forest-based methods, we do not use any information about our estimation target $\tau$ in order to find the best split of the data, which comes with a certain trade-off. On one hand, it is sensible that tailoring the splitting criterion to the target should improve the estimation accuracy; for example, some predictors might affect the conditional distribution of $\bold{Y}$, but not necessarily the estimation target $\tau$ and splitting on such predictors unnecessarily reduces the number of training points used for estimating $\tau$. On the other hand, our approach has multiple benefits: it is easier to use as it does not require any user input such as the likelihood or score functions and it can also be used for very complicated targets for which one could not easily adapt the splitting criterion. Furthermore, only one DRF fit is necessary for producing estimates of many different targets, which has both computational advantage and the practical advantage that the resulting estimates are mutually compatible (see e.g.\ Figure \ref{fig: functionals}).
Interestingly, sometimes it could even be beneficial to split based on a predictor which does not affect the target of estimation, but which affects the conditional distribution. This is illustrated by the following toy example. Suppose that for a bivariate response $(Y_1, Y_2)$ we are interested in estimating the slope of the linear regression of $Y_2$ on $Y_1$ conditionally on $p=30$ predictors $\bold{X}$, i.e. our target is $\tau(\bold{x}) = \Cov(Y_1, Y_2 \mid \bold{X}\myeq\bold{x})/\Var(Y_1 \mid \bold{X}\myeq\bold{x})$. This is one of the main use cases for GRF and its variant which estimates this target is called Causal Forest \citep{wager2018estimation, athey2019generalized}. Let us assume that the data has the following distribution:
\begin{equation} \label{eq: toy_example}
\P\left(\begin{bmatrix}Y_1 \\ Y_2\end{bmatrix} \,\,\middle|\,\, \bold{X}\myeq\bold{x}\right) \sim N\left(\begin{bmatrix} x_1 \\ x_1 \end{bmatrix}, \begin{bmatrix}\sigma^2 &0\\ 0 &\sigma^2\end{bmatrix}\right) \qquad \bold{X} \sim N(\bold{0}, I_p),
\end{equation}
i.e. $X_1$ affects only the mean of the responses, while the other $p-1$ predictors have no effect. In Figure \ref{fig: toy_example} we illustrate the distribution of the data when $n=300, p=30, \sigma=0.2$, together with the DRF and GRF splitting criteria. The true value of the target is $\tau(\bold{x})=0$, but when $\sigma$ is not too big, the slope estimates $\hat{\tau}$ on pooled data will be closer to $1$. Therefore, the difference of $\hat{\tau}_L$ and $\hat{\tau}_R$ between the induced slope estimates for a candidate split, which is used for splitting criterion \eqref{eq: GRF split} of GRF, might not be large enough for us to decide to split on $X_1$, or the resulting split might be too unbalanced. This results in worse forest estimates for this toy example, see Figure \ref{fig: toy_example}.
\begin{figure}[h]
\hspace*{-0.5cm}
\includegraphics[width=1.04\textwidth]{toy_example2.png}
\caption{Top left: Illustration of data distribution for the toy example \eqref{eq: toy_example} when $n=300,\, p=30$. Bottom: The corresponding MMD \eqref{eq: MMD splitcrit} (left) and GRF \eqref{eq: GRF split} splitting criteria (right) at the root node. The curves of different colors correspond to different predictors, with $X_1$ denoted in black. Top right: Comparison of the estimates of DRF and Causal Forest \citep{athey2019generalized} which respectively use those splitting criteria. Test points were randomly generated from the same distribution as the training data. Black dashed line indicates the correct value of the target quantity.}
\label{fig: toy_example}
\end{figure}
\subsection{Weighting Function}
\label{sec: weighting function}
Having constructed our forest, just as the standard Random Forest \citep{breiman2001random} can be viewed as the weighted nearest neighbor method \citep{lin2002random}, we can use the induced weighting function to estimate the conditional distribution at any given test point $\bold{x}$ and thus any other quantity of interest $\tau(\bold{x})$. This approach is commonly used in various forest-based methods for obtaining predictions \citet{hothorn2017transformation, pospisil2018rfcde, athey2019generalized}.
Suppose that we have built $N$ trees $\mathcal{T}_1, \ldots, \mathcal{T}_N$. Let $\mathcal{L}_k(\bold{x})$ be the set of the training data points which end up in the same leaf as $\bold{x}$ in the tree $\mathcal{T}_k$. The weighting function $w_{\bold{x}}(\bold{x}_i)$ is defined as the average of the corresponding weighting functions per tree \citep{lin2006random}:
\begin{equation} \label{eq: weighting function}
w_{\bold{x}}(\bold{x}_i) = \frac{1}{N} \sum_{k=1}^N \frac{\1\left(\bold{x}_i \in \mathcal{L}_k(\bold{x})\right)}{|\mathcal{L}_k(\bold{x})|}.
\end{equation}
The weights are positive add up to $1$: $\sum_{i=1}^n w_{\bold{x}}(\bold{x}_i) = 1$. In the case of equally sized leaf nodes, the assigned weight to a training point $\bold{x}_i$ is proportional to the number of trees where the test point $\bold{x}$ and $\bold{x}_i$ end up in the same leaf node. This shows that forest-based methods can in general be viewed as adaptive nearest neighbor methods. The sets $\mathcal{L}_k(\bold{x})$ of DRF will contain data points $(\bold{x}_i, \bold{y}_i)$ such that $\P(\bold{Y}\mid \bold{X}=\bold{x}_i)$ is close to $\P(\bold{Y}\mid \bold{\mathbf{X}=\mathbf{x}})$, thus removing bias due to heterogeneity of $\bold{Y}$ caused by $\bold{X}$. On the other hand, since the trees are constructed randomly and are thus fairly independent \citep{breiman2001random}, the leaf sets $\mathcal{L}_k(\bold{x})$ will be different enough so that the induced weights $w_{\bold{x}}(\bold{x}_i)$ are not concentrated on a small set of data points, which would lead to high estimation variance. Such good bias-variance tradeoff properties of forest-based methods are also implied by their asymptotic properties \citep{biau2012analysis, wager2014asymptotic}, even though this is a still active area of research and not much can be shown rigorously.
One can estimate the conditional distribution $\P(\bold{Y}\mid\bold{X}=\bold{x})$ from the weighting function by using the corresponding empirical distribution:
\begin{equation} \label{eq: distributional estimate}
\hat{\P}(\bold{Y} \mid \bold{X}=\bold{x}) = \sum_{i=1}^n w_{\bold{x}}(\bold{x}_i)\cdot\delta_{\bold{y}_i},
\end{equation}
where $\delta_{\bold{y}_i}$ is the point mass at $\bold{y}_i$.
The weighting function $w_{\bold{x}}(\bold{x}_i)$ can directly be used for any target $\tau(\bold{x})$ in a second step and not just for estimating the conditional distribution. For example, the estimated conditional joint CDF is given by
\begin{equation}\label{eq: CDF estimate}
\hat{F}_{\bold{Y}\mid\bold{X}\myeq\bold{x}}(\bold{t}) = \hat{\P}(Y_1 \leq t_1, \ldots, Y_d \leq t_d \mid \bold{X}\myeq\bold{x}) = \sum_{i=1}^n w_\bold{x}(\bold{x}_i) \1(\cap_{j=1}^d \{(\bold{y}_i)_j \leq t_j\}).
\end{equation}
It is important to point out that using the induced weighting function for locally weighted estimation is different than the approach of averaging the noisy estimates obtained per tree \citep{wager2018estimation}, originally used in standard Random Forests \citep{breiman2001random}. Even though the two approaches are equivalent for conditional mean estimation, the former approach is often much more efficient for more complex targets \citep{athey2019generalized}, since the number of data points in a single leaf is very small, leading to large variance of the estimates.
For the univariate response, the idea of using the induced weights for estimating targets different than the original target of conditional mean considered in \citet{breiman2001random} dates back to Quantile Regression Forests (QRF) \citep{meinshausen2006quantile}, where a lot of emphasis is put on the quantile estimation, as the number of interesting targets is quite limited in the univariate setting.
In the multivariate case, on the other hand, many interesting quantities such as, for example, conditional quantiles, conditional correlations or various conditional probability statements can easily be directly estimated from the weights.
By using the weights as an input for some other method, we can accomplish some more complicated objectives, such as conditional independence testing, causal effect estimation, semiparametric learning, time series prediction or tail-index estimation in extreme value analysis. As an example, suppose that our data $\bold{Y}$ come from a certain parametric model, where the parameter $\theta$ is not constant, but depends on $\bold{X}$ instead, i.e. $\bold{Y}\mid\bold{\mathbf{X}=\mathbf{x}} \sim f(\theta(\bold{x}), \cdot)$, see also \citet{zeileis2008model}. One can then estimate the parameter $\theta(\bold{x})$ by using weighted maximum likelihood estimation: $$\hat{\theta}(\bold{x}) = \argmax_{\theta \in \Theta} \sum_{i=1}^n w_\bold{x}(\bold{x}_i) \log f(\theta, \bold{y}_i).$$
Another example is heterogeneous regression, where we are interested in the regression fit of an outcome $Y \in \R$ on certain predicting variables $\bold{W}\in \R^s$ conditionally on some event $\{\bold{X} = \bold{x}\}$. This can be achieved by weighted regression of $Y$ on $\bold{W}$, where the weight $w_\bold{x}(\bold{x}_i)$ assigned to each data point $(\bold{w}_i, y_i)$ is obtained from DRF with the multivariate response $(Y, \bold{W}) \in \R^{s+1}$ and predictors $\bold{X} \in \R^p$, for an illustration see Section \ref{sec: causality}.
\begin{figure}[h]
\hspace*{-0.4cm}
\includegraphics[width=1.03\textwidth]{visualization.png}
\caption{Top: the characteristics of the important training sites, for a fixed test site whose position is indicated by a black star and whose characteristics are indicated in the title. The total weight assigned corresponds to the symbol size. Bottom: estimated joint conditional distribution of two pollutants NO$_2$ and PM$2.5$, where the weights correspond to the transparency of the data points. Green area corresponds to 'Good' air quality category ($\text{AQI} \leq 50$).}
\label{air_data}
\end{figure}
The weighting function of DRF is illustrated on the air quality data in Figure \ref{air_data}. Five years ($2015-2019$) of air pollution measurements were obtained from the US Environmental Protection Agency (EPA) website. Six main air pollutants (nitrogen dioxide (NO$_2$), carbon monoxide (CO), sulphur dioxide (SO$_2$), ozone (O$_3$) and coarse and fine particulate matter (PM$10$ and PM$2.5$)) that form the air quality index (AQI) were measured at many different measuring sites in the US for which we know the longitude, latitude, elevation, location setting (rural, urban, suburban) and how the land is used within a $1/4$ mile radius. Suppose we would want to know the distribution of the pollutant measurements at some new, unobserved, measurement site. The top row illustrates for a given test site, whose characteristics are indicated in the plot title, how much weight in total is assigned to the measurements from a specific training site. We see that the important sites share many characteristics with the test site and that DRF determines the relevance of each characteristic in a data-adaptive way. The bottom row shows the corresponding estimates of the joint conditional distribution of the pollutants (we choose $2$ of them for visualization purposes) and one can clearly see how the estimated pollution levels are larger for the suburban site than for the rural site. The induced weights can be used, for example, for estimating the joint density (whose contours can be seen in the plot) or for estimating the probability that the AQI is below a certain value by summing the weights in the corresponding region of space.
\subsection{Distributional Metric}
In order to determine the best split of a parent node $P$, i.e. such that the distributions of the responses $\bold{Y}$ in the resulting child nodes $C_L$ and $C_R$ differ the most, one needs a good distributional metric $\mathcal{D}(\cdot,\cdot)$ (see Equation \eqref{eq: splitcrit}) which can detect change in distribution of the response $\bold{Y}$ when additionally conditioning on an event $\{X_j > l\}$. Testing equality of distributions from the corresponding samples is an old problem in statistics, called two-sample testing problem. For univariate data, many good tests exist such as Wilcoxon rank test \citep{wilcoxon1946individual}, Welch's t-test \citep{welch1947generalization}, Wasserstein two-sample testing \citep{ramdas2017wasserstein}, Kolmogorov-Smirnov test \citep{massey1951kolmogorov} and many others, but obtaining an efficient test for multivariate distributions has proven to be quite challenging due to the curse of dimensionality \citep{friedman1979multivariate, baringhaus2004new}.
Additional requirement for the choice of distributional metric $\mathcal{D}(\cdot,\cdot)$ used for data splitting is that it needs to be computationally very efficient as splitting is used extensively in the algorithm. If we construct $N$ trees from $n$ data points and in each node we consider $\text{mtry}$ candidate variables for splitting, the complexity of the standard Random Forest algorithm \citep{breiman2001random} in the univariate case is $\O(N \times \text{mtry} \times n \log n)$ provided our splits are balanced. It uses the CART splitting criterion, given by:
\begin{equation} \label{eq: CART}
\frac{1}{n_P}\left(\sum_{\bold{x}_i \in C_L} (y_i - \overline{y}_L)^2 + \sum_{\bold{x}_i \in C_R} (y_i - \overline{y}_R)^2 \right),
\end{equation}
where $\overline{y}_L = \tfrac{1}{n_L}\sum_{\bold{x}_i \in C_L} y_i$ and $\overline{y}_R$ is defined analogously. This criterion has an advantage that not only it can be computed in $\O(n_P)$ complexity, but this can be done for all possible splits $\{X_j \leq l\}$ as cutoff level $l$ varies, since updating the splitting criterion when moving a single training data point from one child node to the other requires only $\O(1)$ computational steps (most easily seen by rewriting the CART criterion as in \eqref{eq: CART equivalent}).
If the time complexity of evaluating the DRF splitting criterion \eqref{eq: splitcrit} for a single splitting candidate $X_j$ and all cutoffs $l$ of interest (usually taken to range over all possible values) is at least $n^c$ for some $c>1$, say $\O(f(n_P))$ for some function $f:\R \to \R$, then by solving the recursive relation we obtain that the overall complexity of the method is given by $\O(N \times \text{mtry} \times f(n))$ \citep{akra1998solution}, which can be unfeasible even for moderately large $n$ if $f$ grows too fast.
The problem of sequential two-sample testing is also central to the field of change-point detection \citep{wolfe1984nonparametric, brodsky2013nonparametric}, with the slight difference that in the change-point problems the distribution is assumed to change abruptly at certain points in time, whereas for our forest construction we only are interested in finding the best split of the form $\{X_j \leq l\}$ and the conditional distribution $\P(\bold{Y}\mid \{\bold{X}\in P\} \cap\{X_j \leq l\})$ usually changes gradually with $l$. The testing power and the computational feasibility of the method play a big role in change-point detection as well. However, the state-of-the-art change-point detection algorithms \citep{li2015scan, matteson2014nonparametric} are often too slow for our purpose as sequential testing is done $\O(N \times \text{mtry} \times n)$ times for forest construction, much more frequently than in change-point problems.
\subsubsection{MMD splitting criterion}
\label{sec: MMD splitcrit}
Even though DRF could in theory be constructed with any distributional metric $\mathcal{D}(\cdot, \cdot)$, as a default choice we propose splitting criterion based on the Maximum Mean Discrepancy (MMD) statistic \citep{gretton2007kernel}. Let $(\mathcal{H}, \langle\cdot,\cdot\rangle_\mathcal{H})$ be the RKHS of real-valued functions on $\R^d$ induced by some positive-definite kernel $k$, and let $\varphi:\R^d \to \mathcal{H}$ be the corresponding feature map satisfying that $k(\bold{u}, \bold{v}) = \langle\varphi(\bold{u}), \varphi(\bold{v})\rangle_\mathcal{H}$.
The MMD statistic $\mathcal{D}_{\text{MMD}(k)}\left(U, V\right)$ for kernel $k$ and two samples $U = \{\bold{u}_1, \ldots, \bold{u}_{|U|}\}$ and $V = \{\bold{v}_1, \ldots, \bold{v}_{|V|}\}$ is given by:
\begin{equation} \label{eq: MMD}
\mathcal{D}_{\text{MMD}(k)}\left(U, V\right) = \frac{1}{|U|^2}\sum_{i,j=1}^{|U|} k(\bold{u}_i, \bold{u}_j) + \frac{1}{|V|^2}\sum_{i,j=1}^{|V|} k(\bold{v}_i, \bold{v}_j)
- \frac{2}{|U||V|}\sum_{i=1}^{|U|}\sum_{j=1}^{|V|} k(\bold{u}_i, \bold{v}_j).
\end{equation}
MMD compares the similarities, described by the kernel $k$, within each sample with the similarities across samples and is commonly used in practice for two-sample testing. It is based on the idea that one can assign to each distribution $\mathcal{P}$ its embedding $\mu(\mathcal{P})$ into the RKHS $\mathcal{H}$, which is the unique element of $\mathcal{H}$ given by
\begin{equation} \label{eq: RKHS embedding}
\mu(\mathcal{P}) = \E_{\bold{Y} \sim \mathcal{P}}[\varphi(\bold{Y})].
\end{equation}
The MMD two-sample statistic \eqref{eq: MMD} can then can then equivalently be written as the squared distance between the embeddings of the empirical distributions with respect to the RKHS norm $\norm{\cdot}_\mathcal{H}$:
\begin{equation} \label{eq: MMD embedding}
\mathcal{D}_{\text{MMD}(k)}\left(U, V\right) = \norm*{\mu\left(\frac{1}{|U|}\sum_{i=1}^{|U|} \delta_{\bold{u}_i}\right) - \mu\left(\frac{1}{|V|}\sum_{i=1}^{|V|} \delta_{\bold{v}_i}\right)}_\mathcal{H}^2,
\end{equation}
recalling that $\delta_{\bold{y}}$ is the point mass at $\bold{y}$.
As the sample sizes $|U|$ and $|V|$ grow, the MMD statistic \eqref{eq: MMD embedding} converges to its population version, which is the squared RKHS distance between the corresponding embeddings of the data-generating distributions of $U$ and $V$. Since the embedding map $\mu$ is injective for characteristic kernel $k$, we see that MMD is able to detect any difference in the distribution.
Even though the power of the MMD two sample test also deteriorates as the data dimensionality grows, since the testing problem becomes intrinsically harder \citep{reddi2014decreasing}, it still has good empirical power compared to other multivariate two-sample tests for a wide range of $k$ \citep{gretton2012kernel}.
However, the $\O((|U|+|V|)^2)$ complexity for computing $\mathcal{D}_{\text{MMD}(k)}(U, V)$ from \eqref{eq: MMD} is too large for many applications. For that reason, several fast approximations of MMD have been suggested in the literature \citep{gretton2012kernel, zaremba2013b}. As already mentioned, the complexity of the distributional metric $\mathcal{D}(\cdot,\cdot)$ used for DRF is crucial for the overall method to be computationally efficient, since the splitting step is used extensively in the forest construction. We therefore propose splitting based on an MMD statistic computed with an approximate kernel $\tilde{k}$, which is also a fast random approximation of the original MMD \citep{zhao2015fastmmd}.
Bochner's theorem (see e.g.\ \citet[Theorem 6.6]{wendland_2004}) gives us that any bounded shift-invariant kernel can be written as
\begin{equation} \label{eq: bochner}
k(\bold{u}, \bold{v}) = \int_{\R^d} e^{i\boldsymbol{\omega}^T(\bold{u}-\bold{v})}d\nu(\boldsymbol{\omega}),
\end{equation}
i.e.\ as a Fourier transform of some measure $\nu$. Therefore, by randomly sampling the frequency vectors $\boldsymbol{\omega}_1, \ldots, \boldsymbol{\omega}_B$ from normalized $\nu$, we can approximate our kernel $k$ by another kernel $\tilde{k}$ (up to a scaling factor) as follows:
$$k(\bold{u}, \bold{v}) = \int_{\R^d} e^{i\boldsymbol{\omega}^T(\bold{u}-\bold{v})}d\nu(\boldsymbol{\omega}) \approx \frac{1}{B} \sum_{b=1}^B e^{i\boldsymbol{\omega}_b^T(\bold{u}-\bold{v})} = \tilde{k}(\bold{u}, \bold{v}),$$
where we define $\tilde{k}(\bold{u}, \bold{v}) = \langle \bold{\widetilde{\varphi}}(\bold{u}), \bold{\widetilde{\varphi}}(\bold{v})\rangle_{\mathbb{C}^B}$ as the kernel function with the feature map given by $$\bold{\widetilde{\varphi}}(\bold{u}) = \frac{1}{\sqrt{B}}\left(\tilde{\varphi}_{\boldsymbol{\omega}_1}(\bold{u}),
\ldots, \tilde{\varphi}_{\boldsymbol{\omega}_B}(\bold{u})\right)^T = \frac{1}{\sqrt{B}}\left(e^{i\boldsymbol{\omega}_1^T\bold{u}},
\ldots, e^{i\boldsymbol{\omega}_B^T\bold{u}}\right)^T,$$
which is a random vector consisting of the Fourier features $\widetilde{\varphi}_{\boldsymbol{\omega}}(\bold{u}) = e^{i\boldsymbol{\omega}^T\bold{u}} \in \mathbb{C}$ \citep{rahimi2008random}.
Such kernel approximations are frequently used in practice for computational efficiency \citep{rahimi2009weighted, le2013fastfood}. As a default choice of $k$ we take the Gaussian kernel with bandwidth $\sigma$, since in this case we have a convenient expression for the measure $\nu$ and we sample $\boldsymbol{\omega}_1, \ldots, \boldsymbol{\omega}_B \sim N_{d}(\bold{0}, \sigma^{-2}I_{d})$. The bandwidth $\sigma$ is chosen as the median pairwise distance between all training responses $\{\bold{y}_i\}_{i=1}^n$, commonly referred to as the 'median heuristic' \citep{gretton2012optimal}.
From the representation of MMD via the distribution embeddings \eqref{eq: MMD embedding} we can obtain that MMD two-sample test statistic $\mathcal{D}_{\text{MMD}(\tilde{k})}$ using the approximate kernel $\tilde{k}$ is given by
\begin{equation*}
\mathcal{D}_{\text{MMD}(\tilde{k})}\left(\{\bold{u}_i\}_{i=1}^{|U|}, \{\bold{v}_i\}_{i=1}^{|V|}\right) = \frac{1}{B}\sum_{b=1}^B\left\lvert \frac{1}{|U|}\sum_{i=1}^{|U|} \tilde{\varphi}_{\boldsymbol{\omega}_b}(\bold{u}_i) - \frac{1}{|V|}\sum_{i=1}^{|V|} \tilde{\varphi}_{\boldsymbol{\omega}_b}(\bold{v}_i) \right\rvert^2.
\end{equation*}
Interestingly, $\mathcal{D}_{\text{MMD}(\tilde{k})}$ is not only an MMD statistic on its own, but can also be viewed as a random approximation of the original MMD statistic $\mathcal{D}_{\text{MMD}(k)}$ \eqref{eq: MMD} using kernel $k$; by using the kernel representation \eqref{eq: bochner}, it can be written (the derivation can be found in the Appendix \ref{appendix: proofs}) as
\begin{equation*}
\mathcal{D}_{\text{MMD}(k)}\left(\{\bold{u}_i\}_{i=1}^{|U|}, \{\bold{v}_i\}_{i=1}^{|V|}\right) = \int_{\R^d}\left\lvert \frac{1}{|U|}\sum_{i=1}^{|U|} \tilde{\varphi}_{\boldsymbol{\omega}}(\bold{u}_i) - \frac{1}{|V|}\sum_{i=1}^{|V|} \tilde{\varphi}_{\boldsymbol{\omega}}(\bold{v}_i) \right\rvert^2 d\nu(\boldsymbol{\omega}).
\end{equation*}
Finally, our DRF splitting criterion $\mathcal{D}(\cdot, \cdot)$ \eqref{eq: splitcrit} is then taken to be the (scaled) MMD statistic $\tfrac{n_Ln_R}{n_P^2}\mathcal{D}_{\text{MMD}(\tilde{k})}\left(\{\bold{y}_i\mid \bold{x}_i \in C_L\} \,,\, \{\bold{y}_i \mid \bold{x}_i \in C_R\}\right)$ with the approximate random kernel $\tilde{k}$ used instead of $k$, which can thus be conveniently written as:
\begin{align} \label{eq: MMD splitcrit}
\frac{1}{B}\sum_{b=1}^B \frac{n_Ln_R}{n_P^2} \left\lvert \frac{1}{n_L}\sum_{\bold{x}_i \in C_L} \tilde{\varphi}_{\boldsymbol{\omega}_b}(\bold{y}_i) - \frac{1}{n_R}\sum_{\bold{x}_i \in C_R} \tilde{\varphi}_{\boldsymbol{\omega}_b}(\bold{y}_i) \right\rvert^2,
\end{align}
where we recall that $n_P = |\{i \mid \bold{x}_i \in P\}|$ and $n_L, n_R$ are defined analogously.
The additional scaling factor $\tfrac{n_Ln_R}{n_P^2}$ in \eqref{eq: MMD splitcrit} occurs naturally and compensates the increased variance of the test statistic for unbalanced splits; it also appears in the GRF \eqref{eq: GRF split} and CART (see representation \eqref{eq: CART equivalent}) splitting criteria.
The main advantage of the splitting criterion based on $\mathcal{D}_{\text{MMD}(\tilde{k})}$ is that by using the representation \eqref{eq: splitcrit} it can be easily computed for every possible splitting level $l$ in $\O(Bn_P)$ complexity, whereas the MMD statistic $\mathcal{D}_{\text{MMD}(k)}$ using kernel $k$ would require $\O(n_P^2)$ computational steps, which makes the overall complexity of the algorithm $\O\left(B \times N \times \text{mtry} \times n \log n\right)$ instead of much slower $\O\left(N \times \text{mtry} \times n^2\right)$.
We do not use the same approximate random kernel $\tilde{k}$ for different splits; for every parent node $P$ we resample the frequency vectors $\{\omega_b\}_{b=1}^B$ defining the corresponding feature map $\tilde{\varphi}$. Using different $\tilde{k}$ at each node might help to better detect different distributional changes. Furthermore, having different random kernels for each node agrees well with the randomness of the Random Forests and helps making the trees more independent. Since the MMD statistic $\mathcal{D}_{\text{MMD}(\tilde{k})}$ used for our splitting criterion is not only an approximation of $\mathcal{D}_{\text{MMD}(k)}$, but is itself an MMD statistic, it inherits good power for detecting any difference in distribution of $\bold{Y}$ in the child nodes for moderately large data dimensionality $d$, even when $B$ is reasonably small. One could even consider changing the number of random Fourier features $B$ at different levels of the tree, as $n_P$ varies, but for simplicity we take it to be fixed.
There is some similarity of our MMD-based splitting criterion \eqref{eq: MMD splitcrit} with the standard variance reduction CART splitting criterion \eqref{eq: CART} when $d=1$, which can be rewritten as:
\begin{equation} \label{eq: CART equivalent}
\frac{n_Ln_R}{n_P^2}\left(\frac{1}{n_L}\sum_{\bold{x}_i \in C_L} y_i - \frac{1}{n_R}\sum_{\bold{x}_i \in C_R} y_i \right)^2.
\end{equation}
The derivation can be found in Appendix \ref{appendix: proofs}.
From this representation, we see that the CART splitting criterion \eqref{eq: CART} is also equivalent to the GRF splitting criterion \eqref{eq: GRF split} when our target is the univariate conditional mean $\tau(\bold{x}) = \E[Y \mid \bold{X}\myeq\bold{x}]$ which is estimated for $C_L$ and $C_R$ by the sample means $\hat{\tau}_L = \overline{y}_L$ and $\hat{\tau}_R = \overline{y}_R$.
Therefore, as it compares the means of the univariate response $Y$ in the child nodes, the CART criterion can only detect changes in the response mean well, which is sufficient for prediction of $Y$ from $\bold{X}$, but might not be suitable for more complex targets. Similarly, for multivariate applications, aggregating the marginal CART criteria \citep{kocev2007ensembles, segal2011multivariate} across different components $Y_i$ of the response can only detect changes in the means of their marginal distributions. However, it is possible in the multivariate case that the pairwise correlations or the variances of the responses change, while the marginal means stay (almost) constant. For an illustration on simulated data, see Figure \ref{fig: gaussian_copula}. Additionally, aggregating the splitting criteria over $d$ components of the response $\bold{Y}$ can reduce the signal size if only the distribution of a few components change. Our MMD-based splitting criterion \eqref{eq: MMD splitcrit} is able to avoid such difficulties as it implicitly inspects all aspects of the multivariate response distribution.
If we take a trivial kernel $k_{\text{id}}(y_i, y_j) = y_i y_j$ with the identity feature map $\varphi_{\text{id}}(y) = y$, the corresponding distributional embedding \eqref{eq: RKHS embedding} is given by $\mu(\mathcal{P}) = \E_{Y\sim\mathcal{P}} Y$ and thus the corresponding splitting criterion based on $\mathcal{D}_{\text{MMD}(k_{\text{id}})}$ \eqref{eq: MMD embedding} corresponds exactly to the CART splitting criterion \eqref{eq: CART}, which can be seen from its equivalent representation \eqref{eq: CART equivalent}.
Interestingly, Theorem \ref{thm: MMD_CART_equivalence} in Section \ref{sec: theory} below shows that the MMD splitting criterion with kernel $k$ can also be viewed as the abstract CART criterion in the RKHS $\mathcal{H}$ corresponding to $k$ \citep{fan2010cw}. Moreover, it is also shown that DRF with the MMD splitting criterion can thus be viewed asymptotically as a greedy minimization of the squared RKHS distance between the corresponding embeddings of our estimate $\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})$ and the truth $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$ averaged over $\bold{x}$, thus justifying the proposed method. In Section \ref{sec: theory}, we exploit this relationship to derive interesting theoretical properties of DRF with the MMD splitting criterion.
\section{Theoretical Results}
\label{sec: theory}
In this section we exploit the properties of kernel mean embedding in order to relate DRF with the MMD splitting criterion to an abstract version of the standard Random Forest with the CART splitting criterion \citep{breiman2001random} when the response is taking values in the corresponding RKHS. We further exploit this relationship to adapt the existing theoretical results from the Random Forest literature to show that our estimate \eqref{eq: distributional estimate} of the conditional distribution of the response is consistent with respect to the MMD metric for probability measures and with a good rate. Finally, we show that this implies consistency of the induced DRF estimates for a range interesting targets $\tau(\bold{x})$, such as conditional CDFs or quantiles. The proofs of all results can be found in the Appendix \ref{appendix: proofs}.
Recalling the notation from above, let $\left(\mathcal{H}, \langle\cdot,\cdot\rangle_{\mathcal{H}}\right)$ be the Reproducing kernel Hilbert space induced by the positive definite kernel $k:\R^d\times\R^d \to \R$ and let $\varphi:\R^d \to \mathcal{H}$ be its corresponding feature map. The kernel embedding function $\mu:\mathcal{M}_{b}(\R^d) \to \mathcal{H}$
maps any bounded signed Borel measure $\mathcal{P}$ on $\R^d$ to an element $\mu(\mathcal{P}) \in \mathcal{H}$ defined by
$$\mu(\mathcal{P}) = \int_{\R^d}\varphi(\bold{y})d\mathcal{P}(\bold{y}),$$
see \eqref{eq: RKHS embedding}. Boundedness of $k$ ensures that $\mu$ is indeed defined on all of $\mathcal{M}_{b}(\R^d)$, while continuity of $k$ ensures that $\mathcal{H}$ is separable \citet{hilbertspacebook}.
By considering the kernel embedding $\mu(\cdot)$ and using its linearity, we can write the embedding of the distributional estimate $\mu(\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x}))$ of DRF \eqref{eq: distributional estimate} as the average of the embeddings of the empirical distributions of $\bold{Y}$ in the leaves containing $\bold{x}$ over all trees:
\begin{equation} \label{eq: embedding estimate}
\mu(\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})) = \frac{1}{N}\sum_{k=1}^N \mu\left( \frac{1}{|\mathcal{L}_k(\bold{x})|}\sum_{\bold{x}_i \in \mathcal{L}_k(\bold{x})} \delta_{\bold{y}_i}\right) = \frac{1}{N}\sum_{k=1}^N \frac{1}{|\mathcal{L}_k(\bold{x})|}\sum_{\bold{x}_i \in \mathcal{L}_k(\bold{x})} \mu(\delta_{\bold{y}_i}).
\end{equation}
This is analogous to the prediction of the response for the standard Random Forest, but where we average the embeddings $\mu(\delta_{\bold{y}_i}) = \varphi(\bold{y}_i)$ instead of the response values themselves. Therefore, by using the kernel embedding, we can shift the perspective to the RKHS $\mathcal{H}$ and view DRF as the analogue of the original Random Forest for estimation of $\mu(\P(\bold{Y}\mid\bold{X}=\bold{x})) = \E[\varphi(\bold{Y})\mid\bold{X}\myeq\bold{x}]$ in some abstract Hilbert space.
With this viewpoint, we can relate the MMD splitting criterion to the original CART criterion \eqref{eq: CART}, which measures the mean squared prediction error for splitting a certain parent node $P$ into children $C_L$ and $C_R$. On one hand, from Equation \eqref{eq: CART equivalent} we see that the CART criterion measures the squared distance between the response averages $\tfrac{1}{n_L}\sum_{\bold{x}_i \in C_L} y_i$ and $\tfrac{1}{n_R}\sum_{\bold{x}_i \in C_R} y_i$ in the child nodes, but on the other hand, Equation \eqref{eq: MMD embedding} shows that the MMD splitting criterion measures the RKHS distance between the embeddings of the empirical response distributions in $C_L$ and $C_R$. This is summarized in the following theorem, which not only shows that the MMD splitting criterion can be viewed as the abstract CART criterion in the RKHS $\mathcal{H}$ \citep{fan2010cw}, but also that DRF with the MMD splitting criterion can be viewed as greedy minimization of the average squared distance between the estimated and true conditional distributions, as measured by the RKHS norm between the corresponding embeddings to $\mathcal{H}$:
\begin{restatable}{theorem}{MMDCART}\label{thm: MMD_CART_equivalence}
For any split of a parent node $P$ into child nodes $C_L$ and $C_R$, let $\hat{\P}_\text{split}(\bold{x}) = \sum_{j\in\{L, R\}}\1(\bold{x}\in C_j) \tfrac{1}{n_j} \sum_{\bold{x}_i \in C_j} \delta_{\boldsymbol{y}_i}$ denote the resulting estimate of the distribution $\P(\bold{Y} \mid \bold{X}\myeq\bold{x})$ when $\bold{x} \in P$. Then the MMD splitting criterion is equivalent to the abstract version of the CART criterion \eqref{eq: CART} on $\mathcal{H}$:
\small
\begin{equation*}
\argmax_{\text{split}} \frac{n_Ln_R}{n_P^2}\mathcal{D}_{\text{MMD}(k)}\left(\{\bold{y}_i \mid \bold{x}_i \in C_L\} , \{\bold{y}_i\mid \bold{x}_i\in C_R\}\right) = \argmin_{\text{split}} \frac{1}{n_P}\sum_{\bold{x}_i \in P} \norm*{\mu(\delta_{\bold{y}_i}) - \mu(\hat{\P}_{\text{split}}(\bold{x}_i))}_\mathcal{H}^2.
\end{equation*}
\normalsize
Moreover, for any node $P$ and any fixed distributional estimator $\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})$, we have:
\small
\begin{equation*}
\frac{1}{n_P}\sum_{\bold{x}_i \in P} \norm*{\mu(\delta_{\bold{y}_i}) - \mu(\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x}_i))}_\mathcal{H}^2 = V_P + \E\left[\norm{\mu(\hat{\P}(\bold{Y}\mid\bold{X})) - \mu(\P(\bold{Y} \mid \bold{X}))}_\mathcal{H}^2 \mid \bold{X} \in P\right] + \O_{p}(n^{-1/2}),
\end{equation*}
\normalsize
where $V_P = \E\left[\norm{\mu(\delta_\bold{Y}) - \mu(\P(\bold{Y} \mid \bold{X}))}_\mathcal{H}^2\mid \bold{X} \in P\right]$ is a deterministic term not depending on the estimates $\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})$.
\end{restatable}
In conclusion, DRF with the MMD splitting criterion can be viewed as the standard Random Forest with the CART splitting criterion, but with the response $\mu(\delta_{\bold{Y}})$ taking values in an abstract RKHS $\mathcal{H}$ instead of $\R$. Therefore, one could in principle derive properties of DRF by adapting any theoretical result for standard Random Forests from the literature. However, a lot of care is needed for making the results rigorous in this abstract setup, as many useful properties of $\R$ need not hold for infinite-dimensional $\mathcal{H}$. The remaining part of this section is inspired by the results from \citet{wager2018estimation}.
We suppose that the forest construction satisfies the following properties, which significantly facilitate the theoretical considerations of the method and ensure that our forest estimator is well behaved, as stated in \citet{wager2018estimation}:
\begin{itemize}
\item[\textbf{(P1)}] (\textit{Data sampling}) The bootstrap sampling with replacement, usually used in forest-based methods, is replaced by a subsampling step, where for each tree we choose a random subset of size $s_n$ out of $n$ training data points. We consider $s_n$ going to infinity with $n$, with the rate specified below.
\item[\textbf{(P2)}] (\textit{Honesty}) The data used for constructing each tree is split into two parts; the first is used for determining the splits and the second for populating the leaves and thus for estimating the response.
\item[\textbf{(P3)}] (\textit{$\alpha$-regularity}) Each split leaves at least a fraction $\alpha \leq 0.2$ of the available training sample on each side. Moreover, the trees are grown until every leaf contains between $\kappa$ and $2\kappa - 1$ observations, for some fixed tuning parameter $\kappa \in \N$.
\item[\textbf{(P4)}] (\textit{Symmetry}) The (randomized) output of a tree does not depend on the ordering of the training samples.
\item[\textbf{(P5)}] (\textit{Random-split}) At every split point, the probability that the split occurs along the feature $X_j$ is bounded below by $\pi/p$, for some $\pi > 0$ and for all $j=1,\ldots,p$.
\end{itemize}
The validity of the above properties are easily ensured by the forest construction. For more details, see Appendix \ref{appendix: impdetails}.
From Equation \eqref{eq: embedding estimate}, the prediction of DRF for a given test point $\mathbf{x}$ can be viewed as an element of $\mathcal{H}$. If we denote the $i$-th training observation by $\mathbf{Z}_i=(\mathbf{x}_i, \mu(\delta_{\mathbf{y}_i})) \in \R^p \times \mathcal{H}$, then by \eqref{eq: embedding estimate} we estimate the embedding of the true conditional distribution $\mu(\P(\bold{Y}\mid\bold{X}\myeq\bold{x}))$ by the average of the corresponding estimates per tree:
\begin{equation*}\label{rewriting}
\mu(\hat{\P}(\mathbf{Y} \mid \mathbf{X} \myeq \mathbf{x})) = \frac{1}{N} \sum_{j=1}^N T(\bold{x}; \varepsilon_j, \mathcal{Z}_j),
\end{equation*}
where $\mathcal{Z}_k$ is a random subset of $\{\mathbf{Z}_i\}_{i=1}^n$ of size $s_n$ chosen for constructing the $j$-th tree $\mathcal{T}_j$ and $\varepsilon_j$ is a random variable capturing all randomness in growing $\mathcal{T}_j$, such as the choice of the splitting candidates. $T(\bold{x}; \varepsilon, \mathcal{Z})$ denotes the output of a single tree: i.e. the average of the terms $\mu(\delta_{\bold{Y}_i})$ over all data points $\bold{Z}_i$ contained in the leaf $\mathcal{L}(\bold{x})$ of the tree constructed from $\varepsilon$ and $\mathcal{Z}$.
Since one can take the number of trees $N$ to be arbitrarily large, we consider an ``idealized'' version of our estimator, as done in \citet{wager2017estimation}, which we denote as $\hat{\mu}_n(\mathbf{x})$:
\begin{equation}\label{finalestimator}
\hat{\mu}_n(\mathbf{x}) = \binom{n}{s_n}^{-1} \sum_{i_1 < i_2 < \ldots < i_{s_n}} \E_{\varepsilon} \,\,T(\mathbf{x}; \varepsilon; \{\bold{Z}_{i_1}, \ldots, \bold{Z}_{i_{s_n}}\}),
\end{equation}
where the sum is taken over all $\binom{n}{s_n}$ possible subsets of $\{\mathbf{Z}_{i}\}_{i=1}^n$. We have that $\mu(\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})) \to \hat{\mu}_n(\mathbf{x})$ as $N\to \infty$, while keeping the other variables constant, and thus we assume for simplicity that those two quantities are the same.
Our main result shows that, under similar assumptions as in \citet{wager2017estimation}, the embedding of our conditional distribution estimator $\mu_n(\bold{x}) = \mu(\hat{\P}(\mathbf{Y} \mid \mathbf{X} \myeq \mathbf{x}))$ consistently estimates $\mu(\bold{x}) \coloneqq \mu(\P(\mathbf{Y} \mid \mathbf{X} \myeq \mathbf{x}))$ with respect to the RKHS norm with a certain rate:
\begin{restatable}{theorem}{consistency}\label{thm: consistency}
Suppose that our forest construction satisfies properties \textbf{(P1)}--\textbf{(P5)}. Assume additionally that $k$ is a bounded and continuous kernel and that we have a random design with $\bold{X}_1,\ldots, \bold{X}_n$ independent and identically distributed on $[0,1]^p$ with a density bounded away from $0$ and infinity. If the subsample size $s_n$ is of order $n^\beta$ for some $0 < \beta < 1$, the mapping
$$\mathbf{x} \mapsto \mu(\mathbf{x})=\E[ \mu(\delta_\bold{Y}) \mid \mathbf{X} \myeq \mathbf{x}] \in \mathcal{H},$$
is Lipschitz and $ \sup_{\mathbf{x} \in [0,1]^p} \E[ \|\mu(\delta_\bold{Y})\|_\mathcal{H}^2\mid \mathbf{X} \myeq \mathbf{x}] < \infty$, we obtain the consistency w.r.t. the RKHS norm:
\begin{equation}\label{eq: rate}
\norm{\hat{\mu}_n(\mathbf{x})- \mu(\mathbf{x})}_\mathcal{H} = o_{p}\left(n^{-\gamma} \right),
\end{equation}
for any $\gamma < \frac{1}{2} \min\left( 1- \beta, \frac{\log(1-\alpha)}{\log(\alpha)} \frac{\pi}{p} \cdot \beta \right)$.
\end{restatable}
\begin{remark}
The rate in \eqref{eq: rate} is analogous to the one from \citet{wager2018estimation}, who used it further to derive the asymptotic normality of the random forest estimator in $\R$. Indeed, one can show in our case that there exists a sequence of real numbers $\sigma_n \to 0$, such that $(\hat{\mu}_n(\mathbf{x})- \mu(\mathbf{x}))/\sigma_n$ as a random element of $\mathcal{\mathcal{H}}$ is ``asymptotically linear'', in the sense that it is indistinguishable asymptotically from an average of independent random elements in $\mathcal{H}$. Unfortunately, this alone is not enough to establish asymptotic normality of $(\hat{\mu}_n(\mathbf{x})- \mu(\mathbf{x}))/\sigma_n$ as an element of $\mathcal{H}$, a task left for future research.
\end{remark}
The above result shows that DRF estimate $\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x})$ converges fast to the truth $\P(\bold{Y}\mid\bold{X}\myeq\bold{x})$ in the MMD distance, i.e.\ the RKHS distance between the corresponding embeddings. Even though this is interesting on its own, ultimately we want to relate this result to estimation of certain distributional targets $\tau(\bold{x}) = \tau(\P(\bold{Y}\mid\bold{X}\myeq\bold{x})).$
For any $f \in \mathcal{H}$, we have that the DRF estimate of the target $\tau(\bold{x}) = \E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}]$ equals the dot product $\langle f, \hat{\mu}_n(\mathbf{x}) \rangle_\mathcal{H}$ in the RKHS:
\begin{equation*}
\langle f , \hat{\mu}_n(\mathbf{x}) \rangle_\mathcal{H} = \left\langle f,\, \int_{\R^d} \varphi(\mathbf{y}) d \hat{\P}(\mathbf{y} \mid \mathbf{X} \myeq \mathbf{x})\right\rangle_\mathcal{H} = \int_{\R^d} f(\mathbf{y}) \, d\hat{\P}(\mathbf{y} \mid \mathbf{X} \myeq \mathbf{x}) = \sum_{i=1}^n w_{\mathbf{x}}(\mathbf{x}_i) f(\mathbf{y}_i),
\end{equation*}
where we recall the weighting function $w_{\mathbf{x}}(\cdot)$ induced by the forest \eqref{eq: weighting function}.
Therefore, the consistency result \eqref{eq: rate} in Theorem \ref{thm: consistency} directly implies that
\begin{equation} \label{eq: target consistency}
\sum_{i=1}^n w_{\mathbf{x}}(\mathbf{x}_i) f(\mathbf{y}_i) = \langle f, \hat{\mu}_n(\mathbf{x}) \rangle_\mathcal{H} \stackrel{p}{\to} \langle f, \mu(\mathbf{x}) \rangle_\mathcal{H} = \E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}] \quad \text{ for any $f \in \mathcal{H}$},
\end{equation}
i.e. that the DRF consistently estimates the targets of the form $\tau(\bold{x}) = \E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}]$, for $f \in \mathcal{H}$. From \eqref{eq: rate} we also obtain the rate of convergence when $s_n \asymp n^\beta$:
\begin{equation*}
\left|\sum_{i=1}^n w_{\mathbf{x}}(\mathbf{x}_i) f(\mathbf{y}_i) - \E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}]\right| = o_{p}\left( n^{-\gamma} \norm{f}_\mathcal{H}\right),
\end{equation*}
for $\gamma$ as in Theorem \ref{thm: consistency}.
When $k$ is continuous, it is well known that all elements of $\mathcal{H}$ are continuous, see e.g.\ \citet{hilbertspacebook}. Under certain assumptions on the kernel and its input space, holding for several popular kernels, (e.g.\ the Gaussian kernel) \citep{optimalestimationofprobabilitymeasures}, we can generalize the convergence result \eqref{eq: target consistency} to any bounded and continuous function $f:\R^d \to \R$, as the convergence of measures $\hat{\P}(\bold{Y}\mid\bold{X}\myeq\bold{x}) \to \P(\bold{Y}\mid\bold{X}\myeq\bold{x})$ in the MMD metric will also imply their weak convergence, i.e. $k$ metrizes weak convergence \citep{optimalestimationofprobabilitymeasures, simon2018kernel, simon2020metrizing}:
\begin{corollary} \label{cor: metrizing weak convergence}
Assume that one of the following two sets of conditions holds:
\begin{itemize}
\item[(a)] The kernel $k$ is bounded, (jointly) continuous and has
\begin{align}\label{characteristiccondapp}
\int \int k(\mathbf{x},\mathbf{y}) d\mathcal{P}(\mathbf{x}) d\mathcal{P}(\mathbf{y}) > 0 \ \ \forall \mathcal{P} \in \mathcal{M}_b(\R^d)\setminus \{0\}.
\end{align}
Moreover, $\mathbf{y} \mapsto k(\mathbf{y}_0, \mathbf{y})$ is vanishing at infinity, for all $\mathbf{y}_0 \in \R^d$.
\item[(b)] The kernel $k$ is bounded, shift-invariant, (jointly) continuous and $\nu$ in the Bochner representation in \eqref{eq: bochner} is supported on all of $\R^d$. Moreover, $\mathbf{Y}$ takes its values almost surely in a closed and bounded subset of $\R^d$.
\end{itemize}
Then, under the conditions of Theorem \ref{thm: consistency}, we have for any bounded and continuous function $f:\R^d \to \R$ that DRF consistently estimates the target $\tau(\bold{x})=\E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}]$ for any $\bold{x} \in [0,1]^p$:
$$\sum_{i=1}^n w_{\mathbf{x}}(\mathbf{x}_i) f(\mathbf{y}_i) \,\stackrel{p}{\to}\, \E[f(\bold{Y})\mid\bold{X}\myeq\bold{x}].$$
\end{corollary}
Recalling the Portmanteau Lemma on separable metric spaces, see e.g.\ \citet[Chapter 11]{dudley}, this has several other interesting consequences, such as the consistency of CDF and quantile estimates; Let $F_{\mathbf{Y}\mid\bold{X}\myeq\bold{x}}(\cdot)$ be the conditional CDF of $\mathbf{Y}$ and for any index $1\leq i\leq d$, let $F_{Y_i\mid\bold{X}\myeq\bold{x}}(\cdot)$ be the conditional CDF of $Y_i$ and $F_{Y_i\mid\bold{X}\myeq\bold{x}}^{-1}(\cdot)$ its generalized inverse, i.e. the quantile function. Let $\hat{F}_{Y_i\mid\bold{X}\myeq\bold{x}}(\cdot)$ and $\hat{F}_{Y_i\mid\bold{X}\myeq\bold{x}}^{-1}(\cdot)$ be the corresponding DRF estimates via weighting function \eqref{eq: CDF estimate}. Then we have the following result:
\begin{corollary}\label{cor: cdfresult}
Under the conditions of Corollary \ref{cor: metrizing weak convergence}, we have
\begin{align*}
\hat{F}_{\mathbf{Y} \mid\bold{X}\myeq\bold{x}}(\mathbf{t}) \, &\stackrel{p}{\to}\, F_{\mathbf{Y}\mid\bold{X}\myeq\bold{x}}(\mathbf{t}) \\
\hat{F}_{Y_i\mid\bold{X}\myeq\bold{x}}^{-1}(t) \, &\stackrel{p}{\to}\, F_{Y_i\mid\bold{X}\myeq\bold{x}}^{-1}(t)
\end{align*}
for all points of continuity $\mathbf{t} \in \R^d$ and $t \in \R$ of $F_{\mathbf{Y}\mid\bold{X}\myeq\bold{x}}(\cdot)$ and $F_{Y_i\mid\bold{X}\myeq\bold{x}}^{-1}(\cdot)$ respectively.
\end{corollary}
\section{Applications and Numerical Experiments}
\label{sec: applications}
The goal of this section is to demonstrate the versatility and applicability of DRF for many practical problems. We show that DRF can be used not only as an estimator of the multivariate conditional distribution, but also as a two-step method to easily obtain out-of-the box estimators for various, and potentially complex, targets $\tau(\bold{x})$.
Our main focus lies on the more complicated targets which cannot be that straightforwardly approached by conventional methods. However, we also illustrate the usage of DRF for certain applications for which there already exist several well-established methods. Whenever possible in such cases, we compare the performance of DRF with the specialized, task-specific methods to show that, despite its generality, there is at most a very small loss of precision. However, we should point out that for many targets such as, that can not be written in a form of a conditional mean or a conditional quantile, for example, conditional correlation, direct comparison of the accuracy is not possible for real data, since no suitable loss function exists and the ground truth is unknown. Finally, we show that, in addition to directly estimating certain targets, DRF can also be a very useful tool for many different applications, such as causality and fairness.
Detailed descriptions of all data sets and the corresponding analyses, together with additional simulations can be found in the appendix.
\subsection{Estimation of Conditional Multivariate Distributions}
\label{sec: benchmarks}
In order to provide good estimates for any target $\tau(\bold{x}) = \tau(\P(\bold{Y}\mid\bold{X}\myeq\bold{x}))$, our method needs to estimate the conditional multivariate distribution $\P(\bold{Y}\mid \bold{X}\myeq\bold{x})$ well. Therefore, we first investigate here the accuracy of the DRF estimate \eqref{eq: distributional estimate} of the full conditional distribution and compare its performance with the performance of several existing methods.
There are not many algorithms in the literature that nonparametrically estimate the multivariate conditional distribution $\P(\bold{Y}\mid \bold{\mathbf{X}=\mathbf{x}})$. In addition to a few simple methods such as the $k$-nearest neighbors or the kernel regression, which locally weight the training points, we also consider advanced machine learning methods such as the Conditional Generative Adversarial Network (CGAN) \citep{mirza2014conditional, aggarwal2019benchmarking}, Conditional Variational Autoencoder (CVAE) \citep{sohn2015learning} and Masked Autoregressive Flow \citep{papamakarios2017masked}. It is worth mentioning that the focus in the machine learning literature has been more on applications where $d$ is very large (e.g.\ pixels of an image) and $p$ is very small (such as image labels). Even though some methods do not provide the estimated conditional distribution in a form as simple as DRF, one is still able to sample from the estimated distribution and thus perform any subsequent analysis and make fair comparisons between the methods.
\begin{figure}[h]
\hspace{-0.8cm}
\includegraphics[width=1.06\linewidth]{vignette.png}
\caption{The illustration of the estimated joint conditional distribution obtained by different methods for the toy example \eqref{eq: vignette example}. For $1000$ randomly generated test points $\bold{X}_{\text{test}} \sim U(0, 1)^p$ the top row shows the estimated distribution of the response component $Y_1$, whereas the bottom row shows the estimated distribution of $Y_2$. The $0.1$ and $0.9$ quantiles of the true conditional distribution are indicated by a dashed black line, whereas the conditional mean is shown as a black solid line.}
\label{fig: distribution illustration}
\end{figure}
We first illustrate the estimated distributions of the above method on a toy example where $n=1000, p=10, d=2$ and
\begin{equation} \label{eq: vignette example}
Y_1 \indep Y_2\mid \bold{X}\myeq\bold{x}, \quad Y_1 \mid \bold{X}\myeq\bold{x} \sim U(x_1, x_1 + 1), \quad Y_2 \mid \bold{X}\myeq\bold{x} \sim U(0, x_2), \quad \bold{X} \sim U(0,1)^p.
\end{equation}
That is, in the above example $X_1$ affects the mean of $Y_1$, whereas $X_2$ affects the both mean and variance of $Y_2$, and $X_3,\ldots, X_p$ have no impact. The results can be seen in Figure \ref{fig: distribution illustration} for the above methods. We see that, unlike some other methods, DRF is able to balance the importance of the predictors $X_1$ and $X_2$ and thus to estimate the distributions of $Y_1$ and $Y_2$ well.
One can do a more extensive comparison on real data sets. We use the benchmark data sets from the multi-target regression literature \citep{mtr} together with some additional ones created from the real data sets described throughout this paper. The performance of DRF is compared with the performance of other existing methods for nonparametric estimation of multivariate distributions by using the Negative Log Predictive Density (NLPD) loss, which evaluates the logarithm of the induced multivariate density estimate \citep{quinonero2005evaluating}. As the number of test points grows to infinity, NLPD loss becomes equivalent to the average KL divergence between the estimated and the true conditional distribution and is thus able to capture how well one estimates the whole distribution instead of only its mean.
In addition to the methods mentioned above, we also include the methods that are intended only for mean prediction, by assuming that the distribution of the response around its mean is homogeneous, i.e. that the conditional distribution $\P\left(\bold{Y} - \E[\bold{Y} \mid \bold{X}] \mid \bold{X}\myeq\bold{x}\right)$ does not depend on $\bold{x}$. This is fitted by regressing each component of $\bold{Y}$ separately on $\bold{X}$ and using the pooled residuals. We consider the standard nonparametric regression methods such as Random Forest \citep{breiman2001random}, XGBoost \citep{Chen:XGB}, and Deep Neural Networks \citep{goodfellow2016deep}.
\begin{table}[h]
\centering
\begin{tabularx}{\textwidth}{lcccccccccccccc}
\rot[0][3.5em]{}
& \rot[45][1.8em]{jura}
& \rot[45][1.8em]{slump}
& \rot[45][1.8em]{wq}
& \rot[45][1.8em]{enb}
& \rot[45][1.8em]{atp1d}
& \rot[45][1.8em]{atp7d}
& \rot[45][1.8em]{scpf}
& \rot[45][1.8em]{sf1}
& \rot[45][1.8em]{sf2}
& \rot[45][1.8em]{copula}
& \rot[45][1.8em]{wage}
& \rot[45][1.8em]{births1}
& \rot[45][1.8em]{births2}
& \rot[45][1.8em]{air} \\
\bottomrule
\multicolumn{1}{c|}{$n$} & 359 & 103 & 1K & 768 & 337 & 296 & 143 & 323 & 1K & 5K & 10K & 10K & 10K & 10K\\
\multicolumn{1}{c|}{$p$} & 15 & 7 & 16 & 8 & 370 & 370 & 8 & 21 &22 & 10 & 73 & 23 & 24 & 15 \\
\multicolumn{1}{c|}{$d$} & 3 & 3& 14 &2 & 6 & 6 & 3 & 3 & 6 & 2 & 2 & 2 &4 & 6\\
\hline
\multicolumn{1}{c|}{$\text{DRF}$} & \textbf{3.9} & \textbf{4.0} & \textbf{22.5} & 2.1 & 7.3 & \textbf{7.0} & \textbf{2.0} & \textbf{-24.2} & \textbf{-24.3} & \textbf{2.8} & \textbf{2.8} & 2.5 & \textbf{4.2} & \textbf{8.5} \\
\multicolumn{1}{c|}{$\text{CGAN}$} & 10.8 & 5.3 & 27.3 & 3.5 & 10.4 & 363 & 4.8 & 9.8 & 21.1 & 5.8 & 360 & \textbf{2.4} & {\small $>$1K} & 11.8 \\
\multicolumn{1}{c|}{$\text{CVAE}$} & 4.8& 37.8 & 36.8 & 2.6 & {\small $>$1K} & {\small $>$1K} & 108.8 & 8.6 & {\small $>$1K} & 2.9 & {\small $>$1K} & {\small $>$1K} & 49.7 & 9.6 \\
\multicolumn{1}{c|}{$\text{MAF}$} & 4.6 & 4.5 & 23.9& 3.0 & 8.0 & 8.1 & 2.6 & 4.7 & 3.8 & 2.9 & 3.0 & 2.5 & {\small $>$1K} & 8.5 \\
\multicolumn{1}{c|}{$\text{k-NN}$} & 4.5 & 5.0 & 23.4 & 2.4 & 8.8 & 8.6 & 4.1 & -22.4 & -19.7 & 2.9 & \textbf{2.8} & 2.7 & 4.4 & 8.8 \\
\multicolumn{1}{c|}{$\text{kernel}$} & 4.1 & 4.2 & 23.0 & \textbf{2.0} & \textbf{6.6} & 7.1 & 2.9 & -23.0 & -20.6 & 2.8 & 2.9 & 2.6 & 4.3 & 8.4 \\
\multicolumn{1}{c|}{$\text{RF}$} & 7.1 & 12.1 & 35.2 & 5.7 & 12.7 & 13.3 & 16.7 & 3.9 & 2.2 & 5.8 & 6.1 & 5.0 & 8.3 & 13.9 \\
\multicolumn{1}{c|}{$\text{XGBoost}$} & 11.4 & 38.3 & 25.9 & 3.0 & {\small $>$1K} & {\small {\small $>$1K}} & {\small $>$1K} & 0.3 & 1.6 & 3.5 & 2.9 & {\small $>$1K} & {\small $>$1K} & 12.8 \\
\multicolumn{1}{c|}{$\text{DNN}$} & 4.0 & 4.2 & 23.3 & 2.6 & 8.6 & 8.7 & 2.6 & 2.3 & 2.2 & 2.9 & 3.0 & 2.6 & 5.4 & 8.6 \\
\hline
\end{tabularx}
\caption{
NLPD loss computed on out-of-sample observations for the estimated conditional distributions obtained by several different methods (corresponding to rows) for many real data sets (corresponding to columns). The best method is indicated in bold. Detailed description of both the data sets and the competing methods can be found in Appendix \ref{appendix: simulation details}.}
\label{benchtable}
\end{table}
The results are shown in Table \ref{benchtable}. We see that DRF performs well for a wide range of sample size and problem dimensionality, especially in problems where $p$ is large and $d$ is moderately big. It does so without the need for any tuning or involved numerical optimization. More detailed analysis and descriptions of each competing method and the loss function can be found in the Appendix \ref{appendix: simulation details}.
\subsection{Estimation of Statistical Functionals}
Because DRF represents the estimated conditional distribution $\hat{\P}(\bold{Y} \mid \bold{X}=\bold{x}) = \sum_{i} w_{\bold{x}}(\bold{x}_i)\cdot \delta_{\bold{y}_i}$ in a convenient form by using weights $w_{\bold{x}}(\bold{x}_i)$, a plug-in estimator $\tau(\hat{\P}(\bold{Y} \mid \bold{X}=\bold{x}))$ of many common real valued statistical functionals $\tau(\P(\bold{Y} \mid \bold{X}=\bold{x})) \in \R$ can be easily constructed from $w_{\bold{x}}(\cdot)$.
\begin{figure}[h]
\hspace{-0.7cm}
\includegraphics[width=1.05\linewidth]{univariate.png}
\caption{Scatter plot of predictions of the $0.1, 0.5$ and $0.9$ quantiles against $X_1$ for randomly generated $500$ test data points $\bold{X}_{\text{test}} \sim U(-1, 1)^p$. The true values of the quantiles are displayed by black dashed lines. The columns corresponds to different methods DRF (red), GRF (green), QRF (blue), TRF (purple). The rows correspond to different simulation scenarios. The first two are taken from \citet{athey2019generalized}.}
\label{fig: univariate}
\end{figure}
We first investigate the performance for the classical problem of univariate quantile estimation on simulated data. We consider the following three data generating mechanisms with $p=40, n=2000$ and $\bold{X}_i \iid U(-1,1)^p$:
\begin{itemize}
\item Scenario 1: $Y \sim N(0.8\cdot \1(X_1 > 0), 1)$ (mean shift based on $X_1$)
\item Scenario 2: $Y \sim N(0, (1+\1(X_1>0))^2)$ (variance shift based on $X_1$)
\item Scenario 3: $Y \sim \1(X_1\leq 0) \cdot N(1, 1) + \1(X_1 > 0) \cdot \text{Exp}(1)$ (distribution shift based on $X_1$, constant mean and variance)
\end{itemize}
The first two scenarios correspond exactly to the examples given in \citet{athey2019generalized}.
In Figure \ref{fig: univariate} we can see the corresponding estimates of the conditional quantiles for DRF, Quantile Regression Forest (QRF) \citep{meinshausen2006quantile}, which uses the same forest construction with CART splitting criterion as the original Random Forest \citep{breiman2001random} but estimates the quantiles from the induced weighting function, Generalized Random Forests (GRF) \citep{athey2019generalized} with a splitting criterion specifically designed for quantile estimation and Transformation Forests (TRF) \citep{hothorn2017transformation}. We see that DRF is performing very well even compared to methods that are specifically tailored to quantile estimation. For more detailed analysis and some additional examples, such as the univariate mean regression, we refer the reader to Appendix \ref{appendix: additional examples}.
The multivariate setting is however more interesting, as one can use DRF to compute much more interesting statistical functionals $\tau(\bold{x})$. We illustrate this in Figure \ref{fig: functionals} for the the air quality data set, described in Section \ref{sec: weighting function}. The left plot shows one value of the estimated multivariate CDF, specifically the estimated probability of the event that the air quality index (AQI) is at most $50$ at a given test site. This corresponds to the "Good" category and means that the amount of every air pollutant is below a certain threshold determined by the EPA. Such probability estimates can be easily obtained by summing the weights of the training points belonging to the event of interest.
\begin{figure}[h]
\hspace*{-0.3cm}
\includegraphics[width=1.04\textwidth]{functionals.png}
\caption{Estimates of the probability $\P(\text{AQI} \leq 50 \mid \text{test site})$ (left) and the conditional correlation (right) derived from the DRF estimate of the multivariate conditional distribution.}
\label{fig: functionals}
\end{figure}
In order to investigate the accuracy of the conditional CDF obtained by DRF, we compare the estimated probabilities with estimates of the standard univariate classification forest \citep{breiman2001random} with the response $\1(\text{AQI} \leq 50)$. In the left plot of Figure \ref{comparison}, we can see that the DRF estimates of the $\P(\text{AQI} \leq 50 \mid {\mathbf{X}=\mathbf{x}})$ (also visualized in Figure \ref{fig: functionals}) are quite similar to the estimates of the classification forest predicting the outcome $\1(\text{AQI} \leq 50)$. Furthermore, the cross-entropy loss evaluated on the held-out measurements equals $0.4671$ and $0.4663$ respectively, showing almost no loss of precision. In general, estimating the simple functionals from the weights provided by DRF comes usually at a small to no loss compared to the classical methods specifically designed for this task.
In addition to the classical functionals $\tau(\bold{x})$ in the form of an expectation $\mathbb{E}(f(\bold{Y})\mid \bold{X}=\bold{x})$ or a quantile $Q_{\alpha}(f(\bold{Y})\mid \bold{X}=\bold{x})$ for some function $f: \mathbb{R}^{d} \rightarrow \mathbb{R}$, which can also be computed by solving the corresponding one-dimensional problems, additional interesting statistical functionals with intrinsically multivariate nature that are not that simple to estimate directly are accessible by DRF, such as, for example, the conditional correlations $\Cor(Y_i,\, Y_j \mid \bold{X}\myeq\bold{x})$. As an illustration, the estimated correlation of the sulfur dioxide ($\text{SO}_2$) and fine particulate matter (PM2.5) is shown in the right plot of Figure \ref{fig: functionals}. The plot reveals also that the local correlation in many big cities is slightly larger than in its surroundings, which can be explained by the fact that the industrial production directly affects the levels of both pollutants.
\begin{figure}[h]
\hspace*{-0.4cm}
\includegraphics[width=1.02\textwidth, height=3.5cm]{classification_comparison.png}
\caption{Left: Comparison of the CDF estimates obtained by DRF (displayed also in the left plot of Figure \ref{fig: functionals}) and by the classification forest. Right: Example how the CDF estimated by using the classification forest (blue) need not be monotone, whereas the DRF estimates (red) are well-behaved.}
\label{comparison}
\end{figure}
A big advantage of the target-free forest construction of DRF is that all subsequent targets are computed from same the weighting function $w_\bold{x}$ obtained from a single forest fit. First, this is computationally more efficient, since we do not need for every target of interest to fit the method specifically tailored to it. For example, estimating the CDF with classification forests requires fitting one forest for each function value. Secondly and even more importantly, since all statistical functionals are plug-in estimates computed from the same weighting function, the obtained estimates are mathematically well-behaved and mutually compatible. For example, if we estimate $\Cor(Y_i, Y_j \mid \bold{X}\myeq\bold{x})$ by separately estimating the terms $\Cov(Y_i, Y_j \mid \bold{X}\myeq\bold{x})$, $\Var(Y_i \mid \bold{X}\myeq\bold{x})$, and $\Var(Y_j \mid \bold{X}\myeq\bold{x})$, one can not in general guarantee the estimate to be in the range $[-1, 1]$, but this is possible with DRF. Alternatively, the correlation or covariance matrices that are estimated entrywise are guaranteed to be positive semi-definite if one uses DRF. As an additional illustration, Figure \ref{comparison} shows that the estimated (univariate) CDF using the classification forest need not be monotone due to random errors in each predicted value, which can not happen with the DRF estimates.
\subsection{Conditional Copulas and Conditional Independence Testing}
\label{sec: copulas}
One can use the weighting function not only to estimate certain functionals, but also to obtain more complex objects, such as, for example, the conditional copulas. The well-known Sklar's theorem \citep{sklar1959fonctions} implies that at a point $\textbf{x} \in \mathbb{R}^{p}$, the conditional CDF $\P(\bold{Y} \leq \bold{y} \mid \bold{X}=\bold{x}) = \P(Y_1 \leq y_1, \ldots, Y_d \leq y_d \mid \bold{X}=\bold{x})$ can be represented by a CDF $C_{\bold{x}}$ on $[0,1]^{d}$, the conditional copula at $\bold{x}$, and $d$ conditional marginal CDFs $F_{Y_i\mid\bold{X}\myeq\bold{x}}(y) = \P(Y_{i} \leq y \mid \bold{X}=\bold{x})$ for $1 \leq i \leq d$, as follows:
\begin{equation} \label{eq: copula decomposition}
\P(\bold{Y} \leq \bold{y}\mid \bold{X}=\bold{x}) =
C_{\bold{x}}\left( F_{Y_1\mid\bold{X}\myeq\bold{x}}(y_1), \ldots, F_{Y_d\mid\bold{X}\myeq\bold{x}}(y_d) \right).
\end{equation}
Copulas capture the dependence of the components $Y_i$ by the joint distribution of the corresponding quantile levels of the marginal distributions: $F_{Y_i\mid\bold{X}\myeq\bold{x}}(Y_i) \in [0,1]$. Decomposing the full multivariate distribution to marginal distributions and the copula is a very useful technique used in many fields such as risk analysis or finance \citep{cherubini2004copula}. Using DRF enables us to estimate copulas conditionally, either by fitting certain parametric model or nonparametrically, directly from the weights.
\begin{figure}[h]
\centering
\hspace*{-0.6cm}
\includegraphics[width=1.04\textwidth]{copula+dist.png}
\caption{Estimated conditional joint distribution of $(Y_1, Y_2)$ and conditional copulas obtained by DRF at different test points $\bold{x}$, where $x_1$ equals $0.25$ and $0.75$ respectively. The red lines are the contours of the true multivariate density function.}
\label{fig: gaussian_copula}
\end{figure}
To illustrate this, consider an example where the $5$-dimensional $\bold{Y}$ is generated from the equicorrelated Gaussian copula $\bold{Y}= (Y_{1},\ldots,Y_{5}) \mid \bold{X}=\bold{x} \sim C^{\text{Gauss}}_{\rho(\bold{x})}$ conditionally on the covariates $\bold{X}$ with distribution $\mathbf{X}_i \iid U(0,1)^{p}$, where $p=30$ and $n\myeq 5000$. All $Y_i$ have a $N(0,1)$ distribution marginally, but their conditional correlation for $i \neq j$ is given by $\text{Cor}(Y_i, Y_j) = \rho(\bold{x}) = x_1$. Figure \ref{fig: gaussian_copula} shows that DRF estimates the full conditional distribution at different test points $\bold{x}$ quite accurately and thus we can obtain a good nonparametric estimate of the conditional copula as follows. First, for each component $Y_i$, we compute the corresponding marginal CDF estimate $\hat{F}_{Y_i\mid\bold{X}\myeq\bold{x}}(\cdot)$ from the weights. Second, we map each response $\bold{y}_i \to \bold{u}_i \coloneqq \left(\hat{F}_{Y_1\mid\bold{X}\myeq\bold{x}}\left((\bold{y}_i)_1\right), \ldots, \hat{F}_{Y_d\mid\bold{X}\myeq\bold{x}}\left((\bold{y}_i)_d\right)\right)$. The copula estimate is finally obtained from the weighted distribution $\sum_{i=1}^n w_{\bold{x}}(\bold{x}_i) \delta_{\bold{u}_i}$, from which we sample the points in Figure \ref{fig: gaussian_copula} in order to visualize the copula.
If we want to instead estimate the copula parametrically, we need to find the choice of parameters for a given model family which best matches the estimated conditional distribution, e.g.\ by weighted maximum likelihood estimation (MLE). For the above example, the correlation parameter of the Gaussian copula can be estimated by computing the weighted correlation with weights $\{w_\bold{x}(\bold{x}_i)\}_{i=1}^n$.
The left plot in Figure \ref{fig: conditional-independence} shows the resulting estimates of the conditional correlation $\text{Cor}\left(Y_1, Y_2 \mid \bold{X}=\bold{x}\right)$ obtained from $\text{DRF}_{\text{MMD}}$, which uses the MMD splitting criterion \eqref{eq: MMD splitcrit} described in Section \ref{sec: MMD splitcrit}, and $\text{DRF}_{\text{CART}}$, which aggregates the marginal CART criteria \citep{kocev2007ensembles, segal2011multivariate}. We see that $\text{DRF}_{\text{MMD}}$ is able to detect the distributional heterogeneity and provide good estimates of the conditional correlation. On the other hand, $\text{DRF}_{\text{CART}}$ cannot detect the change in distribution of $\bold{Y}$ caused by $X_1$ that well. The distributional heterogeneity can not only occur in marginal distribution of the responses (a case extensively studied in the literature), but also in their interdependence structure described by the conditional copula $C_{\bold{x}}$, as one can see from decomposition \eqref{eq: copula decomposition}. Since $\text{DRF}_\text{MMD}$ relies on a distributional metric for its splitting criterion, it is capable of detecting any change in distribution \citep{gretton2007kernel}, whereas aggregating marginal CART criteria for $Y_1, \ldots, Y_d$ in $\text{DRF}_\text{CART}$ only captures the changes in the marginal means.
\begin{figure}[h]
\centering
\hspace*{-0.6cm}
\includegraphics[width=0.9\textwidth]{conditional-independence.png}
\caption{Estimated conditional correlation of $Y_1$ and $Y_2$ (left) and estimated conditional dependence quantified by HSIC statistic (right), obtained by $\text{DRF}_{\text{MMD}}$ (blue) and $\text{DRF}_{\text{CART}}$ (red) respectively. For every test point, we set $X_j=0.5, j \neq 1$. Black dashed curve indicates the population values.}
\label{fig: conditional-independence}
\end{figure}
This is further illustrated for a related application of conditional independence testing, where we compute some dependence measure from the obtained weights. For example, we can test the independence $Y_1 \indep Y_2$ conditionally on the event $\bold{X}=\bold{x}$ by using the Hilbert Schmidt Independence Criterion (HSIC) \citep{HSIC}, which measures the difference between the joint distribution and the product of the marginal distributions. The right plot of Figure \ref{fig: conditional-independence} shows that the $\text{DRF}_{\text{MMD}}$ estimates are quite close to the population value of the HSIC, unlike the ones obtained by $\text{DRF}_\text{CART}$.
\subsection{Heterogeneous Regression and Causal Effect Estimation}
\label{sec: causality}
In this and the following section, we illustrate that, in addition to direct estimation of certain targets, DRF can also be a useful tool for complex statistical problems and applications, such as causality.
Suppose we would like to investigate the relationship between some (univariate) quantity of interest $Y$ and certain predictors $\bold{W}$ from heterogeneous data, where the change in distribution of $(\bold{W}, Y)$ can be explained by some other covariates $\bold{X}$. Very often in causality applications, $\bold{W}$ is a (multivariate) treatment variable, $Y$ is the outcome, which is commonly, but not necessarily, discrete, and $\bold{X}$ is a set of observed confounding variables for which we need to adjust if we are interested in the causal effect of $\bold{W}$ on $Y$. This is illustrated by the following causal graph:
\begin{center}
\begin{tikzpicture}[
> = stealth,
shorten > = 1pt,
auto,
node distance = 3cm,
semithick
]
\tikzstyle{every state}=[
draw = black,
thick,
fill = white,
minimum size = 4mm
]
\node[circle, draw=black] (W) at (0,0) {$\bold{W}$};
\node[circle, draw=black] (X) at (2,1.3) {$\boldsymbol{X}$};
\node[circle, draw=black] (Y) at (4, 0) {$Y$};
\path[->,line width=1.4pt] (W) edge node {} (Y);
\path[->] (X) edge node {} (W);
\path[->] (X) edge node {} (Y);
\end{tikzpicture}
\end{center}
The problem of nonparametric confounding adjustment is hard; not only can the marginal distributions of $Y$ and $\bold{W}$ be affected by $\bold{X}$, thus inducing spurious associations due to confounding, but the way how $\bold{W}$ affects $Y$ can itself depend on $\bold{X}$, i.e.\ the treatment effect might be heterogeneous. The total causal effect can be computed by using the adjustment formula \citep{pearl2009causality}:
\begin{align}
\E[Y\mid do(\bold{W}\myeq\bold{w})] &= \int \E[Y\mid do(\bold{W}\myeq\bold{w}), \bold{X}\myeq\bold{x}]\,\P(\bold{X}\myeq\bold{x}\mid do(\bold{W}\myeq\bold{w}))d\bold{x} \nonumber\\
&=\int \E[Y\mid \bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x}]\,\P(\bold{X}\myeq\bold{x})d\bold{x}. \label{eq: causal effect}
\end{align}
However, implementing do-calculus for finite samples and potentially non-discrete data might not be straightforward and comes with certain difficulties. The standard approach would be to estimate the conditional mean $\E[Y\mid \bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x}]$ nonparametrically by regressing $Y$ on $(\bold{X}, \bold{W})$ with some method of choice and to average out the estimates over different $\bold{x}$ sampled from the observed distribution of $\bold{X}$. Using DRF for this is not necessary, but has an advantage that one can easily estimate the full interventional distribution $\P(Y\mid do(\bold{W}\myeq\bold{w}))$ and not only the interventional mean.
Another way to compute the causal effect is explained in the following, which allows to add more structure to the problem. We use DRF to first fit the forest with the multivariate response $(\bold{W}, Y)$ and the predictors $\bold{X}$. In this way, one can for any point of interest $\bold{x}$ obtain the joint distribution of $(\bold{W}, Y)$ conditionally on the event $\bold{X}\myeq\bold{x}$ and then the weights $\{w_\bold{x}(\bold{x}_i)\}_{i=1}^n$ can be used as an input for some regression method for regressing $Y$ on $\bold{W}$ in the second step. This conditional regression fit might be of separate interest and it can also be used for estimating the causal effect $\E[Y \mid do(\bold{W}\myeq\bold{w})]$ from \eqref{eq: causal effect}, by averaging the estimates $\E[Y \mid \bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x}]$ over $\bold{x}$, where $\bold{x}$ is sampled from the empirical observation of $\bold{X}$.
In this way one can efficiently exploit and incorporate any prior knowledge of the relationship between $\bold{W}$ and $Y$, such as, for example, monotonicity, smoothness or that it satisfies a certain parametric regression model, without imposing any assumptions on the effect of $\bold{X}$ on $(\bold{W}, Y)$. Furthermore, one might be able to better extrapolate to the regions of space where $\P(\bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x})$ is small, compared to the standard approach which computes $\E[Y \mid \bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x}]$ directly, by regressing $Y$ on $(\bold{W}, \bold{X})$. Extrapolation is crucial for causal applications, since for computing $\E[Y\mid do(\bold{W}\myeq\bold{w})]$ we are interested in what would happen with $Y$ when our treatment variable $\bold{W}$ is set to be $\bold{w}$, regardless of the value achieved by $\bold{X}$. However, it can easily happen that for this specific combination of $\bold{X}$ and $\bold{W}$ there are very few observed data points, thus making the estimation of the causal effect hard \citep{pearl2009causality}.
\begin{figure}[h]
\hspace{-0.6cm}
\includegraphics[width=1.05\linewidth]{causal_effect_example.png}
\caption{Left: Visualization of heterogeneous synthetic example \eqref{eq: causal example}. Middle: Gray points depict joint distribution of $(W, Y)$ conditionally on $\bold{X}\myeq\bold{x}$, for some choices of $\bold{x}$ indicated in the top left corner. Black curve indicates the true conditional mean $\E[Y\mid W\myeq w, \bold{X}\myeq\bold{x}]$, the blue curve represents the estimate obtained by DRF with response $(W, Y)$ and predictors $\bold{X}$ in combination with smoothing splines regression, whereas the red curve represents the estimate obtained by standard Random Forest. Right: The corresponding estimates for both methods of the causal effect $\E[Y\mid do(W\myeq w)]$ computed from \eqref{eq: causal effect}. The true causal effect is denoted by a black dashed curve.}
\label{fig: causal_effect}
\end{figure}
As an illustration, we consider the following synthetic data example, with continuous outcome $Y$, continuous univariate treatment $W$, $n=5000$ and $p=20$:
\begin{equation} \label{eq: causal example}
\bold{X} \sim U(0,5)^p,\quad W \mid \bold{X} \sim N(X_2, 1),\quad Y \mid \bold{X}, W \sim N(X_2 + X_1 \sin(W), 1).
\end{equation}
A visualization of the data can be seen on the left side of Figure \ref{fig: causal_effect}; treatment $W$ affects $Y$ non-linearly, $X_2$ is a confounding variable that affects the marginal distributions of $Y$ and $W$ and $X_1$ makes the treatment effect heterogeneous. The middle part of Figure \ref{fig: causal_effect} shows the conditional regression fits, i.e.\ the estimates of $\E[Y\mid W\myeq w, \bold{X}\myeq\bold{x}]$ as $w$ varies and $\bold{x}$ is fixed. We see that combination of DRF with response $(Y, W)$ and predictors $\bold{X}$ with the smoothing splines regression of $Y$ on $W$ (blue curve) is more accurate than the estimates obtained by standard Random Forest \citep{breiman2001random} with response $Y$ and predictors $(W, \bold{X})$ (red curve). Furthermore, we see that the former approach can extrapolate better to regions with small number of data points, which enables us to better estimate the causal effect $\E[Y \mid do(W \myeq w)]$ from \eqref{eq: causal effect}, by averaging the corresponding estimates of $\E[Y\mid W\myeq w, \bold{X}\myeq\bold{x}]$ over observed $\bold{x}$, as shown in the right plot of Figure \ref{fig: causal_effect}.
The conditional regression fit $\E[Y\mid\bold{W}\myeq\bold{w}, \bold{X}\myeq\bold{x}]$ is related to the concept of the conditional average treatment effect (CATE) as it quantifies the effect of $\bold{W}$ on $Y$ for the subpopulation for which $\bold{X}=\bold{x}$. There exist many successful methods in the literature for estimating the causal effects and the (conditional) average treatment effects for a wide range of settings \citep{abadie2006large, chernozhukov2018double, wager2018estimation, kunzel2019metalearners}. Due to its versatility, DRF can easily be used when the underlying assumptions of existing methods are violated, when some additional structure is given in the problem or for the general, nonparametric, settings \citep{imbens2004nonparametric, ernest2015marginal, kennedy2017nonparametric}. Appendix \ref{appendix: additional examples} contains additional comparisons with some existing methods for causal effect estimation.
\subsubsection{Births data}
We further illustrate the applicability of DRF for causality-related problems on the natality data obtained from the Centers for Disease Control and Prevention (CDC) website, where we have information about all recorded births in the USA in 2018. We investigate the relationship between the pregnancy length and the birthweight, an important indicator of baby's health. Not only is this relationship complex, but it also depends on many different factors, such as parents' race, baby's gender, birth multiplicity (single, twins, triplets...) etc. In the left two plots of Figure \ref{birthweight} one can see the estimated joint distribution of birthweight and pregnancy length conditionally on many different covariates, as indicated in the plot. The black curves denote the subsequent regression fit, based on smoothing splines and described in detail in Appendix \ref{appendix: simulation details}. In addition to the estimate of the mean, indicated by the solid curve, we also include the estimates of the conditional $0.1$ and $0.9$ quantiles, indicated by dashed curves, which is very useful in practice for determining whether a baby is large or small for its gestational age. Notice how DRF assigns less importance to the mother's race when the point of interest is a twin (middle plot), as in this case more weight is given to twin births, regardless of the race of the parents.
\begin{figure*}[h]
\hspace*{-0.7cm}
\includegraphics[width=1.06\textwidth]{births.png}
\caption{Left and middle: estimated relationship of pregnancy length and birthweight, conditionally on the criteria indicated in the upper left corner. Right: estimated interventional effect of twin birth on the birthweight for a fixed pregnancy length. In all plots the solid curves denote the estimated conditional mean and the dashed denote the estimated $0.1$ and $0.9$ quantiles.}
\label{birthweight}
\end{figure*}
Suppose now we would like to understand how a twin birth $T$ causally affects the birthweight $B$, but ignoring the obvious indirect effect due to shorter pregnancy length $L$. For example, sharing of resources between the babies might have some effect on their birthweight. We additionally need to be careful to adjust for other confounding variables $\bold{X}$, such as, for example, the parents' race, which can affect $B, T$ and $L$. We assume that this is represented by the following causal graph:
\begin{center}
\begin{tikzpicture}[
> = stealth,
shorten > = 1pt,
auto,
node distance = 3cm,
semithick
]
\tikzstyle{every state}=[
draw = black,
thick,
fill = white,
minimum size = 4mm
]
\node[circle, draw=black] (T) at (0,0) {$T$};
\node[circle, draw=black] (X) at (2,1) {$\boldsymbol{X}$};
\node[circle, draw=black] (L) at (2,-1) {$L$};
\node[circle, draw=black] (B) at (4, 0) {$B$};
\path[->] (T) edge node {} (L);
\path[->,line width=1.4pt] (T) edge node {} (B);
\path[dashed,->] (X) edge node {} (T);
\path[dashed,->] (X) edge node {} (L);
\path[dashed,->] (X) edge node {} (B);
\path[->] (L) edge node {} (B);
\end{tikzpicture}
\end{center}
In order to answer the above question, we investigate the causal quantity $\P(B \mid do(T\myeq t, L\myeq l))$. Even though one cannot make such do-intervention in practice, this quantity describes the total causal effect if the birth multiplicity and the length of the pregnancy could be manipulated and thus for a fixed pregnancy length $l$, we can see the difference in birthweight due to $T$. We compute this quantity as already stated above, by using DRF with subsequent regression fits (described in detail in Appendix \ref{appendix: simulation details}), which has the advantage of better extrapolating to regions with small probability, such as long twin pregnancies (see middle plot of Figure \ref{birthweight}). In the right plot of Figure \ref{birthweight} we show the mean and quantiles of the estimated interventional distribution and we see that, as one might expect, a twin birth causes smaller birthweight on average, with the difference increasing with the length of the pregnancy.
\subsection{Fairness}
Being able to compute different causal quantities with DRF could prove useful in a range of applications, including fairness \citep{kusner2017counterfactual}. We investigate the data on approximately $1$ million full-time employees from the 2018 American Community Survey by the US Census Bureau from which we have extracted the salary information and all covariates that might be relevant for salaries. In the bottom left plot of Figure \ref{fig: wage} one can see the distribution of hourly salary of men and women (on the logarithmic scale). The overall salary was scaled with working hours to account for working part-time and for the fact that certain jobs have different working hours. We can see that men are paid more in general, especially for the very high salaries. The difference between the median hourly salaries, a commonly used statistic in practice, amounts $17\%$ for this data set.
We would like to answer whether the observed gender pay gap in the data is indeed unfair, i.e.\ only due to the gender, or whether it can at least in part be explained by some other factors, such as age, job type, number of children, geography, race, attained education level and many others. Hypothetically, it could be, for example, that women have a preference for jobs that are paid less, thus causing the gender pay gap.
In order to answer this question, we assume that the data is obtained from the following causal graph, where $G$ denotes the gender, $W$ the hourly wage and all other factors are denoted by $\bold{X}$:
\begin{center}
\begin{tikzpicture}[
> = stealth,
shorten > = 1pt,
auto,
node distance = 3cm,
semithick
]
\tikzstyle{every state}=[
draw = black,
thick,
fill = white,
minimum size = 4mm
]
\node[circle, draw=black] (G) at (0,0) {$G$};
\node[circle, draw=black] (X) at (2,1) {$\boldsymbol{X}$};
\node[circle, draw=black] (W) at (4, 0) {$W$};
\path[->,line width=1.4pt] (G) edge node {} (W);
\path[->] (G) edge node {} (X);
\path[->] (X) edge node {} (W);
\end{tikzpicture}
\end{center}
i.e.\ $G$ is a source node and $W$ is a sink node in the graph.
In order to determine the direct effect of the gender on wage that is not mediated by other factors, we would like to compute the distribution of the nested counterfactual $W(\text{male},\, \bold{X}(\text{female}))$, which is interpreted in the data-generating process as the women's wage had they been treated in same way as men by their employers for determining the salary, but without changing their propensities for other characteristics, such as the choice of occupation. Therefore, it can be obtained from the observed distribution as follows:
\begin{align}
\P\left(W(\text{male},\, \bold{X}(\text{female}))\right) &= \int \P\left(W( G\myeq\text{male},\, \bold{X}\myeq\bold{x})\right)\P(\bold{X}\myeq\bold{x}\mid G\myeq\text{female})d\bold{x} \nonumber\\
&= \int \P\left(W\mid G\myeq\text{male},\, \bold{X}\myeq\bold{x}\right)\P(\bold{X}\myeq\bold{x}\mid G\myeq\text{female})d\bold{x},
\label{eq: nested_counterfactual}
\end{align}
Put in the language of the fairness literature, it quantifies the unfairness when all variables $\bold{X}$ are assumed to be resolving \citep{kilbertus2017avoiding}, meaning that any difference in salaries directly due to factors $\bold{X}$ is not viewed as gender discrimination. For example, one does not consider unfair if people with low education level get lower salaries, even if the gender distribution in this group is not balanced.
\begin{figure}[h]
\hspace*{-0.6cm}
\includegraphics[width=1.05\textwidth]{wage_data2.png}
\caption{Top row: Estimated joint distribution of wage and gender for some fixed values of other covariates $\bold{X}$ indicated in the top left part of each plot. Bottom row: observed overall distribution of salaries (left), estimated counterfactual distribution $\P\left(W(\text{male},\, \bold{X}(\text{female}))\right)$ of women's salaries (middle) and the quantile comparison of the counterfactual distribution of women's salaries and the observed distribution of men's salaries (right).}
\label{fig: wage}
\end{figure}
There are several ways how one can compute the distribution of $W(\text{male},\, \bold{X}(\text{female}))$ from \eqref{eq: nested_counterfactual} with DRF. The most straightforward option is to take $W$ as the response and $(G, \bold{X})$ as predictors in order to compute the conditional distribution $\P\left(W\mid G\myeq\text{male},\, \bold{X}\myeq\bold{x}\right)$.
However, with this approach it could happen that for predicting $\P\left(W\mid G\myeq\text{male},\, \bold{X}\myeq\bold{x}\right)$ we also assign weight to training data points for which $G\myeq\text{female}$. This happens if in some trees we did not split on variable $G$, which is likely, for example, if $\P(G=\text{male} \mid \bold{X}\myeq\bold{x})$ is low. Using salaries of both genders to estimate the distribution of men's salaries might be an issue if our goal is to objectively compare how women and men are paid.
Another approach is to take $(W, G)$ as a multivariate response and $\bold{X}$ as the predictors for DRF and thus obtain joint distribution of $(W, G)$ conditionally on the event $\bold{X}\myeq\bold{x}$. In this way we can also quantify the gender discrimination of a single individual with characteristics $\bold{x}$ by comparing his/her salary to the corresponding quantile of the salary distribution of people of the opposite gender with the same characteristics $\bold{x}$ \citep{plevcko2019fair}. This is interesting because the distribution of salaries, and thus also the gender discrimination, can be quite different depending on other factors such as the industry sector or job type, as illustrated for a few choices of $\bold{x}$ in the top row of Figure \ref{fig: wage}.
Finally, by averaging the DRF estimates of $\P\left(W\mid\bold{X}\myeq\bold{x},\, G\myeq\text{male}\right)$, conveniently represented via the weights, over different $\bold{x}$ sampled from the distribution $\P(\bold{X}\mid G\myeq\text{female})$, we can compute the distribution of the nested counterfactual $W(\text{male},\, \bold{X}(\text{female}))$. In the middle panel in the bottom row of Figure \ref{fig: wage} a noticeable difference in the means, also called natural direct effect in the causality literature \citep{pearl2009causality}, is still visible between the observed distribution of women's salaries and the hypothetical distribution of their salaries had they been treated as men, despite adjusting for indirect effects of the gender via covariates $\bold{X}$. By further matching the quantiles of the counterfactual distribution $\P\left(W(\text{male},\, \bold{X}(\text{female}))\right)$ with the corresponding quantiles of the observed distribution of men's salaries in the bottom right panel of Figure \ref{fig: wage}, we can also see that the adjusted gender pay gap even increases for larger salaries. Median hourly wage for women is still $11\%$ lower than the median wage for the hypothetical population of men with exactly the same characteristics $\bold{X}$ as women, indicating that only a minor proportion of the actually observed hourly wage difference of $17\%$ can be explained by other demographic factors.
\section{Conclusion}
We have shown that DRF is a flexible, general and powerful tool, which exploits the well-known properties of the Random Forest as an adaptive nearest neighbor method via the induced weighting function. Not only does it estimate multivariate conditional distributions well, but it constructs the forest in a model- and target-free way and is thus an easy to use out-of-the-box algorithm for many, potentially complex, learning problems in a wide range of applications, including also causality and fairness, with competitive performance even for problems with existing tailored methods.
\newpag
| e48dee1b159e6813828b07cad4c8191da6da9e17 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
One of the major motivations in the study of finite-temperature Quantum ChromoDynamics~(QCD) is to shed light on the phase structure of the strong interaction matter. Lattice-QCD data has confirmed the chiral and deconfinement phase transition is a crossover for small baryon chemical potential~\cite{Aoki:2006we,Borsanyi:2010bp}. The sign problem at large baryon chemical potential~\cite{Fodor:2001au,deForcrand:2010ys} prevents lattice-QCD from giving precise predictions about the phase transition at finite density. Model calculations based on NJL model, quark meson model, and various beyond mean field frameworks have all predicted that the chiral phase transition at finite density is a first order phase transition~\cite{Scavenius:2000qd,Schaefer:2007pw,Fukushima:2008wg,Herbst:2010rf}. Thermodynamic theorem then predicts a critical end point~(CEP) between the crossover and the first order phase transition, which is a second order phase transition. However, due to various approximations adopted in the model calculations, there is not an agreement on the location of the CEP on the phase diagram.
The exploration of the QCD phase diagram is also one of the most important goals for relativistic heavy-ion experiments. Through a systematic measurement over a range of beam energies, the beam energy scan~(BES) program makes it possible to search for the CEP in the QCD phase diagram~\cite{Aggarwal:2010wy,Luo:2012kja,Adamczyk:2013dal}. Besides the ongoing BES program at RHIC, several other programs at other facilities such as FAIR and NICA have also contributed to the searching of QCD critical end point. The related experiments are mainly driven by measurements of net-proton or net-charge multiplicity fluctuations~\cite{Stephanov:1998dy,Stephanov:1999zu,Stephanov:2008qz} which are expected to show characteristic non-monotonic behavior near the phase transition and especially near the CEP~\cite{Stephanov:2011pb,Luo:2017faz}. The fast dynamics in the fireball renders it difficult to bridge the gap between the experiment and the theories. Since from the theoretical aspect, the QCD phase structure is investigated in an equilibrium, long-lived, extremely large and homogeneous system. On the contrary, the fireball in the heavy ion collision is a highly dynamical system, characterized by very short lifetime, extremely small size and fast dynamical expansion. The finite-size effect and off-equilibrium effect prevent a divergent correlation length, and thus weaken the critical phenomenon which takes place in the equilibrium system. On one hand, the fast expansion and cooling during the evolution of the fireball tends to drive the system out of local equilibrium. On the other hand, the relaxation time diverges around the critical point~\cite{Berdnikov:1999ph}, the critical slowing down renders it harder for the system to reach local equilibrium. A thorough understanding of phase transitions in the dynamical environment is thus of fundamental necessary to make profound predictions from the BES project.
Various models have been applied to study the chiral phase transition and related critical phenomenon in an out-of-equilibrium system. In Ref.~\cite{Abada:1994mf, Abada:1996bw}, the authors investigate the chiral phase transition in a free-streaming quark-antiquark system by solving the Vlasov equation through the test particle method. The chiral fields are included considering their mean field values and their equations of motion. The Vlasov equation for the quark-antiquark system is also analytically solved in Ref.~\cite{Greiner:1996md}, assuming constant quark mass, and a shell-like structure at late evolution times in the center-of-mass (CM) frame is discovered. In Ref.~\cite{Yang:2003pz} the nonequilibrium and collision effects on the deconfinement phase transition is investigated by solving the Vlasov equation assuming Bjorken symmetry. The elastic two-body collisions for the quarks and antiquarks is included by simulating a Vlasov-type of equation with MonteCarlo test-particle approach~\cite{vanHees:2013qla,Meistrenko:2013yya,Wesp:2017tze}. Among the aforementioned works, although the force term is also considered in the test particle method~\cite{Abada:1994mf, Abada:1996bw,vanHees:2013qla,Meistrenko:2013yya,Wesp:2017tze}, its effect has not been carefully examined, which we find plays an important role in the evolution of the system around the phase transition.
In order to study the chiral phase transition in the non-equilibrium state, we investigate an expanding quark-anitquark gas system by solving the coupled Vlasov equation as well as the gap equation. This paper is arranged as follows. In section III, we introduce and analyze the coupled Vlasov equation and gap equation which we are going to solve, and give the related thermal quantities that can be obtained as momentum integrals of distribution function. In section III, we consider a longitudinal boost invariant and transverse rotational symmetric system, and derive the transport equation under such condition. In section IV, we present our numerical process to solve the coupled equations and then analyze the numerical result. In section V, we summarize this work and give a brief outlook.
\section{Vlasov equation and thermal quantities}
The partons in an off-equilibrium systems with background field and collisions can generally be described by the Vlasov equation
\begin{eqnarray}
\partial_t f^\pm \mp \boldsymbol{F} \cdot \boldsymbol{\nabla_p} f^\pm \pm \boldsymbol{v}\cdot \boldsymbol{\nabla_x} f^\pm = \mathcal{C}[f].
\end{eqnarray}
The above equation is applicable when the external fields and the interactions between the (quasi-)particles are sufficiently weak, so each particle can be considered to be moving along a classical trajectory, punctuated by rare collisions. As an example of an evolving global symmetry in an expanding parton system, we here consider the chiral symmetry in an off-equilibrium quark-antiquark system. At classical level, the velocity is $\boldsymbol{v}=\boldsymbol{p}/E_p$, the energy of the quasi-particle is $E_p=\sqrt{p^2+m(x)^2}$, where $m$ is the effective mass of the quark and antiquark, which is space-time dependent and is determined by the evolving chiral symmetry. The chiral mean field acts as a background field, and affects the motion of quarks through the gradient of the field energy $F=\boldsymbol{\nabla_x} E_p$, which is a continuous force on the quarks. The mass of the quasi-particles $m(x)$ is no longer a free parameter but is determined by the space-time dependent chiral symmetry. The constituent quark mass serves as the order parameter of chiral symmetry. Its temperature dependence in equilibrium can be found in lattice-QCD simulation~\cite{Aoki:2006we,Borsanyi:2010bp} and other model calculations. We here consider the $SU(2)$ Nambu--Jona-Lasinio Lagrangian (NJL) model~\cite{Nambu:1961tp,Klevansky:1992qe,Hatsuda:1994pi},
\begin{equation}
\mathcal{L}=\bar{\psi}(i\gamma^\mu \partial_\mu-m_0)\psi+G\Big[(\bar{\psi}\psi)^2+(\bar{\psi}i\gamma_5{\bf \tau}\psi)^2\Big],
\end{equation}
where $\psi=(u,d)^T$ is the two-component quark field in the flavor space, $m_0$ is the degenerate current mass of the quarks, $\tau$ is the Pauli matrix in the isospin space. The transport equations can be derived from a first principle theory or an effective model in the framework of Wigner function~\cite{Vasak:1987um,BialynickiBirula:1991tx, Zhuang:1995pd,Zhuang:1995jb,Zhuang:1998bqx,Guo:2017dzf}. At the classical level, the quarks are treated as quasi-particles, and the chiral field is approximated by mean field, hence the Vlasov equation is coupled to the gap equation~\cite{Guo:2017dzf}. The disoriented chiral condensate (DCC)~\cite{Felder:2000hj,Cooper:1994ji,Gavin:1993bs} is negligible here because it appears as a quantum effect, and the $\sigma$ condensate appears as the mean field and couples to the Vlasov equation through the inhomogeneous quark mass. In order to investigate the collisions and non-equilibrium effect, we take the relaxation time approximation for the collision terms. The distribution function of the quark/antiquark number density $f^\pm(t,\boldsymbol{x},\boldsymbol{p})$ satisfies the coupled Vlasov equation and gap equation,
\begin{eqnarray}
&& \partial_t f^\pm \mp \frac{\boldsymbol{\nabla_{r}}m^2}{2E_p} \cdot \boldsymbol{\nabla_p} f^\pm \pm \frac{\boldsymbol{p}}{E_p}\cdot \boldsymbol{\nabla_r} f^\pm = \mathcal{C}[f], \\
&& m \Big(1 + 2G \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3} \frac{f^+(x,\boldsymbol{p})-f^-(x,\boldsymbol{p})}{E_p} \Big) = m_0,
\end{eqnarray}
where the $+$($-$) sign stands for quark (antiquark).
In this work, we take the relaxation time approximation for the collision term, $\mathcal{C}[f]=-(f^\pm-f^{\pm}_\mathrm{eq})/\tau_\theta$, with $f^{\pm}_\mathrm{eq}$ represents the corresponding local-equilibrium distribution function, while $\tau_\theta$ is the relaxation time. It is worth noticing that one can make the substitution $\widetilde{f}^-(t,\boldsymbol{x},\boldsymbol{p}) \equiv 1 - f^-(t,\boldsymbol{x},-\boldsymbol{p}) $ which follows the same equation of motion as $f^+$. The transport equations can be further simplified by adopting the recombination $f = f^+ + \widetilde{f}^-$ and $g = f^+ - \widetilde{f}^-$. In such way, the evolution of $f$ and $g$ can be separated. Since $\widetilde{f}^-$ and $f^+$ satisfy the same transport equation, $f(t,\boldsymbol{x},\boldsymbol{p})$ and $g(t,\boldsymbol{x},\boldsymbol{p})$ also satisfy the same transport equation, while the gap equation depends only on $f$ but not $g$. One thus solve the coupled transport equation of distribution function $f(t,\boldsymbol{x},\boldsymbol{p})$ and $g(t,\boldsymbol{x},\boldsymbol{p})$ as well as the gap equation for a finite density system, and solve transport equation of distribution function $f(t,\boldsymbol{x},\boldsymbol{p})$ together with gap equation for a system with vanish baryon density,
\begin{eqnarray}
&& \partial_t f - \frac{\boldsymbol{\nabla_{r}}m^2}{2E_p} \cdot \boldsymbol{\nabla_p} f + \frac{\boldsymbol{p}}{E_p}\cdot \boldsymbol{\nabla_r} f = -\frac{f-f_\mathrm{eq}}{\tau_\theta}, \\
\label{transportf}
&& \partial_t g - \frac{\boldsymbol{\nabla_{r}}m^2}{2E_p} \cdot \boldsymbol{\nabla_p} g + \frac{\boldsymbol{p}}{E_p}\cdot \boldsymbol{\nabla_r} g = -\frac{g-g_\mathrm{eq}}{\tau_\theta}, \\
\label{gapeq}
&& m \Big(1 + 2G \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3} \frac{f(x,\boldsymbol{p})}{E_p}- 2G \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3} \frac{1}{E_p} \Big) = m_0.
\end{eqnarray}
In the numerical procedures, we consider zero density system in this paper and solve both the transport equation of $f(t,\boldsymbol{x},\boldsymbol{p})$ and the gap equation, and eventually get the time and space dependence of the distribution function and the quark mass. The second integral in the gap equation is the vacuum part, which has ultra-violet divergence and needs to be regularized. Here we take the hard cut-off regularization only for the vacuum part,
\begin{eqnarray}
\int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3} \frac{1}{E_p}
&=& 2\pi \int_0^{\Lambda}\int_0^{\Lambda} \frac{p_T\mathrm{d}p_T \mathrm{d}p_z}{(2\pi)^3} \frac{1}{E_p} \\
&=& \frac{1}{4\pi^2} \bigg[\Lambda (\sqrt{m^2+2\Lambda^2} -\sqrt{m^2+\Lambda^2})
+ m^2 \ln\Big(\frac{m}{\Lambda+\sqrt{m^2+\Lambda^2}}\Big) + (m^2+\Lambda^2) \ln\Big(\frac{\Lambda+\sqrt{m^2+2\Lambda^2}}{\sqrt{m^2+\Lambda^2}}\Big) \bigg].\nonumber
\end{eqnarray}
The same cutoff $\Lambda=496$~MeV is adopted for the longitudinal momentum and the transverse momentum in the integral. The NJL coupling constant is set to be $G=1.688/\Lambda^2$, so as to guarantee the quark mass $m\sim 300$~MeV in the vacuum. The momentum integral of the finite temperature part is free from divergence, and is left unregularized. Under such choice of parameters, the temperature dependence of the quark mass in an equilibrium system can be directly calculated from the gap equation~(\ref{gapeq}) by taking Fermi-Dirac distribution. The quark mass for chiral limit $m_0=0$ and real case $m_0=3.7$~MeV are presented in Fig.~\ref{phase transition}, the critical temperature is about $156$~MeV at vanishing baryon chemical potential.
\begin{figure}[H]\centering
\includegraphics[width=0.3\textwidth]{fig_orderparameter1.pdf}\qquad
\includegraphics[width=0.3\textwidth]{fig_orderparameter2.pdf}
\caption{Temperature dependence of quark mass in the real case and chiral limit for various baryon chemical potential in the equilibrium state.}
\label{phase transition}
\end{figure}
The chiral phase transition at vanishing baryon chemical potential is a crossover in the real case, and is a second order phase transition in the chiral limit. For a crossover, the phase transition can be defined at the maximum susceptibility $\mathrm{d}m/\mathrm{d}T$; for a second order phase transition, the susceptibility diverges at critical point. In an equilibrium system, the order parameter displays critical scaling $m\propto((T_c-T)/T_c)^\beta$ in the vicinity of a second order phase transition which can be described by critical exponent $\beta$. The divergence of the susceptibility is expected to affect the transport phenomenon of the system through the force term $\boldsymbol{\nabla_{r}}m^2 \cdot \boldsymbol{\nabla_p} f$, in which the force is provided by the ingredient of mass $\boldsymbol{\nabla_{r}}m^2$. In an equilibrium system with zero baryon density, the mass is determined by temperature alone, and the force can be expressed as $\boldsymbol{\nabla_{r}}m^2=2m(\mathrm{d}m/\mathrm{d}T)\boldsymbol{\nabla_{r}}T$. Since the susceptibility exhibits a peak around the phase transition, the temperature gradient $\boldsymbol{\nabla_{r}}T$ in a realistic system is nonzero, the force term is expected be very large at the phase transition point in the time-space. However, when the system is close to the second order phase transition, the relaxation time may diverge~\cite{Berdnikov:1999ph}, the critical slowing down takes place, makes it harder for the system to reach local equilibrium. When the system has not yet reached local equilibrium, the temperature is not well-defined, and neither for the expression $2m(\mathrm{d}m/\mathrm{d}T)\boldsymbol{\nabla_{r}}T$ of the force term. If the phase transition takes place in an out-of equilibrium system, the aforementioned effects of the force term may have been overestimated. In the following, we study the chiral phase transition in both local equilibrium and out-of-equilibrium systems, controlled by the relaxation time. Then we analyze the influence of the phase transition and the force term on the evolution of the system.
The thermodynamic quantities can be constructed from the distribution function. For single component medium, positive particle $f^+$ for example, the current and energy-momentum stress tensor are defined by:
\begin{eqnarray}
J_+^\mu(t,\boldsymbol{x}) &=& \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p} p^\mu f^+(t,\boldsymbol{x},\boldsymbol{p}) , \nonumber \\
T_+^{\mu\nu} (t,\boldsymbol{x})&=& \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p}p^\mu p^\nu f^+(t,\boldsymbol{x},\boldsymbol{p}),
\end{eqnarray}
which should satisfy that the conservation laws $\partial_\mu J^\mu=0$ and $\partial_\nu T^{\mu\nu}=0$. Since the total energy density and entropy is the sum of that of positive particles and negative particles, while the net-quark number density is the difference of positive particles and negative particles, the above equations can be rewritten using the redefined distribution function $f = f^+ + \widetilde{f}^-$ and $g = f^+ - \widetilde{f}^-$,
\begin{eqnarray}
\begin{split}
J^\mu(t,\boldsymbol{x}) =& \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p} p^\mu g(t,\boldsymbol{x},\boldsymbol{p}) , \\
T^{\mu\nu} (t,\boldsymbol{x})=& \int \frac{\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p}p^\mu p^\nu f(t,\boldsymbol{x},\boldsymbol{p}) , \\
S^\mu(t,\boldsymbol{x})=&-\int\frac{d^3\boldsymbol{p}}{(2\pi)^3E_p}p^\mu
\Big(f^+(t,\boldsymbol{x},\boldsymbol{p})\ln f^+(t,\boldsymbol{x},\boldsymbol{p})+(1-f^+(t,\boldsymbol{x},\boldsymbol{p}))\ln(1-f^+(t,\boldsymbol{x},\boldsymbol{p})) \\
&\qquad\qquad\qquad\;\;+f^-(t,\boldsymbol{x},\boldsymbol{p})\ln f^-(t,\boldsymbol{x},\boldsymbol{p})+(1-f^-(t,\boldsymbol{x},\boldsymbol{p}))\ln(1-f^-(t,\boldsymbol{x},\boldsymbol{p}))
\Big).
\end{split}
\end{eqnarray}
Using the Landau frame definition, the fluid velocity can be determined as the {\it time-like} eigenvector ($u^\mu u_\mu >0$) of the stress tensor $T^{\mu}_{\;\;\nu} u^\nu = \epsilon\; u^\mu$, with energy density $\epsilon$ being the corresponding eigenvalue. One could further obtain the particle number density and entropy density as $n =u_\mu J^\mu$ and $s=S^\mu u_\mu$. The time-space evolution of $\epsilon(t,\textbf{x})$, $n(t,\textbf{x})$ and $u^z(t,\textbf{x})$ would give us quantitative idea about how the system evolves.
By taking the relaxation time approximation for collision kernel, we also need the corresponding local-equilibrium distribution function for any given time and space point,
\begin{equation}
f^\pm_\mathrm{eq}(x,\boldsymbol{p})=\frac{1}{e^{(\pm u_\mu p^\mu + \mu_\mathrm{eq})/T_\mathrm{eq}}+1},
\label{feq}
\end{equation}
In the above distribution, the temperature and chemical potential are determined by matching the energy and number density, i.e. $\epsilon = \epsilon_\mathrm{eq}$ and $n = n_\mathrm{eq}$, where
\begin{eqnarray}
\epsilon_\mathrm{eq} &\equiv&
\int \frac{ (u \cdot p)^2\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p}
f_\mathrm{eq}(t,\boldsymbol{x},\boldsymbol{p}) \,, \\
n_\mathrm{eq} &\equiv&
\int \frac{ (u \cdot p)\mathrm{d}^3\boldsymbol{p}}{(2\pi)^3E_p}
g_\mathrm{eq}(t,\boldsymbol{x},\boldsymbol{p}) \,,
\end{eqnarray}
with $E_p=\sqrt{m(T_\mathrm{eq},\mu_\mathrm{eq})^2+p^2}$.
Similarly, we define the equilibrium limit of the entropy density as
\begin{equation}
s_\mathrm{eq}=-\int\frac{(u \cdot p) d^3\boldsymbol{p}}{(2\pi)^3E_p} \Big(
f_\mathrm{eq}^+ \ln f_\mathrm{eq}^+ +(1-f_\mathrm{eq}^+)\ln(1-f_\mathrm{eq}^+)
+f_\mathrm{eq}^- \ln f_\mathrm{eq}^- +(1-f_\mathrm{eq}^-)\ln(1-f_\mathrm{eq}^-)
\Big) .
\end{equation}
The difference between $s_\mathrm{eq}$ and the actual entropy density $s$ quantifies how close the system is to the equilibrium state.
In a zero chemical potential limit considered in this paper, one would automatically find $g=0$, hence $J=0$, $n=n_\mathrm{eq}=0$, $\mu_\mathrm{eq}$ = 0.
\section{Symmetry and Simplification}
For numerical simplicity, we will focus on the {\it longitudinal boost-invariant} and {\it transversal rotational-symmetric} systems, which is a good approximation for ultra-central relativistic heavy-ion collisions. Under such symmetries, two constraints are applied to the system, and the distribution function $f$ five degrees of freedom. We introduce a new set of coordinates $(\tau,\eta,\rho,\phi, p_\perp,\xi,\theta)$, the original coordinates and the new ones can be transformed through the following relation,
\begin{eqnarray}
t &=& \tau \cosh\eta, \qquad p_t~=~\sqrt{m(\tau,\rho)^2+p_\perp^2}\cosh(\xi+\eta),\nonumber\\
z &=& \tau \sinh\eta, \qquad p_z~=~\sqrt{m(\tau,\rho)^2+p_\perp^2}\sinh(\xi+\eta),\nonumber\\
x &=& \rho \cos\phi, \qquad~ p_x~=~p_\perp\cos(\phi+\theta),\nonumber\\
y &=& \rho \sin\phi, \qquad~ p_y~=~ p_\perp \sin (\phi+\theta),
\end{eqnarray}
where $\rho\in[0,+\infty)$, $p_\perp\in[0,+\infty)$ and $\xi \in(-\infty,+\infty)$. Under the new set of coordinate, the longitudinal boost-invariance and transversal rotational-symmetry of the distribution function can be translated into its independence of $\phi$ and $\eta$, namely $\partial_\phi f=\partial_\eta f=0$. The phase space of the distribution function becomes $(\tau,\rho, p_\perp,\xi,\theta)$. The transport equation~(\ref{transportf}) is then reduced to
\begin{eqnarray}
\partial_\tau f
+\frac{p_\perp\cos\theta}{E_p}\partial_\rho f
-\frac{m(\partial_\rho m)\cos\theta}{E_p}\partial_{p_\perp} f
-\tanh\xi\left(\frac{1}{\tau}+\frac{m(\partial_\tau m)}{p_\perp^2+m^2}\right)\partial_\xi f
-\frac{\sin\theta}{E_p}\left(\frac{p_\perp}{\rho}-\frac{m(\partial_\rho m)}{p_\perp}\right)\partial_\theta f
=-\frac{f-f_\mathrm{eq}}{\tau_\theta},
\label{transport_gubser}
\end{eqnarray}
where $f_\mathrm{eq}$ is the equilibrium distribution.
To clearly analyze the $\theta$-dependence of the distribution function, and to simplify the calculation, we take the Fourier expansion of $f(\tau,\rho, p_\perp,\xi,\theta)$ with respect to $\theta$, and also the equilibrium distribution function $f_\mathrm{eq}(\tau,\rho, p_\perp,\xi,\theta)$,
\begin{eqnarray}
\begin{split}
f(\tau,\rho, p_\perp,\xi,\theta) =&\; a_0(\tau,\rho, p_\perp,\xi)+2\sum_{n=1}^{\infty}\left[a_n(\tau,\rho, p_\perp,\xi)\cos(n\theta)+b_n(\tau,\rho, p_\perp,\xi)\sin(n\theta)\right],\\
f_\mathrm{eq}(\tau,\rho, p_\perp,\xi,\theta) =&\; A_0(\tau,\rho, p_\perp,\xi)+2\sum_{n=1}^{\infty}\left[A_n(\tau,\rho, p_\perp,\xi)\cos(n\theta)+B_n(\tau,\rho, p_\perp,\xi)\sin(n\theta)\right],
\end{split}\label{fourier}
\end{eqnarray}
where the Fourier coefficients are obtained by definition $a_n(\tau,\rho, p_\perp,\xi)\equiv (2\pi)^{-1}\int_{-\pi}^\pi f(\tau,\rho, p_\perp,\xi,\theta)\cos(n \theta) \mathrm{d}\theta$ and $b_n(\tau,\rho, p_\perp,\xi)\equiv(2\pi)^{-1}\int_{-\pi}^\pi f(\tau,\rho, p_\perp,\xi,\theta)\sin(n \theta) \mathrm{d}\theta$, and similarly for $A_n$ and $B_n$.
In a realistic system, one can further expect its symmetry under reflection along either $\boldsymbol{\hat{x}}$-, $\boldsymbol{\hat{y}}$-, or $\boldsymbol{\hat{z}}$-direction. Under such condition, one can show the $\theta$-odd components of $f$ and $f_\mathrm{eq}$ vanish, $b_n \equiv B_n \equiv 0$, as well as $a_n(-\xi) = a_n(\xi)$, $A_n(-\xi) = A_n(\xi)$. Substituting the Fourier expansion~(\ref{fourier}) back into the Vlasov equation~(\ref{transport_gubser}), we then reduce the Vlasov equation to the transport equations of the corresponding Fourier components $a_n$,
\begin{eqnarray}
\label{gubser_couple}
\begin{split}
&(\partial_\tau a_0)-\tanh\xi\left(\frac{1}{\tau}+\frac{m(\partial_\tau m)}{p_\perp^2+m^2}\right)(\partial_\xi a_0)+\frac{p_\perp}{E_p}(\partial_\rho a_1)-\frac{m(\partial_\rho m)}{E_p}(\partial_{p_\perp} a_1)+\frac{1}{E_p}\left(\frac{p_\perp}{\rho}-\frac{m(\partial_\rho m)}{p_\perp}\right)a_1
=-\frac{a_0-A_0}{\tau_\theta},\\
&(\partial_\tau a_n)-\tanh\xi\left(\frac{1}{\tau}+\frac{m(\partial_\tau m)}{p_\perp^2+m^2}\right)(\partial_\xi a_n)+\frac{p_\perp}{2E_p}\partial_\rho\Big( a_{n-1}+ a_{n+1}\Big)
-\frac{m(\partial_\rho m)}{2E_p}\partial_{p_\perp}\Big( a_{n-1}+a_{n+1}\Big)\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad~~~~
-\frac{1}{2E_p}\left(\frac{p_\perp}{\rho}-\frac{m(\partial_\rho m)}{p_\perp}\right)\Big((n-1)a_{n-1}-(n+1)a_{n+1}\Big)
=-\frac{a_n-A_n}{\tau_\theta}
\end{split}
\label{transport2}
\end{eqnarray}
where the energy of quasi-particle is $E_p(\tau,\rho)=\sqrt{m^2(\tau,\rho)+p_\perp^2}\cosh(\xi)$, with mass $m(\tau,\rho)$ obtained from the gap equation, which in the new coordinate becomes,
\begin{eqnarray}
m \Big(1 + 2N_dG \int \frac{p_\perp\mathrm{d}p_\perp\mathrm{d}\xi}{2(2\pi)^2}
\big(a_0(\tau,\rho, p_\perp,\xi)-1\big)\Big) = m_0.
\end{eqnarray}
The momentum integral in the gap equation only relates to the zeroth component $a_0$, while the first and second order Fourier components contribute to the currents and energy-momentum tensor.
The energy-momentum tensor is
\begin{eqnarray}
T^{\mu}_{\;\;\nu} = \left( \begin{array}{cccc}
T^{\tau\tau} & 0 & - T^{\tau\rho} & 0 \\
0 & -\tau^2T^{\eta\eta} & 0 & 0 \\
T^{\rho\tau} & 0 & -T^{\rho\rho} & 0 \\
0 & 0 & 0 & -\rho^2T^{\phi\phi} \\
\end{array}\right),
\end{eqnarray}
with the non-vanishing components are
\begin{eqnarray}
\begin{split}
T^{\tau\tau}(\tau,\rho) =& ~~~\int_pa_0(\tau,\rho, p_\perp,\xi) \left(m(\tau,\rho)^2+p_\perp^2\right)\cosh^2(\xi), \\
T^{\eta\eta}(\tau,\rho) =&\frac{1}{\tau^2}\int_p a_0(\tau,\rho, p_\perp,\xi) \left(m(\tau,\rho)^2+p_\perp^2\right)\sinh^2(\xi), \\
T^{\tau\rho}(\tau,\rho) =&~~~\int_pa_1(\tau,\rho, p_\perp,\xi) p_\perp\sqrt{m(\tau,\rho)^2+p_\perp^2}\cosh(\xi), \\
T^{\rho\rho}(\tau,\rho) =&~~~\int_p\left(a_0(\tau,\rho, p_\perp,\xi) +a_2(\tau,\rho, p_\perp,\xi)\right)p^2_\perp/2,\\
T^{\phi\phi}(\tau,\rho) =&\frac{1}{\rho^2}\int_p \left(a_0(\tau,\rho, p_\perp,\xi) -a_2(\tau,\rho, p_\perp,\xi)\right)p^2_\perp/2,
\end{split}
\label{tensor}
\end{eqnarray}
where $\int_p$ is the abbreviation for $\int \frac{p_\perp\mathrm{d}p_\perp\mathrm{d}\xi}{4(2\pi)^2}$. For this energy-momentum tensor, one can explicitly write down the flow velocity and energy density, being the time-lime eigenvector and the corresponding eigenvalue:
\begin{eqnarray}
\begin{split}
\epsilon&=\Big(T^{\tau\tau}-T^{\rho\rho}+\sqrt{(T^{\tau\tau}+T^{\rho\rho})^2-4(T^{\tau\rho})^2}\Big)/2,\\
u^\mu&\equiv \{u^\tau, 0, u^\rho, 0\}=
\bigg\{\Big( \frac{T^{\tau\tau}+T^{\rho\rho}}{2\sqrt{(T^{\tau\tau}+T^{\rho\rho})^2-4(T^{\tau\rho})^2}} + \frac{1}{2}\Big)^{1/2},0,
\Big( \frac{T^{\tau\tau}+T^{\rho\rho}}{2\sqrt{(T^{\tau\tau}+T^{\rho\rho})^2-4(T^{\tau\rho})^2}} - \frac{1}{2}\Big)^{1/2},
0\bigg\},
\label{eigen}
\end{split}
\end{eqnarray}
Noting that the velocity has only two non-zero components, $u^\tau$ and $u^\rho$, the equilibrium distribution function~(\ref{feq}) can be expressed as
\begin{eqnarray}
f_\mathrm{eq}(\tau,\rho, p_\perp,\xi,\theta)=\frac{1}{\exp\left(T_\mathrm{eq}^{-1} u_\tau \sqrt{m^2+p_\perp^2}\cosh\xi - T_\mathrm{eq}^{-1} u_\rho p_\perp\cos\theta\right)+1}.
\end{eqnarray}
\section{Numerical Procedure and Result}
In the numerical procedures, we solve the finite difference versions of transport equations. The distribution function is discretized on a fixed grid in the calculation frame. The phase space is discretized as follows, take 200 points for $\rho$ within the range $\rho/\rho_0\in[-3,3]$, 100 points for $p_T$ within $p_T/T_0\in[0,8]$, 100 points for $\xi$ within the range $\xi\in[0,6]$. The Fourier expansion of distribution function $f$ with respect to $\theta$ is taken with maximum $n=7$ to guarantee the convergence. To eschew the numerical instability around $\tau=0$, we take the initial time as $\tau_0=0.5$~fm, the time step in the evolution is taken to be $\mathrm{d}\tau=0.0005$~fm to guarantee the stability. The calculation at the discrete time step n+1 involves only quantities at the previous time step. At each time step, we solve both the transport equation and the gap equation, and eventually get the time and space dependence of the distribution function and the quark mass. This numerical framework is verified to be reliable by comparing the result with the analytical solution in a spherical symmetric system~\cite{Greiner:1996md}, as well as checking the conservation of the particle number and energy-momentum.
The initial state of the fireball is a highly off-equilibrium system, the evolution towards local equilibrium quark gluon plasma is also an interesting problem. However, this is not our concern in this paper, since the temperature is not well-defined in such initial stage, the discussion of phase transition is also questionable. We here discuss the evolution of the system from local equilibrium towards off-equilibrium state after the formation of quark gluon plasma. A local equilibrium initial state is adopted, and a local temperature could be assigned. The large gradient in the initial condition drives towards off-equilibrium state, while the collisions drives the system towards local equilibrium. In order to describe the hot chiral restored medium in the inner part and the cold chiral symmetry broken medium in the outer part, we choose a Gaussian temperature profile $T(\rho)=T_0\exp\left(-\rho^2/\rho_0^2\right)$ for the initial state, with $T_0=300$~MeV and $\rho_0=2$~fm. For the local equilibrium initial state, the distribution function is the Fermi-Dirac distribution, $a_0(\tau_0, \rho, p_\perp,\xi)=2(e^{E_p/T(\rho)}+1)^{-1}$, where energy is $E_p=\sqrt{m^2+p_\perp^2}\cosh\xi$, and $T(\rho)$ is the above initial temperature profile. One can easily check that all other Fourier components vanish, $a_i (\tau_0)=0$ for $i\geq 1$. In the real case, the current mass is chosen to be $m_0=3.7$~MeV, and in the chiral limit, we have the current mass $m_0=0$.
The chiral phase transition is characterized by the chiral order parameter $\sigma$, or the quark constituent mass. The constituent mass is generated by the gap equation at each space-time point, and enters the transport equation through three ways: the energy $E_p=\sqrt{m^2+p_\perp^2}\cosh\xi$, the evolution rate of constituent mass $\partial_\tau m$, and through the spatial gradient of mass $\partial_\rho m$. In order to illustrate the influence of the phase transition and the force term in the transport equation, we consider the following three different conditions. First, when solving the transport equation and the gap equation concurrently, the effect of the phase transition and the force terms are both taken into consideration. Second, for a comparison, we solve the transport equation alone and keep the quark mass as a constant, for instance $m=150$~MeV. In this case, there is no phase transition nor force term. Third, in order to further illustrate the influence of the force term, we solve the transport equation and the gap equation at each step, but ignore the force term in the transport equation, namely assuming $\partial_\tau m=0$ and $\partial_\rho m=0$. We also study the influence of out-of-equilibrium effect by comparing the results of different relaxation time, for small relaxation time the system stays close to local equilibrium; while for large relaxation time, the system is away from equilibrium.
\subsection{Thermodynamical quantities}
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{gub_t_real_0005.pdf}\qquad
\includegraphics[width=0.35\textwidth]{gub_t_real_10.pdf}
\caption{The energy density (top), entropy density (middle) and equilibrium temperature (bottom) in the expanding system. The left and right panel corresponds to small and large relaxation time. The lines are rainbow colored, representing different evolution time. In the figure of entropy density, the solid and dashed lines correspond to realistic and equilibrium entropy.}
\label{gub_thermal}
\end{figure}
First, we self-consistently solve the coupled transport equation as well as the gap equation with finite current mass, and adopt different relaxation time. At each time step, the distribution function is obtained by evolving the transport equation, the energy density and entropy density are calculated by definition Eq.~(\ref{current}) and Eq.~(\ref{tensor}). The temperature is obtained by matching the realistic energy density to that of the equilibrium state. The energy density, entropy and temperature are presented in Fig.~\ref{gub_thermal}. The left panel corresponds to the evolution with small relaxation time, and the right panel corresponds to those of large relaxation time. The lines are rainbow colored which represents different evolution time, from the red line to the purple line represent the initial distribution to the distribution at later time.
The initial condition of the system is chosen to be an equilibrium distribution, with a gaussian distribution temperature profile, the inner part (small $\rho$) has higher temperature and the outer part (large $\rho$) has lower temperature. With the expansion of the system, the temperature of the core area gradually decrease, and the temperature of the outer area increases. The initial large gradient of the energy density drives the system away from the local equilibrium, while the collisions bring the system back to equilibrium. The entropy tells whether the system has reached local equilibrium, the solid line represent the realistic entropy density and the dashed lines represent the entropy of the equilibrium state. Since the equilibrium state takes the maximum entropy. When the collisions are not strong enough, the system takes longer time in the out-of-equilibrium state, where the realistic entropy is smaller than that of the equilibrium state. If the relaxation time is small,
the collisions are strong enough to keep the system at local equilibrium, the entropy density of the state is the same as that of the equilibrium state.
\subsection{Constituent mass and Phase boundary}
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{gub_m_real_0005.pdf}
\includegraphics[width=0.35\textwidth]{gub_m_real_10.pdf}\\
\includegraphics[width=0.35\textwidth]{gub_m_chiral_0005.pdf}
\includegraphics[width=0.35\textwidth]{gub_m_chiral_10.pdf}
\caption{The evolution of quark mass $m_\psi(\tau,\rho)$ and transverse velocity $v_\rho(\tau,\rho)$ in the real case and chiral limit, with a comparison between large and small relaxation time. The lines are rainbow colored representing different evolution times. The phase transition point is marked out by the dotted line.}
\label{gub_mass}
\end{figure}
In the equilibrium state, the quark constituent mass $m_\psi$ serves as the order parameter of the chiral symmetry, which tells to what extent the chiral symmetry is broken. Here in an expanding quark-antiquark system, the space-time dependent quark mass $m_\psi(\tau,\rho)$ signals the chiral symmetry in the space-time. In the initial state, the inner part of the system has higher temperature and is in the chiral symmetry restored phase, the constituent mass is small; the outer part of the system has lower temperature and is in the chiral symmetry broken phase with large constituent mass. By solving together the coupled transport equation and gap equation, the evolution of constituent quark mass in the time-space can be obtained self-consistently. The evolution of both the real case and chiral limit are investigated for large and small relaxation time, the results are presented in Fig.~\ref{gub_mass}.
The upper two figures present the quark mass $m_\psi(\tau,\rho)$ and transverse velocity $v_\rho(\tau,\rho)\equiv(u^\rho/u^\tau)$ in the real case, the lower two figures correspond to those of chiral limit. With the expansion of the system, the quark mass in the inner area grows with time, indicating the gradually restoring of chiral symmetry; the quark mass of outer area decreases, indicating the breaking of chiral symmetry. In the equilibrium chiral phase transition, the phase transition point in the chiral limit is well-defined, while that of crossover does not has a strict definition, one of the usually used definition is the maximum of susceptibility $\mathrm{d}m_\psi/\mathrm{d}T$. In the expanding system, the phase transition point in the chiral limit can still be defined by the time-space point where that quark mass reached zero. For the real case, we here take phase boundary as the space-time point where the equilibrium temperature is around the critical temperature and $\mathrm{d}m/\mathrm{d}x$ takes the maximum. The phase transition points are marked out by the dotted lines in the Fig.~\ref{gub_mass}.
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{boundary.pdf}
\caption{The phase transition hypersurface obtained by self-consistent solution of a chiral limit system. Different lines represent different relaxation time.}
\label{gub_boundary}
\end{figure}
The phase diagram of equilibrium strong interaction matter in the $T-\mu$ plane is a map for the chiral symmetry. For non-equilibrium quark matter, the chiral symmetry breaking and restoration can be described by a phase diagram in space-time, with the phase boundary hypersurface indicates where the symmetry changes in the time-space. In the chiral limit, the phase transition hypersurface can still be defined by $x^\mu(m_\psi=0)$ no matter whether the system is in the local equilibrium or not. Fig.~\ref{gub_boundary} presents the hypersurface obtained by self-consistent solution of a chiral limit system, with different colored lines represent different relaxation time. The collisions cast the kinetic energy into internal energy, thus decelerates the expansion of the system. In the initial state $(\tau-\tau_0)/\tau_0=0$, the core area $\rho/\rho_0<0.8$ is in the chiral symmetry restored phase. With the expansion of the system, the core area gradually cools, the area of chiral symmetry restored phase shrinks. The free streaming system expands the fastest, after $(\tau-\tau_0)/\tau_0>2.2$, the chiral restored phase disappears. While the system with small relaxation time expands slower, it takes longer time for the chiral restored phase to disappear. It appears from the evolution of mass (Fig.~\ref{gub_mass}) and the phase boundary (Fig.~\ref{gub_boundary}) that the various relaxation time does not have obvious influence on the quark mass. Since the order parameter describes the long-range correlation and the overall property of the system, while the collision is the local process in a system and is a short range correlation, the collision does not have big impact on the global symmetry of the system.
\subsection{Kink in velocity}
When solving together the transport equation and the gap equation, kinks in the velocity $u_\tau$ and $u_\rho$ are discovered. The velocity along $\rho$-direction $v_\rho \equiv u_\rho/u_\tau$ for different current mass and relaxation time are also presented in Fig.~\ref{gub_mass}. As we have mentioned above, the quark mass enters the transport equation through both the energy $E_p$ and the force term $\boldsymbol{\nabla} E_p\cdot\boldsymbol{\nabla_p} f$. The gradient of the field energy acts as a continuous force on the quarks, describing the interaction between the quarks and the mean fields. This interaction changes the quarks' momenta. Although this force term is also considered in the simulations such as test particle method~\cite{Abada:1994mf, Abada:1996bw,vanHees:2013qla,Meistrenko:2013yya,Wesp:2017tze}, its influence on the phase transition and the expansion has not been carefully investigated. For a chiral phase transition at low density, it is either a crossover or a second order phase transition, in both case the order parameter changes continuously. In comparison, the gradient of order parameter diverges at a second order phase transition, hence the force could be extremely strong.
We first present the velocity $v_\rho$ in the scenario of constant mass so as to illustrate the influence of the phase transition on the evolution of the system, see Fig.~\ref{gub_mass_const}. We take constant homogeneous mass $m_\psi(\tau,\rho)=150$MeV and solve only the transport equation for a free streaming system, the force term vanishes since $\partial_\tau m=0$ and $\partial_\rho m=0$. In this scenario, the kink does not appear, indicating the kink arises from the inhomogeneous mass distribution and thus the force term.
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{gub_v_fixed_150.pdf}
\caption{Velocity in $\rho$-direction of the expansion with constant mass $m_\psi(\tau,\rho)=150$~MeV and infinite relaxation time. The lines are rainbow-colored represents different time in the evolution.}
\label{gub_mass_const}
\end{figure}
\begin{figure}[H]\centering
\includegraphics[width=0.6\textwidth]{gub_v_all.pdf}
\caption{The transverse velocity $v_\rho$ for real case and chiral limit, with different relaxation time. The solid lines are transverse velocity $v_\rho$ obtained by self-consistent solution. The dashed lines are transverse velocity $v_\rho$ obtained by solving together the transport equation and gap equation, but ignore the force terms in the transport equation.}
\label{gub_noforce}
\end{figure}
In order to further illustrate the influence of the force term, we now concurrently solve the transport equation and the gap equation, but removes the force terms in the transport equation by fixing $\partial_\tau m=0$ and $\partial_\rho m=0$. The evolution of mass and thermodynamic quantities has no obvious difference compared to those of self-consistent solutions, however, the velocity $v_\rho$ is quite different, which is presented as dashed lines in Fig.~\ref{gub_noforce}, with the solid lines are the velocity of self-consistent solution. As shown in each figures Fig.~\ref{gub_noforce}, when the force terms are ignored, the velocity has a bump around the phase transition. Namely if there is phase transition but no force term, the quark near the phase boundary are accelerated. In comparison, when the force term is considered, the quark near the phase boundary are slowed down. The appearance of kinks is closely related to the phase transition and the spatial distribution of the quark mass. The velocity kinks are more obvious in scenarios with small current mass or small relaxation time. This phenomenon can be understood as follows, since the force term is related to the susceptibility as well as the gradient of temperature $\boldsymbol{\nabla} m=(\mathrm{d}m/\mathrm{d}T)\boldsymbol{\nabla} T$, the phase transition of the chiral limit has divergent susceptibility while that of real case is finite, thus the kinks in the chiral limit panels are more obvious compared with those in real case panel. The expansion starts with equilibrium distribution and gradually becomes out-of-equilibrium due to the huge pressure gradient. With large relaxation time, the system spends longer time in the out-of-equilibrium state, the critical effect is further washed out.
\subsection{Distribution function}
The phase transition hypersurface affects the motion of quarks around it, and has an influence on the $p_T$ spectrum of the quarks in the system. This can be directly revealed from the distribution function. Under the given symmetry in section III, the distribution function is defined on a $4+1$d phase space, $(\rho,p_\perp,\theta,\xi)$ as well as $\tau$, where $\theta$ is the angle between the transverse coordinate $\rho$ and the transverse momentum $p_T$. The $\theta$-dependence has been expanded to a series of Fourier coefficients. The zeroth component $a_0$, which is obtained by integrating the distribution function $f(\rho, p_T, \theta, \xi)$ over $\theta$ is related to the density distribution. While the first component $a_1$ obtained by integrating $f(\rho, p_T, \theta, \xi)\cos\theta$ over $\theta$ gives hint to the direction of the particle velocity.
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{fig_fp_real_10_1.pdf}\qquad
\includegraphics[width=0.35\textwidth]{fig_fr_real_10_1.pdf}
\caption{Zeroth and the first Fourier components in the case the collisions are weak $\tau_\theta=10/T_\mathrm{eq}$, the solid lines are the distribution function from the self-consistent solution, the dashed lines are the distribution function when the force term is ignored. The left panel are $a_0$ and $a_1$ as functions of transverse momentum $p_T$ at various given transverse coordinates $\rho/\rho_0$, the right panel shows $a_0$ and $a_1$ as functions of transverse coordinates $\rho/\rho_0$ at various given transverse momentum $p_T$.}
\label{distribution1}
\end{figure}
In Fig.~\ref{distribution1}, we present the zeroth and the first Fourier components in the scenario with weak collisions, namely with large relaxation time $\tau_\theta=10/T_\mathrm{eq}$.
The solid lines correspond the self-consistent distribution function, while the dashed lines are the distribution function where the force term is ignored.
Coefficients $a_0$ and $a_1$ as a function of transverse momentum $p_T$ at various given transverse coordinates $\rho/\rho_0$ are presented in the left panel.
The right panel shows the coefficients $a_0$ and $a_1$ as a function of transverse coordinates $\rho/\rho_0$ at various given transverse momentum $p_T$.
At some evolution time $\tau=1.0$ fm or $(\tau-\tau_0)/\tau_0=1$, the phase transition takes place around the position $\rho/\rho_0\sim 0.7$, which is also the location of the kink in the velocity, see Fig.~\ref{gub_mass}.
From the left panel of Fig.~\ref{distribution1}, in the self-consistent solution (solid lines), the distribution function $a_1$ is negative at low $p_T$ for $\rho/\rho_0$ around 0.6 to 0.8, which means the particles with low momentum are bounced back by the phase transition ``wall''; while the dashed lines reveals that without force term, the particles cannot see the ``wall''.
This effect is also obvious from the $a_1$ in the right panel, for small momentums, the distribution function changes sign around $\rho/\rho_0=0.7$; for large momentums, the distribution function stays positive but has smaller values. This indicates that the particles with small momentum are bounced back by the phase transition wall around $\rho/\rho_0=0.7$, particles with large momentum go through the ``wall'', but have been slowed down.
The integral of $a_0(\rho, p_T, \xi)$ over $p_T$ at fixed $\rho$ corresponds to the number density of particles somewhere in the transverse plane, while integral of $a_0(\rho, p_T, \xi)$ over $\rho$ at fixed $p_T$ is the number density of particles of some fixed momentum. It can be observed from the left panel, whether the force term is ignored or not, the number density away from the ``wall'' ($\rho/\rho_0=0.7$) is similar, while the force term would collect more particles around the ``wall''. From the right panel, there are more low momentum particles kept inside the wall because of the force term.
\begin{figure}[H]\centering
\includegraphics[width=0.35\textwidth]{fig_fp_real_0005_1.pdf}\qquad
\includegraphics[width=0.35\textwidth]{fig_fr_real_0005_1.pdf}
\caption{Zeroth and the first Fourier components in the case the collisions are strong $\tau_\theta=0.005/T_\mathrm{eq}$, the solid lines are the distribution function from the self-consistent solution, the dashed lines are the distribution function when the force term is ignored. The left panel shows $a_0$ and $a_1$ as functions of transverse momentum $p_T$ at various given transverse coordinates $\rho/\rho_0$, the right panel shows $a_0$ and $a_1$ as functions of transverse coordinates $\rho/\rho_0$ at various given transverse momentum $p_T$.}
\label{distribution2}
\end{figure}
As is presented in Fig.~\ref{distribution2}, the effect of the phase transition ``wall'' is smoothed by the collision. Since the particles are relaxed into thermal distribution, the direction of a single particle is aligned along the collective velocity. With strong collisions, the influence of the force term on number density almost disappear, giving the same $a_0$ whether the force term is taken into consideration. The first order components $a_1$ is still affected by the force term. Inside the ``wall'' ($\rho/\rho_0<0.7$), the force term does not has any influence. Outside the ``wall'' ($\rho/\rho_0>0.7$), the distribution $a_1$ is smaller when the force term is considered, which means the particles are slowed down by the force term.
\section{Summary}
Quite different from the equilibrium thermodynamics, the realistic chiral phase transition in heavy ion collision takes place in a highly inhomogeneous, fast evolving dynamical system. This requires the understanding of the chiral phase transition in an expanding, out-of-equilibrium system.
In this work, we investigate the evolution of an expanding quark-antiquark system with self-consistently taking into account the dynamical quark constituent mass. In order to reduce the dimension of the phase space, we consider a longitudinal boost invariant and transversal rotational symmetric system, which is a good approximation for ultra-central heavy ion collisions. In the numerical process, both Vlasov and gap equations are solved concurrently, giving a self-consistent evolution of both the quark-antiquark distribution function and the quark constituent mass. The spacetime-dependent constituent mass serves as the chiral order parameter and affects the evolution of the quark distribution function through the force term. In order to investigate the off-equilibrium effects, we introduce relaxation time approximation for the collision term, and compare the local equilibrium and out-of-equilibrium result by considering small and large relaxation time.
The evolution of the quark mass illustrates the chiral phase transition, and defines the phase diagram in the space-time. A kink in the transverse velocity in observed around the phase transition boundary, which appears because the large force term around the phase transition. The kink is more obvious for smaller current mass and smaller relaxation time, which means that the crossover and non-equilibrium effect tends to smooth out the kink. The influence of the phase transition hypersurface is further investigated by directly analyzing the distribution function. It is observed that, the phase transition wall would bounce back the low momentum particles, thus gives a kink in the velocity and may enhance the low $p_T$ part of the momentum spectrum. The kink indicates that the parton decelerates when propagating from hot to cold region because of the increase in its effective mass. This effect may have impact on the produced hadrons and other observables in experiment, such as reducing the mean $p_T$ and the enhancement of particle production yield, which will be investigated in future work.
\noindent {\bf Acknowledgement}: The work is supported by the NSFC Grant Number 11890712. ZyW is supported by the Postdoctoral Innovative Talent Support Program of China, and SS by the Natural Sciences and Engineering Research Council of Canada.
\bibliographystyle{elsarticle-num}
| 274f17a34820cf100b9a3d679d25d63df75cc46c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A successful peg-hole insertion includes a series of sub-tasks like aligning the peg w.r.t hole and inserting in the hole. The task involves concurrent understanding of vision, depth and haptic modalities. This has been shown to pose significant difficulties for the robots. \cite{b1} claims that having a single modality, like vision, does not solve the peg-hole insertion problem accurately and precisely. This is the motivation behind our multi-modal perception system. In this work focus on using vision and optionally proprioception for robotic control, as well as trying to understand precision limits of a trainable vision-based controller.
Reinforcement Learning (RL) allows learning control policies that are difficult to model explicitly. These policies can generalise with respect to the geometry of the peg and the hole, as stated in \cite{b1}.
A UR5e robot used in our work is trained in the CoppeliaSim \cite{coppeliaSim} to learn a peg-hole insertion policy. We are using different model-free RL algorithms embedded in the Catalyst \cite{Catalyst} framework. In our experiments TD3\cite{TD3} converged to higher returns than DDPG\cite{DDPG} or SAC\cite{SAC} as shown in Fig. \ref{model_free} in section \ref{experiments}. Our deterministic control policy is learned from a multi-modal representation consisting of RGB-D images and proprioception. The policy generalises to different colours, as well as the 3D position of the peg and the block.
\section{Related Work and Background} \label{related_work}
\cite{RL_and_imitation} uses a hybrid approach of model-free deep RL and imitation learning to achieve zero-shot Sim2Real transfer on six different tasks with a single agent. The method shows that pre-training on expert demonstrations leads to faster training than learning the policy from scratch.
\cite{GraspGAN} learns to refine synthetic images (domain adaptation) to achieve good grasping success rate using less real-world data. \cite{13} trains a network that maps input image to the robot's motor torque. \cite{b1} uses self-supervision to learn a compact and multimodal representation of the sensory inputs for the peg-hole insertion task. While all aforementioned works incorporate vision, none of them is using eye-in-hand approach which we consider much more suitable specifically for RL, where a policy can learn to optimize observations along the way, making the vision "active".
\section{Problem Statement and Method Overview}\label{problem_statement}
The main objective of our work is to make the robot learn a peg-hole insertion policy, which is invariant to the pose and color of the mating part or peg. This is done by training a robot in the simulation and then directly transferring the model to the real robot. For a better Sim2Real transferability, we experiment with image augmentation and include different modalities. We benchmark our approach against a proposed baseline model, introduced in section \ref{models}.
\section{Baseline and RL Models}\label{models}
\subsection{Baseline Model}
Our baseline for the peg-in-hole insertion uses visual feedback together with classical computer vision tools to control the insertion of the peg. Firstly, the robot uses color information to segment out the mating part from the image. This allows to filter all the irrelevant visual noise. Secondly, the centre of the hole is being estimated. Finally, the end effector is guided to minimize the distance between the pose of the peg and the estimated pose of the hole.
\subsection{Reinforcement Learning}
We employ the actor-critic method, where both the actor and the critic are parameterized using neural networks.
\begin{equation}
R_{total} = R_{distance} + R_{success} + R_{collision} + R_{time}
\label{reward_design}
\end{equation}
The total reward stated in Eq. \ref{reward_design} for each step consists of following components.
\begin{itemize}
\item \textbf{Distance Reward:} Based on the distance between the the peg and the target (the hole of the mating part). The smaller the distance, the higher the reward.
\item \textbf{Success Reward:} A sparse reward based on the distance between the peg and the desired position in the hole. The agent is being rewarded once the peg gets very close to the target (specified by a threshold).
\item \textbf{Collision Reward:} To discourage the robot from colliding with the block, a negative reward is obtained every time the agent collides with the mating part.
\item \textbf{Time Reward:} To coerce the agent to complete the task as fast as possible, it obtains a small penalty per every step.
\end{itemize}
The scene in the CoppeliaSim and the real robot in starting position are shown in Fig. \ref{vrep}. On the real robot we use RealSense D435 RGB-D active stereo sensor. The mating part is being randomly placed on the table (always fully or partially visible for the robot in the starting position), imitating a flexible assembly process.
\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{figures/scene.png}
\caption{\textit{Left:} Rendered simulator scene from CoppeliaSim showing the UR5e robot with Robotiq gripper, a peg and a block. \textit{Right:} Real UR5e robot in our lab.}
\label{vrep}
\end{figure}
\section{Experiments: Results} \label{experiments}
First, we train the agent in the simulation. The learned policy is directly transferred to our real UR5e robot. Evaluation of the models is shown in Table \ref{tab:metrics}. We did a total of $30$ rollouts for each model, in which the mating part pose is changed for every 5 rollouts. These 6 mating poses are predefined, with a maximum distance of $39$ cm and $21$ cm between the poses in x- and y-axis respectively (in the plane of the table). The robot starts at a predefined home position for every rollout in both simulation and real world. The comparison of different model-free RL algorithms can be found in Fig. \ref{model_free}. Fig. \ref{ablation} shows which modalities are most important for the performance in the simulation.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{figures/model-comp.png}
\caption{ TD3 yielding higher return than SAC and DDPG.}
\label{model_free}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{figures/ablation.png}
\caption{Ablation study showing different configurations of training in learning a peg-hole insertion policy in simulation. It can be seen that all the configurations converged to successful insertion in simulation. The performance of these configurations on real robot can be seen in Table \ref{tab:metrics}. }
\label{ablation}
\end{figure}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{Model Type}} & \multicolumn{2}{c|}{\textbf{Mean Error {[}mm{]}}} & \multirow{2}{*}{\textbf{Insertion {[}\%{]}}} \\ \cline{2-3}
& x & y & \\ \hline
Baseline & \hspace{1mm} $2.4\pm1.7$ & \hspace{0.1mm} $0.3\pm3.0$ & $100$ \\ \hline
RGB w/ augs & $-2.0\pm3.8$ &$-9.8\pm11.2$ & $55$
\\ \hline
RGB-D w/ augs & \hspace{1mm} $17.7\pm12.3$ & \hspace{0.1mm} $2.7\pm1.8$ & $0$
\\ \hline
RGB-D w/o augs & $-124.1\pm166.9$ & \hspace{0.1mm} $65.6\pm94.4$ & $0$
\\ \hline
\end{tabular}
\caption{\textbf{Mean insertion error and insertion rate on the real robot.} The mean error between the ground truth position of the inserted peg versus the actual position at the end of an episode. Insertion rate is the percentage of the rollouts which end with successful insertion. The mean is taken over $30$ runs for different positions of the mating part on the table.}
\label{tab:metrics}
\end{table}
\section{Discussion and Conclusion}\label{conclusion}
Our initial hypothesis was that image augmentation is crucial for successful transfer from simulation to the real-world. Table \ref{tab:metrics} shows that augmentations improve the precision significantly. However, the RGB-D model still does not achieve steady, successful insertions. That could be because the depth image has a very specific structural noise (depth values are computed from stereo cameras) that can hardly be modelled with standard augmentation algorithms. Thus, RGB-only model shows much better performance. Additionally, the ablation study in Fig. \ref{ablation} shows that proprioceptive features are not critical for model performance for this task, neither in simulation, nor in real-life. Finally, we show that a purely eye-in-hand, image-based controller can be trained in simulation to perform peg-hole insertion with sub-centimetre accuracy. However, we still experience instabilities caused by changing light conditions and other environmental factors, that can be possibly improved by enhancing rendering in simulation and richer augmentation.
For the future work we consider enhancing the vision-based controller with force-based one (trained end-to-end, as a single model), that can perform precise insertions with large initial uncertainty.
| ca696c5cfd6742bc130782d9a1a3f0c953d07728 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
There has been an increasing trend in recent years to leverage machine learning methods to automate decision-making processes. However, for many applications, the adoption of such methods cannot rely solely on their prediction performance. For example, the European Union's General Data Protection Regulation, which became enforceable on 25 May 2018, introduces a right to explanation for all individuals so that they can obtain ``meaningful explanations of the logic involved'' when automated decision-making has ``legal effects'' on individuals or similarly ``significantly affecting'' them\footnote{\url{https://ec.europa.eu/info/law/law-topic/data-protection_en}}. Therefore, in addition to their prediction performance, machine learning methods have to be assessed on how they can supply their decisions with explanations.
The performance of a machine learning method can be assessed by the extent to which it correctly predicts unseen instances. A metric like the accuracy score commonly measures the performance of a classification model. However, there is no standard approach to assess explainability.
First, there is no mathematical definition of explainability. A definition proposed by~\cite{Miller19} states that the higher the explainability of a machine learning algorithm, the easier it is for someone to comprehend why certain decisions or predictions have been made. Second, there are several methods belonging to different categories (explainability-by-design, post-hoc model-specific explainability and post-hoc model-agnostic explainability)~\cite{Du20}, which provide their own form of explanations.
The requirements for explainable machine learning methods are dependent upon the application and to whom the explanations are intended for~\cite{Tomsett18,Bohlender19}. In order to match these requirements and conduct experiments to validate the usefulness of the explanations by the end-users, there is a need to have a comprehensive assessment of the explainability of the existing methods.~\citeauthor{Doshi17} [2017] claim that creating a shared language is essential for the evaluation and comparison of machine learning methods, which is currently challenging without a set of explanation characteristics. As far as we have seen, there is no existing framework which defines a set of explanation characteristics that systematize the assessment of the explainability of existing machine learning methods.
Hence, in this paper, we propose a new framework to assess and benchmark the performance-explainability characteristics of machine learning methods. The framework hypothesizes a set of explanation characteristics, and as emphasized in~\cite{Wolf19}, focuses on what people might need to understand about machine learning methods in order to act in concert with the model outputs. The framework does not claim to be exhaustive and excludes application-specific implementation constraints like time, memory usage and privacy. It could be a basis for the development of a comprehensive assessment of the machine learning methods with regards to their performance and explainability and for the design of new machine learning methods.
Due to space constraint, we limit the illustration of the use of the framework to one category of machine learning methods and we choose the Multivariate Time Series (MTS) classifiers. Multivariate data which integrates temporal evolution has received significant interests over the past decade, driven by automatic and high-resolution monitoring applications (e.g. healthcare~\cite{Li18}, mobility~\cite{Jiang19}, natural disasters~\cite{Fauvel20}). Moreover, the available explainability solutions to support the current state-of-the-art MTS classifiers remain limited, so this category of methods appears meaningful to assess for us.
The contributions of this paper are the following:
\begin{itemize}
\vspace{-0.4em}
\item We present a new performance-explainability analytical framework to assess and benchmark machine learning methods;
\vspace{-0.3em}
\item We detail a set of characteristics that systematize the performance-explainability assessment of existing machine learning methods;
\vspace{-0.3em}
\item We illustrate the use of the framework by benchmarking the current state-of-the-art MTS classifiers.
\vspace{-0.3em}
\end{itemize}
\section{Related Work}
In this section, we first position this paper in the related work and introduce the different categories of explainability methods as a background to the notions that will be discussed in the framework. Then, we present the state-of-the-art machine learning methods that will be used to illustrate the framework, i.e. MTS classifiers.
\vspace{-0.3em}
\subsection{Explainability}
Multiple taxonomies of explainability methods have been derived from different frameworks~\cite{Guidotti2018,Ventocilla18,Du20}.
However, none of them defines a set of explanation characteristics that systematize the assessment of the explainability of existing machine learning methods.
~\cite{Guidotti2018} provides a classification of the main problems addressed in the literature with respect to the notion of explanation and the type of machine learning systems. ~\cite{Ventocilla18} proposes a high-level taxonomy of interpretable and interactive machine learning composed of six elements (Dataset, Optimizer, Model, Predictions, Evaluator and Goodness).
And,~\cite{Du20} categorizes existing explainability methods of machine learning models into either by design or post-hoc explainability.
As our framework aims to cover all types of methods, we do not present the frameworks focusing on a particular type of explainability methods (e.g.~\cite{Lundberg17,Ancona18,Henin19}).
A five-step method to understand the requirements for explainable AI systems has been published in~\cite{Hall19}. The five steps are: explainee role definition, explanation characteristics identification, requirements collection, existing methods assessment and requirements/existing methods mapping. Our framework can be positioned as a further development of the fourth step of the method by detailing a set of explanations characteristics that systematize the assessment of existing methods. Our framework does not include application-specific implementation constraints like time, memory usage and privacy.
As a background to the notions that will be discussed in the framework, we introduce the three commonly recognized categories (explainability-by-design, post-hoc model-specific explainability and post-hoc model-agnostic explainability)~\cite{Du20} to which all of the explainability methods are belonging to.
First, some machine learning models provide explainability-by-design. These self-explanatory models incorporate explainability directly to their structures. This category includes, for example, decision trees, rule-based models and linear models.
Next, post-hoc model-specific explainability methods are specifically designed to extract explanations for a particular model. These methods usually derive explanations by examining internal model structures and parameters. For example, a method has been designed to measure the contribution of each feature in random forests~\cite{Palczewska13}; and another one has been designed to identify the regions of input data that are important for predictions in convolutional neural networks using the class-specific gradient information~\cite{Selvaraju19}.
Finally, post-hoc model-agnostic explainability methods provide explanations from any machine learning model. These methods treat the model as a black-box and does not inspect internal model parameters. For example, the permutation feature importance method~\cite{Altmann10} and the methods using an explainable surrogate model~\cite{Lakkaraju17,Lundberg17,Ribeiro18,Guidotti19} belong to this category.
The explainability methods presented reflect the diversity of explanations generated to support model predictions, therefore the need for a framework in order to benchmark the machine learning methods explainability. The next section present the MTS classifiers that will be used to illustrate the framework.
\vspace{-0.3em}
\subsection{Multivariate Time Series Classifiers}
\label{rw_MTS}
The state-of-the-art MTS classifiers consist of a diverse range of methods which can be categorized into three families: similarity-based, feature-based and deep learning methods.
Similarity-based methods make use of similarity measures to compare two MTS.
Dynamic Time Warping (DTW) has been shown to be the best similarity measure to use along the k-Nearest Neighbors (k-NN)~\cite{Seto15}.
There are two versions of kNN-DTW for MTS: dependent (DTW$_{D}$) and independent (DTW$_{I}$). Neither dominates over the other~\cite{Shokoohi17} from an accuracy perspective but DTW$_{I}$ allows the analysis of distance differences at feature level.
Next, feature-based methods include shapelets (gRSF~\cite{Karlsson16}, UFS~\cite{Wistuba15}) and bag-of-words (LPS~\cite{Baydogan16}, mv-ARF~\cite{Tuncel18}, SMTS~\cite{Baydogan14}, WEASEL+MUSE~\cite{Schafer17}) models.
WEASEL+MUSE shows better results compared to gRSF, LPS, mv-ARF, SMTS and UFS on average (20 MTS datasets). WEASEL+MUSE generates a bag-of-words representation by applying various sliding windows with different sizes on each discretized dimension (Symbolic Fourier Approximation) to capture features (unigrams, bigrams, dimension identification). Following a feature selection with chi-square test, it classifies the MTS based on a logistic regression classifier.
Then, deep learning methods use Long-Short Term Memory (LSTM) and/or Convolutional Neural Networks (CNN). According to the results published, the current state-of-the-art model (MLSTM-FCN) is proposed in~\cite{Karim19} and consists of a LSTM layer and a stacked CNN layer along with Squeeze-and-Excitation blocks to generate latent features.
Therefore, we choose to benchmark the performance-explainability of the best-in-class for each similarity-based, feature-based and deep learning category (DTW$_{I}$, WEASEL+MUSE and MLSTM-FCN classifiers). The next section introduces the performance-explainability framework, which is illustrated with the benchmark of the best-in-class MTS classifiers in section~\ref{sec:results}.
\vspace{-0.5em}
\section{Performance-Explainability Framework}
The framework aims to respond to the different questions an end-user may ask to take an informed decision based on the predictions made by a machine learning model: \textit{What is the level of performance of the model? Is the model comprehensible? Is it possible to get an explanation for a particular instance? Which kind of information does the explanation provide? Can we trust the explanations? What is the target user category of the explanations?} The performance-explainability framework that we propose is composed of the following components, which will also be translated into terms specific to our application (MTS classifiers) whenever relevant:
\vspace{-0.3em}
\paragraph{Performance} \textit{What is the level of performance of the model?}
The first component of the framework characterizes the performance of a machine learning model. Different methods (e.g. holdout, k-fold cross-validation) and metrics (e.g. accuracy, F-measure, Area Under the ROC Curve) exist to evaluate the performance of a machine learning model~\cite{Witten16}. However, there is no consensus on an evaluation procedure to assess the performance of a machine learning model. Recent work suggests that the definition of such an evaluation procedure necessitates the development of a measurement theory for machine learning~\cite{Flach19}. Many of the problems stem from a limited appreciation of the importance of the \textit{scale} on which the evaluation measures are expressed.
Then, in current practices, the choice of a metric to evaluate the performance of a machine learning model depends on the application. According to the application, a metric aligned with the goal of the experiments is selected, which prevents the performance comparison of machine learning models across applications.
Therefore, the performance component in the framework is defined as a first step towards a standard procedure to assess the performance of machine learning models. It corresponds to the relative performance of a model on a particular application. More specifically, it indicates the relative performance of the models as compared to the state-of-the-art model on a particular application and an evaluation setting. This definition allows the categorization of the models' performance on an application and an evaluation setting. In the case of different applications with a similar machine learning task, the performance component can give the list of models which outperformed current state-of-the-art models on their respective application. Thus, it points to certain models that could be interesting to evaluate on a new application, without providing guarantee that these models would perform the same on this new application. We propose an assessment of the performance in three categories:
\begin{enumerate}
\vspace{-0.4em}
\item[$\bullet$] \textit{Best}: best performance. It corresponds to the performance of the first ranked model on the application following an evaluation setting (models, evaluation method, datasets);
\vspace{-0.3em}
\item[$\bullet$] \textit{Similar}: performance similar to that of the state-of-the-art models. Based on the same evaluation setting, it corresponds to all the models which do not show a statistically significant performance difference with the second ranked model. For example, the statistical comparison of multiple classifiers on multiple datasets is usually presented on a critical difference diagram~\cite{Demsar06};
\vspace{-1em}
\item[$\bullet$] \textit{Below}: performance below that of the state-of-the-art models. It corresponds to the performance of the remaining models with the same evaluation setting.
\end{enumerate}
\paragraph{Model Comprehensibility} \textit{Is the model comprehensible?}
The model comprehensibility corresponds to the ability for the user to understand how the model works and produces certain predictions. Comprehensibility is tightly linked to the model complexity; yet, there is no consensus on model complexity assessment~\cite{Guidotti2018}. Currently, two categories of models are commonly recognized: ``white-box'' models, i.e. easy-to-understand models, and ``black-box'' models, i.e. complicated-to-understand models~\cite{Lipton16}. For example, many rule-based models and decision trees are regarded as ``white-box'' models while ensemble methods and deep learning models are ``black-box'' models. Not all rule-based models or decision trees are ``white-box'' models. Cognitive limitations of humans place restrictions on the complexity of the approximations that are understandable to humans. For example, a decision tree with a hundred levels cannot be considered as an easy-to-understand model~\cite{Lakkaraju17}.
Nevertheless, the distinction between ``white-box'' models and ``black-box'' models is clear among the machine learning methods of this paper. The state-of-the-art MTS classifiers are all ``black-box'' except one which is an easy-to-understand similarity-based approach. Therefore, due to space limitation, we propose a first assessment of the comprehensibility in two categories and we plan to further elaborate this component in future work:
\begin{enumerate}
\item[$\bullet$] \textit{Black-Box}: ``black-box'' model, i.e. complicated-to-understand models;
\vspace{-0.3em}
\item[$\bullet$] \textit{White-Box}: ``white-box'' model, i.e. easy-to-understand models.
\end{enumerate}
\paragraph{Granularity of the Explanations} \textit{Is it possible to get an explanation for a particular instance?}
The granularity indicates the level of possible explanations. Two levels are generally distinguished: global and local~\cite{Du20}. Global explainability means that explanations concern the overall behavior of the model across the full dataset, while local explainability informs the user about a particular prediction. Some methods can provide either global or local-only explainability while other methods can provide both (e.g. decision trees). Therefore, we propose an assessment of the granularity in three categories:
\begin{enumerate}
\item[$\bullet$] \textit{Global}: global explainability;
\vspace{-0.3em}
\item[$\bullet$] \textit{Local}: local explainability;
\vspace{-0.3em}
\item[$\bullet$] \textit{Global \& Local}: both global and local explainability.
\end{enumerate}
\paragraph{Information Type} \textit{Which kind of information does the explanation provide?}
The information type informs the user about the kind of information communicated. The most valuable information is close to the language of human reasoning, with causal and counterfactual rules~\cite{Pearl18}. Causal rules can tell the user that certain observed variables are the causes of specific model predictions. However, machine learning usually leverages statistical associations in the data and do not convey information about the causal relationships among the observed variables and the unobserved confounding variables. The usual statistical associations discovered by machine learning methods highly depend on the machine learning task. Therefore, we first give a generic high-level definition of the information type and then we detail and illustrate it for the application case of this paper (MTS classification). We propose a generic assessment of the information type in 3 categories from the least to the most informative:
\begin{enumerate}
\item[$\bullet$] \textit{Importance}: the explanations reveal the relative importance of each dataset variable on predictions. The importance indicates the statistical contribution of each variable to the underlying model when making decisions;
\vspace{-0.3em}
\item[$\bullet$] \textit{Patterns}: the explanations provide the small conjunctions of symbols with a predefined semantic (patterns) associated with the predictions;
\vspace{-0.3em}
\item[$\bullet$] \textit{Causal}: the most informative category corresponds to explanations under the form of causal rules;
\end{enumerate}
In this paper, the issue of Multivariate Time Series (MTS) classification is addressed. A MTS $M=\{x_1,...,x_d\} \in \mathcal{R}^{d*l}$ is an ordered sequence of $d \in \mathcal{N}$ streams with $x_i=(x_{i,1},...,x_{i,l})$, where $l$ is the length of the time series and $d$ is the number of multivariate dimensions. Thus, considering the MTS data type, the information can be structured around the features, i.e. the observed variables, and the time. We propose to decompose the 3 categories presented into 8 categories. In addition, we will illustrate each of these categories with an application in the medical field. Figure~\ref{fig:framework_MTS_example} shows the first MTS of the UEA Atrial Fibrilation~\cite{Bagnall18} test set that belongs to the class \textit{Non-Terminating Atrial Fibrilation}. This MTS is composed of two dimensions (two channels ECG) with a length of 640 (5 second period with 128 samples per second). It is worth noting that the explanations provided to illustrate each category are assumptive rather than validated, they are given as illustrative in nature.
\begin{figure}[!htpb]
\centering
\includegraphics[width=.95\linewidth]{./Images/MTS_AtrialFibrilation.png}
\caption{The first MTS sample of the UEA Atrial Fibrilation test set. It belongs to the class \textit{Non-Terminating Atrial Fibrilation} and is composed of two channels ECG on a 5 second period (128 samples per second).}
\label{fig:framework_MTS_example}
\vspace{-1.5em}
\end{figure}
\begin{table*}[!htpb]
\centering
\scriptsize
\begin{threeparttable}
\caption{Summary of framework results of the state-of-the-art MTS classifiers.}
\label{tab:MTS}
\vspace{-1em}
\begin{tabularx}{.71\linewidth}{l>{\centering}m{3cm}>{\centering}m{3cm}>{\centering\arraybackslash}m{3cm}}
\toprule
\textbf{} & \textbf{Similarity-Based} & \textbf{Feature-Based} & \textbf{Deep Learning}\\
\textbf{} & DTW$_{I}$ & WEASEL+MUSE with SHAP & MLSTM-FCN with SHAP\\
\midrule\midrule[.1em]
Performance & Below\tnote{1} & Similar\tnote{1} & Best\tnote{1}\\
Comprehensibility & White-Box & Black-Box & Black-Box\\
Granularity & Local & Both Global \& Local & Both Global \& Local\\
Information & Features+Time & Features+Time & Features+Time\\
Faithfulness & Perfect & Imperfect & Imperfect\\
User & Domain Expert & Domain Expert & Domain Expert\\
\bottomrule
\end{tabularx}
\begin{tablenotes}
\item[1] Predefined train/test splits and an arithmetic mean of the accuracies on 35 public datasets [Karim et al., 2019]. As presented in section~\ref{rw_MTS}, the models evaluated in the benchmark are: DTW$_{D}$, DTW$_{I}$, gRSF, LPS, MLSTM-FCN, mv-ARF, SMTS, UFS and WEASEL+MUSE.
\end{tablenotes}
\vspace{-1.6em}
\end{threeparttable}
\end{table*}
\begin{enumerate}
\item[$\bullet$] \textit{Features} (\textit{Importance}): the explanations reveal the relative importance of the features on predictions. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the channel 2 has a greater importance on the prediction than the channel 1;
\vspace{-0.3em}
\item[$\bullet$] \textit{Features + Time} (\textit{Importance}): the explanations provide the relative importance of the features and timestamps on predictions. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the channel 2 has a greater importance on the prediction than the channel 1 and that the timestamps are in increasing order of importance on the prediction;
\vspace{-0.3em}
\item[$\bullet$] \textit{Features + Time + Values} (\textit{Importance}): in addition to the relative importance of the features and timestamps on predictions, the explanations indicate the discriminative values of a feature for each class. For example in Figure~\ref{fig:framework_MTS_example}, the explanations could give the same explanations as the previous category, plus, it could tell the user that the timestamps with the highest importance are associated with high values (values above 0.15) on the channel 2;
\vspace{-0.3em}
\item[$\bullet$] \textit{Uni Itemsets} (\textit{Patterns}): the explanations provide patterns under the form of groups of values, also called itemsets, which occur per feature and are associated with the prediction. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the following itemsets are associated with the prediction: \{channel 1: extremely high value (above 1); channel 1: low value (below -0.05)\} and \{channel 2: high value (above 0.15); channel 2: extremely low value (below -0.1)\}. The first itemset can be read as: the prediction is associated with the occurence on the channel 1 of an extremely high value being above 1 and a low value being below -0.05 at another moment, without information on which one appears first;
\vspace{-0.3em}
\item[$\bullet$]\textit{Multi Itemsets} (\textit{Patterns}): the explanations provide patterns under the form of multidimensional itemsets, i.e. groups of values composed of different features, which are associated with the prediction. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the following itemset is associated with the prediction: \{channel 1: extremely high value (above 1); channel 2: high value (above 0.15)\};
\vspace{-0.3em}
\item[$\bullet$] \textit{Uni Sequences} (\textit{Patterns}): the explanations provide patterns under the form of ordered groups of values, also called sequences, which occur per feature and are associated with the prediction. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the following sequences are associated with the prediction: $<$channel 1: extremely high value (above 1); channel 1: low value (below -0.05)$>$ and $<$channel 2: high values (above 0.15) with an increase during 1 second$>$. The first sequence can be read as: the prediction is associated with the occurrence on the channel 1 of an extremely high value being above 1 followed by a low value being below -0.05;
\vspace{-0.3em}
\item[$\bullet$] \textit{Multi Sequences} (\textit{Patterns}): the explanations provide patterns under the form of multidimensional sequences, i.e. ordered groups of values composed of different features, which are associated with the prediction. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the following sequence is associated with the prediction: $<$channel 1: extremely high value (above 1); channel 2: high values (above 0.15) with an increase during 1 second$>$;
\vspace{-0.3em}
\item[$\bullet$] \textit{Causal}: the last category corresponds to explanations under the form of causal rules. For example, in order to support a model output from the MTS of the Figure~\ref{fig:framework_MTS_example}, the explanations could tell the user that the following rule applies: if (channel 1: extremely high value (above 1)) \& (channel 2: high values (above 0.15) with an increase during 1 second), then the MTS belongs to the class \textit{Non-Terminating Atrial Fibrilation}.
\end{enumerate}
\paragraph{Faithfulness} \textit{Can we trust the explanations?}
The faithfulness corresponds to the level of trust an end-user can have in the explanations of model predictions, i.e. the level of relatedness of the explanations to what the model actually computes. An explanation extracted directly from the original model is faithful by definition. Some post-hoc explanation methods propose to approximate the behavior of the original ``black-box'' model with an explainable surrogate model. The explanations from the surrogate models cannot be perfectly faithful with respect to the original model~\cite{Rudin19}. The fidelity criteria is used to quantify the faithfulness by the extent to which the surrogate model imitates the prediction score of the original model~\cite{Guidotti2018}.
In this paper, two MTS classifiers use an explainable surrogate model among the three state-of-the-art methods presented in section~\ref{sec:results}. However, there is no need to distinguish between the degree of fidelity of the surrogate models for the purpose of the comparison in this paper. Therefore, due to space limitation, we propose a first assessment of the faithfulness in two categories and we plan to further elaborate this component in future work:
\begin{enumerate}
\item[$\bullet$] \textit{Imperfect}: imperfect faithfulness (use of an explainable surrogate model);
\vspace{-0.3em}
\item[$\bullet$] \textit{Perfect}: perfect faithfulness.
\end{enumerate}
\paragraph{User category} \textit{What is the target user category of the explanations?}
The user category indicates the audience to whom the explanations are accessible.
The user's experience will affect what kind of \textit{cognitive chunks} they have, that is, how they organize individual elements of information into collections~\cite{Neat03}. Thus, it could be interesting to categorize the user types and associate with the model to whom the explanations will be accessible to. The broader the audience, the better are the explanations. Therefore, we propose an assessment in three categories:
\begin{enumerate}
\item[$\bullet$] \textit{Machine Learning Expert};
\vspace{-0.3em}
\item[$\bullet$] \textit{Domain Expert}: domain experts (e.g. professionals, researchers);
\vspace{-0.3em}
\item[$\bullet$] \textit{Broad Audience}: non-domain experts (e.g. policy makers).
\vspace{-0.3em}
\end{enumerate}
\begin{figure*}[!htpb]
\centering
\includegraphics[width=0.65\linewidth]{./Images/Parallel_Coordinates_MTS}
\vspace{-0.5em}
\caption{Parallel coordinates plot of the state-of-the-art MTS classifiers. Performance evaluation method: predefined train/test splits and an arithmetic mean of the accuracies on 35 public datasets [Karim et al., 2019]. As presented in section~\ref{rw_MTS}, the models evaluated in the benchmark are: DTW$_{D}$, DTW$_{I}$, gRSF, LPS, MLSTM-FCN, mv-ARF, SMTS, UFS and WEASEL+MUSE.}
\label{fig:coordinates_example}
\vspace{-1.3em}
\end{figure*}
In order to compare the methods visually using the proposed framework, the different aspects can be represented on a parallel coordinates plot. A parallel coordinate plot allows a 2-dimensional visualization of a high dimensional dataset and is suited for the categorical data of this framework. The next section presents an example of parallel coordinates plots comparing the state-of-the-art MTS classifiers.
\vspace{-0.7em}
\section{Application to Multivariate Time Series Classifiers}
\label{sec:results}
This section shows how the framework presented in the previous section can be used to assess and benchmark the state-of-the-art MTS classifiers. As introduced in section~\ref{rw_MTS}, the state-of-the-art MTS classifiers are: DTW$_{I}$, MLSTM-FCN and WEASEL+MUSE. The results of the assessment are summarized in Table~\ref{tab:MTS}, illustrated in Figure~\ref{fig:coordinates_example} and detailed in the following paragraphs.
The first MTS classifier belongs to the similarity-based category and is the one-nearest neighbor MTS classifier with DTW distance (DTW$_{I}$). DTW$_{I}$ classifies MTS based on the label of the nearest sample and a similarity calculated as the cumulative distances of all dimensions independently measured under DTW. For each MTS, the explanation supporting the classification is the ranking of features and timestamps in decreasing order of their DTW distance with the nearest MTS.
Based on predefined train/test splits and an arithmetic mean of the accuracies, DTW$_{I}$ underperforms the current state-of-the-art MTS classifiers on the 35 public datasets (Performance: \textit{Below}). The results from~\cite{Karim19} shows that DTW$_{I}$ has a statistically significant lower performance than MLSTM-FCN and WEASEL+MUSE. Furthermore, DTW$_{I}$ supports its predictions with limited information (Information: \textit{Features+Time}) that needs to be analyzed by a domain expert to ensure that it is relevant for the application (User: \textit{Domain Expert}). However, DTW$_{I}$ is an easy-to-understand model (Comprehensibility: \textit{White-Box}) which provides faithful explanations (Faithfulness: \textit{Perfect}) for each MTS (Granularity: \textit{Local}).
Then, we can analyze MLSTM-FCN and WEASEL+ MUSE together. First, based on predefined train/test splits and an arithmetic mean of the accuracies, MLSTM-FCN exhibits the best performance on the 35 public datasets (Performance: \textit{Best})~\cite{Karim19}, followed by WEASEL+MUSE (Performance: \textit{Similar}). Second, both MLSTM-FCN and WEASEL+MUSE are ``black-box'' classifiers without being explainable-by-design or having a post-hoc model-specific explainability method. Thus, the explainability characteristics of these models depend on the choice of the post-hoc model-agnostic explainability method. We have selected SHapley Additive exPlanations (SHAP)~\cite{Lundberg17}, a state-of-the-art post-hoc model-agnostic explainability method offering explanations at all granularity levels. SHAP method measures how much each variable (Features+Time) impacts predictions and comes up with a ranking of the variables which could be exploited by domain experts. The combination of MLSTM-FCN and WEASEL+MUSE with SHAP enables them to outperform DTW$_{I}$ while reaching explanations with a similar level of information (Information: \textit{Features+Time}, DTW$_{I}$: \textit{Features+Time}), in the meantime remaining accessible to the same user category (User: \textit{Domain Expert}, DTW$_{I}$: \textit{Domain Expert}). However, as opposed to DTW$_{I}$, SHAP relies on a surrogate model which cannot provide perfectly faithful explanations (Faithfulness: \textit{Imperfect}, DTW$_{I}$: \textit{Perfect}).
Therefore, based on the performance-explainability framework introduced, if a ``white-box'' model and perfect faithfulness are not required, it would be preferable to choose MLSTM-FCN with SHAP instead of the other state-of-the-art MTS classifiers on average on the 35 public datasets. In addition to its better level of performance, MLSTM-FCN with SHAP provides the same level of information and at all granularity levels.
However, the imperfect faithfulness of the explanations could prevent the use of MLSTM-FCN with a surrogate explainable model on numerous applications. In addition, the level of information provided to support the predictions remains limited (Information: \textit{Features+Time}). Therefore, based on the assessment of the current state-of-the-art MTS classifiers with the framework proposed, it would be valuable for instance to design some new high-performing MTS classifiers which provide faithful and more informative explanations. For example, it could be interesting to work in the direction proposed in~\cite{Fauvel20LCE}. It presents a new MTS classifier (XEM) which reconciles performance (Performance: \textit{Best}) and faithfulness while providing the time window used to classify the whole MTS (Information: \textit{Uni Sequences}). XEM is based on a new hybrid ensemble method that combines an explicit approach to handle the bias-variance trade-off and an implicit approach to individualize classifier errors on different parts of the training data~\cite{Fauvel19}.
Nevertheless, the explanations provided by XEM are only available per MTS (Granularity: \textit{Local}) and the level of information could be further improved. As suggested by the authors, it could be interesting to analyze the time windows identified for each class to determine if they contain some common multidimensional sequences (Information: \textit{Multi Sequences}, Granularity: \textit{Both Global \& Local}). These patterns could also broaden the audience as they would summarize the key information in the discriminative time windows.
\vspace{-0.7em}
\section{Conclusion}
We have presented a new performance-explainability analytical framework to assess and benchmark the machine learning methods. The framework details a set of characteristics that systematize the performance-explainability assessment of machine learning methods. In addition, it can be employed to identify ways to improve current machine learning methods and to design new ones. Finally, we have illustrated the use of the framework by benchmarking the current state-of-the-art MTS classifiers.
With regards to future work, we plan to further elaborate the definition of the different components of the framework (especially the \textit{Model Comprehensibility}, the \textit{Information Type} and the \textit{Faithfulness}) and evaluate the relevance of integrating new components. Then, we plan to apply the framework extensively to assess the different types of existing machine learning methods.
\section*{Acknowledgments}
This work was supported by the French National Research Agency under the Investments for the Future Program (ANR-16-CONV-0004) and the Inria Project Lab ``Hybrid Approaches for Interpretable AI'' (HyAIAI).
\bibliographystyle{named}
| c3fdedb009a59c2d512e54de778f277a8506589b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the past few decades, graphene has attracted much attention due to its peculiar dispersion relation at low energies, similar to the spectrum of relativistic particles described by the Dirac theory\cite{wallace,neto}. More precisely, the spectrum has two cones, the so-called “Dirac cones” in the vicinity of two non-equivalent points $K_1$ and $K_2$ in the reciprocal space. Anisotropy in graphene was another interesting aspect which was discussed long ago by Pauling\cite{pauling}, and could be induced by uniaxial stress or bending of a graphene sheet. The main motive is to tune the hopping energy between neighbouring carbon atoms with precision, which was later found to be feasible in optical lattices \cite{tarruell} via controlling the lattice potential in order to have a handle on the effective mass in a honeycomb lattice. In a tight-binding model for graphene, if one of the three nearest-neighbor hopping energies is tuned, the two Dirac points with opposite chiralities approach each other and merge into one forming the so-called semi-Dirac point. The band dispersion simultaneously exhibits massless Dirac (linear) and massive fermionic (quadratic) features along two different directions, thereby producing a highly anisotropic electronic dispersion\cite{dietl,pickett}. The materials that host such anisotropic dispersion are phosphorene under pressure and doping \cite{rodin,guan}, electric fields \cite{rudenko,dut}, TiO2/VO2 superlattices \cite{pickett, pardo1, pardo2}, graphene under deformation \cite{mon}, and BEDT-TTF$_{2}$I$_{3}$ salt under pressure \cite{konno,pichon}. Experimentally, semi-Dirac dispersion has been observed in a few-layer black phosphorene by means of the in situ deposition of potassium atoms \cite{kim}. A straightforward approach to realize semi-Dirac materials can be achieved by breaking the hexagonal symmetry of the honeycomb lattice, e.g, by strain. However, directly applying strain to realize the transition in materials, such as graphene or silicene is prohibited by the exorbitant magnitude of the strain required, which would eventually disintegrate them \cite{si,zhao}. Some successfully synthesized graphene-like honeycomb materials, such as silicene \cite{vogt,meng}, germanene \cite{li} and stanene \cite{zhu}, are found to be easily oxidized or they chemically absorb other atoms because of their buckling geometries \cite{molle, spencer, houselt, ciraci}. It is these absorbed atoms that will modify the hopping energies in the honeycomb lattice which is essentially applying a strain that creates a differential hopping.\par
\begin{figure}[h]
\begin{center}
\subfloat{\includegraphics[width=0.45\textwidth]{Fig_1.pdf}\label{fig:1}}
\caption{(Color online) Schematic diagram of hexagonal lattice geometry of a semi-Dirac system with different hopping parameters $t$ and $t_2$ is shown. The black corresponds to hopping, $t$ whereas the green corresponds to $t_2$. Two sublattices are denoted by two different colors (blue and magenta). $\vec{\delta_1}$, $\vec{\delta_2}$ and $\vec{\delta_3}$ are the nearest neighbor real space vectors.}
\label{fig:1}
\end{center}
\end{figure}
On the other hand, behaviors of electrons in graphene, exposed to a strong perpendicular magnetic field played an important role not only for the discovery of quantum Hall effects\cite{vp,zheng}, but also for proving the existence of massless Dirac particles\cite{firsov,zhang}. The unconventional Hall conductivity was found to be quantized as $\sigma_{xy}=2(2n+1) e^2/h$\cite{firsov,zhang}, where both the spin and the valley degeneracies are taken into account. Experimental measurements confirm that the Landau levels of a monolayer graphene obey the relation, $E_n=sgn(n)\sqrt{2 \hbar v_{F}e|n|B}$, where $v_{F}$=$10^6$ m/s is the Fermi velocity, $B$ is the magnetic field and $n$ denote Landau level indices\cite{sadowski1,sadowski2}.
Recently, quite a few studies on Landau levels and transport properties in presence of a magnetic field in phosphorene have been reported \cite{chang,roldan}. More precisely, they have found that the anisotropic band structure that leads to Hall quantization in presence of a perpendicular magnetic field is similar to that of a conventional two-dimensional electron gas (2DEG). Since phosphorene may be considered as a realistic material that possesses semi-Dirac properties, it is necessary to pursue quantum Hall studies on the semi-Dirac systems. As discussed above, the energy dispersion of phosphorene is similar to that of the semi-Dirac systems, it is likely that other properties too show similar characteristics.\par
In this work, we have explored the influence of magnetic field for a semi-Dirac system using a tight-binding Hamiltonian on a honeycomb lattice. We study the Landau level spectrum and Hofstadter butterfly using a nanoribbon in order to show that the semi-Dirac system has quite distinct properties as compared to Dirac fermions. We also calculate the density of states (DOS) via the tight-binding propagation method\cite{yuan,yuan2}, which is a sophisticated numerical tool used in large-scale calculations for any realistic system. We have implemented the recently developed real-space order-$N$ quantum transport approach to calculate the Kubo conductivities as a function of the Fermi energy for moderate as well as very high values of the magnetic field\cite{rappoprt}. The Hall conductivity in a semi-Dirac system shows the {\it {standard}} quantization, namely, $\sigma_{xy}\propto 2n$ as compared to the previously observed {\it {anomalous}} quantization, that is, $\sigma_{xy}\propto 4(n+1/2)$ for a Dirac system. The longitudinal conductivities show highly anisotropic behavior in one direction compared to the other, which is obviously absent for Dirac systems.\par
The paper is organized as follows. The low energy tight-binding Hamiltonian is described in sec.~\ref{II}. We have further studied the Landau level spectra and the Hofstadter butterfly for a nanoribbon in presence of a magnetic field in sec.~\ref{III}. The transport properties are investigated by computing the Hall and the longitudinal conductivities in sec.~\ref{IV}. We conclude with a brief summary in sec.~\ref{V}.
\section{Model Hamiltonian}\label{II}
We study the tight-binding model on the honeycomb lattice with anisotropic hopping that leads to semi-Dirac electronic spectra at low energy.
More precisely, the hopping energy to one of the neighbours ($t_2$) is different than the other two ($t$) as shown in Fig.~\ref{fig:1}.
It is also instructive to look at the full dispersion with the following three nearest neighbor vectors in real space,
$\vec{\delta_1} = \big(0, a\big)$;
$\vec{\delta_2} = \Big(\frac{\sqrt{3}a}{2}, -\frac{a}{2}\Big)$ and
$\vec{\delta_3} = \Big(-\frac{\sqrt{3}a}{2}, -\frac{a}{2}\Big)$, where $a$ is the lattice constant.
\begin{widetext}
The dispersion relation for a semi-Dirac system can be written as,
\begin{align}
E(k)=
\pm \sqrt{2t^2 + t_2^2 +2t^2\cos \sqrt{3}k_xa + 4tt_2 \cos(3k_ya/2) \cos(\sqrt3k_xa/2)}.
\label{eq.3}
\end{align}
\end{widetext}
The above expression in Eq.~(\ref{eq.3}) is plotted in Fig.~\ref{fig:2a}. The Brillouin zone with high-symmetry points for $t_2=t$ is shown in Fig.~\ref{fig:2b}. For the Dirac case (that is, $t_2=t$), the dispersion shows that the Dirac points touch at the $K_1$ and $K_2$ points at the Brillouin zone corners as shown in Fig.~\ref{fig:2c}. With increasing the strength of the parameter $t_{2}$, the two Dirac points originally located at $K_1$ ($\frac{2\pi}{3a}$,$\frac{2\pi}{\sqrt{3}a}$) and $K_2$ ($-\frac{2\pi}{3a}$,$\frac{2\pi}{\sqrt{3}a}$) move closer till they merge at the $M$ point resulting in a semi-Dirac spectrum (see Fig.~\ref{fig:2d}). As mentioned earlier, such manipulation of the Dirac points and their eventual merger have been achieved in honeycomb optical lattices\cite{tarruell}.
Thus by fixing $t_{2} = 2t$ and focusing on
the $M$ point (0,$\frac{2\pi}{\sqrt{3}a}$), the low-energy effective Hamiltonian based on the tight-binding model for a semi-Dirac system, apart from a constant term, can be written as\cite{kush,chen,firoz},
\begin{figure*}[!ht!]
\begin{center}
\subfloat[]{\includegraphics[width=0.37\textwidth]{Fig_2a.pdf}\label{fig:2a}} \hspace*{2.5 cm}
\subfloat[]{\includegraphics[width=0.38\textwidth]{Fig_2b.pdf}\label{fig:2b}}\\
\subfloat[]{\includegraphics[width=0.35\textwidth]{Fig_2c.pdf}\label{fig:2c}} \hspace*{2.5 cm}
\subfloat[]{\includegraphics[width=0.35\textwidth]{Fig_2d.pdf}\label{fig:2d}}
\caption{(Color online) (a) Anisotropic energy band dispersion of a semi-Dirac system is shown. The dispersion is linear along the $y$-direction and quadratic along the $x$-direction. Here $a$ is set to be unity. (b) Brillouin zone with different high-symmetry points is shown for Dirac. (c) and (d) The dispersion along the high symmetry points $\Gamma$ $\rightarrow$ $K_1$ $\rightarrow$ $M$ $\rightarrow$ $K_2$ $\rightarrow$ $\Gamma$ for different strength of hopping parameters $t_2=t$ (Dirac) and $t_2=2t$ (semi-Dirac) respectively. Here we put $t=2.8$eV.}
\label{fig:2}
\end{center}
\end{figure*}
\begin{equation}
{\it{H_0}}=\frac{p^2_x \sigma_x}{2m^*} + v_F p_y \sigma_y
\label{eq.1}
\end{equation}
where $p_x$ and $p_y$ are the momenta along the $x$ and the $y$ directions respectively. $\sigma_x$ and $\sigma_y$ are the Pauli spin matrices in the pseudospin space. The velocity along the $p_y$ direction, $v_F$, and the effective mass, $m^*$ corresponding to the parabolic dispersion along $p_x$ are expressed as $v_F= {3ta}/{\hbar}$ and $m^*=2\hbar/3ta^2$. Henceforth we set $a=1$. The dispersion relation corresponding to Eq.~(\ref{eq.1}) ignoring a constant shift in energy can be written as,
\begin{equation}
E=\pm \sqrt{(\hbar v_{F}k_{y})^2+\bigg(\frac{\hbar^2k_{x}^2}{2m^*}\bigg)^2}
\label{eq.2}
\end{equation}
where `+' denotes for the conduction band and `-' stands for the valence band. Equation (\ref{eq.2}) shows that the dispersion is linear (Dirac-like) along $y$-direction, whereas the dispersion along the $x$-direction is quadratic (non-relativistic), the combination of which results in the semi-Dirac dispersion. The three-dimensional plot in Fig.~\ref{fig:2a} indicates the anisotropic band structure in a semi-Dirac system.\par
\section{The Landau Levels}\label{III}
To include a magnetic field, we shall work with a semi-Dirac nanoribbon which is infinitely long along $x$, but has a finite width along $y$. We apply a uniform magnetic field, $\mathbf{B}=B\hat{z}$ perpendicular to the plane of the ribbon. Owing to the presence of the vector potential $\vec{A}$, each tight-binding wave-function picks up an extra phase term. We have chosen the Landau gauge as $\vec{A} = (-By, 0, 0)$ such that the translational invariance along the $x$-direction remains unaltered under the choice of the gauge. Hence, the momentum along the $x$-direction is conserved and acts as a good quantum number. To make $k_x$ a dimensionless quantity, we have absorbed the lattice spacing $a$ into the definition of $k_x$. The ribbon width is such that it has $N$ unit cells along the $y$-axis (where the index $n$ for the unit cells takes $\in 0$.....$N-1$) as shown in Fig.~\ref{fig:3}.
The tight-binding Hamiltonian in the presence of magnetic field has the form,
\begin{align}
H= -\sum_{\langle{ij}\rangle} (t_{ij}a^{\dagger}_{i} b_{j} +h.c.)
\label{ham}
\end{align}
where $a^{\dagger}_{i}$ ($b_{j}$) creates (annihilates) an electron on sublattice $A$ ($B$). $t_{ij}$ is the hopping amplitude between nearest neighbor sites, which obtain a phase due to the magnetic field by the Peierls substitution, namely, $ t_{ij} = t \rightarrow te^{2i\pi\phi_{ij}}$ (here $t$ denotes both $t$ or $t_2$). $\phi_{ij}$ is the magnetic flux and is given by the line integral of the vector potential from a site $i$ to a site $j$, namely, $\phi_{ij}=e/h\int_{i}^{j} \vec{A}.d\vec{l}$. The flux is usually denoted in terms of the flux quantum $\phi_0=h/e$ ($h$ is Planck$'$s constant and $e$ is the magnitude of the electron charge).
\begin{figure}[h]
\begin{center}
\subfloat{\includegraphics[width=0.45\textwidth]{Fig_3.pdf}\label{fig:3}}
\caption{(Color online) Zigzag nanoribbon of a honeycomb lattice is shown. The magenta and blue circles represent the $A$ and $B$ sublattices respectively. ${\vec {a}}_{1}$ and ${\vec{a}}_2$ are the primitive vectors. $(m,n)$ labels the positions of the unit cells along $x$ and $y$ directions. The ribbon is infinite along the $x$-direction shown by the arrow on both side.}
\label{fig:3}
\end{center}
\end{figure}
Thus the tight-binding Hamiltonian in presence of the perpendicular magnetic field can be written in terms of $m$ and $n$ (where $m$ increases along the $x$-direction and $n$ increases along the negative $y$-direction) (see Fig.~\ref{fig:3})\cite{castro},
\begin{align}
\mathcal{H} = &
- \sum\limits_{\langle{mn}\rangle} \Big[t e^{i\pi(\phi/\phi_0)n[(1+\alpha)/2]}a^{\dagger}(m,n)b(m,n) \nonumber
\\
&
+ t e^{-i\pi(\phi/\phi_0)n} a^{\dagger}(m,n)b(m-1,n-(1-\alpha)/2)\nonumber
\\
&
+ t_2~e^{i\pi(\phi/\phi_0)n[(\alpha-1)/2]} a^{\dagger}(m,n)b(m,n-\alpha)+h.c.\Big]
\label{h1}
\end{align}
where the summation $\langle{mn}\rangle$ is over the nearest neighbors. $a^{\dagger}(m,n)$ and $b(m,n)$ denote the creation and annihilation operators at the ($m$,$n$) site, respectively.
Equation ({\ref{h1}}) for a zigzag semi-Dirac ribbon ($\alpha=1$) reduces to
\begin{align}
\mathcal{H} = &
- \sum\limits_{\langle{mn}\rangle} \Big[t e^{i\pi(\phi/\phi_0)n}a^{\dagger}(m,n)
b(m,n)+ t e^{-i\pi(\phi/\phi_0)n} \nonumber
\\
&
a^{\dagger}(m,n)b(m-1,n)
+ t_2~a^{\dagger}(m,n)b(m,n-1)+h.c.\Big]
\label{h2}
\end{align}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[width=0.24\textwidth]{Fig_4a.jpg}\label{fig:4a}} \hspace*{0.01 cm}
\subfloat[]{\includegraphics[width=0.24\textwidth]{Fig_4b.jpg}\label{fig:4b}}
\caption{(Color online) Hofstadter butterfly spectrum is plotted for as a function of $\phi/\phi_0$ for (a) $t_2=2t$ (semi-Dirac) and (b) $t_2=t$ (Dirac).}
\label{fig:4}
\end{center}
\end{figure}
Using the above Hamiltonian as mentioned in Eq.~(\ref{h2}) we have numerically calculated the Hofstadter butterfly\cite{rammal} as well as the Landau level spectrum for the number of unit cells $N=100$. Figure~\ref{fig:4a} shows fractal spectra plotted as a function of magnetic flux, $\phi/\phi_{0}$ for a semi-Dirac nanoribbon. It can be seen clearly that there occurs opening of a central gap with a flat band at zero energy. The gap gets larger along with the two identical spectra that emerge from the conduction and the valence bands by tuning $t_2$. For comparison, the same is plotted for the Dirac system ($t_2=t$) as shown in Fig.~\ref{fig:4b}. We can see that there is no gap at zero energy with the flat band when one goes from $t_2=2t$ to $t_2=t$.
\begin{figure*}[!ht]
\centering
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5a.pdf}\label{fig:5a}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5b.pdf}\label{fig:5b}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5c.pdf}\label{fig:5c}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5d.pdf}\label{fig:5d}}\\
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5e.pdf}\label{fig:5e}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5f.pdf}\label{fig:5f}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5g.pdf}\label{fig:5g}} \hspace{0.02 cm}
\subfloat[]{\includegraphics[width=0.243\textwidth]{Fig_5h.pdf}\label{fig:5h}}
\caption{(Color online) Evolution of Landau levels for a finite strip with $N=100$ unit cells in the presence of a magnetic flux $\phi=\frac{\phi_0}{100}$, $\frac{\phi_0}{200}$,$\frac{\phi_0}{500}$ and $\frac{\phi_0}{1600}$ for (a-d) $t_2=2t$ (semi-Dirac) and (e-h) $t_2=t$ (Dirac).}
\label{fig:5}
\end{figure*}
Figure~\ref{fig:5} shows the Landau level spectra for different values of the magnetic flux (such as $\phi/\phi_{0} = 1/100, 1/200, 1/500, 1/1600$) for the semi-Dirac and the Dirac systems. Figures~\ref{fig:5a}-\ref{fig:5d} show the evolution of the energy levels in presence of the magnetic field for a semi-Dirac nanoribbon ($t_2=2t$). For comparison, we have also plotted the Landau level spectrum for the Dirac case ($t_2=t$) using the same values of the magnetic flux as shown in Fig.~\ref{fig:5e}-\ref{fig:5h}. It is to be noted that in a semi-Dirac system, there is no zero energy bulk state, which implies that the zero energy state in Fig.~\ref{fig:5} is an edge state. On the other hand, zero energy bulk states exist in a Dirac system.
Further, for $t_2=2t$, the Landau levels are not equidistant, since their energies vary as $(n+\frac{1}{2})^{2/3}$ ($n$ being the Landau level index)\cite{pickett} which lies intermediate to the behavior of the Dirac system and the conventional 2DEG. As a consequence, the gap between the Landau levels shrinks as one considers larger $n$. In the case of a Dirac system, since the energies of the Landau levels go as $\sqrt{n}$, its non-equidistant Landau spectra can have a different quantitative behavior from a semi-Dirac system. For a large value of the flux, $\phi$ such as $\phi=\phi_0/100$, the flatness of the energy bands are observed for both the semi-Dirac and the Dirac systems owing to shrinking of the cyclotron radius (see Fig.~\ref{fig:5a} and \ref{fig:5e}). The energy bands become parabolic for $\phi=\phi_0/200$ as seen from Fig.~\ref{fig:5b}. The flatness of the Landau spectrum in the bulk completely vanishes in the semi-Dirac system as compared to a Dirac one (see Fig.~\ref{fig:5b} and \ref{fig:5f}). With lower values of $\phi/\phi_0$, the Landau levels demonstrate a dispersive behavior and start getting flatter for large values of $n$ for $t_2=2t$ (see Fig.~\ref{fig:5c}). In the case of a Dirac system ($t_2=t$), the Landau levels show quite a distinct behavior, where the flat bands become dispersive in the bulk corresponding to larger values of $n$ and lower values of $\phi/\phi_0$ (see Fig.~\ref{fig:5g}). For a small value of the magnetic field, such that the flux is given by, $\phi=\phi_0/1600$, the energy bands eventually become flat in the bulk for the semi-Dirac case as shown in Fig.~\ref{fig:5d}. This is not the case for a Dirac system (see Fig.~\ref{fig:5h}). For all values of $\phi/\phi_0$, the zero-energy mode is completely separated from the bulk bands for a semi-Dirac system as compared to the Dirac case (see any of Fig.~\ref{fig:5}).
\section{Transport Properties}\label{IV}
To study the transport properties in the presence of a perpendicular magnetic field, we consider a large sample of a lattice model of the semi-Dirac system consisting of millions of atoms. The contribution in transport comes from both the off-diagonal and the diagonal terms as appear in the Kubo formula\cite{kubo1}. The former contributes to the Hall conductivity ($\sigma_{xy}$), whereas the latter leads to individual longitudinal conductivity in different directions ($\sigma_{xx}$ and $\sigma_{yy}$).
\subsection{Methodology}
In this section, we shall describe the numerical approach, developed by Garcia and his co-workers \cite {rappoprt} which is based on a real-space implementation of the Kubo formalism, where both the diagonal and the off-diagonal conductivities are treated on the same footing. It is known that in the momentum space, the Hall conductivity can be easily obtained in terms of the Berry curvature associated with the bands \cite{kohmoto}. The Kubo formalism can be implemented in real space for obtaining the Hall conductivity\cite{rappoprt} which uses Chebyshev expansions to compute the conductivities. The components of the dc conductivity tensor ($\omega \rightarrow 0$ limit of the ac conductivity) for the non-interacting electrons are given by the Kubo-Bastin formula\cite{kubo1, kubo2} which can be written as\cite{bastin,rappoprt,ortmann},
\begin{align}
\sigma_{\alpha\beta} (\mu, T) = & \frac{i e^2 \hbar}{A}\int_{-\infty}^{\infty} d\varepsilon f(\varepsilon) Tr \Big<v_{\alpha} \delta(\varepsilon-H)v_\beta \frac{dG^{+}(\varepsilon)}{d\varepsilon} \nonumber
\\
&
-v_{\alpha} \frac{dG^{-}(\varepsilon)}{d\varepsilon}v_\beta\delta(\varepsilon-H)\Big>
\label{eq.7}
\end{align}
where $T$ is the temperature, $\mu$ is the chemical potential, $v_\alpha$ is the $\alpha$ component of the velocity operator, $A$ is the area of the sample, $f(\varepsilon)$ is the Fermi-Dirac distribution and $G^{\pm}(\varepsilon, H)= \frac{1}{\varepsilon-H\pm i\eta}$ are the advanced (+) and retarded (-) Green's functions. Using the Kernel Polynomial Method (KPM)\cite{weisse}, the rescaled delta and Green's function can be expanded in terms of the Chebyshev polynomials, whence Eq.~(\ref{eq.7}) becomes,
\begin{equation}
\sigma_{\alpha\beta} (\mu, T) = \frac{4 e^2 \hbar}{\pi A} \frac{4}{(\Delta E)^2} \int_{-1}^{1} d{\tilde{\varepsilon}} \frac{f(\tilde{\varepsilon})}{(1-{\tilde{\varepsilon}^2})^2} \sum_{m,n} \Gamma_{nm} (\tilde{\varepsilon}) \mu_{nm}^{\alpha\beta} (\tilde{H})
\label{eq.8}
\end{equation}
where $\Delta E$ is the range of the energy spectrum, $\tilde{\varepsilon}$ is the rescaled energy whose upper and lower bounds are $+1$ and $-1$ respectively and $\tilde{H}$ is the rescaled Hamiltonian. $\Gamma_{nm} (\tilde{\varepsilon})$ and $\mu_{nm}^{\alpha\beta} (\tilde{H})$ are the functions of rescaled energy and Hamiltonian respectively. The energy dependent scalar function, $\Gamma_{nm} (\tilde{\varepsilon})$ can be written as,
\begin{align}
\Gamma_{nm} (\tilde{\varepsilon})\equiv &(\tilde{\varepsilon}-in\sqrt{1-\tilde{\varepsilon}^2})e^{in~arccos(\tilde{\varepsilon})} T_m(\tilde{\varepsilon}) \nonumber
\\
&
+ (\tilde{\varepsilon}+im\sqrt{1-\tilde{\varepsilon}^2})e^{-im~arccos(\tilde{\varepsilon})} T_n(\tilde{\varepsilon})
\end{align}
and the Hamiltonian-dependent term which involves products of polynomial expansions can be written as,
\begin{equation}
\mu_{nm}^{\alpha\beta}(\tilde{H})= \frac{g_m g_n}{(1+\delta_{n0})(1+\delta_{m0})} Tr[v_{\alpha}T_m(\tilde{H})v_{\beta}T_n(\tilde{H})]
\end{equation}
where the Chebyshev polynomials, $T_m(x)$ obey the recurrence relation,
\begin{equation}
T_m(x)= 2x T_{m-1}(x)-T_{m-2}(x)
\end{equation}
The Jackson kernel, $g_m$ is used to smoothen out the Gibbs oscillations which arise due to the truncation of the expansion in Eq.~(\ref{eq.8}) \cite{weisse,silvera}.\par
\begin{figure}[ht]
\begin{center}
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_6a.pdf}\label{fig:6a}} \hspace*{0.01 cm}
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_6b.pdf}\label{fig:6b}}
\caption{(Color online) Density of states (in units of 1/eV) is plotted as a function of energy, $E$ (in units of eV) for (a) $t_2=2t$ (semi-Dirac) and (b) $t_2=t$ (Dirac). We put $t=2.8$eV in the calculation.}
\label{fig:6}
\end{center}
\end{figure}
\begin{figure*}[!ht]
\centering
\subfloat[]{\includegraphics[width=0.35\textwidth]{Fig_7a.pdf}\label{fig:7a}} \hspace*{ 2.5 cm}
\subfloat[]{\includegraphics[width=0.35\textwidth]{Fig_7b.pdf}\label{fig:7b}}\\
\caption{(Color online) Hall conductivity, $\sigma_{xy}$ (in units of $2e^2/h$) is plotted as a function of Fermi energy, $E$ (in units of eV) for (a) $t_2$ = $2t$ (semi-Dirac) and (b) $t_2$ = $t$ (Dirac) for a very high field (400T) and a moderate field (30T and 50T). Here $t$ is taken as 2.8eV.}
\label{fig:7}
\end{figure*}
The density of states (DOS) can be calculated using an efficient algorithm based on the evolution of the time-dependent Schr$\ddot{o}$dinger equation. We use a random superposition of all basis states as an initial state $|\phi(0)\rangle$,
\begin{equation}
|\phi(0)\rangle=\sum_{i}a_i|i\rangle
\end{equation}
where $|i\rangle$ denote the basis states and $a_i$ are the normalized random complex numbers. Applying the Fourier transformation to the correlation function, $\langle\phi(0)|e^{-iHt}|\phi(0)\rangle$ we get the DOS as \cite{yuan},
\begin{equation}
DOS=\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{i\epsilon t}\langle\phi(0)|e^{-iHt}|\phi(0)\rangle dt
\end{equation}
where $t$ denotes time.
\subsection{Longitudinal and Hall conductivities}
Using the above mentioned efficient numerical approach, we calculate the DOS in the absence and the presence of a magnetic field, the longitudinal conductivity in both $x$ ($\sigma_{xx}$) and $y$ ($\sigma_{yy}$) directions and the Hall conductivity ($\sigma_{xy}$) for the semi-Dirac system ($t_2=2t$). To compare between the Dirac and the semi-Dirac systems, we have also shown results for the Dirac ($t_2=t$) case simultaneously. In our simulation, we consider a lattice of $5120$ unit cells in each of the $x$ and $y$ directions (that is, a sample size denoted by ($L_{x},L_{y}$) to be ($5120, 5120$)). We apply periodic boundary conditions for all our numerical results. We set the nearest neighbor hopping parameter $t=2.8$eV. We adopt a large number of Chebyshev moments, $M$, since the energy resolution of the KPM and the convergence of the peaks of $\sigma_{xx}$ depend on $M$. We have used $M=6144$ here\cite{rappoprt}. The system size and the truncation order can be enhanced to reduce the fluctuations.\par
We have plotted the results for the DOS in the absence of a magnetic field for semi-Dirac and Dirac systems as shown in Fig.~\ref{fig:6}. In the case of semi-Dirac systems, the DOS is proportional to $\sqrt{|E|}$ near zero energy (see Fig.~\ref{fig:6b}), whereas for the Dirac case the DOS near the zero energy varies as $|E|$ as shown in Fig.~\ref{fig:6a}. The energy bounds for the semi-Dirac are $E_{min}=-4t$ and $E_{max}=4t$.\par
Next, we have shown the results for Hall conductivity ($\sigma_{xy}$) for moderate values of the magnetic field as well as extremely high fields in Fig.~\ref{fig:7}. Generally higher values of the magnetic field require smaller system size and hence the fewer number of Chebyshev moments have to be computed. This yields faster convergence of the Hall conductivity in the limit of a large magnetic field. For $t_2=2t$ (semi-Dirac), the Hall conductivity ($\sigma_{xy}$) is plotted as a function of Fermi energy, $E$ for the large value of the field $B=400$T (as shown in the main frame of Fig.~\ref{fig:7a}). To relate this result to the recent experiments \cite{young,dean,hunt} performed for realistic values of magnetic field on a Dirac system, we have also plotted the Hall conductivity for moderate values of the field, namely 30T (green curve) and 50T (pink curve) as shown in the inset of Fig.~\ref{fig:7a}. The quantization of the plateaus is similar to that of a conventional 2DEG with a parabolic band dispersion in a sense that the conductance quantization happens at $\sigma_{xy}=2ne^2/h$ where $n$ takes integer values $0$, $\pm1$, $\pm2$, $\pm3$, $\pm4$... in units of $2e^2/h$. The plot in the inset shows that the plateau step can be obtained with good accuracy even in the case of realistic values of the magnetic field. The difference between the semi-Dirac and the Dirac cases lies in the fact that the fluctuations in the plateau step become prominent with lowering of the field, especially at higher values of the Fermi energy. Further lowering of the magnetic field will reduce the sharpness of the plateaus due to the effect of finite energy resolution and finite size of the sample. These are the artifacts of the method used here. In the Dirac system ($t_2=t$), the well-known Hall quantization at $\sigma_{xy}= 2(2n+1)e^2/h$ is observed for very high magnetic field, namely $B=400$T as shown in the main frame of Fig.~\ref{fig:7b}. The inset shows the same for realistic values of the magnetic field. The Hall conductivity plot ensures that there is a transition from a half-integer to an integer quantum Hall effect as we go from a Dirac to a semi-Dirac system by tuning $t_2$ (Fig.~\ref{fig:7a} and \ref{fig:7b}). The reason can be drawn from the fact that although the band dispersion in the semi-Dirac is linear in one direction, the quadratic behavior in the other direction seemingly dominates over the linear term which results in a similar characteristic of conductance quantization of a 2DEG.\par
\begin{figure}[]
\begin{center}
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_8a.pdf}\label{fig:8a}} \hspace*{- 0.2 cm}
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_8b.pdf}\label{fig:8b}}\\
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_8c.pdf}\label{fig:8c}} \hspace*{- 0.2 cm}
\subfloat[]{\includegraphics[width=0.25\textwidth]{Fig_8d.pdf}\label{fig:8d}}
\caption{(Color online) Hall conductivity, $\sigma_{xy}$ (in units of $2e^2/h$) and the DOS (in units of 1/eV) is plotted as a function of Fermi energy, $E$ (in units of eV) for different cases (a) $t_2$ = $2t$, $B=50$T (b) $t_2$ = $t$, $B=50$T (c) $t_2$ = $2t$, $B=400$T and (d) $t_2$ = $t$, $B=400$T.}
\label{fig:8}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\subfloat[]{\includegraphics[trim=0 0 0 0,clip,width=0.28\textwidth,height=0.20\textwidth]{Fig_9a.pdf}\label{fig:9a}} \hspace*{-0.3 cm}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Fig_9b.pdf}\label{fig:9b}}\\
\subfloat[]{\includegraphics[trim=0 0 0 0,clip,width=0.28\textwidth,height=0.20\textwidth]{Fig_9c.pdf}\label{fig:9c}} \hspace*{- 0.3 cm}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Fig_9d.pdf}\label{fig:9d}}\\
\subfloat[]{\includegraphics[trim=0 0 0 0,clip,width=0.28\textwidth,height=0.20\textwidth]{Fig_9e.pdf}\label{fig:9e}} \hspace*{-0.3 cm}
\subfloat[]{\includegraphics[width=0.23\textwidth]{Fig_9f.pdf}\label{fig:9f}}
\caption{(Color online) Longitudinal conductivity, $\sigma_{xx}$ and $\sigma_{yy}$ (in units of $2e^2/h$) is plotted as a function of Fermi energy, $E$ (in units of eV) for different cases (a) $t_2$ = $2t$, $B=50$T (b) $t_2$ = $t$, $B=50$T (c) $t_2$ = $2t$, $B=400$T (d) $t_2$ = $t$, $B=400$T (e) $t_2$ = $2t$, $B=0$T and (f) $t_2$ = $t$, $B=0$T.}
\label{fig:9}
\end{center}
\end{figure}
In Fig.~\ref{fig:8}, we have shown the Hall conductivity, $\sigma_{xy}$ and the DOS for $B=50$T and $400$T in the same frame. In the presence of the magnetic field, the DOS consists of peaks of discrete energy levels (Landau levels) as shown in Fig.~\ref{fig:8}. The DOS vanishes in the plateau region and shows a sharp peak corresponding to a Landau level when the Hall conductivity goes through a transition from one plateau to another. However we get broad DOS peaks at lower values of the magnetic field ($B = 50$T) which is particularly visible for the semi-Dirac case owing to the small energy separation between the Landau levels (less than 3 meV). Sharper peaks will require computation of very large number of Chebyshev moments. Figure~\ref{fig:8a} shows that there is no Landau level peak at zero energy for $t_2=2t$ which is also characteristic of a conventional 2DEG in contrast to $t_2=t$ case, where the zero-energy peak is well-observed (see Fig.~\ref{fig:8b}). The presence of a zero-energy peak for the Dirac case is related to the chiral anomaly present there. Figures~\ref{fig:8c} and \ref{fig:8d} show that several Landau levels can be observed with the same qualitative behavior for very high magnetic field $B=400$T for both the semi-Dirac and the Dirac systems. The Hall conductance is quantized due to the quantized Landau level. It is interesting to note that although the energy does not depend linearly on the Landau level index, $n$ and magnetic field, $B$ in the case of a semi-Dirac system, as said earlier, the quantized value of $\sigma_{xy}$ of a semi-Dirac material is analogous to that of a 2DEG. It is once again pertinent to mention that the Landau level spectra in phosphorene in a perpendicular magnetic field depending on the index, $n$ (an additional factor of `2' will appear for spin degeneracy) has connections with the semi-Dirac physics \cite{chang,roldan}.\par
To investigate the magneto-transport, we further calculate the longitudinal conductivity along the $x$ ($\sigma_{xx}$) and $y$ ($\sigma_{yy}$) directions. Figures~\ref{fig:9a} and \ref{fig:9b} show the longitudinal conductivity, $\sigma_{xx}$ (green curve) and $\sigma_{yy}$ (magenta curve) as a function of the Fermi energy, $E$ for moderate values of the $B$ field $B$=50T for a semi-Dirac and a Dirac system respectively. The longitudinal conductivity reveals largely anisotropic behavior owing to the presence of anisotropy in the dispersion for $t_2=2t$. The component of $\sigma$ in the $x$ ($\sigma_{xx}$) and $y$ ($\sigma_{yy}$) directions are quantitatively different in nature. The magnitude of $\sigma_{yy}$ is larger than that of $\sigma_{xx}$. This is definitely a contrasting feature compared to the case of a Dirac system where the magnitudes of $\sigma_{xx}$ and $\sigma_{yy}$ are same as shown in Fig.~\ref{fig:9b}. Moreover, the absence of zero-energy peak also supports our discussion on the Landau level results (see Fig.~\ref{fig:8c}) for the semi-Dirac material. Figures~\ref{fig:9c} and \ref{fig:9d} show similar results for very high values of the magnetic field, namely, $B=400$T with the well-observed sharp peaks at larger values of energy. The amplitudes increase at large values of the Fermi energy owing to the increase in the scattering rate of the Landau levels as one goes to higher values of $n$ for both the semi-Dirac and the Dirac systems. Since the Landau levels are not equidistant for both of the cases, the interval between the peaks are not spaced equally. It can be seen that the longitudinal conductivity is non-vanishing only when the Fermi energy is within a Landau band where the backscattering processes are present. \par
To compare and contrast further between the two cases, we plot the longitudinal conductivities ($\sigma_{xx}$ and $\sigma_{yy}$) in the absence of any magnetic field ($B = 0$) in Figs.~\ref{fig:9e} and \ref{fig:9f} for the semi-Dirac and the Dirac systems respectively. Apart from the suppression of the conductivities by one order of magnitude by the magnetic field, one can take a note of the linear dependence of the conductivity on the Fermi energy, $E$ for the Dirac case\cite{novak}, while they appear with different exponents for the semi-Dirac one. The feature is qualitatively same as that observed for $B \neq 0$. However, the peaks in the conductance spectra vanish at $B = 0$ owing to the absence of Landau levels. \par
\section{Summary}\label{V}
In this work, we have studied the influence of a perpendicular magnetic field with a semi-Dirac dispersion within the framework of a tight-binding model of a honeycomb lattice. In order to compare and contrast with a prototype Dirac material, such as, graphene, we have presented our results for both cases. We consider a semi-Dirac nanoribbon with a finite width and study the Hofstadter butterfly and properties of the Landau level spectra. We have observed two identical gapped spectra symmetrically placed above and a below $E=0$ and along with a zero-energy mode in the Hofstadter butterfly spectrum in contrast to what is observed for a Dirac system. The Landau level becomes fully dispersive in the bulk for moderate values of the magnetic flux which is not true for the Dirac case. Furthermore, we explore the magneto-transport properties using the Kubo formula. We observed that the Hall conductivity shows standard quantization similar to that of a conventional semiconductor 2DEG with a parabolic band, which is highly contrasted with respect to a Dirac system. The zero Landau level peak is absent in the case of a semi-Dirac system. The longitudinal conductivities, $\sigma_{xx}$ and $\sigma_{yy}$ show anisotropic behaviour due to the distinct dispersion in two longitudinal directions.
\section{ACKNOWLEDGMENTS}
PS acknowledges Sim$\tilde{a}$o M. Jo$\tilde{a}$o, SK Firoz Islam, Abbout Adel for useful discussions and Rafiqul Rahaman for helping in computation. SB thanks SERB, India for financial support under Grant No. EMR/2015/001039. SM is supported by JSPS KAKENHI under Grant No. JP18H03678.
| a4fa96b146a414f9942076fb2d78f3bcb8135515 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Deep neural networks are now central to many real-world applications including image classification~\cite{ResNet,VGGnet,xu2019positive}, object detection~\cite{ren2016object,SSD}, low-level vision~\cite{SRCNN}, and natural language processing (NLP)~\cite{hochreiter1997long,mikolov2010recurrent,vaswani2017attention}. Their performance usually depends on the availability of large amounts of finely labeled training data. For example, the ImageNet dataset~\cite{deng2009imagenet}, which is used for image classification, includes over 1.2M images in 1000 different categories; the MegaFace face recognition dataset~\cite{kemelmacher2016megaface} has over 1M photos from 690k unique users; and the CoQA dataset~\cite{reddy2019coqa} used for NLP contains over 127k questions with answers collected from over 8k conversations.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{./figs/Intro}
\caption{Diagram of the proposed DC-NAS. The given search space will be converted to a series of feature representations of different architectures. Then, a k-means algorithm is utilized to divide the searching problem, and the best architecture will be obtained by merging selected model candidates.}
\label{Fig:pipeline}
\end{figure}
\begin{figure*}[t]
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.29\textwidth]{./figs/eval_5.pdf}&
\includegraphics[width=0.29\textwidth]{./figs/eval_10.pdf}&
\includegraphics[width=0.29\textwidth]{./figs/eval_15.pdf}
\\
\small{(a)} & \small{(b)} & \small{(c)}\\
\includegraphics[width=0.29\textwidth]{./figs/eval_20.pdf}&
\includegraphics[width=0.29\textwidth]{./figs/eval_200.pdf}&
\includegraphics[width=0.29\textwidth]{./figs/final.pdf}
\\
\small{(d)} & \small{(e)} & \small{(f)}\\
\end{tabular}
\caption{Illustrative results of searching different neural architectures using conventional early stopping. Five architectures with different widths (\emph{i.e. }, $4$, $2$, $1$, $\frac{1}{2}$, and $\frac{1}{4}$) are established in each panel. Their performance is tested on the CIFAR-100 validation set using different strategies. (a)-(e) are the results learned using different epochs on the $10\%$ training set, and (f) shows the exact performance of these networks.}
\label{Fig:Intro}
\end{figure*}
In addition to large amount of data, well-designed architectures are critical for effective deep neural networks, especially convolutional neural networks (CNNs) for computer vision tasks. There are many network architectures. Simonyan and Zisserman presented VGGNet~\cite{VGGnet}, which contains over $10^9$ learnable parameters. He~\emph{et.al.}~\cite{ResNet} presented a shortcut operation for training deep neural networks with over $50$ layers. ResNeXt~\cite{ResNeXt} combines the shortcut operation with the 'split-transform-merge' operation by first splitting the channel of each layer, then applying transformation separately, before finally merging the outputs with an addition operation. Huang~\emph{et.al.}~presented DenseNet~\cite{huang2017densely}, which strengths information flow by adding a shortcut to each pair of blocks. In addition, some modern neural architectures have significantly lower computational costs, allowing their use on mobile devices. MobileNet~\cite{Mobilenet} reduces the computational complexity by replacing the traditional convolution operation with a combination of depth-wise convolution and point-wise convolution. ShuffleNet~\cite{Shufflenet} enables information flow between groups by adding a channel shuffle operation after group convolution. Mobilenet-v2~\cite{sandler2018mobilenetv2} further uses a linear bottleneck and an inverted residual block to reuse the features and to overcome the manifold collapse problem produced by the ReLU operation. Han~\emph{et.al.}~developed GhostNet~\cite{han2019ghostnet} and built efficient neural networks by generating more features from some inexpensive operations.
Despite the success of these networks and their variations, their design and implementation still require considerable human and computational resources. To address this, a number of neural architecture search (NAS) algorithms have been proposed to automatically search for optimal neural architectures in the given search space and dataset. Zoph \emph{et.al.}~\cite{zoph2016neural} generated the model descriptions of neural networks using a recurrent network and produced higher-accuracy architectures by training the recurrent neural network (RNN) with reinforcement learning. Real \emph{et.al.}~\cite{real2019regularized} used a modified evolutionary algorithm to search for better neural architectures in a given search space by introducing an age property to favor younger genotypes. Liu \emph{et.al.}~\cite{liu2018darts} represented the network architecture based on continuous relaxation, which allowed efficient architecture searching using gradient descent. Wang~\emph{et.al.}~\cite{wang2018towards} proposed an evolutionary method to automatically identify redundant filters in pre-trained deep neural networks. Pham \emph{et.al.}~\cite{pham2018efficient} reduced the search cost by searching for an optimal subgraph within a large computational graph with a controller.
While these NAS methods have significantly contribute to searching for better neural architectures, they have a significant limitation: in order to achieve an acceptable search cost, each searched architecture is not precisely optimized. To this end, early stopping is widely used in NAS for fast comparison. However, considering all sub-networks together in early stopping is highly unstable since they each have variable computational power. Figure~\ref{Fig:Intro} illustrates a toy experiment comparing five simple networks on the CIFAR-100 dataset with different early stopping settings. Since the convergence of models with fewer parameters and lower computational complexity is usually faster, the early stopping results in the entire search space are biased. In addition, a consensus was recently reached~\cite{sciuto2019evaluating,li2019random,yang2019evaluation} that random searching sometimes performs better than objective function-oriented NAS methods, also suggesting that the NAS evaluation procedure requires further refinement.
In this paper, we present a divide-and-conquer NAS (DC-NAS) scheme to better evaluate and handle the huge search space of neural architectures. In doing so, we highlight that the unified early stopping strategy for evaluating all networks sampled from the entire search space is unreliable, since the networks all have different numbers of parameters and computational power. We first observe the gradient change of each network's layers and operations and cluster them into several categories using the conventional k-means algorithm. Then, network comparisons are conducted in the same cluster by exploiting the usual early stopping strategy. Optimal networks in each cluster are finally merged and carefully trained to derive the best neural architecture in the given search space. The DC-NAS pipeline is shown in Figure~\ref{Fig:pipeline}. Since the gradient change can be directly calculated from the super-network with only limited cost and the k-means approach is also very efficient, DC-NAS does not obviously increase searching costs. Our extensive experimental results show that the proposed method can effectively solve the inaccurate evaluation problem and more efficiently perform NAS. Compared to state-of-the-art approaches, DC-NAS finds better performing networks in the same search space.
\section{Related Works}
To solve the problem of inaccurate evaluation of different networks, this work aims to divide the complex NAS problem into several sub-problems and obtain the optimal architecture by merging their results. Current NAS algorithms can be divided into two evaluation approaches: the early stopping strategy for evaluation and predictor based architecture searching.
\subsection{Early Stopping in NAS}
In order to automatically design high-performance deep neural networks, a series of algorithms have been proposed to search neural architectures in a given search space. Zoph~\emph{et.al.}~\cite{zoph2016neural} first proposed the concept of large-scale image classifier searching, which encoded different operations (\emph{e.g. }, convolution and pooling) and RNN connections to search for effective architectures using reinforcement learning (RL) by treating the accuracy of the current architecture as a reward and selecting policies to obtain the new architecture. Real~\emph{et.al.}~\cite{real2019regularized} proposed an evolutionary-based algorithm (EA) to search the architectures. Specifically, each individual neural network was regarded as an architectural component. Searching was performed across generations by using mutations and re-combinations of architectural components, with components showing better performance on the validation set picked and inherited by the next generation during evolution. Pham~\emph{et.al.}~\cite{pham2018efficient} developed efficient-NAS (ENAS), which significantly reduced the search cost by training a large computational graph and searching for an optimal subgraph that maximized the expected reward on a validation set using the weight-sharing method.
Note that the reward used in RL and component performance in EA require training a large number of neural networks to convergence, which makes the search cost expensive. Thus, the early stopping strategy is commonly used to reduce the search cost, which trains the neural network over only a few epochs and uses the intermediate accuracy as the surrogate reward or performance of the component.
Early stopping is a very effective method for quickly evaluating or predicting the intermediate neural architectures during searches. However, effectiveness is compromised, since for different architectures, the relationship between the intermediate accuracy and the final accuracy is unclear. Specifically, taking different networks sampled from the entire search space and applying the same early stopping strategy is ineffective, since networks with more learnable parameters are usually harder to optimize.
In addition to the EA-based~\cite{yang2019cars} and RL-based NAS methods with early stopping, differentiable neural architecture searching (\emph{e.g. }, DARTS~\cite{liu2018darts}, Hit-Detector~\cite{guo2020hit}) also provides an efficient way to search deep neural networks. DARTS regarded the given super network as a continuous space and simultaneously maintained two sets of parameters with respect to the weights of the desired network and the weights for selecting different operations between nodes.
SNAS~\cite{xie2018snas} replaced the feedback mechanism triggered by constant rewards with gradient feedback from a generic loss. FBNet~\cite{wu2019fbnet} proposed a hardware-aware efficient conv-net designing method by integrating the network architecture latency into the loss function.
Differentiable NAS algorithms significantly increase the search speed by sharing weights across all sub-networks. However, determining the importance of each layer during the search is difficult. In other words, the parameter optimization results for the searched architectures are unreliable.
\subsection{Network Performance Predictor}
To avoid expensive search costs, a series of predictor-based methods have been proposed that better predict the latter part of the learning curve given the former part. For example, Domhan~\emph{et.al.}~\cite{domhan2015speeding} extrapolated network architecture performance from the first part of a learning curve by mimicking the early termination of bad runs using a probabilistic model. Klein~\emph{et.al.}~\cite{klein2016learning} studied the use of Bayesian neural networks to improve performance with a specialized learning curve layer.
Swersky~\emph{et.al.}~\cite{swersky2014freeze} assumed that the training loss during the fitting procedure roughly followed an exponential decay towards an unknown final value and proposed a regression model based on Bayesian optimization. Baker~\emph{et.al.}~\cite{baker2017accelerating} proposed a sequential regression model to predict the validation accuracy using features based on network architectures, hyper-parameters, and time-series validation performance data.
However, good performance relies on the assumption that the learning curve should be smooth, which is often not the case in reality, since the learning curve may abruptly change when the learning rate changes.
Some other predictors try to directly predict architecture performance before training~\cite{tang2020semi}. Deng~\emph{et.al.}~\cite{deng2017peephole} proposed using long short-term memory (LSTM) to predict the performance of a given architecture. Specifically, each layer of the architecture was embedded into a feature vector according to some predefined rules, and the vectors were then concatenated to produce the LSTM input. Sun~\emph{et.al.}~\cite{sun2019surrogate} encoded the network architecture by assuming that it is consisted of ResNet block and DenseNet block, extracted features based on it, and then predicted the accuracy with random forests. Xu~\emph{et.al.}~\cite{xu2019renasrelativistic} argued that the rankings between different models were much more important than absolute performance, and proposed a predictor based on a pairwise ranking-based loss function.
Although a well-trained neural architecture predictor can replace the early stopping approach and avoid the mixed evaluation problem, collecting the ground-truth labels (\emph{i.e. }, the exact network performance) to train the predictor is very expensive. Moreover, the predictor is sensitive to the number of training networks and the sampling strategy. Thus, NAS urgently requires an effective and accurate approach for comparing different searched architectures.
\section{Divide-and-conquer Searching}
We first revisit the evaluation of different sampled neural architectures in NAS, which is the main reason that some excellent architectures cannot be maintained during the search. Then, we develop divide-and-conquer NAS method. The original search space will be split to several clusters according to sub-network similarity.
\subsection{Evaluation Problem in NAS}
NAS represents a series of well-design approaches for obtaining the optimal neural architecture for the target task and related dataset from a huge search space. For a given search space $\mathcal S$ containing $p$ different neural architectures, \emph{i.e. }, $\mathcal{S} = \{\mathcal{N}_1,...,\mathcal{N}_p\}$, where $\mathcal N_i$ is the $i$-th sampled architecture from $\mathcal S$, each architecture is composed of a number of neural operations proven to be useful for deep learning (\emph{e.g. }, convolution, pooling, short-cuts, \emph{etc}.). Denoting the target dataset $\mathbf{D}$, the performance of each neural architecture can be calculated as
\begin{equation}
y_i = T(\mathcal{N}_i,\mathbf{D}),
\end{equation}
where $y_i$ is the accuracy of $\mathcal{N}_i$ on the test set and $T(\cdot)$ is the training function, which is tuned for the sampled network from $S$. In general, the training of a modern deep neural architecture is an intensive process (\emph{e.g. }, over eight GPU*days on ImageNet to train only one sub-network), and a large number of sub-networks exist in the search space. Evaluating all of these architectures individually is almost impossible. Thus, most NAS algorithms adopt an early stopping approach on a reduced dataset to evaluate each sub-network, \emph{i.e. },
\begin{equation}
E(\mathcal{N}_i) = \tilde T(\mathcal{N}_i,\tilde{\mathbf{D}}),
\label{Fcn:early}
\end{equation}
where $\tilde T(\cdot)$ is the training function with the early stopping strategy and $\tilde{\mathbf{D}}$ is the smaller dataset containing $\sigma$ (\emph{e.g. }, $5\%$) of the original training dataset to promote efficiency.
Although Eq.~\ref{Fcn:early} can significantly reduce the time taken to evaluate the searched networks, it is very hard to construct a smaller dataset and training method to satisfy the following objective function:
\begin{equation}
\min_E\frac{1}{p}\sum_{i=1}^p||E(\mathcal{N}_i)-y_i||_2^2,
\end{equation}
where $||\cdot||_2$ is the conventional $\ell_2$-norm. This is because deep neural networks are highly complex and hard to train on large-scale datasets, and there is no ranking-preserving relationship between early stopping performance and the final ground-truth accuracy of an arbitrary neural network.
It is in fact unnecessary to predict the exact accuracy (or ground-truth accuracy) of the searched networks. If the order of all sub-networks can be preserved, the evaluation function $E(\cdot)$ can also accurately help us to identify networks with better performance, significantly improving NAS efficiency. Therefore, the objective function for optimizing the evaluating function can be written as
\begin{equation}
\min_E\sum_{i=1}^p\sum_{j=i+1}^p \text{sgn}(y_j-y_i)\times\text{sgn}(E(\mathcal{N}_i)-E(\mathcal{N}_j)),
\label{Fcn:obj}
\end{equation}
where $\text{sgn}(\cdot)$ is the sign function. Then, our goal is to find an efficient evaluator $E(\cdot)$ and embed it into mainstream NAS algorithms.
Figure~\ref{Fig:Intro} shows that applying the same early stopping strategy to all networks cannot achieve the desired functionality, because the optimization routes of the various networks are different. For instance, some light-weight neural networks will first achieve a relatively high accuracy but their subsequent improvements are not obvious. However, if some neural networks have very close representations, their convergence statuses are similar. Therefore, the global early stopping strategy is unsuitable for constructing $E(\cdot)$.
\subsection{Architecture Clustering}
As discussed in Eq.~\ref{Fcn:obj}, an ideal evaluation function is defined to compare different network architectures during the search. In fact, Eq.~\ref{Fcn:obj} can also be regarded as a predictive problem, with some existing works addressing this~\cite{deng2017peephole,sun2019surrogate,xu2019renasrelativistic}. However, these methods require a certain number of fully-trained networks sampled from the given search space to construct the performance predictor, which is also very resource intensive.
Handling the performance comparison task using a universal early stopping strategy is unreliable; nevertheless, early stopping is still the cheapest way to compare a set of architectures without too many differences. Therefore, we propose to conduct NAS in a divide-and-conquer manner to reduce the difficulty in comparing various architectures during the search. In practice, we divide all the sub-networks in the given search space $\mathcal{S}$ into $K$ categories using $K$ cluster centers to form $K$ different groups, \emph{i.e. }, $\mathbf{G}=\{\mathbf{G}_1,...,\mathbf{G}_K\}$. The corresponding objective function can be written as
\begin{equation}
\arg\min_{\mathbf{G}} \sum_{k=1}^K\sum_{\mathcal{N}\in\mathcal{G}_k} ||F(\mathcal{N})-F(\tilde{\mathcal{N}}_k)||_2^2,
\label{Fcn:arc_cl}
\end{equation}
where $\tilde{\mathcal{N}}_k$ denotes the $k$-th clustering center of group $\mathbf{G}_k$, and $F(\cdot)$ is the feature representation of each input deep neural architecture.
Establishing the relationship between deep neural architectures and their performance is difficult, and how to represent deep networks is also extremely complicated due to the various structures and billions of trainable parameters. Fortunately, the current mainstream search space is often regularized to a super-network~\cite{liu2018darts,xie2018snas,wu2019fbnet}, \emph{i.e. }, the depths and types of connections are usually fixed, and all the operations needing to be searched are merged in it. Thus, we suggest representing each sampled network from the given search space $\mathcal{S}$ using a code with fixed length $L$, \emph{i.e. }, the number of layers.
One straightforward way to represent neural architectures is to regard the operations in each layer (\emph{e.g. }, ``conv'' and ``ReLU''~\cite{deng2017peephole,sun2019surrogate}) or computational complexity (\emph{i.e. }, FLOPS and memory~\cite{xu2019renasrelativistic,istrate2019tapas}) as neural architecture features. However, directly using these features does not capture the relationship between any two layers and operations, and the computational complexity cannot be computed for some parameter-free layers, \emph{e.g. }, pooling and short-cut layers. In addition, the phenomenon shown in Figure~\ref{Fig:Intro} motivates us to encode the optimization difficulty of different neural architectures. Therefore, given $L$ as the number of layers in the neural architectures and $\eta$ as the maximum epochs trained on the training set, we propose using the following representation function to encode an arbitrary network $\mathcal{N}_i$ into an $(L\times\eta)$-dimensional matrix, \emph{i.e. },
\begin{equation}
\hat F_{l,e}(\mathcal{N}_{i})=\frac{<{\theta_{i}^{l,1}}, {\theta}_{i}^{l,e}>}{||\theta_{i}^{l,1}||\cdot||{\theta}_{i}^{l,e}||},
\label{Fcn:pfeat}
\end{equation}
where $\theta_{i}^{l,e}$ are the parameters in the $l$-th layer in network $N_i$ trained on the reduced training set $\tilde{\mathbf{D}}$ after $e$ epochs. The above function calculates the changes in parameters of a given network between the first and the $e$-th epochs, which can reflect the sensitivity of each layer in the sampled network. For instance, a layer with fewer parameters will converge faster than another layer with more parameters due to the different optimization difficulty, as shown in Figure~\ref{Fig:Intro}. Thus, we can use these features to distinguish neural architectures with different properties and divide the huge search space $\mathcal{S}$ accordingly.
However, Eq.~\ref{Fcn:pfeat} also faces the same problem of not being able to represent some layers (such as pooling and short-cut) without trainable parameters. Thus, we further extend it to the output of each layer as the feature representation of a given neural network, \emph{i.e. },
\begin{equation}
F_{l,e}(\mathcal{N}_{i})=\frac{<{o_{i}^{l,1}}, {o}_{i}^{l,e}>}{||o_{i}^{l,1}||\cdot||{o}_{i}^{l,e}||},
\label{Fcn:feat}
\end{equation}
where $o_{i}^{l,e}$ is the averaged output feature of the $l$-th layer in network $\mathcal{N}_{i}$ for a certain amount of data trained on $\tilde{\mathbf{D}}$ with $e$ epochs. To save the calculation of Eq.~\ref{Fcn:feat} and balance the information of each dimensionality, we further apply a global average pooling on the feature representation ${o}$ of each layer before calculating $F$.
Then, we can utilize conventional k-means clustering or hierarchical clustering ~\cite{yaari1997segmentation,alsabti1997efficient} to divide the original search space into $K$ clusters according to the feature similarity between any two sampled neural architectures.
\begin{algorithm}[t]
\renewcommand\baselinestretch{1.1}\selectfont
\caption{DC-NAS for Searching Neural Architectures.}
\label{Alg:main}
\begin{algorithmic}[1]
\Require An arbitrary search space $\mathcal{S}$ containing $p$ sub-networks $\{\mathcal{N}_1,...,\mathcal{N}_p\}$ with $L$ layers. The target dataset $\mathbf{D}$ and it's shrunk version $\tilde{\mathbf{D}}$, the clustering number $K$, the number of epochs $\eta$ for extracting feature representations, and the number of sampled networks $s\leq p$.
\State \textbf{Search Space Clustering:}
\State Randomly generate $s$ neural architectures from $\mathcal{S}$;
\For{$i = 1$ to $s$}
\State Initialized the parameters in $\mathcal{N}_i$;
\State Train the network $\mathcal{N}_i$ on $\tilde{\mathbf{D}}$ with $\eta$ epochs;
\For{$l = 1$ to $L$}
\For{$e = 1$ to $\eta$}
\State Obtain features $o_{i}^{l,e}$ of the $l$-th layer after training $e$ epochs;
\State Calculate $F_{l,e}(\mathcal{N}_{i})$ using Eq.~\ref{Fcn:feat};
\EndFor
\EndFor
\EndFor
\State Cluster $s$ networks into $K$ groups $\{\mathcal{G}_1,...,\mathcal{G}_K\}$ using Eq.~\ref{Fcn:arc_cl};
\State \textbf{Merging Architectures:}
\For{$i = 1$ to $K$}
\State Find the best architecture $\hat{\mathcal{N}}_i$ in $\mathcal{G}_i$ using the conventional early stopping (Eq.~\ref{Fcn:res1} for architectures trained separately, Eq.~\ref{Fcn:res3} for architectures sampled from a super network);
\EndFor
\Ensure The optimal network $\mathcal{N}^*$ among the $K$ groups.
\end{algorithmic}
\end{algorithm}
\subsection{Divide-and-conquer Network Comparison}
After obtaining $K$ deep neural architecture clusters, the evaluation difficulty is naturally reduced, since each cluster represents deep models with similar optimization paths and performance. Therefore, we propose to first distinguish the best architecture in each group and then merge them to produce the final result.
For architectures in the same cluster, we can reuse the conventional early stopping strategy to distinguish searched architectures in the same cluster with similar gradient changes. Therefore, for $K$ clusters, the $K$ optimal architecture can easily be solved by using
\begin{equation}
\hat{\mathcal{N}}_k = \mathop{\arg\max}_{\mathcal{N}\in\mathcal{G}_k}E(\mathcal{N}) = \mathop{\arg\max}_{\mathcal{N}\in\mathcal{G}_k}\tilde T(\mathcal{N},\tilde{\mathbf{D}}),
\label{Fcn:res1}
\end{equation}
where $\hat{\mathcal{N}}_k$ is the best network architecture in the $k$-th group selected by the reduced dataset $\tilde{\mathbf{D}}$.
Eq.\ref{Fcn:res1} is suitable when evaluating different neural networks that have been trained separately. However, for the sub-networks extracted from the pre-trained super network, the performance of each sub-network without further training is usually poor and might not represent its true fitting ability. Note that in a given pre-trained super-network, the operation parameters in each layer give the probabilities of choosing different operations in that layer. Thus, we use the sub-network with the largest probability of being chosen in a cluster as the representation of this cluster. Given ${ a(\mathcal N)}=\{a(\mathcal N^1),a(\mathcal N^2),\cdot\cdot\cdot,a(\mathcal N^L)\}$ as the operation probability parameters of each layer in a sub-network $\mathcal N$, we have:
\begin{equation}
\hat{\mathcal{N}}_k = \mathop{\arg\max}_{\mathcal{N}\in\mathcal{G}_k}\sum_{l=1}^L{a(\mathcal N^l)},
\label{Fcn:res3}
\end{equation}
Then, the optimal neural architecture in the given search space $\mathcal{S}$ can be easily derived using the following function:
\begin{equation}
\mathcal{N}^* = \mathop{\arg\max}_{\hat{\mathcal{N}}_i}T(\hat{\mathcal{N}}_i,\mathbf{D}), \quad \forall\;\; i = 1,...K,
\label{Fcn:res2}
\end{equation}
where, the performance of each network selected in the specific cluster will be derived by fully training them on entire dataset $\mathbf{D}$ for an accurate comparison.
Note that although the training procedure on $\mathbf{D}$ is somewhat expensive, the number of clusters $K$ is relatively small and Eq.~\ref{Fcn:res2} does not dominate our scheme. The detailed searching procedure of DC-NAS for accurately searching neural architectures is summarized in Algorithm~\ref{Alg:main}.
\section{Experiments}
We have therefore developed a novel NAS framework using the divide-and-conquer approach, splitting the huge search space into several groups to reduce the difficulty in evaluating the searched neural architectures. Here we verify the effectiveness of the proposed method on several NAS search spaces.
\subsection{Validations on Toy Search Space}
\paragraph{\textbf{Search Space Definition.}} We first validate the proposed method using a very simple search space, as shown in Figure~\ref{Fig:Intro}.
The baseline network includes six convolutional layers and a fully-connected layer. We extend this baseline network by adjusting the channel ratio in different layers, \emph{i.e. }, $\frac{1}{4}$, $\frac{1}{2}$, $1$, $2$, $4$. Thus, the entire search space contains $5^6 = 15625$ neural architectures with different parameters and computational complexity. CIFAR-100~\cite{krizhevsky2009learning} is selected as the target dataset.
Since the search space is relatively small, we first train all of these networks on $10\%$ of CIFAR-100, \emph{i.e. }, $\tilde{\mathbf{D}}$ in Eq.~\ref{Fcn:early}. The number of epochs used in the original baseline network is 200. For efficiency, these architectures are trained separately for 20 epochs. Then, all the images in the smaller training dataset $\bf\tilde D$ are used to extract the feature representations of each layer of these architectures after training different epochs.
\paragraph{\textbf{Architecture Clustering and Searching.}}
After obtaining the features of these neural architectures, we obtain a feature tensor $\mathcal T\in \mathbb R^{s\times L\times\eta}$ in which $L=6$. Specifically, we set $s=15625$ and $\eta\in\{5,10,20,30\}$. Then, we apply the k-means algorithm on this matrix to generate a series of architecture groups. After that, we evaluate each network cluster using Eq.\ref{Fcn:res1}, and $K$ selected networks will be fully trained on the entire training dataset $\bf D$ to obtain the final result, \emph{i.e. }, the optimal architecture in the given search space.
\renewcommand{\multirowsetup}{\centering}
\begin{table*}[t]
\begin{center}
\renewcommand\arraystretch{1.0}
\caption{Searched results of DC-NAS with different clustering number $K$ and evaluating epochs $\eta$ on the toy search space.}
\begin{tabular}{l|c|c|c|c|c|c}
\hline
DC-NAS & $K = 1$ & $K = 3$ & $K = 5$ & $K = 10$ & $K = 20$ &$K = 50$ \\
\hline\hline
Top-1 acc(\%) (RS) & 75.01 & 75.36 & 75.49 & 75.58 & 75.71 & 75.83\\
\hline
Top-1 acc(\%) ($\eta = 5$) & 71.74 & 73.96 & 73.96& 73.96& 75.38& 75.47 \\
\hline
Top-1 acc(\%) ($\eta = 10$) & 71.13 & 75.42 & 75.42& 75.42& 75.83& 75.83\\
\hline
Top-1 acc(\%) ($\eta = 20$) & 73.32 & 75.91 & 75.91 &75.91 & 75.91&76.07\\
\hline
Top-1 acc(\%) ($\eta = 30$) & 74.07 & 75.91 & 75.91 &75.91 & 75.91&76.07\\
\hline
\end{tabular}
\label{Tab:kmeans}
\end{center}
\end{table*}
In Table~\ref{Tab:kmeans}, we test the impact of parameter $K\in\{1,3,5,10,20,50\}$ in dividing the search space according to the feature representations of the architectures. We also examine the impact of parameter $\eta$ ranging from $\{5,10,20,30\}$. When $K=1$, DC-NAS is equivalent to the conventional early stopping strategy for evaluating searched neural architectures, which is used in almost all NAS algorithms. An increasing $K$ is more likely to obtain the global optimum of the search space, while increasing the evaluation time. When $K=15625$, the proposed method degrades to a traversal search in which the global optimum is guaranteed, but the evaluation time is unaffordable. We further compare the effectiveness of the proposed method with random search (RS). Note that we have trained $s=15625$ different architectures on $10\%$ of the original dataset with $10\%$ of the original epochs. Thus, for a given parameter $K$, the RS method randomly selects $156+K$ different neural architectures from the search space, and the architectures are fully trained to obtain the validation result. The architecture with the best validation result is selected, and this experiment is repeated $20$ times to overcome randomness.
As shown in Table~\ref{Tab:kmeans}, the best search results are achieved when $\eta=30$, which is slightly higher than that of $\eta=20$ only when $K=1$. On the one hand, a larger $\eta$ indicates a higher dimension of the feature representation which increases the amount of information to help generate better clusters. On the other, a larger $\eta$ alleviates the randomness of the feature representation, and the feature can more precisely represent the degree of convergence of the layer. However, a larger $\eta$ will also increase the search cost of the proposed DC-NAS algorithm, and the architecture search problem is exactly a method of exhaustion when $\eta \rightarrow \inf$. Thus, we suggest to set $\eta = 20$ (\emph{i.e. }, $10\%$ of the epochs for fully training) for a trade-off between the performance of neural architecture and search cost.
\begin{figure*}[t]
\centering
\subfloat[K=1]{
\begin{minipage}[t]{0.28\linewidth}
\centering
\includegraphics[width=0.63\linewidth]{figs/cluster_1.png}
\end{minipage}
}
\centering
\subfloat[K=3]{
\begin{minipage}[t]{0.56\linewidth}
\centering
\includegraphics[width=0.99\linewidth]{figs/cluster_3.png}
\end{minipage}
}
\centering
\subfloat[K=5]{
\begin{minipage}[t]{0.84\linewidth}
\centering
\includegraphics[width=1\linewidth]{figs/cluster_5.png}
\end{minipage}
}
\caption{Visualization of the best deep neural architectures in each cluster with $K=1$, $K=3$ and $K=5$, respectively.}
\label{figure_4}
\end{figure*}
Moreover, $K$ is also a very important parameter in our method. When $K=1$, the random search method achieves the best result, which means that the conventional early stopping method is not comparable to RS. However, the proposed method delivers significant gains even with a small cluster number ($K=3$), improving the search result from $73.32\%$ to $75.91\%$ and outperforming the RS method. This shows that DC-NAS can achieve a promising local optimum with only a slightly increase in the evaluation cost. We found that all the selected networks $\{\hat{\mathcal N}_k\}_{k=1}^K$ with a smaller $K$ also appear when $K$ increases as shown in Figure~\ref{figure_4}. The best network searched using $K=5$ includes the results using $K=3$, which demonstrates that the performance of the searched architecture will not decrease when $K$ becomes larger. In addition, we also find that a larger $K$ will not obviously enhance the accuracy of the searched network, since the number good candidates with sufficient discrimination \emph{w.r.t. } the number of intrinsic clusters is limited, and we keep $K=5$ in the following experiments.
In order to explicitly understand the architecture clustering procedure, we detail the selected networks in each cluster using intra-cluster early stopping with $K = 5$ and $\eta=20$ (Table~\ref{Tab:resp}). Meanwhile, the searched architectures using different $K$ are also drawn in Figure~\ref{Fig:Intro}, selected networks in different clusters present variant architectures. Revising Figure~\ref{Fig:Intro}, for the given search space, the conventional early stopping method will search for the architecture with a small channel ratio, \emph{i.e. }, network architectures with few parameters and FLOPs, since a smaller network will converge faster on reduced dataset $\bf\tilde D$ at the beginning of training, thus outperforming larger networks with better ground-truth performance when applying the early stopping strategy. However, in Table~\ref{Tab:resp}, we show that there is an approximately $6$-times difference in the parameters and FLOPs between the largest and smallest models selected by DC-NAS, demonstrating that the proposed method can effectively solve the problem shown in Figure~\ref{Fig:Intro}. Furthermore, it can be found in Table~\ref{Tab:resp}, $\tilde{N}_3$ is about $1.87\times$ larger than $\tilde{N}_2$ with a slightly lower performance. This result illustrates that a larger model does not necessarily have higher accuracy. Overall, the proposed DC-NAS can distinguish models with different optimization difficulties and accurately evaluate them to obtain better architectures.
\begin{table}[t]
\begin{center}
\renewcommand\arraystretch{1.2}
\caption{Optimal architectures in each cluster when $K = 5$.}
\begin{tabular}{c|c|c|c|c}
\hline
Network & Channel Ratio & Param & FLOPs & acc(\%) \\
\hline\hline
$\tilde{N}_2$ & $4,2,1,2,0.25,2$ & 4.30M & 0.46G & 73.64\\
\hline
$\tilde{N}_1$ & $2,2,1,4,0.5,0.5$ & 4.76M & 0.61G & 73.50\\
\hline
$\tilde{N}_4$ & $2,4,4,2,0.25,1$ & 5.69M & 1.14G & 73.60\\
\hline
$\tilde{N}_3$ & $2,4,1,0.5,4,0.5$ & 8.08M & 0.72G & 73.32\\
\hline
$\tilde{N}_5$ & $2,4,4,2,4,1$ & 23.38M & 2.27G & 75.91 \\
\hline
\end{tabular}
\label{Tab:resp}
\vspace{-1.5em}
\end{center}
\end{table}
In addition, we sometimes need to sample some examples from the smaller dataset $\bf \tilde D$ in order to reduce the time taken to compute the averaged output feature of a specific layer in a given network (Eq.~\ref{Fcn:feat}). In order to verify the sensitivity of the proposed method in choosing different examples for extracting features, we repeat the search $10$ times with $K = 5$ and $\eta = 20$, each time randomly sampling 1000 images from the reduced training dataset $\bf \tilde D$ accordingly. The largest gap between the accuracy of the best and worst searched network is only about $0.2\%$, showing that the feature representation generation method utilized in Eq.~\ref{Fcn:feat} is very robust for distinguishing sub-networks in the given search space.
\begin{table*}[t]
\begin{center}
\small
\renewcommand\arraystretch{1.0}
\caption{Searched results of DC-NAS and state-of-the-art methods on ImageNet.}
\vspace{-0.5em}
\begin{tabular}{c|cc|ccc|cc}
\hline
\multirow{2}{*}{Method} & Search & Search & Latency & FLOPs & Params&Top-1 & Top-5 \\
& methods & space & (ms) & (M) & (M) & acc(\%) & acc(\%)\\
\hline\hline
Random Search &-&-&-&276&4.3&72.0&90.6\\
MobileNetV1~\cite{howard2017mobilenets} &handcraft & - & -&- & 4.2 & 70.6 & 89.5 \\
MobileNetV2 1.0$\times$~\cite{sandler2018mobilenetv2} & handcraft &- &-& 300 & 3.4 & 72.0 & 91.0\\
ShuffleNetV1 1.5$\times$ (g=3)~\cite{zhang2018shufflenet} & handcraft & &-& 292 & - & 71.5 & -\\
ShuffleNetV2 1.5$\times$~\cite{ma2018shufflenet} & handcraft & -&-& 299 & 3.5 & 72.6 & - \\
MobileNetV2 1.4$\times$~\cite{sandler2018mobilenetv2} & handcraft &-& -& 585 & 6.9 & 74.7 & - \\
ShuffleNetV2 2.0$\times$~\cite{ma2018shufflenet} & handcraft & -&- & 591 & 7.4 & 74.9 & - \\
\hline
ChamNet-B~\cite{dai2019chamnet} & predictor & layer-wise &-& 323 & - & 73.8 & -\\
NASNet-A~\cite{zoph2018learning} & RL & cell &-& 564 & 5.3 & 74.0 & - \\
MnasNet~\cite{tan2019mnasnet} & RL & stage-wise & -&317 & 4.5 & 74.0 & -\\
FBNet-B~\cite{wu2019fbnet} & gradient & layer-wise & 87.07&295 & 4.5 & 74.1 & - \\
FBNet-B (our impl.) & gradient & layer-wise & 104.25& 326 & 4.7 & 73.7 & 91.5 \\
ProxylessNAS-R~\cite{cai2018proxylessnas} & gradient & layer-wise &-& -& 5.8 & 74.6 & 92.2\\
MnasNet-92~\cite{tan2019mnasnet} & RL & stage-wise& -&388 & 4.4 & 74.8 & -\\
FBNet-C~\cite{wu2019fbnet} & gradient & layer-wise &102.83& 375 & 5.5 & 74.9 & - \\
FBNet-C (our impl.) & gradient & layer-wise &116.55& 406 & 5.5 & 74.8 & 92.1 \\
\hline
DC-NAS-A & & \multirow{5}{*}{layer-wise} & 117.89 &319 & 4.7 & 73.4& 91.5\\
DC-NAS-B & gradient & & 84.19 &328 & 4.9 & 73.7& 91.5\\
DC-NAS-C & + & & 104.10&341 & 5.0 & 74.2& 91.6\\
DC-NAS-D & cluster & & 103.39&369 & 4.9 & 74.7& 92.1\\
DC-NAS-E & & & 118.12&428 & 5.5 & \textbf{75.1}& \textbf{92.2}\\
\hline
\end{tabular}
\label{Tab:imgnet}
\end{center}
\end{table*}
\subsection{Validations on ImageNet}
After conducting experiments on CIFAR-100 data and analyzing the impact of each parameters in the proposed DC-NAS, we further employ the new method on the challenging large-scale ImageNet dataset~\cite{ImageNet}. This dataset contains over 1.2M images from 1000 categories. FBNet~\cite{wu2019fbnet} is selected as the baseline method because of the excellent performance on both model accuracy and latency.
\paragraph{\textbf{Search Space Definition.}} We use the same search space proposed in FBNet~\cite{wu2019fbnet}, which is a layer-wise search space with a fixed super-network. The super-network architecture is composed of a $3\times3$ convolution layer, followed by seven SB blocks, a $1\times1$ convolution layer, a $7\times7$ average pooling layer and a fully-connected layer. Here SB block is the block that needs to be searched. Specifically, the block structure is fixed as a $1\times1$ convolution followed by a $k\times k$ depthwise convolution and another $1\times1$ convolution. ReLU activation is used after each layer, except for the last $1\times1$ convolution. When the stride of the block $s=2$, the stride of the first $1\times1$ convolution is 2. When $s=1$, a skip connection is added between the input and the output of the block.
When searching for the SB block, the expansion ratio $e$, which determines the channel size expansion from input to output of the first $1\times1$ convolution, can be searched from $e\in\{1,3,6\}$, and the kernel size $k$ of the depthwise convolution is $k\in\{3,5\}$. Furthermore, the group convolution followed by a channel shuffle operation can be used in the first and last $1\times1$ convolutions. Finally, a skip operation can be applied to the SB block, which actually cancels the block and reduces the architecture depth. There are $9$ different candidate operations forming a search space of size $9^{22}$.
\paragraph{\textbf{Results on ImageNet.}} Following~\cite{wu2019fbnet}, we first train the super-network on ImageNet-100 with 90 epochs. The first 10 epochs are trained with fixed operation parameters and the next 80 epochs are trained normally. In each epoch, the operator weight $w$ is first trained on $80\%$ of ImageNet-100 using SGD, and the probability parameters $a$ are trained on the remaining $20\%$ of the training set using Adam with an initial learning rate of 0.01 and weight decay of 0.0005. We then sample $10^7$ different architectures from the search space according to the operation probability of the pre-trained super-network. Specifically, we normalize the probabilities of the operations in the same layer, and then select operation of that layer based on the normalized probability. After that, we extract features using Eq.~\ref{Fcn:feat} with $10000$ images sampled from the ImageNet-100 training set. Finally, the feature representations of $10^7$ different samples are clustered into $K$ different groups using the k-means algorithm, and the representative sub-network from each group is selected using Eq.~\ref{Fcn:res3}. Considering the evaluation cost, we manually set a small cluster number $K=5$ in our experiment. When increasing $K$ to a larger number (\emph{e.g. }, $K=10$), we do not discover a significant increase of the search result. The five representative sub-networks are fully trained on ImageNet, and the search results are shown in Table~\ref{Tab:imgnet}, \emph{i.e. }, DC-NAS-A to DC-NAS-E.
\paragraph{\textbf{Compare to state-of-the-art methods.}} The results of the proposed DC-NAS are compared to state-of-the-art models designed manually and automatically. The evaluation metrics are top-1/top-5 accuracies on the ImageNet validation set, and FLOPs and parameters of the searched models. During the experiment, we found that the results of FBNet~\cite{wu2019fbnet} reported in the original paper is different from the results that is implemented by ourselves, which was also pointed out in ~\cite{stamoulis2019single}. Thus, we report both the evaluation metrics of the searched results mentioned in the original paper and the results implemented by us (denote as `our impl.'). Table~\ref{Tab:imgnet} shows that our DC-NAS achieves a top-1 accuracy of \textbf{$75.1\%$}, which is a new stat-of-the-art ImageNet accuracy among hardware-efficient NAS models. Meanwhile, we also compare the proposed DC-NAS with a number of state-of-the-art methods including NASNet~\cite{zoph2018learning}, MnasNet~\cite{tan2019mnasnet}, ProxylessNAS~\cite{cai2018proxylessnas}, \emph{etc}. The comparison results shown in Table~\ref{Tab:imgnet} also demonstrate that the proposed DC-NAS outperforms other methods under the same range of FLOPs.
\paragraph{\textbf{Search cost.}} The training process of super-net is exactly the same as in FBNet~\cite{wu2019fbnet}, which is 216 GPU hours. In this experiment, we do not need to adjust the hyper-parameter $\eta$. Extract features of $10^7$ different architectures using Eq.~\ref{Fcn:res3} takes about 15 GPU hours. Finally, clustering architectures into different groups by applying k-means on the extracted features can be done in a few minutes. Thus, the total search cost of DC-NAS is 231 GPU hours, which is $1.07\times$ compared to that of the baseline FBNet~\cite{wu2019fbnet}. In addition, it is easy to embed the proposed divide-and-conquer strategy into other frameworks to obtain better performance.
\paragraph{\textbf{Latency.}} Besides the FLOPS and model sizes, we also report the latency of different model on an ARM based mobile device for a fair comparison. The latency of the DC-NAS-E with the highest performance for predicting a $224\times 224$ image is $118.12ms$, which is very close to that ($116.55ms$) of the FBNet-C. At the same time, we can find from Table~\ref{Tab:imgnet}, the latency of a neural network is not linearly related to FLOPS and model size. Thus, the hardware-aware strategy is essential for neural architecture search. Overall, the proposed DC-NAS can provide networks with the state-of-the-art performance on the given search space.
\section{Conclusions}
In this paper we present a divide-and-conquer neural architecture search (DC-NAS) approach for effectively searching deep neural architectures. The proposed method overcomes the evaluation problem inherent in traditional NAS methods, namely they choose architectures that perform well on smaller datasets using the early stopping strategy while rejecting architectures that are actually acceptable after full training. Specifically, DC-NAS first extracts features that represent the speed of convergence of sub-networks according to the change in output features of each layer after each training epoch. It then calculates the similarity between two different architectures according to the feature representations. After that, traditional k-means clustering is used to divide the huge search space into several groups, and the best architectures in each group are further compared to obtain the best searched architecture. Our experimental results show that the proposed DC-NAS can significantly improve the performance of the searched architecture with only a slightly increase in the evaluation cost compared to traditional NAS methods.
| 44498aa344312f123ee83fbfff8259e200f41b71 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and Summary}
The area theorem, or second law of black holes, has pervasive implications in all of black hole physics. It puts absolute bounds on gravitational wave emission in collisions (Hawking's original motivation in \cite{Hawking:1971tu}) and limits other classical black hole evolutions, but also, through the identification of horizon area as entropy \cite{Bekenstein:1974ax}, it gives an entry into quantum gravity, holography, and applications of the latter to strongly coupled systems.
Investigating the growth of black hole entropy should throw interesting light into complex dynamical black hole processes. How does the second law constrain the possible final states? Are there phenomena where it can provide more than bounds on allowed outcomes, for instance, indicating their likelihood, according to how much entropy they generate? Since the area of the event horizon can be computed outside stationary equilibrium, one may even study the mechanisms that drive its growth at different stages.
Unfortunately, computing this entropy during a highly dynamical process, such as a black hole merger, is in general very complicated and requires sophisticated numerical calculations. In this article we resort to an approach that simplifies enormously the task: the effective theory of black holes in the limit of a large number of dimensions, $D\to\infty$ \cite{Emparan:2013moa,Emparan:2020vyfinr}, developed in \cite{Emparan:2015hwa,Bhattacharyya:2015dva,Emparan:2015gva,Bhattacharyya:2015fdk}. We use the equations of \cite{Emparan:2015gva,Emparan:2016sjk} for the study of asymptotically flat black holes, their stability, and collisions between them \cite{Andrade:2018nsz,Andrade:2018yqu,Andrade:2019edf}.
We examine in detail the production of entropy---its total increase, but also its generation localized in time and in space on the horizon---in processes where an unstable black hole (a black string \cite{Gregory:1993vy,Lehner:2010pn} or an ultraspinning black hole \cite{Myers:1986un,Emparan:2003sy}) decays and fissions, and in collisions where two black holes fuse into a single horizon. If the total angular momentum in the collision is not too large, the fusion ends on a stable rotating black hole. However, as shown in \cite{Andrade:2018yqu,Andrade:2019edf}, when $D$ is large and if the total angular momentum is also large enough, the merger does not end in the fusion, but proceeds to fission: the intermediate merged horizon is unstable, pinches at a neck---in a mild violation of cosmic censorship \cite{Andrade:2019edf,Emparan:2020vyf}---and arguably breaks up into two (or possibly more) black holes that then fly apart.
The phenomenon, up until the formation of the singular pinch, has been verified to occur in $D=6,7$ through numerical analysis in \cite{Andrade:2020}. The importance of the intermediate phase in the evolution of the system was noticed in the earlier studies \cite{Andrade:2018yqu,Andrade:2019edf}, but here we will go significantly further in revealing how it controls the outcome.
Before we present our main results, we shall discuss general issues related to area growth, its identification with entropy, and its computation in the large $D$ effective theory.
\paragraph*{Black hole entropy and its growth.}
In General Relativity the horizon area increases through two effects: the addition of new generators to the horizon (at caustics or crossover points, on spacelike crease sets), and the expansion of pencils of existing generators. In this article, the methods and approximations that we employ allow to study the latter, {i.e.,\ } how the area expands smoothly. The effective theory of large $D$ black brane dynamics provides explicit entropy production formulas for viscous dissipation on the horizon \cite{Emparan:2016sjk,Herzog:2016hob,Bhattacharyya:2016nhn,Dandekar:2017aiv}, which we apply to the evolution of unstable asymptotically flat black holes, and to the fusion and fission in black hole collisions.
The addition of generators through caustics is actually suppressed when $D$ is large\footnote{This can be seen, for instance, in extreme-mass-ratio mergers with the methods of \cite{Emparan:2016ylg}.}. We expect that, at finite $D$, the entropy growth through this addition in the merger of two black holes is important only during the first instants of the fusion, and much less so during the relaxation. In fission, the break-up of the horizon across a naked but mild singularity involves a region of very small horizon area\footnote{A Planckian area, which vanishes in the classical limit.} and therefore loses only a few generators in any dimension, and even fewer as $D$ grows large. This process is controlled by quantum gravity, but we have argued elsewhere that the effects upon the classical evolution should be negligibly small \cite{Emparan:2020vyf}.
The notions of entropy and entropy current that we use are associated to the area of the event horizon, and in this respect they are closely related to the ones in the Fluid/Gravity correspondence in \cite{Bhattacharyya:2008jc,Bhattacharyya:2008xc}. Indeed, we expect that the discussion in that context carries over to the large $D$ formulation: it is possible that other currents with non-negative divergence can be constructed. A related concern is whether one should identify the entropy with the area of the apparent horizon, as there are general arguments in favor of this \cite{Engelhardt:2017aux}, and it has been possible to identify corresponding currents in Fluid/Gravity \cite{Booth:2011qy}. While it would be interesting to further investigate this in the large $D$ effective theory, we will not be concerned with it here. The divergence of the entropy current that we use is as expected for a physical fluid (from viscous dissipation of shear and expansion of the fluid), so at the very least our results will not be unreasonable. Moreover, we expect that the growth properties of different entropy notions will be very similar. Within the large $D$ effective theory, the system evolves smoothly and continuously and so we expect the entropy to do that too. With these methods one does not capture the less smooth features ({e.g.,\ } caustics) of the event horizon in the first stages of the merger, and large discontinuous jumps in the area of the apparent horizon are not expected; these should be suppressed when $D\gg 1$. So, despite the ambiguities in the definition of out-of-equilibrium entropy, we expect that our conclusions remain qualitatively valid for other viable notions of it.
The other main aspect of finite $D$ physics that is not captured by our methods is the production of gravitational waves, which implies that in our calculations the total energy and angular momentum of the black holes are conserved, making it easier for us to characterize the evolutions. Again, we expect that this radiation is stronger during the initial instants when the black holes first come together. Radiation effects should quickly become less relevant as the number of dimensions grows \cite{Cardoso:2002pa,Emparan:2013moa,Bhattacharyya:2016nhn,Andrade:2019edf}.
The upshot is that we expect that the patterns of entropy production that we find are broadly applicable in $D\geq 6$, and possibly even qualitatively valid for fusion in $D=4$. We will return to this last point near the end.
\paragraph*{Main conclusions.}
All the collisions we study are symmetric, {i.e.,\ } between black holes of equal mass and spin.\footnote{In $2\to 1\to 2$ collisions, the two outgoing black holes will both have the same mass and spin, but the initial and final spins will in general be different. Mass conservation to leading order in $1/D$ implies that the final masses are the same as the initial ones.} This is mostly for simplicity; our methods allow to collide black holes with generic parameters.
Our analysis shows that:
\begin{itemize}
\item Black hole fusion generates comparatively much more entropy, and at faster rates, than black hole fission.
\item Unstable black strings decay with a simple pattern of entropy production which is reproduced in other fission processes.
\item Merger collisions of two black holes have a critical value of the total angular momentum per unit total mass\footnote{These $J$ and $M$ are defined in the effective theory; the corresponding physical quantities are given in appendix~\ref{app:phys}.}
\begin{equation}\label{JMc}
\left(\frac{J}{M}\right)_c\approx 2.66
\end{equation}
that divides low $J/M$ collisions $2\to 1$ that end in fusion, from higher $J/M$ collisions $2\to 1\to N\geq 2$ that evolve to fission. The bound holds except for `grazing' mergers with large initial impact parameters.
\item $2\to 1\to 2$ collisions are dominated by the formation of intermediate, long-lived, quasi-stationary, bar-like entropic attractors (fig.~\ref{fig:attract}):
{\renewcommand{\labelenumi}{(\alph{enumi})}
\begin{itemize}
\item The intermediate quasi-thermalization largely erases the memory of the initial state, so the final outgoing states are almost independent of the initial parameters, other than the total conserved $J/M$.
\item This attractor effect is stronger the closer $J/M$ is to the critical value \eqref{JMc}.
\item The attractor can be approximately (but not exactly) predicted by maximizing the entropy generation among possible outgoing black holes.
\end{itemize}
}
\item Entropy is produced through viscous dissipation of shear and expansion of the effective velocity field. In fusion generically, and in fission always, both enter in almost equal proportion. The formation of the intermediate bar phase can be dominated by shearing depending on initial conditions.
\end{itemize}
\begin{figure}[t]
\centering
\vcenteredhbox{\includegraphics[width=0.55\linewidth]{fig/NeutralClusteringJ2p8}}
\hspace{1em}
\vcenteredhbox{\includegraphics[width=0.4\linewidth]{fig/EntropicAttractorJ2p8}}
\caption{\small Entropic attractor in black hole collisions $2\to 1\to 2$. \emph{Left}: red dots represent initial states, connected by an arrow line to the corresponding final states (green) after dynamical evolution. The initial and final pairs of black holes are characterized by their rotation (spin) parameter $a$ and their linear velocity $u$ (ingoing or outgoing). The (conserved) total angular momentum per unit mass in these collisions is fixed to $J/M=2.8$. We see that, independently of the initial states, the final states cluster on $(a,u)\approx (0.3,0.6)$. This is due to the formation of an intermediate, long-lived, quasi-thermalized phase. The contour colors correspond to the (NLO) entropy $\mc{S}_1$ of the configuration. Entropy would be maximized at the lower-left corner, but this would correspond to infinite impact parameter $b$, which is unphysical since $b$ is constrained by the geometric size of the collision.
The thin purple strip is the region where $b$ takes on geometrically-allowed values for final states. \emph{Right}: entropy along the central value of the purple strip. The attractor is close to the maximum possible final entropy; larger values of $a$ (smaller values of $u$) would be entropically disfavoured, while in the opposite direction the entropy gain would be very small.
\label{fig:attract}}
\end{figure}
The reason the attractor effect is stronger when $J/M$ is near criticality is that the intermediate state is closer to a marginally stable solution. When $J/M$ is higher, the intermediate black hole is shorter-lived and its features are less precisely defined, so its decay outcomes show larger spread. We emphasize that the attractor is a feature of $2\to 1\to 2$ collisions with intermediate fusion; this requires that the initial impact parameter and initial velocities are not too large, otherwise the two black holes fly by each other. Fusionless $2\to 2$ collisions are not included in fig.~\ref{fig:attract}, and are little studied in this article.
Let us elaborate on the near-maximization of entropy in collisions $2\to 1\to 2$. After fixing an overall scale by setting the total mass to one, the outgoing black holes are characterized by their spin parameter $a$, outgoing velocity $u$, and outgoing impact parameter $b$.\footnote{There is one more outgoing parameter: the scattering angle. However, this is not affected by conservation laws nor by entropic considerations, and we shall have little to say about it.} We argue that these can be well predicted by considering three different constraints:
{\renewcommand{\labelenumi}{(\roman{enumi})}
\begin{enumerate}
\item Kinematic: the total final angular momentum must be the same as the initial one, which imposes a relation between $a$, $b$ and $u$.
\item Geometric: the outgoing impact parameter is limited by the geometrical size of the collision. This yields a constraint between $b$ and $a$ (appendix~\ref{app:bout}).
\item Entropic: after imposing (i) and (ii), near-maximization of the final entropy gives a good approximation to the final values of $a$, $b$, and $u$.
\end{enumerate}
}
In fig.~\ref{fig:attract}, (i) is included by considering states with given total $J/M=2.8$, while (ii) restricts final states to the lie along the purple band; the graph on the right then makes (iii) apparent.
The most remarkable of these constraints is entropy maximization: it provides a simple proxy for the complex dynamics that drives the system to its final state. We do not have an answer to why it is not a perfect predictor---other than there is no reason that it should be---but given our results it is natural to wonder how accurate it becomes as $J/M$ approaches from above the critical value \eqref{JMc}.
The principle of entropy maximization actually holds quite well in four-dimensional black hole collisions: in the merger, the final entropy would be maximal (consistent with conservation of energy and angular momentum) if no gravitational waves were emitted. The fact that the radiated energy is typically only a few-percent fraction of the total energy means that the final entropy is only a few percent off the maximum. In the limit $D\to\infty$, the $2\to 1$ fusion trivially maximizes the entropy since radiation is absent. The fission of an unstable black object, instead, has a range of possible outcomes. For instance, black strings can split up into several blobs, and the final entropy is larger for fewer blobs. The decay of ultraspinning black holes (MP, bars, and dumbbells) is more similar to the fission stage in $2\to 1\to 2$ collisions, but the evolution of the instability is sensitive to the specific perturbation that triggers it. The process starts with an unstable system and looks more contrived and less natural than a collision, so in our study we have focused mostly on the latter. Note, however, that the decay of the critical, marginally stable solution at \eqref{JMc} may illuminate the question of how closely can entropy be maximized. This deserves closer examination.
\paragraph*{Entropy generation and irreversibility in the leading order (LO) large $D$ effective theory.}
Readers familiar with the large $D$ effective theory of black holes and branes may be surprised that the entropy growth can be computed with its equations to leading order in the $1/D$ expansion. This theory is known to exactly conserve the entropy of the system: the LO entropy current is divergence-free, and entropy generation is suppressed by a factor $1/D$ \cite{Emparan:2015gva,Bhattacharyya:2016nhn}. A simple illustration of this property is the fusion of two equal mass Schwarzschild black holes \cite{Caldarelli:2008mv,Emparan:2013moa}, each with entropy
\begin{equation}\label{bhentD}
S(M)\propto M^{\frac{D-2}{D-3}}\,,
\end{equation}
which merge into a single one, so that (since losses into radiation are suppressed non-perturbatively in $1/D$)
\begin{equation}\label{fusionent}
\frac{S_\text{final}}{S_\text{initial}}=\frac{S(2M)}{2S(M)}=1+\frac{\ln 2}{D} +\ord{\frac1{D^2}}\,,
\end{equation}
{i.e.,\ } the entropy increase is $\propto 1/D$. This feature extends to all of the dynamics of black branes at large $D$ described by the LO effective theories of \cite{Emparan:2015gva,Bhattacharyya:2016nhn}.
This seems to make it impossible to see entropy growth unless one employs the next-to-leading order (NLO) theory. It also raises a puzzle: if the LO entropy does not grow, how can we characterize the irreversibility of the evolution in the LO theory?
We will argue that there exists a quantity $S_1(t)$ in the LO theory such that the evolution equations imply $\partial_t S_1(t)\geq 0$, and hence characterizes the irreversibility of this theory. This $S_1$ is actually the NLO ($1/D$ suppressed) entropy density, but we might not (indeed, need not) have known this, since $S_1$ and its variations are all given by LO magnitudes.
The argument of \eqref{fusionent} is still valid when the black holes rotate, since rotation effects in the entropy are suppresed by $1/D$ \cite{Emparan:2015gva,Andrade:2018nsz}. However, it ceases to apply if the black holes carry charge. Correspondingly, the effective theory of large $D$ charged black branes \cite{Emparan:2016sjk,Andrade:2018rcx} allows entropy production, through charge diffusion (resistive Joule heating), at leading order in $1/D$.\footnote{However, entropy production in the theory of charged membranes in \cite{Bhattacharyya:2015fdk} is zero at LO.} The study of entropy production in charged collisions is much simpler than in the neutral case, since it is largely dictated by conservation laws, but it is still illustrative and confirms the conclusions above. We will discuss it after our analysis of neutral collisions.
\paragraph{Outline.} In section~\ref{sec:efftheory} we introduce the basic elements of the large $D$ effective theory of black branes, specifically how entropy generation can be studied within the context of the LO theory. In sec.~\ref{sec:bhcoll} we discuss localized black hole solutions in this effective theory and how their stability properties influence the outcome of collisions. In sec.~\ref{sec:entpro} we perform numerical simulations of evolutions of instabilities and collisions. We investigate in detail the generation of entropy, in time and in space, and use it to characterize the different stages in the collision. The study in sec.~\ref{sec:bhscat} of the scattering of black holes reveals the role as an attractor of the intermediate state which nearly maximizes entropy production. Sec.~\ref{sec:chadif} describes how entropy is produced through charge diffusion in collisions between charged black holes. We conclude in sec.~\ref{sec:concl}.
\section{Entropy production in the large $D$ effective theory}\label{sec:efftheory}
We begin with a discussion of the large $D$ effective theory of black branes, with a focus on entropy and its generation. As we will review in the next section, this theory can be used to study localized, asymptotically flat black holes (Schwarzschild-Tangherlini, Myers-Perry, and others) at large $D$, and their collisions. The extension of the results in this section to the large $D$ effective theory of AdS black branes is straightforward (appendix~\ref{app:AdS}).
The field variables of the effective theory of large $D$ black branes are the mass density of the brane $m(t,\mathbf{x})$ and the velocity field $v^i(t,\mathbf{x})$, where $\mathbf{x} = (x^i)$, and $i=1,\dots, p$, label directions on the flat worldvolume of the brane. Only these $p$ spatial directions have non-trivial dynamics, while the remaining
\begin{equation}
n=D-p-3
\end{equation}
dimensions play a passive role, serving to perform the large-$D$ localization of dynamics near the horizon of the black hole.
We refer to \cite{Emparan:2015gva,Emparan:2016sjk} for how $m(t,\mathbf{x})$ and $v^i(t,\mathbf{x})$ determine the geometry of a black brane in a spacetime with a large number of dimensions $D$.
In order to have a solution of the Einstein equations in the large $D$ limit, they must solve the effective field equations \cite{Emparan:2015gva,Emparan:2016sjk,Dandekar:2016jrp}
\begin{equation}\label{dtm}
\partial_t m+\partial_i\left( m v^i\right)=0\,,
\end{equation}
\begin{equation}\label{dtmv}
\partial_t (m v^i) +\partial_j \left( m v^i v^j+\tau^{ij}\right)
=0\,,
\end{equation}
with
\begin{equation}\label{stressneut}
\tau_{ij}= - m\,\delta_{ij} -2m \partial_{(i}v_{j)}- m\,\partial_j\partial_i \ln m\,.
\end{equation}
Written in this form, they are the equations of a non-relativistic, compressible fluid with mass-energy density $m$, velocity $v^i$, and stress tensor $\tau_{ij}$.
The first two terms in the stress tensor \eqref{stressneut} correspond, respectively, to (negative) pressure
\begin{equation}
P=- m\,,
\end{equation}
and to viscous terms, with shear and bulk viscosities
\begin{equation}
\eta=m\,,\qquad
\zeta=\frac{2}{p}\eta\,.
\end{equation}
The last term in \eqref{stressneut}, in a hydrodynamic interpretation, is a second-order transport term.\footnote{It is more easily understood in the elastic interpretation of the effective equations as coming from the extrinsic curvature of the brane \cite{Emparan:2016sjk}.} In the large $D$ limit it must not be assumed to be small: it enters at the same order in $1/D$ as the other terms in \eqref{stressneut}.
The entropy density and the temperature can be obtained from the area density and surface gravity of the black brane. They are
\begin{equation}\label{sTneut}
s=4\pi m\,,\qquad T=\frac1{4\pi}\,,
\end{equation}
and they satisfy the expected thermodynamic relations for the system,
\begin{equation}
m=Ts\,,\qquad dm=Tds\,.
\end{equation}
\subsection{Irreversibility in the effective theory}
Let us now address in detail an elementary puzzle of this effective theory that we alluded to in the introduction. The equations \eqref{dtm}, \eqref{dtmv} contain viscous dissipation, which is expected to render the evolution irreversible.\footnote{This is also apparent using the variable $p_i=mv_i +\partial_i m$, in which \eqref{dtm} and \eqref{dtmv} take the form of inhomogeneous heat equations \cite{Emparan:2015gva}.} This is further confirmed by the spectrum of linearized perturbations, which has quasinormal frequencies with imaginary parts. After all, these equations describe horizons, which are dissipative systems par excellence, but for the moment let us forget black holes and regard by itself the system that these effective equations describe.
On very general grounds, we expect that dissipation in a thermodynamic system creates entropy, reflecting the irreversibility of the evolution. However, in the theory described by \eqref{dtm} and \eqref{dtmv} the total entropy
\begin{equation}
S(t)=\int d^p x\, s(t,x)
\end{equation}
remains constant in time, since \eqref{sTneut} implies that it is exactly proportional to the total mass and this is conserved by \eqref{dtm}.
So, if the entropy is not growing, what, then, characterizes the irreversibility?
Remarkably, we can identify a quantity in this theory that is strictly non-decreasing under time evolution. Define the density\footnote{The factors $4\pi$ that we carry over have their origin in $T$ in \eqref{sTneut}.}
\begin{equation}\label{s1}
s_1=4\pi\left( -\frac12 m v_i v^i -\frac1{2m}\partial_i m\,\partial^i m +m \ln m\right)\,.
\end{equation}
We will justify the choice presently, but for now note that using the field equations \eqref{dtm} and \eqref{dtmv} it follows that
\begin{equation}\label{dts1}
\partial_t s_1 +\partial_i {j_1}^i =8\pi m\left( \partial_{(i} v_{j)}\right)\left(\partial^{(i} v^{j)}\right)\,,
\end{equation}
where
\begin{equation}\label{j1}
{j_1}^i =s_1 v^i -4\pi \left( v^j \tau^{ij}+(\partial_j m)(\partial^j v^i)\right) \,.
\end{equation}
Since the right hand side of \eqref{dts1} is non-negative, we conclude that
\begin{equation}\label{S1}
S_1(t)=\int d^p x\, s_1(t,x)
\end{equation}
is a non-decreasing function in time,
\begin{equation}\label{dtS1}
\partial_t S_1(t)\geq 0\,.
\end{equation}
This characterizes the irreversibility of the evolution in the effective theory.
Observe that the growth rate
\begin{eqnarray}\label{dtSetazeta}
\partial_t S_1(t) &=& 8\pi \int d^p x\, m \left( \partial_{(i} v_{j)}\right)\left(\partial^{(i} v^{j)}\right)\nonumber\\
&=&\int d^p x\, \left( \frac{2\eta}{T}\sigma_{ij}\sigma^{ij} +\frac{\zeta}{T}\left(\partial_i v^i\right)^2\right)
\end{eqnarray}
is that of a hydrodynamic entropy generated by viscous heating, with contributions from dissipation of shear
\begin{equation}
\sigma_{ij}=\partial_{(i}v_{j)} -\frac1{p}\delta_{ij}\, \partial_k v^k
\end{equation}
and dissipation of expansion $\partial_i v^i$. This is also a feature of the entropy in the Fluid/Gravity correspondence \cite{Bhattacharyya:2008xc,Dandekar:2017aiv} (see \cite{Emparan:2020vyfinr} for further discussion in these contexts).
\subsection{Entropy at next-to-leading order}\label{subsec:NLOent}
The explanation for these properties of $S_1$ is that it is actually the leading $1/D$ contribution to the black brane entropy. Namely, the entropy density obtained from the event horizon area of the black brane is, up to a total divergence (see appendix~\ref{app:nlo})
\begin{equation}\label{nloent}
s(t,x)=4\pi\bar{m}(t,x)\left( 1+\frac{c_s}{D}\right)+\frac1{D} s_1(t,x)\,,
\end{equation}
where $\bar{m}$ is the energy density including NLO corrections, and $c_s$ is a constant (which we determine in appendix~\ref{app:phys}) that accounts for the fact that, in order to simplify the form of $s_1$, we have subtracted a term $\propto m$ without changing the right-hand side of \eqref{dts1}. The total entropy is
\begin{equation}\label{SS1}
S(t)=4\pi M\left( 1+\frac{c_s}{D}\right) +\frac1{D}S_1(t)\,.
\end{equation}
Eq.~\eqref{dtSetazeta} then gives the production rate of entropy to NLO in the $1/D$ expansion. The point to notice is that, since the LO entropy is proportional to the energy, which is constant to all orders, the time derivative of the entropy at NLO can be computed using only quantities of the LO effective theory. This is what allows us to identify within this theory the quantities $s_1$ and $S_1$ which behave irreversibly.
Observe that we can write \eqref{nloent} as
\begin{equation}\label{nloent2}
s(t,x)=4\pi \left( \bar{m}-\frac1{D} \left( \frac12 m v^2 +\frac1{2m}(\partial m)^2-c_s\right)\rp^{1+1/D}\,,
\end{equation}
which we can understand as follows. The dependence of entropy on mass \eqref{bhentD} for a Schwarzschild black hole at finite $D$ is $S\propto M^{1+1/(D-3)}$, which is like \eqref{nloent} at large $D$. The term $-\frac12 m v^2$ is a kinetic energy\footnote{Physical velocities in the effective theory are rescaled by a factor $1/\sqrt{D}$ \cite{Emparan:2015gva}, which explains why the term is $1/D$ suppressed.}. It appears here because, out of the total energy of the black hole, only its rest (irreducible) mass contributes to entropy. Observe that this motion could be linear, as in a boosted black hole, or circular, as in a rotating black hole: both reduce the `heat' fraction of the total energy. The last term in \eqref{nloent2} is (likely) a correction from curvature of the horizon due to the difference between the radial position measured by $m$ and the actual area density.
The entropy density \eqref{s1} simplifies for stationary solutions which rotate rigidly, such that \cite{Emparan:2016sjk}
\begin{equation}
\partial_t m +v^i\partial_i m=0\,,\qquad \partial_t v^i=0\,,\qquad \partial_{(i}v_{j)}=0\,.
\end{equation}
In this case the effective equations reduce to the `soap bubble equation'
\begin{equation}\label{soapbubble}
\frac12 m v^i v_i+m\ln m+\partial_i\partial^i m-\frac1{2m}\partial_i m\,\partial^i m=c\, m\,,
\end{equation}
where $c$ is an integration constant that corresponds to a choice of scale for the total mass. We set it to zero, since in the end we will work with scale-invariant quantities where $c$ would disappear. Using this equation we obtain, after dropping a boundary term from a total derivative,
\begin{equation}
S_1(t)=-4\pi\int d^p x\, m v_i v^i\,.
\end{equation}
For a solution that rotates along independent angles $\phi_a$ with velocities $v^{\phi_a}=\Omega^a$, this gives
\begin{equation}
T S_1= -\Omega^a J_a\,,
\end{equation}
where $T$ is the LO temperature \eqref{sTneut}. It is easy to verify that, when added to the LO entropy, this reproduces the Smarr relation for black holes at NLO in $1/D$.
\subsection{Measuring the entropy}
When we compare the entropy of different solutions---{e.g.,\ } ingoing and outgoing black holes---we will do it between configurations with the same total mass. For this purpose, if $\mathbf{S}$ and $\mathbf{M}$ are the physical entropy and mass of the black hole in $D$ dimensions, one works with a mass-normalized, scale-invariant, dimensionless entropy of the form
\begin{equation}\label{mcS}
\mc{S}=C \frac{\mathbf{S}}{\mathbf{M}^{\frac{D-2}{D-3}}}\,.
\end{equation}
Here
\begin{equation}
C =\left( \frac{(D-2)\Omega_{D-2}}{16\pi G}\right) ^{1/(D-3)}\frac{D-2}{4\pi}
\end{equation}
is a suitable convention to simplify later expressions; it could be set to one by adequately choosing Newton's constant $G$.
Similarly, in the effective theory we define a mass-normalized, scale-invariant entropy,
\begin{equation}\label{calS1}
\mc{S}_1(t)=\frac{S_1(t)}{4\pi M}-\ln \frac{M}{2\pi e^2}\,.
\end{equation}
Subtraction of the term $\ln M$ makes this quantity independent of the choice of mass scale, in particular of the value of $c$ in \eqref{soapbubble}. We have also added a constant $\ln(2\pi e^2)$ to simplify later expressions. One can then verify (see appendix~\ref{app:phys}) that the physical mass-normalized entropy \eqref{mcS} is given in terms of the effective theory one \eqref{calS1} by
\begin{equation}\label{physSeftS1}
\mc{S}= 1+ \frac1{D}\mc{S}_1(t)+\ord{\frac1{D^2}}\,.
\end{equation}
\section{General features of black hole collisions}\label{sec:bhcoll}
In the following we restrict to black holes with rotation on a single plane, which will also be the plane on which the black holes move and collide. Then, in the effective theory we study configurations with non-trivial dependence in only $2+1$ dimensions, {i.e.,\ } on a 2-brane.
\subsection{Brane blobology
The effective equations have stationary solutions that describe localized black holes rotating with angular velocity $\Omega$ \cite{Andrade:2018nsz} such that
\begin{equation}
v_i=\Omega\, \varepsilon_{ij}x^j\,.
\end{equation}
The Myers-Perry (MP) black holes are given by
\begin{equation}\label{MPblob}
m(x_1,x_2)=m_0\exp \left(-\frac{x_i x^i}{2(1+a^2)}\right)\,,\qquad \Omega =\frac{a}{1+a^2}\,.
\end{equation}
The mass is
\begin{equation}
M=2\pi m_0 (1+a^2)\,.
\end{equation}
Here $m_0$ measures the horizon radius at the rotation axis, and with our choice of $c=0$ in \eqref{soapbubble} it is
\begin{equation}\label{m0C}
m_0=\exp \left( \frac2{1+a^2}\right)\,.
\end{equation}
While $M$ depends on the choice of scale of normalization,
more relevant are scale-invariant quantities, namely the spin per unit mass
\begin{equation}\label{JMa}
\frac{J}{M}=2a
\end{equation}
and the NLO mass-normalized entropy
\begin{equation}\label{calS1MP}
\mc{S}_1= -\ln (1+a^2)\,.
\end{equation}
The fact that $\mc{S}_1$ becomes more negative with larger $a$ is the familiar decrease of the black hole entropy as the spin grows, in the large $D$ limit.
Another exact solution describes rotating black bars,
\begin{equation}
m= \exp\left( 1-\frac{x_c^2}{4}\left( 1+\sqrt{1-4\Omega^2} \right)-\frac{y_c^2}{4}\left( 1-\sqrt{1-4\Omega^2} \right) \right)\,,
\end{equation}
where $x_c$, $y_c$ are corotating coordinates given by
\begin{align}
x_c&=x^{1} \cos \Omega t+x^{2} \sin \Omega t\, , \nonumber\\
y_c&=x^{2} \cos \Omega t-x^{2} \sin \Omega t\,,
\end{align}
and our scale normalization is again consistent with $c=0$ in \eqref{soapbubble}. This solution has
\begin{equation}
M=\frac{2 \pi} {\Omega}\,,
\end{equation}
and
\begin{equation}
\frac{J}{M}=\Omega^{-1}\,,\qquad
\mc{S}_1=\ln \Omega\,.
\end{equation}
For $\Omega=1/2$ this branch of solutions joins the MP family.
Ref.~\cite{Licht:2020odx} constructed numerically large classes of other stationary solutions. The most relevant for us are rotating dumbbells. They can be regarded as black bars with a pinch in their middle, see fig.~\ref{fig:critdumb}.
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \linewidth]{fig/CriticalDumbbell}
\caption{\small Profile along the long axis of the `critical' dumbbell solution with $J/M$ equal to \eqref{JMc}. \label{fig:critdumb}}
\end{figure}
\subsection{Phase diagrams and the outcomes of collisions}\label{subsec:phases}
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\textwidth]{fig/JOmegaDiagram}
\end{center}
\caption{\small Phases of blobs and their stability as relevant to outcomes of mergers. Solid/dashed lines are stable/unstable stationary blobs. Blue: Myers-Perry black holes, stable up to $J/M=2$. Black: black bars, stable up to $J/M=4/\sqrt{3}\approx 2.31$. Red: black dumbbells, stable up to $J/M=(J/M)_c\approx 2.66$ (dumbbells along the dashed line are unstable binaries of blobs). The background shading indicates the expected outcome of a merger for an initial value of $J/M$. No stable stationary blobs exist for $J/M>(J/M)_c$ (yellow), so, if a merger occurs in this region, it can only evolve to a multi-blob state. The numerical solutions for dumbbells are from \cite{Licht:2020odx}. The same color coding is used in the next figures. \label{fig:JOmegaDiagram}
}
\end{figure}
Fig.~\ref{fig:JOmegaDiagram} is a phase diagram that summarizes the main properties of these solutions, and the implications for the possible initial and, more importantly, final states of a collision. Depending on the value of $J/M$ (distinguished by band-colouring in the figure) the stable phases of stationary single blobs are:
{\renewcommand{\arraystretch}{1.3}
\begin{center}
\begin{tabular}{rl}
$0 \leq J/M < 2$:& MP black holes\\
$2 \leq J/M < \frac{4}{\sqrt{3}}\approx 2.31$:& Black bars \\
$\frac{4}{\sqrt{3}} \leq J/M < (J/M)_c$:& Black dumbbells \\
$(J/M)_c \leq J/M$:& No stable single black hole
\end{tabular}
\end{center}}
\noindent where the numerically determined upper limit $(J/M)_c$ for the existence of stable phases is \eqref{JMc}.
Let us clarify an aspect of the stability of phases in this diagram that was not discussed in \cite{Licht:2020odx}. In that article a second branch of dumbbells (lower in $\Omega$, shown dashed in fig.~\ref{fig:JOmegaDiagram}) was found to exist, starting from $J/M=0$ until it joins the first, upper branch at $(J/M)_c$. In this second branch, the dumbbells are more like slowly rotating black hole binaries, consisting of two gaussian blobs joined by a thin, long tube between them. All these solutions have the same LO entropy, but the NLO entropy $\mc{S}_1$ \eqref{calS1} distinguishes between them.
In fig.~\ref{fig:JEntropyDiagram} we show that lower-branch dumbbells have less entropy than the upper branch. They are therefore thermodynamically unstable. Moreover, a Poincar\'e turning point argument tells us they must have one more negative mode than the upper-branch, and hence be dynamically unstable. This is indeed consistent with two other observations: (i) in our numerical collisions, we never observe a lower-branch dumbbell forming (while upper-branch dumbbells do form); (ii) stationary Keplerian binaries in $D\geq 6$ exist but are unstable.
\begin{figure}[h!]
\centering
\includegraphics[width=0.6 \linewidth]{fig/JEntropyDiagram.pdf}
\caption{\small Phase diagram depicting the entropy $\mathcal{S}_1$ of the different configurations close to the first black bar zero-mode, as a function of angular momentum. For $J/M> 2$ the black bar (black) is entropically favored over the MP black hole (blue). At the zero-mode $J/M = 4/\sqrt{3}$, a branch of stable dumbbells (red) appears with $\mathcal{S}_\text{inv}$ slightly higher that that of unperturbed black bars. This phase dominates entropically up to the turning point at $(J/M)_c\approx 2.66$, where stable dumbbells cease to exist, and the system typically evolves to a fission. \label{fig:JEntropyDiagram}}
\end{figure}
In contrast, upper-branch dumbbells resemble (segments of) stable non-uniform black strings (fig.~\ref{fig:critdumb}). Although other more non-uniform phases were found in \cite{Licht:2020odx}, by generic turning-point/bifurcation arguments they are expected to have more negative modes and hence be dynamically unstable. Therefore, no other stable solutions are expected to exist besides those shown in fig.~\ref{fig:JOmegaDiagram}.
Fig.~\ref{fig:JOmegaDiagram} is then a major guide to predicting the outcome of a collision between two blobs, based only on the total angular momentum $J$ and total mass $M$ of the system, which are conserved. If $J/M<2.66$, two blobs that merge can relax into a stable single blob, namely the only one that is stable for the corresponding value of $J/M$. Bear in mind that they will not necessarily do so, since the dynamical evolution may avoid that endpoint.
If $J/M>(J/M)_c$ the final state must consist of more than a single blob. That is, if there is fusion it will be followed by fission.
If $J/M$ is less than $\approx 4$, we observe that the end state is always two outgoing black holes, {i.e.,\ } $2\to 1\to 2$. At higher $J/M$, a third, smaller black hole can appear between them, the apparent reason being that at these angular momenta there exists an unstable branch of three-bumped bars. End states with more than two blobs are entropically disfavored but nevertheless are dynamically possible. In particular, in collisions at large $J$ with large initial orbital angular momentum the evolution can exhibit complicated patterns. Their investigation would take us beyond the scope of this article.
\subsection{$2\to 1$ vs.\ $2\to 1\to N$}\label{subsec:2to}
In the next sections we will study collisions between two initial MP black holes, of equal mass and equal initial spins, with initial rotation parameter $a_{\text{in}}$ within the stability range of MP black holes $0\leq a_{\text{in}}<1$.\footnote{We do not consider initial black bars and dumbbells. They are expected to exist at finite $D\geq 6$ but not be completely stable, not even stationary, since they must radiate gravitational waves.} Their initial velocities will be $\pm u_{\text{in}}$ and the impact parameter $b_{\text{in}}$. We select the collisions where there is fusion; the cases where the two black holes scatter without ever merging may also be of interest but their physics is different than we intend to explore here and we will barely discuss them\footnote{Since the effective theory describes a continuous horizon, the distinction is not perfectly clear-cut and involves discretionary choices. However, the more ambiguous cases are only marginal to our analysis.}.
For some values of the initial parameters $(a_{\text{in}}, u_{\text{in}}, b_{\text{in}})$ the system fissions.
\begin{figure}[t]
\centering
\includegraphics[width=0.4 \linewidth]{fig/CCVBoundaryNoSpin.pdf}
\qquad
\includegraphics[width=0.4 \linewidth]{fig/CCVBoundarySpinning.pdf}
\caption{\small Outcome of symmetric collisions of two black holes with initial spin, velocity, and impact parameter $(a_{\text{in}}, u_{\text{in}}, b_{\text{in}})$. The dots (joined by dashed lines) separate between initial conditions that lead to $2\to 1$ fusion events (below the dots) and $2\to 1\to N$ cosmic-censorship-violating fission events (above the dots). The colors distinguish between the stable phases available (same color coding as in fig.~\ref{fig:JOmegaDiagram}). For small enough $b_{\text{in}}$ the system always settles down into the available stable single blob, but for very large $b_{\text{in}}$ the dynamical evolution passes too far from the stable phase and proceeds to fission. \label{fig:CCVBoundary}}
\end{figure}
Figure \ref{fig:CCVBoundary} shows the numerically determined boundary (dots joined by dashed lines) between initial conditions that lead to a $2\to 1$ collision, leaving a rotating central object, and initial conditions that produce more than one outgoing object, usually two but possibly more.
As already noted in previous papers \cite{Andrade:2018yqu,Andrade:2019edf}, the boundary follows a curve of constant $J/M = (J/M)_c$, as long as the impact parameter is below some threshold (which depends on $a_{\text{in}}$). This means that for these values of $b_{\text{in}}$ the merger always settles down into the unique stable solution that is available. However, if $b_{\text{in}}$ is large enough, this possible end state is avoided: the two colliding black holes form a horizon that is too elongated to find its path to the stable blob. In these cases, the colliding black holes attract each other deflecting their trajectories,\footnote{Despite the absence of stable Keplerian orbits in $D\geq 5$, we find that the two blobs can perform more than one revolution around each other before either flying apart or merging. In this respect, these collisions resemble four-dimensional ones more than one might have expected.} and then they either fly apart or suddenly fall onto each other to form a stable central object. It is suggestive that this $2\to 2$ scattering might be understood as the formation of an unstable, long dumbbell (two gaussian blobs joined by a long tube), which then either collapses or breaks apart.
\subsection{Kinematics, entropy and geometry of the collisions}\label{subsec:kinent}
Let us now be more specific about the collisions we study.
The two initial black holes are MP blobs on a 2-brane like \eqref{MPblob} that start in the $(x,y)$ plane at
\begin{equation}
(x,y)=\left(\pm \infty,\pm \frac{b_{\text{in}}}2\right)\,,
\end{equation}
with velocities
\begin{equation}
(v_x,v_y)=(\mp u_{\text{in}},0)\,.
\end{equation}
The latter are achieved by applying a Galilean boost to each blob. The entropy of each individual boosted black hole, normalized by its own mass, is
\begin{equation}\label{S1MP}
\mc{S}_1=-\frac12 u^2 -\ln (1+a^2)
\end{equation}
(see eq.~\eqref{S1boost}). The presence of the term $\propto u^2$ is due to the fact, already mentioned, that the kinetic energy must be subtracted from the total energy of the black hole, since only the rest (irreducible) mass contributes to entropy. It is directly related to the fact that the horizon area of a black hole does not change through Lorentz contraction \cite{Horowitz:1997fr}, which is also a property of entropy in the effective theory, as proved in appendix~\ref{app:boost}.
The entropy of a system of two equal MP black holes, now normalized by their combined mass $M$, is
\begin{equation}\label{2bhS1}
\mc{S}_1=-\frac12 u^2 -\ln 2(1+a^2)\,.
\end{equation}
and, using \eqref{JMa}, their total angular momentum is\footnote{Each black hole has spin $2(M/2)a$ and orbital angular momentum $(M/2) u (b/2)$.}
\begin{equation}\label{angmom2bh}
\frac{J}{M}=2 a+\frac{b u}2\,.
\end{equation}
The conservation
of mass-energy (which includes kinetic energy to NLO) and angular momentum imposes restrictions on the possible outgoing final states, and on how much entropy can be produced.
The analysis can be made entirely within the large $D$ effective theory, but since we are only considering initial and final states that are MP black holes, we could also consider the properties of the known solutions exactly in $D$. That is, we could work at finite $D$ using physical magnitudes, expand in $1/D$ and translate into effective theory magnitudes. The two methods of calculation are easily seen to agree.\footnote{Similar considerations were made in \cite{Emparan:2003sy}.}
\paragraph{$2\to 1$: fusion.} If the black holes fuse and then relax into a single, stable blob, this final state can be read from fig.~\ref{fig:JOmegaDiagram} as the unique stable solution with the value
\begin{equation}\label{angmomfus}
\frac{J}{M}=2 a_{\text{in}}+\frac{b_{\text{in}} u_{\text{in}}}2\,.
\end{equation}
If this end state is an MP black hole or a black bar, the final entropy will be
\begin{eqnarray}
\mc{S}_1^{\rm(MP)}&=&-\ln\left( 1+\frac14 \left(\frac{J}{M}\right)^2\right)^2\,,\\
\mc{S}_1^{\rm(bar)}&=&-\ln\left(\frac{J}{M}\right)\,.
\end{eqnarray}
The total entropy production in the fusion will be the difference between these and \eqref{2bhS1}.
We do not have analytical expressions for the entropy of black dumbbells.
\paragraph{$2\to 1\to 2$: fusion $\Rightarrow$ fission.} The final states of the $2\to 1\to 2$ collision are always two equal MP black holes\footnote{In principle, these could also be stable bars and dumbbells. However, the spin of the outgoing states that we observe is always below the range of their existence.} with outgoing parameters $0\leq a_{\text{out}}<1$ and
\begin{equation}\label{xyout}
(x,y)=R(\theta)\left(\pm \infty,\pm \frac{b_{\text{out}}}2\right)\,,
\end{equation}
\begin{equation}\label{vout}
(v_x,v_y)=R(\theta)(\pm u_{\text{out}},0)\,,
\end{equation}
where $R(\theta)$ is a rotation matrix with angle $\theta$.
Conservation of mass and angular momentum implies
\begin{equation}\label{angmomcons}
\frac{J}{M}=2 a_{\text{in}}+\frac{b_{\text{in}} u_{\text{in}}}2=2 a_{\text{out}}+ \frac{b_{\text{out}} u_{\text{out}}}2\,.
\end{equation}
The collision is characterized by seven parameters: $(a,u,b)_\text{in/out}$ plus the scattering angle $\theta$. The latter is not affected by conservation laws and it does not enter into entropic arguments, so we will leave it aside in the following discussion.
Of the six parameters $(a,u,b)_\text{in/out}$, only five are independent once \eqref{angmomcons} is imposed. We can regard the three initial parameters as given, and then two outgoing parameters, say, $u_{\text{out}}$ and $a_{\text{out}}$, are unconstrained by conservation laws, that is they will be determined by the dynamical evolution of the system.
The difference in the entropy between the initial and final states is
\begin{equation}\label{deltas}
\Delta \mc{S}_1=\frac{u_{\text{in}}^2-u_{\text{out}}^2}{2}+\ln \frac{1+a_{\text{in}}^2}{1+a_{\text{out}}^2}\,.
\end{equation}
The entropy of the final state will be larger if the outgoing velocities and spins are as small as possible, since both $a$ and $u$ tend to reduce the entropy of a MP black hole with fixed mass. However, they cannot be made arbitrarily small. The total angular momentum must be conserved, and even though \eqref{angmomcons} seems to allow for two unconstrained outgoing parameters, we cannot expect to have $a_{\text{out}},u_{\text{out}}\to 0$. For any $J\neq 0$ this would require that the outgoing impact parameter diverges, $b_{\text{out}}\to \infty$, which is unreasonable: if the two initial black holes do indeed collide and merge, the outgoing impact parameter will be comparable to the size of an intermediate (unstable) state with the same $J$ and $M$. From the distance between the two peaks in the critical dumbbell, fig.~\ref{fig:critdumb}, we can expect that
\begin{equation}\label{bdumb}
b_\mathrm{out}\approx 7\,,
\end{equation}
and probably a little larger after the fission.
This is indeed a good predictor for the actual values we find below. Another well-motivated and better defined geometric estimate is obtained by demanding that $b_\mathrm{out}$ is approximately equal to twice the radius of the two outgoing black holes. In appendix~\ref{app:bout} we find that this gives
\begin{equation}\label{boutguess}
b_\mathrm{out}\approx 2\sqrt{2(1+a_\mathrm{out}^2)\ln\epsilon_b^{-1}}\,.
\end{equation}
It depends on a small number that we estimate to be $\epsilon_b\approx 10^{-3}$. If entropy is to be maximized, then $b$ will be close to this upper bound.
Eqs.~\eqref{angmomcons} and \eqref{boutguess} leave one unconstrained degree of freedom: an equation between $u_{\text{out}}$ and $a_{\text{out}}$. The last constraint that fixes them will be discussed in sec.~\ref{sec:bhscat}.
\section{Entropy production}\label{sec:entpro}
With our methods we can easily track the entropy production, in space and in time, during the evolution of three different kinds of phenomena:
\begin{itemize}
\item $1\to N$: decay and fission of unstable black holes
\item $2\to 1$: fusion of two black holes
\item $2\to 1\to 2$: fusion of two black holes followed by fission
\end{itemize}
Understanding entropy production in the first two will give us insight into the third.
We evolve the equations numerically, using two different codes (the same as in \cite{Andrade:2018yqu,Andrade:2019edf}, now using finite differences instead of FFT differentiation), until the system either settles into a stable single blob, or breaks up into blobs that fly apart. By keeping track of $m(t,\mathbf{x})$ and $v^i(t,\mathbf{x})$ we can then compute all physical magnitudes.
\subsection{$1\to N$: decay and fission of unstable black strings and black holes}
Here we follow the non-linear evolution of the decay of an unstable, fissile blob. We have chosen three important examples which most clearly exhibit the physics relevant for other more complex evolutions, see fig.~\ref{fig:decays}: the black string; an ultraspinning MP black hole with $a=2$ decaying through an intermediate black bar; and an MP black hole with $a=3$ decaying through an intermediate black ring. The latter evolutions are triggered by choosing different inital perturbations, which excite different unstable modes of the ultraspinning black hole.
\begin{figure}[h!]
\centering
\includegraphics[width=.95\linewidth]{fig/EntropyFramesString}
\\
\medskip
\includegraphics[width=.95\linewidth]{fig/EntropyFramesMyersPerry}
\\
\medskip
\includegraphics[width=.95\linewidth]{fig/EntropyFramesRing}
\medskip
\caption{\small Entropy production, as a function of time and space, during the decay of: unstable black string (top); ultraspinning MP black hole through intermediate black bar (middle); ultraspinning MP black hole through intermediate black ring (bottom). The blue curves give the production rate of the NLO entropy, $\partial_t \mc{S}_1$ (integrated in space); the dashed red curves are the production rate through dissipation of shear (and not of expansion), $\partial_t \mc{S}_1^{\rm{(shear)}}$. The density plots show the time derivative of the entropy density (colors in log scale). The thin black contours serve to guide the eye to where the black hole blobs are, and correspond to $m(x,y) = 0.001 M$. We can see that the pattern of entropy production in the black string decay is reproduced in the second peak of the decays of the MP black hole. The first peak is mostly due to shearing when the intermediate black bar or black ring forms.
\label{fig:decays}}
\end{figure}
Let us emphasize that our simulations of the decay of unstable black strings are not expected to reproduce the details of the late-time evolution in the (much more expensive) numerical evolutions in \cite{Lehner:2010pn}, nor of the related and more complex simulations in \cite{Figueras:2015hkb,Figueras:2017zwa,Bantilan:2019bvf}. This has been discussed in detail in \cite{Emparan:2018bmi}. In particular, the large $D$ effective theory does not reveal the cascading formation of small `satellites'. However, our concern here is with how entropy is produced in this decay, and this appears to occur mostly in the intermediate stages of the evolution. At late times, most of the mass and area reside in the large, black-hole-like blobs, and little on the satellites and thin tubes inbetween them. Therefore, we expect that our study accounts for the main contributions to entropy production.
The analysis of black string decay reveals generic aspects of entropy production in fission. The pattern we see in fig.~\ref{fig:decays} (top) will be present in all subsequent fission phenomena: a single peak in the entropy production rate, midway along the fission, with dissipation equally shared into shear and expansion.
The two decays of the ultraspinning MP black hole in fig.~\ref{fig:decays} show qualitative similarities between themselves: first, a long-lived but ultimately unstable configuration forms---a black bar, or a black ring.\footnote{When $D$ is not large enough this bar radiates away its excess spin fast enough to return to stability \cite{Shibata:2010wz}.}
Entropy is generated mostly through shear dissipation. In fact, it is clear that a dominant shearing motion must be driving the evolution to a bar, while the formation of the ring should also involve some compression. Both features are visible in the entropy production curves. Afterwards, this intermediate state decays following the pattern of black string fission.
The second peak in entropy production appears to have universal features. This confirms that the physics of black string decay also controls the fission of the blob. Observe, however, that in the MP decay the peak is a little higher---hence more irreversible---than in the black string. This could be expected since the latter is a more symmetric configuration.
Finally, we see that the duration of the string break up is on the order of
\begin{equation}\label{tfission}
(\Delta t)_\textrm{fission} \approx 20 M\,.
\end{equation}
This will be a characteristic of other fissions.
\subsection{$2\to 1$: fusion $\Rightarrow$ thermalization}
In a $2\to 1$ collision the final state is completely determined by the conserved initial value of $J/M$: the system settles into the only stable stationary black hole with that value: an MP black hole, a black bar, or a black dumbbell.
In fig.~\ref{fig:fusion} we present an illustrative example: a symmetric collision of two black holes with $(b_\text{in},u_\text{in},a_\text{in}) = (2, 1.0, 0.6)$, so $J/M=2.2$, resulting in the formation of a stable black bar.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{fig/EntropyFramesMerger}
\caption{\small Entropy production during a collision with fusion into a stable black bar. The first, large fusion peak is followed by a smaller peak for the thermalization to the black bar. The height of this second peak is comparable to that in fig.~\ref{fig:decays} (middle). \label{fig:fusion}}
\end{figure}
The figure shows that when the two black holes first meet and fuse there is a large production of entropy. There follows a phase in which the system equilibrates (thermalizes) into the final stable black bar. This second phase is similar to the formation of the (unstable) bar in the decay of the ultraspinning MP black hole in fig.~\ref{fig:decays} (middle), with both peaks having similar height ($\approx 0.03$, in mass-normalized entropy rate). The duration of this phase in the decay of the MP black hole is much longer, since the system there starts in stationary, but unstable, equilibrium, while the merged horizon is farther from equilibrium.
The contributions to dissipation from shear and expansion vary at different stages in the evolution. During the initial fusion phase, one or the other may dominate depending on the initial parameters, but generically we observe both expansion (and compression) and shearing motion of the blob, which contribute roughly equally to entropy production. During the formation of the intermediate quasi-thermalized bar, the proportions of shear and expansion can vary, depending on how much the preceding blob is already bar-like or not. In the simulation shown in fig.~\ref{fig:fusion}, from $t\approx 9$ to $1$ the blob has to undergo less shearing to acquire the bar shape than in fig.~\ref{fig:decays} (middle) from $t\approx 20$ to $40$. Presumably this explains the lower presence of shear dissipation.
\subsection{$2\to 1\to 2$: fusion $\Rightarrow$ quasi-thermalization $\Rightarrow$ fission}
With large enough total angular momentum, the fusion results in a fissile intermediate state.
\paragraph{Stages in the evolution.}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{fig/EntropyFramesCollision}
\caption{\small Entropy production during a collision with fusion followed by fission. The entire process can be divided in three stages: $t\approx 4$ to $8$, fusion; $t\approx 9$ to $14$, quasi-thermalization into a bar; $t\approx 15$ to $30$, fission of the bar, similar to black string decay. \label{fig:EntropyFrames}}
\end{figure}
In fig.~\ref{fig:EntropyFrames} we present the evolution of entropy production in a collision with initial parameters $(b_\text{in},u_\text{in},a_\text{in}) = (3.4, 1.0, 0.8)$, so $J/M=3.3$. It can be interpreted by combining what we have learned so far:
\begin{enumerate}
\item \emph{Fusion:} $t\approx 4$ to $8$. This is a strongly irreversible phase, very similar to the first peak in $2\to 1$ fusion, fig.~\ref{fig:fusion}.
\item \emph{Quasi-thermalization:} $t\approx 9$ to $14$. The fused blob follows qualitatively the evolution of unstable MP black holes in fig.~\ref{fig:decays}~(middle): a quasi-thermalization phase with the (faster) formation of a long-lived bar.
\item \emph{Fission:} $t\approx 15$ to $30$. The intermediate bar fissions into two outgoing black holes, in a manner similar to fig.~\ref{fig:decays}~(middle), ultimately patterned after the decay of black strings, fig.~\ref{fig:decays}~(top). It lasts for a time comparable to \eqref{tfission}.
\end{enumerate}
Observe that not only the qualitative features of the decay of the MP bhs are reproduced in the $2\to 1\to 2$ collision: also the height of the peaks in the mass-normalized entropy production rates, after the fusion peak in fig.~\ref{fig:EntropyFrames}, are quantitatively similar.
The proportions of viscous dissipation from shear and expansion also follow what we have seen before.
\section{Scattering of black holes and entropic attractors}
\label{sec:bhscat}
We now turn to a more complete investigation of collisions, in particular those that result in fission, and the role that the entropy increase plays in them. For this purpose
we have performed an extensive, although not exhaustive, study of symmetric collisions of black holes for wide ranges of the initial parameters $(a_{\text{in}}, u_{\text{in}}, b_{\text{in}})$.
In our simulations we verify that the final blobs can be identified with known stable stationary blobs. In $2\to 1$ and $2\to 1\to 2$ events, the final spin parameter, $a_{\text{out}}$, is extracted from the width of the gaussian blobs by linear regression of $\ln m$ as a function of $r^2$, where $r$ is the distance to the center of the blob (see \eqref{MPblob}). In fission, we extract the parameters
$u_{\text{out}}$ and $b_{\text{out}}$ of the outgoing blobs from the velocity field at their centers,
As we have seen, fusion into a single stable black hole is fully determined by the conservation of mass and angular momentum. In our simulations we have been able to verify that the integration over space and time of the entropy density reproduces correctly the exact predictions from sec.~\ref{subsec:kinent}. This is a good check on the accuracy of our methods.
\subsection{$2\to 1\to 2$: In \& Out}\label{subsec:innout}
In contrast to $2\to 1$ fusion, here there is a continuous two-dimensional range of out states that are allowed by the conservation laws.
\begin{figure}[t]
\centering
\includegraphics[width= \linewidth]{fig/innout.png}
\caption{\small $2\to 1\to 2$ collisions: in 100 simulations, the outgoing parameters $(a_{\text{out}}, u_{\text{out}}, b_{\text{out}})$ show little correlation with initial ingoing parameters $(a_{\text{in}}, u_{\text{in}}, b_{\text{in}})$. This is a consequence of quasi-thermalization in an intermediate stage in the collision. Outgoing parameters cluster in a relatively narrow range, the more so the lower the total $J/M$, for which the intermediate phase lasts longer.
\label{fig:innout}}
\end{figure}
In figs.~\ref{fig:innout} we present the results for the outgoing parameters $(a_{\text{out}}, u_{\text{out}}, b_{\text{out}})$ of 100 simulations with randomly chosen values of the ingoing parameters $(a_{\text{in}}, u_{\text{in}}, b_{\text{in}})$.\footnote{We discard events where the intermediate interaction is too weak to involve fusion. We implement this by removing from our analysis events for which the initial impact parameter is large ($b_{\text{in}} >7$)
and the outgoing parameters change less than $5\%$ relative to the initial ones. In these cases the spin $a$ changes very little, while $u$ and $b$ can change more, as expected if the process is one of direct (fusionless) $2\to 2$ scattering.} We also present, as color shading, the value of the the conserved $J/M$ for each event. Since our sampling is not exhaustive, we have not attempted to perform detailed statistical analyses, but nevertheless there are several discernible patterns in these plots that are worth remarking on.
First, there is a clear clustering of the out parameters. It is stronger the lower $J/M$ is, with $(u_{\text{out}}, a_{\text{out}}, b_{\text{out}})$ being essentially unique for the lowest $J/M$ (slightly above $(J/M)_c=2.66$). The latter are the cases where an unstable but very long-lived intermediate state forms in the collision. The dissipation that happens in this intermediate phase effectively erases the memory of the initial state parameters, other than the conserved $J/M$. As $J/M$ grows larger, the intermediate state is less long-lived and less precisely defined, and the system retains more memory of the initial configuration (for instance, there is some correlation between the values of $a_{\text{in}}$ and $a_{\text{out}}$), resulting in more dispersion in the plots.
More generally, the plots show that the out parameters lie approximately in the following ranges:\footnote{These central values and variances are indicative and should not be taken literally; the dispersion is strongly correlated with $J/M$, and it is very low near the critical value.}
Spin:
\begin{equation}\label{arange}
a_{\text{out}} \approx 0.3\,\substack{+0.2 \\ -0.1}\,,
\end{equation}
with the upper bound being fairly robust.
Velocity:
\begin{equation}\label{vrange}
u_{\text{out}}\approx 0.6\,\substack{+0.4 \\ -0.1}\,,
\end{equation}
with a strong bias towards the lower value, which is never below $\approx 0.5$.
Impact parameter:
\begin{equation}\label{brange}
b_{\text{out}}\approx 8\,\substack{+2 \\ -1}\,.
\end{equation}
The scant correlation of these results with the initial values other than the conserved $J/M$ is indicative of intermediate thermalization.
The clustering values might possibly be compared with those in the decay states of unstable blobs in the yellow regions close to $(J/M)_c$ in fig.~\ref{fig:JOmegaDiagram}. An indication in this direction is \eqref{bdumb}, but we have not attempted to go further since this requires additional extensive numerical studies.
Regarding the scattering angle in the final states \eqref{xyout}, \eqref{vout}, we have observed that $\theta$ can be robustly obtained from the numerical simulations. In particular, it is independent of the scheme implementing a regulator at small $m$, and of the values of the regulator as this is decreased. We expect, therefore, that in collisions at finite $D$ this scattering angle can also be obtained from the classical evolution before the naked singularity forms. However, this angle does not play any role in our study in this paper.
\subsection{Entropy increase}\label{subsec:totent}
Recall that the initial and final states are characterized by the values of $(a,u,b)$, one of which can be traded for the value of $J$, which is common for the initial and final states (we always set the total mass $M=1$). We find convenient to eliminate $b$, so in fig.~\ref{fig:NeutralCollisionEntropy}, for a given value of $J$, we represent the initial and final states each one as a point (red and green, respectively) in the plane $(u,a)$. On this plane, we also show colour contours for the value of the mass-normalized entropy $\mc{S}_1$. We exclude the region of ultraspins $a>1$ since these MP black holes are unstable.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\linewidth]{fig/NeutralCollisionEntropy1.pdf}
\includegraphics[width=0.9\linewidth]{fig/NeutralCollisionEntropy2.pdf}
\caption{\small Total entropy growth in symmetric $2\to 1\to 2$ collisions. \emph{Left}: coloured-contour plot for the total entropy of configurations with velocity and spin $(u,a)$, with a given value of $J/M$. Red and green dots indicate the initial and actual final states in the dynamical evolution. The hashed part marks final states forbidden by the second law. The purple band are states with impact parameter in the geometric range \eqref{boutguess}. We see that the actual final states lie very close to (although not exactly at) the maximum entropy with $b$ in this range. In particular, larger values of $a_\text{out}$ and smaller of $u_\text{out}$ would be entropically very disfavoured. \emph{Right}: evolution in time of $\mc{S}_1$ along the simulation. The initial and final parameters are (top) $(b,u,a)_\text{in} = (3.4, 1.0, 0.8)$, $(b,u,a)_\text{out} = (7.9, 0.65, 0.36)$ and (bottom) $(b,u,a)_\text{in} = (3.0, 2.0, 0.0)$, $(b,u,a)_\text{out} = (9.25, 0.5, 0.34)$. \label{fig:NeutralCollisionEntropy}}
\end{figure}
We already mentioned that the entropy cannot be fully maximized, since this happens at $(u,a,b)=(0,0,\infty)$. The geometric constraint \eqref{boutguess} (particularly good for low $J/M$) selects a set of possible final states, which we mark as a purple band in the $(u,a)$ plane.
In fig.~\ref{fig:NeutralCollisionEntropy} the entropy changes between initial and final states are shown in two illustrative cases.\footnote{The top right curve for $\mc{S}_1(t)$ is the integral of the blue curve in fig.~\ref{fig:EntropyFrames}.}
The salient aspects of these plots are:
\begin{itemize}
\item Final states in the hashed region are excluded by the second law. High final velocities are then excluded. In particular, if the initial black holes are spinless, then the outgoing velocity cannot be higher than the ingoing one.
\item The entropy increases significantly, an in particular it is close to being maximized (but not fully maximized) among the possible outgoing states with geometrically-constrained impact parameter \eqref{boutguess}.\footnote{All purple bands are centered at $\epsilon_b = 0.001$, as defined in \eqref{meps}. The bands in figs.\ref{fig:attract} and \ref{fig:NeutralCollisionEntropy}~(up) span the range $\ln(\epsilon_b) = \ln(0.001) \pm 0.5$. In figs.~\ref{fig:NeutralCollisionEntropy}~(down) and \ref{fig:hiJattract} the range is wider, $\ln(\epsilon_b) = \ln(0.001) \pm 1.5$.}
\end{itemize}
In fig.~\ref{fig:NeutralCollisionEntropy}~(right) we show the time evolution of the entropy. In the first one (top) the total entropy change is approximately equally subdivided between the sharp production at the beginning of the collision and the slower subsequent production rate. In the second one (bottom), which starts with very high initial velocity, most of the entropy is quickly produced in the initial stages.
\subsection{Entropic attractors}
We can now combine the analyses of sec.~\ref{subsec:innout} and sec.~\ref{subsec:totent} to obtain a global perspective on the role of total entropy production in the evolution of the system.
In figs.~\ref{fig:attract} and \ref{fig:hiJattract} we show the results of sampling a large number of collisions $2\to 1\to 2$ with two specific values of $J/M$: a low value close to $(J/M)_c$ in fig.~\ref{fig:attract}, and a quite higher one in fig.~\ref{fig:hiJattract}. The clustering of the final states seen in sec.~\ref{subsec:innout} is even more clearly visible here, and also the near-maximization of the entropy that we discovered in sec.~\ref{subsec:totent}. The attractor that funnels the evolution is stronger the closer to the critical value of the conserved $J/M$, but fig.~\ref{fig:hiJattract} shows that it, and the near-maximization of the entropy, are present even when $J/M$ is quite far from criticality.
\begin{figure}[t]
\centering
\includegraphics[width=0.6\linewidth]{fig/NeutralClusteringJ3p8}
\caption{\small Entropic attractor in collisions $2\to 1\to 2$ with $J/M=3.8$. See fig.~\ref{fig:attract} for the explanation. The attractor is less strong as $J/M$ grows larger. Note that initial states to the left of the purple band have large initial impact parameters and the black holes do not merge, so we do not include them.
\label{fig:hiJattract}}
\end{figure}
We conclude that the dynamical outcome of the collision can be approximately predicted, after imposing kinematic and geometric constraints, by near-maximization of entropy generation. The maximization is not exact, but this principle is a powerful guide to the end result of a complex dynamical process.
\section{Charge diffusion in black holes}\label{sec:chadif}
Since the entropy of neutral black holes is proportional to their mass in the limit $D\to\infty$, it can only be generated at NLO in the $1/D$ expansion---although, as we have argued, this production can be computed using the LO effective theory.
However, when charge is present, the entropy of a black hole when $D\to\infty$ is no longer proportional to the mass. Instead of \eqref{bhentD}, we have
\begin{eqnarray}\label{SMQ}
S(M,Q)&\propto& \left( M+M\sqrt{1-2\left(\frac{Q}{M}\right)^2}\right)^{\frac{D-2}{D-3}}\nonumber\\
&=& M+M\sqrt{1-2\left(\frac{Q}{M}\right)^2}+\ord{\frac1{D}}\,.
\end{eqnarray}
This is easily seen to imply that in the fusion between two black holes with different charge-to-mass ratios, $Q_1/M_1\neq Q_2/M_2$ (including the charge sign), the charge redistribution that occurs gives rise to entropy production, even when $D\to\infty$. The mechanism that drives it is not viscous dissipation, but Joule heating through charge diffusion.
This gives us the opportunity to explore a different mechanism for entropy production, and also provides a simpler set up where we can confirm the general picture that we have developed in the previous sections. It will be easy, and interesting, to consider asymmetric collisions, where the two initial black holes have different parameters, in particular different charge-to-mass ratios.
\subsection{Entropy generation in charged fusion and fission}
Our discussion will be succint, and for more details we refer to \cite{Emparan:2016sjk} and \cite{Andrade:2018rcx}. The effective theory for a charged black brane has as its variables, besides the mass density and the velocity, the charge density $q(t,\mathbf{x})$. In terms of these, the entropy density is
\begin{equation}
s=2\pi \left( m+\sqrt{m^2-2q^2}\right) \,.
\end{equation}
The chemical potential, conjugate to the charge, and the temperature are
\begin{equation}
\mu=\frac{2q}{m+\sqrt{m^2-2q^2}}\,,\qquad T=\frac1{2\pi}\frac{\sqrt{m^2-2q^2}}{ m+\sqrt{m^2-2q^2}}\,.
\end{equation}
The effective equations then imply that
\begin{equation}
\partial_t s +\partial_i j_s^i=\kappa_q \partial_i \left(\frac{\mu}{T}\right) \partial^i \left(\frac{\mu}{T}\right) \,,
\end{equation}
where the expressions for the entropy current $j_s^i$ and the charge diffusion coefficient $\kappa_q$ can be found in \cite{Emparan:2016sjk}.
The term on the right generates entropy when there is a gradient of
\begin{equation}
\frac{\mu}{T}=4\pi\frac{q/m}{\sqrt{1-2(q/m)^2}}\,,
\end{equation}
that is, when $q/m$ is not homogeneous so there can be charge diffusion. Observe that, unlike in the neutral case, the temperature need not be uniform: it is smaller where $|q/m|$ is larger.
The effective equations admit exact solutions for charged blobs that are easy extensions of the neutral ones, in particular charged rotating black holes and black bars \cite{Andrade:2018rcx}. The former are the large $D$ limit of the Kerr-Newman (KN) black hole.
The entropy of a KN black hole or black bar at large $D$ is (cf.~\eqref{SMQ})
\begin{equation}
S=2\pi M \left( 1+\sqrt{1-2\mathfrak{q}^2}\right)\,,
\end{equation}
where we introduce the charge-to-mass ratio of the black hole,
\begin{equation}\label{fracq}
\mathfrak{q}=\frac{Q}{M}\,.
\end{equation}
Observe that, in this limit of $D\to\infty$, the entropy is independent of the spin. The KN black hole and the charged black bar differ in how the spin is related to the mass and charge, but are entropically equivalent.
Consider now a configuration of two KN black holes, labelled $1$ and $2$, with masses $M_{1,2}$ and charges $Q_{1,2}$. We want to study the total entropy of the system for fixed total mass $M=M_1+M_2=1$ and fixed total charge $Q=Q_1 + Q_2$. For the two remaining free parameters in the system, we use $\Delta M$ and $\Delta \mathfrak{q}$, such that
\begin{equation}
M_{1,2}=\frac12\pm \Delta M\,,
\end{equation}
\begin{equation}
\mathfrak{q}_{1,2} =Q-\left(\Delta M \mp \frac12\right)\Delta\mathfrak{q}\,,
\end{equation}
{i.e.,\ }
\begin{equation}
\mathfrak{q}_1-\mathfrak{q}_2=\Delta\mathfrak{q}
\end{equation}
where
\begin{equation}
\mathfrak{q}_{1,2} =\frac{Q_{1,2}}{M_{1,2}}\,.
\end{equation}
The entropy of the two-black hole system is
\begin{equation}
\frac{S_{(2)}}{2\pi}=\left( \frac12+\Delta M\right)\left( 1+\sqrt{1-2\mathfrak{q}_1^2}\right)+\left( \frac12-\Delta M\right)\left( 1+\sqrt{1-2\mathfrak{q}_2^2}\right)\,.
\end{equation}
We now ask what values of $\Delta M$ and $\Delta \mathfrak{q}$ maximize this entropy for fixed total $Q$ and total $M=1$. The answer is that the maximum is reached for
\begin{equation}
\Delta\mathfrak{q} =0
\end{equation}
for any value of $\Delta M$, and this maximum is equal to
\begin{equation}
\frac{S_{(0)}}{2\pi}=1+\sqrt{1-2Q^2}\,,
\end{equation}
which is the entropy of a single black hole or black bar with mass $M=1$ and charge $Q$ (so $\mathfrak{q}=Q$). That is, a system of two black holes, with possibly different masses and charges but both of them having the same charge-to-mass ratio $\mathfrak{q}$, has the same entropy as a single black hole or black bar with that same value of $\mathfrak{q}$; and a system of two black holes with different charge-to-mass ratios has lower entropy than a single black hole or black bar of the same total mass and charge.
The consequences of this for processes of black hole fusion and fission are then clear:
\begin{itemize}
\item Fission processes where an unstable black hole or black bar decays into two black holes are isentropic (adiabatic), and the final black holes will have the same charge-to-mass ratios $\mathfrak{q}$ as the initial one.
\item When two black holes with the same values of $\mathfrak{q}$ (but possibly different masses and charges) collide and merge, the subsequent process will necessarily be isentropic, regardless of whether a long-lived intermediate state forms or not, and (if there is fission) regardless of what the final outgoing black holes are.
\item When two black holes with different values of $\mathfrak{q}$ collide and merge, entropy is produced. If a stable black hole forms, then it will definitely have more entropy than the initial states. The entropy production will be given by $S_{(0)}-S_{(2)}$ above.
\item If the final state consists of two outgoing black holes, then more entropy will be produced the closer the intermediate state is to a single stationary black hole or black bar ({i.e.,\ } the longer-lived the intermediate state is, so there is time to diffuse charge uniformly to maximize the entropy). In that case, entropy will reach a value close to saturation during the intermediate phase, with little entropy production in the subsequent fission stage.
\item If there is no long-lived, almost stationary intermediate state, then entropy production will be less than maximal, but we still expect that it will happen mostly during the fusion process (where the charge-to-mass-ratios are more different) and less so in the fission.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{fig/EntropyFramesCharged}
\caption{\small Entropy production during a charged collision with fusion into a stable black bar. At LO in the effective theory, entropy is generated through charge diffusion only. The initial parameters are $q_1 = - q_2 = -0.6$, $a_1 = a_2 = 0.4$, $u_1 = - u_2 = 1$, $b = 2$, with total mass $M=1$.
\label{fig:charged}}
\end{figure}
These features (to LO at large $D$) are similar to what we have found in the neutral case (at NLO), but with stronger suppression of entropy generation in fission compared to fusion. It is easy to run numerical simulations of collisions that confirm this picture, but since conservation laws constrain much more the phenomena, they are less illustrative than in the neutral case. Therefore we only show one example of the fusion of two oppositely charged black holes, fig.~\ref{fig:charged}. It is qualitatively similar to fig.~\ref{fig:EntropyFrames}, even in its duration.
\section{Final comments and outlook}\label{sec:concl}
Our study has produced a consistent picture of the phenomena of fusion and fission of black holes, including the entropy generation mechanisms at different stages. One of the main results has been to highlight the attractor role that the intermediate, long-lived, quasi-thermalized black bar phase plays in a $2\to 1\to 2$ collision with fission, and how it is connected to a principle of total entropy maximization.
It is surprising that entropy maximization is somehow driving the dynamics. The system obeys time-irreversible equations that imply that, if at some moment in the evolution entropy is generated, then there is no coming back. But, in principle, the dynamics does not force the system, at least not in any manifest way, to evolve in a direction where entropy grows---it only forbids it to decrease---and even less so that entropy should grow as much as possible compatibly with conservation laws. The surprise that we find is that the evolution does lead to an end state very close to maximum entropy among a continuum of kinematically and geometrically allowed states. In statistical and quantum mechanics systems evolve stochastically sampling nearby configurations---and then, final thermodynamic equilibrium is achieved when entropy reaches a maximum. Here, however, we have a completely deterministic classical system whose equations seem to drive it in a direction that almost maximizes the rather non-obvious quantity $S_1$. The time scale involved is much shorter than, say, the scrambling time for a black hole (which, in the strict classical limit, is an infinite time). And moreover, for all we can see, entropy is in general almost, but not quite, maximized among possible final states. Indeed, by adjusting the initial conditions (e.g., to make a black string break up into multiple static black holes, or colliding black holes with very large $J/M$) the difference to the maximum can be made larger. Maximal entropy provides only an approximate criterion, but a remarkably accurate and powerful one.
In addition, we have also produced a detailed temporal and spatial tomography of entropy generation. Fusion, as might be expected, is highly irreversible. The production of entropy during fusion is relatively featureless, characterized by an initial peak in the production rate, where both shear and bulk viscosity contribute. If the system then settles down to a (long-lived) bar, a second stage with smaller entropy production can appear. Fission of unstable configurations follows the pattern of the decay of unstable black strings, with a duration of the order of \eqref{tfission}, and approximately equal amounts of dissipation of shear and expansion.
Our results are expected to be most applicable for black hole evolution in $D\geq 6$, but we would like to elaborate more on their possible qualitative relevance in four dimensions. The most important differences between $D=4$ and $D\geq 6$\footnote{The case $D=5$ is in some respects closer to $D=4$ and in others to $D\geq 6$. We will not refer to its peculiarities here.} refer to the dynamics of rotation: (i) quasi-stable Keplerian orbits (in General Relativity they are not fully stable due to gravitational wave emission), which are important in the evolution towards a merger, are possible in $D=4$ but not in $D\geq 6$; (ii) the properties of rotating black holes differ markedly in the two cases: in $D=4$ their spin is bounded above, while in $D\geq 6$ it is unbounded. Moreover, in $D=4$ the Kerr black hole (and not, {e.g.,\ } a black bar) will always be the endpoint of fusion, without any instability that would lead to its fission. If the total angular momentum in the system is above the Kerr bound for the final black hole, then the excess will be shed off into radiation, possibly involving an `orbital hang up' stage that delays the merger \cite{Campanelli:2006uy,Baumgarte:2010ndz}. In $D\geq 6$ instead, the upper bound on the angular momentum is not absolute but dynamical and set by an instability, so the excess angular momentum does not result in hang up but triggers fission.
We do not expect our studies of fission, nor of the relaxation to a stable black bar, to have application to collisions between Kerr black holes. But when studying fusion into a stable rotating MP black hole, the differences we have mentioned are less important than they may appear. The reason is that in $D=4$ the final plunge before two black holes merge occurs when their orbit becomes unstable. And when the spin of MP black holes is below the ultraspinning bound, their properties are similar to the Kerr black hole. It is interesting that in our simulations of collisions with relatively large impact parameters, we have observed that, prior to coalescence, the black hole trajectories are deflected into what looks like an approximately circular orbit (unstable dumbbells appear to describe such configurations), until the two black holes finally plunge towards each other. Qualitatively at least, this resembles the four-dimensional evolution.
So, as long as the angular momenta involved are moderate, the dynamics of black hole collisions and mergers in $D\geq 6$ are qualitatively similar to $D=4$, and our study of how entropy is produced (with the caveats that concern gravitational wave emission) should provide at least a guide to what to expect in that case.
\section*{Acknowledgments}
Work supported by ERC Advanced Grant GravBHs-692951, MEC grant FPA2016-76005-C2-2-P, and AGAUR grant 2017-SGR 754.
RL is supported by the Spanish Ministerio de Ciencia, Innovaci\'on y Universidades Grant FPU15/01414.
RS is supported by JSPS KAKENHI Grant Number JP18K13541 and partly by Osaka City University Advanced Mathematical Institute (MEXT Joint Usage/Research Center on Mathematics and Theoretical Physics).
\newpage
| 4af7a8855a5d90a3a44120afd31ba6a2bc55c837 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The advent of quantum annealing devices~\cite{johnson:11-ea,bunyk:14} has
spawned a renewed interest in the development of heuristic approaches to
solve discrete optimization problems. On the one hand, the quantum
revolution has inspired remarkable classical algorithms (see, for
example, Ref.~\cite{mandra:16b}) running on
conventional CMOS hardware that have raised the bar for emerging quantum
annealing hardware. On the other hand, substantial progress has been
made in recent years on the development of programmable devices based on
alternative technologies such as, for example, the coherent Ising
machine based on optical parametric
oscillators~\cite{wang:13b,hamerly:18x}, digital MemComputing machines
based on self-organizing logic gates~\cite{diventra:18,traversa:15}, and
the ASIC-based Fujitsu Digital
Annealer~\cite{matsubara:17,tsukamoto:17,aramon:19}.
The rapid progression of novel computing platforms and algorithms
demands tunable hard optimization problems for benchmarking purposes.
Much effort has been devoted to generate binary synthetic benchmark
problems whose optimal configurations are known \textit{a priori}
\cite{barthel:02,hen:15a,king:15,marshall:16,hamze:18,albash:18,hen:19a,hamze:20}.
These are frequently referred to as problems with \textit{planted
solutions}. In particular, benchmark problems that are easily scalable
to large system sizes, and ideally with tunable hardness, facilitate
systematic comparison of optimizers. Aside from their practical
significance, theoretical investigation of such problem ensembles
reveals intriguing insights into the nature of disordered systems, in
particular, the interplay between frustration, thermodynamic behavior,
and computational complexity~\cite{hamze:20,perera:20}.
In this paper we introduce Chook (version 0.2.0), a comprehensive
Python-based tool that integrates multiple solution planting schemes to
provide a unified framework for generating benchmark problems for binary
optimization problems. Chook currently supports tile
planting~\cite{hamze:18,perera:20}, Wishart planting~\cite{hamze:20},
deceptive cluster loops~\cite{mandra:18}, and equation
planting~\cite{hen:19a}. In addition, the software allows for the
construction of planted solutions for problems with higher-order ($k>2$)
interactions, by combining problems with lower-order ($k \leq 2$)
interactions.
The paper is organized as follows. In Sec.~\ref{sec:planting_schemes}
we present an overview of the solution planting schemes included in
Chook. Section~\ref{sec:chook} provides instructions on installation
and usage of the software, along with a detailed description of
parameters and options, followed by concluding remarks.
\section{Solution planting schemes} \label{sec:planting_schemes}
In solution planting, the goal is to construct a binary cost function
such that the minimizing configurations are known \textit{a priori}. In
the most general form, a cost function in Ising form with variables
$\boldsymbol{s} = (s_1, \dotsc, s_N)$, $s_i \in \{\pm 1\}$ is given
by
\begin{equation} \label{eq:main_hamiltonian}
\mathcal{H}(\boldsymbol{s}) = \sum_{j \in V} h_j s_j + \sum_{k=2}^n \; \sum_{( i_1, i_2, \dotsc, i_k ) \in E} J_{i_1 i_2 \dotso i_k} s_{i_1} s_{i_2} \dotsm s_{i_k},
\end{equation}
where the hypergraph $H = (V,E)$ with vertices $V$ and hyperedges $E$
captures the connectivity of the problem, and $\{h_j\}$ are the local
fields. A term containing a product of $k$ spins, $J_{i_1 i_2 \dotso
i_k} s_{i_1} s_{i_2} \dotsm s_{i_k}$, is referred to as a $k$-local
interaction with $J_{i_1 i_2 \dotso i_k}$ being the coupling constant.
Equation \eqref{eq:main_hamiltonian} can be readily mapped onto a cost
function of Boolean variables $\boldsymbol{x} = (x_1, x_2, \dotsc,
x_N)$, $x_i \in \{0, 1\}$ in the form of a high-order polynomial
unconstrained binary optimization (HOBO) problem via the transformation
$s_i \rightarrow 1-2x_i$.
Except for a few platforms like Azure Quantum, most of the software for
binary optimization problems mainly targets up to 2-local interactions,
in which case Eq.~\eqref{eq:main_hamiltonian} reduces to
\begin{equation}
\mathcal{H}(\boldsymbol{s}) = \sum_{j \in V} h_j s_j + \sum_{(i,j) \in E} J_{ij} s_i s_j,
\end{equation}
where $G=(V,E)$ is the underlying problem graph. Among the planting
schemes supported by Chook, tile planting, Wishart planting, and
deceptive cluster loops methods construct cost functions with $2$-local
interactions. Equation planting and $k$-local planting methods are
capable of generating problems with higher-order ($k>2$) local
interactions.
In what follows, we provide a summary of each solution planting scheme
supported by Chook. For a detailed description, the reader is referred
to the the original references introducing the methods.
\subsection{Tile planting}
In tile planting~\cite{hamze:18} one seeks to decompose the problem
graph into edge-disjoint vertex-sharing subgraphs and embed elementary
Ising subproblems that share a common ground state over the subgraphs.
The method produces scalable problems with highly-tunable complexity on
cubic and square lattice topologies~\cite{hamze:18,perera:20}.
It also extends to arbitrary graph structures, e.g., via lattice animals.
Consider a decomposition of the problem graph $G=(V,E)$ into subgraphs
$\{ G_l =(V_l, E_l) \}$ that ensures no edges are shared among the
subgraphs (i.e., edge-disjoint). For each subgraph, we define an Ising
cost function of the form
\begin{equation}
\mathcal{H}_l = \sum_{(i,j) \in E_l} J_{ij} s_i s_j.
\end{equation}
The subproblem Hamiltonians $\{\mathcal{H}_l\}$ are added to obtain the
complete Hamiltonian $\mathcal{H}_{_\text{TP}}$ over $G$ as
\begin{equation}
\mathcal{H}_{_\text{TP}} = \sum_{l} \mathcal{H}_l.
\end{equation}
The subproblems are constructed such that they share a common ground
state configuration $\boldsymbol{t}$, therefore the entire problem has a
ground state characterized by the same local configuration
$\boldsymbol{t}$ occupying the constituent subgraphs. For simplicity,
$\boldsymbol{t}$ is taken to be the ferromagnetic ground state, i.e.,
$\{+1, +1, \dotsc, +1\}$ (modulo $\mathbb{Z}_2$ symmetry). Once the
problem is constructed, the planted ferromagnetic ground state can be
concealed by a gauge randomization in which the couplers $\{J_{ij}\}$
are transformed as $J_{ij}^\prime \leftarrow J_{ij} q_i q_j$, where
$\boldsymbol{q}$ is an arbitrary ground state.
Chook supports the generation of tile-planted problems on square and
cubic lattices with periodic boundary conditions. The regular structure
of these lattices allows for a problem-graph decomposition that
naturally renders a subset of the unit cells as the subgraphs. Figure
\ref{fig:checker_board} illustrates this decomposition for square
lattices, in which the resulting unit-cell subgraphs (dark color) form a
checkerboard pattern.
For square lattices we define four subproblem classes $\{C_i\}$, $i \in
\{1,2,3,4\}$ that correspond to unit cycles (plaquettes) with different
levels of frustration [see Fig.~\ref{fig:subproblem_classes}(a) for an
illustration]. A subproblem from the class $C_i$ is constructed by
assigning the antiferromagnetic value $+1$ to a chosen coupler, the
ferromagnetic value $-1$ to randomly selected $i-1$ couplers, and $-2$
to the remaining couplers. Planted instances are generated by first
assigning a subproblem type for each subgraph in the lattice, followed
by a random rotation of the plaquette to allow for more disorder. We
define instance classes based on the probability distribution over
classes $\{C_i\}$ according to which subproblem types are assigned to
subgraphs. Specifically, we denote $p_i$ to be the probability of
choosing subproblems from class $C_i$, and uniquely define each instance
class based on the three probability parameters $\{p_1, p_2, p_3\}$,
where $p_1 + p_2 + p_3 \leq 1$. Multiple complexity transitions have
been observed in the multidimensional phase space defined by these
parameters~\cite{perera:20}.
For cubic lattices, Chook uses three subproblem types $F_{22}$,
$F_{42}$, and $F_6$, each having two, four, and six frustrated facets,
respectively [see Fig.~\ref{fig:subproblem_classes}(b)]. Each
subproblem class consists of $48$ members that are equivalent by
octahedral symmetry under the operations of rotation and
reflection~\cite{comment:subproblem_classes}. Similar to the square
lattice case, problem construction begins by assigning subproblem types
for the subgraphs according to the chosen probability distribution over
classes $\{F_{ij}\}$. Then for each subproblem type, one of the $48$
members are selected randomly. Instance classes are defined based on the
two probability parameters $\{ p_{F_{22}}, p_{F_{42}} \}$,
$p_{F_{22}}+p_{F_{42}} \leq 1$, where $p_{F_{22}}$ and $p_{F_{42}}$ are
the probabilities of choosing subproblems from the $F_{22}$ class and
$F_{42}$ class, respectively. By varying these parameters one can
achieve drastic changes in typical complexity, with higher
concentrations of $F_6$ subproblems resulting in problems that are many
orders of magnitude harder than random problems with bimodal and
Gaussian disorder~\cite{hamze:18}.
\begin{figure}[tb!]
\includegraphics[width=0.5\linewidth]{checker_board}
\caption{
Decomposition of a square lattice with periodic boundary conditions into
edge-disjoint, unit-cell subgraphs (shaded cells) on which Ising
subproblems are embedded. Under periodic boundary conditions, each
vertex is shared by exactly two subgraphs.
}
\label{fig:checker_board}
\end{figure}
\begin{figure}[tb!]
\includegraphics[width=\linewidth]{subproblem_classes}
\caption{
Class representatives of the unit-cell subproblems for (a) square-lattice
and (b) cubic-lattice topologies. Straight lines represent ferromagnetic
couplers with values $-1$ (thin lines) and $-2$ (thick lines), while
wavy lines denote antiferromagnetic couplers with value $+1$. All
subproblems have the ferromagnetic state as one of the ground states.
Classes $C_2$ and $C_3$ each have three equivalent representations (only
a single member shown) based on the ways the ferromagnetic bonds can be
distributed. $F_{22}$, $F_{42}$, and $F_6$ classes each have $48$
members that are equivalent under the octahedral symmetry.
}
\label{fig:subproblem_classes}
\end{figure}
\subsection{Wishart planting}
Wishart planting~\cite{hamze:20} generates Ising Hamiltonians on
complete graphs of the form
\begin{equation}
\mathcal{H}_{_\text{WP}}(\boldsymbol{s}) = \frac{1}{2} \sum_{i \neq j} J_{ij} s_i s_j = \frac{1}{2} \boldsymbol{s}^T \boldsymbol{J} \boldsymbol{s},
\end{equation}
in which the coupler matrix $\boldsymbol{J}$ follows a Wishart
distribution, a matrix generalization of the $\chi^2$ distribution. The
model exhibits a first-order phase transition in temperature, and allows
for dramatic variations in typical hardness over many orders of
magnitude as a control parameter is varied.
In Wishart planting, one defines the coupler matrix $\boldsymbol{J}$ in
terms of a $N \times M$ real-valued matrix $\boldsymbol{W}$, where $N$
is the number of spins and $M$ ($M \ge 1$) is a tunable parameter which
regulates the typical hardness and the thermodynamic behavior of the
problem ensemble. Specifically, we define $\tilde{\boldsymbol{J}}$ as
\begin{equation}
\tilde{\boldsymbol{J}} = \frac{1}{N} \boldsymbol{W} \boldsymbol{W}^T,
\end{equation}
and zero the diagonal to obtain $\boldsymbol{J} =
\tilde{\boldsymbol{J}}- \text{diag} (\tilde{\boldsymbol{J}})$. It can
be shown that the problem Hamiltonian $\mathcal{H}_{_\text{WP}}$ then
becomes
\begin{equation}
\mathcal{H}_{_\text{WP}}(\boldsymbol{s}) = \frac{1}{N} \frac{1}{2} \|\boldsymbol{W}^T \boldsymbol{s}\|^2 - \frac{1}{2} \text{Tr} (\tilde{\boldsymbol{J}}).
\end{equation}
Because the first term is in positive semidefinite quadratic form,
$\mathcal{H}(\boldsymbol{s})$ is minimized when $\boldsymbol{s} =
\boldsymbol{t}$ at which $\boldsymbol{W}^T \boldsymbol{t} = 0$. Thus,
the goal is to construct the matrix $\boldsymbol{W}$ for a given planted
solution $\boldsymbol{t}$ such that the condition $\boldsymbol{W}^T
\boldsymbol{t} = 0$ is satisfied. We achieve this by drawing the $M$
columns $\{\boldsymbol{\omega}^\mu\}$ of $\boldsymbol{W}$ from a
correlated Gaussian distribution ${\boldsymbol{\omega}^\mu} \sim
\mathcal{N}(0, \boldsymbol{\Sigma})$, where $\boldsymbol{\Sigma} = N[
\boldsymbol{I}_N - \boldsymbol{t} \boldsymbol{t}^T/N ]/(N-1)$ is the
covariance matrix. More precisely, we successively sample uncorrelated
Gaussian variates ${\boldsymbol{z}^\mu} \sim \mathcal{N}(0,
\boldsymbol{I}_N)$ and then multiply by the square root of
$\boldsymbol{\Sigma}$ to obtain ${\boldsymbol{\omega}^\mu}$, i.e.,
${\boldsymbol{\omega}^\mu} = \boldsymbol{\Sigma}^{\frac{1}{2}}
\boldsymbol{z}^\mu$. It can be shown that $\langle
\boldsymbol{\omega}^\mu, \boldsymbol{t} \rangle = 0$ for all $\mu$
generated in this manner. The matrix $\boldsymbol{W} \boldsymbol{W}^T$
is known to follow a Wishart distribution. Analogous to the
clause-to-variable ratio in Boolean satisfiability problems, we define
an equation-to-variable ratio $\alpha = M/N$ for modulating the
typical computational hardness of the problem ensemble. The class
exhibits a pronounced easy-hard-easy complexity transition as $\alpha$
is varied~\cite{hamze:20}.
One can construct problems with discrete coupler values via a simple
modification of the sampling procedure. When generating
$\{\boldsymbol{\omega}^\mu\}$, instead of using Gaussian variates
${\boldsymbol{z}^\mu} \sim \mathcal{N}(0, \boldsymbol{I}_N)$, one can
sample from a Rademacher distribution, i.e., draw samples independently
and uniformly from $\{-1, 1\}$. It can be shown that the scaled couplers
$J_{ij}^\prime = N^2 (N-1) J_{ij}$ assume values in the set of
equally-spaced integers given by
\begin{equation}
J_{ij}^\prime \in \{ -4M(N-1)^2, \dotsc, -4, 0, 4, \dots, 4M(N-1)^2 \}.
\end{equation}
\subsection{Deceptive cluster loops}
Deceptive cluster loops (DCL)~\cite{mandra:18} are a class of benchmark
problems designed for the Chimera topology of the D-Wave 2000Q and
D-Wave 2X quantum annealers, although the ideas can be generalized to
other topologies. DCL problems are derived from the conventional
frustrated cluster loops (FCL) problems~\cite{king:17}.
They have an additional control parameter that conceals the logical
structure of the problem for a particular range of values. Such a
feature makes these problems hard to solve using algorithms that exploit
the underlying logical structure.
Both DCL and FCL problems are based on the traditional frustrated loop
problems introduced in Ref.~\cite{hen:15a}. Frustrated loop problems are
constructed by generating $M = \alpha N$ loops on a graph $G=(V,E)$,
where $N$ is the number of nodes and $\alpha$ is the loop-to-node ratio.
The loops are generated by placing random walkers on random nodes. A
random walk is terminated when it crosses its own path, and the trailing
tail segment is discarded to form a closed loop. On each loop $k$, the
coupler values $J_{ij}^k$ are set to the ferromagnetic value $-1$ except
for a single, randomly-chosen coupler for which we assign the
antiferromagnetic value $+1$. The total Hamiltonian is formed by adding
up the couplers belonging to all the loops
\begin{equation}
\mathcal{H}_{_\text{DCL}} = \sum_{(i,j) \in E} \; \sum_{k=1}^M J_{ij}^k s_i s_j.
\end{equation}
The problem is discarded if $|\sum_{k=1}^M J_{ij}^k| > R$ for any edge
$(i,j) \in E$, where $R$ is referred to as the ``ruggedness.'' In
frustrated cluster loop (FCL) problems~\cite{king:17}, frustrated loops
are generated on a two-dimensional logical lattice embedded on a Chimera
graph. Here, all couplers within each Chimera unit cell are set to the
ferromagnetic value $-1$, forcing all physical spins within the unit
cell to behave as a logical spin. The frustrated loops are then
generated on the resulting $L_x \times L_y$ two-dimensional logical
lattice, where $L_x$ and $L_y$ are the linear dimensions of the parent
Chimera graph.
DCL problems are an extension of FCL problems. Here, all inter-cell
couplers between Chimera unit cells are multiplied by a scaling factor
$\lambda$, while leaving all intra-cell couplers intact. For small
values of $\lambda$ ($\lambda \sim 1$), DCL problems can be described by
a virtual planer model with each Chimera unit-cell behaving as a single
virtual variable. For $\lambda \gg 1$, the model behaves as a virtual
fully-connected bipartite problem. For intermediate values of
$\lambda$, the problem behavior is nontrivial and no logical structures
(as it happens for either small or large values of $\lambda$) can be
determined to simplify the optimization of the problem; see
Ref.~\cite{mandra:18} for details.
\subsection{Equation planting}
Systems of linear equations modulo 2, also known as ``exclusive or
satisfiability'' (XORSAT) equations, are solvable in polynomial time by
Gaussian elimination, but when formulated as optimization problems, they
are often challenging for heuristic
solvers~\cite{joerg:10,guidetti:11,farhi:12}. Equation
planting~\cite{hen:19a} casts XORSAT problems as $k$-local ($k>1$) Ising
cost functions.
Consider a linear system of equations modulo 2 with $N$ Boolean
variables $\{x_1, x_2, \dots, x_N\}$ and $M$ equations
\begin{equation}
\sum_{j=1}^N a_{ij} x_j = b_i,
\end{equation}
for $i \in \{1, \dotsc, M \}$, where the coefficients $\{a_{ij},
b_i\} \in \{0, 1\}$. Alternatively, each equation can be written in
terms of the bit-wise XOR operation as
$a_{i1} x_1 \oplus \dotsm \oplus a_{iN} x_N = b_i$.
When expressed in the Ising form, the above equation becomes
\begin{equation}
\prod_{j: a_{ij}=1} s_j = (-1)^{b_i} \qquad s_j \in \{\pm 1\},
\end{equation}
in which only the variables with non-zero coefficients are included in
the product. The linear system can be cast as an optimization problem
with an Ising cost function, i.e.,
\begin{equation}
\mathcal{H}^\prime = \sum_{i=1}^M \left[ \prod_{j: a_{ij}=1} s_j - (-1)^{b_i} \right]^2,
\end{equation}
which, after dropping irrelevant constants, reduces to
\begin{equation}
\mathcal{H}_{_\text{XORSAT}} = -\sum_{i=1}^M (-1)^{b_i} \prod_{j: a_{ij}=1} s_j.
\end{equation}
The ground state of the Ising cost function corresponds to the solution
of the linear system, given that the linear system is solvable. Here we
limit our attention to $k$-regular $k$-XORSAT problems, where each
equation contains exactly $k$ randomly selected variables out of the $N$
variables (hence $k$-XORSAT), and each variable appears in exactly $k$
equations (hence $k$-regular). Such a linear system consists of $N$
equations (i.e., $M=N$), and the resultant Ising Hamiltonian contains
$N$ $k$-local terms. If the linear system has solutions, the
ground-state energy of the Ising cost function is $-M$, and the
ground-state degeneracy (i.e., the number of minimizing configurations)
is given by $g = 2^{N-r}$, where $r$ is the number of linearly independent
(in arithmetic modulo 2) rows of the matrix $\boldsymbol{A}=(a_{ij})$.
For random $k$-regular $k$-XORSAT instances, $r$ is typically $\mathcal{O}(N)$
and hence the number of ground states grows sub-exponentially
~\cite{mandra:16c, alamino:09, alamino:08}.
\subsection{$k$-local planting}
Here we introduce a method of planting solutions for Hamiltonians with
higher-order ($k>2$) interactions by combining Hamiltonians with
lower-order ($k \leq 2$) interactions and known ground states.
Consider a set of $n$ problems described by the Hamiltonians $\{
\mathcal{H}^{(1)}(\boldsymbol{s}^{(1)}),
\mathcal{H}^{(2)}(\boldsymbol{s}^{(2)}), \dots,
\mathcal{H}^{(n)}(\boldsymbol{s}^{(n)}) \}$ whose ground state energies
are $\{E_0^{(1)}, E_0^{(2)}, \dots, E_0^{(n)}\}$.
The composite Ising cost function defined by the product
\begin{equation}
\mathcal{H}_{_\text{comp}} = \prod_{i=1}^n \left[
\mathcal{H}^{(i)} - E_0^{(i)} \right]
\end{equation}
is minimized for $\boldsymbol{s}^{(i)} = \boldsymbol{t}^{(i)}$, $i
\in \{1, \dotsc, n\}$, where $\boldsymbol{t}^{(i)}$ is a ground state of
$\mathcal{H}^{(i)}$.
To reduce the total number of variables in the composite problem,
we allow spins to be shared among subproblems.
The composite problem hence contains
$N = \text{max} \, \{ N^{(1)}, N^{(2)}, \cdots, N^{(n)}\}$ spins,
where $N^{(i)}$ is the number of spins in the $i$th problem.
The locality of the highest-order term in the composite
Hamiltonian $\mathcal{H}_{_\text{comp}}$ is $k_\text{max} = \sum_{i=1}^n
k_\text{max}^{(i)}$, with $k_\text{max}^{(i)}$ being the locality of the
highest order term in $\mathcal{H}^{(i)}$.
Chook constructs higher-order ($k>2$) cost functions with even
$k_\text{max}$ by using tile-planted problems and Wishart-planted
problems as constituent problems \cite{comment:precision}.
Because these problem types consist of
only $2$-local interactions, one cannot construct composite problems
with odd $k_\text{max}$ using these two problem types alone. Therefore
when generating problems with odd $k_\text{max}$, we also use a trivial
Hamiltonian with $1$-local interactions, namely, Ising spins coupled to
a bimodal random field, given by $\mathcal{H}_\text{1-local} =
\sum_{i=1}^{N_l} h_i s_i$, where $N_l$ is the number of spins and
$\{h_i\}$ are the random fields drawn independently and uniformly from
$\{+1, -1\}$. The ground state of $\mathcal{H}_\text{1-local}$ is
trivially obtained by aligning all spins with their random fields.
\section{Chook}
\label{sec:chook}
Chook is a platform-independent, Python-based tool distributed under the
Apache License 2.0. The code is openly available on the GitHub software
sharing platform (\code{https://github.com/dilinanp/chook}), as well as
being included in this submission (see ancillary files). It is also hosted
on the Python Package Index (PyPI) for easy installation via the package
management system pip. We ask you to please cite this work if you choose
to use Chook.
\subsection{Installation}
Chook requires Python version 3.4 or above for installation. We
recommend installing Chook in a Python virtual environment to avoid
potential conflicts with packages installed system-wide. If the Python
package management system pip is available, Chook can be installed from
the command line by running the command:
\begin{lstlisting}
$ pip install chook
\end{lstlisting}
Alternatively, chook can be directly downloaded from PyPI or the GitHub
repository and can be installed with the command:
\begin{lstlisting}
$ python setup.py install
\end{lstlisting}
During installation, the prerequisite Python libraries numpy, scipy, and
more-itertools will be automatically installed if they are not already
available.
\subsection{Usage}
Chook can be executed from the command line providing two required
positional arguments \code{<problem_type>} and \code{<config_file>},
followed by zero or more optional arguments. The basic usage ignoring
the optional arguments is
\begin{lstlisting}
$ chook <problem_type> <config_file>
\end{lstlisting}
where the different types of \code{<problem_type>}
are listed in Table~\ref{tab:problem_types},
and \code{<config_file>} is a configuration file in INI format that
contains the problem-type specific parameters (see
Sec.~\ref{sec:config_file}). Table~\ref{tab:arguments} shows the
complete list of optional arguments that can be appended to change the
default behavior of Chook.
\begin{table}
\caption{
Solution planting types supported by Chook and the
positional argument \code{<problem_type>} which is also the section
header in the configuration file under which the problem-type specific
parameters are listed. Note that for the problem type, the brackets
should be removed, e.g., for a tile-planted problem the \code{<problem_type>}
is \code{TP} and the section header \code{[TP]}.
\label{tab:problem_types}}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}} l l }
\hline \hline
Problem type & INI section header \\ [0.5ex]
\hline
Tile planting & \code{[TP]} \\
Wishart planting & \code{[WP]} \\
Deceptive cluster loops & \code{[DCL]} \\
Equation planting & \code{[XORSAT]} \\
$k$-local planting & \code{[K\_LOCAL]} \\
\hline
\hline
\end{tabular*}
\end{table}
\begin{table*}
\centering
\caption{Optional command-line arguments supported by Chook.
\label{tab:arguments}}
\begin{tabular}{@{\extracolsep{\fill}} p{4cm} p{12cm}}
\hline \hline
Argument flag & Description \\ [0.5ex]
\hline
\code{-n <num_instances>} & \code{<num_instances>} is the number of instances to be generated (default value: 10). \\
\code{-o <output_format>} & \code{<output_format>} can either be \code{ising} or \code{hobo},
and specifies whether the problem is expressed in
the Ising form or the binary HOBO form (default value: \code{ising}).\\
\code{-f <file_format>} & \code{<file_format>} can either be \code{txt} or \code{json},
and specifies whether the output files are in text
format or the JSON format (default value: \code{txt}). \\
\code{-h, --hel}\code{p} & Show a help message and exit.\\
\hline
\hline
\end{tabular}
\end{table*}
\subsection{Output}
Chook stores the generated problem instances in a subdirectory under the
current working directly, which is named according to the problem type
and major problem-specific parameters. For example, tile-planted problems on
a square lattice with linear lattice size $L=32$ and tuning parameters
$p_1=0.2$, $p_2=0.5$, and $p_3=0.1$ will be stored in a subdirectory
named \code{tile_planting_2D_L_32_p1_0.2_p2_0.5_p3_0.1}. Each problem
instance will be stored in a separate file within the said subdirectory.
Except for deceptive cluster loop problems for which the ground states
cannot be decoded (except in specific limits), the ground-state energies
will be stored in a separate file \code{gs_energies.txt} with two
columns: the instance file name and the ground state energy. For the
case of XORSAT problems, \code{gs_energies.txt} will contain a third
column, the ground state degeneracy.
The generated instances are expressed in the Ising form by default,
but can be cast in the binary HOBO form by setting the optional argument
\code{-o hobo}. By default, the instance files are in plain text format.
Each line in the instance file corresponds to a single $k$-local
interaction term following the format
\begin{lstlisting}
<i_1> <i_2> ... <i_k> <J>
\end{lstlisting}
where the first $k$ elements are the indices of the spin (Boolean)
variables, and the last element \code{<J>} is the coupling constant.
For the case of field terms (i.e., $1$-local interactions), the
entry consists of two terms: the index of the spin (Boolean) variable
followed by the field value. The user has the option to save the
instances in the JavaScript Object Notation (JSON) format by setting the
optional argument \code{-f json}.
\subsection{Configuration file}
\label{sec:config_file}
As the second command-line argument, Chook requires a configuration file
in INI format that specifies the parameters associated with the selected
problem type. Chook is distributed with a sample configuration file
\code{params.i}\code{n} which includes sample parameter specifications
for all problem types. Each parameter is specified as an INI property
with a \code{name} and a \code{value}, separated by an equal sign, i.e.,
\code{name = value}. Properties are grouped into sections according to
the problem types they are associated with, and the sections begin with
section headers of the form \code{[<problem_type>]}.
Table~\ref{tab:problem_types} shows the section headers associated with
the supported problem types. A section ends when the next section
header is encountered, or when the end of file is reached. When Chook
is executed, only the section associated with the chosen problem type is
read from the configuration file, and the rest of the file is ignored.
Table~\ref{tab:properties} shows a comprehensive list of supported
problem-specific properties. We recommend the users to modify the
provided \code{params.i}\code{n} file to meet their needs rather than
scripting their own configuration file to avoid potential errors.
We now describe the parameter specification for the $k$-local planting
method. In Chook, $k$-local problems are constructed by combining a
sequence of constituent problems (or ``subproblems'') of three supported
types, namely, tile planting, Wishart planting, and bimodal
random field terms. Parameter specifications for the subproblems should
be grouped into separate INI sections with user-defined headers and
should be listed below the main section \code{[K_LOCAL]}. The section
\code{[K_LOCAL]} contains two properties: \code{k_max} which represents
the locality $k_\text{max}$ of the highest-order term in the composite
Hamiltonian, and \code{subproblem_id_list} that accepts a list of
identifiers representing the subproblems. Each subproblem identifier
\code{<subproblem_id>}, enclosed within square brackets, i.e.,
\code{[<subproblem_id>]}, is used as the header of the section under
which the subproblem properties are listed. Subproblem identifiers
should begin with a letter, and may contain a combination of letters and
numbers. One can reuse the same subproblem (with the same parameter
specifications) multiple times in the construction procedure, in which
case the corresponding subproblem identifier should be repeated
accordingly in the \code{subproblem_id_list}. In addition to the usual
problem-specific properties, each section defining a subproblem should
contain an additional property \code{subproblem_type}, which is used to
identify the type of the subproblem being defined. The allowed values
are, ``\code{RF}'' for the bimodal random field terms, ``\code{TP}'' for tile planting,
and ``\code{WP}'' for Wishart planting. As the bimodal random field terms can
be trivially optimized, Chook enforces the user to minimize its usage,
and it is only allowed when constructing problems with odd values of
$k_\text{max}$. Therefore, for even values of $k_\text{max}$, one should use
$k_\text{max}/2$ subproblems with $2$-local interactions, which can be
tile-planted problems and/or Wishart problems. For odd values of
$k_\text{max}$, one should use $(k_\text{max}-1)/2$ subproblems with
$2$-local interactions, and a single subproblem with a $1$-local term
(i.e., a bimodal random field term).
\begin{table*}[t]
\centering
\caption{Problem-specific properties defined in the configuration file.
Note that the term ``variables'' refers to either spins (in Ising form) or Boolean variables (in HOBO form).
\label{tab:properties}}
\begin{tabular}{@{\extracolsep{\fill}} p{3cm} p{4cm} p{10.5cm}}
\hline \hline
Section header & Property name & Description \\ [0.5ex]
\hline
\multirow{14}{4em}{\code{[TP]}}
& \code{dimension} & Spatial dimension $D$ of the periodic lattice on which problems are constructed.
Should be set to $2$ for a square-lattice and $3$ for a cubic-lattice. \\
& \code{L} & Linear lattice size $L$.
Must be an even integer greater than two.
The number of variables in the planted problem is $N=L^D$.\\
& \code{p1, p2, p3} & Probabilities $\{p_1, p_2, p_3\}$ for constructing problems on square-lattice topology.
$p_i$ is the probability of choosing subproblems from class $C_i$.
Should satisfy the conditions $p_1+p_2+p_3 \leq 1.0$ and $0 \leq p_i \leq 1$, $i \in \{1,2,3\}$.
These parameters are ignored when \code{dimension = 3}. \\
& \code{ pF22, pF42} & Probabilities $\{p_{F_{22}}, p_{F_{42}}\}$ for constructing problems on a cubic-lattice.
$p_{F_{22}}$ and $p_{F_{42}}$ are the probabilities of selecting subproblems from classes $F_{22}$ and
$F_{42}$, respectively.
Should satisfy the conditions $p_{F_{22}} + p_{F_{42}} \leq 1.0$
and $0 \leq p_{F_{22}}, p_{F_{42}} \leq 1$.
These parameters are ignored when \code{dimension = 2}. \\
& \code{gauge_transform} & If set to \code{yes}, the planted ferromagnetic ground state will be concealed
via a gauge randomization. Allowed values: \{\code{yes}, \code{no}\} \\
\hline
\multirow{10}{4em}{\code{[WP]}}
& \code{N} & The number of variables $N$ in the planted problem. \\
& \code{alpha} & Equation-to-variable ratio $\alpha=M/N$, where $M$ is the
number of columns in the matrix $\boldsymbol{W}$.
Note that $M$ is determined from $\alpha$ as $M=\alpha N$, and will be rounded to the nearest
non-zero integer.
Thus the value of $\alpha$ internally represented by Chook can be different
from the user-provided value. \\
& \code{discretize_couplers} & If set to \code{yes}, the code generates problems with discrete couplers by
sampling from a Rademacher distribution instead of a Gaussian distribution
when constructing the matrix $\boldsymbol{W}$.
Allowed values: \{\code{yes}, \code{no}\} \\
& \code{gauge_transform} & If set to \code{yes}, the planted ferromagnetic ground state will be concealed
via a gauge randomization. Allowed values: \{\code{yes}, \code{no}\} \\
\hline
\multirow{13}{4em}{\code{[DCL]}}
& \code{Lx, Ly} & Linear dimensions $L_x$ and $L_y$ of the Chimera graph
on which problems are constructed ($ 1 \leq L_x, L_y \leq 16$) .
Frustrated loops are generated on the $L_x \times L_y$ logical lattice which treats
Chimera unit cells as logical variables.
The number of (physical) variables is $N = 8 L_x L_y$. \\
& \code{alpha} & Loop-to-node ratio $\alpha$ defined by $\alpha=M/N_l$, where $N_l = L_x \times L_y$
is the number of logical variables,
and $M$ is the number of loops generated on the logical lattice.
Note that $M$ is determined from $\alpha$ as $M=\alpha N_l$, and will be rounded to the nearest
non-zero integer.
Thus the value of $\alpha$ internally represented by Chook can be different
from the user-provided value. \\
& \code{R} & Ruggedness $R$ that limits the range of coupler strength as $|\sum_{k=1}^M J_{ij}^k| > R$.
Must be an integer greater than zero.\\
& \code{lambda} & Scaling factor $\lambda \geq 1$ by which the inter-cell couplers between Chimera
unit cells are scaled.
For $\lambda = 1$, DCL problems are equivalent to FCL problems. \\
\hline
\multirow{4}{4em}{\code{[XORSAT]}}
& \code{k} & Locality $k$ of the terms in the resultant Ising cost function
(equivalent to the number of variables per equation in $k$-regular $k$-XORSAT).\\
& \code{N} & The total number of variables $N$ in the problem ($N \geq k$).
This is equivalent to the number of equations in the $k$-regular $k$-XORSAT representation. \\
\hline
\multirow{8}{4em}{\code{[K_LOCAL]}}
& \code{k_max} & Locality $k_\text{max}>2$ of the highest-order term
in the composite Hamiltonian. \\
& \code{subproblem_id_list} & A list of user-defined, comma-delimited identifiers representing the subproblems.
Subproblem identifiers should begin with a letter and can include alphanumeric characters.
Each identifier enclosed by square brackets, i.e., \code{[<subproblem_id>]},
is used as the section header under which the properties of the subproblem are grouped.
A subproblem identifier can be repeated multiple times in \code{subproblem_id_list},
if one chooses to reuse the corresponding subproblem specification multiple times.\\
\hline
\multirow{4}{4em}{\code{[<subproblem_id>]}}
& \code{subproblem_type} & A numeric code used to identify the type of subproblem being defined.
Should be included with every INI section defining a subproblem,
in addition to the subproblem-specific properties.
Allowed values:
\code{RF} -- Bimodal random field , \code{TP} -- Tile planting, \code{WP} -- Wishart planting \\
& \code{<subproblem-specific properties>} & Define all the properties associated with the selected subproblem type.
If the subproblem type is a bimodal random field, define a single property
\code{N}, which specifies the number of variables in the subproblem. \\
\hline
\hline
\end{tabular}
\end{table*}
\section{Summary}
\label{sec:summary}
We have presented a Python-based suite, Chook, for generating discrete
optimization problems with known solutions using multiple popular
solution planting schemes. Chook is distributed freely via Python
Package Index (PyPI) and GitHub software sharing platform, and allows
for fast and easy installation on any platform. The code supports the
construction of cost functions with tunable hardness and local
interactions spanning $2$-local to arbitrary higher-order. We believe
that Chook will be highly beneficial for generating benchmark problem
sets for current and future generations of programmable devices for
discrete optimization.
\begin{acknowledgments}
D.P. would like to acknowledge Chris Pattison for suggestions during the
initial code development phase. H.G.K.~would like to thank Nacho Kilber
for inspiration. F.H. is indebted to Julia Child for stimulating his
work in this area through her classic and pioneering oeno-thermalized
Chook algorithm. Part of this research is based upon work supported in
part by the Office of the Director of National Intelligence (ODNI),
Intelligence Advanced Research Projects Activity (IARPA), via MIT
Lincoln Laboratory Air Force Contract No.~FA8721-05-C-0002. The views
and conclusions contained herein are those of the authors and should not
be interpreted as necessarily representing the official policies or
endorsements, either expressed or implied, of ODNI, IARPA, or the
U.S.~Government. The U.S.~Government is authorized to reproduce and
distribute reprints for Governmental purpose notwithstanding any
copyright annotation thereon.
\end{acknowledgments}
| 5270136c7fad7c659b996f3269420c4e96e6856b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{}
\documentclass[10pt]{article}
\usepackage{hyperref}
\hypersetup{colorlinks=true,linkcolor=black,citecolor=black,urlcolor=black,breaklinks=true}
\usepackage{float}
\usepackage[symbol]{footmisc}
\usepackage[superscript,sort]{cite}
\usepackage{array}
\usepackage{bm}
\usepackage{longtable}
\newcolumntype{x}[1]{%
>{\centering\hspace{0pt}}p{#1}}%
\usepackage[normalem]{ulem}
\usepackage{amsmath,amssymb}
\newcommand{\textup{\AA}}{\textup{\AA}}
\usepackage{amsthm,amscd,amsxtra,amsfonts,amsmath,amssymb,multirow}
\usepackage{wrapfig}
\usepackage[tiny,compact]{titlesec}
\usepackage{threeparttable}
\usepackage{helvet}
\renewcommand{\familydefault}{\sfdefault}
\usepackage{booktabs}
\usepackage[table]{xcolor}
\usepackage{xcolor,colortbl}
\usepackage{placeins}
\usepackage{tikz}
\usetikzlibrary{shapes,arrows}
\usepackage[T1]{fontenc}
\usepackage[linesnumbered,ruled]{algorithm2e}
\setlength{\oddsidemargin}{-0.2in}
\setlength{\textwidth}{6.8in}
\setlength{\topmargin}{0.0in}
\setlength{\headheight}{0in}
\setlength{\headsep}{0in}
\setlength{\textheight}{9.0in}
\setlength{\footskip}{0.3in}
\providecommand{\e}[1]{\ensuremath{\times 10^{#1}}}
\setlength{\parindent}{0.15in}
\renewcommand\floatpagefraction{.9}
\renewcommand\topfraction{.9}
\renewcommand\bottomfraction{.9}
\renewcommand\textfraction{.1}
\setcounter{totalnumber}{50} \setcounter{topnumber}{50}
\setcounter{bottomnumber}{50} \setlength{\floatsep}{0.05in}
\setlength{\textfloatsep}{0.05in} \setlength{\intextsep}{0.05in}
\setlength{\abovecaptionskip}{0.05in}
\setlength{\belowcaptionskip}{0.05in}
\titlespacing*{\section}{0pt}{*0}{*0}
\titlespacing*{\subsection}{0pt}{*0}{*0}
\titlespacing*{\subsubsection}{0pt}{*0}{*0}
\titlespacing{\paragraph}{0pt}{*0}{*1}
\renewcommand{\thefootnote}{\roman{footnote}}
\newcommand{\warning}[1]{\footnote{\textcolor{red}{{\textbf{#1}}}}}
\newcommand{\inlinewarning}[1]{{\textcolor{red}{{\textbf{#1}}}}}
\newcommand{\ntodo}[1]{\footnote{\textcolor{blue}{{\textbf{Nathan to do: #1}}}}}
\newcommand{\inlinentodo}[1]{{\textcolor{blue}{{\textbf{Nathan to do: #1}}}}}
\definecolor{MyPurple}{rgb}{1,0,1}
\newcommand{\gtodo}[1]{\footnote{\textcolor{MyPurple}{{\textbf{Guowei to do: #1}}}}}
\newcommand{\inlinegtodo}[1]{{\textcolor{MyPurple}{{\textbf{Guowei to do: #1}}}}}
\newcommand{\inlineanswer}[1]{{\textcolor{blue}{{\textbf{Answer: #1}}}}}
\newcommand{\answer}[1]{\footnote{\textcolor{blue}{{\textbf{Answer: #1}}}}}
\newcommand{\myrevision}[1]{{\marginpar{\Large $\Leftarrow$}}{\textbf{{#1}}}}
\newcommand{\myrevisionfig}[1]{{\textbf{{#1}}}}
\newcommand{\myparagraph}[1]{{\vskip 5pt \noindent{\bf [{#1}]}}}
\renewcommand{\vec}[1]{{\mathbf{{#1}}}}
\renewcommand{\thesection}{\Roman{section}}
\renewcommand{\thesubsection}{{\thesection}.\Alph{subsection}}
\renewcommand{\thesubsubsection}{{\thesubsection}.\arabic{subsubsection}}
\newcommand{\norm}[1]{{\left| {#1} \right\|}}
\newcommand{\beq}[1]{\begin{equation} \label{#1}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{\begin{array}{ll}}{\begin{array}{ll}}
\newcommand{\end{array}}{\end{array}}
\newcommand{\displaystyle}{\displaystyle}
\newcommand{&\!\!\!\disp}{&\!\!\!\displaystyle}
\def\S {\S}
\usepackage{longtable}
\usepackage{threeparttable}
\usepackage{array}
\newcolumntype{P}[1]{>{\centering\arraybackslash}p{#1}}
\usepackage{multirow}
\usepackage{color, colortbl}
\definecolor{Lightblue}{rgb}{0.867,0.914,0.961}
\definecolor{Lightgreen}{rgb}{0.883,0.934,0.848}
\DeclareGraphicsExtensions{.pdf,.jpeg,.png}
\title{Generative network complex for the automated generation of druglike molecules
}
\author{Kaifu Gao$^{1}$, Duc Duy Nguyen$^{1}$, Meihua Tu$^{2}$, and Guo-Wei Wei$^{1,3,4,}$\footnote{
Corresponding to Guo-Wei Wei. Email: [email protected]}\\
$^1$ Department of Mathematics,
Michigan State University, MI 48824, USA.\\
$^2$ Pfizer Medicine Design, 610 Main St, Cambridge, MA 02139, USA.\\
$^3$ Department of Electrical and Computer Engineering,
Michigan State University, MI 48824, USA. \\
$^4$ Department of Biochemistry and Molecular Biology,
Michigan State University, MI 48824, USA. \\
}
\date{\today}
\begin{document}
\maketitle
\begin{abstract}
Current drug discovery is expensive and time-consuming. It remains a challenging task to create a wide variety of novel compounds with desirable pharmacological properties and cheaply available to low-income people. In this work, we develop a generative network complex (GNC) to generate new drug-like molecules based on the multi-property optimization via the gradient descent in the latent space of an autoencoder. In our GNC, both multiple chemical properties and similarity scores are optimized to generate and predict drug-like molecules with desired chemical properties.
To further validate the reliability of the predictions, these molecules are reevaluated and screened by independent 2D fingerprint-based predictors to come up with a few hundreds of new drug candidates. As a demonstration, we apply our GNC to generate a large number of new BACE1 inhibitors, as well as thousands of novel alternative drug candidates for eight existing market drugs, including Ceritinib, Ribociclib, Acalabrutinib, Idelalisib, Dabrafenib, Macimorelin, Enzalutamide, and Panobinostat.
\end{abstract}
\section{Introduction}
Drug discovery ultimately tests our understanding of molecular biology, medicinal chemistry, genetics, physiology, and pharmacology, the status of biotechnology, the utility of computational sciences, and the maturity of mathematical biology. Technically, drug discovery involves target discovery, lead discovery, lead optimization, preclinical development, three phases of clinical trials, and finally, launching to the market only if a drug can be demonstrated to be safe and effective in every stage. Among them, lead discovery, lead optimization, and preclinical development disqualify tens of thousands of compounds based on their binding affinities, solubility, partition coefficients, clearances, permeability, toxicities, pharmacokinetics, etc., leaving only about tens of them to clinical trials. Therefore, drug discovery is expensive and time-consuming currently: it takes about \$2.6 billion dollars and more than ten years, on average, to bring a new drug to the market \cite{dimasi2016innovation}. Reducing the expense and speeding up the process is one of the top priorities of the pharmaceutical industry.
One of the key challenges in small molecule drug discovery is to find novel chemical compounds with desirable properties. Much effort has been taken to optimize this critical step in the drug discovery pipeline. For example, the development of high-throughput screening (HTS) has led to an unprecedented increase in the number of potential targets and leads \cite{hughes2011principles}. HTS can quickly conduct millions of tests to identify active compounds of interest using compound libraries rapidly \cite{macarron2011impact}. While there has been an increase in the number of potential targets and leads, the number of newly generated molecular entities has remained stable because of a high attrition rate by the elimination of leads with inappropriate physicochemical or pharmacological properties during preclinical development and clinical phases \cite{graul2009overcoming,tareq2010predictions}.
Rational drug design (RDD) approaches are proposed to better identify candidates with the highest probability of success \cite{waring2015analysis}. These methods aim at finding new medications based on the knowledge of biologically druggable targets \cite{dimasi2016innovation,stromgaard2017textbook}.
More recently, computer-aided drug design (CADD) has emerged as a useful approach in reducing the expense and period of drug discovery \cite{alqahtani2017silico}. Computational techniques have been developed for both virtual screening (VS) and optimizing the ADME properties of lead compounds. Primarily, these methods are designed as {\it in silico} filters to eliminate compounds with undesirable properties. These filters are widely applied for the assembly of compound libraries using combinatorial chemistry \cite{balakin2009compound}. The integration of early ADME profiling of lead chemicals has contributed to speeding up lead selection for phase-I trials without large amounts of revenue loss \cite{kapetanovic2008computer}. Currently, compounds are added in libraries based on target-focused design or diversity considerations \cite{huang2016protein}. VS and HTS can screen compound libraries to a subset of compounds whose properties are in agreement with various criteria \cite{szymanski2012adaptation}.
Despite these efforts, current databases of chemical compounds remain small when compared with the chemical space spanned by all possible energetically stable stoichiometric combinations of atoms and topologies in molecules. It is estimated that there are about 10$^{60}$ distinct molecules; among them, roughly 10$^{30}$ are drug-like \cite{macarron2011impact}. As a result, computational techniques are also being developed for {\it de novo} design of drug-like molecules \cite{schneider2005computer} and generating large virtual chemical libraries, which can be screened more efficiently for {\it in silico} drug discovery.
Among the available computational techniques, deep neural networks (DNN) have gathered much interest for their ability to extract features and learn physical principles from training data. Currently, DNN-based architectures have been successfully applied in a wide variety of fields in the biological and biomedical sciences \cite{min2017deep,mamoshina2016applications}.
More interestingly, many deep generative models based on sequence-to-sequence autoencoders (Seq2seq AEs) \cite{sutskever2014sequence}, variational autoencoders (VAEs) \cite{kingma2013auto}, adversarial autoencoders (AAEs) \cite{makhzani2015adversarial}, generative adversarial networks (GANs) \cite{goodfellow2014generative} or reinforcement learning \cite{kaelbling1996reinforcement}, etc. have been proposed for exploring the vast drug-like chemical space and generating new drug-like molecules. Winter {\it et al.} \cite{winter2019efficient,winter2019learning} performed the optimization based on particle swarm optimization on the continuous latent space of a Seq2seq AE to generate new molecules with desired properties. Gomez-Bombarelli {\it et al.} \cite{gomez2018automatic} used a VAE to encode a molecule in the continuous latent space for exploring associated properties. Skalic {\it et al.} \cite{skalic2019shape} combined a conditional VAE and a captioning network to generate previously unseen compounds from input voxelized molecular representations. Kadurin {\it et al.} \cite{kadurin2017drugan} built an AAE to create new compounds. Sattarov {\it et al.} \cite{sattarov2019novo} combined deep autoencoder RNNs with generative topographic mapping to carry out {\it de novo} molecule design. A policy-based reinforcement learning approach was proposed to tune RNNs for episodic tasks \cite{olivecrona2017molecular,popova2018deep}, and extended to design desired molecules \cite{kang2018conditional}. Zhou {\it et al.} \cite{zhou2019optimization} also proposed a strategy to optimize molecules by combining domain knowledge of chemistry and state-of-the-art reinforcement learning techniques.
However, the generative strategies mentioned above are not drug-specified. What is vital to drug discovery is to design potential drug candidates for specific drug targets. In a regular drug discovery procedure, the starting point is target identification, followed by lead generation. Then, lead optimization is performed to make lead compounds more drug-like \cite{hughes2011principles}.
It is useful to find new lead compounds to replace existing market drugs for several reasons. First, existing market drugs might not be optimal. For example, they might not be potent enough, be too toxic and harmful to human health, or be too hard to synthesize. Additionally, they might have side effects, insufficient aqueous solubility (log S) that reduces the absorption of drugs \cite{kerns2008vitro}, or higher partition coefficients (log P) that lead to improper drug distributions within the body \cite{leo1971partition}. Moreover, for very expensive drugs, it is desirable to find cheap alternatives.
With the generation of new alternative lead compounds in mind, one can make use of existing drug datasets to develop drug-specified generative models. In this process, it is critical to apply a similarity restraint to generate hundreds or even thousands of new drug-like molecules inside the chemical space close to the reference molecule. This similarity restraint enables us to generate new molecules that remain effective to the target. Moreover, from the viewpoint of machine learning, higher similarities to existing data always lead to more reliable predictions in generating molecules. The generative model can also realize lead optimization: by incorporating optimizers, generated molecules are designated to have one or more chemical properties better than the reference molecule. As a result, a large number of alternative drug candidates are created by drug-specified generative models. These candidates could be an effective and specified library to further screen for better or cheaper drug alternatives.
Therefore, in this work, we develop a generative network complex (GNC) based on the multiple-property optimization via gradient descent in the latent space to automatically generate new drug-like molecules. One workflow of our GNC consists of three following stages
\begin{enumerate}
\item The SMILES string of a seed molecule are encoded into a vector in the latent space by a pre-trained encoder.
\item Starting with the latent vector of the seed molecule, a DNN-based drug-like molecule generator modifies the vector via gradient descent, so that new latent vectors that satisfy multiple property restraints including chemical properties and similarity scores to the desired objectives are created.
\item A pre-trained decoder decodes these new latent vectors into the SMILES strings of newly generated molecules.
\end{enumerate}
The rest of the paper is organized as follows. Section \ref{sec:methods} introduces our new GNC framework formed by the seq2seq AE and drug-like molecule generator. Section \ref{sec:experiments} discusses its reliability test on the BACE1 target, and more importantly, presents the performance of our GNC on a variety of existing drug targets. The insight into the roles of the multiple property restraints is offered in Section \ref{sec:discussions}. The conclusion is given in Section \ref{sec:conclusion}.
\section{Methods}\label{sec:methods}
\subsection{The sequence to sequence antoencoder (seq2seq AE)}
The seq2seq model is an autoencoder architecture originated from natural language processing \cite{sutskever2014sequence}. It has been demonstrated as a breakthrough in language translation. The basic strategy of the seq2seq AE is to map an input sequence to a fixed-sized vector in the latent space using a gated recurrent unit (GRU) \cite{cho2014learning} or a long short-term memory (LSTM) network \cite{hochreiter1997long}, and then map the vector to a target sequence with another GRU or LSTM network. Thus the latent vector is an intermediate representation containing the ``meaning" of the input sequence.
In our case, input and output sequences are both SMILES strings -- a one-dimensional "language" of chemical structures \cite{weininger1988smiles}. The seq2seq AE is trained to have a high reconstruction ratio between inputs and outputs so that the latent vectors contain faithful information of the chemical structures (see the upper part of Figure \ref{fig:gen-seq2seq}). Here we utilize a pre-trained seq2seq model from a recent work \cite{winter2019learning}.
\subsection{The drug-like molecule generator based on the multi-property optimization}
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{gen-seq2seq.png}
\caption{A schematic illustration of our generative network complex 2.}
\label{fig:gen-seq2seq}
\end{figure}
In our new GNC, we design a drug-like molecule generator elaborately, so that generated molecules not only satisfy desired properties but also share common pharmacophores with reference compounds. From a seed molecule, one generative workflow of the GNC is depicted in Figure \ref{fig:gen-seq2seq} and described as below.
\begin{enumerate}
\item Randomly pick a low-binding-affinity molecule from a target-specified training set as the seed, then the SMILES string of the seed molecule is encoded by a pre-trained encoder (in our case a GRU encoder) into a latent vector.
\item The latent vector of the seed is fed into our DNN molecule generator. In every epoch, the generator comes up with a new vector $X \in \mathbb{R}^n$, and the deep learning network is instructed to evaluate $X$ via the following loss function
\begin{equation}\label{eq:dnn_loss}
{\cal L}(X)=\frac{1}{n}\sum_{i=1}^n k_i| \hat{y}_i(X) - y_{i0}|,
\end{equation}
where, $k_i$ is the $i$th predefined weight serving the purpose of emphasizing or deemphasizing different property restraints, $\hat{y}_i(X)$ is the predicted $i$th property value by a pre-trained predictor ${\cal M}_i$. Additionally, $y_{i0}$ is the objective value of the $i$th property. The restrained properties can be binding affinity (BA), the similarity score (Sim) to a reference molecule or others such as partition coefficient (Log P), Lipinski's rule of five \cite{lipinski1997experimental}, etc. Some guideline for $y_{i0}$ includes, in the BA restraint, one often sets $y_{\Delta G}$ < -9.6 kcal/mol, in the Log P condition, it is common to set $y_{\rm LogP}$ < 5, etc.
\item Gradient descent is used to minimize the loss function in Eq. (\ref{eq:dnn_loss}) until the maximum number of epochs is reached
\item The generated latent vectors satisfying the desired restraints are decoded into SMILES strings through a pre-trained decoder, as shown in Figure \ref{fig:gen-seq2seq}.
\end{enumerate}
To create a variety of novel drug-like molecules originated from leads or existing drugs (reference molecules), one can adopt different seed molecules as well as different objective values to achieve desired properties and similarity scores. The ultimate purpose of our molecule generator is to keep modifying the latent vector to satisfy the multiple druggable property restraints. Figure \ref{fig:gen-mechanism} illustrates the general mechanism of our generator. Notably, the pre-trained predictor weights inside our model stay intact throughout the backpropagation of the training process to the generator. The loss function converges when all conditions are met, and then, resulting latent vectors are decoded into SMILES strings.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{mechanism-gen2.png}
\caption{The illustration of the latent-space molecule generator.}
\label{fig:gen-mechanism}
\end{figure}
\subsection{The parameters of the molecule generator} \label{sec:param-gen}
In our model, the dimension of the latent space is 512, so the input and output dimensions of the DNN molecule generator are also 512. The DNN generator has two hidden layers, with 1024 neurons in each layer. The activation function is tanh, the learning rate is 0.1, and the momentum is also 0.1. In this work, we are interested in binding affinity and similarity score restraints. The regularization coefficients of these two restraints ($k_{\rm \Delta G}$ and $k_{\rm Sim}$) are set to 1 and 10, respectively. The similarity score restraints is determined via the Tanimoto coefficient between a generated latent vector and the latent vector of a reference molecule.
The binding affinity restraints rely on pre-trained binding affinity predictors. A pre-trained binding affinity predictor (LV-BP) takes latent vectors as its inputs and return predicted binding affinities. Therefore, typically, the input dimension of the predictor is 512, the output dimension is 1. The DNN predictor has three hidden layers with 1024, 1536, and 1024 neurons, respectively. The ReLU activation function is applied. The learning rate is 0.001, the number of training epochs is 4000, the batch size is 4. The predictor network is trained on target-specified datasets carefully selected from public databases such as ChEMBL \cite{gaulton2011chembl}. The generator and predictor are both programmed in the framework of PyTorch (Version 1.0.0) \cite{paszke2017pytorch}.
In the current work, for each generation task, we randomly pick 50 low-binding-affinity molecules from the preselected dataset as seeds. For each seed, the generative network is run in a total of 2000 epochs, which takes less than 10 minutes under the supercomputer equipped with one Nvidia K80 GPU card. In practice, to quickly fill up the potential chemical search space, one can use more seeds and run more epochs for each seed.
\subsection{Binding affinity reevaluation by the 2D-fingerprint predictor} \label{sec:2DFP-BP}
Besides generating new molecules, the LV-BP in our GNC also predicts their binding affinities. However, no experimental values are available to validate these predicted affinities. Therefore, we cross-validate them using alternative binding affinity predictors. In the present work, we construct machine learning predictors based on 2D fingerprints (2DFP-BPs) to reevaluate the affinities of generated compounds. The 2D fingerprints computed from their SMILES strings are inputs to these 2DFP-BPs. If the predictions from the LV-BP and 2DFP-BPs are consistent, we regard the predictions as reliable.
According to our previous tests \cite{gao20202d}, the consensus of ECFP4 \cite{rogers2010extended}, Estate1\cite{hall1995electrotopological} and Estate2 \cite{hall1995electrotopological} fingerprints performs best on binding-affinity prediction tasks. Therefore, this work also makes use of this consensus. We employ the RDKit software (version 2018.09.3) \cite{landrum2006rdkit} to generate 2D fingerprints from SMILES strings. Since the training sets in our current cases are not so large, we apply gradient boosting decision tree (GBDT) \cite{schapire2003boosting} model due to its accuracy and speed when handling small datasets. This GBDT predictor is constructed using the gradient boosting regressor module in scikit-learn (version 0.20.1) \cite{pedregosa2011scikit} and the following parameters: n\_estimators=10000, max\_depth=7, min\_samples\_split=3, learning\_rate=0.01, subsample=0.3, and max\_features=sqrt.
The criteria used in our reevaluation are the root mean square error (RMSE), Pearson correlation coefficient ($R_p$), and active ratio. Here, the active ratio means the ratio of the number of the active molecules indicated both by the 2DFP-BP and LV-BP to the number of the active ones indicated by the LV-BP.
\subsection{Multitask DNN predictors} \label{sec:param-multi}
Multitask DNN predictors \cite{caruana1997multitask} for both latent vectors and 2D fingerprints are built for the drug Ribociclib with two different targets.
The latent-vector based model has three hidden layers with 1024, 1536, and 1024 neurons. For the 2D-fingerprint based models, because the three different 2D fingerprints ECFP4, Estate1, Estate2 have 2048, 79, and 79 features, respectively, two different network architectures are used. For ECFP4, the numbers of neurons in the three hidden layers are 2500, 1500, and 500, respectively. For Estate1 and Estate2, their numbers of neurons are 500, 1000, and 500. Other parameters are the same as those of our single task predictors. These multitask models are also programmed in the framework of PyTorch (Version 1.0.0).
\subsection{Drug-target interaction and common pharmacophore analysis}
The interactions between drugs and their targets, as well as the pharmacophores of the drugs, are investigated. The purpose is to explore whether our generated molecules can still bind to the drug targets.
Drug-target interactions are analyzed via the protein-ligand interaction profiles \cite{salentin2015plip}. It can identify drug-target interactions as well as their types such as hydrogen bonds, hydrophobic interactions, etc.
However, the interaction analysis itself could not determine whether interactions are critical or not to the drug-target binding. By using the Phase module in Schr\"{o}dinger (version 2018-4) \cite{dixon2006phase}, we build pharmacophore models via searching common pharmacophores in all the active compounds to the target. Since these pharmacophores are widespread to all the active compounds, they are critical to the binding. Thus, if generated molecules still contain such pharmacophores, we can believe they are potential binders.
It could be time-consuming to recognize common pharmacophores of hundreds of compounds. To avoid this obstacle, we group compounds into 50 clusters via the k-means algorithm implemented by scikit-learn \cite{pedregosa2011scikit}. We, then, collect the centroids of these 50 clusters for the common pharmacophore search.
\subsection{Datasets}
In this work, first, we explore the effect of different objective values in our generator on the binding-affinity prediction reliability to generated molecules. We carry out this reliability test on the Beta-Secretase 1 (BACE1) dataset.
BACE1 is a transmembrane aspartic-acid protease human protein encoded by the BACE1 gene. It is essential for the generation of beta-amyloid peptide in neural tissue \cite{vassar2009beta}, a component of amyloid plaques widely believed to be critical in the development of Alzheimer's, rendering BACE1 an attractive therapeutic target for this devastating disease \cite{prati2017bace}. We download 3916 BACE1 compounds from the ChEMBL database. In the seq2seq autoencoder we utilized, there is a molecule filter that only selects organic molecules with more than 3 heavy atoms, their weights between 12 and 600, and their Log P values between -5 and 7 \cite{winter2019learning}. As a result, a total of 3151 molecules in the BACE1 dataset pass this filter and are used as the training set.
\begin{table}[h]
\centering
\begin{tabular}{|P{2.0cm}|P{1.5cm}|P{3.2cm}|P{1.4cm}|P{1.4cm}|P{2.0cm}|P{2.6cm}|}
\hline
\rowcolor{Lightgreen}
\textbf{Drug name} & \textbf{ChEMBL ID} & \textbf{Treatment} & \textbf{FDA approval year} & \textbf{$\Delta G$ (kcal/mol)} & \textbf{Filtered training set size} & \textbf{$\Delta G$ range of training set (kcal/mol)} \\
\hline
Ceritinib & 2403108 & Non-small cell lung cancer & 2014 & -10.77 & 1203 & -5.26 to -13.93 \\ \hline
Ribociclib & 3545110 & Breast cancer & 2017 & -10.98 \& -10.16 & 918 \& 289 & -6.31 to -12.98 \& -4.11 to -14.13 \\
\hline
Acalabrutinib & 3707348 & Mantle cell lymphoma & 2017 & -11.67 & 1451 & -4.21 to -14.05 \\ \hline
Idelalisib & 2216870 & Blood cancers & 2014 & -10.43 & 1959 & -1.92 to -14.16 \\ \hline
Macimorelin & 278623 & Adult growth hormone deficiency & 2017 & -10.48 & 608 & -4.18 to -14.68 \\ \hline
Dabrafenib & 2028663 & Cancers with a mutated gene BRAF & 2013 & -12.35 & 2254 & -5.49 to -14.68 \\ \hline
Enzalutamide & 1082407 & Prostate cancer & 2012 & -10.53 & 1386 & -3.83 to -13.72 \\ \hline
Panobinostat & 483254 & Cancers & 2015 & -12.46 & 1645 & -2.91 to -12.46 \\
\hline
\end{tabular}
\caption{The information about the eight market drugs used in the present study.}
\label{table:eight-drug-inf}
\end{table}
More importantly, we employ our GNC to design alternative promising drug candidates for the eight drugs on the market. For each drug, we construct a dataset of the compounds that bind to the same drug target from the ChEMBL database. The collected compounds are also filtered by the filter in the seq2seq autoencoder. Table \ref{table:eight-drug-inf} lists information regarding these eight drugs, namely drug name, ChEMBL ID, FDA approval year, experimental drug affinity ($\Delta G$), filtered training set size, and affinity range of the training set.
The SciFinder is one of the most comprehensive databases for chemical literature and substances (\url{ https://scifinder-n.cas.org}). It contains more than 143 million organic and inorganic substances. Here we search our generated molecules in this database to confirm they are new and never existed before.
\section{Experiments} \label{sec:experiments}
\subsection{Designing BACE1 inhibitors}
\subsubsection{The accuracy of the seq2seq AE and predictors}
We first test the accuracy of the seq2seq autoencoder and the LV-BP and 2DFP-BP predictors.
When performing the seq2seq model on the filtered BACE1 dataset with 3151 molecules, the reconstruction ratio is 96.2\%. This high ratio guarantees the essential information of these input molecules is encoded into the corresponding latent vectors.
Subsequently, these latent vectors are used as the features to train our latent-vector DNN binding affinity predictor (LV-BP), the labels are their corresponding experimental binding affinities. In a 5-fold cross-validation test on the BACE1 dataset, the LV-BP achieves an average Pearson correlation coefficient ($R_p$) of 0.871 and an average RMSE of 0.704 kcal/mol.
The 2D-fingerprint GBDT binding affinity predictor (2DFP-BP) is used to reevaluate the predictions from the LV-BP. We also test this 2DFP-BP by the 5-fold cross-validation. The average $R_p$ and RMSE are 0.874 and 0.692 kcal/mol, respectively, quite comparable to the LV-BP.
\subsubsection{Convergence analysis}
\begin{figure}[!htp]
\centering
\includegraphics[width=0.6\textwidth]{pathway-loss.png}
\caption{The convergence of the loss function, the LV-BP predicted $\Delta G$s, and the similarity score to the reference molecule during a molecule generation course. In this example, the $\Delta G$s of the seed molecule and the reference molecule are -6.81 kcal/mol and -12.02 kcal/mol, respectively. To force generated molecules evolving towards the reference molecule, we set the similarity score objective and the $\Delta G$ objective to be 1.00 and -12.02 kcal/mol, respectively.}
\label{fig:pathway-loss}
\end{figure}
\begin{figure}[!htp]
\centering
\includegraphics[width=0.8\textwidth]{gen-pathway_reduced.png}
\caption{A series of generated molecules over the evolution from the seed to the reference molecule.}
\label{fig:gen-pathway}
\end{figure}
Here we conduct our GNC to generate new molecules and analyze how these new molecules evolve over a generation course. We start this experiment with a seed molecule picked from the BACE1 dataset; this molecule is far from active with the binding free energy ($\Delta G$) = -6.81 kcal/mol. The reference molecule is also from the BACE1 dataset, it is highly active with $\Delta G$ = -12.02 kcal/mol. The binding affinity objective $y_{\Delta G}$ is set to be -12.02 kcal/mol, and the similarity score to the reference molecule is targeted to $y_{\rm sim}=1.0$.
Figure \ref{fig:pathway-loss}a depicts the loss function values computed at every epoch; Figure \ref{fig:pathway-loss}b and c illustrate the LV-BP predicted binding affinities and similarity scores, respectively. From these figures, one can observe that our GNC produces new potential BACE1 inhibitors with desirable binding affinities in less than 3000 epochs.
Figure \ref{fig:gen-pathway} shows a series of generated molecules over the evolution from the seed to the reference molecule. The starting point is the seed molecule, its binding affinity and similarity score to the reference molecule are as low as -6.81 kcal/mol and 0.01, respectively. By receiving the feedback from the gradient descent in the generator, the similarity score gradually rises to 1.0. The improvement to the binding affinity is even faster: while a created molecule has a similarity score of 0.28, its LV-BP predicted $\Delta G$ already reaches -11.30 kcal/mol; while the similarity score is 0.90, the LV-BP predicted $\Delta G$ is -12.00 kcal/mol, which is essentially the same as the reference molecule's $\Delta G$ of -12.02 kcal/mol.
\subsubsection{Reliability test on the designed BACE1 inhibitors}\label{sec:BACE1-reliablity-test}
Using our GNC, we also generate millions of compounds targeting a wide range of binding affinities and similarity scores. Then the prediction reliability with these different ranges of binding affinities and similarity scores is tested. Individually, the similarity score objectives, $y_{\rm sim}$, vary from 0.50 to 0.95 with an increment of 0.025; the binding affinity objectives, $y_{\Delta G}$, receive values from -9.6 kcal/mol to -13.1 kcal/mol with an increment of -0.25 kcal/mol. Here we select -9.6 kcal/mol as the starting point since this value is a widely accepted threshold to identify active compounds; the endpoint of $\Delta G$ = -13.1 kcal/mol is the highest binding affinity value in the BACE1 dataset (see Figure \ref{fig:bace_dist}).
\begin{figure}[!htb]
\centering
\includegraphics[width=0.4\textwidth]{filtered-bace_set-dist.png}
\caption{The experimental $\Delta G$ distribution of the filtered BACE1 dataset.}
\label{fig:bace_dist}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=1.0\textwidth]{bace-reeval.png}
\caption{The reliability test to our GNC generator on the BACE1 dataset. The prediction reliability with different binding affinity objectives ($y_{\Delta G}$) and similarity score objectives ($y_{\rm sim}$) is tested. The discrepancies between the LV-BP and the 2DFP-BP predictions are evaluated by different criteria: (a) The RMSE; (b) The activate ratio; (c) The Pearson correlation; (d) The loss function values.}
\label{fig:bace-reeval}
\end{figure}
The reliability test is based on reevaluating the LV-BP predicted binding affinities by the 2DFP-BP. The reliability criteria are the RMSE, active ratio, and $R_p$ between the LV-BP and 2DFP-BP prediction. The heatmaps in Figure \ref{fig:bace-reeval} shows these evaluation metric values corresponding to different similarity score and binding affinity restraints. Few blank points are present in each heatmap due to no available generation meeting these specific restraints.
Figure \ref{fig:bace-reeval}a plots the RMSE metrics. It reveals the most reliable region, i.e., having low RMSE, is $y_{\rm sim}$ above 0.925, and the $y_{\Delta G}$ between -10.1 kcal/mol and -12.1 kcal/mol. This is expectable since machine learning models can render accurate predictions if generated structures are highly similar to the training data (see Figure \ref{fig:bace_dist}). Besides, a large population of training samples have $\Delta G$s between -10.1 kcal/mol and -12.1 kcal/mol, leading more reliable predictions to molecules generated inside this range. Outside this range, as both $y_{\Delta G}$ and $y_{\rm sim}$ decreasing, the RMSEs between the LV-BP and 2DFP-BP predictions increase. Specifically, if $y_{\Delta G} < -12.1$ kcal/mol and $y_{\rm sim} < 0.675$, the RMSEs are always over 3.2 kcal/mol.
Figure \ref{fig:bace-reeval}c depicts the $R_p$s between the LV-BP and 2DFP-BP predictions with respect to $y_{\Delta G}$ and $y_{\rm sim}$. Similar to the manner of the RMSE distribution, $-$12.1 kcal/mol $ \leq y_{\Delta G} \leq 10.1$ kcal/mol and $0.9\leq y_{\rm sim} \leq 0.95$ lead to the $R_p$ values consistently higher than 0.8.
The last component of our reliable analysis regards the loss function magnitudes of our GNC generator plotted in Figure \ref{fig:bace-reeval}d. In most cases, the loss function values are less than 0.05. However, in some special situations, our network cannot maintain the loss values lower than 0.15. At these points, we cannot find any generated molecules subject to the multi-property restraints simultaneously, which renders the blank spots in the criteria plots in Figures \ref{fig:bace-reeval}a, b, and c.
In summary, to generate molecules with reliable predictions, one should set the binding affinity objective $y_{\Delta G}$ in a region filled with a large population of training data. Besides, the similarity score restraint $y_{\rm sim}$ should be high. However, in some circumstances, the generated molecules that have high predicted affinities should also be included in further consideration.
\subsection{Designing alternative drug candidates}
In this section, we utilize our GNC to produce alternative drug-like molecules with high binding affinities to the existing drugs' targets, which provides effective libraries for further improvement or searching cheaper drug alternatives. This work discusses eight drugs and their targets with the information regarding the names, ChEMBL IDs, energies, etc. summarized in Table \ref{table:eight-drug-inf}. All of these drugs were approved by the FDA in the recent decade to treat critical diseases, especially a variety of cancers. Notably, the drug Ribociclib has two different targets (Cyclin-dependent kinase 4 and Cyclin-dependent kinase 6), so Ribociclib has two sets of $\Delta G$s and two sets of training compounds.
\subsubsection{The single-target drug: Ceritinib}
\paragraph{The statistics of the drug Ceritinib.} The brand name of Ceritinib (ChEMBL ID: CHEMBL2403108) is Zykadia. It was developed by Novartis and approved by the FDA in April 2014 to treat different types of non-small cell lung cancers. Ceritinib is extremely expensive, with the monthly cost of Ceritinib-based treatment in the US being approximately \$11,428.
Ceritinib inhibits the ALK tyrosine kinase receptor (ALK, CHEMBL ID: CHEMBL4247). The ChEMBL database provides 1407 molecules with experimental binding affinity labels to this target available. After going through the filter in our model, 1203 molecules are left to train our generator and predictor. Figure \ref{fig:chembl2403108-training-dis}a depicts the $\Delta G$ distribution of this training set; it unveils that, hundreds of training samples have their binding affinities close to or even higher than Ceritinib. The similarity score distribution in Figure \ref{fig:chembl2403108-training-dis}b indicates, the training set includes 355 samples with their similarity scores to Ceritinib over 0.3, 56 samples with such scores over 0.6. These promising analytical assessments enable our GNC to design potential inhibitors to the target ALK.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.7\textwidth]{chembl2403108-training-dis.png}
\caption{(a) The $\Delta G$ distribution of the filtered training set to the target ALK. The red bar indicates the interval containing the $\Delta G$ of Ceritinib. (b) The similarity scores of the other molecules in this set to Ceritinib.}
\label{fig:chembl2403108-training-dis}
\end{figure}
\paragraph{Designing new druglike molecules.}
Here we use Ceritinib as the reference molecule to design alternative Ceritinib drugs. Section \ref{sec:BACE1-reliablity-test} suggests, to generate new molecules with desirable properties, the binding affinity objectives should be inside a region with plenty of training data, and the similarity scores to the reference molecule $y_{\rm sim}$ should be restrained to be high. However, high similarity scores could lead to quite limited chemical space. Our solution to this drawback is, first, extend similarity score restraint to a broader range, then reevaluate generated compounds using the 2DFP-BP, and only pick the ones with low discrepancies between the LV-BP and 2DFP-BP predicted affinities.
Following this strategy, we set the similarity score restraints varying from 0.3 to 0.9 with an increment of 0.025; this is because more than 300 training samples have their similarity scores over 0.3, which is supportive to generate compounds in this similarity range. The binding affinities are aimed at the interval from -10.5 kcal/mol to -12.25 kcal/mol with an increment of -0.25 kcal/mol; this binding affinity region covers hundreds of training samples as well as the drug Ceritinib itself. After reevaluating by the 2DFP-BP, any generated molecules with the relative errors between the LV-BP and 2DFP-BP predicted affinities over 5 \% are thrown away.
\paragraph{The 2DFP-BP reevaluation.} The details of the 2DFP-BP are offered in Section \ref{sec:2DFP-BP}. With Ceritinib as the reference, our GNC model creates 1095 novel active drug-like molecules, upon eliminating the ones with high discrepancies in their LV-BP and 2DFP-BP predictions, 629 molecules are left. Figure \ref{fig:corr-2d-dl-2403108} indicates, for these 629 molecules, the correlation between the two predictions is quite promising, with the RMSE = 0.28 kcal/mol, $R_p$ = 0.80, and active ratio = 0.95, respectively. This statistical information endorses the drug-likeness potential of our AI-generated molecules.
\begin{figure}[h]
\centering
\includegraphics[width=0.55\textwidth]{corr-2d-dl-2403108.png}
\caption{The correlation plot between the LV-BP and 2DFP-BP predicted $\Delta G$s of the generated molecules to the target ALK, the ones having high relative errors between the two predictions (>5 \%) are already eliminated.}
\label{fig:corr-2d-dl-2403108}
\end{figure}
The LV-BP and 2DFP-BP predicted binding affinities of these 629 molecules are also averaged, and their distribution is shown in \ref{fig:chembl2403108-gen-dis}a. This figure reveals the preferred affinities of the generated compounds is from -9.8 kcal/mol to -10.8 kcal/mol, which is also the most popular affinity region of the training samples. Figure \ref{fig:chembl2403108-gen-dis}b illustrates their similarity score distribution to the reference drug Ceritinib.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.8\textwidth]{chembl2403108-gen-dis.png}
\caption{The average predicted binding affinities to the target ALK of the generated molecules and their similarity scores to the reference drug Ceritinib. (a) The distribution of the averages of the LV-BP and 2DFP-BP predicted $\Delta G$s to the target ALK. (b) The similarity score distribution to the reference drug Ceritinib.}
\label{fig:chembl2403108-gen-dis}
\end{figure}
\paragraph{Top 6 drug candidates.} Ranked by the average predicted binding affinities of these 629 molecules, we select the top 6 drug candidates. Their 2D draws are plotted in Figure \ref{fig:chembl2403108-top6}. The relative errors between their LV-BP and 2DFP-BP predictions are 1.3 \%, 3.1 \%, 0.1 \%, 4.0 \%, 3.5 \%, 3.9 \%, respectively. It is delighted to see that their average predicted affinities are much higher than that of the reference drug Ceritinib. Moreover, these candidates have similarity scores to the reference drug from 0.54 to 0.69.
They are brand new and do not exist in the SciFinder database. We also calculate their log P values and synthetic accessibility scores (SAS) using RDKit, log S values by Alog PS 2.1 \cite{tetko2005virtual}, and report them in Figure \ref{fig:chembl2403108-top6}; these values are comparable to that of the reference drug and in reasonable ranges.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.85\textwidth]{chembl2403108-top6.png}
\caption{The drug Ceritinib and its top 6 generated molecules. The predicted $\Delta G$s to the target ALK, similarity scores (Sim) to the drug, calculated log P, log S values, and synthetic accessibility scores (SASs) are also present.}
\label{fig:chembl2403108-top6}
\end{figure}
\paragraph{The interaction and pharmacophore analysis.} One can observe from Figure \ref{fig:chembl2403108-top6}, even though the similarity scores are only between 0.56 and 0.69, these generated compounds share some moieties with the reference drug Ceritinib. This designates these common moieties to possibly involve critical interactions with the binding site of the target.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{interaction-pharm-chem2403108.png}
\caption{The pharmacophore analysis to the ALK tyrosine kinase's binding site. (a) The interaction plot between the drug Ceritinib and ALK tyrosine kinase from the 3D experimental structure with the PDB ID 4MKC. (b) The 3D alignments of the common pharmacophores obtained from all the active compounds to the target ALK with the 3D experimental structure of the drug Ceritinib, the top 6 generated molecules are also aligned to it.}
\label{fig:interaction-pharm-chem2403108}
\end{figure}
To verify our generated compounds still contain critical moieties to the binding, from experimental structures, we investigate the drug-target interactions, as well as the common pharmacophores among all the active compounds to the target. Figure \ref{fig:interaction-pharm-chem2403108}a shows the interaction details between the drug Ceritinib and its target in the 3D crystal structure of their complex (PDB ID 4MKC \cite{friboulet2014alk}). It reveals their interactions include, one hydrogen bond with one N atom in the pyrimidine of the drug (marked by 1 in the Figure \ref{fig:interaction-pharm-chem2403108}a) as the acceptor, one hydrogen bond with one N atom in the chain of the drug (marked by 2 in the Figure \ref{fig:interaction-pharm-chem2403108}a) as the donor, and another hydrogen bond with one O atom in the drug as the acceptor, one hydrophobic interaction between one benzene ring (marked by 3 in the Figure \ref{fig:interaction-pharm-chem2403108}a) and the target, as well as other hydrophobic interactions.
The common pharmacophore analysis in Figure \ref{fig:interaction-pharm-chem2403108}b is consistent with the interaction analysis. The N atom in the pyrimidine of the drug (marked by 1 in the Figure \ref{fig:interaction-pharm-chem2403108}a and \ref{fig:interaction-pharm-chem2403108}b) and the N atom in the chain of the drug (marked by 2 in the Figure \ref{fig:interaction-pharm-chem2403108}a and \ref{fig:interaction-pharm-chem2403108}b)) are critical pharmacophores, they plays the roles of an acceptor and a donor of hydrogen bonds, respectively. Another pharmacophore is the benzene ring forming a hydrophobic interaction with the target. The pharmacophore analysis also reveals more potential interaction modes, such as the hydrophobic interaction between the Cl atom and target, and the hydrogen bond with the other N atom in the chain of the drug as its donor.
As illustrated in Figure \ref{fig:interaction-pharm-chem2403108}b, all these critical pharmacophores are retained in our top 6 generated compounds, which strongly supports that these compounds are potential inhibitors to the target.
\subsubsection{The double-target drug: Ribociclib}
\paragraph{The statistics of the drug Ribociclib.} Here, we test the generative power of our GNC on multi-target drugs. In this case study, the multi-property restraints consist of multiple binding affinity criteria and the similarity score to a reference molecule. The drug we test here is Ribociclib (ChEMBL ID: CHEMBL3545110, brand name: Kisqali). It was developed by Novartis and Astex Pharmaceuticals and approved by the FDA in 2017 to treat certain kinds of breast cancers. The monthly cost of Ribociclib treatment is \$10,950 in the US.
Ribociclib inhibits two different targets, namely the Cyclin-dependent kinase 4 (CDK4, CHEMBL ID: CHEMBL331) and Cyclin-dependent kinase 6 (CDK6, CHEMBL ID: CHEMBL2508). In the ChEMBL database, 919 molecules have CDK4 binding data, 289 molecules have CDK6 binding data. After filtered out, 918 and 289 molecules are retained, respectively, providing a small training set to the CDK6. Figure \ref{fig:chembl3545110-training-dis}a and b plot the $\Delta G$ distributions of these two training sets. It reveals, in both the CDK4 and CDK6 sets, hundreds of samples have binding affinities close to or even higher than Ribociclib. Figure \ref{fig:chembl3545110-training-dis}c and d show the similarity score distributions to Ribociclib of the two training sets. Both these sets include more than 200 samples with the similarity scores to the drug over 0.3, more than 60 samples with such scores over 0.6.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.8\textwidth]{chembl3545110-training-dis.png}
\caption{(a) The experimental $\Delta G$ distribution of the training set to the target CDK4 with the interval containing the $\Delta G$ of the drug Ribociclib to CDK4 marked in red. (b) The experimental $\Delta G$ distribution of the CDK6 training set with the interval containing the $\Delta G$ of Ribociclib to CDK4 in red. (c) The similarity score distribution of the other samples in the CDK4 training set to Ribociclib. (d) The similarity score distribution of the other samples in the CDK6 set to Ribociclib.}
\label{fig:chembl3545110-training-dis}
\end{figure}
\paragraph{The multitask predictor.}
Since the CDK6 training set is small, it is challenging to train an accurate predictor for this target. However, the two targets, CDK4 and CDK6, are similar due to the calculation via SWISS-MODEL \cite{schwede2003swiss} with the sequence identity being 71.1\%. Therefore, multitask deep learning can enhance reliability.
In our multitask architecture, the binding affinity predictor in our generator offers two outputs, each for one of the two targets, so the predictor is trained by the two training sets simultaneously. As a result, in a 5-fold crossing-validation test, the multitask model significantly improves the performance on the small dataset.
\begin{table}[!htb]
\centering
\begin{tabular}{|l|l|l|l|l|}
\hline
\rowcolor{Lightgreen}
\multirow{2}{*}{} & \multicolumn{2}{l|}{Target CDK4} & \multicolumn{2}{l|}{Target CDK6} \\
\cline{2-5}
\rowcolor{Lightgreen}
& Single task & Multitask & Single task & Multitask \\ \hline
LV-BP & 0.791 & 0.804 & 0.524 & 0.811 \\ \hline
2FP-BP & 0.824 & 0.836 & 0.485 & 0.779 \\ \hline
\end{tabular}
\caption{The $R_p$s of $\Delta G$ predictions from the 5-fold cross-validation tests on the two targets of the drug Ribociclib by the single task and multitask predictors.}
\label{table:5-fold_ST_MT}
\end{table}
As illustrated in Table \ref{table:5-fold_ST_MT}, the target CDK4 with a 918-molecule training set does not benefit so much from the multitask. However, to the target CDK6 only with a 289-molecule training set, the improvement is dramatic: the $R_p$s rise from 0.524 to 0.811 by the LV-BP and from 0.485 to 0.779 by the 2FP-BP. These results demonstrate the efficiency of the multitask architecture.
\paragraph{The generation of new druglike molecules.}
To design alternative Ribociclib drugs in broader chemical space, we set the similarity score restraints to Ribociclib from 0.30 to 0.90, with an increment of 0.025. The binding affinities to the target CDK4 are aimed at the interval between -10.80 kcal/mol and -12.00 kcal/mol with an increment of -0.2 kcal/mol; this interval covers the $\Delta G$s of Ribociclib as well as lots of other training samples. For the same reason, the objectives of the binding affinities to the target CDK6 are set from -10.2 kcal/mol to -11.0 kcal/mol with an increment of -0.2 kcal/mol. Totally, following this scheme, we create 1080 novel molecules.
\paragraph{The reevaluation by 2DFP-BP.}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.9\textwidth]{corr-2d-dl-two-targets.png}
\caption{The correlations between the LV-BP and 2DFP-BP predicted $\Delta G$s of the generated molecules to the targets CDK4 and CDK6, the ones having high relative errors between the two predictions (>5 \%) are already eliminated.}
\label{fig:corr-2d-dl-two-targets}
\end{figure}
The predicted bind affinities of these 1080 generated compounds are reevaluated by the 2DFP-BP model incorporated with the multitask DNN. The parameters of this architecture are introduced in Section \ref{sec:param-multi}.
After excluding the ones with high discrepancies between the LV-BP and 2DFP-BP predictions, 271 molecules are left.
Figure \ref{fig:corr-2d-dl-two-targets} depicts the correlation plots between their LV-BP and 2DFP-BP predicted $\Delta G$s to the two targets. The correlations of the $\Delta G$s to both the two targets are promising. Specifically, the $\Delta G$s to the target CDK4 can achieve an RMSE of 0.29 kcal/mol, a $R_p$ of 0.69, and an active ratio as high as 0.98; the $\Delta G$s to the target CDK6 also have an RMSE of 0.27 kcal/mol, a $R_p$ of 0.65 and an active ratio of 0.78.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.8\textwidth]{chembl3545110-gen-dis.png}
\caption{The average predicted $\Delta G$ to the targets CDK4 and CDK6 of the generated molecules and their similarity scores to the reference drug Ribociclib: (a, b) The distributions of the averages of the LV-BP and 2DFP-BP predicted $\Delta G$s to the two targets, respectively. (c) The similarity score distribution to the reference drug Ribociclib.}
\label{fig:chembl3545110-gen-dis}
\end{figure}
Their LV-BP and 2DFP-BP predicted binding affinities to the two targets are also averaged, and the distributions are shown in Figure \ref{fig:chembl3545110-gen-dis}a,b. Figure \ref{fig:chembl3545110-gen-dis}c illustrates their similarity score distribution to the reference drug Ribociclib.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.7\textwidth]{chembl3545110-top6.png}
\caption{The drug Ribociclib and its top 6 generated molecules. The predicted $\Delta G$s to the targets CDK4 and CDK6, similarity scores (Sim) to the drug, calculated log P, log S values, and synthetic accessibility scores (SASs) are also reported.}
\label{fig:chembl3545110-top6}
\end{figure}
\paragraph{Top 6 drug candidates.}
According to the average predicted binding affinities of these 271 molecules, we select the top 6 drug candidates. Figure \ref{fig:chembl3545110-top6} represents their 2D plots. To the target CDK4, the relative errors between their LV-BP and 2DFP-BP predictions are 4.6 \%, 3.0 \%, 3.2 \%, 0.2 \%, 1.6 \% and 0.7 \%, respectively; to CDK6, the relative errors between the two predictions are 1.7 \%, 1.2 \%, 3.7 \%, 3.4 \%, 4.8 \% and 1.7 \%, respectively. Most of them have better binding affinities than the reference drug Ribociclib. Such as, to CDK4, the affinities of the first and fourth compounds are predicted to be higher than that of Ribociclib, the other three have similar ones with Ribociclib; to CDK6, all the top 6 candidates have better binding affinities. Moreover, their similarity scores to the reference drug are between 0.65 and 0.75. They are not in the SciFinder database, which means these six candidates are novel. The log P, log S values and synthetic accessibility scores (SASs) calculated by RdKit and Alog PS are also showed in the figure; their log P, log S values and SASs are comparable to that of the reference drug.
\begin{figure}[!htb]
\centering
\includegraphics[width=1.0\textwidth]{interaction-pharm-3545110-target2_reduced.png}
\caption{(a) The illustration of the interactions between the drug Ribociclib and the target CDK6 extracted from the experimental structure (PDB ID 5L2T). (b) The 3D alignment of the common pharmacophores obtained from all the active compounds to the target CDK6 with the 3D experimental structure of the drug Ribociclib, the top 6 generated molecules are also aligned to it.}
\label{fig:interaction-pharm-3545110-target2}
\end{figure}
\paragraph{The interaction and pharmacophore analysis.}
Figure \ref{fig:interaction-pharm-3545110-target2}(a) shows the interaction details between the drug Ribociclib and the target CDK6 in the 3D crystal structure of their complex (PDB ID 5L2T \cite{chen2016spectrum}). It indicates the main interactions are hydrogen bonds and hydrophobic interactions. The pharmacophore analysis in Figure \ref{fig:interaction-pharm-3545110-target2}(b) reveals that the critical pharmacophores are the pyrrolopyrimidine, pyridine, and cyclohexane, which can form hydrogen bonds and hydrophobic interactions with the binding site.
Among the top 6 candidates, except the first one, the others contain all these critical pharmacophores; the first one has most of them. This suggests all the six compounds are potential inhibitors to the targets.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.8\textwidth]{corr-other-drugs_reduced.png}
\caption{The correlation plots between the LV-BP and 2DFP-BP predicted $\Delta G$s of the generated molecules to the six drug targets, the ones with high relative errors between the two predictions are already eliminated.}
\label{fig:corr-other-drugs}
\end{figure}
\subsubsection{The tests on the other single-target drugs}
We also apply our GNC to the rest six drugs listed in Table \ref{table:eight-drug-inf} and design novel drug candidates. The similarity score restraints are from 0.30 to 0.90, with an increment of 0.025. The $\Delta G$ objective ranges are chosen to contain the $\Delta G$s of the drugs as well as lots of other training samples. We also only collect the generated molecules with the relative errors between the LV-BP and 2DFP-BP predicted $\Delta G$s below 5 \%.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.9\textwidth]{detg-dis-others.png}
\caption{The sizes and experimental $\Delta G$ distributions of the training sets to the six drug targets. The red bars indicate the intervals containing the $\Delta G$s of the six reference drugs.}
\label{fig:detg-dis-others}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.9\textwidth]{sim-dis-others.png}
\caption{ The similarity score distributions of other molecules in the six training sets to the reference drugs.}
\label{fig:sim-dis-others}
\end{figure}
Figure \ref{fig:corr-other-drugs} indicates, since the ones with high discrepancies between the two predictions are eliminated, our generated compounds to all these six drug targets have promising correlations.
Now, we focus on whether highly active drug candidates are created or not. This relies on the binding affinity and similarity score distributions of the training sets. Figure \ref{fig:detg-dis-others} provides the sizes and the $\Delta G$ distributions of these six target-specified training sets; the red bars indicate the intervals containing the $\Delta G$s of the reference drugs. Figure \ref{fig:sim-dis-others} plots their similarity score distributions.
Figures \ref{fig:detg-dis-others}a shows that the training set of the drug Acalabrutinib contains 1459 compounds. Among them, more than 400 compounds have higher binding affinities than that of the drug, which is beneficial to generate compounds more potent than the drug. Also, Figure \ref{fig:sim-dis-others}a reveals that in the training set, over 500 molecules have the similarity scores to the drug over 0.3, which is quite helpful to generate novel compounds with similarity scores also in this range. These promising statistical information helps to explain why our GNC can create as many as 879 new compounds for this drug target with high confidence, and 17 of them have higher or similar binding affinities with that of Acalabrutinib; the best $\Delta G$ is -11.90 kcal/mol (see Figures S1 and S3).
As revealed by \ref{fig:detg-dis-others}b, the training set of the drug Idelalisib is larger than that of Acalabrutinib; and over 500 molecules in this set have higher binding affinities than that of Idelalisib. As a result, 73 of the 794 generated compounds are more potent than Idelalisib, with the best $\Delta G$ being -11.27 kcal/mol (see Figures S1 and S4).
Figure \ref{fig:detg-dis-others}c exhibits that the dataset of Macimorelin is quite small. However, the correlation of the generated molecules is still satisfactory, and 13 generated molecules have similar $\Delta G$s with that of the drug (see Figure S1). The reason is, its training set contains 205 compounds with higher binding affinities than that of the drug, and also 182 compounds with the similarity scores to the drug over 0.3.
Figures \ref{fig:detg-dis-others}d and e depict the datasets of the drugs Dabrafenib and Enzalutamide. The affinity values of these two drugs are close to the upper boundaries of their training sets' affinity domains. In other words, few molecules in their training sets are more active than the drugs. Since the interpolation nature of machine learning models tends to provide reliable predictions inside the populated range, it is tough to generate compounds more potent than these drugs. Although their training set sizes and similarity scores are favorable, the numbers of generated compounds with better binding affinities than that of the drugs are only 2 and 4, respectively (see Figures S1, S6 and S7).
We perform our last experiment on the drug Panobinostat. Panobinostat is an extreme example: as illustrated in Figure \ref{fig:detg-dis-others}f, its binding affinity is the highest in its training set. Therefore, although our GNC model generates 319 molecules, the top active one among them only reaches a $\Delta G$ of -11.56 kcal/mol, which is far less potent than the drug itself ($\Delta G$=-12.46 kcal/mol).
In the supporting information, Figures S3 to S8 provide the top six generated alternative compounds for the drugs Acalabrutinib, Idelalisib, Macimorelin, Dabrafenib, Enzalutamide, and Panobinostat, as well as their predicted $\Delta G$s.
\section{Discussions}\label{sec:discussions}
With the availability of deep learning technologies, more and more {\it in silico} molecule generation models have been proposed. These models can be classified into three categories: randomized output, controlled output, and optimized output \cite{grow2019generative}. One of the challenges is how to generate new molecules with desired chemical properties, especially drug-like molecules. Another challenge is how to improve the utility of {\it in silico} molecule generation without direct experimental validations.
To address these challenges, we propose a new GNC to generate drug-like molecules based on multi-property optimizations via gradient descent.
\subsection{The essential circumstances for reliable and desired molecule generation}
Based on our experiments, there are two essential circumstances to generate molecules with reliable and desired predicted chemical properties:
\begin{enumerate}
\item An objective property value should always be in a region with lots of training samples. Based on the nature of machine learning methods, the predictor can be built with high accuracy at an objective value if lots of training samples around it. It can be seen from Section \ref{sec:BACE1-reliablity-test} that when a binding affinity objective is in the middle of the training set's distribution, the latent space generator can always generate novel molecules with reliable predictions confirmed by the 2DFP-BP. However, when an objective is close to or even at the edge of the training-set distribution, one may still generate many novel compounds, but the predictions to them are quite risky, such as high discrepancies between the LV-BP and 2DFP-BP predictions.
\item Generated compounds should have some high similarity scores to some existing molecules in the training set. If some molecules (not necessarily reference molecules) in the training set are similar to generated compounds, then the predictions are reliable and can be verified by the 2DFP-BP.
\end{enumerate}
\subsection{The necessity of both similarity and property value restraints}
The two points above also explain why both similarity score and property value restraints should be included in our generator.
The goal of property restraint is two-fold. First, it restricts property values to desired ones. Additionally, it can be used to achieve high reliability. As discussed above, if the property value is limited to a populated region of the training-set distribution, the resulting generated molecules will most likely have reliable predictions.
Similarity restraint is also to ensure prediction accuracy. High similarity scores to existing molecules make predictions more accurate and reliable. Moreover, in a regular drug discovery procedure for a given target, one typically starts from some lead compounds or even current drugs and then carries out lead optimizations to make candidates more "drug-like", e.g., higher activity and lower side effects \cite{hughes2011principles}. Therefore, similarly, in a drug-specified generative model, the similarity to a reference compound or drug must be controlled to guarantee that new drug-like compounds still bind to the target. For example, with similarity restraint, generated molecules share pharmacophores with the reference drug.
\subsection{Multiple property restraints}
Drug design is sophisticated. To develop a drug, plenty of properties must be carefully studied, such as binding affinity, toxicity, partition coefficient (log P), aqueous solubility (log S), off-target effect. The failure in any one of these properties can prevent drug candidates from the market. In other words, drug design is a multi-property optimization process.
Technically, our generator can handle this multi-property optimization. In our framework, the restraint to each property is realized by one term in the loss function. Therefore, multi-property optimization can be satisfied simultaneously in our GNC. In this work, multiple property restraints are tested on one drug with two targets (Ribociclib). It turns out, the generated new candidates have ideal binding affinities to the two targets simultaneously; this means, our model can work on the multiple property restraints. Multiple properties can also be specified as toxicity, log P, log S, etc. To avoid side effects, one can also control drug candidates to have a high binding affinity to one target but low binding affinities to other targets.
\section{Conclusion} \label{sec:conclusion}
Searching alternative drugs is important for improving the quality of existing drugs and making new drugs cheaply available to low-income people.
In this work, we develop a new generative network complex (GNC) for automated generation of drug-like molecules based on the multi-property optimization via gradient descent in the latent space. In this new GNC, multiple chemical properties, particularly binding affinity and similarity score, are optimized to generate new molecules with desired chemical and drug properties.
To ensure the prediction reliability of these new compounds, we reevaluate them by independent 2D-fingerprint based predictors. Molecules without consistent predictions from the latent-vector model and the 2D-fingerprint model are not accepted. After the consistent check, hundreds of potential potent drug candidates are reported. Performed on a supercomputer, the generation process from one seed to a large number of new molecules takes less than 10 minutes. Therefore, our GNC is an efficient new paradigm for discovering new drug candidates. To demonstrate the utility of the present GNC, we first test its reliability on the BACE1 target and then, further apply this model to generate thousands of new alternative drug candidates for a few market existing drugs, namely, Ceritinib, Ribociclib, Acalabrutinib, Idelalisib, Dabrafenib, Macimorelin, Enzalutamide, and Panobinostat.
We also discuss the keys to generate drug-like candidates with reliable predictions. First, an objective property value should be in a populated region of the training-set distribution. Second, generated molecules need to have good similarity scores to some existing compounds in the training set.
\vspace{1cm}
\section*{Supplementary materials}
Figures S1 to S8 are the $\Delta G$ and similarity score distributions of the generated drug-like molecules for the drugs Acalabrutinib, Idelalisib, Dabrafenib, Macimorelin, Enzalutamide, and Panobinostat, as well as the 2D plots and predicted $\Delta G$s of the top 6 among them.
\vspace{1cm}
\section*{Acknowledgments}
This work was supported in part by NSF Grants DMS-1721024, DMS-1761320, and IIS1900473, NIH grant GM126189, and Michigan Economic Development Corporation. DDN and GWW are also funded by Bristol-Myers Squibb and Pfizer.
\clearpage
\vspace{1cm}
| e2ef643ab7c9db8e47cbfe457c39becc848dcb0b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The main aim of this work is to establish large deviation principle and small time asymptotics for the stochastic Navier-Stokes equation with anisotropic viscosity. We consider the following stochastic Navier-Stokes equation with anisotropic viscosity on the two dimensional (2D) torus $\mathbb{T}^2=\mathbb{R}^2/(2\pi\mathbb{Z})^2$:
\begin{equation}\label{equation before projection}\aligned
&du=\partial^2_1 udt-u\cdot \nabla u dt+\sigma(t, u)dW(t)-\nabla pdt,\\
&\text{div } u=0,\\
&u(0)=u_0,
\endaligned
\end{equation}
where $u(t,x)$ denotes the velocity field at time $t\in[0,T]$ and position $x\in\mathbb{T}^2$, $p$ denotes the pressure field, $\sigma$ is the random external force and $W$ is an $l^2$-cylindrical Wiener process.
Let's first recall the classical Navier-Stokes (N-S) equation which is given by
\begin{equation}\label{equation classical}\aligned
&du=\nu\Delta udt-u\cdot \nabla u dt-\nabla pdt,\\
&\text{div } u=0,\\
&u(0)=u_0,
\endaligned
\end{equation}
where $\nu>0$ is the viscosity of the fluid. (\ref{equation classical}) describes the time evolution of an incompressible fluid. In 1934, J. Leray proved global existence of finite energy weak solutions for the deterministic case in the whole space $\mathbb{R}^d$ for $d=2, 3$ in the seminar paper \cite{L33}. For more results on deterministic N-S equation, we refer to \cite{CKN86}, \cite{Te79}, \cite{Te95}, \cite{KT01} and reference therein. For the stochastic case, there exists a great amount of literature too. The existence and uniqueness of solutions and ergodicity property to the stochatic 2D Navier-Stokes equation have been obtained (see e.g. \cite{FG95},
\cite{MR05}, \cite{HM06}). Large deviation principles for the two-dimensional stochastic N-S equations have been established in \cite{CM10} and \cite{SS06}.
Compared to \eqref{equation classical}, \eqref{equation before projection} only has partial dissipation, which can be viewed as an intermediate equation between N-S equation and Euler equation. System of this type appear in in geophysical fluids (see for instance \cite{CDGG06} and \cite{Ped79}). Instead of putting the classical viscosity $-\nu\Delta$ in (\ref{equation classical}), meteorologist often modelize turbulent diffusion by putting a viscosity of the form: $-\nu_h\Delta_h-\nu_3\partial^2_{x_3}$, where $\nu_h$ and $\nu_3$ are empiric constants, and $\nu_3$ is usually much smaller than $\nu_h$. We refer to the book of J. Pedlovsky \cite[Chapter 4]{Ped79} for a more complete discussion. For the 3 dimensional case there is no result concerning global existence of weak solutions.
In the 2D case, \cite{LZZ18} investigates both the deterministic system and the stochastic system (\ref{equation before projection}) for $H^{0,1}$ initial value (For the definition of space see Section 2). The main difference in obtaining the global well-posedness for \eqref{equation before projection} is that the $L^2$-norm estimate is not enough to establish $L^2([0,T], L^2)$ strong convergence due to lack of compactness. In \cite{LZZ18}, the proof is based on an additional $H^{0,1}$-norm estimate. In this paper, we want to establish the large deviation principles for the two-dimensional stochastic Navier-Stokes equations with anisotropic viscosity both for small noise and for short time.
The large deviation theory concerns the asymptotic behavior of a family of random variables $X_\varepsilon$ and we refer to the monographs \cite{DZ92} and \cite{Str84} for many historical remarks and extensive references. It asserts that for some tail or extreme event $A$, $P(X_\varepsilon \in A)$ converges to zero exponentially fast as $\varepsilon\rightarrow 0$ and the exact rate of convergence is given by the so-called rate function. The large deviation principle was first established by Varadhan in \cite{V66} and he also studied the small time asymptotics of finite dimensional diffusion processes in \cite{Var67}. Since then, many important results concerning the large deviation principle have been established. For results on the large deviation principle for stochastic differential equations in finite dimensional case we refer to \cite{FW84}. For the extensions to infinite dimensional diffusions or SPDE, we refer the readers to \cite{BDM08}, \cite{CM10}, \cite{DM09}, \cite{Liu09}, \cite{LRZ13}, \cite{RZ08}, \cite{XZ08}, \cite{Zh00} and the references therein.
We first study the small noise large deviations by using the weak convergence approach. This approach is mainly based on a variational representation formula for certain functionals of infinite dimensional Brownian Motion, which was established by Budhiraja and Dupuis in \cite{BD00}. The main advantage of the weak convergence approach is that one can avoid some exponential probability estimates, which might be very difficult to derive for many infinite dimensional models. To use the weak convergence approach, we need to prove two conditions in Hypothesis \ref{Hyp}. In \cite{Liu09} and \cite{LRZ13}, the authors use integration by parts and lead to some extra condition on diffusion coefficient. In \cite{CM10}, the authors use time discretization and require time-regularity of diffusion coefficient. In this paper, we use the argument in \cite{WZZ}, in which the authors prove a moderate deviation principle by this argument, i.e. we first establish the convergence in $L^2([0,T],L^2)$ and then by using this and It\^o's formula, $L^\infty([0,T], L^2)\bigcap L^2([0,T], H^{1,0})$ convergence can be obtained. By this argument, we can drop the extra condition on diffusion coefficient in \cite{Liu09} and \cite{CM10}.
For the small time asymptotics (large deviations) of the two-dimensional stochastic Navier-Stokes equations with anisotropic viscosity. This describes the limiting behaviour of the solution in time interval $[0,t]$ as $t$ goes to zero. Another motivation will be to get the following Varadhan identity through the small time asymptotics:
$$\lim_{t\rightarrow 0}2t\log P(u(0)\in B, u(t)\in C)=-d^2(B,C),$$
where $d$ is an appropriate Riemannian distance associated with the diffusion generated by the solutions of the two-dimensional stochastic Navier-Stokes equations with anisotropic viscosity. The small time asymptotics is also theoretically interesting, since the study involves the investigation of the small noise and the effect of the small, but highly nonlinear drift.
To prove the small time asymptotics, we follow the idea of \cite{XZ08} to prove the solution to (\ref{equation before projection}) is exponentially equivalent to the solution to the linear equation. The main difference compared to \cite{XZ08} is that similar to \cite{LZZ18} $L^2$-norm estimate is not enough due to less dissipation and we have to do $H^{0,1}$-norm estimate.
\vskip.10in
{\textbf{Organization of the paper}}
In Section 2, we introduce the basic notation, definition and recall some preliminary results. In Section 3, we will build the small noise large deviation principle. In Section 4, we prove the small time asymptotics for the the two-dimensional stochastic Navier-Stokes equations with anisotropic viscosity.
\vskip.10in
{\textbf{Acknowledgement}}
The authors would like to thank Rongchan Zhu and Siyu Liang for helpful discussions.
\section{Preliminary}
\vskip.10in
{\textbf{Function spaces on $\mathbb{T}^2$}}
We first recall some definitions of function spaces for the two dimensional torus $\mathbb{T}^2$.
Let $\mathbb{T}^2=\mathbb{R}/2\pi\mathbb{Z}\times\mathbb{R}/2\pi\mathbb{Z}=(\mathbb{T}_h, \mathbb{T}_v)$ where $h$ stands for the horizonal variable $x_1$ and $v$ stands for the vertical variable $x_2$. For exponents $p,q\in [1, \infty)$, we denote the space $L^p(\mathbb{T}_h, L^q(\mathbb{T}_v))$ by $L^p_h(L^q_v)$, which is endowed with the norm
$$\|u\|_{L^p_h(L^q_v)(\mathbb{T}^2)}:=\{\int_{\mathbb{T}_h}(\int_{\mathbb{T}_v}|u(x_1, x_2)|^qdx_2)^{\frac{p}{q}}dx_1\}^{\frac{1}{p}}.$$
Similar notation for $L^p_v(L^q_h)$. In the case $p, q=\infty$, we denote $L^\infty$ the essential supremum norm.
Throughout the paper, we denote various positive constants by the same letter $C$.
For $u\in L^2(\mathbb{T}^2)$, we consider the Fourier expansion of $u$:
$$u(x)=\sum_{k\in\mathbb{Z}^2}\hat{u}_ke^{ik\cdot x} \text{ } \text{with} \text{ } \hat{u}_k=\overline{\hat{u}_{-k}},$$
where $\hat{u}_k:=\frac{1}{(2\pi)^2}\int_{[0, 2\pi]\times[0, 2\pi]}u(x)e^{-ik\cdot x}dx$ denotes the Fourier coefficient of $u$ on $\mathbb{T}^2$.
Define the Sobolev norm:
$$\|u\|_{H^s}^2:=\sum_{k\in\mathbb{Z}^2}(1+|k|^2)^s |\hat{u}_k|^2,$$
and the anisotropic Sobolev norm:
$$\|u\|^2_{H^{s, s'}}=\sum_{k\in\mathbb{Z}^2}(1+|k_1|^2)^s(1+|k_2|^2)^{s'}|\hat{u}_k|^2,$$
where $k=(k_1, k_2)$. We define the Sobolev spaces $H^s(\mathbb{T}^2)$, $H^{s,s'}(\mathbb{T}^2)$ as the completion of $C^\infty(\mathbb{T}^2)$ with the norms $\|\cdot\|_{H^s}$, $\|\cdot\|_{H^{s,s'}}$ respectively.
To formulate the stochastic Navier-Stokes equations with anisotropic viscosity, we need the following spaces:
$$H:=\{u\in L^2(\mathbb{T}^2; \mathbb{R}^2); \text{div}\text{ } u=0\},$$
$$V:=\{u\in H^1(\mathbb{T}^2; \mathbb{R}^2); \text{div}\text{ } u=0\},$$
$$\tilde{H}^{s,s'}:=\{u\in H^{s,s'}(\mathbb{T}^2;\mathbb{R}^2); \text{div}\text{ } u=0\}.$$
Moreover, we use $\langle\cdot, \cdot\rangle $ to denote the scalar product (which is also the inner product of $L^2$ and $H$)
$$\langle u, v\rangle =\sum_{j=1}^2\int_{\mathbb{T}^2}u^j(x)v^j(x)dx$$
and $\langle\cdot, \cdot\rangle_X$ to denote the inner product of Hilbert space $X$ where $X=l^2$, $V$ or $\tilde{H}^{s,s'}$.
Due to the divergence free condition, we need the Larey projection operator $P_H: \text{ } L^2(\mathbb{T}^2)\rightarrow H$:
$$P_H: u\mapsto u-\nabla \Delta^{-1}(\text{div } u).$$
By applying the operator $P_H$ to (\ref{equation before projection})
we can rewrite the equation in the following form:
\begin{equation}\label{equation after projection}\aligned
&du(t)=\partial^2_1 u(t)dt-B(u(t))dt+\sigma(t, u(t))dW(t),\\
&u(0)=u_0,
\endaligned
\end{equation}
where the nonlinear operator $B(u,v)=P_H(u\cdot \nabla v)$ with the notation $B(u)=B(u,u)$. Here we use the same symbol $\sigma$ after projection for simplicity.
For $u,v,w\in V$, define
$$b(u,v,w):=\langle B(u,v), w\rangle.$$
We have $b(u,v,w)=-b(u,w,v)$ and $b(u,v,v)=0$.
We put some estimates of $b$ in the Appendix.
\vskip.10in
{\textbf{Large deviation principle}}
We recall the definition of the large deviation principle. For a general introduction to the theory we refer to \cite{DZ92}, \cite{DZ98}.
\begin{definition}[Large deviation principle]
Given a family of probability measures $\{\mu_\varepsilon\}_{\varepsilon>0}$ on a metric space $(E,\rho)$ and a lower semicontinuous function $I:E\rightarrow [0,\infty]$ not identically equal to $+\infty$. The family $\{\mu_\varepsilon\}$ is said to satisfy the large deviation principle(LDP) with respect to the rate function $I$ if\\
(U) for all closed sets $F\subset E$ we have
$$\limsup_{\varepsilon\rightarrow 0}\varepsilon\log\mu_\varepsilon(F)\leqslant-\inf_{x\in F}I(x),$$
(L) for all open sets $G\subset E$ we have
$$\liminf_{\varepsilon\rightarrow 0}\varepsilon\log\mu_\varepsilon(G)\geqslant-\inf_{x\in G}I(x).$$
A family of random variable is said to satisfy large deviation principle if the law of these random variables satisfy large deviation princple.
Moreover, $I$ is a good rate function if its level sets $I_r:=\{x\in E:I(x)\leqslant r\}$ are compact for arbitrary $r\in (0,+\infty)$.
\end{definition}
\vskip.10in
\begin{definition}[Laplace principle]
A sequence of random variables $\{X^\varepsilon\}$ is said to satisfy the Laplace principle with rate function $I$ if for each bounded continuous real-valued function $h$ defined on $E$
$$\lim_{\varepsilon\rightarrow 0}\varepsilon\log E\left[e^{-\frac{1}{\varepsilon}h(X^\varepsilon)}\right]=-\inf_{x\in E}\{h(x)+I(x)\}.$$
\end{definition}
\iffalse
It is well known that the large deviation principle and the Laplace principle are equivalent if $E$ is a Polish space and the rate function is good. The equivalence is essentially a consequence of Varadhan's lemma (\cite[Theorem 4.3.1]{DZ98}) and Bryc's converse theorem (\cite[Theorem 4.4.2]{DZ98}). Note that \cite{DZ98} proves that the equivalence holds when $E$ is a completely regular topological space.
\fi
Given a probabilty space $(\Omega,\mathcal{F},P)$, the random variables $\{Z_\varepsilon\}$ and $\{\overline{Z}_\varepsilon\}$ which take values in $(E,\rho)$ are called exponentially equivalent if for each $\delta>0$,
$$\lim_{\varepsilon\rightarrow 0}\varepsilon\log P(\rho(Z_\varepsilon,\overline{Z}_\varepsilon)>\delta)=-\infty.$$
\vskip.10in
\begin{lemma}[{\cite[Theorem 4.2.13]{DZ98}}]\label{EXEQ}
If an LDP with a rate function $I(\cdot)$ holds for the random variables $\{Z_\varepsilon\}$, which are exponentially equivalent to $\{\overline{Z}_\varepsilon\}$, then the same LDP holds for $\{\overline{Z}_\varepsilon\}$.
\end{lemma}
\vskip.10in
{\textbf{Existence and uniqueness of solutions}}
We introduce the precise assumptions on the diffusion coefficient $\sigma$. Given a complete probability space $(\Omega, \mathcal{F}, P)$ with filtration $\{\mathcal{F}_t\}_{t\geqslant 0}$. Let $L_2(l^2,U)$ denotes the Hilbert-Schmidt norms from $l^2$ to $U$ for a Hilbert space $U$. We recall the following conditions for $\sigma$ from \cite{LZZ18}:
(i) \textbf{Growth condition}
There exists nonnegative constants $K'_i$, $K_i$, $\tilde{K}_i$ ($i=0,1,2$) such that for every $t\in[0,T]$:
\vskip.10in
(A0) $\|\sigma(t,u)\|^2_{L_2(l^2, H^{-1})}\leqslant K_0'+K_1'\|u\|^2_{H}$;
(A1) $\|\sigma(t,u)\|^2_{L_2(l^2, H)}\leqslant K_0+K_1\|u\|^2_{H}+K_2\|\partial_1 u\|^2_H$;
(A2) $\|\sigma(t,u)\|^2_{L_2(l^2, H^{0,1})}\leqslant \tilde{K}_0+\tilde{K}_1\|u\|^2_{H^{0,1}}+\tilde{K}_2(\|\partial_1 u\|_H^2+\|\partial_1\partial_2u\|^2_H)$;
\vskip.10in
(ii)\textbf{Lipschitz condition}
There exists nonnegative constants $L_1, L_2$ such that:
\vskip.10in
(A3) $\|\sigma(t,u)-\sigma(t,v)\|^2_{L_2(l^2, H)}\leqslant L_1\|u-v\|^2_{H}+L_2\|\partial_1(u-v)\|^2_H$.
\vskip.10in
The following theorem from \cite{LZZ18} gives the well-posedness of equation (\ref{equation after projection}):
\begin{lemma}[{\cite[Theorem 4.1, Theorem 4.2]{LZZ18}}]\label{ex. and uniq. of solution}
Under the assumptions (A0)-(A3) with $K_2<\frac{2}{11}, \tilde{K}_2<\frac{2}{5}, L_2<\frac{2}{5}$, equation (\ref{equation after projection}) has a unique strong solution $u\in L^\infty([0,T], \tilde{H}^{0,1})\cap L^2([0, T], \tilde{H}^{1,1})\cap C([0,T], H^{-1})$ for $u_0\in\tilde{H}^{0,1}$.
\end{lemma}
\vskip.10in
{\textbf{A martingale lemma}}
The following remarkable result is from \cite{BY82} and \cite{D76}:
\begin{lemma}\label{martingale lemma}
There exists a universal constant $c$ such that, for any $p\geqslant 2$ and for all continuous martingale $(M_t)$ with $M_0=0$ and stopping times $\tau$,
$$\|M^*_\tau\|_p\leqslant cp^\frac{1}{2}\|\langle M\rangle^\frac{1}{2}_\tau\|_p,$$
where $M^*_t=\sup_{0\leqslant s\leqslant t}|M_s|$ and $\|\cdot\|_p$ stands for the $L^p$ norm with respect to the probability space.
\end{lemma}
\section{Large deviation principle}
In this section, we consider the large deviation principle for the stochastic Navier-Stokes equations with anisotropic viscosity. We will use the weak convergence approach introduced by Budhiraja and Dupuis in \cite{BD00}. First we recall it. The starting point is the equivalence between the large deviation principle and the Laplace principle. This result was first formulated in \cite{P93} and it is essentially a consequence of Varadhan's lemma \cite{V66} and Bryc's converse theorem \cite{B90}.
\begin{remark}
By \cite{DZ98} we have the the equivalence between the large deviation principle and the Laplace principle in completely regular topological spaces. In \cite{BD00} the authors give the weak convergence approach on a Polish space. Since the proof does not depend on the separability and the completeness, the result also holds in metric spaces.
\end{remark}
Let $\{W(t)\}_{t\geqslant 0}$ be a cylindrical Wiener process on $l^2$ w.r.t. a complete filtered probability space $(\Omega, \mathcal{F}, \mathcal{F}_t, P)$ (i.e. the path of $W$ take values in $C([0,T]; U)$, where $U$ is another Hilbert space such that the embedding $l^2\subset U$ is Hilbert-Schmidt). For $\varepsilon>0$, suppose $g^\varepsilon$: $C([0,T], U)\rightarrow E$ is a measurable map and $u^\varepsilon:=g^\varepsilon(W(\cdot))$. Let
$$\mathcal{A}:=\left\{v: v \text{ is }l^2\text{-valued }\mathcal{F}_t\text{-predictable process and }\int^T_0\|v(s)(\omega)\|^2_{l^2}ds<\infty\text{ a.s.}\right\},$$
$$S_N:=\left\{\phi\in L^2([0,T],l^2):\text{ }\int^T_0\|\phi(s)\|^2_{l^2}ds\leqslant N\right\},$$
$$\mathcal{A}_N:=\left\{v\in\mathcal{A}:\text{ } v(\omega)\in S_N\text{ P-a.s.}\right\}.$$
Here we will always refer to the weak topology on $S_N$ in the following if we do not state it explicitly.
Now we formulate the following sufficient conditions for the Laplace principle of $u^\varepsilon$ as $\varepsilon\rightarrow 0$.
\begin{hyp}\label{Hyp}
There exists a measurable map $g^0: C([0,T], U)\rightarrow E$ such that the following two conditions hold:\\
1. Let $\{v^\varepsilon:\varepsilon>0\}\subset \mathcal{A}_N$ for some $N<\infty$. If $v^\varepsilon$ converge to $v$ in distribution as $S_N$-valued random elements, then
$$g^\varepsilon\left(W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds\right)\rightarrow g^0\left(\int^\cdot_0v(s)ds\right)$$
in distribution as $\varepsilon\rightarrow 0$.\\
2. For each $N<\infty$, the set
$$K_N=\left\{g^0\left(\int^\cdot_0\phi(s)ds\right): \phi\in S_N\right\}$$
is a compact subset of $E$.
\end{hyp}
\begin{lemma}[{\cite[Theorem 4.4]{BD00}}]\label{weak convergence method}
If $u^\varepsilon=g^\varepsilon(W)$ satisfies the Hypothesis \ref{Hyp}, then the family $\{u^\varepsilon\}$ satisfies the Laplace principle (hence large deviation principle) on $E$ with the good rate function $I$ given by
\begin{equation}
I(f)=\inf_{\{\phi\in L^2([0,T], l^2):f=g^0(\int^\cdot_0\phi(s)ds)\}}\left\{\frac{1}{2}\int^T_0\|\phi(s)\|^2_{l^2}ds\right\}.
\end{equation}
\end{lemma}
Consider the following equation:
\begin{equation}\label{LDP eq.}\aligned
&du^\varepsilon(t)=\partial^2_1 u^\varepsilon(t)dt-B(u^\varepsilon(t))dt+\sqrt{\varepsilon}\sigma(t, u^\varepsilon(t))dW(t),\\
&u^\varepsilon(0)=u_0.
\endaligned
\end{equation}
By Lemma \ref{ex. and uniq. of solution}, under the assumptions (A0)-(A3), (\ref{LDP eq.}) has a unique strong solution $u^\varepsilon\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})\bigcap C([0,T], H^{-1})$ for $u_0\in \tilde{H}^{0,1}$. It follows from Yamada-Watanabe theorem (See \cite[Appendix E]{LR15}) that there exists a Borel-measurable function
$$g^\varepsilon: C([0,T],U)\rightarrow L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$$
such that $u^\varepsilon=g^\varepsilon(W)$ a.s..
Let us introduce the following skeleton equation associated to (\ref{LDP eq.}), for $\phi\in L^2([0,T], l^2)$:
\begin{equation}\label{skeleton eq.}\aligned
&dz^\phi(t)=\partial^2_1 z^\phi(t)dt-B(z^\phi(t))dt+\sigma(t, z^\phi(t))\phi(t)dt,\\
&\text{div }z^\phi=0,\\
&z^\phi(0)=u_0.
\endaligned
\end{equation}
An element $z^\phi\in L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$ is called a (weak) solution to (\ref{skeleton eq.}) if for any $\varphi\in (C^\infty_0([0,T]\times \mathbb{T}^2))^2$ with $\text{div} \varphi=0$, and $t>0$,
$$\langle z^\phi(t),\varphi(t)\rangle=\langle u_0,\varphi(0)\rangle+\int^t_0\langle z^\phi, \partial_t\varphi\rangle-\langle \partial_1z^\phi,\partial_1 \varphi\rangle+\langle- B(z^\phi)+\sigma(s, z^\phi)\phi, \varphi\rangle ds.$$
The existence of the weak solution to (\ref{skeleton eq.}) can be obtained by the same method as in \cite{LZZ18} (see Lemma \ref{solution to skeleton eq.} in the following).
Define $g^0:C([0,T],U)\rightarrow L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$ by
\begin{align*}
g^0(h):=
\left\{
\begin{array}{ll}
z^\phi, &\text{ if }h=\int^\cdot_0\phi(s)ds \text{ for some }\phi\in L^2([0,T],l^2); \\
0, &\text{ otherwise.}
\end{array}
\right.
\end{align*}
Then the rate function can be written as
\begin{equation}\label{rate function LDP}
I(z)=\inf\left\{\frac{1}{2}\int^T_0\|\phi(s)\|^2_{l^2}ds: \text{ }z=z^\phi,\text{ }\phi\in L^2([0,T],l^2)\right\},
\end{equation}
where $z\in L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$.
The main result of this section is the following one:
\begin{Thm}\label{main result LDP}
Assume (A0)-(A3) hold with $L_2=0$ and $u_0\in \tilde{H}^{0,1}$, then $u^\varepsilon$ satisfies a large deviation principle on $L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$ with the good rate function $I$ given by (\ref{rate function LDP}).
\end{Thm}
The proof is divided into the following lemmas.
\vskip.10in
\begin{lemma}\label{solution to skeleton eq.}
Assume (A0)-(A3) hold with $L_2=0$. For all $u_0\in\tilde{H}^{0,1}$ and $\phi\in L^2([0,T], l^2)$ there exists a unique solution $$z^\phi\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})\bigcap C([0,T],H^{-1})$$ to (\ref{skeleton eq.}).
\end{lemma}
\begin{proof}
First we give some a priori estimates for $z^\phi$. By taking $H$ inner product of (\ref{skeleton eq.}) with $z^\phi$ and using $\text{div }z^\phi=0$, we have
\begin{align*}
&\|z^\phi(t)\|^2_H+2\int^t_0\|\partial_1z^\phi(s)\|^2_H ds\\
=&\|u_0\|^2_H+2\int^t_0\langle z^\phi(s), \sigma(s, z^\phi(s))\phi(s)\rangle ds\\
\leqslant& \|u_0\|^2_H+2\int^t_0\|z^\phi(s)\|_H\|\sigma(s,z^\phi(s))\|_{L_2(l^2, H)}\|\phi(s)\|_{l^2}ds\\
\leqslant& \|u_0\|^2_H+2\int^t_0\left(\|z^\phi(s)\|^2_H\|\phi(s)\|^2_{l^2}+K_0+K_1\|z^\phi(s)\|^2_H+K_2\|\partial_1z^\phi(s)\|^2_H\right)ds,
\end{align*}
where we used (A1) in the last inequality.
Hence by Gronwall's inequality, we have
\begin{equation}\label{a priori z1}
\|z^\phi(t)\|^2_H+\int^t_0\|\partial_1 z^\phi(s)\|^2_{H}ds\leqslant (\|u_0\|^2_H+C)e^{C\int^t_0(\|\phi(s)\|^2_{l^2}+1)ds}.
\end{equation}
Similarly, we have
\begin{align*}
&\|z^\phi(t)\|^2_{\tilde{H}^{0,1}}+2\int^t_0(\|\partial_1z^\phi(s)\|^2_H+\|\partial_1\partial_2z^\phi(s)\|^2_H)ds\\
=&\|u_0\|^2_{\tilde{H}^{0,1}}-2\int^t_0\langle \partial_2z^\phi(s), \partial_2(z^\phi\cdot \nabla z^\phi)(s)\rangle ds+2\int^t_0\langle z^\phi(s), \sigma(s, z^\phi(s))\phi(s)\rangle_{\tilde{H}^{0,1}}ds\\
\leqslant&\|u_0\|^2_{\tilde{H}^{0,1}}+\int^t_0(\frac{1}{5}\|\partial_1\partial_2z^\phi(s)\|^2_H+C(1+\|\partial_1z^\phi(s)\|^2_H)\|\partial_2z^\phi(s)\|^2_H)ds\\
&+2\int^t_0(\|z^\phi(s)\|^2_{\tilde{H}^{0,1}}\|\phi(s)\|_{l^2}^2+\|\sigma(s, z^\phi(s))\|_{L_2(l^2, \tilde{H}^{0,1})}^2)ds,
\end{align*}
where we used Lemma \ref{estimate for b with partial_2} in the last inequality.
Hence by (A2) we deduce that
\begin{align*}
&\|z^\phi(t)\|^2_{\tilde{H}^{0,1}}+\int^t_0\|z^\phi(s)\|^2_{\tilde{H}^{1,1}}ds\\
\leqslant& \|u_0\|^2_{\tilde{H}^{0,1}}+C+C\int^t_0(1+\|\partial_1z^\phi(s)\|^2_H+\|\phi(s)\|^2_{l^2})\|z^\phi(s)\|^2_{\tilde{H}^{0,1}}ds.
\end{align*}
Then by Gronwall's inequality and (\ref{a priori z1}) we have
\begin{equation}\label{a priori z2}
\|z^\phi(t)\|^2_{\tilde{H}^{0,1}}+\int^t_0\|z^\phi(s)\|^2_{\tilde{H}^{1,1}}ds\leqslant (\|u_0\|^2_{\tilde{H}^{0,1}}+C)e^{C(t,\phi,u_0)},
\end{equation}
where
$$C(t,\phi,u_0)=C\left(\int^t_0(1+\|\phi(s)\|^2_{l^2})ds+(\|u_0\|^2_H+1)e^{C\int^t_0(1+\|\phi(s)\|^2_{l^2})ds}\right).$$
Now consider the following approximate equation:
\begin{equation}\label{approxi. eq.}\aligned
\left\{
\begin{array}{lll}
&dz_\epsilon^\phi(t)=\partial^2_1 z_\epsilon^\phi(t)dt+\epsilon^2\partial_2^2z_\epsilon^\phi(t)dt-B(z_\epsilon^\phi(t))dt+\sigma(t, z_\epsilon^\phi(t))\phi(t)dt,\\
&\text{div} z_\epsilon^\phi=0,\\
&z_\epsilon^\phi(0)=u_0*j_\epsilon,
\end{array}
\right.
\endaligned
\end{equation}
where $j$ is a smooth function on $\mathbb{R}^2$ with
$$j(x)=1, \text{ } |x|\leqslant 1;\text{ }j(x)=0,\text{ }|x|\geqslant 2,$$
and
$$j_\epsilon(x)=\frac{1}{\epsilon^2}j(\frac{x}{\epsilon}).$$
It follows from classical theory on Navier-Stokes system that (\ref{approxi. eq.}) has a unique global smooth solution $z^\phi_\epsilon$ for any fixed $\epsilon$. Furthermore, along the same line to (\ref{a priori z1}) and (\ref{a priori z2}) we have
\begin{equation}\label{a priori z}\aligned
&\|z_\epsilon^\phi(t)\|^2_H+\int^t_0\|\partial_1 z_\epsilon^\phi(s)\|^2_{H}ds+\epsilon^2\int^t_0\|\partial_2z_\epsilon^\phi(s)\|^2_Hds\leqslant (\|u_0\|^2_H+C)e^{C\int^t_0(\|\phi(s)\|^2_{l^2}+1)ds},\\
&\|\partial_2 z_\epsilon^\phi(t)\|^2_{H}+\int^t_0\|\partial_1\partial_2 z_\varepsilon^\phi(s)\|^2_{H}ds+\epsilon^2\int^t_0\|\partial_2^2 z_\epsilon^\phi(s)\|^2_Hds\leqslant (\|u_0\|^2_{\tilde{H}^{0,1}}+C)e^{C(t,\phi,u_0)},
\endaligned
\end{equation}
The following follows a similar argument as in the proof of \cite[Theorem 3.1]{LZZ18}. By (\ref{a priori z}), we have $\{z_\epsilon^\phi\}_{\epsilon>0}$ is uniformly bounded in $L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})$, hence bounded in $L^4([0,T],H^{\frac{1}{2}})$ (by interpolation) and $L^4([0,T], L^4(\mathbb{T}^2))$ (by Sobolev embedding). Thus $B(z^\phi_\epsilon)$ is uniformly bounded in $L^2([0,T],H^{-1})$. Let $p\in(1,\frac{4}{3})$, we have
\begin{align*}
\int^T_0\|\sigma(s,z^\phi_\epsilon(s))\phi(s)\|^p_{H^{-1}}ds\leqslant &\int^T_0\|\sigma(s,z^\phi_\epsilon(s))\|^p_{L_2(l^2,H^{-1})}\|\phi(s)\|^p_{l^2}ds\\
\leqslant &C\int^T_0(1+\|\sigma(s,z^\phi_\epsilon(s))\|^4_{L_2(l^2,H^{-1})}+\|\phi(s)\|^2_{l^2})ds\\
\leqslant&C\int^T_0(1+\|z^\phi_\epsilon(s))\|^4_{H}+\|\phi(s)\|^2_{l^2})ds<\infty,
\end{align*}
where we used Young's inequality in the second line and (A0) in the third line.
It comes out that
\begin{equation}\label{bded in Lp(H-1)}
\{\partial_tz_\epsilon^\phi\}_{\epsilon>0}\text{ is uniformly bounded in } L^p([0,T],H^{-1}).\end{equation}
Thus by Aubin-Lions lemma (see \cite[Lemma 3.6]{LZZ18}), there exists a $z^\phi\in L^2([0,T], {H})$ such that
$$z_\epsilon^\phi\rightarrow z^\phi \text{ strongly in }L^2([0,T],H)\text{ as }\epsilon\rightarrow 0 \text{ (in the sense of subsequence)}.$$
Since $\{z_\epsilon^\phi\}_{\epsilon>0}$ is uniformly bounded in $L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})$, there exists a $\tilde{z}\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})$ such that
$$z_\epsilon^\phi\rightarrow \tilde{z} \text{ weakly in }L^2([0,T],\tilde{H}^{1,1})\text{ as }\epsilon\rightarrow 0 \text{ (in the sense of subsequence)}.$$
$$z_\epsilon^\phi\rightarrow \tilde{z} \text{ weakly star in }L^\infty([0,T],\tilde{H}^{0,1})\text{ as }\epsilon\rightarrow 0 \text{ (in the sense of subsequence)}.$$
By the uniqueness of weak convergence limit, we deduce that $z^\phi=\tilde{z}$. By (\ref{bded in Lp(H-1)}) and \cite[Theorem 2.2]{FG95}, we also have for any $\delta>0$
$$z_\epsilon^\phi\rightarrow z^\phi \text{ strongly in }C([0,T],H^{-1-\delta})\text{ as }\epsilon\rightarrow 0 \text{ (in the sense of subsequence)}.$$
Now we use the above convergence to prove that $z^\phi$ is a solution to (\ref{skeleton eq.}). Note that for any $\varphi\in C^\infty([0,T]\times \mathbb{T}^2)$ with $\text{div} \varphi=0$, for any $t\in[0,T]$, $z_\epsilon^{\phi}$ satisfies
\begin{equation}\label{eq of approximate solution}
\langle z^\phi_\epsilon(t), \varphi(t)\rangle =\langle u_0,\varphi(0)\rangle+\int^{t}_0\langle z^\phi_\epsilon, \partial_t\varphi\rangle-\langle \partial_1z^\phi_\epsilon,\partial_1 \varphi\rangle-\epsilon^2\langle \partial_2z^\phi_\epsilon,\partial_2 \varphi\rangle+\langle- B(z^\phi_\epsilon)+\sigma(s, z^\phi_\epsilon)\phi, \varphi\rangle ds.
\end{equation}
By \cite[Chapter 3, Lemma 3.2]{Te79} we have
$$\int^{t}_0\langle- B(z^{\phi}_\epsilon), \varphi\rangle ds\rightarrow \int^{t}_0\langle- B(z^\phi), \varphi\rangle ds\text{ as }\epsilon\rightarrow 0.$$
For the last term in the right hand side of (\ref{eq of approximate solution}), we have
\begin{align*}
&\int^{t}_0\langle \sigma(s, z^{\phi}_\epsilon)\phi-\sigma(s,z^\phi)\phi, \varphi\rangle ds\\
\leqslant& \int^{t}_0 \|(\sigma(s, z^{\phi}_\epsilon)-\sigma(s,z^\phi))\phi\|_{H}\|\varphi\|_Hds\\
\leqslant & C\int^{t}_0 \|\sigma(s, z^{\phi}_\epsilon)-\sigma(s,z^\phi)\|_{L_2(l^2,H)}\|\phi\|_{l^2}ds\\
\leqslant & C\left(\int^{t}_0 \|z^{\phi}_\epsilon-z^\phi\|^2_{H}ds\right)^\frac{1}{2}\left(\int^t_0\|\phi(s)\|^2_{l^2}ds\right)^\frac{1}{2},
\end{align*}
where we used H\"older's inequality and (A3) with $L_2=0$ in the last inequality.
Thus let $\epsilon\rightarrow 0$ in (\ref{eq of approximate solution}), we have $z^\phi\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})$ and
$$\partial_t z^\phi=\partial_1^2z^\phi-B(z^\phi)+\sigma(t,z^\phi(t))\phi.$$
Since the right hand side belongs to $L^p([0,T],H^{-1})$, we deduce that $$z^\phi\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})\bigcap C([0,T],H^{-1}).$$
For uniqueness, let $z^\phi_1, z^\phi_2\in L^\infty([0,T], \tilde{H}^{0,1})\bigcap L^2([0,T], \tilde{H}^{1,1})\bigcap C([0,T],H^{-1})$ be two solutions to (\ref{skeleton eq.}) and $w^\phi=z^\phi_1-z^\phi_2$. Then we have
\begin{align*}
&\|w^\phi(t)\|^2_H+2\int^t_0\|\partial_1 w^\phi(s)\|^2_Hds\\
=&\|w^\phi(0)\|^2_H-2\int^t_0\langle w^\phi(s), B(z^\phi_1)(s)-B(z^\phi_2)(s)\rangle ds\\
&+2\int^t_0\langle w^\phi(s), \sigma(s, z^\phi_1(s))\phi(s)-\sigma(s, z^\phi_2(s))\phi(s)\rangle ds\\
\leqslant& \|w^\phi(0)\|^2_H-2\int^t_0b(w^\phi(s), z^\phi_2(s), w^\phi(s))ds\\
&+2\int^t_0\|w^\phi(s)\|_H\| \sigma(s, z^\phi_1(s))-\sigma(s, z^\phi_2(s))\|_{L_2(l^2,H)}\|\phi(s)\|_{l^2} ds\\
\leqslant& \|w^\phi(0)\|^2_H+\int^t_0\frac{1}{5}\|\partial_1 w^\phi(s)\|^2_Hds+C\int^t_0(1+\|z^\phi_2(s)\|^2_{\tilde{H}^{1,1}})\|w^\phi(s)\|^2_Hds\\
&+\int^t_0(\|w^\phi(s)\|^2_H \|\phi(s)\|^2_{l^2}+L_1\|w^\phi(s)\|^2_H)ds,
\end{align*}
where we used Lemma \ref{anisotropic estimate for b} in the sixth line and (A3) with $L_2=0$ in the last line.
Then by Gronwall's inequality we have
$$\|w^\phi(t)\|^2_H\leqslant \|w^\phi(0)\|^2_He^{C\int^t_0(1+\|z^\phi_2(s)\|^2_{\tilde{H}^{1,1}}+\|\phi(s)\|^2_{l^2})ds},$$
which along with the fact that $z_2^\phi\in L^2([0,T], \tilde{H}^{1,1})$ and $\phi\in L^2([0,T],l^2)$ implies that $w^\phi(t)=0$. That is: $z_1^\phi=z_2^\phi$.
\end{proof}
\iffalse
\begin{lemma}\label{compact estimate for z}
For any $N>0$ and $\alpha\in(0,\frac{1}{2})$, there exists constants $C_{N, u_0}$ and $C_{\alpha, N,u_0}$ such that
\begin{equation}\label{uniform bded 1 for z}
\sup_{\phi\in\mathcal{S}_N}\{\int^T_0\|z^\phi(s)\|^2_{H^{1,1}}ds\}\leqslant C_{N,u_0},
\end{equation}
and
\begin{equation}\label{uniform bded 2 for z}
\sup_{\phi\in\mathcal{S}_N}\|z^\phi\|_{W^{\alpha, 2}([0,T], H^{-1})}\leqslant C_{\alpha, N,u_0}
\end{equation}
\end{lemma}
\begin{proof}
The first inequality is a direct consequence of (\ref{a priori z2}) since $\phi\in\mathcal{S}_N$. Now we prove (\ref{uniform bded 2 for z}). Notice that
\begin{align*}
z^\phi(t)&=u_0+\int^t_0\partial_1^2z^\phi(s)ds-\int^t_0B(z^\phi(s))ds+\int^t_0\sigma(s,z^\phi(s))\phi(s)ds\\
&=:u_0+I_1(t)+I_2(t)+I_3(t)
\end{align*}
By (\ref{a priori z1}) and H\"older's inequality we have
\begin{align*}
\int^T_0\|I_1(s)\|^2_{H^{-1}}ds\leqslant&\int^T_0\left(\int^s_0\|\partial^2_1z^\phi(l)\|_{H^{-1}}dl\right)^2ds \\
\leqslant& C\int^T_0\|\partial^2_1z^\phi(l)\|^2_{H^{-1}}dl\leqslant C\int^T_0\|\partial_1z^\phi(l)\|^2_{H}dl<+\infty,
\end{align*}
and
\begin{align*}
\int^T_0\int^T_0\frac{\|I_1(t)-I_1(s)\|^2_{H^{-1}}}{|t-s|^{1+2\alpha}}dtds\leqslant&C\int^T_0\int^T_0\frac{|t-s|}{|t-s|^{1+2\alpha}}\int^T_0\|\partial_1z^\phi(l)\|^2_Hdldtds\\
\leqslant&C\int^T_0\|\partial_1z^\phi(l)\|^2_Hdl<+\infty,
\end{align*}
where we use that $\alpha<\frac{1}{2}$.
Hence we have
$$\|I_1\|^2_{W^{\alpha,2}([0,T],H^{-1})}\leqslant C_{N,u_0}.$$
For $I_2$, by Lemma \ref{b(u,v,w)}, we have
$$\|B(z^\phi)\|_H^2=b(z^\phi,z^\phi,B(z^\phi))\leqslant C\|z^\phi\|^2_{\tilde{H}^{1,1}}\|B(z^\phi)\|_H,$$
which by (\ref{a priori z2}) implies that
\iffalse
for $t,s\in[0,T]$
\begin{align*}
\|\int^t_sB(z^\phi(l))dl\|^2_{H^{-1}}\leqslant \left(\int^t_s\|B(z^\phi(l))\|_{H^{-1}}dl\right)^2\leqslant C\left(\int^t_s\|z^\phi(l)\|_{H^{1,1}}^2dl\right)^2,
\end{align*}
Then we have
\fi
\begin{align*}
\int^T_0\|I_2(s)\|^2_{H^{-1}}ds\leqslant C\left(\int^T_0\|z^\phi(l)\|_{H^{1,1}}^2dl\right)^2.
\end{align*}
From the proof the Lemma \ref{solution to skeleton eq.} we know that by interpolation and Sobolev embedding, a similar argument shows $B(z^{\phi})$ is uniformly bounded in $L^2([0,T],H^{-1})$, by H\"older's inequality we have
\begin{align*}
\int^T_0\int^T_0\frac{\|I_2(t)-I_2(s)\|^2_{H^{-1}}}{|t-s|^{1+2\alpha}}dtds\leqslant \int^T_0\int^T_0\frac{1}{|t-s|^{2\alpha}}dtds\int^T_0\|B(z^\phi(l))\|^2_{H^{-1}}dl<+\infty.
\end{align*}
Thus by (\ref{a priori z2}) we have
$$\|I_2\|^2_{W^{\alpha,2}([0,T],H^{-1})}\leqslant C_{N,u_0}.$$
It remains to deal with the last term $I_3$. Since $\phi\in\mathcal{S}_N$, we have
\begin{align*}
\|\int^t_s\sigma(l,z^\phi(l))\phi(l)dl\|^2_{H^{-1}}\leqslant& \left(\int^t_s\|\sigma(l,z^\phi(l))\|_{L_2(l^2,H^{-1})}\|\phi(l)\|_{l^2}dl\right)^2\\
\leqslant& \left(\int^t_s\|\sigma(l,z^\phi(l))\|^2_{L_2(l^2,H^{-1})}dl\right)\left(\int^t_s\|\phi(l)\|^2_{l^2}dl\right)\\
\leqslant& CN\left(\int^t_s(1+\|z^\phi(l)\|^2_{H})dl\right),
\end{align*}
where we use H\"older's inequality in the second inequality and (A0) in the last one. Thus similar as $I_2$, by (\ref{a priori z1}) we deduce that
$$\|I_3\|^2_{W^{\alpha,2}([0,T],H^{-1})}\leqslant C_{N,u_0}.$$
Combing the above estimates, we obtain (\ref{uniform bded 2 for z}).
\end{proof}
\fi
The following Lemma shows that $I$ is a good rate function. The proof follows essentially the same argument as in \cite[Proposition 4.5]{WZZ}.
\begin{lemma}\label{good rate function}
Assume (A0)-(A3) hold with $L_2=0$. For all $N<\infty$, the set
$$K_N=\left\{g^0\left(\int^\cdot_0 \phi(s)ds\right): \phi\in S_N\right\}$$
is a compact subset in $L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$.
\end{lemma}
\begin{proof}
By definition, we have
$$K_N=\left\{z^\phi: \phi\in L^2([0,T], l^2), \text{ }\int^T_0\|\phi(s)\|^2_{l^2}ds\leqslant N\right\}.$$
Let $\{z^{\phi_n}\}$ be a sequence in $K_N$ where $\{\phi_n\}\subset S_N$. Note that (\ref{a priori z2}) implies that $z^{\phi_n}$ is uniformly bounded in $L^\infty([0,T], H^{1,0})\cap L^2([0,T], H^{1,1})$. Thus by weak compactness of $S_N$, a similar argument as in the proof of Lemma \ref{solution to skeleton eq.} shows that there exists $\phi\in\mathcal{S}_N$ and $z'\in L^2([0,T],H)$ such that the following convergence hold as $n\rightarrow \infty$ (in the sense of subsequence):
$\phi_n\rightarrow \phi$ in $\mathcal{S}_N$ weakly,
$z^{\phi_n}\rightarrow z'$ in $L^2([0,T], H^{1,0})$ weakly,
$z^{\phi_n}\rightarrow z'$ in $L^\infty([0,T], H)$ weak-star,
$z^{\phi_n}\rightarrow z'$ in $L^2([0,T], H)$ strongly.
$z^{\phi_n}\rightarrow z'$ in $C([0,T], H^{-1-\delta})$ strongly for any $\delta>0$.
Then for any $\varphi\in C^\infty([0,T]\times \mathbb{T}^2)$ with $\text{div} \varphi=0$ and for any $t\in[0,T]$, $z^{\phi_n}$ satisfies
\begin{equation}\label{eq in good rate function}
\langle z^{\phi_n}(t), \varphi(t)\rangle=\langle u_0,\varphi(0)\rangle+\int^{t}_0\langle z^{\phi_n}, \partial_t\varphi\rangle-\langle \partial_1z^{\phi_n},\partial_1 \varphi\rangle+\langle- B(z^{\phi_n})+\sigma(s, z^{\phi_n})\phi_n, \varphi\rangle ds.
\end{equation}
Let $n\rightarrow \infty$, we have
\begin{align*}
&\int^{t}_0\langle \sigma(s, z^{\phi_n})\phi_n-\sigma(s,z')\phi, \varphi\rangle ds\\
=&\int^{t}_0\langle [\sigma(s, z^{\phi_n})-\sigma(s,z')]\phi_n+\sigma(s,z')(\phi_n-\phi), \varphi\rangle ds\\
\leqslant& \int^{t}_0 \|(\sigma(s, z^{\phi_n})-\sigma(s,z'))\phi_n\|_{H}\|\varphi\|_Hds+\int^{t}_0\langle\sigma(s,z')(\phi_n-\phi), \varphi\rangle ds\\
\leqslant & C\int^{t}_0 \|\sigma(s, z^{\phi_n})-\sigma(s,z')\|_{L_2(l^2,H)}\|\phi_n\|_{l^2}ds+\int^{t}_0\langle\sigma(s,z')(\phi_n-\phi), \varphi\rangle ds\\
\leqslant & C\left(\int^{t}_0 \|z^{\phi_n}-z'\|^2_{H}ds\right)^\frac{1}{2}\left(\int^t_0\|\phi_n(s)\|^2_{l^2}ds\right)^\frac{1}{2}+\int^{t}_0\langle\sigma(s,z')(\phi_n-\phi), \varphi\rangle ds\\
\rightarrow&\text{ }0,
\end{align*}
where we used H\"older's inequality and (A3) with $L_2=0$ in the last inequality. By \cite[Chapter 3, Lemma 3.2]{Te79} we also have
$$\int^{t}_0\langle- B(z^{\phi_n}), \varphi\rangle ds\rightarrow \int^{t}_0\langle- B(z'), \varphi\rangle ds.$$
Then we deduce that
$$\langle z'(t), \varphi(t)\rangle=\langle u_0,\varphi(0)\rangle+\int^{t}_0\langle z', \partial_t\varphi\rangle-\langle \partial_1z',\partial_1 \varphi\rangle+\langle- B(z')+\sigma(s, z')\phi, \varphi\rangle ds,$$
which implies that $z'$ is a solution to (\ref{skeleton eq.}). By the uniqueness of solution, we deduce that $z'=z^\phi$.
Our goal is to prove $z^{\phi_n}\rightarrow z^\phi$ in $L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$.
Let $w^n=z^{\phi_n}-z^\phi$, by a direct calculation, we have
\begin{align*}
&\|w^n(t)\|^2_H+2\int^t_0\|\partial_1w^n(s)\|^2_Hds\\
=&-2\int^t_0\langle w^n(s), B(z^{\phi_n})(s)-B(z^\phi)(s)\rangle ds\\
&+2\int^t_0\langle w^n(s), \sigma(s, z^{\phi_n}(s))\phi_n(s)-\sigma(s, z^\phi(s))\phi(s)\rangle ds\\
=&-2\int^t_0 b(w^n, z^\phi, w^n)(s)ds+2\int^t_0\langle w^n(s), (\sigma(s,z^{\phi_n}(s))-\sigma(s,z^\phi(s)))\phi_n(s) \rangle ds\\
&+2\int^t_0\langle w^n(s), \sigma(s, z^\phi(s))(\phi_n(s)-\phi(s))\rangle ds\\
\leqslant & \int^t_0 \frac{1}{5}\|\partial_1w^n(s)\|^2_Hds+C\int^t_0(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})\|w^n(s)\|^2_Hds\\
&+C\int^t_0\|w^n(s)\|^2_H\|\phi_n(s)\|_{l^2}ds\\
&+\int^t_0\|w^n(s)\|_H\|\phi_n(s)-\phi(s)\|_{l^2}(K_0+K_1\|z^\phi(s)\|^2_H+K_2\|\partial_1 z^\phi(s)\|^2_H )^\frac{1}{2} ds,
\end{align*}
where we used Lemma \ref{anisotropic estimate for b} in the sixth line, (A3) with $L_2=0$ in the seventh line and (A1) in the last line. Then we have
\begin{align*}
&\sup_{t\in[0,T]}\|w^n(t)\|^2_H+\int^T_0\|\partial_1w^n(s)\|^2_Hds\\
\leqslant &C\int^T_0(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})\|w^n(s)\|^2_Hds\\
+& C(\sup_{t\in[0,T]}\|z^{\phi_n}(t)\|_H+\sup_{t\in[0,T]}\|z^\phi(t)\|_H)\left(\int^T_0\|\phi_n(s)\|^2_{l^2}ds\right)^\frac{1}{2}\left(\int^T_0\|w^n(s)\|^2_Hds\right)^\frac{1}{2}\\
+& C\left(\int^T_0\|\phi_n(s)-\phi(s)\|^2_{l^2}ds\right)^\frac{1}{2}\left(\int^T_0 (1+\|z^\phi(s)\|^2_H+\|\partial_1z^\phi(s)\|^2_H)\|w^n(s)\|^2_Hds\right)^\frac{1}{2}\\
\leqslant &C\int^T_0(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})\|w^n(s)\|^2_Hds+C(N)\left(\int^T_0\|w^n(s)\|^2_Hds\right)^\frac{1}{2}\\
+& CN^\frac{1}{2}\left(\int^T_0 (1+\|z^\phi(s)\|^2_H+\|\partial_1z^\phi(s)\|^2_H)\|w^n(s)\|^2_Hds\right)^\frac{1}{2},
\end{align*}
where we used (\ref{a priori z1}) and the fact that $\phi_n$, $\phi$ are in $\mathcal{S}_N$.
For any $\epsilon>0$, let
$$A_\epsilon:=\{s\in[0,T]; \|z^{\phi_n}(s)-z^\phi(s)\|_H>\epsilon\}.$$
Since $z^{\phi_n}\rightarrow z^\phi$ in $L^2([0,T], H)$ strongly, we have
$$\int^T_0\|w^n(s)\|^2_Hds\rightarrow 0,\text{ as }n\rightarrow\infty$$
and $\lim_{n\rightarrow\infty}Leb(A_\epsilon)=0$, where $Leb(B)$ means the Lebesgue measure of $B\in\mathcal{B}(\mathbb{R})$. Thus we have
\begin{align*}
&\int^T_0(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})\|w^n(s)\|^2_Hds\\
\leqslant&\left(\int_{ A_\epsilon}+\int_{[0,T]\setminus A_\epsilon}\right)(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})\|w^n(s)\|^2_Hds\\
\leqslant& C\epsilon+ 2\int_{A_\epsilon}(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})(\|z^{\phi_n}(s)\|^2_H+\|z^\phi(s)\|^2_H)ds\\
\leqslant&C\epsilon+ C\int_{A_\epsilon}(1+\|z^\phi(s)\|^2_{\tilde{H}^{1,1}})ds\\
\rightarrow & \text{ }C\epsilon\text{ as } n\rightarrow \infty,
\end{align*}
where we used (\ref{a priori z1}) in the forth line and (\ref{a priori z2}) in the last line. A similar argument also implies that
\begin{align*}
\int^T_0 (1+\|z^\phi(s)\|^2_H+\|\partial_1z^\phi(s)\|^2_H)\|w^n(s)\|^2_Hds\leqslant C\epsilon.
\end{align*}
Hence we have
\begin{align*}
\sup_{t\in[0,T]}\|w^n(t)\|^2_H+\int^T_0\|\partial_1w^n(s)\|^2_Hds\leqslant C\epsilon+C\sqrt{\epsilon} \text{ as }n\rightarrow \infty.
\end{align*}
Since $\epsilon$ is arbitrary, we obtain that
$$z^{\phi^n}\rightarrow z^\phi\text{ strongly in }L^\infty([0,T], H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1}).$$
\end{proof}
For next step, consider the following equation:
\begin{equation}\label{eq. for weak convergence}\aligned
dZ^\varepsilon_v(t)&=\partial^2_1 Z^\varepsilon_v(t)dt-B(Z_v^\varepsilon(t))dt+\sigma(t, Z^\varepsilon_v(t))v^\varepsilon(t)dt+\sqrt{\varepsilon}\sigma(t, Z_v^\varepsilon(t))dW(t),\\
\text{div}Z^\varepsilon_v&=0,\\
Z^\varepsilon_v(0)&=u_0,
\endaligned
\end{equation}
where $v^\varepsilon\in\mathcal{A}_N$ for some $N<\infty$.
Here $Z^\varepsilon_v$ should have been denoted $Z^\varepsilon_{v^\varepsilon}$ and the slight abuse of notation is for simplicity.
\begin{lemma}\label{Girsanov thm to prove existence}
Assume (A0)-(A3) hold with $L_2=0$ and $v^\varepsilon\in\mathcal{A}_N$ for some $N<\infty$. Then $Z_v^\varepsilon=g^\varepsilon\left(W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds\right)$ is the unique strong solution to (\ref{eq. for weak convergence}).
\end{lemma}
\begin{proof}
Since $v^\varepsilon\in \mathcal{A}_N$, by the Girsanov theorem (see \cite[Appendix I]{LR15}), $\tilde{W}(\cdot):=W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds$ is an $l^2$-cylindrical Wiener-process under the probability measure
$$d\tilde{P}:=\exp\left\{-\frac{1}{\sqrt{\varepsilon}}\int^T_0v^\varepsilon(s)dW(s)-\frac{1}{2\varepsilon}\int^T_0\|v^\varepsilon(s)\|^2_{l^2}ds\right\}dP.$$
Then $(Z_v^\varepsilon, \tilde{W})$ is the solution to (\ref{LDP eq.}) on the stochastic basis $(\Omega, \mathcal{F}, \tilde{P})$. By (A0) we have
\begin{align*}
\int^T_0\|\sigma(s,Z^\varepsilon_v(s))\|_{H^{-1}}ds<\infty.
\end{align*}
Then $(Z^\varepsilon_v, W)$ satisfies the condition of the definition of weak solution (see \cite[Definition 4.1]{LZZ18}) and hence is a weak solution to (\ref{eq. for weak convergence}) on the stochastic basis $(\Omega, \mathcal{F}, {P})$ and $Z^\varepsilon_v=g^\varepsilon\left(W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds\right)$.
If $\tilde{Z^\varepsilon_v}$ and $Z^\varepsilon_v$ are two weak solutions to (\ref{eq. for weak convergence}) on the same stochastic basis $(\Omega, \mathcal{F}, {P})$. Let $W^\varepsilon=Z^\varepsilon_v-\tilde{Z^\varepsilon_v}$ and $q(t)=k\int^t_0(\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}+\|v^\varepsilon(s)\|^2_{l^2})ds$ for some constant $k$. Applying It\^o's formula to $e^{-q(t)}\|W^\varepsilon(t)\|^2_H$, we have
\begin{align*}
&e^{-q(t)}\|W^\varepsilon(t)\|^2_H+2\int^t_0e^{-q(s)}\|\partial_1 W^\varepsilon(s)\|^2_Hds\\
=&-k\int^t_0e^{-q(s)}\|W^\varepsilon(s)\|^2_H(\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}+\|v^\varepsilon(s)\|^2_{l^2})ds-2\int^t_0e^{-q(s)}b(W^\varepsilon, Z^\varepsilon_v, W^\varepsilon)ds\\
&+2\int^t_0e^{-q(s)}\langle \sigma(s,Z^\varepsilon_v)v^\varepsilon-\sigma(s, \tilde{Z}^\varepsilon_v)v^\varepsilon, W^\varepsilon(s)\rangle ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-q(s)}\langle W^\varepsilon(s), (\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v))dW(s)\rangle\\
&+\varepsilon\int^t_0e^{-q(s)}\|\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v)\|^2_{L_2(l^2,H)}ds.
\end{align*}
By Lemma \ref{anisotropic estimate for b}, there exists constants $\tilde{\alpha}\in(0,1)$ and $\tilde{C}$ such that
$$|b(W^\varepsilon, Z^\varepsilon_v, W^\varepsilon)|\leqslant \tilde{\alpha}\|\partial_1 W^\varepsilon\|^2_H+\tilde{C}(1+\|Z^\varepsilon_v\|^2_{\tilde{H}^{1,1}})\|W^\varepsilon\|^2_H.$$
We also have
\begin{align*}
2|\langle \sigma(s,Z^\varepsilon_v)v^\varepsilon-\sigma(s, \tilde{Z}^\varepsilon_v)v^\varepsilon, W^\varepsilon\rangle|&\leqslant 2\|(\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v))v^\varepsilon\|_H\|W^\varepsilon\|_H\\
&\leqslant \|\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v)\|^2_{L_2(l^2,H)}+\|v^\varepsilon\|^2_{l^2}\|W^\varepsilon\|^2_H.
\end{align*}
Let $k>2\tilde{C}$ and we may assume $\varepsilon<\frac{16}{25}$, by (A3) with $L_2=0$ we have
\begin{align*}
&e^{-q(t)}\|W^\varepsilon(t)\|^2_H+(2-2\tilde{\alpha})\int^t_0e^{-q(s)}\|\partial_1 W^\varepsilon(s)\|^2_Hds\\
\leqslant& C\int^t_0e^{-q(s)}\|W^\varepsilon(s)\|^2_Hds+2\sqrt{\varepsilon}\int^t_0e^{-q(s)}\langle W^\varepsilon(s), (\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v))dW(s)\rangle.
\end{align*}
By the Burkh\"older-Davis-Gundy's inequality (see \cite[Appendix D]{LR15}), we have
\begin{align*}
&2\sqrt{\varepsilon}|E[\sup_{r\in[0,t]}\int^r_0e^{-q(s)}\langle W^\varepsilon(s), (\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v))dW(s)\rangle]|\\
\leqslant &6\sqrt{\varepsilon}E\left(\int^t_0e^{-2q(s)}\|\sigma(s,Z^\varepsilon_v)-\sigma(s, \tilde{Z}^\varepsilon_v)\|^2_{L_2(l^2,H)}\|W^\varepsilon(s)\|^2_Hds\right)^\frac{1}{2}\\
\leqslant& \sqrt{\varepsilon}E(\sup_{s\in[0,t]}(e^{-q(s)}\|W^\varepsilon(s)\|^2_H))+9\sqrt{\varepsilon} E\int^t_0e^{-q(s)}L_1\|W^\varepsilon(s)\|^2_Hds,
\end{align*}
where we used (A3) with $L_2=0$ and assume that $\tilde{\alpha}<1$.
Thus we have
$$E(\sup_{s\in[0,t]}(e^{-q(s)}\|W^\varepsilon(s)\|^2_H))\leqslant CE\int^t_0e^{-q(s)}\|W^\varepsilon(s)\|^2_Hds.$$
By the Gronwall's inequality we obtain $W^\varepsilon=0$ $P$-a.s., i.e. $\tilde{Z^\varepsilon_v}=Z^\varepsilon_v$ $P$-a.s..
Then by the Yamada-Watanabe theorem, we have $Z^\varepsilon_v$ is the unique strong solution to (\ref{eq. for weak convergence}).
\end{proof}
\begin{lemma}\label{estimate eq. for weak convergence}
Assume $Z^\varepsilon_v$ is a solution to (\ref{eq. for weak convergence}) with $v^\varepsilon\in\mathcal{A}_N$ and $\varepsilon<1$ small enough. Then we have
\begin{equation}\label{eq01 in lemma estimate eq. for weak convergence}
E(\sup_{t\in[0,T]}\|Z^\varepsilon_v(t)\|^4_H)+E\int^T_0\|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds+E\int^T_0\|\partial_1 Z^\varepsilon_v(s)\|^2_{H}ds\leqslant C(N,u_0).
\end{equation}
\iffalse
Define a stopping time with respect to $\mathcal{F}_{t}$:
$$\tau^{M,\varepsilon}_1=T\wedge\inf\{t>0: \int^t_0\|\partial_1 Z^\varepsilon_v(s)\|^2_{H}ds>M\},$$
\fi
Moreover, there exists $k>0$ such that
\begin{equation}\label{eq02 in lemma estimate eq. for weak convergence}
E(\sup_{t\in[0,T]}e^{-kg(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}})+E\int^{T}_0e^{-kg(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}ds\leqslant C(N,u_0),
\end{equation}
where $g(t)=\int^t_0\|Z^\varepsilon_v(s)\|^2_{H}ds$ and $C(N,u_0)$ is a constant depend on $N, u_0$ but independent of $\varepsilon$.
\iffalse
We also have
\begin{equation}\label{L4 estimate}
E(\sup_{t\in[0,T]}\|Z^\varepsilon_v(t)\|^4_H)+E\int^T_0\|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds\leqslant C(N, u_0).
\end{equation}
\fi
\end{lemma}
\begin{proof}
We prove (\ref{eq01 in lemma estimate eq. for weak convergence}) by two parts of estimates.
For first step, applying It\^o's formula to $\|Z^\varepsilon_v(t)\|^2_H$, we have
\begin{align*}
&\|Z^\varepsilon_v(t)\|^2_H+2\int^t_0\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds\\
=&\|u_0\|^2_H+2\int^t_0 \langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s)\rangle ds\\
&+2\sqrt{\varepsilon}\int^t_0\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle+\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, H)}ds\\
\leqslant&\|u_0\|^2_H+ \int^t_0(\|Z^\varepsilon(s)\|^2_H\|v^\varepsilon(s)\|^2_{l^2}+\|\sigma(s, Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)})ds\\
&+2\sqrt{\varepsilon}\int^t_0\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle+\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, H)}ds\\
\leqslant &\|u_0\|^2_H+\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|v^\varepsilon(s)\|^2_{l^2}ds+ (1+\varepsilon)\int^t_0(K_0+K_1\|Z^\varepsilon_v\|^2_H+K_2\|\partial_1Z^\varepsilon_v\|_H^2)ds\\
&+2\sqrt{\varepsilon}\int^t_0\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle,
\end{align*}
where we used (A1) in the last inequality.
By Gronwall's inequality and $v^\varepsilon\in \mathcal{A}_N$,
\begin{align*}
&\|Z^\varepsilon_v(t)\|^2_H+(2-(1+\varepsilon)K_2)\int^t_0\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant &(\|u_0\|^2_H+C+2\sqrt{\varepsilon}\int^t_0\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle)e^{N+2K_1T}.
\end{align*}
\iffalse
\begin{align*}
&e^{-\int^t_0\|v^\varepsilon(s)\|^2_{l^2}ds}\|Z^\varepsilon_v(t)\|^2_H+2\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds\\
=&\|u_0\|^2_H-\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|v^\varepsilon(s)\|^2_{l^2}\|Z^\varepsilon_v(s)\|^2_Hds\\
&+2\int^t_0 \langle e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s)\rangle ds\\
&+2\sqrt{\varepsilon}\int^t_0\langle e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(t)\rangle\\
&+\varepsilon\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, H)}ds\\
\leqslant&\|u_0\|^2_H-\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|v^\varepsilon(s)\|^2_{l^2}\|Z^\varepsilon_v(s)\|^2_Hds\\
&+ \int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}(\|Z^\varepsilon(s)\|^2_H\|v^\varepsilon(s)\|^2_{l^2}+\|\sigma(s, Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)})ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle\\
&+\varepsilon\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, H)}ds\\
\leqslant &\|u_0\|^2_H+ 2\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}(K_0+K_1\|Z^\varepsilon_v\|^2_H+K_2\|\partial_1Z^\varepsilon_v\|_H^2)ds\\
&+2\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle,
\end{align*}
where we use (A1) in the last inequality and that $\varepsilon<1$.
\fi
For the term in the right hand side, by the Burkh\"older-Davis-Gundy inequality we have
\begin{align*}
&2\sqrt{\varepsilon}e^{N+K_1T} E\left(\sup_{0\leqslant s\leqslant t}|\int^s_0\langle Z^\varepsilon_v(r), \sigma(r,Z^\varepsilon_v(r))dW(r)\rangle|\right)\\
\leqslant &6\sqrt{\varepsilon}e^{N+K_1T}E\left(\int^t_0\|Z^\varepsilon_v(r)\|^2_H\|\sigma(r,Z^\varepsilon_v(r))\|^2_{L_2(l^2,H)}ds\right)^\frac{1}{2}\\
\leqslant &\sqrt{\varepsilon} E[\sup_{0\leqslant s\leqslant t}(\|Z^\varepsilon_v(s)\|^2_H)]+9\sqrt{\varepsilon}e^{2N+2K_1T}E\int^t_0[K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H]ds,
\end{align*}
where $(9\sqrt{\varepsilon}e^{2N+2K_1T}+1+\varepsilon)K_2-2<0$ (this can be done when $\varepsilon<(\frac{10}{9e^{2N+2K_1T}+1})^2$) and we used (A1) in the last inequality. Thus we have
\begin{align*}
&E[\sup_{s\in[0,t]}(\|Z^\varepsilon_v(t)\|^2_H)]+E\int^t_0\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant&C(\|u_0\|^2_H+1)+C\int^t_0E[\sup_{r\in[0,s]}(\|Z^\varepsilon_v(r)\|^2_H)]ds.
\end{align*}
Then by Gronwall's inequality we have
\iffalse
\begin{align*}
E[\sup_{s\in[0,t]}(e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|Z^\varepsilon_v(t)\|^2_H)]+E\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds\leqslant C(\|u_0\|^2_H+1),
\end{align*}
which implies
\begin{equation}\label{eq1 in lemma estimate eq. for weak convergence}\aligned
&E(\sup_{0\leqslant t\leqslant T}\|Z^\varepsilon_v(t)\|^2_H)+E\int^T_0\|\partial_1 Z^\varepsilon_v(s)\|^2_{H}ds\\
\leqslant&E\left[(\sup_{s\in[0,t]}(e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|Z^\varepsilon_v(t)\|^2_H)+\int^t_0e^{-\int^s_0\|v^\varepsilon(r)\|^2_{l^2}dr}\|\partial_1 Z^\varepsilon_v(s)\|^2_Hds)e^{\int^t_0\|v^\varepsilon(r)\|^2_{l^2}dr}\right]\\
\leqslant &C(1+\|u_0\|^2_H)e^N.
\endaligned
\end{equation}
\fi
\begin{equation}\label{eq1 in lemma estimate eq. for weak convergence}\aligned
E(\sup_{0\leqslant t\leqslant T}\|Z^\varepsilon_v(t)\|^2_H)+E\int^T_0\|\partial_1 Z^\varepsilon_v(s)\|^2_{H}ds\leqslant C(1+\|u_0\|^2_H).
\endaligned
\end{equation}
The second step is similar to \cite[Lemma 4.2]{LZZ18}. By It\^o's formula we have
\begin{equation}\label{L4 estimate step1}\aligned
\|Z^\varepsilon_v(t)\|^4_H=&\|u_0\|^4_H-4\int^t_0\|Z^\varepsilon_v\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
&+4\int^t_0\|Z^\varepsilon_v(s)\|_H^2\langle\sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s),Z^\varepsilon_v(s)\rangle ds\\
&+2\varepsilon\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)}ds\\
&+4\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))^*(Z^\varepsilon_v)\|^2_{l^2}ds\\
&+4\sqrt{\varepsilon}\int^t_0\|Z^\varepsilon_v(s)\|^2_H\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle_H\\
=:&\|u_0\|^4_H-4\int^t_0\|Z^\varepsilon_v\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds+I_1+I_2+I_3+I_4.
\endaligned
\end{equation}
By (A1) we have
\begin{align*}
I_1(t)\leqslant& 4\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\sigma(s,Z^\varepsilon_v(s))\|_{L_2(l^2,H)}\|v^\varepsilon(s)\|_{l^2}\|Z^\varepsilon_v(s)\|_Hds\\
\leqslant& 2\int^t_0\|Z^\varepsilon_v(s)\|^2_H(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H+\|v^\varepsilon(s)\|^2_{l^2}\|Z^\varepsilon_v(s)\|^2_H)ds,
\end{align*}
and
\begin{align*}
I_2+I_3\leqslant &6\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|_{L_2(l^2,H)}^2\|Z^\varepsilon_v(s)\|_H^2ds\\
\leqslant& 6\varepsilon\int^t_0(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H)\|Z^\varepsilon_v(s)\|^2_Hds.
\end{align*}
Thus we have
\begin{align*}
&\|Z^\varepsilon_v(t)\|^4_H+(4-2K_2-6\varepsilon K_2)\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant& \|u_0\|^4_H+I_4+(2+6\varepsilon)K_0\int^t_0\|Z^\varepsilon_v(s)\|^2_Hds+\int^t_0(2K_1+6\varepsilon K_1+2\|v^\varepsilon(s)\|^2_{l^2})\|Z^\varepsilon_v(s)\|^4_H)ds.
\end{align*}
Since $v^\varepsilon\in\mathcal{A}_N$, by Gronwall's inequality we have
\begin{align*}
&\|Z^\varepsilon_v(t)\|^4_H+(4-2K_2-6\varepsilon K_2)\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant &\left(\|u_0\|^4_H+I_4+(2+6\varepsilon)K_0\int^t_0\|Z^\varepsilon_v(s)\|^2_Hds\right)e^{8K_1T+N}.
\end{align*}
The Burkh\"older-Davis-Gundy inequality, the Young's inequality and (A1) imply that
\begin{align*}
E(\sup_{s\in[0,t]}I_4(s))\leqslant& 12\sqrt{\varepsilon} E\left(\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)}\|Z^\varepsilon_v(s)\|^6_Hds\right)^\frac{1}{2}\\
\leqslant& \sqrt{\varepsilon} E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)\\
&+36\sqrt{\varepsilon}E\int^t_0(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H)\|Z^\varepsilon_v(s)\|^2_Hds.
\end{align*}
Let $\varepsilon$ small enough such that $2K_2+6\varepsilon K_2+36\sqrt{\varepsilon}K_2e^{8K_1T+N}<4$ and $\sqrt{\varepsilon}e^{8K_1T+N}<1$ (for instance $\varepsilon<(\frac{10}{3+18e^{8K_1T+N}})^2$). Then the above estimates and (\ref{eq01 in lemma estimate eq. for weak convergence}) imply that
\begin{align*}
&E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)+\int^t_0 \|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds\\
\leqslant &C(N,u_0)+CE\int^t_0\|Z^\varepsilon_v(s)\|^4_Hds,
\end{align*}
which by Gronwall's inequality yields that
\begin{align*}
E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)+\int^t_0 \|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds\leqslant C(N,u_0).
\end{align*}
For (\ref{eq02 in lemma estimate eq. for weak convergence}), let $h(t)=kg(t)+\int^t_0\|v^\varepsilon(s)\|^2_{l^2}ds$ for some universal constant $k$. Applying It\^o's formula to $e^{-h(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}}$, we have
\begin{align*}
&e^{-h(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}}+2\int^t_0e^{-h(s)}(\|\partial_1 Z^\varepsilon_v(s)\|^2_H+\|\partial_1\partial_2 Z^\varepsilon_v(s)\|^2_H)ds\\
=&\|u_0\|^2_{\tilde{H}^{0,1}}-\int^t_0e^{-h(s)}(k\|\partial_1Z^\varepsilon_v(s)\|^2_H+\|v^\varepsilon(s)\|^2_{l^2})\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{0,1}}ds\\
&+2\int^t_0e^{-h(s)}\langle \partial_2 Z^\varepsilon_v(s), \partial_2(Z^\varepsilon_v\cdot \nabla Z^\varepsilon_v)(s)\rangle ds+2\int^t_0e^{-h(s)}\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s)\rangle_{\tilde{H}^{0,1}}ds\\
&+2\sqrt{\varepsilon}\int^t_0 e^{-h(s)}\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(t)\rangle_{\tilde{H}^{0,1}}+\varepsilon\int^t_0e^{-h(s)}\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, \tilde{H}^{0,1})}ds.
\end{align*}
By Lemma \ref{estimate for b with partial_2}, there exists a constant $C_1$ such that
$$|\langle \partial_2 Z^\varepsilon_v, \partial_2(Z^\varepsilon_v\cdot \nabla Z^\varepsilon_v)\rangle|\leqslant \frac{1}{2}\|\partial_1\partial_2Z^\varepsilon_v\|^2_H+C_1(1+\|\partial_1Z^\varepsilon_v\|^2_H)\|\partial_2Z^\varepsilon_v\|^2_H.$$
By Young's inequality,
$$2|\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s)\rangle_{\tilde{H}^{0,1}}|\leqslant \|Z^\varepsilon_v\|^2_{\tilde{H}^{0,1}}\|v^\varepsilon\|^2_{l^2}+\|\sigma(s,Z^\varepsilon_v)\|^2_{L_2(l^2,\tilde{H}^{0,1})}.$$
Choosing $k>2C_1$, we have
\begin{align*}
&e^{-h(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}}+\int^t_0e^{-h(s)}(\|\partial_1 Z^\varepsilon_v(s)\|^2_H+\|\partial_1\partial_2 Z^\varepsilon_v(s)\|^2_H)ds\\
\leqslant&\|u_0\|^2_{\tilde{H}^{0,1}}+C\int^t_0e^{-h(s)}\|\partial_2Z^\varepsilon_v(s)\|^2_Hds+(1+\varepsilon)\int^t_0e^{-h(s)}\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2, \tilde{H}^{0,1})}ds\\
&+2\sqrt{\varepsilon}\int^t_0 e^{-h(s)}\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(t)\rangle_{\tilde{H}^{0,1}}.
\end{align*}
By the Burkh\"older-Davis-Gundy inequality we have
\begin{align*}
&2\sqrt{\varepsilon} E\left(\sup_{s\in[0,t]}|\int^s_0e^{-h(r)}\langle Z^\varepsilon_v(r), \sigma(r,Z^\varepsilon_v(r))dW(r)\rangle_{\tilde{H}^{0,1}}|\right)\\
\leqslant &6\sqrt{\varepsilon} E\left(\int^{t}_0e^{-2h(s)}\|Z^\varepsilon_v(r)\|^2_{\tilde{H}^{0,1}}\|\sigma(r,Z^\varepsilon_v(r))\|^2_{L_2(l^2,\tilde{H}^{0,1})}ds\right)^\frac{1}{2}\\
\leqslant &\sqrt{\varepsilon} E[\sup_{s\in[0,t]}(e^{-h(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{0,1}})]\\
&+9\sqrt{\varepsilon} E\int^{t}_0e^{-h(s)}[\tilde{K_0}+\tilde{K_1}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{0,1}}+\tilde{K_2}(\|\partial_1 Z^\varepsilon_v(s)\|^2_H+\|\partial_1\partial_2 Z^\varepsilon_v(s)\|^2_H)]ds,
\end{align*}
where $(9\sqrt{\varepsilon}+1+\varepsilon)\tilde{K_2}-1<0$ (this can be done if $\varepsilon<\frac{9}{400}$) and we used (A2) in the last inequality.
Combine the above estimates, we have
\begin{align*}
&E(\sup_{s\in[0,{t}]}e^{-h(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{0,1}})+E\int^{t}_0e^{-h(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}ds\\
\leqslant &C(\|u_0\|^2_{\tilde{H}^{0,1}}+1+E\int^{t}_0e^{-h(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{0,1}}ds)
\end{align*}
Then Gronwall's inequality implies that
\begin{align*}
E(\sup_{0\leqslant t\leqslant T}e^{-h(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}})+E\int^{T}_0e^{-h(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}ds\leqslant C(1+\|u_0\|^2_{\tilde{H}^{0,1}}).
\end{align*}
Since $v^\varepsilon\in \mathcal{S}_N$, we deduce that
\begin{equation}\label{eq2 in lemma estimate eq. for weak convergence}
E(\sup_{t\in[0,T]}e^{-kg(t)}\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{0,1}})+E\int^{T}_0e^{-kg(s)}\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,1}}ds\leqslant C(1+\|u_0\|^2_{\tilde{H}^{0,1}})e^{N}.
\end{equation}
\iffalse
The proof of (\ref{eq03 in lemma estimate eq. for weak convergence}) is very similar to that of the deterministic case (\ref{uniform bded 2 for z}). Note that
\begin{align*}
Z^\varepsilon_v(t)&=\int^t_0\partial_1^2Z^\varepsilon_v(s)ds-\int^t_0B(Z^\varepsilon_v(s))ds+\int^t_0\sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s)ds+\sqrt{\varepsilon}\int^t_0\sigma(s,Z^\varepsilon_v(s))dW(s)\\
&=:I_1+I_2+I_3+I_4.
\end{align*}
Similar to the proof of Lemma \ref{compact estimate for z}, by (\ref{eq01 in lemma estimate eq. for weak convergence}) we have
\begin{align*}
E\|I_1\|^2_{W^{\alpha, 2}([0,T], H^{-1})}\leqslant C(\alpha, N, u_0),
\end{align*}
and
\begin{align*}
E\|I_3\|^2_{W^{\alpha, 2}([0,T], H^{-1})}\leqslant C(\alpha, N, u_0).
\end{align*}
By (\ref{eq02 in lemma estimate eq. for weak convergence}) we have
\begin{align*}
E\|1_{\{t\leqslant\tau^{M,\varepsilon_1}\}}I_2\|^2_{W^{\alpha, 2}([0,T], H^{-1})}\leqslant C(\alpha, N, u_0)e^{CM}.
\end{align*}
For $I_4$, by (A0) we have
\begin{align*}
E\|I_4(T)\|^2_{H^{-1}}\leqslant& CE\int^T_0\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2,H^{-1})}ds\\
\leqslant & CE\int^T_0(1+\|Z^\varepsilon_v(s)\|^2_H)ds\\
\leqslant & C(N,u_0),
\end{align*}
then we have
\begin{align*}
E\|I_4\|^2_{W^{\alpha, 2}([0,T], H^{-1})}\leqslant &C(\alpha) E\|I_4(T)\|^2_{H^{-1}}\leqslant C(\alpha, N, u_0).
\end{align*}
Thus (\ref{eq03 in lemma estimate eq. for weak convergence}) follows from the above estimates.
The proof of (\ref{L4 estimate}) is similar to \cite[Lemma 4.2]{LZZ18}. By It\^o's formula we have
\begin{equation}\label{L4 estimate step1}\aligned
\|Z^\varepsilon_v(t)\|^4_H=&\|u_0\|^4_H-4\int^t_0\|Z^\varepsilon_v\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
&+4\int^t_0\|Z^\varepsilon_v(s)\|^2\langle\sigma(s,Z^\varepsilon_v(s))v^\varepsilon(s),Z^\varepsilon_v(s)\rangle_Hds\\
&+2\varepsilon\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)}ds\\
&+4\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))^*(Z^\varepsilon_v)\|^2_{l^2}ds\\
&+4\sqrt{\varepsilon}\int^t_0\|Z^\varepsilon_v(s)\|^2_H\langle Z^\varepsilon_v(s), \sigma(s,Z^\varepsilon_v(s))dW(s)\rangle_H\\
=:&\|u_0\|^4_H-4\int^t_0\|Z^\varepsilon_v\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds+I_1+I_2+I_3+I_4.
\endaligned
\end{equation}
By (A1) we have
\begin{align*}
I_1(t)\leqslant& 4\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\sigma(s,Z^\varepsilon_v(s))\|_{L_2(l^2,H)}\|v^\varepsilon(s)\|_{l^2}\|Z^\varepsilon_v(s)\|_Hds\\
\leqslant& 2\int^t_0\|Z^\varepsilon_v(s)\|^2_H(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H+\|v^\varepsilon(s)\|^2_{l^2}\|Z^\varepsilon_v(s)\|^2_H)ds,
\end{align*}
and
\begin{align*}
I_2+I_3\leqslant &6\varepsilon\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|_{L_2(l^2,H)}^2\|Z^\varepsilon_v(s)\|_H^2ds\\
\leqslant& 6\varepsilon\int^t_0(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H)\|Z^\varepsilon_v(s)\|^2_Hds.
\end{align*}
Thus we have
\begin{align*}
&\|Z^\varepsilon_v(t)\|^4_H+(4-2K_2-6\varepsilon K_2)\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant& \|u_0\|^4_H+I_4+(2+6\varepsilon)K_0\int^t_0\|Z^\varepsilon_v(s)\|^2_Hds+\int^t_0(2K_1+6\varepsilon K_1+2\|v^\varepsilon(s)\|^2_{l^2})\|Z^\varepsilon_v(s)\|^4_H)ds.
\end{align*}
Since $v^\varepsilon\in\mathcal{A}_N$, by Gronwall's inequality we have
\begin{align*}
&\|Z^\varepsilon_v(t)\|^4_H+(4-2K_2-6\varepsilon K_2)\int^t_0\|Z^\varepsilon_v(s)\|^2_H\|\partial_1Z^\varepsilon_v(s)\|^2_Hds\\
\leqslant &\left(\|u_0\|^4_H+I_4+(2+6\varepsilon)K_0\int^t_0\|Z^\varepsilon_v(s)\|^2_Hds\right)e^{8K_1+N}.
\end{align*}
The Burkh\"older-Davis-Gundy inequality, the Young's inequality and (A1) imply that
\begin{align*}
E(\sup_{s\in[0,t]}I_4(s))\leqslant& 16\sqrt{\varepsilon} E\left(\int^t_0\|\sigma(s,Z^\varepsilon_v(s))\|^2_{L_2(l^2,H)}\|Z^\varepsilon_v(s)\|^6_Hds\right)^\frac{1}{2}\\
\leqslant& \sqrt{\varepsilon} E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)\\
&+64\sqrt{\varepsilon}E\int^t_0(K_0+K_1\|Z^\varepsilon_v(s)\|^2_H+K_2\|\partial_1Z^\varepsilon_v(s)\|^2_H)\|Z^\varepsilon_v(s)\|^2_Hds.
\end{align*}
Let $\varepsilon$ small enough such that $2K_2+6\varepsilon K_2+32\sqrt{\varepsilon}K_2e^{8K_1+N}<4$ and $\sqrt{\varepsilon}e^{8K_1+N}<1$ (for instance $\varepsilon<(\frac{10}{3+16e^{8K_1+N}})^2$), then the above estimates and (\ref{eq01 in lemma estimate eq. for weak convergence}) imply that
\begin{align*}
&E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)+\int^t_0 \|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds\\
\leqslant &C(N,u_0)+CE\int^t_0\|Z^\varepsilon_v(s)\|^4_Hds,
\end{align*}
the Gronwall's inequality yields that
\begin{align*}
E(\sup_{s\in[0,t]}\|Z^\varepsilon_v(s)\|^4_H)+\int^t_0 \|Z^\varepsilon_v(s)\|^2_H\|Z^\varepsilon_v(s)\|^2_{\tilde{H}^{1,0}}ds\leqslant C(N,u_0).
\end{align*}
\fi
\end{proof}
Similar as \cite[lemma 4.3]{LZZ18}, we have the following tightness lemma:
\begin{lemma}\label{tightness lemma}
Assume $Z^\varepsilon_v$ is a solution to (\ref{eq. for weak convergence}) with $v^\varepsilon\in\mathcal{A}_N$ and $\varepsilon<1$ small enough.
There exists $\varepsilon_0>0$, such that
$\{Z^\varepsilon_v\}_{\varepsilon\in(0,\varepsilon_0)}$ is tight in the space
$$\chi=C([0,T],H^{-1})\bigcap L^2([0,T],H)\bigcap L^2_w([0,T], H^{1,1})\bigcap L^\infty_{w^*}([0,T], H^{0,1}),$$
where $L^2_w$ denotes the weak topology and $L^\infty_{w^*}$ denotes the weak star topology.
\end{lemma}
\begin{proof}
Let $k$ be the same constant as in the proof of (\ref{eq02 in lemma estimate eq. for weak convergence}) and let
\begin{align*}
K_R:=&\Big{\{} u\in C([0,T],H^{-1}): \sup_{t\in[0,T]}\|u(t)\|^2_H+\int^T_0\|u(t)\|^2_{\tilde{H}^{1,0}}dt+\|u\|_{C^\frac{1}{16}([0,T], H^{-1})}\\
&+\sup_{t\in[0,T]}e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{0,1}}+\int^T_0e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{1,1}}dt\leqslant R\Big{\}},
\end{align*}
where $C^\frac{1}{16}([0,T], H^{-1})$ is the H\"older space with the norm:
$$\|f\|_{C^\frac{1}{16}([0,T], H^{-1})}=\sup_{0\leqslant s< t\leqslant T}\frac{\|f(t)-f(s)\|_{H^{-1}}}{|t-s|^{\frac{1}{16}}}.$$
Then from the proof of \cite[Lemma 4.3]{LZZ18}, we know that for any $R>0$, $K_R$ is relatively compact in $\chi$.
Now we only need to show that for any $\delta>0$, there exists $R>0$, such that $P(Z^\varepsilon_v\in K_R)>1-\delta$ for any $\varepsilon\in(0,\varepsilon_0)$, where $\varepsilon_0$ is the constant such that Lemma \ref{estimate eq. for weak convergence} hold.
By Lemma \ref{estimate eq. for weak convergence} and Chebyshev inequality, we can choose $R_0$ large enough such that
\begin{align*}
P\left(\sup_{t\in[0,T]}\|Z^\varepsilon_v(t)\|^2_H+\int^T_0\|Z^\varepsilon_v(t)\|^2_{\tilde{H}^{1,0}}dt>\frac{R_0}{3}\right)<\frac{\delta}{4},
\end{align*}
and
\begin{align*}
P\left(\sup_{t\in[0,T]}e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{0,1}}+\int^T_0e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{1,1}}dt>\frac{R_0}{3}\right)<\frac{\delta}{4},
\end{align*}
where $k$ is the same constant as in (\ref{eq02 in lemma estimate eq. for weak convergence}).
Fix $R_0$ and let
\begin{align*}
\hat{K}_{R_0}=&\Big{\{}u\in C([0,T],H^{-1}): \sup_{t\in[0,T]}\|u(t)\|^2_H+\int^T_0\|u(t)\|^2_{\tilde{H}^{1,0}}dt\leqslant \frac{R_0}{3}\text{ and }\\
&\sup_{t\in[0,T]}e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{0,1}}+\int^T_0e^{-k\int^t_0\|\partial_1u(s)\|^2_Hds}\|u(t)\|^2_{\tilde{H}^{1,1}}dt\leqslant \frac{R_0}{3}\Big{\}}.
\end{align*}
Then $P(Z^\varepsilon_v\in C([0,T], H^{-1})\setminus \hat{K}_{R_0})<\frac{\delta}{2}$.
Now for $Z^\varepsilon_v\in\hat{K}_{R_0}$, we have $\partial_1^2Z^\varepsilon_v$ is uniformly bounded in $L^2([0,T],H^{-1})$. Similar as in Lemma \ref{solution to skeleton eq.}, $Z^\varepsilon_v$ is uniformly bounded in $L^4([0,T],H^\frac{1}{2})$ and $L^4([0,T], L^4(\mathbb{T}^2))$, thus $B(Z^\varepsilon_v)$ is uniformly bounded in $L^2([0,T],H^{-1})$. By H\"older's inequality, we have
\begin{align*}
\sup_{s,t\in[0,T],s\neq t}\frac{\|\int^t_s\partial_1^2Z^\varepsilon_v(r)+B(Z^\varepsilon_v(r))dr\|^2_{H^{-1}}}{|t-s|}\leqslant \int^T_0\|\partial_1^2Z^\varepsilon_v(r)+B(Z^\varepsilon_v(r))\|^2_{H^{-1}}dr\leqslant C(R_0),
\end{align*}
where $C(R_0)$ is a constant depend on $R_0$. For any $p\in(1,\frac{4}{3})$, by H\"older's inequality, we have
\begin{align*}
\sup_{s,t\in[0,T],s\neq t}\frac{\|\int^t_s\sigma(r,Z^\varepsilon_v(r))v^\varepsilon(r) dr\|^p_{H^{-1}}}{|t-s|^{p-1}}\leqslant& \int^T_0\|\sigma(r,Z^\varepsilon_v(r))v^\varepsilon(r)\|^p_{H^{-1}}dr\\
\leqslant& \int^T_0 \|\sigma(r,Z^\varepsilon_v(r))\|^p_{L_2(l^2,H^{-1})}\|v^\varepsilon(r)\|^p_{l^2}dr\\
\leqslant& C\int^T_0(1+\|Z^\varepsilon_v(r)\|^4_{H}+\|v^\varepsilon(r)\|^4_{l^2})dr\\
\leqslant& C(R_0),
\end{align*}
where we used Young's inequality and (A0) in the third inequality.
Moreover, for any $0\leqslant s\leqslant t\leqslant T$, by H\"older's inequality we have
\begin{align*}
E\|\int^t_s\sigma(r,Z^\varepsilon_v(r))dW(r)\|^4_{H^{-1}}\leqslant& C E\left(\int^t_s \|\sigma(r,Z^\varepsilon_v(r))\|^2_{L_2(l^2,H^{-1})}dr \right)^2\\
\leqslant& C|t-s|E\int^t_s \|\sigma(r,Z^\varepsilon_v(r))\|^4_{L_2(l^2,H^{-1})}dr\\
\leqslant& C|t-s|^2(1+E(\sup_{t\in[0,T]}\|Z^\varepsilon_v(t)\|^4_H))\\
\leqslant &C|t-s|^2,
\end{align*}
where we used (A0) in the third inequality and (\ref{eq01 in lemma estimate eq. for weak convergence}) in the last inequality. Then by Kolmogorov's continuity criterion, for any $\alpha\in(0, \frac{1}{4})$, we have
\begin{align*}
E\left(\sup_{s,t\in[0,T],s\neq t}\frac{\|\int^t_s\sigma(r,Z^\varepsilon_v(r))dW(r)\|^4_{H^{-1}}}{|t-s|^{2\alpha}}\right)\leqslant C.
\end{align*}
Choose $p=\frac{8}{7}, \alpha=\frac{1}{8}$ in the above estimates, we deduce that there exists $R>R_0$ such that
\begin{align*}
P\left(\|Z^\varepsilon_v\|_{C^\frac{1}{16}([0,T],H^{-1})}>\frac{R}{3}, Z^\varepsilon_v\in \hat{K}_{R_0}\right)\leqslant \frac{E\left(\sup_{s,t\in[0,T],s\neq t}\frac{\|Z^\varepsilon_v(t)-Z^\varepsilon_v(s)\|_{H^{-1}}}{|t-s|^{\frac{1}{16}}}1_{\{Z^\varepsilon_v\in \hat{K}_{R_0}\}}\right)}{\frac{R}{3}}< \frac{\delta}{2}.
\end{align*}
Combining the fact that $P(Z^\varepsilon_v\in C([0,T], H^{-1})\setminus \hat{K}_{R_0})<\frac{\delta}{2}$, we finish the proof.
\end{proof}
\begin{lemma}\label{weak convergence}
Assume (A0)-(A3) hold with $L_2=0$. Let $\{v^\varepsilon\}_{\varepsilon>0}\subset \mathcal{A}_N$ for some $N<\infty$. Assume $v^\varepsilon$ converge to $v$ in distribution as $S_N$-valued random elements, then
\begin{align*}
g^\varepsilon\left(W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds\right)\rightarrow g^0\left(\int^\cdot_0v(s)ds\right)
\end{align*}
in distribution as $\varepsilon\rightarrow 0$.
\end{lemma}
\begin{proof}
The proof follows essentially the same argument as in \cite[Proposition 4.7]{WZZ}.
By Lemma \ref{Girsanov thm to prove existence}, we have $Z_v^\varepsilon=g^\varepsilon\left(W(\cdot)+\frac{1}{\sqrt{\varepsilon}}\int^\cdot_0v^\varepsilon(s)ds\right)$. By a similar but simple argument as in the proof of Lemmas \ref{solution to skeleton eq.} and \ref{estimate eq. for weak convergence}, there exists a unique strong solution $Y^\varepsilon\in L^{\infty}([0,T],\tilde{H}^{0,1})\bigcap L^2([0,T],\tilde{H}^{1,1})\bigcap C([0,T],H^{-1})$ satisfying
\begin{align*}
dY^\varepsilon(t)=&\partial_1^2Y^\varepsilon(t)dt+\sqrt{\varepsilon}\sigma(t,Z^\varepsilon_v(t))dW(t),\\
\text{div }Y^\varepsilon =&0,\\
Y^\varepsilon(0)=&0,
\end{align*}
and
\begin{align*}
\lim_{\varepsilon\rightarrow 0}\left[E\sup_{t\in[0,T]}\|Y^\varepsilon(t)\|_{H}^2+E\int^T_0\|Y^\varepsilon(t)\|^2_{\tilde{H}^{1,0}}dt\right]=0,
\end{align*}
\begin{align*}
\lim_{\varepsilon\rightarrow 0}\left[E\sup_{t\in[0,T]}(e^{-kg(t)}\|Y^\varepsilon(t)\|_{\tilde{H}^{0,1}}^2)+E\int^T_0e^{-kg(t)}\|Y^\varepsilon(t)\|^2_{\tilde{H}^{1,1}}dt\right]=0,
\end{align*}
where $g(t)=\int^t_0\|Z^\varepsilon_v(s)\|^2_Hds$ and $k$ are the same as in (\ref{eq02 in lemma estimate eq. for weak convergence}).
Set $$\Xi:=\left(\chi, \mathcal{S}_N, L^\infty([0,T],H)\bigcap L^2([0,T],\tilde{H}^{1,0}) \bigcap C([0,T].H^{-1})\right).$$ The above limit implies that $Y^\varepsilon\rightarrow 0$ a.s. in $ L^\infty([0,T],H)\bigcap L^2([0,T],\tilde{H}^{1,0}) \bigcap C([0,T].H^{-1})$ as $\varepsilon\rightarrow 0$ (in the sense of subsequence). By Lemma \ref{tightness lemma} the family $\{(Z^\varepsilon_v,v^\varepsilon)\}_{\varepsilon\in(0,\varepsilon_0)}$ is tight in $(\chi, \mathcal{S}_N)$. Let $(Z_v, v, 0)$ be any limit point of $\{(Z^\varepsilon_v, v^\varepsilon, Y^\varepsilon)\}_{\varepsilon\in(0,\varepsilon_0)}$. Our goal is to show that $Z_v$ has the same law as $g^0\left(\int^\cdot_0v(s)ds\right)$ and $Z^\varepsilon_v$ convergence in distribution to $Z_v$ in the space $ L^{\infty}([0,T],H)\bigcap L^2([0,T],\tilde{H}^{1,0})\bigcap C([0,T],H^{-1})$.
By the Skorokhod Theorem, there exists a stochastic basis $(\tilde{\Omega}, \tilde{\mathcal{F}}, \{\tilde{\mathcal{F}}_t\}_{t\in[0,T]}, \tilde{P})$ and, on this basis, $\Xi$-valued random variables $(\tilde{Z}_v, \tilde{v}, 0)$, $(\tilde{Z}^\varepsilon_v, \tilde{v}^\varepsilon, \tilde{Y}^\varepsilon)$, such that $(\tilde{Z}^\varepsilon_v, \tilde{v}^\varepsilon, \tilde{Y}^\varepsilon)$ (respectively $(\tilde{Z}_v, \tilde{v}, 0)$) has the same law as $(Z^\varepsilon_v,v^\varepsilon, Y^\varepsilon)$ (respectively $(Z_v,v,0)$), and $(\tilde{Z}^\varepsilon_v, \tilde{v}^\varepsilon, \tilde{Y}^\varepsilon)\rightarrow (\tilde{Z}_v, \tilde{v}, 0)$, $\tilde{P}$-a.s.
We have
\begin{equation}\label{weak convergence step 1}\aligned
d(\tilde{Z}^\varepsilon_v(t)-\tilde{Y}^\varepsilon(t))=&\partial_1^2(\tilde{Z}^\varepsilon_v(t)-\tilde{Y}^\varepsilon(t))dt-B(\tilde{Z}^\varepsilon_v(t))dt+\sigma(t,\tilde{Z}^\varepsilon_v(t))\tilde{v}^\varepsilon(t)dt,\\
\tilde{Z}^\varepsilon_v(0)-\tilde{Y}^\varepsilon(0)=&u_0,
\endaligned
\end{equation}
and
\begin{align*}
&P(\tilde{Z}^\varepsilon_v-\tilde{Y}^\varepsilon\in L^\infty([0,T],H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1}))\\
=&P({Z}^\varepsilon_v-{Y}^\varepsilon\in L^\infty([0,T],H)\bigcap L^2([0,T], \tilde{H}^{1,0})\bigcap C([0,T],H^{-1}))\\
=&1.
\end{align*}
Let $\tilde{\Omega}_0$ be the subset of $\tilde{\Omega}$ such that for $\omega\in\tilde{\Omega}_0$,
$$(\tilde{Z}^\varepsilon_v, \tilde{v}^\varepsilon, \tilde{Y}^\varepsilon)(\omega)\rightarrow (\tilde{Z}_v,\tilde{v}, 0)(\omega) \text{ in }\Xi,$$
and
$$e^{-k\int^{\cdot}_0\|\tilde{Z}^\varepsilon_v(\omega, s)\|^2_Hds}\tilde{Y}^\varepsilon(\omega)\rightarrow 0 \text{ in } L^{\infty}([0,T],\tilde{H}^{0,1})\bigcap L^2([0,T],\tilde{H}^{1,1})\bigcap C([0,T],H^{-1}),$$
then $P(\tilde{\Omega}_0)=1$. For any $\omega\in\tilde{\Omega}_0$, fix $\omega$, we have $\sup_{\varepsilon}\int^T_0\|\tilde{Z}^\varepsilon_v(\omega,s)\|_H^2ds<\infty$, then we deduce that
\begin{equation}\label{convergence of Yvarepsilon to 0}
\lim_{\varepsilon\rightarrow 0} \left(\sup_{t\in[0,T]}\|\tilde{Y}^\varepsilon(\omega,t)\|_{\tilde{H}^{0,1}}+\int^T_0\|\tilde{Y}^\varepsilon(\omega,t)\|^2_{\tilde{H}^{1,1}}dt\right)=0.
\end{equation}
Now we show that
\begin{equation}\label{weak convergence step 2}
\sup_{t\in[0,T]}\|\tilde{Z}^\varepsilon_v(\omega,t)-\tilde{Z}_v(\omega,t)\|^2_H+\int^T_0\|\tilde{Z}^\varepsilon_v(\omega, t)-\tilde{Z}_v(\omega,t)\|^2_{\tilde{H}^{1,0}}dt\rightarrow 0\text{ as }\varepsilon\rightarrow 0.
\end{equation}
Let $Z^\varepsilon=\tilde{Z}^\varepsilon_v(\omega)-\tilde{Y}^\varepsilon(\omega)$, then by (\ref{weak convergence step 1}) we have
\begin{equation}
dZ^\varepsilon(t)=\partial_1^2Z^\varepsilon(t) dt-B(Z^\varepsilon(t)+\tilde{Y}^\varepsilon(t))dt+\sigma(t,Z^\varepsilon(t)+\tilde{Y}^\varepsilon(t))\tilde{v}^\varepsilon(t)dt.
\end{equation}
Since $Z^\varepsilon(\omega)\rightarrow \tilde{Z}_v(\omega)$ in $\chi$, by a very similar argument as in Lemma \ref{good rate function} we deduce that $\tilde{Z}_v=z^{\tilde{v}}=g^0\left(\int^\cdot_0{\tilde{v}}(s)ds\right)$. Moreover, note that $\tilde{Z}^\varepsilon_v(\omega)\rightarrow z^{\tilde{v}}(\omega)$ weak star in $L^{\infty}([0,T],\tilde{H}^{0,1})$, then the uniform boundedness principle implies that
\begin{equation}\label{uniform bded of Z in H01}
\sup_{\varepsilon}\sup_{t\in[0,T]}\|\tilde{Z}^\varepsilon_v(\omega)\|_{\tilde{H}^{0,1}}<\infty.
\end{equation}
Let $w^\varepsilon=Z^\varepsilon-z^{\tilde{v}}$, then we have
\begin{align*}
&\|w^\varepsilon(t)\|^2_H+2\int^t_0\|\partial_1w^\varepsilon(s)\|^2_Hds\\
=&-2\int^t_0\langle w^\varepsilon(s), B(Z^\varepsilon+\tilde{Y}^\varepsilon)-B(z^{\tilde{v}})\rangle ds+2\int^t_0\langle w^\varepsilon(s), \sigma(s,Z^\varepsilon+\tilde{Y}^\varepsilon){\tilde{v}}^\varepsilon(s)-\sigma(s,z^{\tilde{v}}){\tilde{v}}(s)\rangle ds.\\
\end{align*}
By Lemmas \ref{anisotropic estimate for b} and \ref{b(u,v,w)}, we have
\begin{align*}
&\int^t_0\langle w^\varepsilon(s), B(Z^\varepsilon+\tilde{Y}^\varepsilon)-B(z^{\tilde{v}})\rangle ds\\
=&\int^t_0b(\tilde{Y}^\varepsilon,z^{\tilde{v}},w^\varepsilon)+b(\tilde{Y}^\varepsilon,\tilde{Y}^\varepsilon,w^\varepsilon)+b(w^\varepsilon,\tilde{Y}^\varepsilon+z^{\tilde{v}},w^\varepsilon)+b(z^{\tilde{v}},\tilde{Y}^\varepsilon, w^\varepsilon)ds\\
\leqslant &\int^t_0[\frac{1}{2}\|\partial_1w^\varepsilon(s)\|^2_H+\frac{1}{2}\|\tilde{Y}^\varepsilon(s)\|^2_{\tilde{H}^{1,1}}+C(1+\|z^{\tilde{v}}(s)\|_{\tilde{H}^{1,1}}^2+\|\tilde{Y}^\varepsilon(s)\|^2_{\tilde{H}^{1,1}})\|w^\varepsilon(s)\|^2_H]ds\\
&+C\int^t_0\|\tilde{Y}^\varepsilon(s)\|^2_{\tilde{H}^{1,1}}\|w^\varepsilon(s)\|_Hds\\
\leqslant& \int^t_0\frac{1}{2}\|\partial_1 w^\varepsilon(s)\|^2_Hds+C\int^t_0\|\tilde{Y}^\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds+C\int^t_0(1+\|z^{\tilde{v}}(s)\|_{\tilde{H}^{1,1}}^2)\|w^\varepsilon(s)\|^2_Hds,
\end{align*}
where we used the fact that by \eqref{convergence of Yvarepsilon to 0} and \eqref{uniform bded of Z in H01} $w^\varepsilon$ are uniformly bounded in $L^\infty([0,T], H)$ in the last inequality.
By (A1) and (A3) with $L_2=0$ we have
\begin{align*}
&\int^t_0\langle w^\varepsilon(s), \sigma(s,Z^\varepsilon+\tilde{Y}^\varepsilon)v^\varepsilon(s)-\sigma(s,z^{\tilde{v}}){\tilde{v}}(s)\rangle ds\\
=&\int^t_0\langle w^\varepsilon(s), (\sigma(s,Z^\varepsilon+\tilde{Y}^\varepsilon)-\sigma(s,z^{\tilde{v}})) {\tilde{v}}^\varepsilon(s)\rangle ds+\int^t_0\langle w^\varepsilon(s), \sigma(s,z^{\tilde{v}})({\tilde{v}}^\varepsilon(s)- {\tilde{v}}(s))\rangle ds\\
\leqslant &C\int^t_0(\|w^\varepsilon(s)\|_H\|{\tilde{v}}^\varepsilon(s)\|_{l^2}(\|w^
\varepsilon(s)\|^2_H+\|\tilde{Y}^\varepsilon(s)\|^2_H)^\frac{1}{2} ds\\
&+\int^t_0\|w^\varepsilon(s)\|_H\|{\tilde{v}}^\varepsilon(s)-{\tilde{v}}(s)\|_{l^2}(K_0+K_1\|z^{\tilde{v}}(s)\|^2_H+K_2\|\partial_1 z^{\tilde{v}}(s)\|^2_H)^\frac{1}{2} ds\\
\leqslant &C N^\frac{1}{2}\left(\int^t_0(\|w^
\varepsilon(s)\|^2_H+\|\tilde{Y}^\varepsilon(s)\|^2_H ds\right)^\frac{1}{2}\\
&+CN^\frac{1}{2}\left(\int^t_0\|w^\varepsilon(s)\|_H^2(K_0+K_1\|z^{\tilde{v}}(s)\|^2_H+K_2\|\partial_1 z^{\tilde{v}}(s)\|^2_H) ds\right)^\frac{1}{2},
\end{align*}
where we used the fact that $w^\varepsilon$ are uniformly bounded in $L^\infty([0,T], H)$ and that ${\tilde{v}}^\varepsilon$, ${\tilde{v}}$ are in $\mathcal{A}_N$.
Thus we have
\begin{align*}
&\|w^\varepsilon(t)\|^2_H+\int^t_0\|\partial_1w^\varepsilon(s)\|^2_Hds\\
\leqslant& C\int^t_0(1+\|z^{\tilde{v}}(s)\|_{\tilde{H}^{1,1}}^2)\|w^\varepsilon(s)\|^2_H ds+C\int^t_0\|\tilde{Y}^\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\\
&+CN^\frac{1}{2}\left(\int^t_0(\|w^\varepsilon(s)\|^2_H+\|\tilde{Y}^\varepsilon(s)\|^2_H)ds\right)^\frac{1}{2}+CN^\frac{1}{2}\left(\int^t_0(1+\|z^{\tilde{v}}(s)\|^2_{\tilde{H}^{1,1}})\|w^\varepsilon(s)\|^2_Hds\right)^\frac{1}{2}.
\end{align*}
Since $Z^\varepsilon(\omega)\rightarrow z^{\tilde{v}}(\omega)$ strongly in $L^2([0,T],H)$ and $\tilde{Y}^\varepsilon\rightarrow 0$ in $L^2([0,T],\tilde{H}^{1,1})$, the same argument used in Lemma \ref{good rate function} implies
\begin{equation}\label{weak convergence step 3}
\sup_{t\in[0,T]}\|\tilde{Z}^\varepsilon_v(\omega,t)-z^{\tilde{v}}(\omega,t)\|^2_H+\int^T_0\|\tilde{Z}^\varepsilon_v(\omega, t)-z^{\tilde{v}}(\omega,t)\|^2_{\tilde{H}^{1,0}}dt\rightarrow 0\text{ as }\varepsilon\rightarrow 0.
\end{equation}
The proof is thus complete.
\end{proof}
\begin{proof}[{Proof of Theorem \ref{main result LDP}}]
The result holds from Lemmas \ref{weak convergence method}, \ref{good rate function} and \ref{weak convergence}.
\end{proof}
\section{Small time asymptotics}
In this section, we consider the small time behaviour. We need the following additional assumption (A3') and (A4). Note that (A3') is stronger than (A3).
(A3') $\|\sigma(t,u)-\sigma(s,v)\|^2_{L_2(l^2, H)}\leqslant L_0|t-s|^\alpha+L_1\|u-v\|^2_{H}$.
(A4) $\|\sigma(t,u)\|^2_{L_2(l^2, V)}\leqslant \overline{K}_0+\overline{K}_1\|u\|_{V}^2$.
\begin{remark} \text{ }
A typical example of $\sigma$ is similar as in \cite[Remark 4.2]{LZZ18}. For $u=(u^1,u^2)\in H^{1,1}$ and $y\in l^2$, let
$$\sigma(t,u)y=\sum^\infty_{k=1}b_kg(u)\langle y, \psi_k\rangle_{l^2},$$
where $\{\psi_k\}_{k\geqslant 0}$ is the orthonormal basis of $l^2$, $\{b_k\}_{k\geqslant 0}$ are functions from $\mathbb{T}^2$ to $\mathbb{R}$ and $g$ is a differentiable function from $\mathbb{R}^2$ to $\mathbb{R}$. Assume that $|g(x)-g(y)|\leqslant C|x-y|$ for all $x,y\in\mathbb{R}^2$ and some constant $C$ depends on $g$. Also suppose that $\textsl{div}(b_kg(u))=0$ and $b_k, \partial_1 b_k, \partial_2b_k\in L^\infty$, $\sum^\infty_{k=1}\|b_k\|_{L^\infty}^2\leqslant M$, $\sum^\infty_{k=1}\|\partial_1 b_k\|_{L^\infty}^2\leqslant M$ and $\sum^\infty_{k=1}\|\partial_2 b_k\|_{L^\infty}^2\leqslant M$. From the conditions of $g$, it is easy to obtain $|g(u)|\leqslant C|u|+C$, $|\partial_1g(u)|\leqslant C$ and $|\partial_2g(u)|\leqslant C$. In this case,
$\sigma$ satisfies (A0)-(A4) and (A3'):
\begin{align*}
\|\sigma(t,u)\|^2_{L_2(l^2, H)}\leqslant& \sum^\infty_{k=1}\|b_kg(u)\|^2_H\leqslant CM(\|u\|^2_H+1);\\
\|\sigma(t,u)\|^2_{L_2(l^2, H^{0,1})}\leqslant&\sum^\infty_{k=1}\|b_kg(u)\|^2_H+\sum^\infty_{k=1}\|\partial_2(b_kg(u))\|^2_H\\
\leqslant& CM(\|u\|^2_H+1)+\sum^\infty_{k=1}\|\partial_2b_kg(u)+b_k(\partial_1g(u)\partial_2u^1+\partial_2g(u)\partial_2u^2)\|^2_H\\
\leqslant& CM(1+\|u\|^2_H+\|\partial_2u\|^2_H);\\
\|\sigma(t,u)\|^2_{L_2(l^2, V)}\leqslant &C{M}(\|u\|^2_H+1)+\sum^\infty_{k=1}\|\partial_1(b_kg(u))\|^2_H+\sum^\infty_{k=1}\|\partial_2(b_kg(u))\|^2_H\\
\leqslant& C{M}(1+\|u\|^2_H+\|\partial_1u\|^2_H+\|\partial_2u\|^2_H);\\
\|\sigma(t,u)-\sigma(s,v)\|^2_{L_2(l^2, H)}\leqslant&{M}C\|u-v\|^2_H.\\
\end{align*}
\end{remark}
Let $\varepsilon>0$ and $u$ be the solution to (\ref{equation after projection}), by the scaling property of the Brownian motion, $u(\varepsilon t)$ coincides in law with the solution to the following equation:
\begin{equation}\label{eq. for small time}\aligned
&du_\varepsilon=\varepsilon\partial_1^2u_\varepsilon dt-\varepsilon B(u_\varepsilon)dt+\sqrt{\varepsilon}\sigma(\varepsilon t, u_\varepsilon)dW(t),\\
&u_\varepsilon(0)=u_0.
\endaligned
\end{equation}
Define a functional $I^{u_0}$ on $L^\infty([0,T], H)\bigcap C([0,T],H^{-1})$ by
$$I^{u_0}(g)=\inf_{h\in\Gamma_g}\{\frac{1}{2}\int^T_0\|{h}(t)\|_{l^2}^2dt\},$$
where
$$\Gamma_g=\{h\in L^2([0,T],l^2): g(t)=u_0+\int^t_0\sigma(0,g(s)){h}(s)ds, \text{ }t\in[0,T]\}.$$
The main theorem of this section is the following one:
\begin{Thm}\label{main}
Assume (A0), (A1), (A2), (A3'), (A4) hold with $K_2=\tilde{K}_2=0$ and $u_0\in \tilde{H}^{0,1}$, then $u_\varepsilon$ satisfies a large deviation principle on the space $L^\infty([0,T], H)\bigcap C([0,T],H^{-1})$ with the good rate function $I^{u_0}$.
\end{Thm}
We aim to prove that $u_\varepsilon$ is exponentially equivalent to the solution to the following equation:
\begin{equation}\label{eq for v}
v_\varepsilon(t)=u_0+\sqrt{\varepsilon}\int^t_0\sigma(\varepsilon s, v_\varepsilon(s))dW(s).
\end{equation}
Because of the non-linear form $b(\cdot,\cdot,\cdot)$ and the anisotropic viscosity, we split the proof into several lemmas.
\begin{lemma}\label{LDP for v}
Assume $u_0\in \tilde{H}^{0,1}$, then $v_\varepsilon$ satisfies a large deviation principle on the space $L^\infty([0,T], H)\bigcap C([0,T],H^{-1})$ with the good rate function $I^{u_0}$.
\end{lemma}
\begin{proof}
Let $z_\varepsilon$ be the solution to the stochastic equation:
$$z_\varepsilon(t)=u_0+\sqrt{\varepsilon}\int^t_0\sigma(0,z_\varepsilon(s))dW(s).$$
By \cite[Theorem 12.11]{DZ92}, we know that $z_\varepsilon$ satisfies a large deviation principle with the good rate function $I^{u_0}$. Applying It\^o's formula to $\|v_{\varepsilon}-z_\varepsilon\|^2_H$, we obtain
\begin{align*}
\|v_{\varepsilon}(t)-z_\varepsilon(t)\|^2_H=&2\sqrt{\varepsilon}\int^t_0\langle v_{\varepsilon}(s)-z_\varepsilon(s),[ \sigma(\varepsilon s, v_{\varepsilon}(s))-\sigma(0,z_\varepsilon(s)) ]dW(s)\rangle\\
&+\varepsilon\int^t_0\|\sigma(\varepsilon s, v_{\varepsilon}(s))-\sigma(0,z_\varepsilon(s))\|^2_{L_2(l^2,H)}ds.
\end{align*}
Then by (A3') and Lemma \ref{martingale lemma}, we get for $p\geqslant 2$,
\begin{align*}
&\left(E[\sup_{0\leqslant t\leqslant T}\|v_{\varepsilon}(t)-z_\varepsilon(t)\|^{2p}_H]\right)^\frac{2}{p}\\
\leqslant &C\varepsilon \left(E[\sup_{0\leqslant t\leqslant T}\int^t_0\langle v_{\varepsilon}(s)-z_\varepsilon(s),(\sigma(\varepsilon s, v_{\varepsilon}(s))-\sigma(0,z_\varepsilon(s)) )dW(s)\rangle ]^{p}\right)^\frac{2}{p}\\
&+C\varepsilon^2 \left(E[\int^T_0\|\sigma(\varepsilon s, v_{\varepsilon}(s))-\sigma(0,z_\varepsilon(s))\|^2_{L_2(l^2,H)}ds ]^p\right)^\frac{2}{p}\\
\leqslant &C\varepsilon p\left(E\left[\int^T_0\|v_{\varepsilon}(s)-z_\varepsilon(s)\|^2_H\|\sigma(\varepsilon s, v_{\varepsilon}(s))-\sigma(0,z_\varepsilon(s))\|^2_{L_2(l^2, H)}ds\right]^\frac{p}{2}\right)^\frac{2}{p}\\
&+C\varepsilon^2\left(\varepsilon^{2\alpha} T^{2+2\alpha}+T\int^T_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{\varepsilon}(l)-z_\varepsilon(l)\|_H^{2p}]\right)^\frac{2}{p} ds\right)\\
\leqslant&C\varepsilon p\left(\varepsilon^{2\alpha}+\int^T_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{\varepsilon}(l)-z_\varepsilon(l)\|_H^{2p}]\right)^\frac{2}{p} ds\right)\\
&+C\varepsilon^2\left(\varepsilon^{2\alpha}+\int^T_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{\varepsilon}(l)-z_\varepsilon(l)\|_H^{2p}]\right)^\frac{2}{p} ds\right).
\end{align*}
By Gronwall's inequality, we have
$$\left(E[\sup_{0\leqslant t\leqslant T}\|v_{\varepsilon}(t)-z_\varepsilon(t)\|^{2p}_H]\right)^{\frac{2}{p}}\leqslant C(\varepsilon^{1+2\alpha} p+\varepsilon^{2+2\alpha})e^{C(\varepsilon p+\varepsilon^2)}.$$
Then Chebyshev's inequality implies that
\begin{align*}
\varepsilon\log P(\sup_{0\leqslant t\leqslant T}\|v_{\varepsilon}(t)-z_\varepsilon(t)\|^{2}_H>\delta)\leqslant& \varepsilon \log E[\sup_{0\leqslant t\leqslant T}\|v_{\varepsilon}(t)-z_\varepsilon(t)\|^{2p}_H]-\varepsilon p\log \delta\\
\leqslant & \frac{\varepsilon p}{2} (C+C\varepsilon p +C\varepsilon^2+\log (\varepsilon^{1+2\alpha} p+\varepsilon^{2+2\alpha})-2\log \delta).\\
\end{align*}
Let $p=\frac{1}{\varepsilon}$ and $\varepsilon\rightarrow 0$, we get that $v_\varepsilon$ and $z_\varepsilon$ are exponentially equivalent, which by Lemma \ref{EXEQ} implies the result.
\end{proof}
\begin{lemma}\label{lemma1}
Let $F_{u_\varepsilon}(t)=\sup_{0\leqslant s\leqslant t}\|u_\varepsilon(s)\|^2_{H}+\varepsilon\int^t_0\|\partial_1 u_\varepsilon(s)\|^2_{H}ds$, then
$$\lim_{M\rightarrow \infty}\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P(F_{u_\varepsilon}(T)>M)=-\infty.$$
\end{lemma}
\begin{proof}
Since $b(u_\varepsilon,u_\varepsilon,u_\varepsilon)=0$, applying It\^o's formula to $\|u_\varepsilon(t)\|^2_H$, we have
\begin{align*}
&\|u_\varepsilon(t)\|^2_H+2\varepsilon\int^t_0\|\partial_1 u_\varepsilon(s)\|^2_{H}ds\\
=&\|u_0\|^2_H+2\sqrt{\varepsilon}\int^t_0\langle u_\varepsilon(s), \sigma(\varepsilon s, u_\varepsilon(s)dW(s)\rangle+\varepsilon\int^t_0\|\sigma(\varepsilon s, u_\varepsilon(s))\|^2_{L_2(l^2,H)}ds.
\end{align*}
Then it follows from (A1) with $K_2=0$ that
\begin{align*}
\|u_\varepsilon(t)\|^2_H+\varepsilon\int^t_0\|\partial_1 u_\varepsilon(s)\|^2_{H}ds\leqslant& \|u_0\|^2_{\tilde{H}^{0,1}}+C\varepsilon t+ C\varepsilon \int^t_0\|u_\varepsilon(s)\|^2_Hds\\
&+2\sqrt{\varepsilon}\int^t_0\langle u_\varepsilon, \sigma(\varepsilon s, u_\varepsilon(s))dW(s)\rangle.
\end{align*}
Take supremum over $t$, for $p\geqslant 2$, we have
\begin{align*}
(E[F_{u_\varepsilon}(T)]^p)^\frac{1}{p}\leqslant & \|u_0\|^2_{\tilde{H}^{0,1}}+C\varepsilon T+C\varepsilon \int^T_0 (E[F_{u_\varepsilon}(t)]^p)^\frac{1}{p}dt\\
&+2\sqrt{\varepsilon}(E[\sup_{0\leqslant t\leqslant T}|\int^t_0\langle u_\varepsilon, \sigma(\varepsilon s, u_\varepsilon(s))dW(s)\rangle|]^p)^\frac{1}{p}.
\end{align*}
For the term in the last line, by Lemma \ref{martingale lemma} and \cite[(3.12)]{XZ08}, we have
\begin{align*}
&2\sqrt{\varepsilon}(E[\sup_{0\leqslant t\leqslant T}|\int^t_0\langle u_\varepsilon, \sigma(\varepsilon s, u_\varepsilon(s))dW(s)\rangle|]^p)^\frac{1}{p}\\
\leqslant& C\sqrt{\varepsilon p} \left[\int^T_01+(E\|u_\varepsilon(s)\|^{2p}_H)^\frac{2}{p} ds\right]^\frac{1}{2}.
\end{align*}
Combining the above estimate, we arrive at
\begin{align*}
\left(E[F_{u_\varepsilon}(T)]^p\right)^\frac{2}{p}\leqslant &C\left(\|u_0\|^2_{\tilde{H}^{0,1}}+\varepsilon T\right)^2+C\varepsilon^2\int^T_0\left(E[F_{u_\varepsilon}(t)]^p\right)^\frac{2}{p}ds\\
&+C\varepsilon p T+C\varepsilon p\int^T_0\left(E[F_{u_\varepsilon}(t)]^p\right)^\frac{2}{p}dt.
\end{align*}
Then the Gronwall's inequality implies
\begin{align*}
\left(E[F_{u_\varepsilon}(T)]^p\right)^\frac{2}{p}\leqslant C\left[\|u_0\|^4_{\tilde{H}^{0,1}}+\varepsilon^2+\varepsilon p \right]e^{C\varepsilon^2+C\varepsilon p}.
\end{align*}
Let $p=\frac{1}{\varepsilon}$, by Chebyshev's inequality, we have
\begin{align*}
&\varepsilon\log P(F_{u_\varepsilon}(T)>M)\\
\leqslant& -\log M+ \log \left(E[F_{u_\varepsilon}(T)]^p\right)^\frac{1}{p}\\
\leqslant &-\log M+ \log\sqrt{\|u_0\|^4_{\tilde{H}^{0,1}}+\varepsilon^2 +1}+ C(\varepsilon^2+1).
\end{align*}
Take supremum over $\varepsilon$ and let $M\rightarrow \infty$, we finish the proof.
\end{proof}
\begin{lemma}\label{stopping time}
For M>0, define a random time
$$\tau_{M,\varepsilon}=T\wedge\inf\{t:\|u_\varepsilon(t)\|^2_{H}>M,\text{ or }\varepsilon\int^t_0\|\partial_1 u_\varepsilon(s)\|^2_{H}ds>M\}.$$
Then $\tau_{M,\varepsilon}$ is a stopping time with respect to $\mathcal{F}_{t+}=\cap_{s>t}\mathcal{F}_s$.
Similarly, Let
$$\tau'_{M,\varepsilon}= T\wedge\inf\{t: \|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}>M,\text{ or }\varepsilon\int^t_0\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds>M\},$$
then $\tau'_{M,\varepsilon}$ is a stopping time with respect to $\mathcal{F}_{t+}$.
\end{lemma}
\begin{proof}
The problem comes with the continuity of $u_\varepsilon(t)$. More precisely, since $\int^t_0\|\partial_1 u_\varepsilon(s)\|^2_{H}ds$ is a continuous adapted process, we only need to prove that $\hat{\tau}=\inf\{t>0: \|u_\varepsilon(t)\|_H^2>M\}$ is a stopping time.
Since $u_\varepsilon\in L^\infty([0,T],H)\bigcap C([0, T], H^{-1})$,
$u_\varepsilon(t)$ is weakly continuous on $H$, which implies the lower semi-continuity of $u_{\varepsilon}$ on $H$.
By definition of $\hat{\tau}$, for $t>0$
$$\bigcap_{s\in(0,t]}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}\subset \{\hat{\tau}\geqslant t\}\subset \bigcap_{s\in(0,t)}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}.$$
On the contrary, if $\omega\in\{\hat{\tau}\geqslant t\}$, for any $s<t$, $\|u_\varepsilon(s)(\omega)\|^2_H\leqslant M$. Then lower semi-continuity implies $$\|u_\varepsilon(t)(\omega)\|^2_H\leqslant \liminf_{s<t,s\rightarrow t}\|u_\varepsilon(s)\|^2_H\leqslant M.$$
Hence we have
$$\{\hat{\tau}\geqslant t\}=\bigcap_{s\in(0,t]}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}.$$
Note that for $\omega\in\bigcap_{s\in(0,t]\cap\mathbb{Q}}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}$, we have for any $s\in(0,t]$, by the lower semi-continuity, $$\|u_\varepsilon(s)(\omega)\|^2_H\leqslant \liminf_{s'\rightarrow s}\|u_\varepsilon(s')\|^2_H\leqslant \liminf_{s'\rightarrow s, s'\in\mathbb{Q}}\|u_\varepsilon(s')\|^2_H\leqslant M,$$
which means
$$\bigcap_{s\in(0,t]}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}=\bigcap_{s\in(0,t]\cap\mathbb{Q}}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}.$$
Then we have for $t>0$
$$\{\hat{\tau}\geqslant t\}=\bigcap_{s\in(0,t]}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}=\bigcap_{s\in(0,t]\cap\mathbb{Q}}\{\|u_\varepsilon(s)\|^2_H\leqslant M\}\in\mathcal{F}_t,$$
which implies the result.
For $\tau'_{M,\varepsilon}$, the result follows from the fact that $u_\varepsilon$ is weakly continuous in $\tilde{H}^{0,1}$ since $u_\varepsilon\in L^\infty([0,T],\tilde{H}^{0,1})\bigcap C(0,T],H^{-1})$.
\end{proof}
\begin{lemma}\label{lemma2}
Let $G_{u_\varepsilon}(t)=\sup_{0\leqslant s\leqslant t}\|u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^{t}_0\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds$.
For fixed $M_1$, we have
$$\lim_{M\rightarrow \infty}\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P(G_{u_\varepsilon}(\tau_{M_1,\varepsilon})>M)=-\infty.$$
\end{lemma}
\begin{proof}
Let $k$ be a positive constant and $f_\varepsilon(t)=1+\|\partial_1u_\varepsilon(t)\|^2_H$. Applying It\^o's formula to $e^{-k\varepsilon\int^t_0f_\varepsilon(s) ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}$, we obtain
\begin{align*}
&e^{-k\varepsilon\int^t_0f_\varepsilon(s) ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}+2\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}(\|\partial_1u_\varepsilon(s)\|^2_{H}+\|\partial_1\partial_2u_\varepsilon(s)\|^2_H)ds\\
=&\|u_0\|^2_{\tilde{H}^{0,1}}-k\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}f_\varepsilon(s)\| u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}}ds\\
&-2\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\langle\partial_2u_\varepsilon(s), \partial_2(u_\varepsilon\cdot\nabla u_\varepsilon)(s)\rangle ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\langle u_\varepsilon(s), \sigma(\varepsilon s, u_\varepsilon(s))dW(s)\rangle_{\tilde{H}^{0,1}}\\
&+\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|\sigma(\varepsilon s, u_\varepsilon(s))\|^2_{L_2(l^2, {\tilde{H}^{0,1}})}ds.
\end{align*}
The fourth and the fifth line can be dealt in the same way as in the proof of Lemma \ref{lemma1}. For the third line, by Lemma \ref{estimate for b with partial_2}, we have
\begin{align*}
&|\langle\partial_2u_\varepsilon, \partial_2(u_\varepsilon\cdot\nabla u_\varepsilon)\rangle|
\leqslant \frac{1}{2}\|\partial_1\partial_2 u_\varepsilon\|_H^2+ C_1f_\varepsilon\|\partial_2u_\varepsilon\|_H^2,
\end{align*}
where $C_1$ is a constant.
Therefore by (A2) with $\tilde{K}_2=0$ we get
\begin{align*}
&e^{-k\varepsilon\int^t_0f_\varepsilon(s) ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\\
\leqslant&\|u_0\|^2_{\tilde{H}^{0,1}}-k\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}f_\varepsilon(s)\|u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}}ds\\
&+2C_1\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}f_\varepsilon(s)\|u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}} ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\langle u_\varepsilon(s), \sigma(\varepsilon s, u_\varepsilon(s))dW(s)\rangle_{\tilde{H}^{0,1}}\\
&+\varepsilon\int^t_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}[\tilde{K_0}+(\tilde{K}_1+1)\|u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}}]ds.
\end{align*}
For the last second line, similar to \cite[(3.12)]{XZ08}, we have
\begin{align*}
&2\sqrt{\varepsilon}(E[\sup_{0\leqslant s\leqslant t}|\int^s_0e^{-k\varepsilon\int^r_0f_\varepsilon(l)dl}\langle u_\varepsilon(r), \sigma(\varepsilon r, u_\varepsilon(r))dW(r)\rangle_{\tilde{H}^{0,1}}|]^p)^\frac{1}{p}\\
\leqslant& C\sqrt{\varepsilon p} (E[\int^t_0e^{-2k\varepsilon\int^r_0f_\varepsilon(l)dl}\| u_\varepsilon(r)\|_{\tilde{H}^{0,1}}^2\|\sigma(\varepsilon r, u_\varepsilon(r))\|^2_{L_2(l^2,\tilde{H}^{0,1})}dr]^\frac{p}{2})^\frac{1}{p}\\
\leqslant& C\sqrt{\varepsilon p} (E[\int^t_0e^{-2k\varepsilon\int^r_0f_\varepsilon(l)dl}\| u_\varepsilon(r)\|_{\tilde{H}^{0,1}}^2(1+\|u_\varepsilon(r)\|^2_{\tilde{H}^{0,1}})dr]^\frac{p}{2})^\frac{1}{p}\\
\leqslant& C\sqrt{\varepsilon p} (E[\int^t_0e^{-2k\varepsilon\int^r_0f_\varepsilon(l)dl}(1+\|u_\varepsilon(r)\|^4_{\tilde{H}^{0,1}})dr]^\frac{p}{2})^\frac{1}{p}\\
\leqslant& C\sqrt{\varepsilon p} \left[\int^t_01+(E[e^{-pk\varepsilon\int^r_0f_\varepsilon(l)dl}\|u_\varepsilon(s)\|^{2p}_{\tilde{H}^{0,1}}])^\frac{2}{p} ds\right]^\frac{1}{2},
\end{align*}
where we used (A2) with $K_2=0$ in the third line.
Let $k>2C_1$ and using Lemma \ref{martingale lemma}, we have for $p\geqslant 2$
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant t\wedge\tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^{t\wedge\tau_{M_1,\varepsilon}}_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\right]^p\right)^\frac{2}{p}\\
\leqslant&C(\|u_0\|^2_{\tilde{H}^{0,1}}+\varepsilon)^2+C\varepsilon^2\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s\wedge\tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^r_0f_\varepsilon(l)dl}\|u_\varepsilon(r)\|^2_{\tilde{H}^{0,1}}\right]^p\right)^\frac{2}{p}ds\\
&+C\varepsilon p+C\varepsilon p\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s\wedge\tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^r_0f_\varepsilon(l)dl}\|u_\varepsilon(r)\|^2_{\tilde{H}^{0,1}}\right]^p\right)^\frac{2}{p}ds.
\end{align*}
Applying Gronwall's inequality, we obtain
\begin{align*}
&\left(E\left[\sup_{0\leqslant t\leqslant \tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^t_0f_\varepsilon(s)ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^{\tau_{M_1,\varepsilon}}_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\right]^p\right)^\frac{2}{p}\\
\leqslant &C\left[\|u_0\|^4_{\tilde{H}^{0,1}}+\varepsilon^2+\varepsilon p\right]e^{C(\varepsilon^2+\varepsilon p)}.
\end{align*}
Hence by the definition of $\tau_{M_1,\varepsilon}$, we have
\begin{align*}
&\left(E\left[G_{u_\varepsilon}(\tau_{M_1,\varepsilon})\right]^p\right)^\frac{2}{p}\\
\leqslant& \left(E\left[\left(\sup_{0\leqslant t\leqslant \tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^t_0f_\varepsilon(s)ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^{\tau_{M_1,\varepsilon}}_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\right)^pe^{pk\varepsilon\int^t_0f_\varepsilon(s)ds}\right]\right)^\frac{2}{p}\\
\leqslant&e^{C(M_1+\varepsilon)}\left(E\left[\sup_{0\leqslant t\leqslant \tau_{M_1,\varepsilon}}e^{-k\varepsilon\int^t_0f_\varepsilon(s)ds}\|u_\varepsilon(t)\|^2_{\tilde{H}^{0,1}}+\varepsilon\int^{\tau_{M_1,\varepsilon}}_0e^{-k\varepsilon\int^s_0f_\varepsilon(r)dr}\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds\right]^p\right)^\frac{2}{p}\\
\leqslant &Ce^{C(M_1+\varepsilon)}\left[\|u_0\|^4_{\tilde{H}^{0,1}}+\varepsilon^2+\varepsilon p\right]e^{C(\varepsilon^2+\varepsilon p)}.
\end{align*}
Let $p=\frac{2}{\varepsilon}$, by Chebyshev's inequality, we have
\begin{align*}
&\varepsilon\log P(G_{u_\varepsilon}(\tau_{M_1,\varepsilon})>M)\\
\leqslant &\varepsilon\log\frac{E\left[G_{u_\varepsilon}(\tau_{M_1,\varepsilon})\right]^p}{M^p}\\
\leqslant&-2\log M+C+C(M_1+\varepsilon)+C(\varepsilon^2+\varepsilon p)+\log[\|u_0\|^4_{\tilde{H}^{0,1}}+\varepsilon^2+\varepsilon p].
\end{align*}
Take supremum over $\varepsilon$ and let $M\rightarrow \infty$, we finish the proof.
\end{proof}
Since $V$ is dense in ${\tilde{H}^{0,1}}$, there exists a sequence $\{u^n_0\}\subset V$ such that
$$\lim_{n\rightarrow +\infty}\|u^n_0-u_0\|_{\tilde{H}^{0,1}}=0.$$
Let $u_{n,\varepsilon}$ be the solution to (\ref{eq. for small time}) with the initial data $u^n_{0}$. Similarly, let $v_{n,\varepsilon}$ be the solution to (\ref{eq for v}) with the initial data $u^n_{0}$.
For $M>0$, define a random time (which is also a stopping time with respect to $\mathcal{F}_{t+}$ by Lemma \ref{stopping time})
$$\tau^n_{M,\varepsilon}:=T\wedge\inf\{t:\|u_{n,\varepsilon}(t)\|^2_{H}>M,\text{ or }\varepsilon\int^t_0\|\partial_1 u_{n,\varepsilon}(s)\|^2_{H}ds>M\}.$$
From the proof of Lemma \ref{lemma1} and Lemma \ref{lemma2}, it follows that
\begin{lemma}\label{lemma3}
$$\lim_{M\rightarrow \infty}\sup_n\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P(F_{u_{n,\varepsilon}}(T)>M)=-\infty.$$
For fixed $M_1$, we have
$$\lim_{M\rightarrow \infty}\sup_n\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P(G_{u_{n,\varepsilon}}(\tau_{M_1,\varepsilon}^n)>M)=-\infty.$$
\end{lemma}
\vskip.10in
The following lemma for $v_{n,\varepsilon}$ is from \cite{XZ08}:
\begin{lemma}[{\cite[Lemma 3.2]{XZ08}}]\label{lemma4}
$$\lim_{M\rightarrow\infty}\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P\left(\sup_{0\leqslant t\leqslant T}\|v_{n,\varepsilon}(t)\|^2_{V}>M\right)=-\infty.$$
\end{lemma}
\vskip.10in
\iffalse
\begin{proof}
Applying It\^o's formula to $\|v_{n,\varepsilon}\|^2_V$, we obtain
\begin{align*}
\|v_{n,\varepsilon}(t)\|^2_V=\|u_{n,0}\|^2_V+2\sqrt{\varepsilon}\int^t_0\langle v_{n,\varepsilon}, \sigma(\varepsilon s, v_{n,\varepsilon}(s))dW(s)\rangle_{V}+\varepsilon\int^t_0\|\sigma(\varepsilon s, v_{n,\varepsilon}(s))\|^2_{L_2(l^2,V)}ds.
\end{align*}
Then by (A4) and Lemma \ref{martingale lemma}, we have
\begin{align*}
\left(E[\sup_{0\leqslant t\leqslant r}\|v_{n,\varepsilon}(t)\|^{2p}_V]\right)^\frac{2}{p}\leqslant &2\|u_{n,0}\|^4_V+8C\varepsilon p\left(E\left[\int^r_0\|v_{n,\varepsilon}(s)\|^2_V\|\sigma(\varepsilon s, v_{n,\varepsilon}(s))\|^2_{L_2(l^2, V)}ds\right]^\frac{p}{2}\right)^\frac{2}{p}\\
&+4\varepsilon^2Cr\left(r+\int^r_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{n,\varepsilon}(l)\|_V^{2p}]\right)^\frac{2}{p} ds\right)\\
\leqslant&2\|u_{n,0}\|^4_V+C\varepsilon p\left(r+\int^r_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{n,\varepsilon}(l)\|_V^{2p}]\right)^\frac{2}{p} ds\right)\\
&+C\varepsilon^2r\left(r+\int^r_0\left(E[\sup_{0\leqslant l\leqslant s}\|v_{n,\varepsilon}(l)\|_V^{2p}]\right)^\frac{2}{p} ds\right).
\end{align*}
By Gronwall's inequality, we have
$$\left(E[\sup_{0\leqslant t\leqslant T}\|v_{n,\varepsilon}(t)\|^{2p}_V]\right)^{\frac{2}{p}}\leqslant C(\|u_{n,0}\|_V^4+\varepsilon p+\varepsilon^2)e^{C(\varepsilon p+\varepsilon^2)}.$$
The rest of proof is the same as that of Lemma \ref{lemma1} which follows from Chebyshev's inequality.
\end{proof}
\fi
\begin{lemma}\label{lemma5}
For any $\delta>0$,
$$\lim_{n\rightarrow\infty}\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)=-\infty.$$
\end{lemma}
\begin{proof}
Clearly, for $M_1, M_2>0$
\begin{equation}\label{eq in lemma5}\aligned
&P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)\\
\leqslant &P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta, F_{u_\varepsilon}(T)\leqslant M_1, G_{u_\varepsilon}(T)\leqslant M_2\right)\\
&+P\left(F_{u_\varepsilon}(T)>M_1\right)+P\left(F_{u_\varepsilon}(T)\leqslant M_1, G_{u_\varepsilon}(T)>M_2\right)\\
\leqslant &P\left(\sup_{0\leqslant t\leqslant\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)\\
&+P\left(F_{u_\varepsilon}(T)>M_1\right)+P\left(G_{u_\varepsilon}(\tau_{M_1,\varepsilon})>M_2\right),
\endaligned
\end{equation}
where $\tau_{M_1,\varepsilon}$ and $\tau'_{M_2,\varepsilon}$ are introduced in Lemma \ref{stopping time}.
For the first term on the right hand of (\ref{eq in lemma5}), let $k$ be a positive constant and
\begin{align*}
U_\varepsilon= 1+\|u_\varepsilon\|^2_{\tilde{H}^{1,1}}.
\end{align*}
Applying It\^o's formula to $e^{-\varepsilon k\int^{t}_0U_\varepsilon(s) ds}\|u_\varepsilon(t)-u_{n,\varepsilon}(t)\|_H^2$, we get
\begin{align*}
&e^{-\varepsilon k\int^{t}_0U_\varepsilon(s) ds}\|u_\varepsilon(t)-u_{n,\varepsilon}(t)\|_H^2+2\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|\partial_1(u_\varepsilon(s)-u_{n,\varepsilon}(s))\|^2_Hds\\
=&\|u_0-u_{n,0}\|^2_H-k\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}U_\varepsilon(s)\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_Hds\\
&-2\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\left(b(u_\varepsilon,u_\varepsilon,u_\varepsilon-u_{n,\varepsilon})(s)-b(u_{n,\varepsilon},u_{n,\varepsilon},u_\varepsilon-u_{n,\varepsilon})(s)\right)ds\\
&+\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|\sigma(\varepsilon s, u_\varepsilon(s))-\sigma(\varepsilon s, u_{n,\varepsilon}(s))\|^2_{L_2(l^2, H)}ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\langle u_\varepsilon(s)-u_{n,\varepsilon}(s), (\sigma(\varepsilon s, u_\varepsilon(s))-\sigma(\varepsilon s, u_{n,\varepsilon}(s)))dW(s)\rangle.
\end{align*}
Notice that by the property of the trilinear form $b$ and Lemma \ref{anisotropic estimate for b}, we have
\begin{align*}
&|b(u_\varepsilon,u_\varepsilon,u_\varepsilon-u_{n,\varepsilon})-b(u_{n,\varepsilon},u_{n,\varepsilon},u_\varepsilon-u_{n,\varepsilon})|\\
=&|b(u_\varepsilon,u_\varepsilon,u_\varepsilon-u_{n,\varepsilon})-b(u_{n,\varepsilon},u_\varepsilon,u_\varepsilon-u_{n,\varepsilon})|\\
=&|b(u_\varepsilon-u_{n,\varepsilon},u_\varepsilon,u_\varepsilon-u_{n,\varepsilon})|\\
\leqslant& \frac{1}{2}\|\partial_1(u_\varepsilon-u_{n,\varepsilon})\|^2_H+C_1U_\varepsilon\|u_\varepsilon-u_{n,\varepsilon}\|^2_H,
\end{align*}
where $C_1$ is a constant.
Therefore,
\begin{align*}
&e^{-\varepsilon k\int^{t}_0U_\varepsilon(s) ds}\|u_\varepsilon(t)-u_{n,\varepsilon}(t)\|_H^2\\
\leqslant& \|u_0-u_{n,0}\|^2_{\tilde{H}^{0,1}}-k\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}U_\varepsilon(s)\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_Hds\\
&+2\varepsilon C_1 \int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}U_\varepsilon(s)\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_Hds\\
&+L\varepsilon\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\langle u_\varepsilon(s)-u_{n,\varepsilon}(s), (\sigma(\varepsilon s, u_\varepsilon(s))-\sigma(\varepsilon s, u_{n,\varepsilon}(s)))dW(s)\rangle,
\end{align*}
where we used (A3') in the forth line.
Choosing $k>2C_1$ and using Lemma \ref{martingale lemma} and (A3'), by the similar calculation as in the proof of Lemma \ref{lemma2} we have for $p\geqslant 2$
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant t\wedge\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}\right]^p\right)^\frac{2}{p}\\
\leqslant& 2\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}+C\varepsilon^2\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s\wedge\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^r_0U_\varepsilon(l) dl}\|u_\varepsilon(r)-u_{n,\varepsilon}(r)\|^2_{H}\right]^p\right)^\frac{2}{p}ds\\
&+C\varepsilon p\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s\wedge\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^r_0U_\varepsilon(l) dl}\|u_\varepsilon(r)-u_{n,\varepsilon}(r)\|^2_{H}\right]^p\right)^\frac{2}{p}ds.\\
\end{align*}
Applying Gronwall's inequality, we obtain
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant t\wedge\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}\right]^p\right)^\frac{2}{p}\leqslant C\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}e^{C(\varepsilon^2+\varepsilon p)}.
\end{align*}
\iffalse
Note that by Young's inequality, we have
\begin{align*}
\varepsilon\int^{\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}_0U_\varepsilon(s)ds&\leqslant C\varepsilon\int^{\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}_0(1+\|\partial_1u_\varepsilon(s)\|^2_H+\|\partial_2u_\varepsilon(s)\|^2_H+\|\partial_1\partial_2u_\varepsilon(s)\|^2_H)ds.\\
&\leqslant C\varepsilon(1+\int^{\tau'_{M_2, \varepsilon}}_0\|u_\varepsilon(s)\|^2_{\tilde{H}^{1,1}}ds)\leqslant C(\varepsilon+M_2).
\end{align*}
\fi
Hence, by the definition of the stopping times,
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant \tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}\right]^p\right)^\frac{2}{p}\\
\leqslant&\left(E\left[\big{(}\sup_{0\leqslant s\leqslant \tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}\big{)}^pe^{kp\varepsilon\int^{\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}_0U_\varepsilon(s)ds}\right]\right)^\frac{2}{p}\\
\leqslant&e^{C(\varepsilon+M_2)k}\left(E\left[\sup_{0\leqslant s\leqslant \tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}e^{-\varepsilon k\int^s_0U_\varepsilon(r) dr}\|u_\varepsilon(s)-u_{n,\varepsilon}(s)\|^2_{H}\right]^p\right)^\frac{2}{p}\\
\leqslant&Ce^{C(\varepsilon+M_2)k}\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}e^{C(\varepsilon^2+\varepsilon p)}.
\end{align*}
Fix $M_1, M_2$, let $p=\frac{2}{\varepsilon}$, then Chebyshev's inequality implies that
\begin{align*}
&\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P\left(\sup_{0\leqslant t\leqslant \tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)\\
\leqslant&\sup_{0<\varepsilon\leqslant 1}\varepsilon\log\frac{E\left[\sup_{0\leqslant t\leqslant \tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^{2p}\right]}{\delta^p}\\
\leqslant& C(\varepsilon+M_2)-2\log\delta+\log\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}+C(\varepsilon^2+\varepsilon p)+C\\
\rightarrow&-\infty, \text{ as }n\rightarrow \infty.
\end{align*}
By Lemma \ref{lemma1}, for any $R>0$, there exists a constant $M_1$ such that for any $\varepsilon\in(0,1]$,
$$P\left(F_{u_\varepsilon}(T)>M_1\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
For such a $M_1$, by Lemma \ref{lemma2}, there exists a constant $M_2$ such that for any $\varepsilon\in(0,1]$,
$$P\left(G_{u_\varepsilon}(\tau_{M_1,\varepsilon})>M_2\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
For such $M_1,M_2$, there exists a positive integer $N$, such that for any $n\geqslant N$ and $\varepsilon\in(0,1]$,
$$P\left(\sup_{0\leqslant t\leqslant\tau_{M_1,\varepsilon}\wedge\tau'_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
Then by (\ref{eq in lemma5}), we see that there exists a positive integer $N$, such that for any $n\geqslant N$, $\varepsilon\in(0,1]$,
$$P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-u_\varepsilon(t)\|_H^2>\delta\right)\leqslant 3e^{-\frac{R}{\varepsilon}}.$$
Since $R$ is arbitrary, the lemma follows.
\end{proof}
The following lemma for $v_\varepsilon$ is from \cite{XZ08}:
\begin{lemma}[{\cite[Lemma 3.4]{XZ08}}]\label{lemma6}
For any $\delta>0$,
$$\lim_{n\rightarrow\infty}\sup_{0<\varepsilon\leqslant 1}\varepsilon\log P\left(\sup_{0\leqslant t\leqslant T}\|v_{n,\varepsilon}(t)-v_\varepsilon(t)\|_H^2>\delta\right)=-\infty.$$
\end{lemma}
\vskip.10in
\iffalse
\begin{proof}
The proof is similar but simple to that of Lemma \ref{lemma5}. Applying It\^o's formula to $\|v_\varepsilon(t)-v_{n,\varepsilon}(t)\|^2_H$, taking Lemma \ref{martingale lemma} and (A4) into consideration, we have
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant t}\|v_\varepsilon(s)-v_{n,\varepsilon}(s)\|_H^{2p}\right]\right)^{\frac{2}{p}}\\
\leqslant& 2\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}+C(\varepsilon p+\varepsilon^2)\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s}\|v_\varepsilon(r)-v_{n,\varepsilon}(r)\|_H^{2p}\right]\right)^{\frac{2}{p}}ds.
\end{align*}
Applying Gronwall's inequality,
$$\left(E\left[\sup_{0\leqslant s\leqslant t}\|v_\varepsilon(s)-v_{n,\varepsilon}(s)\|_H^{2p}\right]\right)^{\frac{2}{p}}\leqslant C\|u_0-u_{n,0}\|^4_{\tilde{H}^{0,1}}e^{C(\varepsilon p+\varepsilon^2)}.$$
Then the lemma follows from Chebyshev's inequality.
\end{proof}
\fi
\begin{lemma}\label{lemma7}
For any $\delta>0$, and every positive integer $n$,
$$\lim_{\varepsilon\rightarrow 0}\varepsilon\log P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)=-\infty.$$
\end{lemma}
\begin{proof}
For $M>0$, recall the definition of $\tau^n_{M, \varepsilon}$ and define the following random time:
$$\tau^{2,n}_{M,\varepsilon}:= T\wedge\inf\{t: \|u_{n,\varepsilon}(t)\|^2_{\tilde{H}^{0,1}}>M,\text{ or }\varepsilon\int^t_0\|u_{n,\varepsilon}(s)\|^2_{\tilde{H}^{1,1}}ds>M\},$$
which is a stopping time with respect to $\mathcal{F}_{t+}$ by Lemma \ref{stopping time}.
Moreover, define
$$\tau^{3,n}_{M,\varepsilon}:= T\wedge\inf\{t: \|v_{n,\varepsilon}(t)\|^2_{V}>M\},$$
$$\tau^{1,n}_{M,\varepsilon}:=\tau^n_{M,\varepsilon}\wedge\tau^{3,n}_{M,\varepsilon}.$$
We should point out that $\tau^{3,n}_{M,\varepsilon}$ is a stopping time with respect to $\mathcal{F}_t$ under the condition $v_{n,\varepsilon}\in C([0,T], V)$. Now we prove that $v_{n,\varepsilon}\in C([0,T], V)$.
\iffalse
Let $\{e_k\}_{k\in\mathbb{Z}, k\geqslant 1}\subset H$ be a smooth orthogonal basis of $V$. For any $f\in H$, we can define $\langle f, e_k\rangle_V:=\langle f, (I-\Delta) e_k\rangle_H$. Set $\varphi_k=\langle v_{n,\varepsilon}, e_k\rangle_V$, by It\^o's formula, for $t\in[0, T]$, we have
\begin{align*}
\varphi^2_k(t)=\varphi_k^2(0)+\int^t_0 2\sqrt{\varepsilon}\varphi_k(s)\langle \sigma(\varepsilon s, v_{n,\varepsilon}(s))dW(s),e_k\rangle_V+\int^t_0\varepsilon\|\sigma(\varepsilon s, v_{n,\varepsilon}(s))^*e_k\|_{l^2}^2ds,
\end{align*}
where $\sigma^*$ denote the adjoint operator of $\sigma$. Summing over $k$, by (A3') and Burkh\"older-Davis-Gundy's inequality, we deduce that
\begin{align*}
E(\sup_{s\in[0,t]}\|v_{n,\varepsilon}(s)\|_V^2)\leqslant C(\sqrt{\varepsilon}+\varepsilon)\int^t_0(1+E(\sup_{r\in[0, s]}\|v_{n,\varepsilon}(r)\|_V^2))ds.
\end{align*}
\fi
By It\^o's formula and Gronwall's inequality there exists a constant $C(\varepsilon)$ such that
\begin{align*}
E(\sup_{s\in[0,t]}\|v_{n,\varepsilon}(s)\|_V^2)\leqslant C(\varepsilon).
\end{align*}
For $0\leqslant s<t\leqslant T$, by (A4) we have
\begin{align*}
E\|v_{n,\varepsilon}(t)-v_{n,\varepsilon}(s)\|_V^2\leqslant& \varepsilon E\int^t_s\|\sigma(\varepsilon r, v_{n,\varepsilon}(r))\|^2_{L_2(l^2,V)}dr\\
\leqslant &\varepsilon \int^t_s(\overline{K}_0+\overline{K}_1E(\sup_{l\in[0,r]}\|v_{n,\varepsilon}(l)\|_V^2))dr\\
\leqslant& \varepsilon(\overline{K}_0+\overline{K}_1C(\varepsilon))|t-s|.
\end{align*}
Then Kolmogorov's continuity criterion implies that $v_{n,\varepsilon}\in C([0,T], V)$.
Now for $M_1, M_2>0$, similarly to (\ref{eq in lemma5}), we have
\begin{equation}\label{eq in lemma7}\aligned
&P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)\\\leqslant&P\left(\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)\\
&+P(F_{u_{n,\varepsilon}}(T)>M_1)+P(G_{u_{n,\varepsilon}}(\tau^n_{M_1,\varepsilon})>M_2)+P\left(\sup_{0\leqslant t\leqslant T}\|v_{n,\varepsilon}(t)\|^2_{V}>M_1\right)
\endaligned
\end{equation}
Let $U_{n,\varepsilon}=1+\|u_{n,\varepsilon}\|^2_{\tilde{H}^{1,1}}$,
applying It\^o's formula to $e^{-k\varepsilon\int^t_0U_{n,\varepsilon}(s)ds}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H$ for some constant $k>0$, we get
\begin{equation}\label{eq2 in lemma7}\aligned
&e^{-k\varepsilon\int^t_0U_{n,\varepsilon}(s)ds}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H+2\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|\partial_1(u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s))\|^2_Hds\\
=&-k\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}U_{n,\varepsilon}(s)\|u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s)\|^2_Hds\\
&+2\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\langle u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s), \partial_1^2 v_{n,\varepsilon}(s)\rangle ds\\
&-2\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}b(u_{n,\varepsilon}(s),u_{n,\varepsilon}(s),u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s))ds\\
&+\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|\sigma(\varepsilon s,u_{n,\varepsilon}(s))-\sigma(\varepsilon s, v_{n,\varepsilon}(s))\|^2_{L_2(l^2,H)}ds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\langle u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s), (\sigma(\varepsilon s,u_{n,\varepsilon}(s))-\sigma(\varepsilon s, v_{n,\varepsilon}(s)))dW(s) \rangle.
\endaligned
\end{equation}
For the second term on the right hand side of (\ref{eq2 in lemma7}), we have
\begin{align*}
&\Big{|}\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\langle u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s), \partial_1^2 v_{n,\varepsilon}(s)\rangle ds\Big{|}\\
\leqslant& \int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|\partial_1( u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s))\|_H\| \partial_1 v_{n,\varepsilon}(s)\|_H ds\\
\leqslant& \frac{1}{4}\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\| \partial_1(u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s))\|^2_Hds+C\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\| v_{n,\varepsilon}(s)\|^2_{V} ds,
\end{align*}
where we use Young's inequality in the last inequality.
For the third term on the right hand side of \eqref{eq2 in lemma7}, by Lemmas \ref{anisotropic estimate for b} and \ref{b(u,v,w)} we have
\begin{equation}\aligned
&|b(u_{n,\varepsilon},u_{n,\varepsilon},u_{n,\varepsilon}-v_{n,\varepsilon})|\\
=&|b(u_{n,\varepsilon}-v_{n,\varepsilon},u_{n,\varepsilon},u_{n,\varepsilon}-v_{n,\varepsilon})+b(v_{n,\varepsilon},u_{n,\varepsilon},u_{n,\varepsilon}-v_{n,\varepsilon}|\\
\leqslant&\frac{1}{4}\|\partial_1(u_{n,\varepsilon}-v_{n,\varepsilon})\|^2_H+CU_{n,\varepsilon}\|u_{n,\varepsilon}-v_{n,\varepsilon}\|^2_H+ C\|v_{n,\varepsilon}\|_{V}\|u_{n,\varepsilon}\|_{\tilde{H}^{1,1}}\|u_{n,\varepsilon}-v_{n.\varepsilon}\|_H\\
\leqslant&\frac{1}{4}\|\partial_1(u_{n,\varepsilon}-v_{n,\varepsilon})\|^2_H+C\|v_{n,\varepsilon}\|^2_{V}+C_1 U_{n,\varepsilon}\|u_{n,\varepsilon}-v_{n,\varepsilon}\|^2_H,
\endaligned
\end{equation}
where $C_1$ is a constant.
Thus we obtain
\begin{align*}
&e^{-k\varepsilon\int^t_0U_{n,\varepsilon}(s)ds}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H+\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|\partial_1(u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s))\|^2_Hds\\
\leqslant&-k\varepsilon\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}U_{n,\varepsilon}(s)\|u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s)\|^2_Hds+C\varepsilon \int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\| v_{n,\varepsilon}(s)\|^2_{V} ds\\
&+C_1\varepsilon \int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}U_{n,\varepsilon}(s)\|u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s)\|^2_Hds\\
&+L_1\varepsilon \int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s)\|^2_Hds\\
&+2\sqrt{\varepsilon}\int^t_0e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\langle u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s), (\sigma(\varepsilon s,u_{n,\varepsilon}(s))-\sigma(\varepsilon s, v_{n,\varepsilon}(s)))dW(s) \rangle,
\end{align*}
where we used (A3') in the fourth line.
Hence, choosing $k>C_1+C_2$, by Lemma \ref{martingale lemma} and the similar techniques in the previous lemma and the definition of stopping times, we deduce that for $p\geqslant 2$
\begin{align*}
&\left(E\left[\sup_{0\leqslant s\leqslant t\wedge\tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}e^{-k\varepsilon\int^s_0U_{n,\varepsilon}(r)dr}\|u_{n,\varepsilon}(s)-v_{n,\varepsilon}(s)\|^2_H\right]^p\right)^{\frac{2}{p}}\\
\leqslant& CM_1^2\varepsilon^2+C(\varepsilon^2+\varepsilon p)\int^t_0\left(E\left[\sup_{0\leqslant r\leqslant s\wedge\tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}e^{-k\varepsilon\int^r_0U_{n,\varepsilon}(l)dl}\|u_{n,\varepsilon}(r)-v_{n,\varepsilon}(r)\|^2_H\right]^p\right)^{\frac{2}{p}}ds.
\end{align*}
Then Gronwall's inequality implies that
\begin{equation}\label{eq3 in lemma7}\aligned
&\left(E\left[\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H\right]^p\right)^{\frac{2}{p}}\\
\leqslant&\left(E\left[\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}(e^{-k\varepsilon\int^t_0U_{n,\varepsilon}(s)ds}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H)^pe^{kp\varepsilon\int^{\tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}_0U_{n,\varepsilon}(s)ds}\right]\right)^{\frac{2}{p}}\\
\leqslant& e^{C(\varepsilon+M_2)}CM_1^2\varepsilon^2 e^{C(\varepsilon^2+\varepsilon p)}.
\endaligned
\end{equation}
By Lemmas \ref{lemma3} and \ref{lemma4}, we know that for any $R>0$, there exists $M_1$ such that
$$\sup_{0<\varepsilon\leqslant 1} \varepsilon \log P\left(F_{u_{n,\varepsilon}}(T)>M_1\right)\leqslant -R,$$
$$\sup_{0<\varepsilon\leqslant 1} \varepsilon \log P\left(\sup_{0\leqslant t\leqslant T}\|v_{n,\varepsilon}(t)\|^2_{V}>M_1\right)\leqslant -R.$$
For such a constant $M_1$, by Lemma \ref{lemma3}, there exists $M_2$ such that
$$\sup_{0<\varepsilon\leqslant 1} \varepsilon \log P\left(G_{u_{n,\varepsilon}}(\tau^n_{M_1,\varepsilon})>M_2\right)\leqslant -R.$$
Then for such $M_1,M_2$, let $p=\frac{2}{\varepsilon}$ in (\ref{eq3 in lemma7}), we obtain
\begin{align*}
&\varepsilon \log P\left(\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)\\
\leqslant&\log \left(E\left[\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H\right]^p\right)^{\frac{2}{p}}-\log\delta^2\\
\leqslant& C(\varepsilon+M_2)+\log[CM^2_1\varepsilon^2]+C(\varepsilon^2+1)-\log\delta^2\\
\rightarrow& -\infty \text{ }\text{ as }\varepsilon\rightarrow 0,
\end{align*}
where we used Chebyshev's inequality in the first inequality. Thus there exists a $\varepsilon_0\in(0,1)$ such that for any $\varepsilon\in(0,\varepsilon_0)$,
$$P\left(\sup_{0\leqslant t\leqslant \tau^{1,n}_{M_1,\varepsilon}\wedge\tau^{2,n}_{M_2,\varepsilon}}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
Putting the above estimate together, by (\ref{eq in lemma7}) we see that for $\varepsilon\in(0,\varepsilon_0)$
$$P\left(\sup_{0\leqslant t\leqslant T}\|u_{n,\varepsilon}(t)-v_{n,\varepsilon}(t)\|^2_H>\delta\right)\leqslant 4e^{-\frac{R}{\varepsilon}}.$$
Since $R$ is arbitrary, we finish the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main}]
By Lemma \ref{LDP for v}, $v_\varepsilon$ satisfies a large deviation principle with the rate function $I^{u_0}$. Our task remain is to show that $u_\varepsilon$ and $v_\varepsilon$ are exponentially equivalent, then the result follows from Lemma \ref{EXEQ}.
By Lemmas \ref{lemma5} and \ref{lemma6}, for any $R>0$, there exists a $N_0$ such that for any $\varepsilon\in(0,1]$,
$$P\left(\sup_{0\leqslant t\leqslant T}\|u_{\varepsilon}(t)-u_{N_0,\varepsilon}(t)\|_H^2>\frac{\delta}{3}\right)\leqslant e^{-\frac{R}{\varepsilon}},$$
and
$$P\left(\sup_{0\leqslant t\leqslant T}\|v_{\varepsilon}(t)-v_{N_0,\varepsilon}(t)\|_H^2>\frac{\delta}{3}\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
Then by Lemma \ref{lemma7}, for such $N_0$, there exists a $\varepsilon_0$ such that for any $\varepsilon\in(0,\varepsilon_0)$,
$$P\left(\sup_{0\leqslant t\leqslant T}\|u_{N_0, \varepsilon}(t)-v_{N_0,\varepsilon}(t)\|_H^2>\frac{\delta}{3}\right)\leqslant e^{-\frac{R}{\varepsilon}}.$$
Therefore we deduce that for $\varepsilon\in(0,\varepsilon_0)$
$$P\left(\sup_{0\leqslant t\leqslant T}\|u_{\varepsilon}(t)-v_{\varepsilon}(t)\|_H^2>\delta\right)\leqslant 3e^{-\frac{R}{\varepsilon}}.$$
Since $R$ is arbitrary, we finish the proof.
\end{proof}
| 314d85a42b6632194e9b09db5304761eea8c39e4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the last two decades, random matrix theory (RMT) has produced numerous results that offer a better understanding of large random matrices. These advances have enabled interesting applications in communication theory and even though it can potentially contribute to many other data-rich domains such as brain imaging or genetic research, it has rarely been applied.
The main barrier to the adoption of RMT may be the lack of concrete statistical results from the probability side. The straightforward adaptation of classical multivariate theory to high dimensions can sometimes be achieved, but such procedures are only valid under strict assumptions about the data such as normality or independence. Even minor differences between the model assumptions and the actual data lead to catastrophic results and such procedures also often do not have enough power.
This paper proposes a statistical procedure for testing the equality of two covariance matrices when the number of potentially dependent data vectors $n$ and the number of variables $m$ are large. RMT denotes the investigation of estimates of covariance matrices $\hat{\Sigma}$ or more precisely their eigenvalues and eigenvectors when both $n$ and $m$ tend to infinity with $\lim \frac{m}{n}=c>0$.
When $m$ is finite and $n$ tends to infinity the behaviour of the random matrix is well known and presented in the books of \cite{multi3}, \cite{multi} and \cite{multi2} (or its original version \cite{multi22}).
In the RMT case, the behaviour is more complex, but many results of interest are known. \cite{Alice}, \cite{Tao} and more recently \cite{bookrecent} contain comprehensive introductions to RMT and \cite{Appliedbook} covers the case of empirical (estimated) covariance matrices.
Although the existing theory builds a good intuition of the behaviour of these matrices, it does not provide enough of a basis to construct a test with good power, which is robust with respect to the assumptions.
Inspired by the spike models, we extend the residual spikes introduced in \cite{mainarticle} and provide a description of the behaviour of this statistic under a null hypothesis when the perturbation is of order $k$. These results enable the user to test the equality of two populations as well as other null hypotheses such as the independence of two sets of variables.
This paper can be seen as a complex particular case of \cite{mainarticle2}. However simulations show that equality between eigenvalues of the perturbation are not necessary for this complex statistic and moreover they show good robustness against perturbations of distributions.\\
The remainder of the paper is organized as follows. In the next section, we develop the test statistic and discuss the problems associated with high dimensions. Then we present the main theorem \ref{TH=Main}. The proof itself is technical and presented in Appendix \ref{appendixproof}. The last section contains an example of an application.
\section{Test statistic}\label{section:teststat}
We compare the spectral properties of two covariance estimators $\hat{\Sigma}_X$ and $\hat{\Sigma}_Y$ of dimension $m\times m$ which can be represented as
\begin{eqnarray*}
\hat{\Sigma}_X=P_X^{1/2} W_X P_X^{1/2} \text{ and } \hat{\Sigma}_Y=P_Y^{1/2} W_Y P_Y^{1/2}.
\end{eqnarray*}
In this equation, $W_X$ and $W_Y$ are of the form
\begin{eqnarray*}
W_X=O_X \Lambda_X O_X \text{ and } W_Y=O_Y \Lambda_Y O_Y,
\end{eqnarray*}
with $O_X$ and $O_Y$ being independent unit orthonormal random matrices whose distributions are invariant under rotations, while $\Lambda_X$ and $\Lambda_Y$ are independent positive random diagonal matrices, independent of $O_X, O_Y$ with trace equal to m and a bound on the diagonal elements. Note that the usual RMT assumption, $\frac{m}{n}=c$ is replaced by this bound! The (multiplicative) spike model of order $k$ determines the form of
$P_X= {\rm I}_m + \sum_{s=1}^k (\theta_{X,s}-1) u_{X,s} u_{X,s}^t$ and $P_Y= {\rm I}_m+ \sum_{s=1}^k(\theta_{Y,s}-1) u_{Y,s} u_{Y,s}^t$ where $\left\langle u_{X,s} ,u_{X,r} \right\rangle=\left\langle u_{Y,s} ,u_{Y,r} \right\rangle=\delta_{s,r}$.
Our results will apply to any two centered data matrices ${\mathbf{X}} \in \mathbb{R}^{m\times n_X}$ and ${\mathbf{Y}} \in \mathbb{R}^{m\times n_Y}$ which are such that
\begin{eqnarray*}
\hat{\Sigma}_X= \frac{1}{n_X}{\mathbf{X}} {\mathbf{X}}^t \text{ and } \hat{\Sigma}_Y= \frac{1}{n_Y}{\mathbf{Y}} {\mathbf{Y}}^t
\end{eqnarray*}
and can be decomposed in the manner indicated. This is the basic assumption concerning the covariance matrices. We will assume throughout that $n_X\geq n_Y$. Because $O_X$ and $O_Y$ are independent and invariant by rotation we can assume without loss of generality that for $s=1,2,...,k$, $u_{X,s}=e_s$ as in \cite{deformedRMT}. Under the null hypothesis we have $P_X=P_Y$ and we use the simplified notation $P_k$ for both matrices where for $s=1,2,...,k$, $\theta_{X,s}=\theta_{Y,s}=\theta_s$ and $u_{X,s}=u_{Y,s}(=e_s)$.
To test $H_0:P_k=P_X=P_Y$ against $H_1:P_X\neq P_Y$ it is natural to consider the extreme eigenvalues of
\begin{eqnarray}
\hat{\Sigma}_X^{-1/2} \hat{\Sigma}_Y \hat{\Sigma}_X^{-1/2}\,.\label{eq:base1}
\end{eqnarray}
We could also swap the subscripts, but it turns our to be preferable to use the inversion on the matrix with larger sample size.
The distributional approximations we will refer to are based on RMT, that is, they are derived by embedding a given data problem into a sequence of random matrices for which both $n$ and $m$ tend to infinity such that $m/n$ tends to a positive constant $c$. The most celebrated results of RMT describe the almost sure weak convergence of the empirical distribution of the eigenvalues (spectral distribution) to a non-random compactly supported limit law. An extension of this theory to the "Spike Model" suggests that we should modify $\hat{\Sigma}$ because estimates of isolated eigenvalues derived from the usual estimates are asymptotically biased. The following corrections will be used.
\begin{Def} \label{Def=unbiased}
Suppose $\hat{\Sigma}$ is of the form described at the start of the section.
The \textbf{unbiased estimator of }$\theta_s$ for $s=1,...,k$ is defined as
\begin{equation} \hat{\hat{\theta}}_s=1+\frac{1}{\frac{1}{m-k} \sum_{i=k+1}^{m} \frac{\hat{\lambda}_{\hat{\Sigma},i}}{\hat{\theta}_s-\hat{\lambda}_{\hat{\Sigma},i}}}\,,\label{eq:corrlambda}
\end{equation}
where $\hat{\lambda}_{\hat{\Sigma},i}$ is the $i^{\text{th}}$ largest eigenvalue of $\hat{\Sigma}$. When $\hat{\Sigma}=P_k^{1/2}WP_k^{1/2}$ as above, it is asymptotically equivalent to replace $\frac{1}{m-k}\sum_{i=k+1}^{m} \frac{\hat{\lambda}_{\hat{\Sigma},i}}{\hat{\theta}-\hat{\lambda}_{\hat{\Sigma},i}}$ in the denominator by $\frac{1}{m}\sum_{i=1}^{m} \frac{\hat{\lambda}_{W,i}}{\hat{\theta}-\hat{\lambda}_W,i}$.\\
Suppose that $\hat{u}_s$ is the eigenvector corresponding to $\hat{\theta}_s$, then the \textbf{filtered estimated covariance matrix } is defined as
\begin{equation}
\hat{\hat{\Sigma}}= {\rm I}_m+\sum_{s=1}^k (\hat{\hat{\theta}}_s-1) \hat{u}_s \hat{u}_s^t\,.\label{eq:corrSigma}
\end{equation}
\end{Def}
The matrix (\ref{eq:base1}) which serves as the basis for the test then becomes either
\begin{equation}
\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2}
\text{ or }\hat{\hat{\Sigma}}_X^{-1/2} \hat{\Sigma}_Y \hat{\hat{\Sigma}}_X^{-1/2} \,.\label{eq:base2}
\end{equation}
In the particular case where ${\mathbf{X}}$ and ${\mathbf{Y}}$ have independent jointly normal columns vector with constant variance $P_k=P_X=P_Y$, the distribution of the spectrum of the second of the above matrices is approximately Marcenko-Pastur distributed (see \cite{deformed}). This follows because $\hat{\hat{\Sigma}}_X$ is a finite perturbation. However, because of the non-consistency of the eigenvectors presented in \cite{deformedRMT}, we may observe \textbf{residual spikes} in the spectra, as shown in Figure \ref{fig=spike}. Thus, even if the two random matrices are based on the same perturbation, we see some spikes outside the bulk. This observation is worse in the last plot because four spikes fall outside the bulk even if there is actually no difference!
This poses a fundamental problem for our test, because we must be able to distinguish the spikes indicative of a true difference from the residual spikes. These remarks lead to the following definition.
\begin{Def}
The \textbf{residual spikes} are the isolated eigenvalues of
$$\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2}\text{ or of }
\hat{\hat{\Sigma}}_X^{-1/2} \hat{\Sigma}_Y \hat{\hat{\Sigma}}_X^{-1/2}$$
when $P_X=P_Y$ (under the null hypothesis).
The \textbf{residual zone} is the interval where a residual spike can fall asymptotically.
\end{Def}
\begin{figure}[htbp]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.26\paperwidth]{residualtheta10.pdf} &
\includegraphics[width=0.26\paperwidth]{residualtheta10c=1.pdf} \\
\includegraphics[width=0.26\paperwidth]{residualtheta10c=02.pdf}&
\includegraphics[width=0.26\paperwidth]{residualthetak.pdf}
\end{tabular}
\caption{Example of residual spikes of $\hat{\hat{\Sigma}}_X^{-1/2} \hat{\Sigma}_Y \hat{\hat{\Sigma}}_X^{-1/2}$ when $\theta=10$ for the first three figures and $\theta_{1,2,3,4}=10,15,20,25$ for the last figure.}\label{fig=spike}
\end{figure}
This paper studies these residual spikes by deriving the distribution of the extreme residual spikes under the null hypothesis. The philosophy is explained in Figure \ref{fig=residualzone0} with illustrations inspired by the i.i.d. normal case. All the eigenvalues inside the residual zone are potentially not indicative of real differences. However, when an eigenvalue is larger, we declare that this spike expresses a true difference.
Most of our plots feature the seemingly more natural matrix
\begin{eqnarray*}
\hat{\hat{\Sigma}}_X^{-1/2} \hat{\Sigma}_Y \hat{\hat{\Sigma}}_X^{-1/2}\,.
\end{eqnarray*}
But, although this choice simplifies the study in terms of convergence in probability when the perturbation is of order $1$, this is no longer the case in more complex situations. In addition, the eigenvectors associated with the residual spikes are more accessible for the matrix in which all estimates are filtered.
\begin{figure}[htbp]
\centering
\scalebox{0.7}{
\begin{tabular}{c}
\input{residualzone} \\ \ \input{residualzone2}
\end{tabular}
}
\caption{Residual zone of $\hat{\hat{\Sigma}}_X^{-1/2} \hat{\Sigma}_Y \hat{\hat{\Sigma}}_X^{-1/2}$ and $\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2}$.}\label{fig=residualzone0}
\end{figure}
Let $\hat\theta_{X,s}$ and $\hat\theta_{Y,s}$ be isolated eigenvalues and construct the asymptotic unbiased estimators as in Equation (\ref{eq:corrlambda})
\begin{eqnarray*}
\hat{\hat{\theta}}_{X,s}=1+\frac{1}{\frac{1}{m-k} \sum_{i=k+1}^{m} \frac{\hat{\lambda}_{\hat{\Sigma}_{X},i}}{\hat{\theta}_{X,s}-\hat{\lambda}_{\hat{\Sigma}_{X},i}}} \text{ and } \hat{\hat{\theta}}_{Y,s}=1+\frac{1}{\frac{1}{m-k} \sum_{i=k+}^m \frac{\hat{\lambda}_{\hat{\Sigma}_{Y},i}}{\hat{\theta}_{Y,s}-\hat{\lambda}_{\hat{\Sigma}_{Y},i}}},
\end{eqnarray*}
where $\hat{\lambda}_{\hat{\Sigma}_{X},i}$ and $\hat{\lambda}_{\hat{\Sigma}_{Y},i}$ are the i$^{th}$ ordered eigenvalue of $\hat{\Sigma}_{X}$ and $\hat{\Sigma}_{Y}$, respectively. The test statistic is then
\begin{eqnarray*}
\lambda_{\min}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) \text{ and } \lambda_{\max}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right)\,,
\end{eqnarray*}
where the filtered matrices are constructed as in (\ref{eq:corrSigma}).
These two statistics provide a basis for a powerful and robust test for the equality of (detectable) perturbations $P_X$ and $P_Y$.
\subsection{Null distribution} \label{sec=test}
Under $H_0$, $\lambda_{\max}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) $ is obviously a function of $\theta_s=\theta_{X,s}=\theta_{Y,s}$ for $s=1,2,...,k$. The suspected \textbf{worst case} occurs in the limit as $\theta_s\to\infty$ for all $s$ and it is this limit which will determine the critical values of the test. A criterion proposed in \cite{mainarticle} allows to check if this scenario is really the worst case. Let
\begin{eqnarray*}
&&\lambda_{\max}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) \leqslant \lim_{\theta \rightarrow \infty} \lambda_{\max}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) =V_{\max} ,\\
&&\lambda_{\min}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) \geqslant \lim_{\theta \rightarrow \infty} \lambda_{\min}\left(\hat{\hat{\Sigma}}_X^{-1/2} \hat{\hat{\Sigma}}_Y \hat{\hat{\Sigma}}_X^{-1/2} \right) =V_{\min}.
\end{eqnarray*}
Because of our focus on the worst case scenario under $H_0$, we will investigate the asymptotic as $\theta_i=\theta p_i$ for fixed $p_i>0$ and $\frac{\theta}{\sqrt{m}} \rightarrow \infty$. We recall that \cite{mainarticle2} also allows some finite $\theta_s$ but it seems intuitive that this scenario will not create a worst case in most situations. This intuition is highlighted by \cite{mainarticle} showing by simulation that the residual spike increase as a function of $\theta$ assuming $W_X=\frac{1}{n_X}{\mathbf{X}}\X^t$.
Our test rejects the null hypothesis of equal populations if
either $\P\left( V_{\max} > \hat{\lambda}_{\max} \right)$ or $\P\left( V_{\min} < \hat{\lambda}_{\min} \right)$ is small, where $\hat{\lambda}_{\max}$ and $\hat{\lambda}_{\min}$ are the observed extreme residual spikes.
The following result describes the asymptotic behavior of the extreme eigenvalues and thus of $V_{\max}$ and $V_{\min}$.
\begin{Th}
\label{TH=Main}
Suppose $W_X,W_Y \in \mathbb{R}^{m\times m}$ are as described at the start of Section 2 and
\begin{enumerate}
\item \cite{mainarticle} have already investigated the case
$P_1= {\rm I}_m+(\theta-1)e_1 e_1^t \in \mathbb{R}^{m\times m}$ with $\frac{\sqrt{m}}{\theta}=o(1)$ with regard to large $m$. Let
\begin{eqnarray*}
\hat{\Sigma}_{X,P_1}=P_1^{1/2} W_X P_1^{1/2} \text{ and } \hat{\Sigma}_{Y,P_1}=P_1^{1/2} W_Y P_1^{1/2}.
\end{eqnarray*}
and $\hat{\hat{\Sigma}}_{X,P_1}$, $\hat{\hat{\Sigma}}_{Y,P_1}$ as described above (see, \ref{Def=unbiased}).
Then, conditional on the spectra $S_{W_X}=\left\lbrace \hat{\lambda}_{W_X,1},\hat{\lambda}_{W_X,2},...,\hat{\lambda}_{W_X,m} \right\rbrace$ and $S_{W_Y}=\left\lbrace \hat{\lambda}_{W_Y,1},\hat{\lambda}_{W_Y,2},...,\hat{\lambda}_{W_Y,m} \right\rbrace$ of $W_X$ and $W_Y$,
\begin{eqnarray*}
\left. \sqrt{m} \frac{\left(\lambda_{\max}\left(\hat{\hat{\Sigma}}_{X,P_1}^{-1/2} \hat{\hat{\Sigma}}_{Y,P_1} \hat{\hat{\Sigma}}_{X,P_1}^{-1/2} \right) -\lambda^+ \right)}{\sigma^+} \right| S_{W_X}, S_{W_Y} \sim \mathbf{N}(0,1)+o_{p}(1),
\end{eqnarray*}
where
\begin{eqnarray*}
&&\lambda^+= \sqrt{M_2^2-1}+M_2, \hspace{30cm}
\end{eqnarray*}
\scalebox{0.77}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
&&{\sigma^+}^2=\frac{1}{\left(M_{2,X}+M_{2,Y}-2\right) \left(M_{2,X}+M_{2,Y}+2\right)} \hspace{30cm} \\
&&\hspace{2cm} \Bigg( 9 M_{2,X}^4 M_{2,Y}+4 M_{2,X}^3 M_{2,Y}^2+4 M_{2,X}^3 M_{2,Y}+2 M_{2,X}^3 M_{3,Y}-2 M_{2,X}^2 M_{2,Y}^3\\
&& \hspace{2cm}+4 M_{2,X}^2 M_{2,Y}^2-11 M_{2,X}^2 M_{2,Y}-8 M_{3,X} M_{2,X}^2 M_{2,Y}+2 M_{2,X}^2 M_{2,Y} M_{3,Y}\\
&& \hspace{2cm}-2 M_{2,X}^2 M_{3,Y}+M_{2,X}^2 M_{4,Y}+4 M_{2,X} M_{2,Y}^3+M_{2,X} M_{2,Y}^2+4 M_{2,X} M_{2,Y}\\
&& \hspace{2cm}-4 M_{3,X} M_{2,X} M_{2,Y}^2-4 M_{3,X} M_{2,X} M_{2,Y}-2 M_{2,X} M_{2,Y}^2 M_{3,Y}-4 M_{2,X} M_{2,Y} M_{3,Y}\\
&& \hspace{2cm}-6 M_{2,X} M_{3,Y}+2 M_{4,X} M_{2,X} M_{2,Y}+2 M_{2,X} M_{2,Y} M_{4,Y}-2 M_{3,X} M_{2,Y}^2\\
&& \hspace{2cm}+2 M_{3,X} M_{2,Y}+M_{4,X} M_{2,Y}^2+4 M_{2,X}^5+2 M_{2,X}^4-4 M_{3,X} M_{2,X}^3-13 M_{2,X}^3\\
&& \hspace{2cm}-2 M_{3,X} M_{2,X}^2+M_{4,X} M_{2,X}^2-2 M_{2,X}^2+10 M_{3,X} M_{2,X}+4 M_{2,X}+4 M_{3,X}\\
&& \hspace{2cm}-2 M_{4,X}+M_{2,Y}^5+2 M_{2,Y}^4-M_{2,Y}^3-2 M_{2,Y}^2+4 M_{2,Y}-2 M_{2,Y}^3 M_{3,Y}\\
&& \hspace{2cm}-2 M_{2,Y}^2 M_{3,Y}+2 M_{2,Y} M_{3,Y}+4 M_{3,Y}+M_{2,Y}^2 M_{4,Y}-2 M_{4,Y}-4 \Bigg)\\
&&\hspace{1.25cm} + \frac{1}{\sqrt{\left(M_{2,X}+M_{2,Y}-2\right) \left(M_{2,X}+M_{2,Y}+2\right)}}\\
&& \hspace{2cm} \Bigg( 5 M_{2,X}^3 M_{2,Y}-M_{2,X}^2 M_{2,Y}^2+2 M_{2,X}^2 M_{2,Y}+2 M_{2,X}^2 M_{3,Y}-M_{2,X} M_{2,Y}^3\\
&& \hspace{2cm}+2 M_{2,X} M_{2,Y}^2-4 M_{2,X} M_{2,Y}-4 M_{3,X} M_{2,X} M_{2,Y}-2 M_{2,X} M_{3,Y}+M_{2,X} M_{4,Y}\\
&& \hspace{2cm}-2 M_{3,X} M_{2,Y}+M_{4,X} M_{2,Y}+4 M_{2,X}^4+2 M_{2,X}^3-4 M_{3,X} M_{2,X}^2-5 M_{2,X}^2\\
&& \hspace{2cm}-2 M_{3,X} M_{2,X}+M_{4,X} M_{2,X}+2 M_{2,X}+2 M_{3,X}+M_{2,Y}^4+2 M_{2,Y}^3+M_{2,Y}^2\\
&& \hspace{2cm}+2 M_{2,Y}-2 M_{2,Y}^2 M_{3,Y}-2 M_{2,Y} M_{3,Y}-2 M_{3,Y}+M_{2,Y} M_{4,Y} \Bigg),
\end{eqnarray*}
\end{minipage}}
\begin{eqnarray*}
&&M_{s,X}= \frac{1}{m} \sum_{i=1}^m \hat{\lambda}_{W_X,i}^s,\hspace{30cm}\\
&&M_{s,Y}= \frac{1}{m} \sum_{i=1}^m \hat{\lambda}_{W_Y,i}^s,\\
&&M_s=\frac{M_{s,X}+M_{s,Y}}{2}.
\end{eqnarray*}
Moreover,
\begin{eqnarray*}
\left.\sqrt{m} \frac{\left(\lambda_{\min}\left(\hat{\hat{\Sigma}}_{X,P_1}^{-1/2} \hat{\hat{\Sigma}}_{Y,P_1} \hat{\hat{\Sigma}}_{X,P_1}^{-1/2} \right) -\lambda^- \right)}{\sigma^-}\right| S_{W_X}, S_{W_Y} \sim \mathbf{N}(0,1)+o_{m}(1),
\end{eqnarray*}
where
\begin{eqnarray*}
&&\lambda^-= -\sqrt{M_2^2-1}+M_2, \hspace{30cm}\\
&&{\sigma^-}^2=\left(\lambda^-\right)^4 {\sigma^+}^2.
\end{eqnarray*}
The error $o_p(1)$ in the approximation is with regard to large values of $m$.
\item Suppose that $P_k= {\rm I}_m+\sum_{s=1}^k(\theta_s-1)e_s e_s^t \in \mathbb{R}^{m\times m}$ with $\theta_s=p_s \theta$, $p_s>0$ and $\frac{\sqrt{m}}{\theta}=o(1)$ with regard to large $m$.
\begin{eqnarray*}
\hat{\Sigma}_{X,P_k}=P_k^{1/2} W_X P_k^{1/2} \text{ and } \hat{\Sigma}_{Y,P_k}=P_k^{1/2} W_Y P_k^{1/2},
\end{eqnarray*}
and $\hat{\hat{\Sigma}}_{X,P_k}$, $\hat{\hat{\Sigma}}_{Y,P_k}$ as described above (see, \ref{Def=unbiased}).
\noindent Then, conditioning on the spectra $S_{W_X}$ and $S_{W_Y}$,\\
\scalebox{0.82}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
\left. \lambda_{\max}\left(\hat{\hat{\Sigma}}_{X,P_k}^{-1/2} \hat{\hat{\Sigma}}_{Y,P_k} \hat{\hat{\Sigma}}_{X,P_k}^{-1/2}\right) \right| S_{W_X},S_{W_Y}=\lambda_{\max}\left( H^+ \right)+1+O_p\left(\frac{1}{m}\right),\\
\left.\lambda_{\min}\left(\hat{\hat{\Sigma}}_{X,P_k}^{-1/2} \hat{\hat{\Sigma}}_{Y,P_k} \hat{\hat{\Sigma}}_{X,P_k}^{-1/2}\right)\right| S_{W_X},S_{W_Y}=\lambda_{\max}\left( H^- \right)+1+O_p\left(\frac{1}{m}\right),
\end{eqnarray*}
\end{minipage}}
where
\begin{eqnarray*}
H^\pm= \zeta_{\infty}^\pm \begin{pmatrix}
\hat{\zeta}_1^\pm / \zeta_{\infty}^\pm & w_{1,2}^\pm & w_{1,3}^\pm & \cdots & w_{1,k}^\pm \\
w_{2,1}^\pm & \hat{\zeta}_2^\pm / \zeta_{\infty}^\pm & w_{2,3}^\pm & \cdots & w_{2,k}^\pm \\
w_{3,1}^\pm & w_{3,2}^\pm & \hat{\zeta}_3^\pm /\zeta_{\infty}^\pm & \cdots & w_{3,k}^\pm \\
\vdots & \vdots & \ddots & \ddots & \vdots \\
w_{k,1}^\pm & w_{k,2}^\pm & w_{k,3}^\pm & \cdots & \hat{\zeta}_k^\pm /\zeta_{\infty}^\pm \\
\end{pmatrix} ,
\end{eqnarray*}
and \\
\scalebox{0.73}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
&&\hat{\zeta}_i^+= \left.\lambda_{\max}\left(\hat{\hat{\Sigma}}_{X,\tilde{P}_i}^{1/2} \hat{\hat{\Sigma}}_{Y,\tilde{P}_i} \hat{\hat{\Sigma}}_{X,\tilde{P}_i}^{1/2} \right)-1 \right| S_{W_X},S_{W_Y},\hspace{30cm}\\
&&\hat{\zeta}_i^-= \left.\lambda_{\min}\left(\hat{\hat{\Sigma}}_{X,\tilde{P}_i}^{1/2} \hat{\hat{\Sigma}}_{Y,\tilde{P}_i} \hat{\hat{\Sigma}}_{X,\tilde{P}_i}^{1/2} \right)-1 \right| S_{W_X},S_{W_Y},\\
&&\zeta_{\infty}^\pm =\underset{m \rightarrow \infty}{\lim} \hat{\zeta}_i^\pm = \lambda^\pm -1,\\
&&w_{i,j}^\pm \sim { \color{red} \mathbf{N}}\left(0,
\frac{1}{m}\frac{2 (M_{2,X}-1) (M_{2,Y}-1)+B_{X}^\pm +B_{Y}^\pm}{\left((\zeta_{\infty}^\pm-2 M_2+1)^2+2 (M_2-1)\right)^2}
\right)+o_p\left( \frac{1}{\sqrt{m}} \right),
\end{eqnarray*}
\end{minipage}}\\
\scalebox{0.73}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
&&B_{X}^+=\left(1-M_2+2 M_{2,X}+\sqrt{M_2^2-1}\right)^2 (M_{2,X}-1)\hspace{20cm}\\
&&\hspace{2cm} +2\left(-1+M_2-2M_{2,x}-\sqrt{M_2^2-1}\right)(M_{3,X}-M_{2,X})+(M_{4,X}-M_{2,X}^2),\\
&&B_{Y}^+=\left(1+M_2+M_{2,Y}-M_{2,X}-\sqrt{M_2^2-1}\right)^2 (M_{2,Y}-1)\\
&&\hspace{2cm} +2\left(-1-M_2-M_{2,Y}-M_{2,X}-\sqrt{M_2^2-1}\right)(M_{3,Y}-M_{2,Y})+(M_{4,Y}-M_{2,Y}^2),\\
&&B_{X}^-=\left(1-M_2+2 M_{2,X}-\sqrt{M_2^2-1}\right)^2 (M_{2,X}-1)\hspace{20cm}\\
&&\hspace{2cm} +2\left(-1+M_2-2M_{2,x}+\sqrt{M_2^2-1}\right)(M_{3,X}-M_{2,X})+(M_{4,X}-M_{2,X}^2),\\
&&B_{Y}^-=\left(1+M_2+M_{2,Y}-M_{2,X}+\sqrt{M_2^2-1}\right)^2 (M_{2,Y}-1)\\
&&\hspace{2cm} +2\left(-1-M_2-M_{2,Y}+M_{2,X}-\sqrt{M_2^2-1}\right)(M_{3,Y}-M_{2,Y})+(M_{4,Y}-M_{2,Y}^2).
\end{eqnarray*}
\end{minipage}}\\
\noindent The matrices $H^+$ and $H^-$ are strongly correlated. However, within a matrix, all the entries are uncorrelated.
\end{enumerate}
\end{Th}
\begin{Rem} \label{Rem=thMain} \
The entries of the matrices $H^+$ and $H^-$ are asymptotically uncorrelated Normal or a sum of two Normals.
\end{Rem}
\paragraph*{Special case}
If the spectra are Marcenko-Pastur distributed, we define $c_X=m/n_X$ and $c_Y=m/n_Y$. Then,
\scalebox{0.79}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
&&c=\frac{c_X+c_Y}{2},\hspace{30cm}\\
&&\lambda^+= c+\sqrt{c (c+2)}+1,\\
&&{\sigma^+}^2= c_X^3+c_X^2 c_Y+3 c_X^2+4 c_X c_Y-c_X+c_Y^2+c_Y\\
&& \hspace{2cm}+\frac{ (8 c_X+2 c_X^2+\left(c_X^3+5 c_X^2+c_X^2 c_Y+4 c_X c_Y+5 c_X +3 c_Y+c_Y^2\right)\sqrt{c(c+2)}}{c+2},\\
&&w_{i,j}^+ \sim {\color{red} \mathbf{N}}\left(0, \frac{\sigma_w^2}{m} \right),\\
&& \sigma_w^2=\frac{2c_X \left(\sqrt{c(c+2)}+2\right)+2 \c_Y \left(-\sqrt{c(c+2)}+2\right)+c_X^2+c_Y^2}
{4c \left(-\sqrt{c \left(c+2\right)}+c+2\right)^2}.
\end{eqnarray*}
\end{minipage}}\\
(Proof Appendix \ref{appendixproof}. The {\color{red} red} character is not proven and is a sum of two asymptotic uncorrelated marginally normal random variables that are certainly independent.)
\subsection{Discussion and simulation}\label{sec:smallsimulationmainth}
The above theorem gives the limiting distribution of $V_{\max}$ and $V_{\min}$. In this subsection, we first check the quality of the approximations in Theorem \ref{TH=Main}. Then we investigate the worst case with regard to $\theta$. Finally we relax some assumption on $\theta_s$ and on the distribution.
\subsubsection{Some simulations}
Assume $\mathbf{X} \in \mathbb{R}^{m\times n_X}$ and $\mathbf{Y} \in \mathbb{R}^{m\times n_Y}$ with $\mathbf{X}=\left(X_1,X_2,...,X_{n_X}\right)$ and $\mathbf{Y}=\left(Y_1,Y_2,...,Y_{n_Y}\right)$.
\scalebox{0.65}{
\begin{minipage}{1\textwidth}
\begin{eqnarray*}
&&X_i \sim \mathbf{N}_m\left(\vec{0},\sigma^2 {\rm I}_m \right) \text{ with } X_1 = \epsilon_{X,1}\text{ and }
X_{i+1}=\rho X_i+\sqrt{1-\rho^2} \ \epsilon_{X,i+1}, \text{ where } \epsilon_{X,i} \overset{i.i.d}{\sim} \mathbf{N}_m\left(\vec{0},\sigma^2 {\rm I}_m\right),\\
&&Y_i \sim \mathbf{N}_m\left(\vec{0},\sigma^2 {\rm I}_m \right) \text{ with } Y_1 = \epsilon_{Y,1}\text{ and } Y_{i+1}=\rho Y_i+\sqrt{1-\rho^2} \ \epsilon_{Y,i+1}, \text{ where } \epsilon_{Y,i} \overset{i.i.d}{\sim} \mathbf{N}_m\left(\vec{0},\sigma^2 {\rm I}_m\right)
\end{eqnarray*}
\end{minipage}}
Let $P_X= {\rm I}_m + (\theta_{X}-1) u_{X} u_{X}^t$ and $P_Y= {\rm I}_m+ (\theta_{Y}-1) u_{Y} u_{Y}^t$ be two perturbations in $\mathbb{R}^{m\times m}$. Then,
\begin{eqnarray*}
\mathbf{X}_P=P_X^{1/2} \mathbf{X} \text{ and } \mathbf{Y}_P=P_Y^{1/2} \mathbf{Y},\\
\hat{\Sigma}_X=\frac{\mathbf{X}_P^t \mathbf{X}_P}{n_X} \text{ and } \hat{\Sigma}_Y=\frac{\mathbf{Y}_P^t \mathbf{Y}_P}{n_Y}.
\end{eqnarray*}
We assume a common and large value for $\theta$ and $P_X=P_Y$.
\begin{table}
\begin{tabular}{ c c }
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image1thetasame.pdf}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image2thetasame.pdf}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 1 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=4, $\vec{\theta}=\left(15'000,5000,2000,500\right)$.
\end{tabular}}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 2 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=4, $\vec{\theta}=\left(5'000,5000,5000,5000\right)$.
\end{tabular}}
\end{minipage} \\
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image51differentk.pdf}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image52differentk.pdf}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 3 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=1, $\vec{\theta}=5'000$.
\end{tabular}}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 4 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=8, $\vec{\theta}=\left(5'000,5'000,...,5'000\right)$.
\end{tabular}}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image53differentk.pdf}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image1image31kest.pdf}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 5 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=15, $\vec{\theta}=\left(5'000,5'000,...,5'000\right)$.
\end{tabular}}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 6 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=4, $\vec{\theta}=\left(15'000,5'000,2'000,500\right)$\\
$k_{est}=7$.
\end{tabular}}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image1image32kest.pdf}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\includegraphics[width=1\textwidth]{image1image33kest.pdf}
\end{minipage}\\
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 7 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=4, $\vec{\theta}=\left(15'000,5'000,2'000,500\right)$\\
$k_{est}=3$.
\end{tabular}}
\end{minipage}
&
\begin{minipage}{.45\textwidth} \centering
\scalebox{0.8}{
\begin{tabular}{c}
Scenario 8 \\
\begin{tabular}{c|c|c}
$\rho=0.5$ & $c_X=0.5$ & $c_Y=2$ \\
\hline
$m=1000$ & $n_X=2000$ & $n_Y=500$ \\
\end{tabular} \\
k=4, $\vec{\theta}=\left(15'000,3'000,8,6\right)$\\
$k_{est}=2$.
\end{tabular}}
\end{minipage}
\end{tabular}
\caption{Empirical distributions of the residual spikes together with the Gaussian densities from the theorem \ref{TH=Main} (in blue).
} \label{tab:residual}
\end{table}
\paragraph*{Multiple eigenvalues} \ \\
Despite the lack of a proof, the maximum residual distribution when the eigenvalues of the perturbations are multiple is well approximated by our Theorem. This can be seen in Table \ref{tab:residual} but also in Appendix \ref{appendix:Tablecomplete}.
\paragraph*{Different values of $k$} \ \\
Scenario 3, 4 and 5 of Table \ref{tab:residual} shows that the result holds for different values of $k$. The less accurate result of scenario 5 is due to the relatively large $k$, whereas the theorem is based on an approximation which considers $k$ to be small compared to $m$. The precision of the asymptotic approximation would be better for $k=15$ when $m=10'000$ with the same $c_X$ and $c_Y$, for example.
\paragraph*{Wrong estimation of $k$} \ \\
Scenario 6, 7 and 8 of Table \ref{tab:residual} shows the impact of using wrong values of $k$. We see in Scenario 6 that a small overestimation of $k$ leads to a small overestimation of the maximum and small underestimation the minimum. This will lead to conservative tests. Scenario 7 shows that underestimation of $k$ can lead to a bad approximation but in this scenario we neglect perturbations of size $500$! Scenario 8 shows that neglecting two small perturbations of size $6$ and $8$, as we could easily do by mistake, still leads to very accurate approximations.\\
The simulation of Table \ref{tab:residual} are done to convince the reader of the usefulness of Theorem \ref{TH=Main}. In practice, we must estimate the parameters needed in the approximation. An arguments based on the Cauchy-interlacing theorem can convinced the reader that we can estimate
\begin{eqnarray*}
M_{X,s}=\frac{1}{m}\sum_{i=1}^m \lambda_{W_X}^s
\end{eqnarray*}
by
\begin{eqnarray*}
\hat{M}_{X,s}=\frac{1}{m-k}\sum_{i=k+1}^m \lambda_{\hat{\Sigma}_X}^s.
\end{eqnarray*}
The impact of using a wrong value for $k_{est}$ is investigate in Table \ref{tab=simulationkestimation} in Appendix \ref{appendix:Simulationkest}.
\paragraph*{Other simulations}
In appendix \ref{appendix:Simulationkest} we also investigate the approximation with estimated spectra for data with distributions that are not invariant by rotation. In some scenarios the approximation succeeds to estimate the location but failed to correctly estimate the variance. In others, the location of the maximum residual spike is overestimated and the minimum residual spike is underestimated. This would again lead to conservative tests, but suggests a lack of power.
\section{An application}
In this section, we apply our procedure developed from Theorem \ref{TH=Main} to data ${\color{red}{{\mathbf{X}}}}$ and ${\color{red}{{\mathbf{Y}}}}$. First, each step is briefly explained. Then, an analysis is presented on simulated data together with the mathematical work and the important plots.\\
This procedure is not unique and other solutions better adapted to the problem could be implemented. For example, the choice of $k$ and the number of perturbations, could certainly be improved. The goal of this section is to provide a procedure as conservative as possible with reasonably good asymptotic power.
\begin{enumerate}
\item First, we center the data with regard to the rows and columns.
\item Then, we need to estimate $k$ and rescale the variance. These two tasks are interconnected. One intuitive way to choose $k$ for each matrix ($k_X$ and $k_Y$) consists in looking at the spectra for spikes and keeping in mind that overestimation is preferable to underestimation of the actual value.\\
Using $k=\max(k_X,k_Y)$, we can then rescale the matrices ${\color{red}{{\mathbf{X}}}}$ and ${\color{red}{{\mathbf{Y}}}}$ to create ${\color{blue}{{\mathbf{X}}}}$ and ${\color{blue}{{\mathbf{Y}}}}$.
\item Next, we apply the procedure to ${\color{blue}{{\mathbf{X}}}}$ and ${\color{blue}{{\mathbf{Y}}}}$, using the above $k$. In our case, this leads to two observed extreme residual spikes.
\item We compute the distribution of the residual spike by assuming $k$ perturbation and estimating $M_{s,X}$.
\item Finally, we can compare the extreme values with their distribution under $\text{H}_0$ for testing purposes.
\end{enumerate}
\begin{Rem} \
Our simulations in Appendix \ref{appendix:Simulationkest}, show that the choice of $k$ does not affect the conservative nature of the test. A strong underestimation of $k$, however, greatly reduces the power. This explains the advice to overestimate $k$.
\end{Rem}
\subsection{Analysis} \label{sec:simulAnalysis}
We observe data ${\color{red}{{\mathbf{X}}}}\in \mathbf{R}^{m \times n_X}$ and ${\color{red}{{\mathbf{Y}}}}\in \mathbf{R}^{m \times n_Y}$ that we suppose is already centred by rows and columns. We choose $k$ by looking at the histogram of the matrices in Figure \ref{Figexample1} where $m=1000$, $n_X=2000$ et $n_Y=500$.
\begin{figure}[hbtp]
\includegraphics[width=0.5\paperwidth]{ExampleSpectrum.pdf}
\caption{Spectra of ${\color{red}{{\mathbf{X}}}}$ and ${\color{red}{{\mathbf{Y}}}}$ with largest isolated eigenvalues indicated by arrows.}\label{Figexample1}
\end{figure}
We try to overestimate a lower bound on $k$ based on Figure \ref{Figexample1}. The spectrum of ${\color{red}{{\mathbf{X}}}}$ seems to have $6$ isolated eigenvalues, but we could argue that two other eigenvalues are perturbations. The spectrum ${\color{red}{{\mathbf{Y}}}}$ clearly shows $5$ isolated eigenvalues and at most $2$ additional ones. We thus set $k=8$ knowing that we probably overestimate the true value. Next, we estimate the variances,
\begin{eqnarray*}
\hat{\sigma}_X^2= \frac{1}{m-k}\sum_{i=k+1}^m \lambda_i\left( \frac{1}{n_X}{\color{red}{{\mathbf{X}}}}\rX^t \right),\\
\hat{\sigma}_Y^2= \frac{1}{m-k} \sum_{i=k+1}^m \lambda_i\left( \frac{1}{n_Y}{\color{red}{{\mathbf{Y}}}}\rY^t \right).
\end{eqnarray*}
We can then rescale the matrices ${\color{red}{{\mathbf{X}}}}$ and ${\color{red}{{\mathbf{Y}}}}$ by $\hat{\sigma}_X$ and $\hat{\sigma}_Y$, respectively, to create the covariance matrices
\begin{eqnarray*}
\hat{\Sigma}_X=\frac{1}{n_X \hat{\sigma}_X^2} {\color{red}{{\mathbf{X}}}} {\color{red}{{\mathbf{X}}}}^t \text{ and }
\hat{\Sigma}_Y=\frac{1}{n_Y \hat{\sigma}_Y^2} {\color{red}{{\mathbf{Y}}}} {\color{red}{{\mathbf{Y}}}}^t.
\end{eqnarray*}
Finally, we filter the matrices as in definition \ref{Def=unbiased}.
\begin{eqnarray*}
&&\hat{\hat{\Sigma}}_X = {\rm I}_m + \sum_{i=1}^k \left(\hat{\hat{\theta}}_{X,i}-1 \right) \hat{u}_{\hat{\Sigma}_X,i}\hat{u}_{\hat{\Sigma}_X,i}^t,\\
&&\hat{\hat{\theta}}_{X,i}= 1+\frac{1}{\frac{1}{m-k} \sum_{j=k+1}^m \frac{\hat{\lambda}_{ \hat{\Sigma}_X,j}}{\hat{\lambda}_{ \hat{\Sigma}_X,i}-\hat{\lambda}_{ \hat{\Sigma}_X,j}}}.
\end{eqnarray*}
The computed residual spikes of $\hat{\hat{\Sigma}}_X^{-1} \hat{\hat{\Sigma}}_Y$ are shown in Table \ref{Tabexample1}.
\begin{table}[hbtp]
\begin{tabular}{l|cccccccc}
$\lambda_{\max}$ & 56.03 & 10.25 & 9.88 & 8.96 & 8.29 & 7.27 & 5.71 & 5.10 \\
$\lambda_{\min}$ & 0.04 & 0.10 & 0.13 & 0.13 & 0.18 & 0.21 & 0.34 & 0.36
\end{tabular}
\caption{Observed residual spikes.}\label{Tabexample1}
\end{table}
Using Figure \ref{Figexample2} a, these values are compared to the theoretical distributions of the extreme residual spikes assuming equality of the perturbations of order $k$. The distribution in blue uses the usual estimator of the spectra and the distribution in orange uses the conservative estimator introduced in Appendix \ref{appendix:Spectrumestimation}.
The moments of the spectra are summarize in Figure \ref{Figexample2} b.
\begin{figure}[hbtp]
\begin{center}
\begin{tabular}{cc}
\begin{minipage}{0.3\paperwidth}
\includegraphics[width=0.3\paperwidth]{ExampleSpectrum2.pdf}
\end{minipage}
&
\begin{minipage}{0.3\paperwidth}
\scalebox{0.7}{
\begin{tabular}{l|cc}
&$\lambda_{\min}$ & $\lambda_{\max}$\\
\color{blue} Usual & (0.085, 0.006) & (10.75, 0.45)\\
\color{orange} Robust & (0.067, 0.006) & (12.67, 0.47)
\end{tabular}}
\end{minipage}\\
a & b
\end{tabular}
\end{center}
\caption{a: Distribution of the extreme residual spike assuming equality of the covariance, $k=8$ and $\theta_i$ large. (Robust estimation of the spectra in orange.) b: Estimated residual spikes moments, $(\mu,\sigma)$ using usual or robust estimators of the spectral moments. }\label{Figexample2}
\end{figure}
We finally clearly detect two residual spikes. Figure \ref{Figexample3} presents the residual eigenvectors of the residual eigenvalues.
\begin{figure}[hbtp]
\begin{center}
\begin{tabular}{cc}
$\lambda=56.03$ & $\lambda=0.04 $\\
\includegraphics[width=0.25\paperwidth]{residualvectormax1.pdf} & \includegraphics[width=0.25\paperwidth]{residualvectormin1.pdf} \\
$\lambda=10.25$ & $\lambda=0.10$ \\
\includegraphics[width=0.25\paperwidth]{residualvectormax2.pdf} & \includegraphics[width=0.25\paperwidth]{residualvectormin2.pdf}
\end{tabular}
\end{center}
\caption{Representation of the entiere residual eigenvectors and only the 20 first entries. }\label{Figexample3}
\end{figure}
We conclude that the differences are in direction $e_3$ and $e_4$. As we see in the figure, two other eigenvectors also exhibit a structure. Without our test, we could have concluded that they also represent significant differences, but this residual structure is merely due to the biased estimation of the eigenvectors.
\begin{center}
\textbf{A structure in a residual eigenvector does not imply a real difference!}
\end{center}
\subsection{Conclusion}
By studying perturbation of order $1$ in \cite{mainarticle} and perturbation of order $k$ in \cite{mainarticle2}, we highlighted the lack of power of the usual procedure to detect differences between two groups. This paper extended the residual spike to perturbations of order $k$ by using tools introduced in previous papers. While this test has weaker power than in \cite{mainarticle2}, it has the important advantage do be able to deal with multiple equal eigenvalues. Additional simulations investigating the robustness of this new procedure are contained in the thesis \cite{mathese} and seems promising.
\pagebreak
| a7d7ca7b235ae26158a575c400645c689b2bc7ff | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\emph{A motivating question and our model.}
In the idealized experiment shown in Figure \ref{fig:idealized experiment}, a pulse of inert gas
at low pressure is pumped into a long but finite tube, which we refer to as the {\em channel.}
The inner surface of the channel has some degree of roughness due to its molecular
structure and surface irregularities.
The experimenter is able to measure the rate of gas outflow using some device such as a mass spectrometer,
which generates data of the kind represented by the graph on the right-hand side of the figure.
From such data, transport characteristics of the gas flow through the channel can be derived, as described in \cite{CFZ2016}.
We assume a sufficiently small pulse, under vacuum conditions, to insure that molecular mean free path is much larger than the diameter of the channel.
Thus collisions between the gas molecules can be ignored while gas-surface interaction is expected to influence transport properties most prominently.
The property of interest here, which can be indirectly measured from such an experiment, is the Knudsen self-diffusivity coefficient of the gas, as explained, for example, in \cite{CFZ2016}.
The central question we wish to address is: How do the surface characteristics affect the Knudsen self-diffusivity?
In this paper, we assume that gas-surface interaction amounts to perfectly elastic, or billiard-like,
collisions between point masses (the {\em gas molecules}, also referred to here as {\em particles}) and the channel
surface, and hence energy exchange between surface and molecules will be ignored.
We assume moreover that the channel is two-dimensional and that its surface microstructure is static and periodic,
and can be described by a relatively small number of geometric parameters.
Thus the mathematical problem we pose here is to determine how the Knudsen self-diffusivity explicitly depends on these parameters.
In the large Knudsen number limit (i.e., for large mean free paths),
molecular trajectories are independent of each other
and the diffusion process is derived from an analysis of individual trajectories of particles undergoing a random flight inside the channel.
This random flight is governed by a Markov operator $P$ that gives,
at each particle-surface collision, the post-collision velocity of the particle as a random function of the pre-collision velocity.
All the information about the periodic surface geometry relevant to the task of obtaining diffusivity is encoded in $P$.
In fact, diffusivity corresponds to the variance of a one-dimensional Wiener process
obtained from the random flight determined by $P$ via a Central Limit Theorem.
(As explained in \cite{CFZ2016},
this variance can be obtained from the mean exit time in the limit of long channel lengths in the context of the above idealized experiment.
The mean exit time, as a function of the channel length, is the only information that needs to be extracted from the exit flow rate data. We won't deal here with this particular aspect of the analysis and assume, in effect, that the channel is infinite in length.)
Our main goals are thus centered around two issues.
First, we aim to establish a functional, analytic relationship among the aforementioned variance, the spectrum of the Markov operator $P$, and parameters of the geometric microstructure.
Second, we aim to obtain effective numerical methods for finding this dependence for any given geometric microstructure.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\columnwidth]{Exit_flow.eps}
\caption{ Idealized experiment for measuring diffusivity of a rarefied gas flow through a channel. In the limit of large mean free path, trajectories of gas molecules (point masses) injected into the channel as a short pulse, are independent of each other and their stochastic behavior provides information about the geometric microstructure of the inner surface of the channel. From exit flow rate data one determines the Knudsen self-diffusivity. The mathematical problem posed in this paper is the explicit determination of the diffusivity constant as a function of geometric parameters defining the microstructure.
\label{fig:idealized experiment}}
\end{center}
\end{figure}
\emph{Main results.}
The main results in the paper are centered around a detailed study of the Markov operator $P$ and
establish analytic and probabilistic properties of $P$ and its corresponding Markov chain.
To begin, we show that for a large class of microstructures, $P$ has a positive spectral gap,
which in turn establishes the ergodicity of the Markov chain
as well as the fact that functionals of the Markov chain satisfy the Central Limit Theorem.
We have shown in previous work \cite{Feres2007, FZ2012}
that $P$ is a self-adjoint, compact or quasi-compact operator on an appropriate Hilbert space,
for microstructures whose sides are concave with curvature bounded away from zero.
However, the present work establishes a positive spectral gap of $P$ for a significantly larger class of microstructures.
Using a conditioning technique, we show that $P$ has positive spectral gap
when only a certain positive measure portion of the billiard phase space is dispersing.
It is now a classical result in the theory of Markov chains \cite{KV1986} that one can obtain an expression
of the diffusivity of the Markov chain corresponding to $P$ in terms of an integral over the spectrum of $P$.
A key insight of the present work is that, for relatively flat microstructures, these quantities,
namely the diffusivity and the spectral gap of $P$,
are directly connected with a single summary geometric parameter
that can be computed in a straightforward way from a description of the surface microstructure,
which we call the surface \emph{flatness parameter} and denote by $h$.
The connection between these three properties of the system is obtained based on a fact which,
to the best of our knowledge, has not been noted previously in the context of computing the Knudsen diffusivity.
When scattering by $P$ is relatively weak (in a sense to be made precise),
it is natural to approximate the operator $P$ in the form $P=I+\mathcal L$,
where $I$ is the identity operator and $\mathcal L$ may be expected to take the form of a differential (velocity diffusion) operator.
We show that $\mathcal L$ has a universal form:
it is a constant (namely, up to a factor of 2, our flatness parameter $h$)
times the Legendre operator, whose (purely discrete) spectrum is known explicitly.
We are able to exploit the approximation of $P$ by the Legendre operator
to give an asymptotic expression and error estimates for the Knudsen diffusivity in terms of $h$.
The conceptual link we obtain between $h$, the Knudsen diffusivity, and the spectral gap of $P$ is, in our opinion,
a new theoretical insight in a very classical subject, which also yields a very effective method of computation,
at least in the case of small values of $h$.
The final concern of this work is to obtain and validate effective numerical methods for computing the Knudsen self-diffusivity in terms of the geometric microstructure parameters, in both the small and large $h$ case.
This will be discussed in detail in Subsections \ref{subsec:diff-approx} and \ref{sec:Numerical-techniques-examples} and Sections \ref{sec:Diffusivity} and \ref{sec:Numerical-techniques}.
A number of numerical experiments involving different microstructures will also be explored.
\emph{Remarks on assumptions.}
We make a few remarks about the assumptions in our model.
The analysis developed in this paper does not require in an essential way all the assumptions made,
but we hope that the greater simplicity of the present set-up will help to make clearer the main points.
For example, we have made a deliberate choice to consider periodic profiles of microstructures with relatively few geometric parameters to emphasize the relationship between geometric parameters, the spectrum of the Markov operator, which in turn establishes the relationship between geometric parameters and Knudsen diffusivity.
While it's possible to give similar formulas relating Knudsen diffusivity and geometric parameters for, say, randomly chosen profiles, such formulas are straightforward but tedious to express, and we fear would muddle the main point.
When studying the Knudsen self-diffusivity,
the observable which measures particle flight between collisions in the channel has infinite variance.
A study of the Central Limit Theorem and Knudsen diffusivity for a different class of random billiard Markov chains with infinite variance observables has been done previously in \cite{CFZ2016}.
While the methods in the present work can be adapted to the case of infinite variance observables,
we have chosen to use a cut-off observable to reduce to the finite variance case for the sake of clarity.
Besides clarity, there are a number of physically relevant reasons for considering the cut-off observable we have used in our examples.
Namely, (1) the cut-off can arise for macroscopic curvature of the channel in which the test particle traverses.
It can also arise due to (2) a finite mean free path resulting from unlikely but non-zero probability particle-particle collisions in the large but finite Knudsen number regime,
and (3) real systems where the channel is of finite length and bounded on either end.
These physically relevant mechanisms are discussed in detail in \cite{ACM2003}.
Finally, we should note that in the case of three dimensional cylinders, the inter-collision distance observable is always of finite variance, so our methods in the current paper serve as prototypes for this generalization.
The techniques we introduce here are in fact not particular to dimension $2$.
Indeed, a multivariable Legendre operator on the unit disc,
whose spectral theory is explicitly known,
plays the same role in higher dimensions as the classical Legendre operator does in the present work.
The details of this approximation and the corresponding models in higher dimensions---
where we consider three-dimensional cylindrical channels, parallel plates, and allow for collisions
that induce energy exchange between gas and surface at a given surface temperature---
are at the core of future work currently in preparation.
\emph{Related work.}
Better understanding of rarefied gas transport has practical implications for a number of engineering fields including high altitude gas dynamics, porous media, vacuum technology, nano- and microfluidics, among others. These applications have stimulated much experimental work. The following list of papers is a far from thorough or systematic sample of such work: \cite{A2014,M2009,PGEM2011,VNHDV2009,YMN2012}. The reader interested in the more applied side of the subject should consult these sources and others cited in them. From a purely mathematical perspective, this is a rich source of well motivated and potentially fruitful problems in the general theory of stochastic processes, and, more specifically, in
the study of the stochastic dynamics of {\em random billiard systems}. This is
our main motivation for studying the subject. We mention from the mathematical literature the following, also necessarily incomplete, list: \cite{ABS2013, BBG2019, CPSV2009, E2001, KY2013}.
\emph{Organization of the paper.}
The rest of this paper is organized as follows. In Section \ref{sec:definitions and results} we
detail our main results after introducing the necessary definitions;
we define what we call the {\em random billiard Markov chain} model in detail and state
some of its basic properties. Among the main results stated in Section \ref{sec:definitions and results}
(and proved in more general form later in the paper) we have that
under certain geometric conditions on the boundary microstructure,
the Markov chain has positive spectral gap and is uniformly ergodic. Numerical
evidence for this is then given for a few examples. With ergodicity
in hand, we discuss the central limit theory of the Markov chain providing explicit expressions for the variance of the limit diffusion
in terms of the Markov operator $P$. The main analytic technique for computing
diffusivity, based on a Galerkin method for solving a Markov-Poisson equation and a key observation that $P$ is closely related to the Legendre differential operator, is also given in this introductory section. This approach for obtaining diffusivity is then compared with
other more straightforward methods for a family of microstructures we call the {\em simple bumps} family.
A few more examples of microstructures are explored, having in mind the relation between geometric parameters, diffusivity, and spectral gap.
Section \ref{sec:Spectral-gap} is dedicated to stating and proving the analytical results of the paper in their general form,
while Section \ref{sec:Diffusivity} details, and adds further information, to the numerical methods and their validation.
\section{Main definitions and results\label{sec:definitions and results}}
\subsection{The billiard cell and its transition operator $P$}
The notation $\mathcal{P}(\Omega)$ will be used below to denote the space of probability measures on a measurable space $\Omega$. If $\mu$
is a measure on $\Omega$ and $f:\Omega\rightarrow \mathbb{R}$ is $\mu$-integrable, we write the integral of $f$ with respect to $\mu$ as
$$\mu(f):=\int_\Omega f(\omega)\, \mu(d\omega). $$
The Hilbert space of square integrable functions with respect to $\mu$
and its subspace of functions with mean zero will be written
$$L^2(\Omega,\mu)=\left\{f: \mu\left(f^2\right)<\infty\right\}, \ L_0^2(\Omega,\mu)=\left\{f\in L^2(\Omega,\mu): \mu(f)=0\right\},$$ with inner product $\langle f,g\rangle_\mu:=\int_\Omega f(\omega)g(\omega)\, \mu(d\omega)$ and norm $\|f\|_\mu:=\langle f,f\rangle_\mu^{1/2}$.
Moreover, we define a norm on the space of square integrable probability measures on $\Omega$ which are absolutely continuous with respect to $\mu$ as follows. Let $\nu$ be such a measure, so that $f$ is the Radon-Nikodym derivative of $\nu$ with respect to $\mu$. Then $\|\nu\|_\mu := \|f\|_\mu$.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.8\columnwidth]{billiard_cell.eps}
\par\end{centering}
\caption{ A periodic microstructure and its billiard cell, with some of the notation used to define the random billiard map and its transition operator $P$.
For some of our results we assume that the boundary curve is the graph of a piecewise smooth function $F:\mathbb{T}\rightarrow \mathbb{R}$.
\label{fig:billiard cell}}
\end{figure}
The general set-up will be that of a two-dimensional {\em random billiard} with static, periodic, geometric microstructure, as in \cite{CFZ2016,CF2012,Feres2007,FNZ2013,FY2004,FZ2012}. The periodic structure
is defined by the choice of a {\em billiard cell} $M$, from which the Markov operator $P$ will be defined. The billiard cell is a subset $M$ of $\mathbb{T}\times \mathbb{R}$, where $\mathbb{T}$ denotes the $1$-dimensional torus (equivalently, the interval $(0,\ell)$ with periodic condition imposed at the endpoints, where $\ell$ will typically be set equal to $1$.)
The boundary of the billiard cell is assumed to be a piecewise smooth curve.
For some of the results given below, the boundary will be the graph of a piecewise smooth function $F:\mathbb{T}\rightarrow \mathbb{R}$, so that $M$ consists of the points $(r,y)$ such that $y\geq F(r)$. Choose an arbitrary value $c$ such that $c>F(r)$ for all $r\in \mathbb{T}$. The line $y=c$ will be called the {\em reference line}. At any point $(r,c)$ on the reference line we define the half spaces $\mathbb{H}^2_-$ and
$\mathbb{H}^2_+$ of incoming and outgoing velocities, respectively. Thus $(r, c, v)\in M\times \mathbb{H}^2_-$ represents the initial conditions of
an incoming particle trajectory. These conditions uniquely specify (for almost every $r$ and $v$) a billiard trajectory: upon hitting a non-corner point on the cell boundary, the particle reflects specularly without changing speed, and upon crossing a vertical boundary line of $M$ (more precisely, a line
separating two adjacent cells, represented
in Figure \ref{fig:billiard cell} by the vertical dashed lines) it reenters the other (dashed) line with unchanged velocity. With probability $1$ on the set of initial conditions (due to Poincar\'e's recurrence), the trajectory returns to the reference line, at which point we register its outgoing velocity
$V(r,v)\in \mathbb{H}^2_+$ and new position $r'$. Without risk of confusion we may identify (through reflection about the reference line) $\mathbb{H}^2_-$ and $\mathbb{H}^2_+$, denoting both by
$\mathbb{H}^2$. We have thus defined a transformation $(r,v)\mapsto (r',V(r,v))$ (for almost all initial conditions $(r,v)$) on $\mathbb{T}\times \mathbb{H}^2$. We call this transformation the {\em return billiard map}.
Note that the vector norms satisfy $|v|=|V|$ since collisions are elastic. We may, without loss of generality, assume that the particle trajectories have unit speed. The incoming or outgoing {\em state space}, consisting of initial or return velocities, can then be taken to be the interval $\mathcal{X}=(0,\pi)$ of angles the particle velocity makes with the reference line. We can (and often will) equivalently define $\mathcal{X}=(-1,1)$ as the set of values of the cosine of those angles. Given an initial velocity $x \in \mathcal X$, we will often denote the return velocity by $X(r,x)\in\mathcal X$ in analogy with the earlier notation of velocities $v$ and $V(r,v)$ in $\mathbb{H}^2$.
Let $\mathcal{P}(\mathcal{X})$ denote the space of probability measures on $\mathcal{X}$. Given an incoming velocity $v$, let us suppose that $r=U$
is a random variable with the uniform distribution over $\mathbb T$. Thus $X(U,x)$ becomes a random variable. We now define the Markov (or transition probabilities) operator $P$ as follows.
Let $f$ be any bounded and continuous function on $\mathcal{X}$ and define
$$\left(Pf\right)(x):=E\left[f(X(U,x))\right] = \int_\mathbb{T} f(X(r,x))\, dr, $$
where $dr$ is the length element of normalized Lebesgue measure on $\mathbb T$.
Equivalently, we define a sequence of random variables $(X_n)_{n\geq 0}$ with a given initial distribution $\mu \in \mathcal P(\mathcal X)$ as follows. Let $(U_n)_{n\geq0}$ be an independent, identically distributed sequence of random variables uniformly distributed on $\mathbb T$, and, for each $n\geq 0$, let
$$X_{n+1} := X(U_n, X_n).$$
The justification for assuming, at each scattering event, that the point $r$ of entry over the opening of a billiard cell is random and uniformly distributed is due to our regarding the billiard cell as being very small relative to other length scales; any small uncertainty in the incoming velocity will make $r$ nearly fully uncertain. See \cite{FY2004} for a more detailed explanation of this point.
We can also regard $P$ as a map from $\mathcal{P}(\mathcal{X})$ to itself: Given any $\mu\in\mathcal{P}(\mathcal{X})$, let $\mu P\in \mathcal{P}(\mathcal{X})$ be such that for any test function $f$ (bounded and continuous),
$$(\mu P)(f) := \mu(Pf).$$
The following summarizes the basic properties of $P$.
For
their proofs, see \cite{Feres2007,FZ2012}. We say that the billiard cell $M$ is {\em bilaterally symmetric} (or simply {\em symmetric}) if it is invariant under reflection through the middle vertical
line. When the boundary of the cell is the graph of a function $F$, this means that
$F(r)=F(\ell-r)$
for all $r\in (0,\ell)$.
\begin{prop}\label{prop:P}
The Markov operator $P$, for any given billiard cell,
has the following properties.
\begin{enumerate}
\item The measure $\pi\in\mathcal{P}(\mathcal{X})$ given by $\pi(d\theta)=1/2\sin\theta\,d\theta$
is stationary of $P$. That is, $\pi P=\pi$.
\item As an operator on $L^{2}(\mathcal{X},\pi)$, $P$ has norm 1.
\item If $M$ is symmetric, $P$ is self-adjoint
and the stationary
Markov chain is reversible.
\end{enumerate}
\end{prop}
Note that when $\mathcal X = (-1,1)$, it is straightforward to see by a change of variables that the stationary measure $\pi$ is given by the uniform measure $\pi(dx) = 1/2\,dx$.
If the the billiard cell is not bilaterally symmetric, the adjoint of $P$ is still closely related to $P$ as described in \cite{CF2012} and much of the analysis developed in this paper still applies. For simplicity, we do not consider the more general type of cells here.
\subsection{Spectral gap and ergodicity\label{sec:Spectral-gap}}
Let $(X_n)_{n\geq 0}$ be the Markov chain with transition operator $P$ and initial distribution $\mu$. Then the measure $\mu P^n$ is the law
of the $n$th step $X_n$. We are interested in the convergence of $\mu P^n$ to the stationary measure $\pi$ in the sense of total variation. Recall that the {\em total variation} of a measure $\mu$ is defined as
$$\|\mu\|_{v}:= \sup_{A\subset \mathcal{X}} |\mu(A)|. $$
\begin{defn}
A Markov chain with stationary distribution $\pi$ is \emph{$\pi$-a.e.
geometrically ergodic }if there exists $0<\rho<1$ such that for $\pi$-a.e.
$x\in\mathcal{X}$ there exists a constant $M_{x}>0$ possibly dependent
on $x$ such that
$
\left\Vert \delta_{x}P^{n}-\pi\right\Vert _{v}\leq M_{x}\rho^{n}
$
for all $n\geq1$.
\end{defn}
The operator
$P$ has \emph{spectral gap }if there exists a constant $0<\rho<1$
such that $$\left\Vert Pf\right\Vert _{\pi}\leq\rho\left\Vert f\right\Vert _{\pi}$$
for all $f\in L_{0}^{2}(\mathcal{X},\pi)$. The value $\gamma:=1-\rho$
is called the {\em spectral gap} of $P$. It is straightforward
to see that for a compact and self-adjoint $P$, $\rho$ is given by the largest
eigenvalue of $P$ restricted to $L_{0}^{2}(\mathcal{X},\pi)$ and
$\gamma>0$. Finally, we note that if $P$ has spectral gap and is
self-adjoint, then for any initial distribution $\mu$ which is absolutely
continuous with respect to $\pi$, there exists a constant $M_{\mu}>0$
such that
\[
\left\Vert \mu P^{n}-\pi\right\Vert _{v}\leq M_{\mu}\rho^{n}.
\]
See \cite{RR1997}. We will prove
geometric ergodicity for a large class of microstructures satisfying
certain geometric conditions.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.3\columnwidth]{theorem_hypothesis.eps}
\par\end{centering}
\caption{An example of billiard cell for which Theorems \ref{thm:geo_erg} and \ref{thm:clt} hold. Other than being bilaterally symmetric, its shape is essentially arbitrary below a line $y=h_0$ whereas above it, the boundary consists of two smooth concave lines with curvature bounded below by some positive number $K$.}
\label{fig:theorem hypothesis}
\end{figure}
The following is a special case of a more general result to be stated and proved in Section \ref{sec:Spectral-gap}. We call the {\em height of the billiard cell} the supremum of the $y$ coordinate function restricted to the boundary of the cell.
\begin{thm}\label{thm:geo_erg}
Let $P$ be the Markov transition operator for a random billiard Markov chain whose billiard cell is symmetric and satisfies the following property:
above a certain $y=h_0$ strictly less than the height of the cell, the cell boundary is the union of smooth, concave curves having curvature bounded away from $0$. Then $P$ is a self-adjoint operator with a positive spectral gap. As a result, there exists a constant $\rho\in (0,1)$ such that for each $\mu \in \mathcal{P}(\mathcal{X})$ with $\|\mu\|_\pi<\infty$,
$$ \left\| \mu P^n -\pi \right\|_v\leq M_\mu \rho^n$$
for some constant
$M_\mu<\infty$ and $n\geq 1$.
\end{thm}
Figure \ref{fig:theorem hypothesis} gives an example of billiard cell for which Theorem \ref{thm:geo_erg} holds.
\subsection{Central Limit and Diffusivity}
Referring back to Figure \ref{fig:idealized experiment}, one expects for a sufficiently long channel that the molecular random flight
can be approximated by a Wiener process whose variance corresponds to the Knudsen self-diffusivity. This is justified by a Central Limit Theorem (CLT). This diffusivity has a convenient expression when the transition operator $P$ is self-adjoint. We
describe this expression here and prove further details later in the paper.
Let $(X_n)_{n\geq 1}$ be, as above, the stationary Markov chain generated by $P$, with stationary probability measure $\pi$. Recall that $X_n$ has values in the space of post-collision velocities $\mathcal{X}$. This space can be parametrized by the values of the cosine of the angle the velocity vector makes with the horizontal reference line $y=c$.
(See Figure \ref{fig:billiard cell}.) Thus we may set $\mathcal{X}=(-1,1)$. Let $f:\mathcal{X}\rightarrow \mathbb{R}$ be the observable
$$\tilde f(x)= 2 r x\left(1-x^2\right)^{-1/2} $$
where $r$ is the radius of the channel. We suppose, in the context of formulating a CLT for molecular trajectories, that the length
of the channel is infinite. Note that $\tilde f(X_n)$ is the distance travelled by the particle along the channel's horizontal axis between the $n$th and the $n+1$st collisions with the channel wall.
The total horizontal displacement up to the $n$th collision is
$$S_n(\tilde f)=\sum_{k=0}^{n-1} \tilde f(X_k).$$
In its standard form, the CLT gives a limit in distribution for expressions of the form $S_n(f)/\sqrt{n}$ where $f$ is an observable having mean zero and finite variance. A simple calculation shows that the horizontal displacement function $\tilde f$ has mean zero but infinite variance. For this reason we consider instead the following modified, cut-off displacement observable:
\begin{equation}\label{eqn:truncated}f_a(x):= \tilde f(x) \mathbbm{1}_{\{|\tilde f|\leq a\}}(x) + a \mathbbm{1}_{\{|\tilde f|>a\}}(x)\end{equation}
for large $a>0$. Here $\mathbbm{1}_I(x)$ denotes the indicator
function of the set $I$, which is defined as $\mathbbm{1}(x)=1$ if $x\in I$ and $0$ if $x\notin I$.
There are a number of physical mechanisms that could be invoked to make this cut-off plausible. For example, the channel might have
a slight curvature along its length, setting an upper bound on the horizontal distance traveled. See \cite{ACM2003} for an outline of other mechanisms.
We should also note that while the CLT with the
usual scaling does not hold for the observable $\tilde f$, the distribution of $\tilde f(X_n)$ is still
in the domain of attraction of the Gaussian law. One can check
that $\tilde f$ is slowly varying and, as a result, a CLT
with nonstandard scaling holds for random billiard Markov chains with
sufficient mixing. See \cite{CFZ2016} for a detailed study of such
Markov chains. The program we outline in this paper to estimate the diffusivity should hold in the
infinite variance case as well, but we have chosen to focus on the
finite variance case for the sake of clarity of exposition. It should
also be noted that for cylindrical channels in dimension 3 (and higher),
the observable that gives the distance traveled along the axis of
the channel is of finite variance.
We suppose the microstructure satisfies the same geometric assumptions of Theorem \ref{thm:geo_erg}. In particular, $P$ is self-adjoint and has positive spectral gap.
Let $\Pi$ be the {\em spectral resolution} of $P$\----the projection-valued
measure on the spectrum $\sigma(P)\subset [-1,1]$ granted by the Spectral Theorem for bounded self-adjoint operators. Then
$$ P=\int_{-1}^1 \lambda\, \Pi(d\lambda).$$
Let $f$ be any observable in $L_0^2(\mathcal{X}, \pi)$ (for example, the truncated displacement function $f_a$) and define the measure
$\Pi_f$ supported on $\sigma(P)\setminus \{1\}$ by $$\Pi_f(d\lambda):=\langle f, \Pi(d\lambda)f\rangle_\pi.$$
The following is a special case of a theorem that will be stated and proved in Section \ref{sec:Diffusivity}.
\begin{thm}\label{thm:clt}
Let $(X_n)_{n\geq 0}$ be a Markov chain taking values in $\mathcal{X}$ with Markov transition operator $P$ and stationary measure $\pi$.
Suppose $P$ is associated to a billiard cell satisfying the same geometric assumptions of Theorem \ref{thm:geo_erg}.
Let $f\in L^2_0(\mathcal{X},\pi)$. Then $S_n(f)/\sqrt{n}$ converges in distribution to a centered Gaussian random variable $\mathcal{N}(0, \sigma_f^2)$, where the variance is given by
$$\sigma_f^2 = \int_{-1}^1 \frac{1+\lambda}{1-\lambda} \Pi_f(d\lambda) = \langle f, f\rangle_\pi + 2\langle f, P(I-P)^{-1}f\rangle_\pi. $$
\end{thm}
The expression for the diffusivity given above
suggests the following approach for computing $\sigma_{f}^{2}$.
Let $L:=P-I$ be the \emph{Markov Laplacian} and $g$ the solution to the
\emph{Markov-Poisson equation $Lg=-f$}. Then
the dimensionless Knudsen self-diffusivity coefficient takes the form
\begin{equation}\label{eqn:eta} \eta=\frac{\sigma^2_f}{\sigma_0^2} = 1 +2 \|f\|_\pi^{-2} \left\langle f, Pg\right\rangle_\pi,\end{equation}
where $\sigma_0^2=\|f\|_\pi^2$ is the diffusivity for the process with independent post-collision velocities with the identical distribution $\pi$.
In the next subsection we explain one approach to carrying out this program by approximating $L$ by an elliptic differential operator
$\mathcal{L}$ whose spectral theory is well
understood.
It turns out that $\mathcal{L}$ has a canonical form as we show next.
\subsection{The Legendre Equation and Diffusion Approximation}\label{subsec:diff-approx}
Our aim now is to show that it is possible to approximate the solution
of the Markov-Poisson equation $Lg=-f$ for a large class of random
billiard microstructures when $P$ is close to the identity operator $I$. We consider families of microstructures
indexed by a scalar quantity $h$ that, in a sense to be made precise,
characterizes a key geometric feature of the microscopic billiard
cell, namely its flatness. For each microstructure with parameter
$h$, the corresponding Markov operator $P_{h}$ defines the dynamics
of the random billiard Markov chain as discussed previously. The key
idea now is that for small values $h$, the operator $P_{h}$ will
act nearly like the identity operator, due to the flatness of the
geometry; the Markov-Laplace operator $L_{h}:=P_{h}-I$,
in the limit as $h\to0$ and under some general assumptions
on the microscopic billiard cell, will then have a canonical approximation by
the classical Legendre differential operator, whose spectral theory is well understood.
In the rest of the subsection, we make explicit the necessary assumptions
on the geometry and give the statement of our operator approximation
result and provide examples.
Let the boundary of the billiard cell be the graph
of a periodic function $F:\mathbb{T}\to\mathbb{R}.$ (See Figure \ref{fig:billiard cell}.)
In order to characterize how flat the microstructure boundary is,
we consider the normal vector field $\mathbbm n:\mathbb{T}\to\mathbb{R}^{2}$
along the graph of $F$, and let $\bar{\mathbbm n} = \bar{\mathbbm n}(r)$ denote its projection
onto its first (horizontal) component. Finally, we let
\begin{equation}
h:=\int_{\mathbb{T}}\bar{\mathbbm n}^{2}\,dr = \int_{\mathbb T}\frac{F'(r)^{2}}{1+F'(r)^{2}}\,dr.\label{eq:flat}
\end{equation}
It will be seen in examples that $h$ captures information
about the curvature of the boundary. For small values of $h$, the collision events
with the boundary will be relatively simple, often resulting in only
a single collision with the cell's boundary and only
a small deviation from specular reflection. This implies little change in the tangential momentum of the particle with high probability. It is in this sense that $h$
can be thought to have a role similar to the accommodation coefficient $\vartheta$ referred to earlier in the paper.
Let $\mathcal X = (-1,1)$ and let $\mathcal{L}$ denote the differential operator acting on smooth functions
$f:\mathcal{X}\to\mathbb{R}$ as
\begin{equation}
\mathcal{L}f(x)=\frac{d}{dx}\left(\left(1-x^{2}\right)\frac{d}{dx}f(x)\right).\label{eq:legendre}
\end{equation}
\begin{thm}
\label{thm:diffusion-app}Let $(F_{h})_{h>0}$ be a
family of piecewise smooth functions $F_{h}:\mathbb{T}\to\mathbb{R}$ defining bilaterally symmetric billiard cells, indexed by the
flatness parameter $h$ introduced in (\ref{eq:flat}).
Let $(P_{h})_{h>0}$ be the corresponding Markov transition operators.
Then for any $f\in C^{3}(\mathcal{X})$,
\[
L_{h}f(x)=2h\mathcal{L}f(x)+O\left(h^{3/2}\right)
\]
holds for each $v$ such that every initial condition with velocity $v$ results in a trajectory that collides only once with the boundary of the cell.
\end{thm}
In the context of Theorem \ref{thm:diffusion-app} we observe that, for each $x \in \mathcal X$, every initial condition with velocity $x$ results in a trajectory
that collides only once cell boundary as long as we take
$h$ to be sufficiently small.
The differential operator $\mathcal{L}$ has a well understood spectral
theory that will be used to obtain information about
$P_{h}$. We recall that the eigenvalue problem $\mathcal{L}f=\lambda f$ has square integrable solutions
if and only if $\lambda$ is of the form $\lambda=-l(l+1)$ for integers
$l\geq0$.
The associated eigenfunctions are the Legendre polynomials $\phi_{l}$,
$l\geq0$,
$$\phi_{0}=1,\ \ \phi_{1}(x)=x,\ \ \phi_{2}(x)=(3x^{2}-1)/2, \dots.$$
The collection $(\phi_{l})_{l\geq0}$ forms
a complete orthogonal basis for $L^{2}(\mathcal X, \pi)$ and
\[
\langle\phi_{n},\phi_{m}\rangle_\pi =\int_{\mathcal X}\phi_{n}(x)\phi_{m}(x)\,\pi(dx)=\frac{1}{2n+1}\delta_{n,m},
\]
where $\delta_{n,m}$ is the Kronecker delta symbol.
As a
first application of the approximation given in Theorem \ref{thm:diffusion-app},
we give an informal estimation of the spectral gap $\gamma_h$ of $P_{h}$
for values of $h$ near 0.
Note that the largest eigenvalue of $P_{h}$
is 1, with eigenfunctions given by the constant functions. So $\gamma_h$ is given by $1-\lambda$ where $\lambda$
is the second largest eigenvalue of $P_{h}$. Using the approximation
in Theorem \ref{thm:diffusion-app},
\[
P_{h}\phi_{l}=\left(1-2hl(l+1)\right)\phi_{l}+O(h^{3/2}),
\]
where $\phi_{l}$ is the Legendre polynomial associated to eigenvalue
$-l(l+1)$. This suggests that the second largest eigenvalue $\lambda$
of $P_{h}$ is given by
$
\lambda\approx1-4h
$. Equivalently, this suggests the following asymptotic estimate of $\gamma_h$:
\begin{equation}\label{eq:gap}\gamma_h\approx 4h. \end{equation}
The idea then will be to use the
approximation $\mathcal{L}$ of the Markov-Laplacian $L$ in order
to give an approximation of the function $g=(I-P)^{-1}f$ that appears
in the equation $$\sigma_f^2= \left\langle f, f\rangle_\pi + 2\langle f, P(I-P)^{-1}f\right\rangle_\pi$$
obtained in Theorem \ref{thm:clt}.
Note that $g$ is a solution
of the Markov-Poisson equation $Lg=-f$. The following thorem shows that a series
solution of the Poisson equation for $\mathcal{L}$ can be given explicitly
in terms of Legendre polynomials.
%
\begin{thm}\label{thm:sigma_approx}
Let $(P_{h})_{h>0}$ be a family of random billiard Markov transition
operators for a family of billiard cells satisfying the geometric assumptions of Theorems \ref{thm:clt} and \ref{thm:diffusion-app}. For any function $f\in L_{0}^{2}(\mathcal{X},\pi)$,
let $\sigma_{f,h}^{2}$ denote the diffusivity corresponding to $P_{h}$.
Then
\begin{equation}
\sigma_{f,h}^{2}=-\langle f,f\rangle_\pi+\frac{1}{h}\sum_{l=1}^{\infty}\frac{2l+1}{l(l+1)}\left\langle \phi_{l},f\right\rangle_\pi^{2}+O(h^{1/2}).\label{eq:variance-h-approx}
\end{equation}
\end{thm}
\begin{rmk}
It should be noted that for the sake of numerical computations, it is natural to consider the quantity given by truncating the series in (\ref{eq:variance-h-approx}) after a fixed number of terms $n \geq 1$, so that
$$\sigma^2_{f,h} = -\langle f, f \rangle_\pi + \frac 1h \sum_{l=1}^n \frac{2l+1}{l(l+1)} \langle \phi_l, f \rangle_\pi^2 + E_{h,n},$$
where $E_{h,n}$ is the tail of the series along with the $O(h^{1/2})$ error term. This quantity can be estimated as follows:
$$E_{h,n} = \frac 1h \sum_{l=n+1}^\infty \frac{2l+1}{l(l+1)}\langle \phi_l, f \rangle_\pi^2 +O(h^{1/2}) \leq \frac{\|f\|^2_\pi}{h} \sum_{l=n+1}^\infty \frac{2l+1}{l(l+1)}\|\phi_l\|_\pi^2 + O(h^{1/2})= \frac{\|f\|_\pi^2}{h(n+1)} + O(h^{1/2}).$$
\end{rmk}
The theorem implies that the dimensionless self-diffusivity coefficient satsifies
$$\eta_f = -1 + \frac{1}{h}\sum_{l=1}^{\infty}\frac{2l+1}{l(l+1)}\left\langle \phi_{l},f/\|f\|_\pi\right\rangle_\pi^{2}+O(h^{1/2}) = -1 + \frac1h C_f + O\left(h^{1/2}\right),$$
where $C_f$ is defined by this identity. Thus, for small $h$,
\begin{equation}\label{eq:approx} \eta_f \approx \frac{C_f -h}{h}.\end{equation}
Then the approximate identity (\ref{eq:gap})
suggests
\begin{equation}\label{eq:approx_gamma} \eta_f \approx \frac{4C_f -\gamma}{\gamma}.\end{equation}
It is interesting to compare this expression with the one obtained under the Maxwell-Smoluchowski model:
$$\eta = \frac{2-\vartheta}{\vartheta}$$
where $\vartheta$ is the accommodation coefficient, defined as the fraction of diffuse collisions.
We thus obtain a conceptual relation linking the purely geometric quantity $h$ (flatness), the spectral quantity $\gamma$ (spectral gap), and
the tangential momentum accommodation coefficient $\vartheta$ defined for a standard and widely used collision model.
Finally, it
is worth comparing these expressions with the exact equation
$$ \eta_f = \int_{-1}^1 \frac{2-\vartheta}{\vartheta} \overline{\Pi}_f(d\vartheta)$$
where $\overline{\Pi}_f(d\vartheta)=\Pi_f(d\vartheta)/\|f\|^2_\pi$, which is obtained from
Theorem \ref{thm:clt} by setting $\vartheta=1-\lambda$.
\subsection{Two Examples}
Consider the microscopic billiard cell, which we will refer to as
the {\em small bumps} microstructure throughout the discussion, whose boundary
is given by arcs of circles as in Figure \ref{fig:bumps}. The
geometric parameter of interest here is the dimensionless curvature
given by $K=\ell/R$, where $R$ is the radius of one of the arcs and $\ell$ is the length of the opening to the billiard cell as shown in the figure.
An elementary computation using (\ref{eq:flat}) gives
\[
h=\frac{K^{2}}{12}.
\]
As a result, the spectral gap, approximated for values of $K$ near
zero, is given by $$1-\lambda\approx4h=K^{2}/3.$$
Figure \ref{fig:bumps-spectral-gap} shows the numerically obtained values for the spectral gap and $\eta$ compared to the respective approximations as functions of the dimensionless curvature parameter $K$.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.4\columnwidth]{bumps.eps}
\par\end{centering}
\caption{The {\em bumps} microstructure with dimensionless curvature parameter $K$.}
\label{fig:bumps}
\end{figure}
\begin{figure}[h]
\begin{centering}
\includegraphics[width=1.0\columnwidth]{bumps_spectral_gap_w_approx}
\par\end{centering}
\caption{Left: the spectral gap of the operator $P$ for the bumps family of microstructures depicted in Figure \ref{fig:bumps},
with dimensionless curvature parameter $K$, compared with the
approximation of the Markov-Laplacian by the Legendre differential operator. The solid
curve is constructed from the numerical approximation detailed in
Section \ref{sec:Numerical-techniques-examples}. Right: comparison of the dimensionless diffusivity coefficient $\eta$ obtained using (\ref{eqn:eta})
and a finite dimensional approximation of $P$ (indicated on the graph by the stars) and the approximation of $\eta$ as a function of the geometric parameter
given by (\ref{eq:approx}). The observable is $f_a$ with cut-off $a=50000$.
\label{fig:bumps-spectral-gap}}
\end{figure}
A similar computation can be done for the microgeometry in Figure
\ref{fig:bumps alpha} that consists of a mixture of the small bumps
geometry together with flat, specularly reflecting lines. In this
case, the family is parameterized by the proportion of initial positions
$\alpha$ that result in reflections with the part of the boundary
with curvature. After expressing the boundary as the graph of an appropriately
defined function and computing an elementary integral, we get that
$h=\alpha/3$.
Generalizing this second example, consider the transition operator $$P_\alpha = \alpha P_1+(1-\alpha)I $$
where $P_1$ is the operator associated to a given microstructure.
Then $P_\alpha$ is associated to the microstructure for which a segment of horizontal line of length $d$ is added to the billiard cell of the first microstructure.
The parameter $\alpha$ is then the probability that an incoming particle will not collide with the flat segment. It is easy to see the effect of the additional parameter $\alpha$. Note that $P_\alpha-I = \alpha(P_1-I)$. An elementary algebraic manipulation starting from the expression
$$ \sigma_{f,\alpha}^{2} =\left\langle f,f\right\rangle_\pi +2\left\langle P_{\alpha}f,(I-P_{\alpha})^{-1}f\right\rangle_\pi$$
gives
$$\eta_{f,\alpha} = \eta_{f,1} + \frac{2(1-\alpha)}{\alpha} \frac{\langle f, (I-P_1)^{-1}f\rangle_\pi}{\|f\|_\pi^2}$$
where $f$ is arbitrary. As it is to be expected, $\eta_{f,\alpha}$ approaches infinity as the probability of specular reflection increases to $1$.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.5\columnwidth]{bumps_alpha.eps}
\par\end{centering}
\caption{Adding a flat segment to a given microstructure, as indicated in this diagram, gives the transition operator $P_\alpha=\alpha P_1+(1-\alpha)I$,
where $P_1$ is the operator associated to the original microstructure. }
\label{fig:bumps alpha}
\end{figure}
\subsection{Summary of the numerical techniques and examples \label{sec:Numerical-techniques-examples}}
In equation (\ref{eq:variance-h-approx}) of Theorem \ref{thm:sigma_approx}, we have given our main numerical approach of the paper with respect to analyzing the regime of small flatness parameter $h$; namely, we estimate the dimensionless self-diffusivity $\eta = \eta_f$ by truncating the series in equation (\ref{eq:variance-h-approx}).
In this subsection we outline an additional numerical approach for computing the dimensionless self-diffusivity $\eta$ (or, equivalently, the variance $\sigma_f^2$ of the Gaussian limit of the random flight in a channel).
This method, which we will refer to as the Galerkin method, requires us to introduce a finite rank approximation,
or discretization, of the Markov operator $P$, which we describe below.
The purpose of introducing this additional approach is two-fold.
First, the Galerkin method serves as numerical verification of the main approach of using equation (\ref{eq:variance-h-approx}).
Additionally, the Galerkin method is applicable to microstructures which fall outside of the small $h$ regime.
As we will see, however, the method has the disadvantage of requiring a discretization of $P$ and, for this reason, is more computationally demanding.
We conclude this subsection with a discussion of some additional examples that show the subtle relationship between the spectral gap, the dimensionless self-diffusivity, and geometric features of the microstructure.
The starting point in computing $\eta$ is the equation
$\sigma_f^2= \left\langle f, f\rangle_\pi + 2\langle f, P(I-P)^{-1}f\right\rangle_\pi,$
which in turn requires that we obtain the solution $g$ to the Markov-Poisson equation $(P-I)g=-f.$
The classical Galerkin method
gives us a general approach for solving this equation as follows
(see \cite{atkinson1987discrete} for a broader discussion of the approach).
For each $n \geq 1$, let $T_n:L_0^2(\mathcal{X},\pi)\rightarrow R_n$ denote the orthogonal projection to
the linear span $R_n=\{\phi_1, \dots, \phi_n\}$ of Legendre polynomials defined on $\mathcal{X}$.
Define $g_n\in L_0^2(\mathcal{X},\pi)$ to be the solution of the finite dimensional linear system
$$ (I-T_nP)g_n = T_nf.$$
Equivalently, we find $g_n\in R_n$ so that
$\left\langle (I-P)g_n, \psi\right\rangle_\pi = \langle f, \psi\rangle_\pi$
for all $\psi\in R_n$, which can be done as follows.
Writing $g_n=\sum_{j=1}^n \alpha_j \phi_j$
and defining $x=(\alpha_1, \dots, \alpha_n)^\intercal,
y= (\langle f, \phi_1\rangle_\pi, \dots, \langle f, \phi_n\rangle_\pi)^\intercal,$
and $G=\left(\langle \phi_j,\phi_i\rangle_\pi - \langle P\phi_j, \phi_i\rangle_\pi\right)_{i,j=1}^n,$
we are left to do two computations.
First, we find the entries of the matrix $G$.
Second, we find the solution $x$ to the linear system $Gx=y$.
This then gives the solution $g_n$ to the finite dimensional linear system, and from it the approximate value $\sigma^2_{\mathrm{GM},n}$.
The following theorem provides an error estimate for this approximation.
A proof is given in Section \ref{sec:Numerical-techniques}.
Figure \ref{fig:Galerkin error} gives numerical verification of the convergence and error bound for $\sigma^2_{\mathrm{GM},n}$
as given in the theorem.
\begin{thm}\label{thm:galerkin}
Let $f\in L_0^2(\mathcal{X},\pi)$, where $\mathcal{X}=(-1,1)$, be such that the first derivative $f'$ is absolutely continuous and the second derivative $f''$ is of bounded variation. Let $\sigma_f^2$ be defined by the equation
$$\sigma_f^2 = \langle f, f\rangle_\pi + 2\left\langle Pf, (I-P)^{-1} f\right\rangle_\pi. $$
Then
$ \lim_{n\rightarrow \infty} \sigma^2_{\mathrm{GM},n} =\sigma_f^2.$
Moreover, we have the following rate of convergence:
$$ \left|\sigma_f^2 -\sigma_{\mathrm{GM},n}^2\right|\leq \frac{C}{4n-6}$$
where $C$ is a constant depending on $f$ and $P$ but independent of $n$.
\end{thm}
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.5\columnwidth]{Galerkin_variance.eps}
\par\end{centering}
\caption{Error bound for the Galerkin approximation of $\sigma_f^2$ given in Theorem \ref{thm:galerkin}.}
\label{fig:Galerkin error}
\end{figure}
In practice, the drawback in the Galerkin method arises in finding the entries of the matrix $G$.
We compute the entries of $G$ by introducing a finite dimensional matrix $P_M$
that approximates the Markov operator $P$
and perform numerical integration.
We now describe how $P_M$ is constructed.
Given the billiard cell $\mathcal{M}$ with phase space $\mathcal{V}=\mathbb{T}\times\mathcal{X}$,
we partition $\mathbb{T}$ and $\mathcal{X}$ into $N$ and $M$
evenly spaced subintervals $\{I_1\ldots, I_N\}$ and $\{J_1\ldots, J_M\}$, respectively.
For each subinterval in the partitions, we choose a representative element,
e.g. the midpoint, to construct the sequences $\{r_k\}_{k=1}^{N}$ and $\{x_\ell\}_{\ell=1}^{M}$, respectively.
For each pair $(r,x)$ in the set $\{(r_k,x_\ell) \;:\; 1\leq k\leq N, 1\leq \ell\leq M \}$
we then simulate the standard billiard motion of a particle in the cell $\mathcal{M}$
with initial conditions $(r,x)$ and record the particle's velocity upon its return to $\mathbb{T}$.
The finite rank approximation $P_M$ is then the $M\times M$ matrix
whose $ij$ entry is the proportion of $N$ trajectories
whose initial velocity $x_i\in J_i$ yield a return velocity in the subinterval $J_j$.
We mention here that $\sigma_{GM,n}^2$ stabilizes for moderately sized values of $M$ (say $M=1000)$,
independent of the choice of $n$.
Moreover, we have used the matrix $P_M$ to give another numerical approximation of $\sigma_f^2$.
For larger values of $h$, it is possible to solve the Markov-Poisson equation $(P_M-I)g=-f$
using a standard numerical linear system solver (which implements the bi-conjugate stabilized method, or BiC method).
We should mention that the BiC method is used simply to give numerical verification for the use
of the Galerkin method in the large $h$ regime. For small values of $h$, the spectral gap of $P_M$ is small
and the condition number of $I-P_M$, propositional to the inverse of the spectral gap of $P_M$, is too high to be reliable.
In Figure \ref{fig:variances}, we have shown a comparison of approximations for $\sigma_f^2$
for the small bumps family introduced in the previous subsection and the observable $f_a$ defined in (\ref{eqn:truncated})
as produced by the three methods: (1) using equation (\ref{eq:variance-h-approx}) truncated to $n$ terms and denoted by $\sigma^2_{\text{Lser},n}$, (2) using the Galerkin method with dimension $n$ and denoted by $\sigma^2_{\text{GM},n}$,
and (3) using the BiC linear system solver.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=1\columnwidth]{New_bumps.eps}
\par\end{centering}
\caption{Comparison of the variance, computed using the methods described in this section,
for the simple bumps family with observable $f$
given by the horizontal displacement function along the length of the channel with cut-off $a=50000$
(there are no qualitative differences for larger choices of the cut-off).
The value of $\sigma^2_{\mathrm{Lser},n}$ is computed using (\ref{eq:variance-h-approx})
where we have used the first $n = 500$ terms in the series.
For $\sigma^2_{\mathrm{GM},n}$ we have used dimension $n=200$ and a finite rank approximation $P_M$
of $P$ with $M=1000$.
The same matrix $P_M$ has been used for $\sigma^2_{\text{BiC}}$.
On the left, the dimensionless curvature parameter $K$ is relatively small, while on the right it is relatively large.
The inset of the graph on the left zooms in on the smallest values of $K$, where the
$\sigma^2_{\mathrm{Lser},n}$ and $\sigma^2_{\mathrm{GM},n}$ approximations are valid.}
\label{fig:variances}
\end{figure}
We conclude the section with the result of two more numerical experiments.
A first example is given by the family of microstructures depicted in Figure \ref{fig:twobumps}. There are two competing curvatures, which are fixed while the height parameter $d$ varies over a range of positive and negative values. When $d<0$, the higher curvature bump is more exposed and when $d>0$
the smaller curvature bump is on top.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.5\columnwidth]{twobumps.eps}
\par\end{centering}
\caption{The two-bumps family. By varying the parameter $d$ keeping the curvatures constant, we can investigate how the two curvatures compete against each other in the determination of the spectral gap and the dimensionless coefficient of self-diffusivity $\eta$. The result is shown in Figure \ref{fig:twobumps_plot}.}
\label{fig:twobumps}
\end{figure}
The numerical results are shown in the plots of Figure \ref{fig:twobumps_plot}. The interpretation is somewhat straightforward: when the bigger curvature bump is more exposed to collision with the particles, scattering is more diffuse, spectral gap is larger, and diffusivity is smaller (slower diffusion), than
when the less curved bump rises above the other.
Perhaps more surprising is the near perfect mirror symmetry between the two graphs.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.6\columnwidth]{two_curvatures_plot.eps}
\par\end{centering}
\caption{Spectral gap and dimensionless coefficient of self-diffusivity for the microstructure of Figure \ref{fig:twobumps}.}
\label{fig:twobumps_plot}
\end{figure}
In the second example we obtain the dimensionless diffusivity and spectral gap for the
one-parameter family of microstructures indicated in Figure \ref{fig:bumps_with_wall_figure}. Here the parameter investigated is the (dimensionless) width
of the flat top wall, while the radius $R$ of the curved part is kept constant. Diffusivity is computed using the Galerkin method (dimension $200$) while the spectral gap is obtained more directly by computing eigenvalues of the finite dimensional approximation of $P$.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.4\columnwidth]{bumps_and_wall_figure.eps}
\par\end{centering}
\caption{Bumps with flat top wall microstructure. The geometric parameters we vary are the relative width $w$ and height $d$ of the flat top wall. }
\label{fig:bumps_with_wall_figure}
\end{figure}
The results are now somewhat harder to interpret. The interplay between the flat wall top, the curvature of the middle bumps, and reflection on the sides
of the walls creates a qualitatively more complicated effect. Nevertheless, both this and the previous example show a marked transition in the values of
diffusivity and spectral gap as the height of the wall (with curved top in the first example and flat top in the second) crosses the height of the adjacent curved segments. Once again, we observe near mirror symmetry in the graphs of spectral gap and diffusivity as functions of the geometric parameter. This is an interesting observation that merits further investigation.
\begin{figure}[h]
\begin{centering}
\includegraphics[width=0.8\columnwidth]{Bumps_with_wall_plot.eps}
\par\end{centering}
\caption{Spectral gap and diffusivity coefficient $\eta$ for the microstructure shown in Figure \ref{fig:bumps_with_wall_figure}.
The geometric parameters being varied are the relative height $d$ of the flat top wall and its relative width $w$. }
\label{fig:bumps_with_wall_plot}
\end{figure}
\section{Spectral gap and ergodicity\label{sec:Spectral-gap}}
The theorems of the previous sections will be strengthened and proved in this and following sections.
We begin this section by introducing a useful technique for decomposing
the operator $P$. The idea will be to condition on the event that
a billiard trajectory within the microscopic cell satisfies certain
properties, which will allow us to focus attention on geometric features
of the microgeometry that create mixing in the dynamics. More specifically,
we show here that under assumptions to be stated, the transition probability
operator $P$ for the random billiard Markov chain has a spectral
gap by showing that for certain components of the decomposition it
is a Hilbert-Schmidt operator. This, along with an additional geometric
assumption that yields a reversible Markov chain, in turn will give
ergodicity.
Let $N:=\mathbb T \times \mathcal X$
be the space of initial conditions of a scattering event and let $N_{1},N_{2},\ldots$
be a measurable partition of $N$. For each $x \in \mathcal X$ and $i\geq 1$, let $N_{i}(x):=\left\{ r\in\mathbb{T}:(r,x)\in N_{i}\right\} $.
Define $\alpha_{i}(x):=|N_{i}(x)|$,
where $|\cdot|$ denotes the size of a set under the normalized Lebesgue measure on $\mathbb T$.
For each $f\in L^{2}(\mathcal{X},\pi)$, define
\[
(P_{i}f)(x)=\begin{cases}
\frac{1}{\alpha_{i}(x)}\int_{N_{i}(x)}f(X(r,x))\,dr & \text{if \ensuremath{\alpha_{i}(x)\neq0}}\\
0 & \text{if \ensuremath{\alpha_{i}(x)=0.}}
\end{cases}
\]
We refer to $P_{i}$ as the \emph{conditional
operator} associated to partition element $N_{i}$. Note that $P_{i}\mathbbm{1}_{A}(x)$ is the
conditional probability that the outgoing velocity vector is in $A\subset\mathcal{X}$
given pre-collision velocity $x$ and given that the event
$N_{i}$ holds. Let $\pi_{i}$ denote the measure on $\mathcal{X}$
such that $\pi_{i}(dx)=\alpha_{i}(x)/(dr\otimes\pi)(N_{i})\,dx$.
Then $\pi_{i}$ is the conditional measure given by $\pi$ conditioned
on the event that $N_{i}$ holds. Finally, observe that for any $f\in L^{2}(\mathcal{X},\pi)$,
it makes sense to decompose $P$ as follows:
\begin{equation}\label{eq:decomposition}
(Pf)(x)=\sum_{i}\alpha_{i}(P_{i}f)(x).
\end{equation}
We now outline some properties of the conditional operators and the
resulting decomposition of $P$. For details of proofs, see \cite{FZ2012}.
\begin{prop}
Let $P_{i}$, $i \geq 1$, be the conditional operators associated to the measurable
partition $N_{1},N_{2},\ldots$ of the space $N$ of initial conditions
of billiard trajectories within billiard microcell $M$, and let $\pi_{j}$
be the conditional measures associated to the partition. Then for each $i \geq 1$,
\begin{enumerate}
\item $P_{i}$ has norm 1.
\item Each term $\alpha_{i}P_{i}$ in the decomposition has norm at most
$\left\Vert \alpha_{i}\right\Vert _{\infty}.$
\item If $N_{i}$ is symmetric---that is, it is invariant under the map
$(r,x)\mapsto(1-r,Jx)$ where $Jx$ denotes the reflection across
the vertical axis in $\mathbb{H}_{-}^{2}$ of the velocity vector corresponding to $x$ and $\mathbb{T}$ is identified
with the unit interval--- then $P_{i}$ is self-adjoint as an operator
on $L^{2}(\mathcal{X},\pi_{i})$.
\end{enumerate}
\end{prop}
The following assumptions will be shown to be sufficient
for ergodicity.
\begin{assumption}
\label{assu:reversibility}The billiard cell is symmetric
with respect to reflection across the vertical axis given by the map
$(x,y)\mapsto(-x,y)$.
\end{assumption}
\begin{assumption}
\label{assu:partition}There exists a measurable partition $N_{1},N_{2},\ldots$ whose elements are symmetric and
such that the following holds for at least one partition element $N_{j}$.
\begin{enumerate}
\item The trajectories with initial conditions in $N_{j}$ collide only
with portions of the boundary of the microscopic billiard cell consisting
piecewise smooth concave curves whose curvatures are bounded below by a constant $K>0$.
\item $\inf_{v\in\mathcal{X}}\alpha_{j}(v)>0$.
\end{enumerate}
\end{assumption}
Note that these assumptions are not optimal---for example,
billiard cells with convex sides have been shown to give
geometrically ergodic random billiard Markov chains in \cite{CFZ2016}---but
capture a large class of examples like those in Section \ref{sec:definitions and results}.
The key idea of Assumption \ref{assu:partition} is that partitioning the phase space
and subsequently decomposing the Markov transition operator into corresponding conditional operators
allows us to focus our study of the operator only on the features that create enough dispersion to yield ergodicity.
\begin{thm}
\label{thm:small-bumps-gap}Let $P$ be the Markov transition operator
for a random billiard Markov chain whose billiard cell
satisfies Assumptions \ref{assu:reversibility} and \ref{assu:partition}.
Then $P$ is a self-adjoint operator with spectral gap. As a result,
there exists a constant $\rho\in(0,1)$ such that for each probability
measure $\mu\in\mathcal{P}(\mathcal{X})$, absolutely continuous with respect to $\pi$ with $\|\mu\|_\pi<\infty$,
there exists a constant $M_{\mu}<\infty$ such that $\left\Vert \mu P^{n}-\pi\right\Vert _{v}\leq M_{\mu}\rho^{n}.$
\end{thm}
Note that Theorem \ref{thm:small-bumps-gap} generalizes Theorem \ref{thm:geo_erg}. Indeed, for billiard cells that satisfy the geometric property in the hypotheses of Theorem \ref{thm:geo_erg}, it is clear that for each $x$, there exists an open set $W_x^1 \subset \mathbb{T}$ such that for each $r \in W_x^1$, the billiard trajectory with initial condition $(r,x)$ results in one collision with the boundary of the billiard cell before returning to the reference line. Letting $N_1 = \{(r,x) : x \in \mathcal{X}, r \in W_x^1\}$ and $N_2 = N\setminus N_1$, it is clear that Assumptions \ref{assu:reversibility} and \ref{assu:partition} are satisfied.
We also note that Theorem \ref{thm:small-bumps-gap} includes as a special case, the case of i.i.d. mixtures of microstructures.
The proof of Theorem \ref{thm:small-bumps-gap} requires a series of lemmas, which we now introduce. Note that these lemmas are adapted from a series of lemmas in \cite{Feres2007} but the present statements have more relaxed hypotheses on the geometry of the billiard cell and thus are stronger.
Before stating the first lemma, we need to introduce some notation. Consider a measurable partition satisfying the conditions in Assumption \ref{assu:partition}, where $N_j$ and $P_j$ are the partition element and corresponding conditional operator that satisfy the restrictions in the assumption. Let $W_x^i := \{r \in \mathbb{T} : (r,x) \in N_i\}$ for each partition element $N_i$. For each $x \in \mathcal X$, we let $X_x : \mathbb T \to \mathcal X$ be the function given by $X_x(r) = X(r,x)$, where $X(r,x)$ is the return velocity at the reference line of the billiard cell for a trajectory with initial condition $(r,x)$.
\begin{lem}
\label{lem:kernel-def}Suppose the billiard cell satisfies Assumption \ref{assu:partition}, with partition element $N_j$ and conditional operator $P_j$ satisfying the conditions in the assumption. Then for all $x \in \mathcal X$, the set $W_x^j = \{r \in \mathbb{T} : (r,x) \in N_j\}$ consists of a countable union of open intervals $W_{x,i} \subset \mathbb{T}$. Moreover, the restriction $X_{x,i} := X_x|_{W_{x,i}}$ is a diffeomorphism from $W_{x,i}$ onto its image $V_{x,i}$. Finally, when we use the convention $\mathcal X = (0,\pi)$, we have that for all $f \in L^2(\mathcal{X}, \pi_j)$, $P_jf(x)=\int_{\mathcal{X}}f(\phi)\,\omega(x,\phi)\,\pi_j(d\phi)$
where
\begin{equation}
\omega(x,\phi):= \frac{(\lambda\otimes\pi)(N_j)}{\alpha_j(x)\alpha_j(\phi)}\sum_{i}\mathbbm{1}_{V_{x,i}}(\phi)\left(\frac{1}{2}\left|X_{x,i}'\left(X_{x,i}^{-1}(\phi)\right)\right|\sin\phi\right)^{-1}\label{eq:kernel-1}
\end{equation}
and $\mathbbm{1}_{V_{x,i}}$ denotes the indicator function of
the set $V_{x,i}$.
\end{lem}
\begin{proof}
We begin by outlining some standard facts in the theory of classical billiards. See \cite{CM2006} for details.
Let $\Gamma$ denote the boundary of the billiard cell $Q$ and note that $\Gamma = \bigcup_i \Gamma_i$ consists of a union of smooth component curves, or walls.
We denote by $\Gamma_0$ the reference line, which is identified with $\mathbb{T}$. Let $\mathcal{M} = \bigcup_i \mathcal{M}_i$ be the collision space, where each set $\mathcal{M}_i$ consists of pairs $(q,v)$ where $q \in \Gamma_i$ and $v$ points into the interior of $Q$.
The billiard map $\mathcal{F}:\mathcal{M} \to \mathcal{M}$ is the map defined so that $\mathcal{F}(q,v)$ gives the pair $(q',v')$ where $q'$ is the first intersection of the ray $q+tv$, $t>0$, with $\partial Q$.
The normalized measure $m\otimes\pi \in \mathcal{P}(\mathcal{M})$, where $m$ is the normalized arclength measure on $\partial Q$, is left invariant by $\mathcal{F}$.
Moreover, if we let $T:N \to N$ be the first return map of billiard orbits, the measure $dr \otimes \pi$, where $dr$ is the normalized Lebesgue measure on $\mathbb{T}$, is left invariant by $T$.
By Poincar\'e recurrence, there is a subset $E_0 \subset N$ of full $dr\otimes\pi$ measure of orbits that start at and return to $\Gamma_0$ in a finite number of steps, and the orbits are non-singular, ie. they do not hit corners of boundary and there are no grazing tangential collisions.
As a result, for each $(q,v)\in E_0$, there is an open neighborhood in $N$ whose elements return to $N$ in the same number of steps as $(q,v)$ and the return map on this set is smooth.
In a similar fashion, it follows that the map $X_x : \mathbb{T} \to \mathcal{X}$ is smooth on an open subset of $\mathbb{T}$ and its restriction to the set $W_x^j$ is likewise a diffeomorphism on an open set which consists of a countable union of open intervals $W_{x,i} \subset \mathbb{T}$. It is also the case that for dispersing billiards, e.g. those billiards for which $\partial Q$ consists of smooth convex curves with positive curvature, the restriction $X_{x,i}$ of $X_x$ to the set $W_{x,i}$ has the property that $X_{x,i}'\neq 0$. Moreover, the summation in (\ref{eq:kernel-1}) is well defined; see \cite[Lemma 5.56]{CM2006}.
We conclude the proof with a verification that the function $\omega$ defined in (\ref{eq:kernel-1}) is a kernel for $P_j$. Let $A \subset \mathcal{X}$ be a measurable set and let $A_{x,i} = \{r \in W_{x,i} : X_{x,i}(r) \in A\}$. Then
\begin{align*}
\int_{A} \omega(x, \phi)\,\pi_j(d\phi) &= \frac{1}{\alpha_j(x)} \sum_{i} \int_{A \cap V_{x,i}} \left(\frac{1}{2}\left|X_{x,i}'\left(X_{x,i}^{-1}(\phi)\right)\right|\sin\phi\right)^{-1}\, \pi(d\phi) \\
&= \frac{1}{\alpha_j(x)} \sum_{i} \int_{X_{x,i}(A_{x,i})} \left(\left|X_{x,i}'\left(X_{x,i}^{-1}(\phi)\right)\right|\right)^{-1}\, d\phi \\
&= \frac{1}{\alpha_j(x)} \sum_{i} \int_{A_{x,i}} dr \\
&= P_j \mathbbm{1}_{A}(x).
\end{align*}
Since this relation holds for indicator functions, it follows by a standard argument using linearity and the density of simple functions in $L^2(\mathcal X, \pi_j)$ that $P_j$ has kernel $\omega$ for all $f \in L^2(\mathcal{X},\pi_j)$.
\end{proof}
The next intermediary lemma gives an estimate on the kernel in Lemma \ref{lem:kernel-def}. Its proof follows from \cite[Lemmas 6.5, 6.6, 6.7]{Feres2007} with only minor modifications.
\begin{lem}
\label{lem:kernel-estimates}
Consider a billiard cell satisfying Assumption \ref{assu:partition} and let $\omega$ be the kernel given in (\ref{eq:kernel-1}). Then $\omega \in L^2(\mathcal{X}\times\mathcal{X}, \pi_j\otimes\pi_j)$.
\end{lem}
The following lemma is adapted from \cite[Theorem 9.9]{W1980}. It will be used to show that for an operator $P$ which admits a decomposition as in (\ref{eq:decomposition}), it suffices to show that one conditional operator is compact in order to prove that $P$ has spectral gap. The notation $\|\cdot\|$ is used to denote the canonical Hilbert space operator norm.
\begin{lem}\label{lem:gap}
Let $K$ and $T$ be bounded self-adjoint operators on a Hilbert space
and suppose that $K$ is compact. Then the essential spectrum of $T+K$
is contained in the essential spectrum of $T$. In particular, if
$\|T+K\| =1$ and $\|T\| <1$,
then the spectral gap of $T+K$ satisfies $\gamma(T+K)\geq\min\left\{ 1-\|T\| ,\gamma(K)\right\} $.
\end{lem}
We conclude with the proof of the section's main theorem.
\begin{proof}[Proof of Theorem \ref{thm:small-bumps-gap}]
That $P$ is self adjoint follows from Assumption \ref{assu:reversibility} and Proposition \ref{prop:P}. To see that $P$ has spectral gap, we apply Lemma \ref{lem:gap}. Using the notation of the lemma, we let $K = \alpha_jP_j$ and $T = \sum_{i \neq j}\alpha_iP_i$. Then, applying Lemmas \ref{lem:kernel-def} and \ref{lem:kernel-estimates}, we have that $K$ is a Hilbert-Schmidt integral operator and hence it is compact. It is clear that $T$ is bounded and self-adjoint and $\|T+K\| = \|P\| = 1$, where $\|\cdot\|$ is the $L^2$-operator norm. Moreover, $1 - \|T\| \geq \inf_{v\in\mathcal{X}} \alpha_j(v) > 0$. It follows that the spectral gap of $P$ is strictly positive. The concluding statement of exponential convergence to the stationary measure in total variation then follows immediately using the Cauchy-Schwarz inequality since $\|\mu\|_v \leq 1/2\|\mu\|_{\pi}$.
\end{proof}
\section{Diffusivity\label{sec:Diffusivity}}
Let $f:\mathcal{X}\to\mathbb{R}$ be a function on the state space
of the random billiard Markov chain $(X_{n})_{n\geq0}$ with Markov
transition operator $P$. We refer to $f$ as an \emph{observable}
(or \emph{functional})\emph{ }of the Markov chain. Without loss of
generality, we suppose that it has mean zero with respect to the stationary
distribution: $\pi(f)=0$. Our focus in this section will be on the
limiting distribution (after appropriate scaling) of partial sums
of the functional of the Markov chain given by
\[
S_{n}(f):=\sum_{k=0}^{n-1}f(X_{k}).
\]
It is well known that under appropriate mixing conditions for the
Markov chain, $S_{n}(f)/\sqrt{n}$ converges in distribution to a
centered Gaussian distribution with variance parameter $\sigma_{f}^{2}$.
As a preliminary result, we show that random billiard Markov chains
with microstructure have sufficiently fast mixing for a (central)
limit theorem of this kind to hold. However, our primary focus will
be to show that the variance $\sigma_{f}^{2}$ of the limiting Gaussian
distribution, which we refer to as the \emph{diffusivity} of the system,
can be rigorously approximated, and formulas can be derived in terms
of geometric parameters for families of random billiard microstructures.
We use here a result adapted from \cite{KV1986}, which states that
the central limit theorem holds for reversible Markov chains satisfying
a nondegeneracy condition on $\sigma_{f}^{2}$.
\begin{thm}
\label{thm:KV1986}Let $(X_{n})_{n\geq0}$ be a Markov chain with
stationary measure $\pi$ and let $f\in L_0^2(\mathcal X, \pi)$. If the
Markov chain is reversible, then $S_{n}(f)/\sqrt{n}$ converges in
distrubtion to a centered Gaussian random variable $\mathcal{N}(0,\sigma_{f}^{2})$
as long as $\sigma_{f}^{2}<\infty$, where $\sigma_f^2$ is given by \eqref{eq:variance-correlations}.
\end{thm}
In the discussion that follows, it will be useful to express $\sigma_{f}^{2}$
in terms of the spectrum of $P$, viewed as an operator on $L^{2}(\mathcal{X},\pi)$.
We first note that since $P$ is a bounded, self-adjoint operator
on $L^{2}(\mathcal{X},\pi)$ with norm 1, there exists a projection-valued
measure $\Pi$, supported on the spectrum $\sigma(P)\subset[-1,1]$
of $P$, defined so that
\[
P=\int_{-1}^{1}\lambda\,\Pi(d\lambda).
\]
For each $f\in L_{0}^{2}(\mathcal{X},\pi)$, we further define a measure
$\Pi_{f}$ supported on $\sigma(P)\setminus\left\{ 1\right\} $ by
$\Pi_{f}(d\lambda):=\left\langle f,\Pi(d\lambda)f\right\rangle _{\pi}.$
Now, observe that
\begin{align}
\sigma_{f}^{2} & =\left\langle f,f\right\rangle _{\pi}+2\sum_{k=1}^{\infty}\left\langle f,P^{k}f\right\rangle _{\pi}\label{eq:variance-correlations} \\
& =\left\langle f,f\right\rangle _{\pi}+2\left\langle f,P(I-P)^{-1}f\right\rangle _{\pi}\label{eq:variance-Poisson}\\
& =\int_{-1}^{1}\frac{1+\lambda}{1-\lambda}\Pi_{f}(d\lambda).\label{eq:variance}
\end{align}
Using the expression in (\ref{eq:variance}), we show that the existence
of a positive spectral gap is sufficient for the central limit theorem
to hold.
\begin{cor}
\label{cor:KV1986-RR1997}Let $(X_{n})_{n\geq0}$ be a Markov chain
with Markov transition operator $P$ and stationary measure $\pi$.
Let $f \in L_0^2(\mathcal X, \pi)$. If the Markov chain is reversible and $P$
has spectral gap $\gamma>0$, then $S_{n}(f)/\sqrt{n}$ converges
in distrubtion to a centered Gaussian random variable $\mathcal{N}(0,\sigma_{f}^{2})$.
\end{cor}
\begin{proof}
Since $P$ has spectral gap, there exists $0<\rho<1$ such that for
every $\lambda\in\text{supp\,}(\Pi_{f})$, $\lambda\leq\rho$. Therefore,
$\sigma_{f}^{2}$, as given by (\ref{eq:variance}), is finite
since $\sigma_{f}^{2}\leq(1+\rho)/(1-\rho)\pi(f^{2})<\infty.$
\end{proof}
\subsection{Diffusion approximation and diffusivity\label{subsec:Diffusion-approximation}}
We now prove Theorem \ref{thm:diffusion-app}.
Recall that the boundary of the billiard cell is assumed to be the graph of a periodic function $F:\mathbb{T} \rightarrow \mathbb{R}$.
Also recall the definitions of $h$ and $\mathcal{L}$ from Subsection 2.4.
\begin{proof}[Proof of Theorem \ref{thm:diffusion-app}]
When only a single boundary surface collision occurs, the relationship
between the initial and return velocity vectors, $v=(x, v_0)$ and $V(r,v)$ respectively, is straightforward. Indeed, let $\mathbbm n:\mathbb T \to \mathbb R^2$ denote the vector field of normal vectors along the boundary of the billiard cell and let $\bar{\mathbbm n}$ and $\mathbbm n_0$ denote the first (horizontal) and second (vertical) components of $\mathbbm n$. If collision with the boundary surface occurs at the point $(r', F(r'))$,
then $V(r,v)=v-2\left\langle v,\mathbbm n(r')\right\rangle \mathbbm n(r')$, where $\langle \cdot, \cdot \rangle$ denotes the Euclidean inner product. Note
that by elementary geometry
\[
\mathbbm n(r')=\frac{1}{\sqrt{1+F'(r')^{2}}}\left(-F'(r'),1\right)^\intercal,\quad r=r'-\left(F(r')-c\right)x/v_{0}.
\]
It now follows that for any smooth function $f:(-1,1)\to\mathbb{R}$
\begin{align*}
Pf(x) & =\int_{\mathbb{T}}f\left(x-2\left\langle v,\mathbbm n(r')\right\rangle \bar{\mathbbm n}(r')\right)\,dr\\
& =\int_{\mathbb{T}}f(x-2(\alpha+\beta)\bar{\mathbbm n}(r'))\left(1+\alpha/\beta\right)\,dr',
\end{align*}
where $\alpha=\bar{\mathbbm n}x,\beta=\mathbbm n_{0}v_{0}.$ Moreover, by Assumption \ref{assu:reversibility}, the symmetry relations
$\bar{\mathbbm n}(\ell-r)=-\bar{\mathbbm n}(r)$ and $\mathbbm n_{0}(\ell-r)=\mathbbm n_{0}(r)$ hold. Using these relations, and suppressing the explicit dependence of $\bar{\mathbbm n}$ on $r'$ for the sake of simplicity of notation, and we get
\[
Pf(x)=\frac{1}{2}\int_{\mathbb{T}}\left[f(x-2(\alpha+\beta)\bar{\mathbbm n})\left(1+\alpha/\beta\right)+f(x+2(-\alpha+\beta)\bar{\mathbbm n})\left(1-\alpha/\beta\right)\right]\,dr'.
\]
From here we use the second order Taylor approximation of $\phi$
centered about $x$. Observe that for $w\in(-1,1)$
\[
f(x+w)=f(x)+f'(x)w+\frac{f''(x)}{2}w^{2}+R_{x}(w),
\]
where $R$ is the usual Taylor remain term $R_{x}(w)=f'''(c)w^{3}/3!$
for some $c$ in the interval between $x$ and $w$. Using this,
together with straightforward algebraic manipulation that we omit
for the sake of clarity of exposition, we get that
\begin{align*}
Pf(x) & =f(x)-4xf'(x)\int_{\mathbb{T}}\bar{\mathbbm n}^{2}\,dr'+f''(x)\int_{\mathbb{T}}\bar{\mathbbm n}^{2}(6\alpha^{2}+2\beta^{2})\,dr'+E(x)\\
& =f(x)-4xf'(x)h+2\left(1-x^{2}\right)f''(x)h+O(h^{2})+E(x)\\
& =f(x)+2h\frac{d}{dx}\left(\left(1-x^{2}\right)f'(x)\right)+O(h^{2})+E(x),
\end{align*}
where $h=\int_{\mathbb{T}}\bar{\mathbbm n}^{2}\,dr'$ and $E$ is an error term. The error term
arises from the remainder $R$ and is bounded as follows: $|E|\leq C_{\phi}p(x, v_0)I_{3}$,
where $C_{\phi}$ is a constant that depends only on the third derivative
of $\phi$, $p(x,v_{0})$ is a polynomial in $x,v_{0}$
of degree at most 3 with coefficients that do not depend on $\phi$,
and $I_{3}:=\int_{\mathbb{T}}\bar{\mathbbm n}^{3}\,dr'$.
\end{proof}
\subsection{Computing the diffusivity}
The differential operator $\mathcal{L}$ defined in (\ref{eq:legendre}) has a well understood spectral
theory. We will take advantage of this in the following to give a method for computing $\sigma_f^2$.
Before going on, we first note a few well known facts about $\mathcal{L}$.
\begin{prop}
Let \emph{$\mathcal{L}$ be the Legendre differential operator defined
in (\ref{eq:legendre}). The following properties
hold.}
\begin{enumerate}
\item The eigenvalue problem $\mathcal{L}f=\lambda f$ has solutions
if and only if $\lambda$ is of the form $\lambda=-l(l+1)$ for integers
$l\geq0$.
\item The solutions of the eigenvalue problem are the polynomials $\phi_{l}$,
$l\geq0$, known as the Legendre polynomials. The first few are given
by $\phi_{0}=1,\phi_{1}(x)=x,\phi_{2}(x)=(3x^{2}-1)/2$.
\item The collection $(\phi_{l})_{l\geq0}$ of Legendre polynomials form
a complete orthogonal basis for $L^{2}(\mathcal X, \pi)$ and
\[
\langle\phi_{n},\phi_{m}\rangle:=\int_\mathcal X\phi_{n}(x)\phi_{m}(x)\,\pi(dx)=\frac{1}{2n+1}\delta_{n,m},
\]
where $\delta_{n,m}$ is the Kronecker delta symbol.
\end{enumerate}
\end{prop}
We are now ready to discuss the diffusivity $\sigma_{f}^{2}$ introduced
at the start of the section. The idea will be to use the diffusion
approximation $\mathcal{L}$ of the Markov-Laplacian $L$ in order
to give an approximation of the function $g=(I-P)^{-1}f$ that arises
in (\ref{eq:variance-Poisson}). Note that $g$ is a solution
of the Markov-Poisson equation $Lg=-f$. We first show that a series
solution of the classical Poisson equation can be given explicitly
in terms of Legendre polynomials.
\begin{lem}
\label{lem:series-solution}For any $f\in L_{0}^{2}(\mathcal X, \pi)$, the equation
$\mathcal{L}g=-f$ has solution given by
\[
g=\sum_{l=1}^{\infty}a_{l}\phi_{l},\quad a_{l}=\frac{2l+1}{2l(l+1)}\left\langle \phi_{l},f\right\rangle_\pi .
\]
\end{lem}
\begin{proof}
Let $f\in L_{0}^{2}(\mathcal X, \pi)$. Since the Legendre functions form a complete
orthogonal basis for $L_{0}^{2}(\mathcal{X},\pi)$, $f=\sum_{l=1}^{\infty}b_{l}\phi_{l}$,
where $b_{l}=(2l+1)\left\langle \phi_{l},f\right\rangle _{\pi}$.
Now, let $g=\sum_{l=1}^{\infty}a_{l}\phi_{l}$, where $a_{l}=b_{l}/(l(l+1))$.
Observe that
$\mathcal{L}g =\sum_{l=1}^{\infty}a_{l}\mathcal{L}\phi_{l}=-\sum_{l=1}^{\infty}a_{l}l(l+1)\phi_{l}=-f.
$
\end{proof}
With the lemma in hand, we now give our main approximation result.
The idea of the proof will be to contruct a series solution approximation
of the Markov-Poisson equation using the series solution of the Poisson
equation along with the diffusion approximation of $P$. We use the
estimates in Theorem \ref{thm:diffusion-app} to control
the error terms in our approximation.
\begin{thm}
Let $(P_{h})_{h>0}$ be a family of random billiard Markov transition
operators for a family of microscopic billiard cells satisfying Assumptions
\ref{assu:reversibility} and \ref{assu:partition}. For any function $f\in L_{0}^{2}(\mathcal X, \pi)$,
let $\sigma_{f,h}^{2}$ denote the diffusivity corresponding to $P_{h}$.
Then
\begin{equation}
\sigma_{f,h}^{2}=-\langle f,f\rangle_\pi+\frac{1}{h}\sum_{l=1}^{\infty}\frac{2l+1}{l(l+1)}\left\langle \phi_{l},f\right\rangle_\pi ^{2}+O(h^{1/2}).\label{eq:variance-h-approx}
\end{equation}
\end{thm}
\begin{proof}
Let $h>0$ and let $g_{h}$ be the solution of the Poisson equation
$\mathcal{L}g=-f/(2h)$. Note that by Lemma \ref{lem:series-solution},
$g_{h}=\sum_{l=1}^{\infty}a_{l,h}\phi_{l}$ where
\[
a_{l,h}=\frac{2l+1}{2hl(l+1)}\left\langle \phi_{l},f\right\rangle_\pi .
\]
By Theorem \ref{thm:diffusion-app}, $Lg_{h}=2h\mathcal{L}g_{h}+O(h^{1/2})=-f+O(h^{1/2}).$
Note that the error in the above expression is of lower order than
that in the theorem because the right hand side in the Poisson equation
contains a factor of $h^{-1}$. Next observe that
\begin{align*}
\left\langle Pf,g_{h}\right\rangle_\pi & =\left\langle f,Pg_{h}\right\rangle_\pi \\
& =\left\langle f,g_{h}\right\rangle_\pi +2h\left\langle f,\mathcal{L}g_{h}\right\rangle_\pi +O(h^{1/2})\\
& =\frac{1}{2h}\sum_{l=1}^{\infty}\frac{2l+1}{l(l+1)}\left\langle \phi_{l},f\right\rangle_\pi ^{2}-\left\langle f,f\right\rangle_\pi +O(h^{1/2}).
\end{align*}
Using the expression above, along with the formula for $\sigma_{f,h}^{2}$
given in (\ref{eq:variance-Poisson}), the result then follows.
\end{proof}
\section{Analysis of the Galerkin method\label{sec:Numerical-techniques}}
In this section, we conclude with an analysis of the Galerkin method introduced in Subsection \ref{sec:Numerical-techniques-examples}, including a proof of Theorem \ref{thm:galerkin}.
We begin with a result on the decay rates of Legendre series truncation which will be useful. It is taken from Theorem 2.2 from \cite{wang2018new}, restated slightly here to fit our notation and context.
We will give estimates in terms of the following weighted semi-norm, defined on the space of functions $u : (-1,1) \to \mathbb R$ such that the following integral is defined:
$$ \left\Vert u\right\Vert_w = \int_{-1}^{1}\frac{|u'(x)|}{(1-x^2)^{\frac{1}{4}}} \,dx.$$
\begin{thm}[Adapted from Theorem 2.2 in \cite{wang2018new}]\label{thm:legendre_decay}
Let $m \geq 1$, and let $u : (-1,1) \to \mathbb R$ be a function such that $u, u',\ldots, u^{(m-1)}$ are absolutely continuous and the $m$-th derivative $ u^{(m)}$ is of bounded variation. Furthermore, assume that $\left\Vert u^{(m)} \right\Vert_w < \infty$. Let $a_n = (2n+1) \langle u, \phi_n \rangle$ be the sequence of coefficients in the Legendre expansion of $u$ such that $u(x) = \sum_{n = 0}^\infty a_n \phi_n(x)$. Then, for $n \geq m+1$,
$$|a_n| \leq \frac{\left\Vert u^{(m)}\right\Vert_w}{\sqrt{\pi(2n - 2m - 1)}} \prod_{k=1}^{m} \frac{2}{2n - 2k + 1}.$$
\end{thm}
We are now ready to prove Theorem \ref{thm:galerkin}. We begin by recalling some notation introduced in Subsection \ref{sec:Numerical-techniques-examples}. Let $T_n:L_0^2(\mathcal{X},\pi) \to R_n$ denote the orthogonal projection to the linear span $R_n=\{\phi_1, \dots, \phi_n\}$ of the first $n$ non-constant Legendre polynomials. The solution of the finite dimensional linear system $ (I-T_nP)x = T_nf$, where $f \in L_0^2(\mathcal X, \pi)$ is the given observable, will be written as $g_n$. Note that $g_n \in R_n$, and writing $g_n=\sum_{j=1}^n \alpha_j \phi_j$, it is straightforward to see that we aim to find a solution $x$ to the system $Gx = y$ where
$$x=(\alpha_1, \dots, \alpha_n)^\intercal, \ \ y= (\langle f, \phi_1\rangle_\pi, \dots, \langle f, \phi_n\rangle_\pi)^\intercal, \ \ G=\left(\langle \phi_j,\phi_i\rangle_\pi - \langle P\phi_j, \phi_i\rangle_\pi\right)_{i,j=1}^n.$$
The solution $g_n$ will in turn be used to give the approximation $\sigma^2_{\mathrm{GM},n} := \langle T_n f, T_n f\rangle + 2\langle T_nPf, g_n\rangle$ of the diffusivity $\sigma_f^2$.
\begin{proof}[Proof of Theorem \ref{thm:galerkin}]
Throughout the proof, we take $P$ to be the restriction of the Markov operator
to the space $L_0^2(\mathcal X, \pi)$ so that $\|P\| < 1$, where $\|\cdot\|$ denotes $L^2$-operator norm.
Observe that from the definition of $\sigma^2_{\mathrm{GM},n}$,
\begin{equation}\label{eq:gm-diff}
\left|\sigma_f^2 - \sigma^2_{\mathrm{GM},n}\right| \leq \left \langle f, f\rangle_\pi - \langle T_nf, T_nf\rangle_\pi \right|
+ 2\left| \langle Pf, g \rangle_\pi - \langle T_n Pf, g_n \rangle_\pi \right|.
\end{equation}
Our aim is to show that the two terms of the right hand side above are bounded by a common factor in terms of $n$,
which we will then show decays as in the statement of the theorem.
For the first term on the right hand side of (\ref{eq:gm-diff}), we see that
\begin{align}
\inner{f}{f}_{\pi} - \inner{T_nf}{T_nf}_{\pi}
& = \|f\|_\pi^2 - \|T_nf\|_\pi^2 \nonumber \\
& \leq 2\|f\|_\pi (\|f\|_\pi -\|T_nf\|_\pi ) \nonumber\\
& \leq 2\|f\|_\pi \|f - T_nf\|_\pi \label{eq:term1},
\end{align}
and for the second term,
\begin{align*}
\left|\inner{Pf}{g}_{\pi}-\inner{T_nPf}{g_n}_{\pi}\right|
&= \left|\int_\mathcal{X} [Pf(x)g(x) - T_nPf(x)g_n(x) ]\,\pi(dx)\right|\\
&\leq \int_\mathcal{X} |Pf(x)(g(x)-g_n(x))|\,\pi(dx) + \int_\mathcal{X}|g_n(x)(Pf(x) -T_nPf(x))|\,\pi(dx)\\
&\leq \|Pf\|_\pi\|g-g_n\|_\pi + \|g_n\|_\pi \|Pf-T_nPf\|_\pi\\
&\leq \|Pf\|_\pi \|(I-T_nP)^{-1} \| \|g-T_n g \|_\pi + \|g_n\|_\pi \|Pf-T_nPf\|_\pi,
\end{align*}
where in the last step we have used the fact that $g - g_n = (I-T_nP)^{-1}(g - T_n g)$.
It is straightforward to see that $T_n P = PT_n$.
Indeed, for any function $f \in L_0^2 (\mathcal X, \pi)$, with Legendre expansion given by $f = \sum_{k=1}^\infty a_k \phi_k$, we have that $Pf(x) = \sum_{k = 1}^\infty a_k P\phi(x).$
Applying $T_n$ to both sides of this equation, we get that
$T_nPf(x) = \sum_{k=1}^n a_k P\phi(x) = PT_nf(x).$
It now follows that $$\|Pf-T_nPf\|_\pi = \|Pf-PT_nf\|_\pi \leq \|P\| \|f-T_nf\|_\pi.$$
Moreover, a similar argument gives that
$$\|g-T_ng\|_\pi \leq \|(I-P)^{-1}\| \|f-T_nf\|_\pi \leq (1-\|P\|)^{-1}\|f-T_nf\|_\pi.$$
Finally, we note that
$$\|(I-T_nP)^{-1} \|
\leq (1-\|T_nP\|)^{-1}
\leq (1-\| P\|)^{-1},$$
and consequently,
$$g_n = (I-T_nP)^{-1}T_nf \leq (1-\|P\|)^{-1}\|f\|_\pi.$$
We then have for the second term on the right hand side of (\ref{eq:gm-diff}),
\begin{equation}\label{eq:term2}
\left|\inner{Pf}{g}_{\pi}-\inner{T_nPf}{g_n}_{\pi}\right| \leq
\|P\|(2-\|P\|)(1-\|P\|)^{-2}\|f\|_\pi\|f-T_nf\|_\pi.
\end{equation}
Applying the estimates in (\ref{eq:term1}) and (\ref{eq:term2}) to (\ref{eq:gm-diff}) and simplifying, we have that
\begin{equation}\label{eq:last-estimate}
\left|\sigma_f^2 - \sigma^2_{\mathrm{GM},n}\right| \leq 2\|f\|_\pi(1-\|P\|)^{-2}\|f - T_n f\|_\pi.
\end{equation}
It is now evident that the convergence rate will depend on the decay rate of $f$ with it's Legendre series truncation.
Observe that
$$\|f-T_nf\|_\pi \leq \sum_{k=n+1}^\infty a_k \|\phi_k\|_\pi = \sum_{k=n+1}^\infty (2k+1)^{-1/2}a_k,$$
where $a_k = (2k+1)|\langle f, \phi_k\rangle_\pi|$. Using Theorem \ref{thm:legendre_decay} with $m=1$ we get
\begin{equation*}
|a_k| \leq \frac{2\|f'\|_{w}}{\sqrt{\pi}(2k-1)\sqrt{2k-3}}.
\end{equation*}
Thus
$$\|f-T_nf\|_\pi
\leq \frac{2s_n\| f'\|_{\mathrm{w}}}{\sqrt{\pi}},$$
where
$$s_n:=\sum_{k=n+1}^{\infty} \frac{1}{(2k-1)\sqrt{2k+1}\sqrt{2k-3}}.$$
Further, for $n>2$ we have
\begin{align*}
s_n &< \sum_{i=n+1}^{\infty} \frac{1}{(2i-3)^{2}} \\
&\leq \int_n^{\infty} \frac{dx}{(2x-3)^{2}}\\
&=\frac{1}{4n-6}.
\end{align*}
The result now follows by applying these estimates to (\ref{eq:last-estimate}).
\end{proof}
| 726087a7c3ac0d31c07aed214fc82886753e2734 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Over past decades, much attentions focus on multitemporal Synthetic Aperture Radar (SAR) image change detection, since a SAR is capable of working in all-time and all-weather without the influence of extremely bad weather and the cloud.
In a past decade, most traditional SAR image change detection methods are developed how to extract changed areas from a difference image (DI), which suppose to include the information of changed regions. The DI calculated by the log-ratio (LR) \cite{bazi2006automatic} is usually subject to the speckle and it is challenging to extract the accurate and clear information on the changed region. To tackle this issue, sparse learning\cite{wang2016sar} was recently proposed learning robust features from the noisy DI. Wang et al.\cite{wang2019can} analyzed the affects of the SAR image speckle on change detection and proposed a sparse learning-based method for SAR image change detection.
Recently, deep neural networks have been successfully employed to computer vision and remote sensing image analysis due to its ability to exploiting essential and robust structural features on categories of objects. It also has been introduced into the field of change detection. Gong et al\cite{Gong2017Change} proposed a deep neural network for SAR image change detection. Gao et al. \cite{gao2016automatic} proposed a simple convolutional neural network, known as PCA-Net, exploring robust features on the changed regions from the noisy DI. However, the performance of these two unsupervised methods are limited without the correct guidance. To tackle this issue, Wang et al. \cite{wang2018imbalanced} proposed a supervised PCA-Net to improve the performance by carefully collecting typical training samples, which obtains state-of-arts performance of bitemporal SAR image change detection. However, this two-layer convolutional neural network is low efficient, since the convolutional kernels are trained or generated by a Principle Component Analysis (PCA) decomposition. Recently, Li et al. \cite{li2019deep} proposed a convolutional neural network (CNN) for SAR image change detection based on both unsupervised and supervised learning. Zhao et al. \cite{zhao2017novel} proposed a bitemporal PolSAR image change detection by a joint classification and a similarity measurement. Currently, it is still an open problem to extract the changed regions from the noisy DI.
{Nowadays, a large volume of SAR images are acquired by satellites and it is imperative to develop an efficient model that can produce promising results of SAR image interpretation. Most above networks are too heavy and cost much computational burden. It is strongly required to develop a lightweight convolutional neural network.}
{Recently, several lite networks are proposed to improve the inference efficiency. Howard et al. and Sandler et al. proposed two lite networks MobileNetV1 \cite{howard2017mobilenets} and MobileNetV2 \cite{sandler2018mobilenetv2} for visual category. Recently, Howard et al. proposed to search for MobileNetV3 \cite{howard2019searching}. Tan et al. \cite{tan2019efficientnet} proposed an efficient network for visual category. These lite networks have been extensively employed to visual category and the experimental results show that they can achieve comparable performance with heavy networks, but with low latency and network capacity. It can be potentially performed on edge devices with low power. Most recently, Chen et al.\cite{chen2020a} proposed a lightweight multiscale spatial pooling network for bitemporal SAR image change detection.
Following the idea of the lightweight neural network, in this letter, we focus on the application of lite networks in SAR image change detection. To achieve this, we propose a lite CNN for SAR image change detection. In the proposed network, bottleneck layers are introduced to reduce the number of output channel. Furthermore, the dilated convolutional layers\cite{li2018csrnet} are introduced to enlarge receptive field with a few of non-zero entries in the kernel, which reduces the number of network parameters. We verify the proposed network by comparing other conventional CNN. Compared with the lightweight network in \cite{chen2020a}, the proposed network is more robust with the residual and bottleneck structure. Experimental results on four sets of bitemporal SAR image show that our proposed method obtain comparable performance with CNN, while being much more efficient than CNNs. }
The rest of paper will be organized as follows. We will introduce our proposed method in Section 2. Then the proposed method will be verified on four datasets in Section 3. Finally, we will draw a conclusion in Section 4.
\section{Proposed Method}
{Given bitemporal SAR images ${\bf I}_1$ and ${\bf I}_2$, the DI can be generated as follows}
\begin{equation}\label{di}
{\bf I}_{DI} = {\bf I}_1 \ominus {\bf I}_2
\end{equation}
{where $\ominus$ denote the difference operator. However, most existing difference operator is subject to the speckle.
Then we will propose a lightweight convolutional neural network to exploit the changed regions from the noisy DI.}
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{fig1.png}
\caption{The framework of the proposed network.(a)Network Architecture. (b) Bottleneck for encoder. (c)Bottleneck for decoder.}
\label{framework}
\end{figure}
\begin{table}[!htbp]
\centering
\caption{The tensors of all the layers.}
\label{tab:network}
\begin{tabular}{cccccccc}
\toprule
Layer Name & Tensor Size & & Layer Name & Tensor Size & & Layer Name & Tensor Size \\
\hline
\multicolumn{8}{c}{Initial Block} \\
Input& (32,32,1) \\
Conv&(16,16,13) & &
Max-pooling & (16,16,1) &&
Concatenation & (16,16,14) \\
\hline
\multicolumn{8}{c}{Group 1} \\
BottleNeck 1.0&(8,8,64) & &
BottleNeck 1.1&(8,8,64) & &
BottleNeck 1.2&(8,8,64)\\
BottleNeck 1.3&(8,8,64) & &
BottleNeck 1.4&(8,8,64) \\
\hline
\multicolumn{8}{c}{Group 2} \\
BottleNeck 2.0&(4,4,128) & &
BottleNeck 2.1&(4,4,128) & &
BottleNeck 2.2&(4,4,128)\\
BottleNeck 2.3&(4,4,128) & &
BottleNeck 2.4&(4,4,128) & &
BottleNeck 2.5&(4,4,128) \\
BottleNeck 2.6&(4,4,128) & &
BottleNeck 2.7&(4,4,128)& &
BottleNeck 2.8&(4,4,128) \\
\hline
\multicolumn{8}{c}{Group 3} \\
BottleNeck 3.0&(4,4,128) & &
BottleNeck 3.1&(4,4,128) & &
BottleNeck 3.2&(4,4,128)\\
BottleNeck 3.3&(4,4,128) & &
BottleNeck 3.4&(4,4,128) & &
BottleNeck 3.5&(4,4,128)\\
BottleNeck 3.6&(4,4,128) & &
BottleNeck 3.7&(4,4,128)\\
\hline
\multicolumn{8}{c}{Group 4} \\
BottleNeck 4.0&(8,8,64) & &
BottleNeck 4.1&(8,8,64) & &
BottleNeck 4.2&(8,8,64)\\
\hline
\multicolumn{8}{c}{Group 5} \\
BottleNeck 5.0&(16,16,16) & &
BottleNeck 5.1&(16,16,16) \\
\hline
\multicolumn{8}{c}{Output} \\
Conv&(32,32,2) \\
\bottomrule
\end{tabular}
\end{table}
The whole framework of the proposed network can illustrated in Fig.\ref{framework}(a). It is shown that the network consists of five groups of bottleneck layers \cite{lin2013network} with an 1$\times$1 kernels, illustrated by variety of colorful bars, among which the former three ones work as the encoder, and the latter two ones as the decoder. {The tensors of all layers are listed in Table \ref{tab:network}.}
In the forward process, the network takes a patch of DI as the input. Firstly, the input data go through a normal convolutional layer and a max-pooling (MP) layer, respectively and then the outputs are concatenated. Next, the contact activations go through the decoder with three groups bottleneck layers. The essential structure of a bottleneck of encoder can be illustrated in Fig.\ref{framework}(b). It is shown that a bottleneck layer is constructed by a small residual block, including a maxpooling path and a convolutional path. More specifically, the convolutional path consists of two 1$\times$1 convolutions and one main convolution. The main convolution will vary with the various function of the bottleneck. It can be a normal convolution, a dilated convolution \cite{li2018csrnet} or an asymmetrical convolution\cite{szegedy2016rethinking}. {The tensors inside the encoding bottleneck are listed in Table \ref{tab:bo_en}.}
\begin{table}[!htbp]
\centering
\caption{The tensors inside an encoding bottleneck layer.}
\label{tab:bo_en}
\begin{tabular}{ccccc}
\toprule
Layer Name & Tensor Size & & Layer Name & Tensor Size \\
\hline
Input& (16,16,14) \\
\hline
\multicolumn{5}{c}{Branch 1} \\
Conv1& (8,8,16) &&
Conv2&(8,8,16) \\
Conv3&(8,8,64) &&
Dropout&(8,8,64) \\
\hline
\multicolumn{5}{c}{Branch 2} \\
Max-pooling&(8,8,14) & &
Padding&(8,8,64)\\
\multicolumn{5}{c}{Output} \\
\hline
Addition & (8,8,64)\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!htbp]
\centering
\caption{The tensors inside a decoding bottleneck layer.}
\label{tab:bo_de}
\begin{tabular}{ccccc}
\toprule
Layer Name & Tensor Size & & Layer Name & Tensor Size \\
\hline
Input& (4,4,128) \\
\hline
\multicolumn{5}{c}{Branch 1} \\
Conv1& (4,4,16) &&
Conv2&(8,8,16) \\
Conv3&(8,8,64) \\
\hline
\multicolumn{5}{c}{Branch 2} \\
Conv&(4,4,64) & &
Up-Sampling&(8,8,64)\\
\multicolumn{5}{c}{Output} \\
\hline
Addition & (8,8,64)\\
\bottomrule
\end{tabular}
\end{table}
The first group consist a down-sampling bottleneck and four normal convolutional bottleneck layers. In the normal bottleneck layers, the main convolution component is default set as the main normal convolutional layer. Especially, in the down-sampling bottleneck, the main convolutional component is set as a normal convolution kernel with 3$\times$3 and the 1$\times$1 convolution component is replaced by a 2$\times$2 one. In the next two groups, to exploit the spatial context, we insert the bottleneck layers with asymmetrical and dilated convolution layers with various kernel sizes among the normal bottleneck, where the kernel sizes are set 2, 4, 8 and 16, respectively, as shown the digits below the green bars. In these bottleneck layers, the main convolutional layers are replaced by the dilated convolutional layers, where the kernels are sparse and most entries are zeros. Furthermore, the bottleneck layers with asymmetric convolution layers are also insert between the dilated and normal convolutional layer, illustrated by the blue bars. After encoding, the sizes of feature maps decrease as the one fourth as the original image.
In the decoding part, the context information collected by the encoder will be propagated to the pixel level. To achieve this, inspired by the idea of U-Net \cite{ronneberger2015u}, we schedule a upsampling layer and two bottleneck layers. The essential structure of bottleneck for decoder can be illustrated in Fig.\ref{framework}(c). Similar to the bottleneck for the encoder, the bottleneck for the decoder contains a pooling path and a convolution path. The former includes a maxpooling layer and a 1$\times$1 convolution, while the latter includes two 1$\times$1 convolutional layers and a 3$\times$3 convolutional layer. Especially, when the bottleneck layer is used for upsampling, the 3$\times$3 convolution component, illustrated by the yellow bar, will be replaced by a 3$\times$3 transpose convolution \cite{dumoulin2016guide}. Then the bottleneck will do the 2x upsampling. Trough two groups of decoding bottleneck layers, the feature map will be recover the same size as the input image. {The tensors inside the decoding bottleneck are listed in Table \ref{tab:bo_de}.} Finally, we put a 2$\times$2 convolutional layer to get the probability map of two categories.
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{fig2.png}
\caption{Four sets of bitemporal SAR images. The first two rows are bitemporal images and the last row is the DIs. (a) YR-A. (b) YR-B. (c) Sendai-A. (d) Sendai-B. }
\label{fig2}
\end{figure}
\section{Experimental Results}
\label{sec:experiment}
\subsection{Experiment Datasets}
In this paper, the proposed method is verified on four sets of bitemporal SAR images. Two scenes (YR-A and YR-B) are from bitemporal Yellow River SAR images \cite{Gong2017Change} acquired by the Radarsat-2 satellite in 2008 and 2009, respectively. Their image sizes are 306 $\times$ 291 and 400 $\times$ 350, respectively. Other two are parts of TerraSAR-X images acquired prior to (on Oct. 20, 2010) and after (on May 6, 2011) the Sendai earthquake in Japan \cite{Cui2016A}. Their sizes (Sendai-A and Sendai-B) are 590 $\times$687 and 689 $\times$ 734, respectively. These four datasets are shown in Fig.\ref{fig2}). These four datasets are quite challenging, such as the linear-shape changed regions in YR-B dataset and complex scene in both Sendai-A and Sendai-B datasets.
\begin{table}[!htbp]
\centering
\caption{The number of training samples.}
\label{tab:nos}
\begin{tabular}{ccccc}
\toprule
Dataset & YR-A & YR-B & Sendai-A & Sendai-B \\
\hline
No. Samples& 3596 & 6205 & 15375 & 20294 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.6\textwidth]{fig3.png}
\caption{The variations of loss and accuracies.}
\label{fig:loss}
\end{figure}
\subsection{Implementations}
{
We have introduced a lite CNN for change detection for bitemporal SAR images. We first generate the DI by the Eq.(\ref{di}) and the difference operator is implemented by the neighborhood-based LR operator \cite{gong2011neighborhood}.
To train the network, we collect a training dataset according to the method in \cite{wang2018imbalanced}. The numbers of training samples for each dataset are listed in Table \ref{tab:nos}.
More specifically, the patchsize of each sample is set as 32 $\times$ 32 and 8 samples are fed in each training step. Additionally, the network is trained by an end-to-end back-propagation manner and the loss is set as {the binary entropy function defined in \cite{sadowski2016notes}}.
The training is optimized by the Adam algorithm \cite{kingma2014adam} in the training stage, where the initial learning rate is set as 0.005.
The training is performed on the PyTorch platform built on the Ubuntu 16.04 installed in a PC with a 16 GB DDR memory and an NVIDIA TITAN Xp Graphics Processing Unit of 11 GB memory. The training process will converge at around 15 epochs. We show the variations of loss values and accuracies of training process in Fig.\ref{fig:loss}.
}
\subsection{Comparison Experiments}
To verify the benefits of the proposed method, it is compared with the unsupervised PCA-Net (U-PCA-Net)\cite{gao2016automatic}, the supervised PCA-Net (S-PCA-Net) \cite{wang2018imbalanced} which achieves the state-of-arts performance on SAR image change detection. We also compare the proposed method with the deep neural network (DNN) method \cite{Gong2017Change} and CNN \cite{li2019deep}. Among these methods, DNN and U-PCA-Net are unsupervised methods,while S-PCA-Net, CNN and the proposed method are supervised ones.
The performance of the compared methods is evaluated by probabilistic Missed Alarm (pMA), probabilistic False Alarm (pFA) and kappa coefficient, where pFA (pMA) are calculated by the ratios between FA (MA) and the number of Non-Changed pixels (NC)\cite{wang2019can}.
\begin{figure}[!thb]
\centering
\includegraphics[width=\textwidth]{fig4.png}
\caption{The visual comparison results. (a)U-PCA-Net. (b)S-PCA-Net. (c)DNN. (d) CNN. (e)Lite CNN. (f)Reference. }
\label{fig3}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{fig5.png}
\caption{The quantitative evaluations of compared methods.(a) MA. (b)FA. (c) Kappa. }
\label{eva}
\end{figure}
\subsection{Experiment Results on Individual Dataset Change Detection}
In this experiment, for the supervised learning methods, we collect the training samples from an individual dataset, which covers 30\% areas of the whole image frame. The rest part is employed for testing.
The visual comparison results are shown in Fig.\ref{fig3}. It is shown that for the YR-A dataset, S-PCA-Net, DNN and Lite CNN get less noisy but more completed changed regions. Lite CNN can get more clear boundary of changed regions than DNN. For the YR-B dataset, Lite CNN can get more completed changed regions, especially the line at the bottom of the image. Other methods can not get the completed changed regions. For the Sendai-A dataset with complex scene, S-PCA-Net and the Lite CNN are less subject to the speckle and the background and get more clear changed regions, while other compared methods are subjected to the speckle and background and they are almost failed. Moreover, compared with S-PCA-Net, Lite CNN get better inner regional consistence. For the Sendai-B dataset, Lite CNN gets more accurate changed regions than other methods.
Moreover, we show the quantitative evaluations in Fig.\ref{eva}. {It is shown in Fig.\ref{eva} (a) that the proposed Lite CNN gets a lower pMA on YR-A, Sendai-A and Sendai-B dataset, while the CNN gets a lower pMA on YR-B dataset. It is shown in Fig.\ref{eva} (b) that the proposed Lite CNN gets a lower pFA on Sendai-B dataset, while S-PCA-Net gets a lower pFA on YR-A, YR-B and Sendai-A dataset. }
It is shown in Fig.\ref{eva} (c) that on the YR-A dataset, S-PCA-Net gets the best kappa among all the methods, while the Lite CNN gets the comparable kappas with other methods, except S-PCA-Net. On the YR-B dataset, Lite CNN gets the comparable kappas with other methods. However, on both Sendai-A and Sendai-B datasets, the Lite CNN performs better than other methods in terms of pMA and kappas.
\begin{figure*}[!hbt]
\centering
\includegraphics[width=\textwidth]{fig6.png}
\caption{The visual comparison results. (a) S-PCA-Net. (b) CNN. (c) Lite CNN. (d) Reference. }
\label{fig4}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=\textwidth]{fig7.png}
\caption{The quantitative evaluations of compared methods.(a) MA. (b)FA. (c) Kappa.}
\label{eva2}
\end{figure*}
\subsection{Experiment Results on Cross-dataset Change Detection}
To further compare the proposed method with other supervised learning methods, S-PCA-Net and CNN, we perform the comparisons on the cross-dataset change detection, where the network trained on several datasets is applied to an unknown testing dataset. More specifically, to achieve this, this experiment is conducted through the leave-one-out manner, i.e. each dataset alternative is selected as the testing dataset and others as training datasets.
The visual comparisons are shown in Fig.\ref{fig4}. It is shown that on the YR-A dataset Lite CNN gets more clear visual result with less noisy spots. On the YR-B dataset, Lite CNN performs better than CNN, but not better than S-PCA-Net. There is many miss alarms in the results of Lite CNN. On both Sendai-A and Sendai-B datasets, Lite CNN gets better results than other two methods.
{The quantitative evaluations in terms of pFA, pMA and Kappa are shown in Fig.\ref{eva2}. It is shown in Fig.\ref{eva2}(a) that the S-PCA-Net gets a lower pMA on all datasets. It is shown in Fig.\ref{eva2}(b) that the Lite CNN gets a lower pFA on the Sendai-A dataset, while S-PCA-Net gets a lower pFA on the YR-A dataset and CNN gets a lower pFA on the YR-B and Sendai-B dataset.} It is shown in Fig.\ref{eva2}(c) that Lite CNN performs better than CNN but comparable with S-PCA-Net on YR-A and YR-B datasets. However, Lite CNN shows great advantages over other two methods on Sendai-A and Sendai-B datasets. It indicates that the Lite CNN performs better than other two methods in model generalization, especially on challenging datasets with complex scenes.
\begin{table}[!htbp]
\centering
\caption{The training times of compared methods.}
\label{time}
\begin{tabular}{lccc}
\toprule
Methods & S-PCA-Net & CNN& Lite CNN \\
\hline
Times& ~3 h & 30 mins. &15 mins. \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Discussion}
From the above comparisons, it is shown that Lite CNN can obtain comparable performance with other method on the YR-A and YR-B datasets. On the challenging datasets, e.g. Sendai-A and Sendai-B, Lite CNN outperforms other methods. Moreover, it has better ability to model generalization. Moreover, Lite CNN is more computationally efficient than S-PCA-Net and CNN. The training times of three supervised learning methods are compared in Table \ref{time}. It is shown that Lite CNN is easy to train and take less time than S-PCA-Net and CNN, while S-PCA-Net takes longer time, since the convolutional kernel is generated by the principle component analysis decomposition.
Overall, Lite CNN can obtain comparable or even better performance than S-PCA-Net and CNN. It is also more computationally efficient than other two methods. It is expected that Lite CNN is more practical in change detection, especially for the requirement of real-time detection.
\section{Conclusion}
\label{sec:conlusion}
In this paper, we develop a lightweight convolutional neural network for bitemporal SAR image change detection. The proposed network consists of groups of bottlenecks layers which exploit the image feature. To verify the benefits of our proposed method, we compare it with several traditional neural networks and the comparisons are performed on fours sets of bitemporal SAR images. The experimental results show that our proposed method Lite CNN performs better than other two methods on cross-dataset change detection, especially when the scene is complex. Furthermore, Lite CNN is a lightweight network, which is more computationally efficient than CNN and S-PCA-Net. In the future, we will further optimize Lite CNN and make it more efficient on the edge device.
\section*{Acknowledgment}
This work was supported by the State Key Program of National Natural Science of China (No. 61836009), the National Natural Science Foundation of China (No. 61701361, 61806154), the Major Research Plan of National Natural Science Foundation of China (No. 91838303), the Open Fund of Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education, Xidian University (Grant No. IPIU2019006), the Natural Science Basic Research Plan in Shaanxi Province of China (No.2018JM6083) and the Open Fund of State Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University (No. 17E02).
\bibliographystyle{spiejour}
| 0a1668e026078a83dd803598842a0e80fef4e02a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
In~\cite{DBLP:journals/Prawitz15} Dag Prawitz proposed a natural deduction system for what was later called {\em Ecumenical logic} (EL), where classical and intuitionistic logic could coexist in peace.
In this system, the classical logician and the intuitionistic logician would share the universal quantifier, conjunction, negation, and the constant for the absurd ({\em the neutral connectives}), but they would each have their own existential quantifier, disjunction, and implication, with different meanings.
Prawitz' main idea is that these different meanings are given by a semantical framework that can be accepted by both parties. While proof-theoretical
aspects were also considered, his work was more focused on investigating the philosophical significance of the fact that classical logic can be translated into intuitionistic logic.
Pursuing the idea of having a better understanding of Ecumenical systems under the proof-theoretical point of view, in~\cite{Syn20} an Ecumenical sequent calculus ($\LEci$) was proposed.
This enabled not only the proof of some important proof theoretical properties (such as cut-elimination and invertibility of rules), but it also provided a better understanding of the Ecumenical nature of consequence: it is intrinsically intuitionistic, being classical only in the presence of classical succedents.
Also in~\cite{Syn20} a Kripke style semantics for EL was provided, and it was shown how worlds in the semantics could be adequately interpreted as nestings in nested sequents.
This also enabled the proposal of an Ecumenical multi-succedent sequent system.
In this work, we propose an extension of EL with the alethic modalities of {\em necessity} and {\em possibility}. There are many choices to be made and many relevant questions to be asked, \eg: what is the ecumenical interpretation of Ecumenical modalities? Should we add classical, intuitionistic, or neutral versions for modal connectives? What is really behind the difference between the classical and intuitionistic notions of truth?
We propose an answer for these questions in the light of Simpson's meta-logical interpretation of modalities~\cite{Sim94} by embedding the expected semantical behavior of the modal operator into the Ecumenical first order logic.
We start by highlighting the main proof theoretical aspects of $\LEci$ (Section~\ref{sec:LEci}). This is vital for understanding how the embedding mentioned above will mold the behavior of Ecumenical modalities, since modal connectives are interpreted in first order logics using quantifiers. In Sections~\ref{sec:modal} and~\ref{sec:emodal}, we justify our choices by following closely Simpson's script, with the difference that we prove meta-logical soundness and completeness using proof theoretical methods only. We then provide an axiomatic and semantical interpretation of Ecumenical modalities in Sections~\ref{sec:ax} and~\ref{sec:bi}. This makes it possible to extend the discussion, in Section~\ref{sec:extensions}, to relational systems with the usual restrictions on the relation in the Kripke model. That Section also brings two very interesting observations about intuitionistic $\K\T$. We end the paper with a discussion about logical Ecumenism in general.
\section{The system $\LEci$}\label{sec:LEci}
The language $\Lscr$ used for Ecumenical systems is described as follows. We will use a subscript $c$ for the classical meaning and $i$ for the intuitionistic, dropping such subscripts when formulae/connectives can have either meaning.
Classical and intuitionistic n-ary predicate symbols ($P_{c}, P_{i},\ldots$) co-exist in $\Lscr$ but have different meanings.
The neutral logical connectives $\{\bot,\neg,\wedge,\forall\}$ are common for classical and intuitionistic fragments, while $\{\iimp,\ivee,\iexists\}$ and $\{\cimp,\cvee,\cexists\}$ are restricted to intuitionistic and classical interpretations, respectively.
The sequent system $\LEci$ was presented in~\cite{Syn20} as the sequent counterpart of Prawitz natural deduction system.
The rules of $\LEci$ are depicted in Fig.~\ref{Fig:LEci}. Observe that the rules $R_c$ and $L_c$ describe the intended meaning of a classical predicate $P_{c}$ from an intuitionistic predicate $P_i$,
$\LEci$ has very interesting proof theoretical properties, together with a Kripke semantical interpretation, that allowed the proposal of a variety of ecumenical proof systems, such as a multi-conclusion and a nested sequent systems, as well as several fragments of such systems~\cite{Syn20}.
\begin{figure}[t]
{\sc Initial and structural rules}
\[
\infer[\init]{A,\Gamma \seq A}{}
\qquad
\infer[{\W}]{\Gamma \seq A}{\Gamma \seq \bot}
\]
{\sc Propositional rules}
\[
\infer[{\wedge L}]{A \wedge B,\Gamma \seq
C}{
A, B,\Gamma \seq C}
\quad
\infer[{\wedge R}]{\Gamma \seq
A \wedge B}{\Gamma
\seq A \quad \Gamma \seq B}
\quad
\infer[{\vee_i L}]{A \vee_i B,\Gamma \seq
C}{A,\Gamma
\seq C \quad B,\Gamma \seq C}
\quad
\infer[{\vee_i R_j}]{\Gamma \seq A_1\vee_i A_2}{\Gamma
\seq A_j}
\]
\[
\infer[\vee_cL]{ A \vee_c B,\Gamma \seq\bot }{A,\Gamma \seq \bot & B, \Gamma
\seq \bot}
\quad
\infer[\vee_cR]{\Gamma
\seq \ A \vee_c B}{\Gamma, \neg A , \neg B \seq \bot}
\quad
\infer[{\iimp L}]{\Gamma, A\iimp B \seq
C}
{A\iimp B,\Gamma \seq A \quad B, \Gamma \seq
C}
\]
\[
\infer[{\iimp R}]{\Gamma \seq
A\iimp B}{\Gamma, A \seq B}
\quad
\infer[\cimp L]{ A \rightarrow_c B,\Gamma \seq\bot }{ A \rightarrow_c B,\Gamma \seq A & B,\Gamma \ \seq
\bot}
\quad
\infer[\cimp R]{\Gamma \seq A
\rightarrow_c B}{\Gamma, A , \neg B \seq \bot}
\]
\[
\infer[{\neg L}]{\neg A,\Gamma \seq \bot
}{ \neg A,\Gamma \seq A}
\quad
\infer[{\neg R}]{\Gamma \seq
\neg A}{\Gamma, A \seq \bot}
\quad
\infer[{\bot L}]{\bot,\Gamma \seq A}{}
\quad
\infer[L_c]{ P_c,\Gamma\seq\bot}{P_i,\Gamma\seq \bot }
\quad
\infer[R_c]{\Gamma \seq P_c}{\Gamma,\neg P_i \seq\bot}
\]
{\sc Quantifiers}
\[
\infer[\forall L]{\forall x.A,\Gamma \seq C }{A[y/x],\forall x.A,\Gamma \seq C}
\quad
\infer[\forall R]{\Gamma \seq \forall x.A }{\Gamma \seq A[y/x]}
\]
\[
\infer[\exists_i L]{ \exists_ix.A,\Gamma \seq C}{A[y/x],\Gamma \seq C}
\quad
\infer[\exists_i R]{\Gamma \seq \exists_ix.A }{\Gamma \seq A[y/x]}
\quad
\infer[\exists_c L]{\exists_cx.A,\Gamma \seq \bot}{A[y/x],\Gamma \seq \bot}
\quad
\infer[\exists_c R]{\Gamma \seq \exists_cx.A}{\Gamma, \forall x.\neg A \seq \bot}
\]
\caption{Ecumenical sequent system $\LEci$. In rules
$\forall R,\iexists L,\cexists L$, the eigenvariable $y$ is fresh.}\label{Fig:LEci}
\end{figure}
Denoting by $\vdash_{\mathsf{S}} A$ the fact that the formula $A$ is a theorem in the proof system $\mathsf{S}$, the following theorems are easily provable in $\LEci$:
\begin{enumerate}
\item\label{neg} $\vdash_{\LEci} (A \to_c \bot) \leftrightarrow_i (A \to_i \bot) \leftrightarrow_i (\neg A)$;
\item\label{and} $\vdash_{\LEci} (A \vee_c B) \leftrightarrow_i \neg (\neg A \wedge \neg B)$;
\item\label{or} $\vdash_{\LEci} (A \to_c B) \leftrightarrow_i \neg (A \wedge \neg B)$;
\item\label{foe1} $\vdash_{\LEci} (\exists_cx.A) \leftrightarrow_i \neg (\forall x.\neg A)$.
\end{enumerate}
Note that~(\ref{neg}) means that the Ecumenical system defined in Fig.~\ref{Fig:LEci} does not distinguish between intuitionistic or classical negations, thus they can be called simply $\neg A$. We prefer to keep the negation operator in the language since the calculi presented in this work make heavy use of it.
Theorems (\ref{and}) to (\ref{foe1}) are of interest since they relate the classical
and the neutral operators: the classical connectives can be defined using negation, conjunction, and the universal quantifier.
On the other hand,
\begin{enumerate}
\setcounter{enumi}{4}
\item $\vdash_{\LEci} (A \to_i B) \to_i (A \to_c B)$ but $\not\vdash_{\LEci} (A \to_c B) \to_i (A \to_i B)$ in general;
\item $\vdash_{\LEci} A\vee_c\neg A$ but $\not\vdash_{\LEci} A\vee_i\neg A $ in general;
\item $\vdash_{\LEci} (\neg\neg A) \to_c A $ but $\not\vdash_{\LEci} (\neg\neg A) \to_i A $ in general;
\item\label{mp} $\vdash_{\LEci} (A\wedge(A \to_i B)) \to_i B$ but $\not\vdash_{\LEci} (A\wedge(A \to_c B)) \to_i B $ in general;
\item\label{foe2} $\vdash_{\LEci} \forall x.A \iimp \neg \exists_cx.\neg A $ but $\not\vdash_{\LEci} \neg\exists_cx.\neg A \to_i \forall x.A$ in general.
\end{enumerate}
Observe that~(\ref{foe1}) and~(\ref{foe2}) reveal the asymmetry between definability of quantifiers: while the classical existential can be defined from the universal quantification, the other way around is not true, in general. This is closely related with the fact that, proving $\forall x.A$ from $\neg\exists_cx.\neg A$ depends on $A$ being a classical formula. We will come back to this in Section~\ref{sec:modal}.
On its turn, the following result states that logical consequence in $\LEci$ is intrinsically intuitionistic.
\begin{proposition}[\cite{Syn20}]\label{prop:ev}
$\Gamma\vdash B$ is provable in $\LEci$ iff $\vdash_{\LEci}\bigwedge\Gamma\iimp B$.
\end{proposition}
To preserve the ``classical behaviour'', \ie, to satisfy all the principles of classical logic \eg\ {\em modus ponens} and the {\em classical reductio}, it is sufficient that the main operator of the formula
be classical (see~\cite{luiz17}). Thus, ``hybrid" formulas, \ie, formulas that contain
classical and intuitionistic operators may have a classical behaviour. Formally,
\begin{definition}
A formula $B$ is called \textit{externally classical} (denoted by $B^{c}$) if and only if $B$ is $\bot$, a classical predicate letter, or its
root operator is classical (that is: $\cimp,\vee_c,\exists_c$). A formula $C$ is {\em classical} if it is built from classical atomic predicates using only the connectives: $\cimp,\vee_c,\exists_c,\neg, \wedge,\forall$, and the unit $\bot$.
\end{definition}
For externally classical formulas we can now prove the following theorems
\begin{enumerate}
\setcounter{enumi}{9}
\item $\vdash_{\LEci} (A \to_c B^{c}) \to_i (A \to_i B^{c})$.
\item $\vdash_{\LEci} (A\wedge(A \to_c B^{c})) \iimp B^{c} $.
\item $\vdash_{\LEci} \neg\neg B^c\iimp B^c$.
\item $\vdash_{\LEci} \neg\exists_cx.\neg B^c \to_i \forall x.B^c$.
\end{enumerate}
Moreover, notice that all classical right rules as well as the right rules for the neutral connectives in $\LEci$ are invertible. Since invertible rules can be applied eagerly when proving a sequent, this entails that classical formulas can be eagerly decomposed.
As a consequence, the Ecumenical entailment, when restricted to classical succedents (antecedents having an unrestricted form), is classical.
\begin{theorem}[\cite{Syn20}]\label{thm:classical-right}
Let $C$ be a classical formula and $\Gamma$ be a multiset of Ecumenical formulas. Then
$$\vdash_{\LEci}\bigwedge\Gamma\cimp C \mbox{ iff } \vdash_{\LEci}\bigwedge\Gamma\iimp C.$$
\end{theorem}
This sums up well, proof theoretically, the {\em ecumenism} of Prawitz' original proposal.
\section{Ecumenical modalities}\label{sec:modal}
In this section we will propose an ecumenical view for alethic modalities. Since there are a number of choices to be made, we will construct our proposal step-by-step.
\subsection{Normal modal logics}\label{sec:nml}
The language of \emph{(propositional, normal) modal formulas} consists of a denumerable set $\Prop$ of propositional symbols and a
set of propositional connectives enhanced with the unary \emph{modal operators} $\square$ and $\lozenge$ concerning necessity and possibility, respectively~\cite{blackburn_rijke_venema_2001}.
The semantics of modal logics is often determined by means of \emph{Kripke models}. Here, we will follow the approach in~\cite{Sim94}, where a modal logic is characterized by the respective interpretation of the modal model in the meta-theory (called {\em meta-logical characterization}).
Formally, given a variable $x$, we recall the standard translation $\tradm{\cdot}{x}$ from modal formulas into first-order formulas with at most one free variable, $x$, as follows: if $P$ is atomic, then
$\tradm{P}{x}=P(x)$; $\tradm{\bot}{x}=\bot$; for any binary connective $\star$, $\tradm{A\star B}{x}=\tradm{A}{x}\star \tradm{B}{x}$; for the modal connectives
\begin{center}
\begin{tabular}{lclc@{\quad}lcl}
$\tradm{\square A}{x}$ & = & $\forall y ( \relfo(x,y) \rightarrow \tradm{A}{y})$ & &
$\tradm{\lozenge A}{x}$ & = & $\exists y (\relfo(x,y) \wedge \tradm{A}{y})$
\\
\end{tabular}
\end{center}
where $\rel(x,y)$ is a binary predicate.
Opening a parenthesis: such a translation has, as underlying justification, the interpretation of alethic modalities in a Kripke model $\m=(W,R,V)$:
\begin{equation}\label{eq:kripke}
\begin{array}{l@{\qquad}c@{\qquad}l}
\m, w \models \square A & \mbox{ iff } & \text{for all } v \text{ such that }w \rel v, \m, v \models A.\\
\m, w \models \lozenge A & \mbox{ iff } & \text{there exists } v \text{ such that } w \rel v \text{ and } \m, v\models A.
\end{array}
\end{equation}
$\rel(x,y)$ then represents the \emph{accessibility relation} $R$ in a Kripke frame. This intuition can be made formal based on the one-to-one correspondence between classical/intuitionistic translations and Kripke modal models~\cite{Sim94}. We close this parenthesis by noting that this justification is only motivational, aiming at introducing modalities. Models will be discussed formally in Section~\ref{sec:bi}.
The object-modal logic OL is then characterized in the first-order meta-logic ML as
$$
\vdash_{OL} A \quad\mbox{ iff } \quad \vdash_{ML} \forall x. \tradm{A}{x}
$$
Hence, if ML is classical logic (CL), the former definition characterizes the classical modal logic $\logick$~\cite{blackburn_rijke_venema_2001}, while if it is intuitionistic logic (IL), then it characterizes the intuitionistic modal logic $\logicik$~\cite{Sim94}.
In this work, we will adopt EL as the meta-theory (given by the system $\LEci$), hence characterizing what we will defined as the {ecumenical modal logic} $\EK$.
\subsection{An Ecumenical view of modalities}\label{sec:view}
The language of \emph{Ecumenical modal formulas} consists of a denumerable set $\Prop$ of (Ecumenical) propositional symbols and the set of Ecumenical connectives enhanced with unary \emph{Ecumenical modal operators}. Unlike for the classical case, there is not a canonical definition of constructive or intuitionistic modal logics. Here we
will mostly follow the approach in~\cite{Sim94} for justifying our choices for the Ecumenical interpretation for {\em possibility} and {\em necessity}.
The ecumenical translation $\tradme{\cdot}{x}$ from propositional ecumenical formulas into $\LEci$ is defined in the same way as the modal translation $\tradm{\cdot}{x}$ in the last section.
For the case of modal connectives, observe that, due to Proposition~\ref{prop:ev}, the interpretation of ecumenical consequence should be essentially {\em intuitionistic}. With this in mind, the semantical description of $ \square A$ given by~\eqref{eq:kripke} should be understood, for an arbitrary $v$, as: {\em assuming that $wRv$, then $A$ is satisfied in $v$}. Or:
\begin{center}
{\em $A$ is satisfied in $v$ is a consequence of the fact that $wRv$.}
\end{center}
This implies that the box modality is a {\em neutral connective}. The diamond, on the other hand, has two possible interpretations: classical and intuitionistic, since its leading connective is an existential quantifier. Hence we should have the ecumenical modalities: $\square, \ilozenge, \clozenge$, determined by the translations
\begin{center}
\begin{tabular}{lclc@{\qquad}lcl}
$\tradme{\square A}{x}$ & = & $\forall y ( \relfo(x,y) \iimp \tradme{A}{y})$\\
$\tradme{\ilozenge A}{x}$ & = & $\iexists y (\relfo(x,y) \wedge \tradme{A}{y})$ & &
$\tradme{\clozenge A}{x}$ & = & $\cexists y (\relfo(x,y) \wedge \tradme{A}{y})$
\\
\end{tabular}
\end{center}
Observe that, due to the equivalence~(\ref{foe1}), we have
\begin{enumerate}
\setcounter{enumi}{13}
\item $\clozenge A\iequiv \neg\square\neg A$
\end{enumerate}
On the other hand, $\square$ and $\ilozenge$ are not inter-definable due to (\ref{foe2}).
Finally, if $A^c$ is externally classical, then
\begin{enumerate}
\setcounter{enumi}{14}
\item $\square A^c\iequiv \neg\clozenge\neg A^c$
\end{enumerate}
This means that, when restricted to the classical fragment, $\square$ and $\clozenge$ are duals.
This reflects well the ecumenical nature of the defined modalities. We will denote by $\EK$ the Ecumenical modal logic meta-logically characterized by $\LEci$ via $\tradme{\cdot}{x}$.
\section{A labeled system for $\EK$}\label{sec:emodal}
\begin{figure}[t]
{\sc Initial and structural rules}
\[
\infer[\init]{ x:A,\Gamma \vdash x:A}{}
\quad
\infer[{\W}]{\Gamma \vdash x:A}{\Gamma \vdash y:\bot}
\]
{\sc Propositional rules}
\[
\infer[\wedge L]{ x:A\wedge B,\Gamma \vdash z:C}{ x:A, x:B,\Gamma \vdash z:C}
\quad
\infer[\wedge R]{\Gamma \vdash x:A \wedge B}
{\Gamma \vdash x:A
&
\Gamma \vdash x:B}
\]
\[
\infer[\ivee L]{x: A \ivee B,\Gamma \vdash z:C}
{x:A,\Gamma \vdash z:C
&
x:B,\Gamma \vdash z:C}
\quad
\infer[\ivee R_j]{\Gamma \vdash x:A_1 \ivee A_2}{\Gamma \vdash x:A_j}
\]
\[
\infer[\cvee L]{x: A \cvee B,\Gamma \vdash x:\bot}
{x:A,\Gamma \vdash x:\bot
&
x:B,\Gamma \vdash x:\bot}
\quad
\infer[\cvee R]{\Gamma \vdash x:A \cvee B}{\Gamma,x:\neg A,x:\neg B \vdash x:\bot}
\]
\[
\infer[\iimp L]{x:A \iimp B,\Gamma \vdash z:C}
{x:A \iimp B,\Gamma \vdash x:A
&
x:B,\Gamma \vdash z:C}
\quad
\infer[\iimp R]{\Gamma \vdash x:A \iimp B}{x:A,\Gamma \vdash x:B}
\]
\[
\infer[\cimp L]{x:A \cimp B,\Gamma \vdash x:\bot}
{x:A \cimp B,\Gamma \vdash x:A
&
x:B,\Gamma \vdash y: \bot}
\quad
\infer[\cimp R]{\Gamma \vdash x:A \cimp B}{x:A,x:\neg B,\Gamma \vdash x: \bot}
\]
\[
\infer[{\neg L}]{x:\neg A,\Gamma \vdash x:\bot
}{x:\neg A,\Gamma \vdash x:A}
\quad
\infer[{\neg R}]{\Gamma \vdash
x:\neg A}{x:A,\Gamma \vdash x:\bot}
\quad
\infer[\bot]{x:\bot, \Gamma \vdash z:C}{}
\]
\[
\infer[L_c]{\Gamma, x:P_c\vdash x:\bot}{\Gamma, x:P_i\vdash x:\bot }
\quad
\infer[R_c]{\Gamma \vdash x:P_c}{\Gamma,x:\neg P_i \vdash x:\bot}
\]
{\sc Modal rules}
\[
\infer[\square L]{xRy, x:\Box A,\Gamma \vdash z:C}{ xRy,y:A,x:\Box A,\Gamma \vdash z:C}
\quad
\infer[\square R]{\Gamma \vdash x:\Box A}{xRy, \Gamma \vdash y:A}
\quad
\infer[\ilozenge L]{x:\ilozenge A,\Gamma \vdash z:C}{xRy, y: A,\Gamma \vdash z:C}
\]
\[
\infer[\ilozenge R]{xRy, \Gamma \vdash x:\ilozenge A}{xRy,\Gamma \vdash y:A}
\quad
\infer[\clozenge L]{x:\clozenge A,\Gamma \vdash x:\bot}{xRy, y: A,\Gamma \vdash x:\bot}
\quad
\infer[\clozenge R]{\Gamma \vdash x:\clozenge A}{x:\Box\neg A,\Gamma \vdash x:\bot}
\]
\caption{Ecumenical modal system $\labEK$. In rules $\square R,\ilozenge L, \clozenge L$, the eigenvariable $y$ does not occur free in any formula of the conclusion.}
\label{fig:EK}
\end{figure}
One of the advantages of having an Ecumenical framework is that some well known classical/intuitionistic systems arise as fragments~\cite{Syn20}. In the following, we will seek for such systems by proposing a labeled sequent system for ecumenical modalities.
The basic idea behind labeled proof systems for modal logic is to internalize elements of the associated Kripke semantics (namely, the worlds of a Kripke structure and the accessibility relation between them) into the syntax.
\emph{Labeled modal formulas}
are either \emph{labeled formulas} of the form $x:A$ or \emph{relational atoms} of the form $x \rel y$, where $x, y$ range over a set of variables and $A$ is a modal formula.
\emph{Labeled sequents}
have the form $\Gamma \vdash x:A$, where $\Gamma$ is a multiset containing labeled modal formulas.
Following~\cite{Sim94}, we will prove the following meta-logical soundness and completeness theorem.
\begin{theorem}\label{thm:meta}
Let $\Gamma$ be a multiset of labeled modal formulas and denote $[\Gamma]=\{\rel(x,y)\mid xRy\in\Gamma\}\cup \{\tradme{B}{x}\mid x:B\in\Gamma\}$.
The following are equivalent:
\begin{enumerate}
\item[1.] $\Gamma\vdash x:A$ is provable in $\labEK$.
\item[2.] $[\Gamma]\seq \tradme{A}{x}$ is provable in $\LEci$.
\end{enumerate}
\end{theorem}
\begin{proof}
We will consider the following translation between $\labEK$ rule applications and $\LEci$ derivations, where the translation for the propositional rules is the trivial one:
\begin{align*}
\vcenter{
\infer[\square L]{xRy,x:\square A,\Gamma\vdash z:C}{xRy,y: A,x:\square A,\Gamma\vdash z:C}
}
&\quad\leadsto\quad
\\
\multispan2{$\hspace*{-2cm}
\vcenter{
\infer=[(\forall L, \iimp L)]{
\rel(x,y),\forall y.(\rel(x,y)\iimp\tradme{A}{y}),[\Gamma]\seq\tradme{C}{z}
}{
\infer{\rel(x,y),\forall y.(\rel(x,y)\iimp\tradme{A}{y}), \rel(x,y)\iimp\tradme{A}{y}, [\Gamma]\seq \rel(x,y)}{}
&
\deduce{\rel(x,y),\tradme{A}{y},\forall y.(\rel(x,y)\iimp\tradme{A}{y}),[\Gamma]\seq\tradme{C}{z}}{}&
}
}
$}
\end{align*}
\begin{align*}
&
\vcenter{
\infer[\square R]{\Gamma\vdash x:\square A}{xRy,\Gamma\vdash y: A}
}
&
\quad\leadsto\quad
&
\vcenter{
\infer=[(\forall R, \iimp R)]{[\Gamma]\seq \forall y.(\rel(x,y)\iimp\tradme{A}{y})}
{\rel(x,y),[\Gamma]\seq \tradme{A}{y}}
}
\\&
\vcenter{
\infer[\lozenge_i L]{x:\ilozenge A,\Gamma\vdash z:C}{xRy,y:A,\Gamma\vdash z:C}
}
&
\quad\leadsto\quad
&
\vcenter{
\infer=[(\iexists L, \wedge L)]{\iexists y.(\rel(x,y)\wedge\tradme{A}{y}),[\Gamma]\seq \tradme{C}{z}}{\rel(x,y),\tradme{A}{y},[\Gamma]\seq \tradme{C}{z}}
}
\\&
\vcenter{
\infer[\lozenge_c L]{x:\clozenge A,\Gamma\vdash x:\bot}{xRy,y:A,\Gamma\vdash x:\bot}
}
&
\quad\leadsto\quad
&
\vcenter{
\infer=[(\cexists L, \wedge L)]{\cexists y.(\rel(x,y)\wedge\tradme{A}{y}),[\Gamma]\seq \bot}
{\rel(x,y),\tradme{A}{y},[\Gamma]\seq \bot}
}
\\&
\vcenter{
\infer[\lozenge_i R]{
xRy,\Gamma\vdash x:\ilozenge A
}{
xRy,\Gamma\vdash y: A}
}
&
\quad\leadsto\quad
&
\vcenter{
\infer=[(\iexists R, \wedge R)]{
\rel(x,y),[\Gamma]\seq \iexists y.(\rel(x,y)\wedge\tradme{A}{y})
}{
\infer{\rel(x,y),[\Gamma]\seq \rel(x,y)}{}
&
\deduce{\rel(x,y),[\Gamma]\seq \tradme{A}{y}}{}
}
}
\\&
\vcenter{
\infer[\lozenge_c R]{\Gamma\vdash x:\clozenge A}{x:\Box\neg A,\Gamma\vdash x:\bot}
}
&
\quad\leadsto\quad
&
\vcenter{
\infer[(\cexists R)]{[\Gamma]\seq \cexists y.(\rel(x,y)\wedge\tradme{A}{y})}
{\forall y.\neg(\rel(x,y)\wedge\tradme{A}{y}),[\Gamma]\seq\bot}
}
\end{align*}
$(1)\Rightarrow(2)$ is then easily proved by induction on a proof of $\Gamma\vdash x:A$ in $\labEK$.
For proving $(2)\Rightarrow(1)$ observe that
\begin{itemize}
\item the rules $\iimp R,\wedge L, \wedge R$ are invertible in $\LEci$ and $\iimp L$ is semi-invertible on the right (\ie~if its conclusion is valid, so is its right premise);
\item $\vdash_{\LEci}\forall y.\neg (\rel(x,y)\wedge \tradme{A}{y})\leftrightarrow_i
\forall y.\rel(x,y)\iimp\tradme{\neg A}{y}$.
\end{itemize}
Hence, in the translated derivations in $\LEci$
{\em provability} is maintained from the end-sequent to the open leaves.
This means that choosing a formula $\tradme{B}{x}$ to work on is equivalent to performing
all the steps of the translation given above.
Therefore,
any derivation of $[\Gamma]\vdash\tradme{A}{x}$ in $\LEci$ can be transformed into a derivation of the same sequent where all the steps of the translation are actually performed.
This is, in fact, one of the pillars of the {\em focusing} method ~\cite{andreoli01apal,DBLP:journals/apal/LiangM11}.
In order to illustrate this, consider
the derivation
\begin{center}
$
\infer[(\forall L)]{\rel(x,y),\tradme{\Box A}{x},[\Gamma]\seq\tradme{C}{z}}
{\deduce{\rel(x,y),\rel(x,y)\iimp\tradme{A}{y},\tradme{\Box A}{x},[\Gamma]\seq\tradme{C}{z}}{\pi}}
$
\end{center}
where one decides to work on the formula $\forall y.(\rel(x,y)\iimp\tradme{A}{y})=\tradme{\Box A}{x}$ obtaining a premise containing the formula $\rel(x,y)\iimp\tradme{A}{y}$, with proof $\pi$. Since $\iimp L$ is semi-invertible on the right and the left premise is straightforwardly provable,
then $\pi$ can be substituted by the proof:
\begin{center}
$
\infer[\iimp L]{
\rel(x,y),\rel(x,y)\iimp\tradme{A}{y},\tradme{\Box A}{x},[\Gamma]\seq\tradme{C}{z}
}{
\infer{\rel(x,y),\rel(x,y)\iimp\tradme{A}{y},\tradme{\Box A}{x},[\Gamma]\seq\rel(x,y)}{}
&
\deduce{\rel(x,y),\tradme{A}{y},\tradme{\Box A}{x},[\Gamma]\seq\tradme{C}{z}}{\pi'}
}
$
\end{center}
Thus, by inductive hypothesis, $[xRy,y:A,x:\Box A,\Gamma\vdash z:C]$ is provable in $\labEK$.
\end{proof}
Finally, observe that, when restricted to the intuitionistic and neutral operators, $\labEK$
matches {\em exactly} Simpson's sequent system $\mathcal L_{\Box\Diamond}$~\cite{Sim94}. The analyticity of $\labEK$ is presented in Appendix~\ref{sec:cut}.
\section{Axiomatization}\label{sec:ax}
So far, we have motivated our discussion on Ecumenical modalities based on Simpson's
approach of meta-logical characterization. But what about the axiomatic characterization of $\EK$?
Classical modal logic $\K$ is characterized as propositional classical logic, extended with the \emph{necessitation rule} (presented in Hilbert style)
$
A/\square A
$
and the \emph{distributivity axiom}
$
\ka:\;\square(A\ra B)\ra (\square A\ra\square B)
$.
Intuitionistic modal logic should then consist of propositional intuitionistic logic plus necessitation and distributivity.
The problem is that there are many variants of axiom $\ka$ that induces classically, but not intuitionistically, equivalent systems. In fact, the following
axioms classically follow from $\ka$ and the De Morgan laws, but not in an intuitionistic setting
\begin{center}
$
\begin{array}{lc@{\qquad}l}
\ka_1:\;\square(A\ra B)\ra (\lozenge A\ra\lozenge B) & & \ka_2:\;\lozenge(A\vee B)\ra (\lozenge A\vee\lozenge B)\\
\ka_3:\;(\lozenge A\ra \square B)\ra \square(A\ra B) & & \ka_4:\;\lozenge \bot\ra\bot
\end{array}
$
\end{center}
The combination of axiom $\ka$ with axioms $\ka_1$ to $\ka_4$ then exactly characterizes intuitionistic modal logic $\IK$~\cite{plotkin:stirling:86,Sim94}.
In the ecumenical setting, there are many more variants from $\ka$, depending on the classical or intuitionistic interpretation of the implication and diamond. It is easy to see that
the intuitionistic versions of the above axioms are provable in $\EK$ (see Appendix~\ref{app}). Hence, by combining this result with cut-elimination (Appendix~\ref{sec:cut}), $\EK$ is {\em complete} w.r.t.~this set of axioms.
Observe that, since the intuitionistic operators imply the classical ones, if we substitute $i$ by $c$, the resulting clause is either not provable in $\EK$ or it is a consequence of the intuitionistic versions.
Next we will show that $\EK$ is also sound w.r.t.~this set of axioms. For that, we propose an Ecumenical birrelational semantics for $\EK$. The proof passes through a translation from $\labEK$ to $\mathcal L_{\Box\Diamond}$, so we remember that this last labeled system is sound and complete w.r.t. the birelational semantics of $\IK$~\cite{Sim94}.
\section{Ecumenical birelational models}\label{sec:bi}
In~\cite{DBLP:journals/jlp/Avigad01}, the negative translation was used to relate cut-elimination theorems for classical and intuitionistic logics. Since part of the argumentation was given semantically, a notion of Kripke semantics for classical logic was stated, via the respective semantics for intuitionistic logic and the double negation interpretation (see also~\cite{DBLP:journals/apal/IlikLH10}). In~\cite{luiz17} a similar definition was given, but under the Ecumenical approach, and it was extended to the first-order case in~\cite{Syn20}. We will propose a birelational Kripke semantics for Ecumenical modal logic, which is an extension of the proposal in~\cite{luiz17} to modalities.
\begin{definition}\label{def:ekripke} A {\em birelational Kripke model} is a quadruple
$\m=(W,\leq,R,V)$ where $(W,R,V)$ is a Kripke model
such that $\wld$ is partially ordered with order $\leq$, the
satisfaction function $V:\langle W,\leq\rangle\rightarrow \langle2^\Prop,\subseteq\rangle$ is monotone and:
\noindent
F1. For all worlds $w,v,v'$, if $wRv$ and $v \leq v'$, there is a $w'$ such that $w \leq w'$ and $w'Rv'$;
\noindent
F2. For all worlds $w', w, v$, if $w \leq w'$ and $wRv$, there is a $v'$ such that $w'Rv'$ and $v \leq v'$.
An {\em Ecumenical modal Kripke model} is a birelational Kripke model such that truth of an ecumenical formula at a point $w$ is the
smallest relation $\modelse$ satisfying
\noindent
$
\begin{array}{l@{\qquad}c@{\qquad}l}
\m,w\modelse P_{i} & \mbox{ iff } & P_{i}\in \val(w);\\
\m,w\modelse A\wedge B & \mbox{ iff } & \m,w\modelse A \mbox{ and } \m,w\modelse B;\\
\m,w\modelse A\vee_i B & \mbox{ iff } & \m,w\modelse A \mbox{ or } \m,w\modelse B;\\
\m,w\modelse A\iimp B & \mbox{ iff } & \text{for all } v \text{ such that }w \leq v,
\m,v\modelse A
\mbox{ implies }
\m,v\modelse B;\\
\m,w\modelse \neg A & \mbox{ iff } & \text{for all } v \text{ such that }w \leq v,
\m,v\not\modelse A;\\
\m,w \modelse \bot & & \mbox{never holds};\\
\m, w \modelse \square A & \mbox{ iff } & \text{for all } v,w' \text{ such that }w\leq w'\text{ and } w'\rel v, \m, v \modelse A.\\
\m, w \modelse \ilozenge A & \mbox{ iff } & \text{there exists } v \text{ such that } w \rel v \text{ and } \m, v\modelse A. \\
\m,w\modelse P_c & \mbox{ iff } & \m,w\modelse \neg(\neg P_i);\\
\m,w\modelse A\vee_c B & \mbox{ iff } & \m,w\modelse \neg(\neg A\wedge \neg B);\\
\m,w\modelse A\cimp B & \mbox{ iff } & \m,w\modelse \neg(A\wedge \neg B).\\
\m, w \modelse \clozenge A & \mbox{ iff } & \m,w\modelse \neg\square\neg A.
\end{array}
$
\end{definition}
Since, restricted to intuitionistic and neutral connectives, $\modelse$ is the usual birelational interpretation $\models$ for $\IK$ (and, consequently, $\mathcal L_{\Box\Diamond}$ \cite{Sim94}), and since the classical connectives are interpreted
via the neutral ones using the double-negation translation, an Ecumenical modal Kripke model is nothing else than the standard birelational Kripke model for intuitionistic modal logic $\IK$.
Hence, it is not hard to prove soundness and completeness of the semantical interpretation above w.r.t. the sequent system $\labEK$.
We start by defining a translation from $\labEK$ to $\mathcal L_{\Box\Diamond}$. We will abuse the notation and represent the connectives of $\IK/\mathcal L_{\Box\Diamond}$ using the neutral/intuitionistic correspondents in $\EK/\labEK$.
\begin{definition}\label{def:transS}
Let $\tradmk{\cdot}$ be the translation between formulas in $\EK$ and $\IK$ recursively defined as
\noindent
$
\begin{array}{lclc@{\qquad}lcl}
\tradmk{P_i} & = & P_i & & \tradmk{P_c} & = & \neg(\neg(P_i))\\
\tradmk{\bot} & = & \bot & &
\tradmk{\neg A} & = & \neg \tradmk{A}\\
\tradmk{A\wedge B} & = & \tradmk{A}\wedge\tradmk{B} & &
\tradmk{A\ivee B} & = & \tradmk{A}\ivee\tradmk{B}\\
\tradmk{A\iimp B} & = & \tradmk{A}\iimp\tradmk{B} & &
\tradmk{A\cvee B} & = & \neg(\neg\tradmk{A}\wedge\neg\tradmk{B})\\
\tradmk{A\cimp B} & = & \neg(\tradmk{A}\wedge\neg\tradmk{B})& &
\tradmk{\Box A} & = & \Box \tradmk{A}\\
\tradmk{\ilozenge A} & = & \ilozenge \tradmk{A} & &
\tradmk{\clozenge A} & = & \neg\Box\neg \tradmk{A}\\
\end{array}
$\\
\noindent
The translation $\tradms{\cdot}:\labEK\to \mathcal L_{\Box\Diamond}$ is defined as
$\tradms{x:A}=x:\tradmk{A}$ and assumed identical on relational atoms.
\end{definition}
Since the translations above preserve the double-negation interpretation of classical connectives into intuitionistic (modal) logic, it is possible to prove:
\begin{lemma}\label{lem:trad1}
$\vdash_{\labEK} \Gamma\vdash x:A$
iff $\vdash_{\labEK}\tradms{\Gamma\vdash x:A}$
iff $\vdash_{\mathcal L_{\Box\Diamond}}\tradms{\Gamma\vdash x:A}$.
\end{lemma}
\begin{lemma}\label{lem:trad3} Let $A$ be a formula in $\EK$. Then
$\modelse A$ iff $\models \tradmk{A}$.
\end{lemma}
\begin{proof}
First of all, note that an Ecumenical modal Kripke model is totally determined by the valuation of the (intuitionistic) propositional variables.
The proof follows then by easy structural induction on the formula $A$.
For example, if $A=\clozenge B$ then
$\modelse A$ iff $\modelse \neg\Box\neg B$. By inductive hypothesis, $\modelse B$ iff $\models \tradmk{B}$. Since $\neg$ and $\Box$ are neutral operators, then
$\modelse\neg\Box\neg B$ and iff $\models \neg\Box\neg \tradmk{B}$.
\end{proof}
Now, observe that every formula in $\IK$ is a formula in $\EK$, hence the following theorem holds.
\begin{theorem}\label{thm:scLEci}
$\EK$ is sound and complete w.r.t.~the Ecumenical modal Kripke semantics, that is, $\vdash_{\EK} A$ iff $\modelse A$.
\end{theorem}
Moreover, $\EK$ is sound and complete w.r.t. the axioms $\ka-\ka_4$ presented in Section~\ref{sec:ax}.
\section{Extensions}\label{sec:extensions}
Depending on the application, several further modal logics can be defined as extensions of $\K$ by simply restricting the class of frames we consider.
Many of the restrictions one can be interested in are definable as formulas of first-order logic, where the binary predicate $\rel(x,y)$ refers to the corresponding accessibility relation.
Table \ref{tab:axioms-FOconditions} summarizes some of the most common logics, the corresponding frame property, together with the modal axiom capturing it~\cite{Sah75}.
In the Ecumenical setting, we adopt the motto that ``relational predicates are second order citizens'' suggested in Section~\ref{sec:view} and interpret the implications in the axioms intuitionisticaly. We will refer to the ecumenical logic satisfying the axioms $F_1,\ldots, F_n$ as $\EK F_1\ldots F_n$.
We conjecture that the semantics of a given logic $\EK F_1 \ldots F_n$ can be inferred from the one for $\EK$ of Definition $\ref{def:ekripke}$: We just consider Ecumenical models whose accessibility relation satisfies the set of properties $\{F_1, \ldots, F_n\}$ in place of generic Ecumenical models.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Axiom} & \textbf{Condition} & \textbf{First-Order Formula} \\
\hline
$\T:\,\square A \rightarrow A \wedge A \rightarrow \lozenge A$ & Reflexivity & $\forall x. \rel(x,x)$ \\
\hline
$\4:\,\square A \rightarrow \square \square A \wedge \lozenge\lozenge A \rightarrow \lozenge A$ & Transitivity & $\forall x,y,z. (\rel(x,y) \wedge \rel(y,z)) \rightarrow \rel(x,z)$ \\
\hline
$\5:\,\square A \rightarrow \square \lozenge A \wedge \lozenge\square A \rightarrow \lozenge A$ & Euclideaness & $\forall x,y,z. (\rel(x,y) \wedge \rel(x,z)) \rightarrow \rel(y,z)$ \\
\hline
$\B:\,A \rightarrow \square \lozenge A \wedge\lozenge \square A \rightarrow A$ & Symmetry & $\forall x,y. \rel(x,y) \rightarrow \rel(y,x)$ \\
\hline
\end{tabular}
\end{center}
\caption{Axioms and corresponding first-order conditions on $\rel$.}
\label{tab:axioms-FOconditions}
\end{table}
Furthermore, following the approaches in~\cite{Sim94,Vig00,Neg05}, we can transform the axioms in Table~\ref{tab:axioms-FOconditions} into rules.
In future work, we would like to investigate how these rules behave w.r.t~the ecumenical setting.
\begin{figure}[t]
\[
\infer[\T]{\Gamma\vdash w:C}{xRx,\Gamma\vdash w:C}
\quad
\infer[\4]{xRy,yRz,\Gamma\vdash w:C}{xRz,xRy,yRz,\Gamma\vdash w:C}
\]\[
\infer[\5]{xRy,xRz,\Gamma\vdash w:C}{yRz,xRy,xRz,\Gamma\vdash w:C}
\quad
\infer[\B]{xRy,\Gamma\vdash w:C}{yRx,xRy,\Gamma\vdash w:C}
\]
\caption{Labeled sequent rules corresponding to axioms in Table~\ref{tab:axioms-FOconditions}.
}
\label{fig:extK}
\end{figure}
As a first step in this research direction, we finish this section with two very interesting observations about the case of axiom $\T$, illustrating the complexity of the interaction of modal axioms and ecumenical connective.
First of all, recall~\cite{Sim94,DBLP:conf/fossacs/Strassburger13} that by itself, $\square A \iimp A$ does not enforce reflexivity of an intuitionistic model. In fact, in frames having the reflexivity property, both $\square A \iimp A$ and $A \iimp \ilozenge A$ are provable.
\[
\infer[\wedge R]{xRx \vdash x: (\square A\iimp A) \wedge (A\iimp \ilozenge A)}
{\infer[\iimp R]{xRx \vdash x:\square A\iimp A}{\infer[\square L]{xRx, x:\square A \vdash x: A}{\infer[\init]{xRx, x: A \vdash x: A}{}}}&\infer[\iimp R]{xRx \vdash x: A \iimp \ilozenge A}{\infer[\ilozenge R]{xRx, x: A \vdash x: \ilozenge A}{\infer[\init]{xRx, x: A \vdash x: A}{}}}}
\]
For the converse, since $\square$ and $\ilozenge$ are not inter-definable, we need to add $A \iimp \ilozenge A$ in order to still be complete w.r.t. reflexive models.
Finally, it is well known that the set intuitionistic propositional operators is ``independent'', \ie, that each operator cannot be defined in terms of the others (Prawitz proposed a syntactical proof of this result in~\cite{prawitz1965}). It is also known that in the case of (many) constructive modal logics the modal operators $\Box$ and $\lozenge$ are independent of each other, as it is the case in $\EK$. But what would be the consequence of adding $\neg\ilozenge\neg A\iimp \Box A$ as an extra axiom to $\EK$? The following derivation shows that the addition of this new axiom has a disastrous propositional consequence.
$$\small
\infer[\cut]{\vdash x:(A\ivee\neg A)}
{\infer[eq]{\vdash x:\Box (A\ivee\neg A)}
{\infer[\neg R]{\vdash x:\neg\ilozenge\neg (A\ivee\neg A)}
{\infer[\ilozenge L]{x:\ilozenge\neg (A\ivee\neg A)\vdash x:\bot}
{\infer=[\neg L,\ivee R2,\neg R]{xRy,y:\neg (A\ivee\neg A)\vdash x:\bot}
{\infer[\neg L,\ivee R1]{xRy,y:A, y:\neg (A\ivee\neg A)\vdash y:\bot}
{\infer[\init]{xRy,y:A, y:\neg (A\ivee\neg A)\vdash y:A}{}}}}}}&
\infer[\T]{x:\Box (A\ivee\neg A)\vdash x:(A\ivee\neg A)}
{\infer[\Box L]{xRx,x:\Box (A\ivee\neg A)\vdash x:(A\ivee\neg A)}
{\infer[\init]{xRx,x: (A\ivee\neg A)\vdash x:(A\ivee\neg A)}{}}}}
$$
where $eq$ represents the proof steps of the substitution of a boxed formula for its
diamond version.\footnote{We have presented a proof with $\cut$ for clarity, remember that $\labEK$ has the cut-elimination property (see Appendix~\ref{sec:cut}).}
That is, if $\Box$ and $\lozenge_i$ are inter-definable, then $A\vee_i\neg A$ is a theorem and intuitionistic $\K\T$ collapses to a classical system!
\section{Discussion and conclusion}\label{sec:conc}
Some questions naturally arise with respect to Ecumenical systems: what (really) are Ecumenical systems? What are they good for? Why should anyone be interested in Ecumenical systems? What is the real motivation behind the definition and development of Ecumenical systems? Based on the specific case of the Ecumenical system that puts classical logic and intuitionist logic coexisting in peace in the same codification, we would like to propose three possible motivations for the definition, study and development of Ecumenical systems.
\paragraph{Philosophical motivation}
This was the motivation of Prawitz. Inferentialism, and in particular, logical inferentialism, is the semantical approach according to which the meaning of the logical constants can be specified by the rules that determine their correct use. According to Prawitz~\cite{DBLP:journals/Prawitz15},
\begin{quote}
{\em ``Gentzen's introduction rules, taken as meaning constitutive of the logical constants of the language of predicate logic, agree, as is well known, with how intuitionistic mathematicians use the constants. On the one hand, the elimination rules stated by Gentzen become all justified when the constants are so understood because of there being reductions, originally introduced in the process of normalizing natural deductions, which applied to proofs terminating with an application of elimination rules give canonical proofs of the conclusion in question. On the other hand, no canonical proof of an arbitrarily chosen instance of the law of the excluded middle is known, nor any reduction that applied to a proof terminating with an application of the classical form of reductio ad absurdum gives a canonical proof of the conclusion.''}
\end{quote}
But what about the use classical mathematicians make of the logical constants? Again, according to Prawitz,
\begin{quote}
{\em ``What is then to be said about the negative thesis that no coherent meaning can be attached on the classical use of the logical constants? Gentzen's introduction rules are of course accepted also in classical reasoning, but some of them cannot be seen as introduction rules, that is they cannot serve as explanations of meaning. The classical understanding of disjunction is not such that $A\vee B$ may be rightly asserted only if it is possible to prove either A or B, and hence Gentzen's introduction rule for disjunction does not determine the meaning of classical disjunction.''}
\end{quote}
As an alternative, in a recent paper~\cite{Murzi2018} Murzi presents a different approach to the extension of inferentialism to classical logic. There are some natural (proof-theoretical) inferentialist requirements on admissible logical rules, such as harmony and separability (although harmonic, Prawitz' rules for the classical operators do not satisfy separability). According to Murzi, our usual logical practice does not seem to allow for an inferentialist account of classical logic (unlike what happens with respect to intuitionistic logic). Murzi proposes a new set of rules for classical logical operators based on: absurdity as a punctuation mark, and Higher-level rules~\cite{DBLP:journals/sLogica/Schroeder-Heister14}. This allows for a ``pure'' logical system, where negation is not used in premises.
\paragraph{Mathematical/computational motivation}
(This was actually the original motivation for proposing Ecumenical systems.) The first Ecumenical system (as far as we know) was defined by Krauss in a technical report of the University of Kassel~\cite{Krauss} (the text was never published in a journal). The paper is divided in two parts: in the first part, Krauss' Ecumenical system is defined and some properties proved. In the second part, some theorems of basic algebraic number theory are revised in the light of this (Ecumenical) system, where constructive proofs of some ``familiar classical proofs'' are given (like the proof of Dirichlet's Unit Theorem).
The same motivation can be found in the final passages of the paper~\cite{DBLP:journals/Dowek16a}, where Dowek examines what would happen in the case of axiomatizations of mathematics. Dowek gives a simple example from Set Theory, and ends the paper with this very interesting remark:
\begin{quote}
{\em ``Which mathematical results have a classical formulation that can be proved from the axioms of constructive set theory or constructive type theory and which require a classical formulation of these axioms and a classical notion of entailment remains to be investigated.''}
\end{quote}
\paragraph{Logical motivation}
In a certain sense, the logical motivation naturally combines certain aspects of the philosophical motivation with certain aspects of the mathematical motivation.
According to Prawitz, one can consider the so-called classical first order logic as {\em ``an attempted codification of a fragment of inferences occurring in [our] actual deductive practice''}. Given that there exist different and even divergent attempts to codify our (informal) deductive practice, it is more than natural to ask about what relations are entertained between these codifications.
Ecumenical systems may help us to have a better understanding of the relation between classical logic and intuitionistic logic.
But one could say that, from a logical point of view, there's nothing new in the ecumenical proposal: Based on translations, the new classical operators could be easily introduced by ``explicit definitions''. Let us consider the following dialogue between a classical logician (CL) and an intuitionistic logician (IL), a dialogue that may arise as a consequence of the translations mentioned above:
\begin{itemize}
\item IL: if what you mean by $(A \vee B)$ is $\neg (\neg A \wedge \neg B)$, then I can accept the validity of $(A \vee \neg A)$!
\item CL: but I do not mean $\neg (\neg A \wedge \neg \neg A)$ by $(A \vee \neg A)$. One must distinguish the excluded-middle from the the principle of non-contradiction. When I say that Goldbach's conjecture is either true or false, I am not saying that it would be contradictory to assert that it is not true and that it is not the case that it is not true!
\item IL: but you must realize that, at the end of the day, you just have one logical operator, the Sheffer stroke (or the Quine's dagger).
\item CL: But this is not at all true! The fact that we can define one operator in terms of other operators does not imply that we don't have different operators! We do have 16 binary propositional operators (functions). It is also true that we can prove $\vdash (A \vee_{c} B) \leftrightarrow \neg(\neg A \wedge \neg B)$ in the ecumenical system, but this doest not mean that we don't have three different operators, $\neg $, $\vee_{c}$ and $\wedge $.
\end{itemize}
Maybe we can resume the logical motivation in the following (very simple) sentence:
\begin{quote}
Ecumenical systems constitute a new and promising instrument to study the nature of different (maybe divergent!) logics.
\end{quote}
Now, what can we say about {\em modal} Ecumenical systems?
Regarding the {\bf philosophical view}, in~\cite{Syn20} we have used invertibility results in order to obtain a sequent system for Prawitz' Ecumenical logic with a minimal occurrences of negations, moving then towards
a ``purer'' Ecumenical system. Nevertheless, negation still plays an important r\^{o}le on interpreting classical connectives. This is transferred to our definition of Ecumenical modalities, where the classical possibility is interpreted using negation. We plan to investigate what would be the meaning of classical possibility without impure rules. For the {\bf mathematical view}, our use of intuitionistic/classical/neutral connectives allows for a more chirurgical detection of the parts of a mathematical proof that are intrinsically intuitionistic, classical or independent. We now bring this discussion to modalities. Finally, concerning the {\bf logical view}, it would be interesting to explore some relations between general results on translations and Ecumenical systems, expanding this discussion to modalities.
To finish, we would like to say a word about our choices. It seems to be a common view, in the proof theory community, that Simpson's view is the more reasonable approach for modalities and intuitionism.
From that, the choice of a labeled proof system for $\EK$ seems only natural. But labeled systems have a very unfortunate feature: it is really tricky
to define an interpretation of sequents into the logical language. This problem often disappears when moving to nested-like systems~\cite{Brunnler:2009kx,Poggiolesi:2009vn,DBLP:conf/fossacs/Strassburger13,Lellmann:2015lns}, since the nestings keep the tree-structure information, matching exactly the history of a backwards proof search in an ordinary sequent calculus.
Also, having an Ecumenical nested system would most probably allow for a comparison, in one system, between the nested sequent for $\IK$~\cite{DBLP:conf/fossacs/Strassburger13} and for $\CK$~\cite{DBLP:journals/corr/ArisakaDS15,DBLP:journals/aml/KuznetsS19}. Hence this is a path worth pursuing, together with the comparison of $\labEK$ with other labelled sequent systems for intuitionistic modal logics, specially the recent ones proposed in~\cite{marin:hal-02390454} and~\cite{DBLP:journals/corr/abs-1901-09812}.
| edc7bc088defaf48cfa0a15df346f8b2fb4c503a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{ Introduction}
In this paper, we study a priori interior Hessian estimates in all dimensions for the Lagrangian mean curvature equation
\begin{equation}
F(D^{2}u)=\sum _{i=1}^{n}\arctan \lambda_{i}=\psi(x) \label{s}
\end{equation}
under the assumption that $|\psi|\geq (n-2)\frac{\pi}{2}+\delta$ where $\delta>0$, and $\psi$ has bounded second derivatives.
Here $\lambda_i$'s are the eigenvalues of the Hessian matrix $D^2u$ and then the phase $\psi$ becomes a potential for the mean curvature of the Lagrangian submanifold $(x,Du(x))\subseteq \mathbb{R}^n\times\mathbb{R}^n$. When the phase $\psi$ is constant, denoted by $c$, $u$ solves the special Lagrangian equation
\begin{equation}
\sum _{i=1}^{n}\arctan \lambda_{i}=c \label{s1}
\end{equation} or equivalently,
\[ \cos c \sum_{1\leq 2k+1\leq n} (-1)^k\sigma_{2k+1}-\sin c \sum_{0\leq 2k\leq n} (-1)^k\sigma_{2k}=0.
\]
Equation (\ref{s1}) originates in the special Lagrangian geometry by Harvey-Lawson \cite{HL}. The Lagrangian graph $(x,Du(x)) \subset \mathbb{R}^n\times\mathbb{R}^n$ is called special
when the argument of the complex number $(1+i\lambda_1)...(1+i\lambda_n)$
or the phase $\psi$ is constant, and it is special if and only if $(x,Du(x))$ is a
(volume minimizing) minimal surface in $(\mathbb{R}^n\times\mathbb{R}^n,dx^2+dy^2)$ \cite{HL}
A dual form of (\ref{s1}) is the Monge-Amp\'ere equation
\begin{equation*}
\sum_{i=1}^n \ln\lambda_i=c.
\end{equation*}
This is the potential equation for special Lagrangian submanifolds in $(\mathbb {R}^n\times \mathbb {R}^n, dxdy)$ as interpreted in \cite{Hi}. The gradient graph $(x,Du(x))$ is volume maximizing in this pseudo-Euclidean space as shown by Warren \cite{W}. In the 1980s, Mealy \cite{Me} showed that an equivalent algebraic form of the above equation is the potential equation for his volume maximizing special Lagrangian submanifolds in $(\mathbb R^n\times \mathbb R^n, dx^2-dy^2)$.
The arctangent operator or the logarithmic operator is concave if $u$ is convex, or if the Hessian of $u$ has a lower bound $\lambda\geq 0$. Certain concavity properties of the arctangent operator are still preserved for saddle $u$. The concavity of the arctangent operator in (\ref{s}) depends on the range of the Lagrangian phase. The phase $(n-2)\frac{\pi}{2}$ is called critical because the level set $\{ \lambda \in \mathbb{R}^n \vert \lambda$ satisfying $ (\ref{s})\}$ is convex only when $|\psi|\geq (n-2)\frac{\pi}{2}$ \cite[Lemma 2.2]{YY}. The concavity of the level set is evident for $|\psi|\geq (n-1)\frac{\pi}{2}$ since that implies $\lambda>0$ and then $F$ is concave. For solutions of (\ref{s1}) with critical and supercritical phases $|\psi|\geq (n-2)\frac{\pi}{2}$, Hessian estimates have been obtained by Warren-Yuan \cite{WY9,WY} and Wang-Yuan \cite{WaY}.
If the phase is subcritical $|\psi|<(n-2)\frac{\pi}{2}$, solutions of (\ref{s1}) fail to have interior estimates as shown in examples of Nadirashvili-Vl\u{a}du\c{t} \cite{NV} and Wang-Yuan \cite{WangY}.
Our main results in this paper are the following:
\begin{theorem}\label{main1}
Let $u$ be a $C^4$ solution of (\ref{s}) on $B_{R}(0)\subset \mathbb{R}^{n}$ where $\psi\in C^{1,1}(B_{R}) $, $|\psi|\geq (n-2)\frac{\pi}{2}+\delta$. Then we have
\begin{equation}
|D^2u(0)|\leq C_1\exp [C_2\max_{B_R(0)}|Du|^{2n-2}/R^{2n-2}] \label{H}
\end{equation}
where $C_1$ and $C_2$ are positive constants depending on $||\psi||_{C^{1,1}(B_{R})}$, $n$, and $\delta$.
\end{theorem}
In order to link the dependence of Hessian estimates
to the potential $u$ itself, we have the following gradient estimate.
\begin{theorem}\label{main2}
Let $u$ be a $C^3$ solution of (\ref{s}) on $B_{3R}(0)\subset \mathbb{R}^{n}$ where $\psi\in C^{1,1}(B_{3R}) $, $|\psi|\geq (n-2)\frac{\pi}{2}+\delta$. Then we have
\begin{equation}
\max_{B_{R}(0)}|Du|\leq C_3 osc_{B_{3R}(0)}\frac{u}{R}+C_4(n) \label{g1}
\end{equation}
where $C_3$ is a positive constant depending on $||\psi||_{C^{1}(B_{3R})}$, $n$, and $\delta$.
\end{theorem}
\begin{remark}
For the constant critical and supercritical phase equation (\ref{s1}), the constants $C_3$ and $C_4$ are only dimensional \cite[Theorem 1.2]{WY}. The constant $C_4$ was further reduced to $0$ in Yuan's unpublished 2015 notes on special Lagrangian equations.
\end{remark}
\begin{remark}
One application of the above estimates is the regularity (analyticity) of
$C^0$ viscosity solutions of (\ref{s}) where $|\psi|\geq (n-2)\frac{\pi}{2}+\delta$; solutions of the Dirichlet problem of (\ref{s}) with continuous boundary data enjoy interior regularity, as shown in \cite[Theorem 1.1]{AB1}. Another application is the regularity of convex $C^0$ viscosity solutions of (\ref{s}) where $\psi\in C^{2,\alpha}(B_1)$, as shown in \cite[Theorem 1.1]{BS}.
\end{remark}
\begin{remark} For Theorem \ref{main1}, an assumption weaker than $C^{1}$ on $\psi$ leads to counterexamples. For example, in two dimensions, we consider a boundary value problem of (\ref{s}) on the unit ball $B_1(0)$ where the phase is in $C^{\alpha}$ with $\alpha\in (0,1)$: $\psi(x)=\frac{\pi}{2}-\arctan (\alpha^{-1}|x|^{1-\alpha})$ and $u(x)=\int_{0}^{|x|}t^{\alpha}dt$ on $\partial B_1$. Now if Hessian estimates hold good for a H\"older continuous phase then by a smooth approximation of the phase and boundary data we would find a solution that is $C^{2,\alpha}$ in the interior of $B_1$. However, this boundary value problem admits a non $C^2$ unique viscosity solution $u$ with gradient $D u=|x|^{\alpha-1}x$, thereby proving a contradiction.
\end{remark}
\begin{remark}
The existence of interior estimates for solutions of (\ref{s}) with critical and supercritical phase where $\psi\in C^{1,\varepsilon_{0}}$, or even $|\psi|\geq (n-2)\frac{\pi}{2}$ where $\psi\in C^{1,1}$ are still open questions. Again if the phase is subcritical then even for the constant phase equation (\ref{s1}), singular $C^{1,\varepsilon}$ viscosity solutions were constructed in \cite{NV, WangY}.
\end{remark}
\medskip
For the two dimensional case, Heinz \cite{H} derived a Hessian bound for solutions of the Monge-Ampère type equation including (\ref{s1}); Pogorelov \cite{P1} derived Hessian estimates for solutions of these equations including (\ref{s1}) with $|\psi|\geq\frac{\pi}{2}$. Later Pogorelov \cite{P2} constructed his famous
counterexamples for the three dimensional Monge-Ampère equation $\sigma_3(D^2u) = \det(D^2u) = 1$, which also
serve as counterexamples for cubic and higher order symmetric $\sigma_k$ equations (see \cite{U1}). Gregori \cite{Gg} extended Heinz’s estimate to a gradient
bound in terms of the heights of the two dimensional minimal surfaces, and for graphs with non-zero mean curvature an additional requirement on the length of the mean curvature vector was assumed. Hessian estimates for solutions to Monge-Ampère equations and the $\sigma_k$
equations for $k\geq 2$ were established by Pogorelov \cite{P2} and Chou-Wang \cite{ChW} under certain strict convexity constraints. For the three dimensional case, Trudinger \cite{T2}, Urbas \cite{U2,U3}, and Bao-Chen
\cite{BC} obtained pointwise Hessian estimates in terms of certain integrals of
the Hessian, for $\sigma_k$ equations and the special Lagrangian equation (\ref{s1}) with $c=\pi$ respectively. Bao-Chen-Guan-Ji \cite{BCGJ} obtained pointwise Hessian estimates for strictly convex solutions to quotient equations $\sigma_n=\sigma_k$ in terms of certain
integrals of the Hessian.
Recently, along the integral way, Qiu \cite{Q} proved Hessian estimates for solutions of the three dimensional quadratic
Hessian equation with a $C^{1,1}$ variable right hand side. Hessian estimates for
convex solutions of general quadratic Hessian equations were obtained via a new pointwise approach by Guan-Qiu \cite{GQ}. For convex viscosity solutions of (\ref{s1}) Hessian estimates have been obtained by Chen-Warren-Yuan \cite{WYJ} and Chen-Shankar-Yuan \cite{RYJ}. Hessian estimates for semiconvex smooth solutions and almost convex viscosity solutions of $\sigma_2(D^2u)=1$ were recently established by Shankar-Yuan in \cite{RY1} and \cite{RYY} respectively.
Our proof of the Hessian estimates goes as follows: we first bound the Hessian of $u$ by its integral followed by
an integral of its gradient, then by the volume of the Lagrangian graph, and lastly,
by the height of the Lagrangian graph, which is the gradient of the solution of (\ref{s}). The presence of $\psi(x)$ in the non-uniformly elliptic equation \eqref{s} presents unique challenges. One of the difficulties is the unavailability of harmonic co-ordinates $\Delta_g x=0$ since the Lagrangian graph $(x,Du(x))\subset \mathbb{R}^n\times\mathbb{R}^n$ is not a minimal surface. As a result, the linearized operator of (\ref{s}) at $u$ does not represent the Laplace-Beltrami operator, like in the constant phase case \cite{WaY,WY}.
Another hurdle is proving Jacobi inequalities, which require differentiating \eqref{s} twice.
In the homogeneous case, one can differentiate \eqref{s1} and recover $D^3u$, but upon differentiating \eqref{s}, we end up with terms involving derivatives of $\psi$, which, a priori, could be large compared to the coefficients $DF(\lambda)$ of $D^3u$. We prove a Jacobi type inequality for $b=\ln\sqrt{1+\lambda_{\max}^2}$ where its Hessian is bounded below by its gradient and the $C^{1,1}$ norm of $\psi$. Applying a mean value inequality for $b$ and certain Sobolev inequalities we estimate the integral of $b$ by a weighted volume of the non-minimal Lagrangian graph. By a conformality identity, the weighted volume element turns out to be a linear combination of the elementary symmetric functions of $D^2u$.
The linear combination poses yet another difficulty but we take advantage of the divergence type structure to bound
the weighted volume of the Lagrangian graph in terms of its height and the $C^{1,1}$ norm of $\psi$.
Through out this paper we assume $\psi\geq (n-2)\frac{\pi}{2}+\delta$ since
by symmetry $\psi\leq- (n-2)\frac{\pi}{2}-\delta$ can be treated similarly. This paper is divided into the following sections: in section two, we introduce some notations and state some well known trigonometric inequalities satisfied by solutions of (\ref{s}), which will be used later in the proofs. In section three, we establish the gradient estimates, thereby proving Theorem \ref{main2}. In section four, we prove the pointwise and integral Jacobi inequality. In section five, we prove a mean value inequality for functions satisfying a Jacobi type inequality on submanifolds with high co-dimension, followed by the proof of Theorem \ref{main1}.\\
\section{Preliminaries}
\subsection{Notations}
We introduce some notations that will be used in this paper.
The induced Riemannian metric on the Lagrangian submanifold $X=(x,Du(x))\subset \mathbb{R}^n\times\mathbb{R}^n$ is given by
\[g=I_n+(D^2u)^2 .
\]
We denote
\begin{align*}
\partial_i=\frac{\partial}{\partial_{x_i}}\\
\partial_{ij}=\frac{\partial^2}{\partial_{x_i}\partial_{x_j}}\\
u_i=\partial_iu\\
u_{ij}=\partial_{ij}u.
\end{align*}
Note that for the functions defined below, the subscripts on the left do not represent partial derivatives\begin{align*}
b_k=(\ln\sqrt{1+\lambda_1^2}+...+\ln\sqrt{1+\lambda_k^2})/k\\
h_{ijk}=\sqrt{g^{ii}}\sqrt{g^{jj}}\sqrt{g^{kk}}u_{ijk}\\
g^{ii}=\frac{1}{1+\lambda_i^2}.
\end{align*}
Here $(g^{ij})$ is the inverse of the matrix $g$ and $h_{ijk}$ denotes the second fundamental form when the Hessian of $u$ is diagonalized.
The volume form, gradient, and inner product with respect to the metric $g$ are given by
\begin{align*}
dv_g=\sqrt{\det g}dx\\
\nabla_g v=g^{ij}v_iX_j\\
\langle\nabla_gv,\nabla_g w\rangle_g =g^{ij}v_iw_j\\
|\nabla_gv|^2=\langle\nabla_gv,\nabla_g v\rangle_g.
\end{align*}
\subsection{Laplace-Beltrami operator and mean curvature formula}
Taking variations of the energy functional $\int |\nabla_g v|^2 dv_g$ with respect to $v$, one has the Laplace-Beltrami operator of the metric $g$:
\begin{align}
\Delta_g =\frac{1}{\sqrt{ g}}\partial_i(\sqrt{ g}g^{ij}\partial_j )
=g^{ij}\partial_{ij}+\frac{1}{\sqrt{g}}\partial_i(\sqrt{g}g^{ij})\partial_j \label{2!}\\
=g^{ij}\partial_{ij}-g^{jp}\psi_q u_{pq} \partial_j. \nonumber
\end{align} The last equation follows from the following intrinsic and then extrinsic computations:
\begin{comment}
If the Hessian $D^2u$ of the $C^4$ function $u$ is diagonalized at $x_0$, $D^2u(x_0)=diag[\lambda_1,...,\lambda_n]$, then at $x_0$, (\ref{2!}) reduces to the following
\begin{equation}
\Delta_g = \sum_{i=1}^n g^{ii}\partial_{ii} -
\sum_{i=1}^ng^{ii}\lambda_i\psi_i\partial_{i}. \label{ho}
\end{equation}
One can verify the above formula from the following observation:
\begin{equation}
\Delta_g=g^{ij}\partial_{ij}+\frac{1}{\sqrt{g}}\partial_i(\sqrt{g}g^{ij})\partial_j \label{rrr}
\end{equation} and
\end{comment}
\begin{align}
\frac{1}{\sqrt{g}}\partial_i(\sqrt{g}g^{ij})=\frac{1}{\sqrt{g}}\partial_i(\sqrt{g})g^{ij}+\partial_ig^{ij}=\frac{1}{2}(\partial_i \ln g)g^{ij}+\partial_kg^{kj}\nonumber\\
=\frac{1}{2}g^{kl}\partial_i g_{kl}g^{ij}-g^{kl}\partial_k g_{lb}g^{bj}\nonumber\\
=-g^{jp}g^{ab}u_{abq}u_{pq}
=-g^{jp}\psi_q u_{pq}\label{lolz}
\end{align}
where the last equation follows from (\ref{111}) and (\ref{linearize}) below.
The first
derivative of the metric $g$ is given by
\begin{align}
\partial_i g_{ab}=\partial_i(\delta_{ab}+u_{ak}u_{kb})=u_{aik}u_{kb}+u_{bik}u_{ka}\overset{\text{at } x_0}{=}u_{abi}(\lambda_a+\lambda_b) \label{111}
\end{align}
assuming the Hessian of $u$ is diagonalized at $x_0$.
On taking the gradient of both sides of the Lagrangian mean curvature equation (\ref{s}), we get
\begin{equation}
\sum_{a,b=1}^{n}g^{ab}u_{jab}=\psi_j .\label{linearize}
\end{equation}
\begin{comment}
intrinsic computation shows
\begin{align}
\frac{1}{\sqrt{g}}\partial_i(\sqrt{g}g^{ij})=\frac{1}{\sqrt{g}}\partial_i(\sqrt{g})g^{ij}+\partial_ig^{ij}=\frac{1}{2}(\partial_i \ln g)g^{ij}+\partial_kg^{kj}\nonumber\\
=\frac{1}{2}g^{kl}\partial_m g_{kl}g^{mj}-\frac{1}{2}g^{kl}\partial_k g_{lm}g^{mj}-\frac{1}{2}g^{lk}\partial_l g_{km}g^{mj}\nonumber\\
=g^{kl}[-\frac{1}{2}g^{jm}(\partial_k g_{lm}+\partial_l g_{km}-\partial_m g_{kl})]=g^{kl}\Gamma^j_{kl}. \label{hot}
\end{align}
\end{comment}
\begin{comment}
Plugging (\ref{111}) and (\ref{linearize}) into (\ref{hot}) we get
\begin{align}
\frac{1}{\sqrt{g}}\partial_i(\sqrt{g}g^{ij})
=g^{jm}\psi_m\lambda_m. \label{rr}
\end{align}
Changing indices and writing $i=j=m=k$, we plug (\ref{rr}) in (\ref{rrr}) to get (\ref{ho}).
Note that in \cite[(2.19)]{HL}, the mean curvature vector $\vec{H}$ of this Lagrangian submanifold was shown to be
\begin{equation}
\vec{H}=J\nabla_g\psi \label{mean}
\end{equation}
where $\nabla_g$ is the gradient operator for the metric $g$ and $J$ is the complex structure, or the $\frac{\pi}{2}$ rotation matrix in $\mathbb{R}^n\times \mathbb{R}^n$. This follows from the following observation.
\end{comment}
The coefficients, given by (\ref{lolz}), are in fact equal to the tangential part of the following decomposition of $X_{ij}$
where $X=(x,Du(x))$
\begin{align*}
X_{ij}=(X_{ij})^T+(X_{ij})^N=\langle X_{ij},X_a\rangle g^{ab}X_b+II_{ij}\\
=u_{ijk}u_{ka}g^{ab}X_b+II_{ij}\overset{\text{define }}=\Gamma^b_{ij}X_b+II_{ij}
\end{align*}
where $\Gamma_{ij}^b$ is the Christoffel symbol.
\begin{comment}
is derived from the following:
\begin{align*}
\langle X_{ij},X_a\rangle= \partial_j\langle X_i,X_a\rangle-\langle X_i,X_{aj}\rangle\\
=\partial_j\langle X_i,X_a\rangle -\partial_a\langle X_i,X_j\rangle +\langle X_{ia},X_j\rangle\\
=\partial_j\langle X_i,X_a\rangle-\partial_a\langle X_i,X_j\rangle+\partial_i\langle X_a,X_j\rangle-\langle X_a,X_{ij}\rangle\\
=\frac{1}{2}(\partial_i g_{ja}+\partial_j g_{ia}-\partial_a g_{ij}).
\end{align*}
\end{comment}
On taking trace with respect to the metric $g$ and projecting to the tangential direction $X_l=(\partial_l, Du_l)$, we get
\[g^{ab}\Gamma^m_{ab}=g^{ab}u_{kab}u_{kl}g^{lm}=\psi_k u_{kl}g^{lm},
\]
which is the coefficient derived in (\ref{lolz}).
In turn, the normal part is
\[\vec{H}=g^{ab}II_{ab}=g^{ab}(\partial_{ab}X-\Gamma^m_{ab}\partial_m X)=\Delta_g X.
\]
On the other hand, directly projecting to the normal direction $J X_l$, we get the mean curvature vector of the Lagrangian submanifold $(x,Du(x))$
\begin{equation}
\vec{H}=g^{ab}\langle(-Du_{ab},0),-X_l\rangle g^{lm}JX_m=g^{ab}u_{lab}g^{lm}JX_m=\psi_l g^{lm}JX_m=J\nabla_g \psi \label{mean}
\end{equation}
where $\nabla_g$ is the gradient operator for the metric $g$ and $J$ is the complex structure, or the $\frac{\pi}{2}$ rotation matrix in $\mathbb{R}^n\times \mathbb{R}^n$ and we used (\ref{linearize}) for the second last equation. Note that the above formula for mean curvature of the Lagrangian submanifold $(x,Du(x))$ was originally found in \cite[(2.19)]{HL}.
\begin{comment}
\begin{lemma}
Suppose that the Hessian $D^2u$ of the $C^4$ function $u$ is diagonalized at $x_0$, $D^2u(x_0)=diag[\lambda_1,...,\lambda_n]$. Then at $x_0$, we have
\begin{equation}
\Delta_g = \sum_{i=1}^n g^{ii}\partial_{ii} -
\sum_{i=1}^ng^{ii}\lambda_i\psi_i\partial_{i}. \label{ho}
\end{equation}
\end{lemma}
\begin{proof}
First note that
\begin{align}
\Delta_g =\frac{1}{\sqrt{\det g}}\sum_{i,j=1}^{n}\partial_i(\sqrt{\det g}g^{ij}\partial_j )\nonumber\\
=\sum_{i,j=1}^{n} g^{ij}[\partial_i\partial_j -\sum_{k=1}^n \Gamma_{ij}^k \partial_k ].\label{rrr}
\end{align}
Now we compute the lower order terms of the above equation. Recall
\begin{align*}
\Gamma_{ij}^k=\frac{1}{2}g^{kl}[\partial_i g_{lj}+\partial_j g_{il}-\partial_l g_{ij}].
\end{align*}
We compute the first derivative of the metric $g$ at $x_0$
\begin{align}
\partial_i g_{ab}=\partial_i(\delta_{ab}+\sum_{k=1}^nu_{ak}u_{kb})=u_{abi}(\lambda_a+\lambda_b). \label{111}
\end{align}
Using the above, we compute
\begin{align}
g^{ij}\Gamma_{ij}^k=\frac{1}{2}g^{ij}g^{kl}[u_{jli}(\lambda_j+\lambda_l)+u_{ilj}(\lambda_i+\lambda_l)-u_{ijl}(\lambda_i+\lambda_j)]\nonumber\\
=\frac{1}{2}g^{ij}g^{kl}[2u_{ijl}\lambda_l]=g^{kl}\psi_l\lambda_l \label{rr}
\end{align}
where we used the fact that on taking the gradient of both sides of the Lagrangian mean curvature equation (\ref{s}), we get
\begin{equation}
\sum_{a,b=1}^{n}g^{ab}u_{jab}=\psi_j .\label{linearize}
\end{equation}
Changing indices and writing $i=j=l=k$, we plug (\ref{rr}) in (\ref{rrr}) to get (\ref{ho}).
In \cite[(2.19)]{HL}, the mean curvature vector $\vec{H}$ of this Lagrangian submanifold was shown to be
\begin{equation}
\vec{H}=J\nabla_g\psi \label{mean}
\end{equation}
where $\nabla_g$ is the gradient operator for the metric $g$ and $J$ is the complex structure, or the $\frac{\pi}{2}$ rotation matrix in $\mathbb{R}^n\times \mathbb{R}^n$. Note that by our assumption on $\psi$, $|H|$ is bounded.
\end{proof}
\end{comment}
\begin{remark}
When $\psi$ is constant, harmonic co-ordinates $\Delta_g x=0$ are available, which
reduces the Laplace-Beltrami operator on the minimal submanifold $\{(x,Du(x))|x\in B_R(0)\}$ to the linearized operator of (\ref{s1}) at $u$.
Also, note that in this paper, by our assumption on $\psi$, $|H|$ is bounded.
\end{remark}
Next we state the following Lemma.
\begin{lemma}\label{y1}
Suppose that the ordered real numbers $\lambda_{1}\geq \lambda_{2}\geq...\geq \lambda_{n}$ satisfy (\ref{s}) with $\psi\geq (n-2)\frac{\pi}{2}$.
Then we have \begin{enumerate}
\item $\lambda_{1}\geq \lambda_{2}\geq...\geq \lambda_{n-1}>0, \lambda_{n-1}\geq |\lambda_{n}|$,
\item $\lambda_{1}+(n-1)\lambda_{n}\geq 0$,
\item $\sigma_{k}(\lambda_{1},...,\lambda_{n})\geq 0$ for all $1\leq k\leq n-1$ and $n\geq 2$,
\item if $\psi\geq (n-2)\frac{\pi}{2}+\delta$, then $D^2u\geq -\cot \delta I_n$.
\end{enumerate}
\end{lemma}
\begin{proof}
Properties (1), (2), and (3) follow from \cite[Lemma 2.2]{WaY}. Property (4) follows from \cite[Pg 1356]{YY}.
\end{proof}
\section{Gradient Estimates}
We prove Theorem \ref{main2}.
\begin{proof}
We assume $R=1$ by scaling $\frac{u(Rx)}{R^2}$.
We set $M=osc_{B_1}u$. W.l.o.g we assume $M>0$. Replacing $u$ by $u-\min_{B_1}u+M$ we now have
\begin{equation}
M\leq u\leq 2M \label{q19}
\end{equation}
in $B_1$. We define
\[w=\eta |Du|+Au^2\] with $\eta=1-|x|^2$ and
\begin{equation}
A=\frac{n}{M}. \label{q29}
\end{equation}
If $w$ attains its supremum on
the boundary, then we are done.
So we assume that $w$ attains its supremum at an interior point $x_0\in B_1$. We choose a co-ordinate system so that $D^2u$ is diagonalized at $x_0$. Let's assume that $u_n\geq \frac{|Du|}{\sqrt{n}}>0$ at $x_0$. For $1 \leq i\leq n$ we have
\begin{equation}
\frac{u_{i}u_{ii}}{|Du|}=|Du|_i=-\frac{\eta_i|Du|+2Auu_i}{\eta}. \label{g11}
\end{equation}
Observe that $u_{nn}<0$ since $A=\frac{n}{M}$. Since $\psi\geq(n-2)\frac{\pi}{2}+\delta$ we must have $\lambda_{min}=\lambda_n$ and $\lambda_k\geq |\lambda_n|$ by Lemma \ref{y1}. So we see that for $1\leq k\leq n-1$
\begin{equation}
g^{nn}=\frac{1}{1+\lambda_n^{2}}\geq \frac{1}{1+\lambda_k^{2}}=g^{kk} \label{g12}
\end{equation}
and
\begin{equation}
\frac{1}{g^{nn}}=1+\lambda_n^{2}<C(\delta) \label{k}.
\end{equation}
Next we define the following operator \[Lu=\sum_{i=1}^{n}g^{ii}u_{ii}.\] Note that this is the Laplace-Beltrami operator on minimal submanifolds.
Using (\ref{g11}) and (\ref{g12}) we have
\[Lu(x_0)\geq g^{nn}u_{nn}=-g^{nn}\frac{|Du|(\eta_n|Du|+2Auu_n)}{nu_n}\]
which shows
\begin{equation}
Lu(x_0)\geq -g^{nn}\frac{|Du|6n}{\eta}. \label{hh}
\end{equation}
Recalling the definition of $w$, we note the following for all $1\leq i\leq n$ at $x_0$
\begin{align}
w_{i}=\eta |Du|_i+\eta_i|Du|+2Auu_i \nonumber\\
Lw=|Du|L\eta+2\sum_{a=1}^{n}g^{aa}\eta_{a}|Du|_{a}+\eta L(|Du|)+2AuLu+2A\sum_{a=1}^{n}g^{aa}u_a^2. \label{g5}
\end{align}
Next we observe the following
\begin{align}
\sum_{a,b=1}^{n}g^{ab}\partial_{ab}|Du|\nonumber\\
=\sum_{a,b,i=1}^{n}g^{ab}\left[ \frac{u_i u_{abi}}{|Du|}+\frac{u_{ai}u_{bi}}{|Du|}-\sum_{j=1}^{n}\frac{u_{aj}u_{bj}u_iu_j}{|Du|^3}\right]\nonumber\\
=\sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}+\sum_{a,b,i=1}^{n}g^{ab}\left[\frac{u_{ai}u_{bi}}{|Du|}-\sum_{j=1}^{n}\frac{u_{aj}u_{bj}u_iu_j}{|Du|^3}\right] \nonumber
\end{align}
where we get the last inequality using (\ref{linearize}).
Assuming $D^2u$ is diagonalized at $x_0$, we get
\begin{equation}
L(|Du|)(x_{0})= \sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}+\sum_{a=1}^{n}g^{aa}\frac{(|Du|^2-u_a^2)\lambda_a^2}{|Du|^3}\geq \sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}. \label{2.7}
\end{equation}
We plug (\ref{2.7}) in (\ref{g5}) and on applying (\ref{hh}), (\ref{g11}), (\ref{g12}) we get the following at $x_0$
\begin{align*}
Lw\geq -2ng^{nn}|Du|-2\sum_{a=1}^n g^{aa}\eta_a(\frac{\eta_a|Du|+2Auu_a}{\eta})+\\
+\eta\sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}-6ng^{nn}\frac{|Du|}{\eta}
+\frac{2A}{n}g^{nn}|Du|^{2}\\
\geq -2ng^{nn}|Du|-8g^{nn}\frac{|Du|}{\eta}-8g^{nn}Au\frac{|Du|}{\eta}+\eta \sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}\\
-6ng^{nn}\frac{|Du|}{\eta}
+\frac{2A}{n}g^{nn}|Du|^2.
\end{align*}
Noting that $Lw(x_0)\leq 0$, we divide the above inequality by $g^{nn}\frac{|Du|}{\eta}$ and on using (\ref{q19}), (\ref{q29}) we get the following at $x_0$
\begin{align*}
0\geq -2n\eta-8-8Au-6n+\frac{2A}{n}\eta|Du|+\sum_{i=1}^{n}\frac{\psi_iu_i}{|Du|}\frac{\eta^2}{g^{nn}|Du|}\\
\implies \eta|Du|\leq (12n+4)M-\sum_{i=1}^{n}\frac{\psi_iu_i}{2|Du|^{2}g^{nn}}\eta^{2}M\\
\implies \eta|Du(x_0)|\leq (12n+4)M+\frac{|D\psi|}{2\eta|Du(x_0)|}\eta^3MC(\delta)
\end{align*}
where the last inequality follows from the Cauchy-Schwarz inequality and (\ref{k}).
Now solving the above quadratic expression in $\eta|Du(x_0)|$, we get
\[\eta|Du|(x_0)\leq C_1M+\sqrt{C_1^2M+C_2M}
\]
where \begin{align*}
C_1=12n+4=C(n)\\
C_2=C(n,||\psi||_{C^{1}(B_1)},\delta).
\end{align*}
This proves the gradient estimate:
\begin{align*}
|Du(0)|\leq w(0)\leq w(x_0)\leq \eta|Du|(x_0)+Au^2(x_0)\\
\leq C_1M+\sqrt{C_1^2M^2+C_2M}+4nM\\
\leq C(n,||\psi||_{C^{1}(B_1)},\delta)osc_{B_1}u+C(n).
\end{align*}
\end{proof}
\begin{remark} The above proof also follows from the observation that when $\psi$ lies in the supercritical range, $u$ is semiconvex [Lemma \ref{y1}]. On modifying $u$ to the convex function $
\tilde{u}(x)=u(x)+\cot(\delta)\frac{|x|^2}{2}$, the gradient estimate (which is independent of the Lipschitz norm of $\psi$) follows from the fact that the gradient of any convex
function is dominated by its oscillation.
\end{remark}
\section{The Jacobi inequality}
In this section we prove the Jacobi inequality and the integral Jacobi inequality, which is essential in proving the Hessian estimates.
\begin{proposition}
Let $u$ be a smooth solution of (\ref{s}) in $\mathbb{R}^{n}$. Suppose that the Hessian $D^{2}u$ is diagonalized and the eigenvalue $\lambda_\gamma$ is distinct from all other eigenvalues of $D^2u$ at point $x_0$. Then we have the following at $x_0$
\begin{equation}
|\nabla_g \ln\sqrt{1+\lambda_\gamma^2}|^2=\sum_{k=1}^{n}\lambda_\gamma^2h_{rrk}^2\label{grad}
\end{equation} and
\begin{align}
\Delta_g\ln\sqrt{1+\lambda_\gamma^2}=\nonumber\\
(1+\lambda_\gamma^2)h_{rrr}^2+\sum_{k\neq r}\left[\frac{2\lambda_\gamma}{\lambda_\gamma-\lambda_k}+\frac{2\lambda_\gamma^2\lambda_k}{\lambda_\gamma-\lambda_k} \right]h_{kkr}^2 \nonumber\\
+\sum_{k\neq r}\left[1+\frac{2\lambda_\gamma}{\lambda_\gamma-\lambda_k}+\frac{\lambda_\gamma^2(\lambda_k+\lambda_\gamma)}{\lambda_\gamma-\lambda_k} \right]h_{rrk}^2\nonumber\\
+\sum_{k>j,k,j\neq r}2\lambda_\gamma\left[ \frac{1+\lambda_k^2}{\lambda_\gamma-\lambda_k}+\frac{1+\lambda_j^2}{\lambda_\gamma-\lambda_j} +(\lambda_j+\lambda_k) \right]h^2_{kjr}\nonumber\\
+\frac{\lambda_\gamma}{1+\lambda_\gamma^2}\psi_{\gamma\gamma}- \sum_{a=1}^n\lambda_ag^{aa}\psi_a\partial_{a}\ln\sqrt{1+\lambda_\gamma^2}. \label{cc}
\end{align}
\end{proposition}
\begin{proof}
Define
\[b_\gamma=\ln\sqrt{1+\lambda_\gamma^2}.
\] We assume $\gamma=1$ for the sake of simplifying notation. On implicitly differentiating the characteristic equation
\[\det(D^2u-\lambda_1I)=0
\] near any point where $\lambda_1$ is distinct from the other eigenvalues we get
\begin{align*}
\partial_e \lambda_1=\partial_e u_{11}\\
\partial_{ee}\lambda_1=\partial_{ee}u_{11}+\sum_{k>1}2\frac{(\partial_eu_{ik})^2}{\lambda_1-\lambda_k}
\end{align*} where $e$ is any arbitrary unit vector in $\mathbb{R}^n$.
On computing the derivatives of the smooth function $b_1$ near $x_0$, we get
\[|\nabla_g b_1|^2=\sum_{k=1}^n g^{kk}\left[\frac{\lambda_1}{1+\lambda^2}\partial_ku_{11} \right]^2=\sum_{k=1}^n\lambda_1^2h_{11k}^2.
\]
We see that
\begin{align*}
\partial_{ee}b_1=\partial_{ee}\ln\sqrt{1+\lambda_1^2}=\frac{\lambda_1}{1+\lambda_1^2}\partial_{ee}\lambda_1+\frac{1-\lambda_1^2}{(1+\lambda_1^2)^2}(\partial_e\lambda_1)^2.\\
\text{ At $x_0$, } \partial_{ee}b_1=\frac{\lambda_1}{1+\lambda_1^2}[\partial_{ee}u_{11}+\sum_{k>1}2\frac{(\partial_eu_{ik})^2}{\lambda_1-\lambda_k}]+\frac{1-\lambda_1^2}{(1+\lambda_1^2)^2}(\partial_eu_{11})^2 .
\end{align*}
We define an operator $L=\sum_{a,b=1}^{n}g^{ab}\partial_{ab}.$ At $x_0$, we have
\begin{align}
Lb_1=\sum_{r=1}^{n}g^{rr}\partial_{rr}b_1\label{label}\\
=\sum_{r=1}^{n}g^{rr}\frac{\lambda_1}{1+\lambda_1^2}\left[\partial_{rr}u_{11}+2\sum_{k>1}\frac{u_{1kr}^2}{\lambda_1-\lambda_k} \right]\nonumber\\
+\sum_{r=1}^n\frac{1-\lambda_1^2}{(1+\lambda_1^2)^2}g^{rr}u_{11r}^2. \label{bbb22}
\end{align}
Combining (\ref{linearize}) with (\ref{111}) we observe the following at $x_0$ for $i,j$ fixed
\begin{align}
Lu_{ij}=\sum_{a,b=1}^{n}g^{ab}u_{ijab}=\psi_{ij}-g^{ab}_i u_{abj}\nonumber\\
=\psi_{ij}+\sum g^{aa}g^{bb}(\lambda_a+\lambda_b)u_{abi}u_{abj}. \label{u}
\end{align}
Next in (\ref{u}) we substitute $\partial_{rr} u_{11}$ in terms of lower order derivatives and $\psi$. Recalling (\ref{2!}), at $x_0$ we have
\begin{equation}
\Delta_g = \sum_{i=1}^n g^{ii}\partial_{ii} -
\sum_{i=1}^ng^{ii}\lambda_i\psi_i\partial_{i}. \label{ho}
\end{equation}
We set $i=j=1$, and plug (\ref{u}) in (\ref{bbb22}) and then regroup the terms $h_{**1}, h_{11*}, h_{*@1}$ to get
\begin{align}
\Delta_g b_1=Lb_1- \sum_{a=1}^n\lambda_ag^{aa}\psi_a\partial_{a}\ln\sqrt{1+\lambda_1^2}=\frac{\lambda_1}{1+\lambda_1^2}\psi_{11}+\\
(1-\lambda_1^2)h_{111}^2+2\sum_{a=1}^{n}\lambda_1\lambda_ah^2_{aa1}+2\sum_{k>1}\frac{\lambda_1(1+\lambda_k^2)}{\lambda_1-\lambda_k}h_{kk1}^2\\
+2\sum_{k>1}\lambda_1(\lambda_1+\lambda_k)h_{k11}^2+\sum_{k>1}(1-\lambda_1^2)h_{11k}+
2\sum_{k>1}\frac{\lambda_1(1+\lambda_k^2)}{\lambda_1-\lambda_k}h_{1k1}^2\\
+2\sum_{k>j>1}\lambda_1(\lambda_j+\lambda_k)h_{jk1}^2+ 2\sum_{j\neq k,j,k>1}\frac{\lambda_1(1+\lambda_k^2)}{\lambda_1-\lambda_k}h_{jk1}^2- \sum_{a=1}^n\lambda_ag^{aa}\psi_a\partial_{a}\ln\sqrt{1+\lambda_1^2}.
\end{align}
On simplifying we get (\ref{cc}).
\end{proof}
\begin{lemma}\label{4.2}
Let $u$ be a smooth solution of (\ref{s}) in $\mathbb{R}^{n}$. Suppose that the Hessian $D^{2}u$ is diagonalized at $x_0$ and that the ordered eigenvalues $\lambda_1\geq\lambda_2\geq...\geq\lambda_n$ of the Hessian satisfy $\lambda_1=...=\lambda_m>\lambda_{m+1}$ at $x_0$. Then the function $b_m=\frac{1}{m}\sum_{i=1}^m\ln\sqrt{1+\lambda_i^2}$ is smooth near $x_0$ and at $x_0$ it satisfies
\begin{equation}\Delta_g b_m\geq c(n)|\nabla_gb_m|^2-C\label{J}
\end{equation}
where $C=C(||\psi||_{C^{1,1}(B_1)},\delta,n)$.
\end{lemma}
\begin{proof}
\begin{itemize}
\item [Step 1.] The function $b_m$ is symmetric in $\lambda_1,..,\lambda_m$. Thus for
$m < n$, $b_m$ is smooth in terms of the matrix entries when $\lambda_m>\lambda_{m+1}$ and it is
still smooth in terms of $x$ after being composed with the smooth function $D^2u(x)$, in particular near $x_0$, at which $\lambda_1=...=\lambda_m>\lambda_{m+1}$. For $m= n$, $b_n$ is
clearly smooth everywhere.
First let's assume that the first $m$ eigenvalues are distinct. Using (\ref{bbb22}) from
the Proposition above, we compute $mL(b_m)$ where $L$ is the operator defined in (\ref{label}). As before after grouping the terms $h_{***},h_{**@},h_{*@!}$ in the summation, we get
\begin{align*}
mL(b_m)(x_0)=\sum_{k\leq m}(1+\lambda_k^2)h^2_{kkk}+(\sum_{i<k\leq m}+\sum_{k<i\leq m})(3+\lambda_i^2+2\lambda_i\lambda_k)h^2_{iik}\\
+\sum_{k\leq m<i}\frac{2\lambda_k(1+\lambda_k\lambda_i)}{\lambda_k-\lambda_i}h^2_{iik}
+\sum_{i\geq m<k}\frac{3\lambda_i+\lambda_k+\lambda_i^2(\lambda_i+\lambda_k)}{\lambda_i-\lambda_k}h^2_{iik}\\
+2\Bigg[\sum_{i<j<k\leq m}(3+\lambda_i\lambda_j+\lambda_j\lambda_k+\lambda_k\lambda_i)h^2_{ijk}\\
+\sum_{i<j\leq m<k}(1+\lambda_i\lambda_j+\lambda_j\lambda_k+\lambda_k\lambda_i+\lambda_i\frac{1+\lambda_k^2}{\lambda_i-\lambda_k}+\lambda_j\frac{1+\lambda^2_k}{\lambda_j-\lambda_k})h^2_{ijk}\\
+\sum_{i\leq m<j<k}\lambda_i[\lambda_j+\lambda_k+\frac{1+\lambda_j^2}{\lambda_i-\lambda_j}+\frac{1+\lambda_k^2}{\lambda_j-\lambda_k}]h_{ijk}^2\Bigg]\\
+ \sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}.
\end{align*}
Observe that as a function of matrices, $b_m$ is $C^2$ at $D^2u(x_0)$ with eigenvalues satisfying $\lambda=\lambda_1=...=\lambda_m>\lambda_{m+1}$. Note that $D^2u(x_0)$ can be approximated by matrices with distinct eigenvalues. This shows that the above expression for $Lb_m(x_0)$ still holds good. Using Lemma \ref{y1} we can further simplify it to
\begin{align}
mL(b_m)(x_0)= \sum_{k\leq m}(1+\lambda^2)h^2_{kkk}+(\sum_{i<k\leq m}+\sum_{k<i\leq m})(3+3\lambda^2)h_{iik}^2+\sum_{k\leq m<i}\frac{2\lambda(1+\lambda\lambda_i)}{\lambda-\lambda_i}h^2_{iik}\nonumber\\
+\sum_{i\leq m<k}\frac{3\lambda-\lambda_k+\lambda^2(\lambda+\lambda_k)}{\lambda-\lambda_k}h_{iik}^2
+2\Bigg[\sum_{i<j<k\leq m}(3+3\lambda^2)h_{ijk}^2
+\sum_{i<j\leq m<k}[1+\frac{2\lambda}{\lambda-\lambda_k}+\nonumber\\
\frac{\lambda^2(\lambda+\lambda_k)}{\lambda-\lambda_k}]h_{ijk}^2
+\sum_{i\leq m<j<k}\lambda[\lambda_j+\lambda_k+\frac{1+\lambda_j^2}{\lambda-\lambda_j}+\frac{1+\lambda_k^2}{\lambda-\lambda_k}]h_{ijk}^2\Bigg]
+\sum_{i=1}^{m}\frac{\lambda}{1+\lambda^2}\psi_{ii}\nonumber\\
\geq \sum_{k\leq m}\lambda^2h_{kkk}^2+(\sum_{i<k\leq m}+\sum_{k<i\leq m})3\lambda^2h^2_{iik}+\sum_{k\leq m<i}\frac{2\lambda^2\lambda_i}{\lambda-\lambda_i}h^2_{iik}\nonumber\\
+\sum_{i\leq m<k}\frac{\lambda^2(\lambda+\lambda_k)}{\lambda-\lambda_k}h^2_{iik}+\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}.\label{star2}
\end{align}
Using the $C^1$ continuity of $b_m$ as a function of matrices at $D^2u(x_0)$, we can simplify (\ref{grad}) at $x_0$ to
\begin{equation}
|\nabla_gb_m|^2(x_0)=\frac{1}{m^2}\sum_{1\leq k\leq n}\lambda^2\left[ \sum_{i\leq m}h_{iik} \right]^2\leq \frac{\lambda^{2}}{m}\sum_{1\leq k\leq n}\left[ \sum_{i\leq m}h^2_{iik}\right]. \label{star1}
\end{equation}
Combining (\ref{star1}) and (\ref{star2}) we get the following at $x_0$:
\begin{align}
m(\Delta_g b_m-\varepsilon(n)|\nabla_gb_m|^2)\geq \nonumber\\
\lambda^2\left[\sum_{k\leq m}(1-\varepsilon)h_{kkk}^2+(\sum_{i<k\leq m} + \sum_{k<i\leq m})(3-\varepsilon)h^2_{iik}+ 2\sum_{k\leq m< i}\frac{\lambda_i}{\lambda-\lambda_i}h^2_{iik} \right] \label{y5.23}\\
+\lambda^2\left[ \sum_{i\leq m \leq k} (\frac{\lambda+\lambda_k}{\lambda-\lambda_k}-\varepsilon)h^2_{iik} \right] \label{y5.24}\\
+\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}\label{11}\\
-m\sum_{i=1}^n\lambda_ig^{ii}\psi_i\partial_i b_m \label{22}
\end{align}
with $\varepsilon(n)$ to be fixed.
\item[Step 2.] Next we estimate each of the terms of the above expression. For each fixed $k$ in the above expression, we set
$t_i=h_{iik}.$
For the sake of simplicity, we use the following notation
\begin{align*}
(\ref{y5.23}+\ref{y5.24})=Z_1+Z_2=Z\\
H^k(x_0)=t_1(x_0)+...+t_{n-1}(x_0)+t_n(x_0)=t'(x_0)+t_{n}(x_0)
\end{align*}
where $H^k$ denotes the $k$th component of the mean curvature vector given by (\ref{mean}), i.e. the component of the mean curvature vector along $J(e_k,Du_{e_k})$ with $e_k$ being the $k^{th}$ eigendirection of $D^2u$.
So far we have
\[m(\Delta_g b_m-\varepsilon(n)|\nabla_gb_m|^2)\geq
Z+\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}-m\sum_{i=1}^n\lambda_ig^{ii}\psi_i\partial_i b_m.\]
\item[Step 2.1. ]We estimate the term $Z$ by first showing that $Z_1\geq-C(||\psi||_{C^{1}(B_1)})$.
For each fixed $k\leq m$ in (\ref{y5.23}), we show that the $[\hspace{.1cm}]_k$ term is $\geq 0$.
For the case where $\lambda_i\geq 0$ for all $i$, the proof follows directly. So we consider only the case where $\lambda_{n-1}>0>\lambda_n$.
For simplifying notation we assume $k=1$. Noting that $t_n(x_0)=H^1(x_0)-t'(x_0)$ we observe the following:
\begin{align}
[\hspace{.1cm}]_1= \Bigg[ (1-\varepsilon)t_1^2+\sum_{i=2}^m(3-\varepsilon)t_i^2+\sum_{i=m+1}^{n-1}\frac{2\lambda_i}{\lambda-\lambda_i}t_i^2 \Bigg]+\frac{2\lambda_n}{\lambda-\lambda_n}t_n^2\nonumber\\
= \Bigg[ (1-\varepsilon)t_1^2+\sum_{i=2}^m(3-\varepsilon)t_i^2+\sum_{i=m+1}^{n-1}\frac{2\lambda_i}{\lambda-\lambda_i}t_i^2 \Bigg]\nonumber\\
+\frac{2\lambda_n}{\lambda-\lambda_n}[(H^1)^2-2H^1t'+t'^2]\nonumber\\
\geq \Bigg[ (1-\varepsilon)t_1^2+\sum_{i=2}^m(3-\varepsilon)t_i^2+\sum_{i=m+1}^{n-1}\frac{2\lambda_i}{\lambda-\lambda_i}t_i^2 \Bigg]\nonumber\\
+\frac{2\lambda_n}{\lambda-\lambda_n}[t'^2(1+\delta)]
+\frac{2\lambda_n}{\lambda-\lambda_n}[(H^1)^2(1+\frac{4}{\delta})]\nonumber
\end{align} where the last inequality follows from Young's inequality. Noting that $\frac{2\lambda_n}{\lambda-\lambda_n}\geq-\frac{2}{n}$ and (\ref{mean}), we see the following
\begin{align}
[\hspace{.1cm}]_1
\geq \Bigg[ (1-\varepsilon)t_1^2+\sum_{i=2}^m(3-\varepsilon)t_i^2+\sum_{i=m+1}^{n-1}\frac{2\lambda_i}{\lambda-\lambda_i}t_i^2 \Bigg]\nonumber\\
+\frac{2\lambda_n}{\lambda-\lambda_n}[t'^2(1+\delta)]-C(||\psi||_{C^{1}(B_1)})\nonumber\\
\geq \Bigg[(1-\varepsilon)t_1^2+\sum_{i=2}^m(3-\varepsilon)t_i^2+\sum_{i=m+1}^{n-1}\frac{2\lambda_i}{\lambda-\lambda_i}t_i^2 \Bigg]\label{q1}\\
\Bigg[1+\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}\big(\frac{1}{1-\varepsilon}+\sum_{i=2}^m\frac{1}{3-\varepsilon}+\sum_{i=m+1}^{n-1}\frac{\lambda-\lambda_i}{2\lambda_i} \big) \Bigg]\label{q2}\\
-C(||\psi||_{C^{1}(B_1)})\nonumber
\end{align}where the last inequality follows from the Cauchy-Schwartz inequality. We see that (\ref{q1}) is positive, so now we need to choose $\varepsilon(n)$ suitably to make (\ref{q2}) positive, thereby proving $Z_1\geq -C(||\psi||_{C^1(B_1)})$.
We have
\begin{align}
\Bigg[1+\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}(\frac{1}{1-\varepsilon}+\sum_{i=2}^m\frac{1}{3-\varepsilon}+\sum_{i=m+1}^{n-1}\frac{\lambda-\lambda_i}{2\lambda_i}) \Bigg]\nonumber\\
=\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}\Bigg[ \frac{\lambda-\lambda_n}{2\lambda_n}-\frac{\lambda-\lambda_n}{(\frac{2}{\delta}+2)\lambda_n} +\frac{1}{1-\varepsilon}+\frac{m-1}{3-\varepsilon}
+\frac{\lambda-\lambda_{m+1}}{2\lambda_{m+1}}+...+\frac{\lambda-\lambda_{n-1}}{2\lambda_{n-1}}\Bigg]\nonumber\\
=\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}\Bigg[ \frac{1}{1-\varepsilon}+\frac{m-1}{3-\varepsilon}+\frac{\lambda}{2}(\frac{1}{\lambda_1}+..+\frac{1}{\lambda_1})-\frac{n}{2} \Bigg]-\delta\nonumber\\
=\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}\Bigg[ \frac{1}{1-\varepsilon}+\frac{m-1}{3-\varepsilon}+\frac{\lambda}{2}\frac{\sigma_{n-1}}{\sigma_n}-\frac{n}{2}
\Bigg]-\delta\nonumber\\
\geq\frac{2(1+\delta)\lambda_n}{\lambda-\lambda_n}\Bigg[ \frac{1}{1-\varepsilon}+\frac{m-1}{3-\varepsilon}-\frac{n}{2}\Bigg]
-\delta \nonumber
\end{align} where we used that the fact $\lambda_1=..=\lambda_m$ and Lemma \ref{y1}. Now as $\delta$ is arbitrarily small,
and $\lambda_n<0$ we choose $\varepsilon(n)>0$ such that
\begin{equation*}
\Bigg[ \frac{1}{1-\varepsilon}+\frac{m-1}{3-\varepsilon}-\frac{n}{2}\Bigg]\leq 0
\end{equation*} which in turn makes (\ref{q2}) positive.
On simplifying, we see that
\begin{equation*}
\varepsilon(n)\leq 2-\frac{m}{n}-\sqrt{(1-\frac{m}{n})^2+\frac{4}{n}}.
\end{equation*}
\item[Step 2.2] Now we estimate the term $Z_2$. For each $k$ between $m$ and $n$, we have $\lambda_k>0$, and the $[\hspace{.1cm}]_k$ in (\ref{y5.24}) satisfies
\begin{align*}
[\hspace{.1cm}]_k=\sum_{i\leq m}\Bigg[\frac{\lambda+\lambda_k}{\lambda-\lambda_k}-\varepsilon \Bigg]t_i^2\\
\geq \sum_{i\leq m}(1-\varepsilon)t_i^2\geq 0
\end{align*}
assuming $\varepsilon\leq 1$.\\
For $k=n$, the $[\hspace{.1cm}]_n$ term in (\ref{y5.24}) becomes \begin{align*}
[\hspace{.1cm}]_n=\sum_{i\leq m}\Bigg[\frac{\lambda+\lambda_k}{\lambda-\lambda_k}-\varepsilon \Bigg]t_i^2\\
\geq \sum_{i\leq m}\Bigg[ \frac{n-2}{n}-\varepsilon\Bigg]t_i^2\geq 0
\end{align*}
where the last inequality follows from Lemma \ref{y1} and the assumption of $\varepsilon\leq \frac{n-2}{n}.$
So far we have shown $Z\geq-C(||\psi||_{C^{1}(B_1)})$ for $n-1\geq m\geq 1$. When $m=n$, we have $\lambda_1=...=\lambda_n>0$ and therefore, $Z\geq-C(||\psi||_{C^{1}(B_1)})$ holds.
\item[Step 2.3] Next we estimate the remaining terms (\ref{11}) and (\ref{22})
\begin{align*}
m(\Delta_g b_m-\varepsilon(n)|\nabla_gb_m|^2)\geq \\
Z+\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}-m\sum_{i=1}^n\lambda_ig^{ii}\psi_i\partial_i b_m\geq\\
\geq-C(||\psi||_{C^{1}(B_1)})+\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}\psi_{ii}-\frac{\delta}{2}m^2|\nabla_g b_m|^2-\sum_{i=1}^m\frac{2}{\delta}g^{ii}\lambda^2_{i}\psi_i^2
\end{align*}
where the last inequality follows from Young's inequality and $Z\geq-C(||\psi||_{C^{1}(B_1)})$.
Denoting $c(n)=\varepsilon(n)+\delta/2$, we get
\begin{align}
\Delta_g b_m-c(n)|\nabla_gb_m|^2\nonumber\\ \geq-C(||\psi||_{C^{1}(B_1)})-|\sum_{i=1}^m\frac{\lambda_i}{1+\lambda_i^2}[\psi_{ii}-\frac{2}{\delta}\lambda_i\psi_i^2]|\nonumber\\
\geq -C(||\psi||_{C^{1,1}(B_1)},\delta,n)=-C. \label{C}
\end{align}
\end{itemize}
\end{proof}
\subsection{The integral Jacobi inequality}
In order to prove Hessian estimates we will need the following integral form of the Jacobi inequality (\ref{J}).
\begin{proposition}
Let $u$ be a smooth solution of (\ref{s}) on $B_R(0)\subset\mathbb{R}^{n}$ with
$\psi\geq (n-2)\frac{\pi}{2}+\delta$. Let
\begin{equation}
b=\ln\sqrt{1+\lambda_{\max}^2} \label{star}
\end{equation} where
$\lambda_{\max}$ is the largest eigenvalue of $D^2u$, namely, $\lambda_{\max}=\lambda_1\geq\lambda_2\geq...\geq\lambda_n$. Then for all non-negative $\phi\in C^{\infty}_0(B_R)$, $b$ satisfies the integral Jacobi inequality
\begin{equation}
\int_{B_R}-\langle \nabla_g \phi,\nabla_g b\rangle_g dv_g\geq c(n)\int_{B_R}\phi|\nabla_gb|^2dv_g-\int_{B_R}C\phi dv_g \label{IJ}
\end{equation}
where $C$ and $c(n)$ are the constants from (\ref{J}).
\end{proposition}
\begin{proof}
If $b_1=\ln\sqrt{1+\lambda_{\max}^2} $ is smooth everywhere, then the pointwise Jacobi inequality (\ref{J}) implies the integral Jacobi inequality (\ref{IJ}). It is
easy to see that $\lambda_{max}$ is always a Lipschitz function of the entries of the Hessian of $u$.
Since $u$ is smooth in $x$, $b$ is Lipschitz in terms of $x$. We now show that (\ref{J}) holds in the viscosity sense. \\
Let $x_0\in B_R(0)$ and let $P(x)$ be a quadratic polynomial such that
\[P(x)\geq b(x) \text{ with equality holding at $x_0$}.
\]
Now if $x_0$ is a smooth point of $b$, then from (\ref{J}), we see that at $x_0$, with $m=1$, the following holds
\[\Delta_g P\geq c(n)|\nabla_gP|^2-C(||\psi||_{C^{1,1}(B_1)},\delta,n).
\]
Or else we would have $m>1$, i.e. $\lambda_1$ is not distinct at $x_0$. Let's suppose that we have
$\lambda_1=...=\lambda_k>\lambda_{k+1}$ at $x_0$. Consider the function $b_k=\frac{1}{k}\sum_{i=1}^k \ln \sqrt{1+\lambda_i^2}$. Note that this is smooth near $x_0$ from Lemma \ref{4.2}. Observe that since
$b(x)\geq b_k(x)$ with equality holding at $x_0$, we must have
\[P(x)\geq b_k(x) \text{ with equality holding at $x_0$}.
\]
So now applying the pointwise Jacobi inequality (\ref{J}) to $b_k$, we see the following holds at $x_0$
\[\Delta_g P\geq c(n)|\nabla_gP|^2-C(||\psi||_{C^{1,1}(B_1)},\delta,n).
\] So far we have shown that (\ref{J}) holds in the viscosity sense. Applying Ishii's result \cite[Theorem 1]{Ish} we see that the viscosity subsolution $b$ of (\ref{J}) should also be a distribution subsolution. Now since $b$ is a Lipschitz function we perform integration by parts on the first term to get
\[\int_{B_R}\Delta_g b \phi dv_g=\int_{B_R}-\langle \nabla_g \phi,\nabla_g b\rangle_g dv_g.
\]
On combining \cite[Theorem 1]{Ish} and the above equation, we get $(\ref{IJ})$.
\end{proof}
\section{Hessian Estimates}
\subsection{Mean Value Inequality}
In \cite[Theorem 3.4]{MS}, Michael-Simon established the mean value inequality for a non-negative subharmonic function on a $m$-dimensional submanifold $M$ of $\mathbb{R}^n$. In order to prove the Hessian estimates, having the following mean value inequality for a variable Lagrangian phase is crucial. For completeness, we include a proof here.
\begin{proposition}
Let $M\subset\mathbb{R}^{n}$ be a $m$-dimensional submanifold, $0\in M$, and $s>0$ satisfies $B_s(0)\cap\partial M=\emptyset$. Suppose that there exists $\Lambda_0>0$ such that $|\vec{H}|\leq \Lambda_0$ where $\vec{H}$ is the mean curvature vector of $M$. If $f$ is a non-negative function on $M$ such that $\Delta_{g}f\geq-\beta f$, then
\begin{equation}
f(0)\leq C(\beta,n,\Lambda_0)\frac{\int_{B_1\cap M}f dv_g}{Vol(B_1\subset\mathbb{R}^{n})}. \label{mvi}
\end{equation}
\end{proposition}
\begin{proof}
The symmetric matrix $\tilde{g}^{ij}(x)$ denotes the projection of $\mathbb{R}^{n}$ onto the $m$-dimensional submanifold $M$, and it satisfies
\begin{align}
\sum_{i=1}^{n}\tilde{g}^{ii}(x)=m \label{gec}\\
0\leq\sum_{i,j=1}^{n}\tilde{g}^{ij}(x)x_ix_j\leq|x|^2 \hspace{.2cm}\forall x\in \mathbb{R}^{n} \nonumber.
\end{align} Let $U$ denote an open subset of $\mathbb{R}^{n}$ that contains $M$.
Let $\phi$ be a non decreasing $C^1(\mathbb{R})$ function such that $\phi(t)=0$ when $t\leq 0$. For each $x_0\in M$, we define the following two functions
\begin{align*}
g_0(y)=\int_M f(x)\phi(y-r)dv_g(x)\\
h_0(y)=\int_Mf(x)|H(x)|\phi(y-r)dv_g(x)
\end{align*}
where $r=|x-x_0|$. Let's assume $d(x_0,\partial U)=1.$
\begin{claim}
For $0<y\leq 1$, we prove that
\begin{equation}
-\frac{d}{dy}\left[ \frac{\ g_0(y)}{y^m}\right]\leq y^{-m-1}\int_{0}^{y}t h'_0(t)dt. \label{main claim}
\end{equation}
\end{claim}
Let $\gamma$ be a real valued function defined by
\[\gamma(s)=\int_s^{\infty}t\phi(y-t)dt.
\]
Then $\gamma(s)=0$ when $s\geq y$. Note that $\gamma(|x-x_0|)$ is in $C^2(U)$ and vanishes outside a compact subset of $U$ if $y<1$. So for $y<1$ we can use $\gamma(r)$ as a test function. Next we observe that
\begin{align*}
\Delta_g \gamma(r)=-\left[\phi(y-r)\sum_{i=1}^{n}\tilde{g}^{ii}-r\phi'(y-r)\sum_{i,j=1}^{n}\tilde{g}^{ij}(\frac{x_i-(x_0)_i}{r})(\frac{x_j-(x_0)_j}{r}) \right]-\\
\phi(y-r)\sum_{i=1}^{n}(x_i-(x_0)_i)H_i
\end{align*} where $H_i$'s denote the components of the mean curvature vector. So then by our assumption $\Delta_{g}f\geq-\beta f$ and (\ref{gec}) we have
\begin{align*}
mg_0(y)-\int_M fr\phi'(y-r)dv_g(x)\leq \int_M f|H|r\phi(y-r)dv_g+\int_M \beta f\gamma(r)dv_g\\
\implies mg_0(y) -\int_M fr\phi'(y-r)dv_g(x)\leq \int_M f|H|r\phi(y-r)dv_g+\int_M \beta f y\phi(y-r) dv_g.
\end{align*}
In the last inequality we used $\gamma(r)\leq \int _{r}^1 t\phi(y-t)dt\leq y\phi(y-r)$ since we need $t\leq y$ for the function to be non-zero.
This gives us
\begin{equation}
mg_0(y)-\int_M fr\phi'(y-r)dv_g(x)\leq \int_M f|H|r\phi(y-r)dv_g+\int_M \beta f y\phi(y-r) dv_g. \label{imp1}
\end{equation}
Using the inequality $r\phi'(y-r)\leq y\phi'(y-r)$ we get
\[\int_M fr\phi'(y-r)dv_g\leq yg_0'(y).
\]
We see that
\begin{align*}
\int_M f|H|r\phi(y-r)dv_g =\int_M f|H|\left[ \int_0^y r\phi'(t-r)dt \right]dv_g \\
\leq \int_M f|H|\left[ \int_0^y t\phi'(t-r)dt \right]dv_g\\
=\int_0^y th_0'(t)dt.
\end{align*}
Therefore, (\ref{imp1}) reduces to
\[mg_0(y)-\int_M fr\phi'(y-r)dv_g(x)\leq \int_0^y th_0'(t)dt+\int_M \beta f y\phi(y-r) dv_g
\] which can be written as
\[-\frac{d}{dy}\left[ \frac{g_0(y)}{y^m}\right]\leq y^{-m-1}\int_0^yth_0'(t)dt+y^{-m}
\int_M \beta f \phi(y-r) dv_g.
\]
Using the fact
\[\int_0^yth_0'(t)dt\leq y\int_0^yh'_0(t)dt=yh_0(y)
\]
we get
\[-\frac{d}{dy}\left[ \frac{g_0(y)}{y^m}\right]\leq \frac{s_0(y)}{y^m}.
\]
Observe that
\begin{align*}
s_0=\int_Mf(x)|H(x)|\phi(y-r)dv_g(x)+\int_M \beta f \phi(y-r) dv_g\\
\leq C\int_M f\phi(y-r)dv_g(x)
\end{align*}
where $C=(\beta,n,\Lambda_0)$.
This shows
\[-\frac{d}{dy}\left[ \frac{g_0(y)}{y^m}\right]\leq C\left[ \frac{g_0(y)}{y^m}\right].
\]
Integrating this we get
\[\sup_{t\in (0,y)}\left[ \frac{g_0(t)}{t^m}\right]\leq e^{Cy}\left[ \frac{g_0(y)}{y^m}\right]
\] for all $y\in (0,1)$. Now we choose $\phi$ such that $\phi(s)=1$ when $s\geq \varepsilon$ and letting $\varepsilon\rightarrow 0+$, (\ref{mvi}) follows.
\end{proof}
\begin{remark} The mean value inequality for $b$ defined in (\ref{star}) also follows from the observation that the potential $u$ is semi convex, so a rotation of Yuan \cite[Pg 125]{YY02} can be performed on
the gradient graph $(x,Du(x))$, which results in a uniformly elliptic Laplace-Beltrami operator on the rotated graph $(\bar x,D\bar u(\bar x))$.
Then on applying the De Giorgi iteration for divergence form equations \cite[Pg 197, (8.58)]{GT} to $b$, and given the invariance of the Jacobi inequality and integral, we obtain a MVI for it.
\end{remark}
Note that the condition $u$ is smooth in section 4 can be clearly replaced by $u\in C^4$. Now we prove our main Theorem.
\subsection{Proof of Theorem \ref{main1}}
\begin{proof}
We first consider the case $n\neq 2$. We verify that the positive function $b$ defined in (\ref{star}) satisfies the requirements of the above mean value inequality. Observe that by our choice of $b$, we have
\begin{equation}
b\geq \ln\sqrt{1+\tan^2(\frac{\pi}{2}-\frac{\pi}{n})}\geq \ln \sqrt{4/3} \label{b}.
\end{equation}
Combining the above with (\ref{J}), we conclude that $b$ satisfies the conditions of the above Theorem:
\begin{align*}
\Delta_g b\geq c(n)|\nabla_gb|^2-C
\geq -\frac{C}{\ln \sqrt{4/3}}b=-Cb
\end{align*}
where $C=C(n,||\psi||_{C^{1,1}(B_1)},\delta, \ln \sqrt{4/3})$ is the positive constant from (\ref{IJ}).
For simplifying notation in the remaining proof, we assume $R=2n+1$ and $u$ is a solution on $B_{2n+1}\subset\mathbb{R}^n$. Then by scaling $v(x)=\frac{u(\frac{R}{2n+1}x)}{(\frac{R}{2n+1})^2}$, we get the estimate in Theorem \ref{main1}. Also, we will denote the above constant $C$ as $C(n,\psi,\delta)$.
\begin{itemize}
\item[Step 1.] We show that the function $b^{\frac{n}{n-2}}$ meets the requirements of the above MVI:
\begin{align*}
\int-\langle\nabla_g \phi,\nabla_g b^{\frac{n}{n-2}}\rangle_g dv_g\\
=\int-\langle \nabla_g(\frac{n}{n-2}b^\frac{2}{n-2}\phi)-\frac{2n}{(n-2)^2}b^{\frac{4-n}{n-2}}\phi\nabla_gb,\nabla_g b \rangle_g dv_g\\
\geq\int (\frac{n}{n-2}c(n)b^{\frac{2}{n-2}}\phi|\nabla_g b|^2+\frac{2n}{(n-2)^2}b^{\frac{4-n}{n-2}}\phi|\nabla_g b|^2)dv_g-\int C\frac{n}{n-2}\phi b^{\frac{2}{n-2}}dv_g \\
\geq -\int C\frac{n}{n-2}\frac{1}{b}b^{\frac{n}{n-2}} \phi dv_g \\
\geq -C(n,\psi,\delta)\int b^{\frac{n}{n-2}} \phi dv_g
\end{align*}
where the last inequality follows from (\ref{b}).
So now by the MVI applied to the Lipschitz function $b^\frac{n}{n-2}$ we get
\begin{equation}b(0)\leq C(n,\psi,\delta)(\int_{\Tilde{B_1}\cap X}b^{\frac{n}{n-2}}dv_g)^{\frac{n-2}{n}}
\leq C(n,\psi,\delta)(\int_{B_1}b^{\frac{n}{n-2}}dv_g)^{\frac{n-2}{n}}\label{bbb}
\end{equation}
where $X=(x,Du(x))\subset \mathbb{R}^n\times\mathbb{R}^n$ is the Lagrangian submanifold, $\Tilde{B_1}$ is the ball with radius $1$ and center at $(0,Du(0))$ in $\mathbb R^n\times \mathbb R^n$, and $B_1$ is
the ball with radius $1$ and center at $0$ in $\mathbb R^n.$
Choose a cut off $\phi\in C^\infty_0(B_2)$ such that $\phi\geq 0$, $\phi=1$ on $B_1$ and $|D\phi|<2$. That gives us
\begin{equation}
[\int_{B_1}b^{\frac{n}{n-2}}dv_g]^{\frac{n-2}{n}}\leq [\int_{B_2}\phi^{\frac{2n}{n-2}} b^{\frac{n}{n-2}}dv_g]^{\frac{n-2}{n}}=[\int_{B_2}(\sqrt{b}\phi) ^{\frac{2n}{n-2}} dv_g]^{\frac{n-2}{n}} \label{b1}.
\end{equation}
We assume $\sqrt{b}\phi$ to be $C^1$ by approximation. Applying the general Sobolev inequality \cite[Theorem 2.1]{MS} to it on this Lagrangian submanifold, and using the mean curvature formula (\ref{mean}), we get
\begin{align}
[\int_{B_2}(\sqrt{b}\phi )^{\frac{2n}{n-2}}dv_g]^{\frac{n-2}{n}}\leq C(n)[\int _{B_2}|\nabla_g(\sqrt{b}\phi )|^2dv_g+\int_{B_2}|\sqrt{b}\phi \nabla_g \psi|^2dv_g]. \label{b2}
\end{align}
Next we observe the following
\begin{align}
|\nabla_g(\sqrt{b}\phi )|^2=|\frac{1}{2\sqrt{b}}\phi\nabla_g b+\sqrt{b}\nabla_g\phi|^2\nonumber\\
\leq \frac{1}{2b}\phi^2|\nabla_g b|^2+2b|\nabla_g\phi|^2\nonumber\\
\leq \phi^2|\nabla_g b|^2+2b|\nabla_g \phi|^2.\label{b3}
\end{align}
Combining (\ref{b1}, \ref{b2}, \ref{b3}) and plugging into (\ref{bbb}), we see that
\begin{align}
b(0)\leq C(n,\psi,\delta)[\int_{B_2}\phi^2|\nabla_g b|^2 dv_g+\int_{B_2}b|\nabla_g \phi|^2dv_g
+\int_{B_2}\phi^2b|\nabla_g \psi|^2dv_g].\label{III}
\end{align}
\item[Step 2.]
Using the integral Jacobi inequality and recalling the constants $C$ and $c(n)$ from (\ref{J}) we get
\begin{align}
\int_{B_2}\phi^2|\nabla_gb|^2dv_g\leq \frac{1}{c(n)}[\int_{B_2}\phi^2\Delta_gbdv_g+\int_{B_2}\phi^2C dv_g] \nonumber \\
=-\frac{1}{c(n)}[\int_{B_2}\langle2\phi\nabla_g\phi,\nabla_gb\rangle dv_g+\int_{B_2}\phi^2Cdv_g] \nonumber\\
\leq \frac{1}{2}\int_{B_2}\phi^2|\nabla_gb|^2dv_g+\frac{2}{c(n)^2}\int_{B_2}|\nabla_g\phi|^2dv_g+\frac{1}{c(n)}\int_{B_2}\phi^2Cdv_g \nonumber\\
\implies \int_{B_2}\phi^2|\nabla_gb|^2dv_g\leq
\frac{4}{c(n)^2}\int_{B_2}|\nabla_g\phi|^2dv_g+\frac{2}{c(n)}\int_{B_2}\phi^2Cdv_g. \label{ii}
\end{align}
Again using (\ref{b}) and plugging the above inequality in (\ref{III}) we get the following
\begin{align}
b(0)\leq
C(n,\psi, \phi,\delta)[\int_{B_2}b\sum_{i=1}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx+\int_{B_2}\sqrt{\det g}dx
].\label{pp2}
\end{align}
Next we choose a suitable test function $\eta$ such that on using the Sobolev inequality we get the following estimate
\begin{align*}
\int_{B_{2}}\sqrt{\det g}dx\leq C(n)[\int_{B_{2}}\eta ^{\frac{2n}{n-2}}dv_g]^{\frac{n-2}{n}}\\
\leq C(n)[\int_{B_{2}} |\nabla_g\eta|^2dv_g+\int_{B_{2}}|\eta\nabla_g\psi|^2dv_g]\\
\leq C(n)[\int_{B_{2}} \sum_{i}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx+ \int_{B_{2}}|\eta\nabla_g\psi|^2dv_g ].
\end{align*}
Using (\ref{b}) we get
\begin{align*}
\int_{B_{2}}\sqrt{\det g}dx\leq C(n,\psi,\eta)\int_{B_{2}}b \sum_{i}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx .
\end{align*}
On combining everything and plugging in (\ref{pp2}), we get
\begin{align}
b(0)\leq C(n,\psi,\delta)\int_{B_{2}} b\sum_{i}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx.\label{b12}
\end{align}
\item[Step 3.] Now we estimate $\int_{B_2}b\sum_{i=1}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx$.\\
We denote $V=\sqrt{\det g}$. We see that (ref: \cite [(3.2)]{WaY}) on differentiating the complex identity
\[\ln V+i\sum_{i=1}^n\arctan\lambda_i=\ln \bigg[ \sum_{0\leq 2k\leq n} (-1)^k\sigma_{2k} +i\sum_{1\leq2k+1\leq n}(-1)^k \sigma_{2k+1}\bigg]
\] we get
\[ \bigg( \frac{1}{1+\lambda_1^2},...,\frac{1}{1+\lambda_n^2}\bigg)V=\bigg( \frac{\partial \Sigma}{\partial \lambda_1},...,\frac{\partial \Sigma}{\partial \lambda_n}\bigg)
\]
where \begin{align*}
\sum=\cos\psi \sum_{1\leq 2k+1\leq n}(-1)^k\sigma_{2k+1}-\sin\psi \sum_{0\leq 2k\leq n}(-1)^k\sigma_{2k}
\end{align*}
where $\psi$ is the Lagrangian phase. On taking the trace, we get
\begin{align*}
\sum_{i=1}^n \frac{1}{1+\lambda_i^2}V=\sum_{i=1}^n\frac{\partial \Sigma}{\partial \lambda_i}\\
=\cos\psi \sum _{1\leq 2k+1\leq n}(-1)^k(n-2k)\sigma_{2k}-\sin\psi\sum_{0\leq2k\leq n}(-1)^k(n-2k+1)\sigma_{2k-1}\\
=c_0(x)+c_{1}(x)\sigma_1+...+c_{n-1}(x)\sigma_{n-1}
\end{align*}
where the variable coefficient $c_i$ now depends on $i,n,\psi$ for all $1\leq i\leq n-1$.
Hence, (\ref{b12}) becomes
\begin{align*}
\int_{B_2}b\sum_{i=1}^{n}\frac{1}{1+\lambda_i^2}\sqrt{\det g}dx\leq C(n,\psi,\delta)\int_{B_2}b(c_0(x)+c_1(x)\sigma_1+...+c_{n-1}(x)\sigma_{n-1})dx.
\end{align*}\\
\item[Step 4.] We next estimate the integrals $\int b\sigma_k dx$ for $1\leq k\leq n-1$ inductively, using the divergence structure of $\sigma_k(D^2u)$. Let $L_{\sigma_k}$ denote the matrix $(\frac{\partial \sigma_k}{\partial u_{ij}})$, then we see that
\begin{align*}
c_k k\sigma_k(D^2u)=c_k\sum_{i,j=1}^n\frac{\partial \sigma_k}{\partial u_{ij}}\frac{\partial^ 2u}{\partial x_i \partial x_j}=c_k\sum_{i,j=1}^n\frac{\partial}{\partial x_i}\left[ \frac{\partial \sigma_k}{\partial u_{ij}}\frac{\partial u}{\partial x_j}\right]\\
=c_kdiv(L_{\sigma_k}Du)=div(c_kL_{\sigma_k}Du)-Dc_k\cdot L_{\sigma_k}Du.
\end{align*}
Let $v$ be a smooth cut-off function on $B_{r+1}$ such that $v=1$ on $B_r$, $0\leq v\leq 1$ and $|Dv|<2$. Note that by Lemma \ref{y1}, we have $\sigma_k>0$ and from (\ref{b}) we have $b>\ln\sqrt{4/3}$, which together imply $c_k>0$. So we have
\begin{align}
\int_{B_r}c_kb\sigma_kdx \leq \int_{B_{r+1}}c_kvb\sigma_k dx \nonumber\\
=\int_{B_{r+1}}vb\frac{1}{k} [div(c_kL_{\sigma_k}Du)-Dc_k\cdot L_{\sigma_k}Du]dx \nonumber\\
=\frac{1}{k}\int_{B_{r+1}}[-\langle bDv+vDb, c_k L_{\sigma_k}Du\rangle -vbDc_k\cdot L_{\sigma_k}Du]dx \nonumber \\
\leq C(n,\psi,\phi)||Du||_{L^{\infty}(B_{r+1})}\{ \int_{B_{r+1}} b\sigma_{k-1}dx+\nonumber \\
\int_{B_{r+1}}(|\nabla_gb|^2+tr(g^{ij}))\sqrt{\det g}dx \} \label{imp}
\end{align}
where the last inequality follows from the argument used in \cite[(3.6)]{WaY}.\\
We use the integral Jacobi inequality (\ref{IJ}) to simplify the last integral.
Combining (\ref{ii}) with (\ref{pp2}) we have
\[\int_{B_2}\phi^2|\nabla_gb|^2dv_g\leq C(n,\psi,\delta)\int_{B_{r+2}}tr(g^{ij})\sqrt{\det g}dx.\]
On rearranging constants, (\ref{imp}) reduces to the following inductive inequality
\[\int_{B_r}c_kb\sigma_kdx \leq C(n,\psi,\delta)||Du||_{L^{\infty}(B_{r+1})}\{ \int_{B_{r+1}} b\sigma_{k-1}dx+\int_{B_{r+2}}tr(g^{ij})\sqrt{\det g}dx\}.
\]
Now applying the argument used in \cite[(3.7)]{WaY} and the trace-conformality identity we conclude that
\[b(0)\leq C(n,\psi,\delta)\left[ ||Du||_{L^{\infty}(B_{2n+1})}+||Du||^{2n-2}_{L^{\infty}(B_{2n+1})}\right].
\]
On exponentiating we get
\[ |D^2 u(0)|\leq C_1\exp[C_2||Du||^{2n-2}_{L^{\infty}(B_{2n+1})}]
\]
where $C_1$ and $C_2$ are positive constants depending on $||\psi||_{C^{1,1}}$, $n$, and $\delta$.
\end{itemize}
Next, we consider the case $n=2$: we may fix $\arctan\lambda_3=\pi/2-\delta/2$ and add $(\pi/2-\delta/2)$ to both sides of the two dimensional supercritical equation (\ref{s}) to get:
\[\arctan\lambda_1+\arctan\lambda_2+\arctan\lambda_3=\psi(x)+\frac{\pi}{2}-\frac{\delta}{2}\geq\frac{\pi}{2}+\frac{\delta}{2}.
\]
This again brings us to the three dimensional supercritical equation (\ref{s}) for which Hessian estimates hold good by the above proof.
This completes the proof of Theorem \ref{main1}.
\end{proof}
\textbf{Acknowledgments.} The author is grateful to Yu Yuan for his guidance, support, and many useful discussions. The author is grateful to Ravi Shankar and Micah Warren for helpful comments.
\bibliographystyle{unsrt}
| 79f79fbb19d8990984c70fa2043d752e91be1e86 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction: Hospital Readmissions and the Pitfalls of Risk-Based Targeting}
Unplanned hospital readmissions represent an undesirable outcome following a hospitalization, but are common, costly, and associated with substantial morbidity and mortality, occurring within 30 days following nearly 20\% of hospitalizations by Medicare beneficiaries \cite{Jencks2009RehospitalizationsProgram}. In 2011, 3.3 million patients in the United States were readmitted to the hospital within 30 days, incurring costs of \$41 billion \cite{Hines2011Conditions2011}. In 2012, responding the growing awareness of the toll of readmissions, the Centers for Medicare and Medicaid Services introduced the Hospital Readmissions Reduction Program (HRRP), which penalizes hospitals with risk-adjusted 30-day readmission rates higher than the average. As a consequence of the HRRP and other value-based care initiatives, many hospitals and health care systems in the United States have since implemented quality improvement (QI) initiatives and population health management programs that rely on risk assessment tools to identify hospitalized patients at high risk of readmission. Tailored interventions can be then targeted to these patients immediately following discharge, with the goal of preventing their readmission. The effectiveness of these interventions in preventing readmissions has been mixed, and the precise mechanisms through which they do so remain unclear. \cite{Leppin2014PreventingTrials,Hansen2011InterventionsReview,Wadhera2018AssociationPneumonia,Kansagara2016SoLiterature,Finkelstein2020HealthTrial,Bates2014BigPatients,Berkowitz2018AssociationJ-CHiP}
Many risk assessment tools used in these efforts apply statistical modeling or supervised machine learning to estimate readmission risk among hospitalized patients based on data prior to discharge. \cite{Kansagara2011RiskReview.,Bayati2014Data-drivenStudy.,Escobar2015NonelectiveMortality,Billings2006CasePatients,Berkowitz2018AssociationJ-CHiP,Bates2014BigPatients} Stakeholders select a risk threshold with respect to resource constraints, which underpins objective criteria (often referred to in the machine learning literature as a \textit{treatment policy}) that determine which patients are selected to receive the intervention. Common threshold-based criteria specify that an intervention is to be delivered to all patients above a prespecified risk threshold, while those below it receive usual care. Underlying many population health management and QI efforts aimed at reducing readmissions is the implicit assumption that the patients most at risk are also those most likely to benefit from the intervention. \cite{Fihn2014InsightsAdministration,Bates2014BigPatients,HealthITAnalytics2016UsingHealthitanalytics.com/features/using-risk-scores-stratification-for-population-health-management,HealthITAnalytics2018TopHttps://healthitanalytics.com/news/top-4-big-data-analytics-strategies-to-reduce-hospital-readmissions} Ostensibly, this assumption has intuitive appeal, given that higher-risk patients appear to have "more room to move the needle", but it is not guaranteed to hold in practice \cite{Ascarza2018RetentionIneffective, Athey2017BeyondProblems}, especially in the context of readmissions \cite{Finkelstein2020HealthTrial,Lindquist2011UnderstandingFactors} and other settings where treatment effect heterogeneity may be present \cite{Athey2017BeyondProblems}.
The need for analytical approaches that estimate patient-level benefit---referred to in some contexts as \textit{impactibility} \cite{Lewis2010ImpactibilityPrograms,Freund2011IdentificationPrograms,Steventon2017PreventingRisk,Flaks-Manov2020PreventingPrediction}---is beginning to be recognized, particularly for readmission reduction programs \cite{Steventon2017PreventingRisk}. However, the distinction between benefit and risk does not yet appear to be widely appreciated by both those developing and applying risk assessment tools. Individual benefit is often expressed in terms of treatment effects, which cannot be estimated by modelling outcome risk. Predicting, for example, a readmission risk of 60\% for a patient provides no information on their counterfactual risk if they were to receive a readmissions reduction intervention. The actual counterfactual risk for this hypothetical patient could be unchanged, on average, corresponding to no effect for the intervention. On the other hand, the effect of this intervention may be heterogeneous across levels of predicted risk, so that, for example, this patient experiences an absolute risk reduction (ARR) of 10\% as a result of the intervention, while another patient at a predicted risk of 30\% experiences an ARR of 20\%. Given limited resources, a decision-maker may wish to give the intervention to the latter patient. Indeed, when it comes to preventing readmissions, there is growing evidence that higher-risk patients---referred to in some contexts as "super-utilizers" \cite{Finkelstein2020HealthTrial}---may be less sensitive to a class of care coordination interventions relative to those at lower risk \cite{Steventon2017PreventingRisk,Lindquist2011UnderstandingFactors,Rich1993PreventionStudy.}.
Moreover, efforts targeting preventative interventions based on predicted risk also fail to take into account that low-risk patients comprise the majority of readmissions \cite{Roland2012ReducingTrack}. That the majority of poor outcomes are experienced by patients at low risk, but who would not have been selected to receive an intervention, is an observation which has also surfaced in a range of predictive modeling problems in population health management \cite{Bates2014BigPatients}. Thus, targeting preventative interventions so as to include lower-risk patients among whom they may be effective, rather than targeting them only to high-risk patients, may potentially prevent more readmissions than the latter strategy \cite{Rose1985SickPopulations,Chiolero2015TheStrategy, McWilliams2017FocusingCosts}. However, even given a hypothetical ideal intervention---one guaranteed to be effective for \textit{all} patients in a population---staffing and other resource constraints may preclude scaling up such an intervention to an entire population. Hence, in order to maximize the net benefit of an intervention, "decoupling" the prediction problem into causal and predictive components---modeling not just heterogeneity in treatment effects, but also the (possibly heterogeneous) "payoffs" associated with successful prevention---may be necessary \cite{Kleinberg2015PredictionProblems}.
Few, if any, analytical approaches to identify "care-sensitive" patients, or those whose outcomes may be most "impactible", currently exist, despite the clear need for such approaches. \cite{Lewis2010ImpactibilityPrograms,Flaks-Manov2020PreventingPrediction} Existing approaches based on off-the-shelf supervised machine learning methods, despite their flexibility and potential predictive power, cannot meet this need. \cite{Athey2017BeyondProblems} In this paper, we propose and demonstrate the feasibility of a causal machine learning framework to identify preventable hospital readmissions with respect to a readmission prevention intervention. In our context, the "preventability" of a readmission is not based on predefined, qualitative criteria, as in prior work (e.g., \cite{Goldfield2008IdentifyingReadmissions,Auerbach2016PreventabilityPatients}). Rather, it is expressed in quantitative terms: the greater the treatment effect on readmission estimated for a patient, the more preventable their potential readmission may be.
To do so, we leverage a rich set of data drawn from before and after the roll-out of a comprehensive readmissions prevention intervention in an integrated health system, seeking to: (1) estimate the heterogeneity in the treatment effect of this intervention; (2) characterize the potential extent of mismatch between treatment effects and predicted risk; and (3) quantify the potential gains afforded by targeting based on treatment effects or benefit instead of risk. Finally, based on our findings, we also outline some possible directions for how population health management programs could be redesigned so as to maximize the aggregate benefit, or overall impact, of these preventative interventions.
\section{Methods}
\subsection{Data and Context}
The data consist of 1,584,902 hospitalizations taking place at the 21 hospitals in Kaiser Permanente's Northern California region (hereafter KPNC) between June 2010 and December 2018. They include patient demographics, diagnosis codes, laboratory-based severity of illness scores at admission and at discharge, and a comorbidity burden score that is updated monthly. The data also record whether a patient experienced a non-elective re-admission and/or death within 30 days. These data are described in greater detail in \cite{Escobar2019MultiyearSystem}.
These data encompass a period where a comprehensive readmissions prevention intervention, known as the \textit{Transitions Program}, began and completed implementation at all 21 KPNC hospitals from January 2016 to May 2017. The Transitions Program had two goals: (1) to standardize post-discharge care by consolidating a range of preexisting care coordination programs for patients with complex care needs; and (2) to improve the efficiency of this standardized intervention by targeting it to the patients at highest risk of the composite outcome of post-discharge re-admission and/or death.
As currently implemented, the Transitions Program relies on a validated predictive model for the risk of this composite outcome \cite{Escobar2015NonelectiveMortality}, which was developed using historical data from between June 2010 and December 2013, before the implementation of the Transitions Program. Following development and validation of this model by teams at KPNC's Division of Research, it was subsequently integrated into KP HealthConnect, KPNC's electronic health record (EHR) system to produce continuous risk scores, ranging from 0 to 100\%, at 6:00 AM on the planned discharge day. These risk scores are used to automatically assign inpatients awaiting discharge to be followed by the Transitions Program over the 30-day period post-discharge. Inpatients with a predicted risk of $\geq\!\!25$ are assigned to be followed by the Transitions Program, and are considered to have received the Transitions Program intervention. On the other hand, inpatients with a predicted risk below 25\% receive usual post-discharge care at the discretion of the discharging physician.
The Transitions Program intervention consists of a bundle of discrete interventions aimed at improving the transition from inpatient care to home or a to a skilled nursing facility. Together, they comprise an interlocking care pathway over the 30 days following discharge, beginning on the morning of the planned discharge day, and which we summarize step-by-step here and in Table \ref{tab:transitions}.
\begin{table}[]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}cccccc@{}}
\toprule
Risk Level & Initial Assessment & Week 1 & Week 2 & Week 3 & Week 4 \\ \midrule
\begin{tabular}[c]{@{}c@{}}High\\ ($\geq 45$\%)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Phone follow-up \\ within 24 to 48 hours \\ \\ \textit{and}\end{tabular} & \begin{tabular}[c]{@{}c@{}}Phone follow-up\\ every other day\end{tabular} & \begin{tabular}[c]{@{}c@{}}2 phone follow-ups; \\ more as needed\end{tabular} & \begin{tabular}[c]{@{}c@{}}Phone follow-up \\ once weekly; \\ more as needed\end{tabular} & \begin{tabular}[c]{@{}c@{}}Phone follow-up \\ once weekly; \\ more as needed\end{tabular} \\ \cmidrule(r){1-1} \cmidrule(l){3-6}
\begin{tabular}[c]{@{}c@{}}Medium\\ ($25-45$\%)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Primary care physician follow-up visit\\ within 2 to 5 days\end{tabular} & \multicolumn{4}{c}{\begin{tabular}[c]{@{}c@{}}Once weekly phone follow-up \\ (with more as needed)\end{tabular}} \\ \midrule
\begin{tabular}[c]{@{}c@{}}Low \\ ($\leq 25$\%)\end{tabular} & \multicolumn{5}{c}{Usual care at discretion of discharging physician} \\ \bottomrule
\end{tabular}%
}
\caption{\footnotesize The Transitions Program intervention pathway. The initial assessment applies to both the medium and high risk groups. Following it, the pathway diverges in terms of the frequency of phone contact.}
\label{tab:transitions}
\end{table}
For inpatients awaiting discharge and who have been assigned to be followed by the Transitions Program, on their planned discharge day, a case manager meets with them at the bedside to provide information on the Transitions Program. Next, once the patient has arrived home, a Transitions case manager calls them within 24 to 48 hours to walk them through their discharge instructions, and to identify any gaps in their understanding of them. If necessary, the case manager can also refer the patient to a pharmacist or social worker for focused follow-up. At the same time, the nurse also works to make an appointment with the patient's primary care physician to take place within 3 to 5 days post-discharge.
Following this initial outreach, the Transitions case manager continues to contact the patient weekly by phone, and remains available throughout if the patient requires further assistance. At 30 days post-discharge, the patient is considered to have "graduated" and is no longer followed by the Transitions Program. All steps of this process are initiated through and documented in the EHR, enabling consistent followup for the patients enrolled. A special category of patients are considered at very high risk if their predicted risk is $\geq \!\! 45$\% or if they lack social support, and receive a more intensified version of the intervention. This version entails follow-up every other day via telephone for the first week post-discharge, followed by $\geq \!\! 2$ times a week the second week, and once a week afterward until "graduation" at 30 days.
Initial analyses of the impacts of the Transitions Program \cite{Marafino2020ASystem}, using a hybrid difference-in-differences analysis/regression discontinuity approach \cite{Walkey2020NovelInitiative}, indicated that it was effective, being associated with approximately 1,200 and 300 fewer annual readmissions and deaths, respectively, within 30 days following their index discharge. Notably, these analyses also suggested some extent of risk-based treatment effect heterogeneity, in that the intervention appeared to be somewhat less effective for patients at higher risk compared to those at relatively lower risk.
As a consequence of the conclusions of these analyses, two questions presented themselves. The first was whether the Transitions Program intervention could be re-targeted more efficiently, leading to greater aggregate benefit, possibly by including patients at lower risk who would not otherwise have received the intervention. The second arose out of the risk-based treatment effect heterogeneity suggested by these analyses---may the intervention in fact be less effective for patients at higher risk? If so, what is the extent of the mismatch between risk and treatment effects? To answer these questions requires modeling individual heterogeneous treatment effects, and not risk, as is commonly done.
In this paper, we use the same data on 1,584,902 patients as that used for the analysis of the effectiveness of the Transitions Program. As in that analysis, we use a subset of 1,539,285 "index stays" which meet a set of eligibility criteria. These criteria include: the patient was discharged alive from the hospital; age $\geq \!\! 18$ years at admission; and their admission was not for childbirth (although post-delivery complications were included) nor for same-day surgery. Moreover, we consider a readmission non-elective if it began in the emergency department; if the principal diagnosis was an ambulatory care-sensitive condition \cite{AgencyforHealthcareResearchandQuality2001AHRQConditions.}; or if the episode of care began in an outpatient clinic, and the patient had elevated severity of illness, based on a mortality risk of $\geq\!\!7.2$\% as predicted by their laboratory-based acuity score (\texttt{LAPS2}) alone. The covariates used in this study (as input to the causal forest below) are summarized in Table \ref{tab:table1}.
This project was approved by the KPNC Institutional Review Board for the Protection of Human Subjects, which has jurisdiction over all the study hospitals and waived the requirement for individual informed consent.
\begin{table}
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{ll}
\toprule
\texttt{AGE} & Patient age in years, recorded at admission \\
\texttt{MALE} & Male gender indicator \\
\texttt{DCO\_4} & Code status at discharge (4 categories) \\
\texttt{HOSP\_PRIOR7\_CT} & Count of hospitalizations in the last 7 days prior to the current admission \\
\texttt{HOSP\_PRIOR8\_30\_CT} & Count of hospitalizations in the last 8 to 30 days prior to the current admission \\
\texttt{LOS\_30} & Length of stay, in days (with stays above 30 days truncated at 30 days) \\
\texttt{MEDICARE} & Indicator for Medicare Advantage status \\
\texttt{DISCHDISP} & Discharge disposition (home, skilled nursing, home health; one-hot encoded) \\
\texttt{LAPS2} & Laboratory-based acuity of illness score, recorded at admission \\
\texttt{LAPS2DC} & Laboratory-based acuity of illness score, recorded at discharge\\
\texttt{COPS2} & Comorbidity and chronic condition score, updated monthly\\
\texttt{HCUPSGDC} & Diagnosis super-group classification (30 groups; one-hot encoded) \\
\midrule
\texttt{W} (or $W_i$) & Treatment: Transitions Program intervention \\
\texttt{Y} (or $Y_i$) & Outcome: Non-elective readmission within 30 days post-discharge \\
\bottomrule
\end{tabular}}
\vspace{0.2em}
\caption{\footnotesize A list of the covariates used in this study.}
\label{tab:table1}
\end{table}
\subsection{From (observational) data to predicted treatment effects: Causal forests}
To identify potentially preventable readmissions, we undertake a causal machine learning approach using data taken from before and after the implementation of the Transitions Program at KPNC. Our causal machine learning approach is distinct from supervised machine learning as it is commonly applied in that it seeks to estimate individual \textit{treatment effects}, and not outcome risk. This objective cannot be readily accomplished with off-the-shelf supervised machine learning methods. \cite{Athey2017BeyondProblems} Compared to other methods for studying treatment effect heterogeneity (e.g. subgroup analyses), causal machine learning methods afford two advantages: one, they avoid strong modeling assumptions, allowing a data-driven approach; and two, they guard against overfitting through applying a form of regularization.
We express these individual treatment effects of the Transitions Program intervention in terms of the estimated conditional average treatment effects (CATE), $\hat{\tau}_i$, which can be interpreted as absolute risk reductions (ARRs). It is through the sign and magnitude of these estimated CATEs that we consider a readmission potentially preventable: a $\hat{\tau}_i < 0$ denotes that the intervention would be expected to lower 30-day readmission risk for that patient, while a $\hat{\tau}_i > 0$ suggests that the intervention would be more likely to result in readmission within 30 days. A larger (more negative) CATE suggests a greater extent of preventability: i.e., $\hat{\tau}_j < \hat{\tau}_i < 0$ implies that patient $j$'s readmission is more "preventable"---their risk is more modifiable by the intervention---compared to patient $i$'s.
To estimate these CATEs, we apply causal forests to the KPNC data described in the previous subsection. Causal forests \cite{AtheyGRF, Wager2018EstimationForests} represent a special case of generalized random forests \cite{Athey2019GeneralizedForests}. They have been used in a variety of applications, including to estimate conditional average partial effects; our overall approach resembles that undertaken in \cite{Athey}, which used them to study treatment effect heterogeneity in an observational setting. Causal forests can be viewed as a form of adaptive, data-driven subgroup analysis, and can be applied in either observational or randomized settings, relying on much of the machinery of random forests \cite{Breiman2001RandomForests}.
These machinery retained from random forests include recursive partitioning, subsampling, and random selection of the splits. However, causal forests differ from random forests in several significant ways. For one, the splitting criterion seeks to place splits so as to maximize heterogeneity, instead of minimizing prediction error as with random forests. Moreover, when placing splits, the causal forest algorithm (depending on implementation details) has access to the data $X_i$ and treatment assignments $W_i$, but not to the outcomes $Y_i$, and a separate sample is used to estimate effects once these splits have been placed---a property often referred to as 'honesty' \cite{Wager2018EstimationForests}.
We describe some necessary assumptions that are required in order to identify the CATEs, as the data are considered observational, and not randomized. The causal forest already fits a treatment assignment model, which deconfounds the subgroups in our study (described below) to some extent, but we also make further assumptions regarding the data, which we describe below, to facilitate deconfounding. We begin with some notation: for each of a set of units $i = 1, \ldots, n$, we observe the triple $(X_i, Y_i, W_i)$, where $X_i \in \mathbb{R}^p$ is a covariate vector, $Y_i \in \{0,1\}$ denotes the observed outcome, and $W_i \in \{0,1\}$ treatment assignment. Following Rubin's potential outcomes framework \cite{Rubin1974EstimatingStudies}, we assume the existence of potential outcomes, $Y_i(1)$ and $Y_i(0)$, for each unit, and define the conditional average treatment effect (CATE) for an unit $i$ with the covariate vector $X_i = x$ as
\begin{equation}
\tau_i(x) = \mathbb{E}[Y_i(1) - Y_i(0) \mid X_i = x].
\end{equation}
Within a leaf $L$, a causal forest estimates this quantity as
\begin{equation}
\hat{\tau}(x) = \frac{1}{\mid\!\! \{j : W_j = 1 \wedge X_j \in L \}\!\! \mid} \sum_{\{j : W_j = 1 \wedge X_j \in L\}} Y_j \quad - \quad \frac{1}{\mid\!\! \{ j : W_j = 0 \wedge X_j \in L \}\!\! \mid} \sum_{\{j : W_j = 0 \wedge X_j \in L\}} Y_j
\end{equation}
for a $x \in L$. Heuristically, the process of fitting a causal forest aims to make these leaves $L$ as small as possible so that the data in each resemble a randomized experiment, while simultaneously maximizing effect heterogeneity. \cite{Wager2018EstimationForests}
However, as we observe only one of the two potential outcomes for each unit, $Y_i = Y_i(W_i)$, we cannot estimate $Y_i(1) - Y_i(0)$ directly from these data. Under some assumptions, however, we can reprovision the units in the data that did experience the counterfactual outcome to estimate $\tau_i$, by having those units serve as 'virtual twins' for an unit $i$. These assumptions entail (1) the existence of these twins; and (2) that, in some sense, these twins look similar in terms of their covariates. These are the \textit{overlap} and \textit{uncounfoundedness} assumptions, respectively. Heuristically, the overlap assumption presumes that these twins could exist, and unconfoundedness posits that these twins are in fact similar in terms of their observed covariates. Together, these assumptions allow us to have some confidence that the $\tau_i(x)$ in fact do identify causal effects.
We address identification with respect to the predicted risk threshold of 25\%, as well as the time period. By time period, recall that the Transitions Program intervention was rolled out to each of the 21 KPNC hospitals in a way that formed two completely disjoint subsets of patients, corresponding to the pre- and post-implementation periods. Also recall that patients were assigned to the intervention if they were discharged during the post-implementation period and their risk score was $>\!\!25$\%, while those below that value received usual care. An useful feature of these data is that all patients in the pre-implementation period were assigned 'shadow' risk scores using the same instantiation of the predictive model, even though none of these patients received treatment. Hence, the data are split into four disjoint subgroups, indexed by risk category and time period:
\[
(\text{pre}, \geq\!\!25), \quad (\text{post}, \geq\!\!25), \quad (\text{pre}, <\!\!25), \quad (\text{post}, <\!\!25),
\]
i.e., the tuple $(\text{pre}, \geq\!\!25)$ denotes the subgroup consisting of patients discharged during the pre-implementation period with a risk score of $\geq\!\!25$\%, and $(.\,, \geq\!\!25)$ denotes all patients with risk $\geq\!\!25$\% in the data. Each hospital discharge belongs to one and only one of these subgroups.
Heuristically, these 'shadow' risk scores allow us to mix data from across periods so that the $(\text{pre}, \geq\!\!25)$ subgroup can be used as a source of counterfactuals for patients in the $(\text{post}, \geq\!\!25)$ subgroup, assuming no unmeasured confounding as a consequence of the shift from "pre" to "post". Stratifying on the risk score allows us to deconfound the potential outcomes of the patients in these two subgroups. Moreover, these two subgroups together can be used to provide plausible counterfactuals for the patients in the $(.\,, <\!\!25)$ risk subgroup, despite none of those patients having been assigned to the intervention. We describe the identification strategy---which relies on standard ignorability assumptions---in more detail below, beginning with the $(.\,, \geq\!\!25)$ subgroup.
First, for each of the four subgroups, we assume overlap: given some $\epsilon > 0$ and all possible $x \in \mathbb{R}^p$,
\begin{equation}
\epsilon < P(W_i = 1 \mid X_i = x) < 1 - \epsilon.
\end{equation}
This assumption means that no patient is guaranteed to receive the intervention, nor are they guaranteed to receive the control, based on their covariates $x$.
For the patients in the $(.\,, \geq\!\!25)$ group, our identification strategy makes use of the balancing properties of risk scores (or prognostic scores) \cite{Hansen2008TheScore} to establish unconfoundedness. Assuming no hidden bias, conditioning on a prognostic score $\Psi(x) = P(Y \mid W = 0, x)$ is sufficient to deconfound the potential outcomes. The risk score used to assign the Transitions intervention is a prognostic score; hence, it is sufficient to deconfound these potential outcomes.
For patients in the $(.\,, <\!\!25)$ subgroup, the picture is slightly more complicated. Among these patients, we cannot assume exchangeability conditional on the risk score $\hat{Y}_i$,
\begin{equation}
\{ Y_i(1), Y_i(0) \} \mathrel{\text{\scalebox{1.07}{$\perp\mkern-10mu\perp$}}} W_i \mid \hat{Y}_i,
\end{equation}
because, again, treatment assignment is contingent on a predicted risk $\hat{Y}_i \geq 0.25$, i.e., $W_i = \mathbf{1}\{\hat{Y}_i \geq 0.25\}$. However, recall that the causal forests are performing estimation in $X_i$-space, and not in $\hat{Y}_i$-space, and note that we can instead impose the slightly weaker assumption of ignorability conditional on some subset $X'_i \subseteq X_i$, which comprise inputs to the score $\hat{Y}_i$; namely, that
\begin{equation}
\{ Y_i(1), Y_i(0) \} \mathrel{\text{\scalebox{1.07}{$\perp\mkern-10mu\perp$}}} W_i \mid X'_i,
\end{equation}
which we can justify \textit{a priori} with the knowledge that no one component predictor predominates in the risk model (see the appendix to \cite{Escobar2015NonelectiveMortality}); that is, no one covariate strongly determines treatment assignment. We provide empirical evidence to establish the plausibility of this assumption, at least in low dimensions, in Figure \ref{fig:overlap}. Together with assuming the existence of potential outcomes, this \textit{unconfoundedness} assumption is sufficient to obtain consistent estimates of $\tau(x)$ \cite{Wager2018EstimationForests}. Moreover, since causal forests perform a form of local estimation, our assumptions are independent for the $(.\,, \geq\!\!25)$ and $(.\,, <\!\!25)$ subgroups in the sense that if the unconfoundedness assumption fails for either subgroup, but not the other, the estimates for the subgroup in which it does hold should not be affected.
Finally, as a formal assessment of treatment effect heterogeneity, we also perform the omnibus test for heterogeneity \cite{Chernozhukov2017}, which seeks to estimate the best linear predictor of the CATE by using the "out-of-bag" predictions from the causal forest, $\hat{\tau}^{-i}$, to fit the following linear model:
\begin{equation}
Y_i - \hat{m}^{-i}(X_i) = \alpha\bar{\tau}(W_i - \hat{e}^{-i}(X_i)) + \beta(\hat{\tau}^{-i}(X_i) - \bar{\tau})(W_i - \hat{e}^{-i}(X_i)) + \epsilon,
\end{equation}
where
\begin{equation}
\bar{\tau} = \frac{1}{n} \sum^n_{i=1} \hat{\tau}^{-i}(X_i),
\end{equation}
and $\hat m(.)$ and $\hat e(.)$ denote the marginal outcome and assignment models estimated by the causal forest, respectively. (The superscript $-i$ denotes that the quantity was computed "out-of-bag", i.e., that the forest was not trained on example $i$). Fitting this linear model yields two coefficient estimates, $\alpha$ and $\beta$; an interpretation of these coefficients is that $\alpha$ captures the average treatment effect, and if $\alpha \approx 1$, then the predictions the forest makes are correct, on average. Likewise, $\beta$ measures how the estimated CATEs covary with the true CATEs; if $\beta \approx 1$, then these CATE estimates are well-calibrated. Moreover, we can use the $p$-value for $\beta$ as an omnibus test for heterogeneity; if the coefficient is statistically significantly greater than zero, then we can reject the null hypothesis of no treatment effect heterogeneity. \cite{AtheyEstimatingApplication}
All analyses were performed in R (version 3.6.2); causal forests and the omnibus test for heterogeneity were implemented using the \texttt{grf} package (version 0.10.4). Causal forests were fit using default settings with $n = 8,000$ trees and per-hospital clusters (for a total of 21 clusters).
\subsection{Translating predictions into treatment policies: Decoupling effects and payoffs}
A relevant question, once predictions have been made---be they of risk or of treatment effects---is how to translate them into treatment decisions. With predicted risk, these decisions are made with respect to some risk threshold, or to a decision-theoretic threshold that takes utilities into account (e.g. as in the approach in \cite{Bayati2014Data-drivenStudy.}). However, both approaches are potentially suboptimal in the presence of treatment effect heterogeneity, requiring strong assumptions to be made regarding the nature of the treatment effect. (Indeed, \cite{Bayati2014Data-drivenStudy.} assumes a constant treatment effect for all patients.)
Here, however, we deal with treatment effects instead of risk; an obvious approach starts by treating all patients $i$ with $\hat{\tau}(X_i) < 0$---that is, by treating all patients who are expected to benefit. However, resources may be constrained so that it is infeasible to treat all these patients, and making it necessary to prioritize from among those with $\hat{\tau}_i < 0$. One way to do so is to incorporate the costs associated with the potential outcome of a readmission, or the "payoffs" $\pi$ associated with successfully preventing a readmission \cite{Kleinberg2015PredictionProblems}, which we denote by $\pi_i = \pi(X_i)$.
There are several ways to characterize these payoffs, which ideally can be done mainly in terms of the direct costs required to provide care for a readmitted patient, as well as financial penalties associated with high readmission rates. However, these data are not available to us, so we instead use length of stay (LOS) from the readmission as a proxy for cost, and assume that higher LOS is associated with higher resource utilization and thus higher costs. A range of payoffs could be specified, such as the risk of in-hospital mortality during the readmission, or an acuity-scaled LOS measure, as well as weighted combinations of these quantities. It is important to emphasize that these payoffs are associated with the characteristics of the readmission following the index stay, if one does occur---not those of the index stay itself.
One approach to estimating these payoffs is to predict them using historical data, i.e., $\hat{\pi}(X_i) = \mathbb{E}[\pi_i \mid X_i = x]$ in a manner similar to that used to derive the risk scores. However, this is beyond the scope of this paper, and so we make some simplifying assumptions regarding the payoffs. Namely, we assume that (1) the individual payoffs $\pi_{i}$ can be approximated by the mean payoff across all patients, $\pi_{i} \approx \mathbb{E}[\pi_{i}]$, and (2) that the payoffs are mean independent of the predicted treatment effects, $\mathbb{E}[\pi_i \mid \tau_i] = \mathbb{E}[\pi_i]$. These two assumptions make it so that the $\hat{\tau}_i$ become the sole decision criterion for the treatment policies we evaluate in this paper, but we briefly outline how to incorporate these payoffs in decision-making.
Given both the predicted treatment effects, $\hat{\tau}_i$, and payoffs, $\hat{\pi}_i$, we can compute the individual expected utilities, $\mathbb{E}[u_i] = \hat{\tau}_i \hat{\pi}_i$ for each patient. We assume that decision-makers are risk-neutral and that the cost to intervene is fixed. Then, given two patients, $i$ and $j$, and their respective expected utilities, we would prefer to treat $i$ over $j$ if $\mathbb{E}[u_i] > \mathbb{E}[u_j]$. Another interpretation (in a population sense) is that ordering the discharges in terms of their $\hat{\tau}_i$ induces one rank ordering, while ordering them in terms of their $\mathbb{E}[u_i]$ induces another. We can treat the top $k$\% of either ordering, subject to resource constraints, but doing so with the latter will result in greater net benefit and thus would be preferred. Under the assumptions we make above, $\mathbb{E}[u_i] \propto \hat{\tau}_i$ for each patient $i$. This particular decision-theoretic approach requires absolute, and not relative outcome measures, such as the relative risk reduction. \cite{Sprenger2017ThreeMeasures}
\subsection{Measuring the impacts of different targeting strategies}
To estimate the impact (in terms of the number of readmissions prevented) of several notional targeting strategies that focus on treating those with the largest expected benefit, and not those at highest predicted risk, we undertake the following approach. We stratify the patients in the dataset into ventiles $V_1, \ldots, V_{20}$ of predicted risk, where $V_1$ denotes the lowest (0 to 5\%) risk ventile, and $V_{20}$ the highest (95 to 100\%). Then, the causal forest is trained on data through the end of 2017, and used to predict CATEs for all patients discharged in 2018.
First, for all patients above a predicted risk of 25\%, we compute the impact of the current risk-based targeting strategy based on the predicted CATEs from 2018, and compare it to that from prior work which estimated based on the average treatment effect of this intervention. \cite{Marafino2020ASystem} This comparison serves as one check of the calibration of the predicted CATEs; the number of readmissions prevented should substantially agree in both cases.
Second, we then use these same predicted CATEs for 2018 to assess the impact of three CATE-based targeting strategies, which treat the top 10\%, 20\%, and 50\% of patients in each risk ventile based on their predicted CATE. These strategies spread the intervention across ventiles as a form of a hedge ("not putting all one's eggs in one basket"), rather than treating the top $k$\% patients in the dataset.
The impacts of all targeting strategies are characterized both in terms of the annual number of readmissions prevented as well as the number needed to treat (NNT) to prevent one readmission. We estimate the annual number of readmissions prevented by training the causal forest on data through the end of 2017, and then summing the predicted CATEs for discharges taking place in 2018. The annual number of readmissions prevented can be expressed as
\begin{equation}
\sum^{20}_{j = 1} \sum_{i \in V^T_j} \hat{\tau}_i,
\end{equation}
where $V^T_j$ denotes the set of patients notionally selected for treatment in ventile $j$, and $\hat{\tau}_i$ the predicted CATE (treatment effect) for the patient $i$.
Finally, the NNT in this setting has a slightly different interpretation than the one commonly encountered in many studies. Often, the NNT is cited as a summary measure based on the overall results of a trial, e.g., the average treatment effect (ATE). In this case, an ATE estimate $\hat{\tau}$, expressed in terms of the ARR, corresponds to a NNT of $1/\overline{\hat{\tau}_i} = 1/\hat{\tau}$. However, here, we instead estimate CATEs at the individual patient level and are interested in the NNTs specific to the subgroups that receive the intervention. Hence, the NNT is a property of these subgroups and can vary across subgroups as the average predicted CATE $\overline{\hat{\tau}_i}$ varies.
\section{Results}
\subsection{Overall characteristics of the cohort}
From June 2010 to December 2018, 1,584,902 hospitalizations took place at the 21 KPNC hospitals represented in this sample. These included both inpatient hospitalizations as well as stays for observation. Further details regarding the overall cohort are presented in Table \ref{tab:a1}. Of these hospitalizations, 1,539,285 met the inclusion criteria, of which 1,127,778 (73.3\%) occurred during the pre-implementation period for the Transitions Program, and 411,507 (26.7\%) during the post-implementation period. Among these 411,507 hospitalizations taking place post-implementation, 80,424 (19.5\%) were predicted to be at high risk of 30-day post-discharge mortality or readmission; these patients were considered to have received the Transitions Program intervention following hospital discharge.
Of the patients whose index stays were included, their mean age was 65.0 years, and 52.5\% were women. The overall 30-day non-elective rehospitalization rate among these index stays was 12.4\%, and 30-day post-discharge mortality was 4.0\%. Other patient-level characteristics are presented in Table \ref{tab:a1} in the Appendix. Notably, based on the distributions of \texttt{COPS2} and \texttt{LAPS2}, a key modeling assumption---that of overlap---appears to have been satisfied (Figure \ref{fig:overlap}).
Patients at low risk (risk score $<\!\!25$\%) represented 63.3\% of all readmissions throughout the study period, while making up 82.9\% of index stays, compared to 36.7\% of all readmissions among those at high risk ($\geq \!\! 25$\%), which represented 17.1\% of index stays. Moreover, the mean length of stay of the readmission following an index stay was approximately constant across ventiles of predicted risk, satisfying another assumption; patients with predicted risk of 5 to 50\% at their index discharge had a mean length of stay during their readmission that ranged from 4.6 to 5.7 days, and these patients represented 90.5\% of all readmissions. (Figure \ref{fig:los-v-risk})
\begin{figure}
\centering
\includegraphics[scale=0.67]{img/overlap.pdf}
\caption{\footnotesize Assessing the "unconfoundedness" assumption: each point denotes an admission, which are colored according to whether they received the Transitions Program intervention post-discharge ($W_i$). The $x$-axis records the value of a laboratory-based acuity score (\texttt{LAPS2DC}) and the $y$-axis the value of a chronic condition score (\texttt{COPS2)}, both at discharge. The extent of overlap displayed here is relatively good, and implies that overlap may be implausible only among patients at very high or very low risk. This plot is based on a random sample of $n = 20,000$ index admissions taken from the post-implementation period.}
\label{fig:overlap}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.60]{img/los-v-risk.pdf}
\caption{\footnotesize Average length of stay (LOS) by risk score ventile. The values in parentheses below the name of each ventile denote the proportion of all 30-day readmissions incurred by patients in that ventile; patients with a predicted risk below 25\% based on their index stay accounted for 63\% of all readmissions. Notably, the average LOS is roughly similar (at 5 days) for patients with predicted risk of 5\% to 80\%. The vertical dotted line represents the 25\% risk threshold used to assign the Transitions Program intervention.}
\label{fig:los-v-risk}
\end{figure}
\subsection{Characterizing the treatment effect heterogeneity of the Transitions Program intervention}
The estimated out-of-bag conditional average treatment effects (CATEs) yielded by the causal forest are presented in Figure \ref{fig:cate-overall}. The distributions of the CATEs are those of the discharges in their respective risk ventiles, but are not drawn on a common scale and so do not reflect the variation in sample size across ventiles. Qualitatively, these distributions exhibit wide spread, and suggest some extent of heterogeneity in the treatment effect of the Transitions Program intervention. In particular, treatment effects appear to be largest for patients discharged with a predicted risk of around 15 to 35\%. These effects also appeared to be somewhat attenuated as risk increased, with the center of mass tending towards zero for patients at higher risk. Notably, particularly among patients at higher risk, some estimated effects were greater than zero, indicating that the intervention was more likely to lead to readmission within 30 days. Finally, we also note that the CATE estimates themselves were well-calibrated in the sense that we identified no cases where an individual's CATE estimate was greater than their predicted risk.
Figure \ref{fig:cate-hcupsg} is similar to the previous, but stratifies the display by Clinical Classification Software (CCS) supergroups. Definitions of these supergroups can be found in Table \ref{tab:b1} in the Appendix. The overall pattern is similar to that in the unstratified plot, in that treatment effects appear to be greatest for patients at low to moderate risk, but the shapes of these distributions vary from supergroup to supergroup. All supergroups appear to exhibit heterogeneity in treatment effect within ventiles, as well, which is more pronounced for some conditions, including hip fracture, trauma, and highly malignant cancers. Qualitatively, some supergroups exhibit bimodal or even trimodal distributions in the treatment effect of the Transitions Program intervention, suggesting identification of distinct subgroups based on these effects. Some ventiles are blank for some supergroups, because there were no patients belonging to those supergroups with predicted risks falling within those ranges.
Quantitatively, fitting the best linear predictor yields estimates of $\hat \alpha = 1.16$ and $\hat \beta = 1.06$, with $p = 5.3 \times 10^{-8}$ and $2.23 \times 10^{-7}$, respectively. Interpreting the estimate of $\beta$ as an omnibus test for the presence of heterogeneity, we can reject the null hypothesis of no treatment effect heterogeneity.
These effects can also be evaluated on a grid of two covariates to assess how the estimated CATE function varies with interaction of these covariates. This yields insight into the qualitative aspects of the surface of the CATE function and may identify subgroups among which the Transitions Program intervention may have been more or less effective. Here, we choose the Comorbidity Point Score (\texttt{COPS2}) and the Laboratory-based Acuity Score at discharge (\texttt{LAPS2DC}), while holding all other continuous covariates at their median values, except for age, which we set to 50 and 80. Categorical covariates were held at their mode, except for the supergroup, which we set to chronic heart failure (CHF). We plot the CATE function from the 10th to 90th percentiles of \texttt{LAPS2DC} and from the 0th to 95th percentiles of \texttt{COPS2}. This is akin to evaluating the CATE function for a set of pseudo-patients with CHF having these values of \texttt{COPS2} and \texttt{LAPS2DC}.
Figure \ref{fig:cate-hcupsg} shows the resulting CATE functions for two choices of patient age: 50 and 80 years. In this region, the estimated CATE ranged from -0.060 to 0.025 (-6.0 to 2.5\%), meaning that the estimated absolute risk reduction of the Transitions Program intervention was as large as -6\% for some patients, while for others, their readmission risk was increased by as much as 2.5\%. The estimated CATE generally was increased in magnitude---suggesting that the Transitions Program intervention became more effective as age increased---at age 80 compared to 50. Moreover, and notably, the estimated CATE tended to increase with increasing \texttt{LAPS2DC}, which measures how acutely ill a patient was upon discharge based on their laboratory test data. This finding suggests that, for patients who were more ill at discharge (indeed, the average \texttt{LAPS2DC} in 2018 was 45.5), enrolling them in the Transitions Program may actually have encouraged them to return to the hospital. While this finding of such an effect may appear surprising, it is unclear if it actually represents "harm" in the sense it is usually interpreted; we discuss this finding in more depth in the Discussion section.
\begin{figure}
\centering
\includegraphics[scale=0.67]{img/overall-cate-estimate.pdf}
\caption{\footnotesize Treatment effect heterogeneity across risk score ventiles. The densities represent the distribution of estimated conditional average treatment effects within each ventile. They are drawn on a common scale, and hence do not reflect the variation in sample size across ventiles.}
\label{fig:cate-overall}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.67]{img/cate-by-hcupsg.pdf}
\caption{\footnotesize Treatment effect heterogeneity across risk score ventiles, stratified by Clinical Classification Software (CCS) supergroups based on the principal diagnosis code at discharge. A full listing of the definitions of these supergroups is given in Table \ref{tab:b1} in the Appendix. Abbreviations: CVD, cerebrovascular disease; AMI, acute myocardial infarction; CAP, community-acquired pneumonia; CHF, congestive heart failure; GI, gastrointestinal; UTI, urinary tract infection.}
\label{fig:cate-hcupsg}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.67]{img/5080_chf.pdf}
\caption{\footnotesize The estimated CATE function as it varies in the dimensions of \texttt{LAPS2DC} and \texttt{COPS2}, for a patient with chronic heart failure at ages 50 and 80. All other continuous variables were fixed at their median, and other categorical variables were fixed at their mode. (The transitions from one cell to another in this figure should not appear smooth; if this is the case, try a different PDF viewer.)}
\label{fig:cate-chf}
\end{figure}
\subsection{Notional estimates of overall impact under different targeting strategies}
Based on these individual CATE estimates, we compute the potential impacts of several notional targeting strategies using these estimated effects, and not predicted risk, to target the Transitions Program intervention. These impacts are expressed in terms of the number of annual readmissions prevented, and are presented in Table \ref{tab:policies}. These quantities are computed by training a model on all data through December 2017, which is then used to predict effects for patients discharged in 2018. These predicted effects are used to compute the numbers of readmissions prevented and number needed to treat (NNT). We also present the estimated number of interventions required under each strategy.
We first confirm the calibration of the individual effect estimates by taking the same group of patients who were intervened upon under the current risk-based strategy, and use this group to estimate the number of readmissions prevented, with the aim of comparing this number to a previous estimate of the impact of this policy using the average treatment effect \cite{Marafino2020ASystem}. This results in an estimate of 1,246 (95\% confidence interval [CI] 1,110-1,381) readmissions prevented annually, which compares favorably to the previous estimate of 1,210 (95\% CI 990-1,430), representing further evidence that these estimates are well-calibrated. The NNT under both of these these strategies is 33, and the number of individual interventions needed is 39,985. (Table \ref{tab:policies})
Next, computing the impacts of the CATE-based strategies, which target the Transitions Program intervention to the top 10\%, 20\%, and 50\% of each risk ventile, we find that all of these policies are estimated to result in greater potential reductions in the absolute number of readmissions prevented. \ref{tab:policies} The top-10\% strategy may prevent 1,461 (95\% CI 1,294-1,628) readmissions annually, and does so more efficiently, as implied by the NNT of 13. Moreover, the top-20\% strategy requires the same total number of interventions as the existing risk-based strategy (39,648 vs. 39,985), yet is estimated to lead to double the number of annual readmissions prevented, at 2,478 (95\% CI 2,262-2,694). This strategy appears to be much more efficient, as evidenced by its estimated NNT of 16.
Even under the most expansive strategy, which treats the top 50\% of each risk ventile and requires 250\% the number of total interventions compared to the risk-based strategy, also represents an improvement in the NNT (23 vs. 33). This strategy is estimated to lead to 4,458 (95\% CI 3,925-4,990) readmissions prevented annually, or nearly four times as many as the existing strategy. Finally, we also note that while there appears to exist a tradeoff in terms of absolute impact and efficiency, all CATE-based strategies substantially improved upon the risk-based targeting strategy in terms of the NNT.
\begin{table}[]
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
Treatment strategy & \begin{tabular}[c]{@{}c@{}}Annual readmissions \\ prevented, $n$\end{tabular} & \begin{tabular}[c]{@{}c@{}}Total interventions,\\ $n$\end{tabular} & NNT \\ \midrule
\textbf{Risk-based targeting} & & & \\
\quad Target to $ \geq\!\! 25$\% (DiD estimate) & 1,210 (990-1,430) & 39,985 & 33 \\
\quad Target to $ \geq\!\! 25$\% (CF estimate) & 1,246 (1,110-1,381) & 39,985 & 33 \\ \midrule
\textbf{CATE-based targeting} & & & \\
\quad Targeting top 10\% & 1,461 (1,294-1,628) & 18,993 & 13 \\
\quad Targeting top 20\% & 2,478 (2,262-2,694) & 39,648 & 16 \\
\quad Targeting top 50\% & 4,458 (3,925-4,990) & 102,534 & 23 \\ \bottomrule
\end{tabular}
\vspace{0.2em}
\caption{\footnotesize Estimates of overall impacts of risk-based and notional CATE-based targeting strategies in terms of the annual numbers of readmissions prevented as well as the numbers needed to treat (NNTs) under each targeting strategy, based on the estimates for index admissions in 2018. The first quantity---the difference-in-differences (DiD) estimate---is based on the results of \cite{Marafino2020ASystem}. All quantities are rounded to the nearest integer. Parentheses represent 95\% confidence intervals. Abbreviations: CATE, conditional average treatment effect; DiD, difference-in-differences; CF, causal forest.}
\label{tab:policies}
\end{table}
\section{Discussion}
In this paper, we have shown the feasibility of estimating individual treatment effects for a comprehensive readmissions prevention intervention using data on over 1.5 million hospitalizations, representing an example of an "impactibility" model \cite{Lewis2010ImpactibilityPrograms}. Even though our analysis used observational data, we found that these individual estimates were well-calibrated, in that none of the individual estimates were greater than the predicted risk. Moreover, these estimates, in aggregate, when used to compute the impact of the risk-based targeting policy, substantially agreed with a separate estimate computed via a difference-in-differences analysis. Notably, our results suggest that strategies targeting similar population health management and quality improvement (QI) interventions based on these individual effects may lead to far greater aggregate benefit compared to targeting based on risk. In our setting, the difference translated to nearly as many as four times the number of readmissions prevented annually over the current risk-based approach.
Our analysis also found both qualitative and quantitative evidence for treatment effect heterogeneity, particularly across levels of predicted risk: the Transitions Program intervention seemed less effective as predicted risk increased. The extent of this mismatch between treatment effect and predicted risk appeared substantial, and may have implications for the design of readmission reduction programs and related population health management programs. Furthermore, our analysis, when stratified by diagnostic supergroup, also appeared to identify distinct subgroups consisting of patients with larger, more negative treatment effects, indicating that the intervention may have been more effective in those subgroups. Patients expected to benefit could be prioritized to receive the intervention, while patients unlikely to benefit could instead receive more targeted care that better meets their needs, including specific subspecialty care, and in some cases, palliative care. More work remains to be done to investigate if these findings hold for other types of preventative interventions, to characterize these subgroups, and to evaluate how to best translate our preliminary findings into practice. Notably, our finding of a risk-treatment effect mismatch is in line with suggestions in the readmissions prevention literature \cite{Finkelstein2020HealthTrial,Steventon2017PreventingRisk,Lindquist2011UnderstandingFactors}.
To the best of our knowledge, this work is the first to apply causal machine learning together with decision analysis to estimate the treatment effect heterogeneity of a population health management intervention, and as such, represents the first example of an end-to-end "impactibility" model \cite{Lewis2010ImpactibilityPrograms}. Our approach is also notable in that we show how to decouple the causal and predictive aspects of this prediction problem \cite{Kleinberg2015PredictionProblems, Ascarza2018RetentionIneffective}. In particular, our approach was principally inspired by a study of the effectiveness of targeting marketing interventions based on risk by Ascarza \cite{Ascarza2018RetentionIneffective}, as well as previous work on the causal aspects of prediction problems by Kleinberg and co-authors \cite{Kleinberg2015PredictionProblems, Kleinberg2018HumanPredictions}, and others taking causal approaches to prediction problems. \cite{Rubin2006EstimatingMethodology, Blake2015ConsumerExperiment} From a modeling perspective, treating such a problem as purely predictive, as is commonly done in studies developing readmission risk tools \cite{Kansagara2011RiskReview.}, relies on an assumption that may be implausible---specifically that treatment effects correlate with risk. On the other hand, a purely causal approach---one that focuses on modeling treatment effects---fails when confronted with prediction problems that involve resource allocation. Even those patients whose readmissions may be the most "impactible" may not also have the most resource-intensive readmissions. Indeed, this potential oversight exemplifies a form of bias referred to as "omitted payoff bias". \cite{Kleinberg2018HumanPredictions}
Previous studies have investigated the extent to which preexisting risk assessment tools may already capture a patient's degree of impactibility, by integrating qualitative assessments from nurses and physicians into these models \cite{Flaks-Manov2020PreventingPrediction}. While there seems to be some overlap between predicted risk and impactibility as expressed by these providers, it appears incomplete, and it is not clear how clinician assessments could be integrated into existing models. \cite{Flaks-Manov2020PreventingPrediction} Further afield, performance metrics for predictive models such as the recently proposed "C statistic for benefit" \cite{vanKlaveren2018TheEffects} can assess the ability of such a model to discriminate between patients expected to benefit from a treatment and those who will not. While useful, these metrics do not facilitate assessment of treatment effect heterogeneity. Attempts have been made to model treatment effect heterogeneity among patients undergoing antihypertensive therapy using a similar causal machine learning approach (the X-learner \cite{Kunzel2019MetalearnersLearning}), and also found some extent of mismatch between treatment effects and risk. \cite{Duan2019ClinicalTherapy}
\subsection{The necessity of estimating heterogeneous treatment effects, and not just outcome risk}
Our results highlight the necessity of estimating the treatment effect heterogeneity associated with preventative interventions, particularly at the population scale. A full appraisal of such heterogeneity can aid in targeting these interventions to patients among whom they might be most effective, while targeting intensified versions, or different versions, to patients who may not be expected to benefit from the intervention as originally implemented. As a result, our analyses suggest that up to nearly as four times as many readmissions could be prevented compared to the current state of affairs, while incurring nominal marginal costs in terms of the total number of interventions delivered. Thus, the current risk-centric modeling methodology employed by many hospitals and health systems, as well as by payers, may limit the full potential of these interventions.
We focus on what could be considered a special case of treatment effect heterogeneity---of that on the absolute risk scale. \cite{Kent2016RiskTrials., Kent2018PersonalizedEffects} Unlike other studies performing similar analyses (e.g. \cite{Kent2008AInfarction} and others cited in \cite{Kent2016RiskTrials.}) finding that a small group of high-risk patients accounted for most of the aggregate benefits, we instead found that the effect of the Transitions Program intervention were largest in the most numerous subgroup of patients, namely those at relatively low to moderate risk. For high-risk patients, the intervention appeared to be less effective, which is in line with the growing body of evidence showing similar findings \cite{Finkelstein2020HealthTrial,Steventon2017PreventingRisk,Lindquist2011UnderstandingFactors,Rich1993PreventionStudy.}. An open question is the extent to which these high-risk/ineffective patients are qualitatively distinct, to the point that they should be considered "futile", as some have proposed \cite{Lewis2010ImpactibilityPrograms}, or if they may be amenable to a more intensified version of the intervention or an entirely different approach altogether.
A notable finding is that some patients had a predicted CATE greater than zero, indicating that the Transitions Program intervention may have encouraged them to return to the hospital. In other settings, a positive treatment effect would often be interpreted as harm, and suggest that the treatment be withheld from these patients. However, we argue that our finding does not readily admit such an interpretation. To see why, we note that this subgroup of patients with a positive CATE appeared to be those who were more acutely ill at discharge, as evidenced by their higher \texttt{LAPS2DC} scores (Figure \ref{fig:cate-chf}). In light of this finding, an alternative interpretation of these weakly positive effects is that they represent readmissions which may have been necessary, and which perhaps may have been facilitated by aspects of the Transitions Program intervention, including instructions to patients outlining the circumstances (e.g. new or worsening symptoms) under which they should return to the hospital. This finding holds particular relevance given increasing concern that readmission prevention programs, in responding to the incentives of the HRRP, may be reducing 30-day hospitalization rates at the expense of increased short- and long-run mortality. \cite{Wadhera2018AssociationPneumonia, Fonarow2017TheReconsider, Gupta2018ThePolicy}
Moreover, this finding also suggests that the CATE estimates may be insufficient to capture the full impact of the Transitions Program intervention on patient outcomes, meaning that the estimated effect of the intervention on readmission alone may not represent a sufficient basis for future targeting strategies. It is plausible that intervening in a patient with a positive estimated effect may be warranted if the readmission would have a positive effect on other outcomes, despite the current emphasis of value-based purchasing programs on penalizing excess 30-day readmissions. For example, in fiscal year 2016, the maximum penalty for excess 30-day mortality was 0.2\% of a hospital's diagnosis-related group (DRG) payments under the Hospital Value-Based Purchasing program, while the maximum penalty for excess 30-day readmission was 3.0\% of DRG payments under the HRRP. \cite{Abdul-Aziz2017AssociationMortality} A more holistic targeting strategy would incorporate estimates of the intervention's effect on short- and long-run mortality and other outcomes, and explicitly include these quantities when computing expected utility. Selecting patients for treatment can then be formulated as an optimization problem that attempts to balance regulatory incentives, organizational priorities, patient welfare, and resource constraints.
\subsection{Causal aspects of prediction problems}
Our approach, in contrast to many studies which have developed readmission risk prediction models \cite{Kansagara2011RiskReview.,Bayati2014Data-drivenStudy.,Bates2014BigPatients}, instead focuses on the causal aspects of the readmissions prediction problem. Our analyses emphasize modeling treatment effect heterogeneity of the Transitions Program intervention and on closing the loop from prediction to decision. However, more work remains to formalize criteria that more clearly delineate the roles of causal inference and prediction in these prediction problems, and the extent to which either is necessary or sufficient for a given problem.
As explored by \cite{Bayati2014Data-drivenStudy.} and \cite{Kleinberg2018HumanPredictions}, fully operationalizing a prediction model entails the following steps:
\begin{equation*}
\text{data} \Rightarrow \text{prediction} \Rightarrow \text{decision}.
\end{equation*}
Historically, many machine learning studies, including those in medicine, have emphasized the $\textit{data} \Rightarrow \textit{prediction}$ link, while neglecting the $\textit{prediction} \Rightarrow \textit{decision}$ link. These studies often evaluate model quality on the basis of performance metrics, including the area under the reciever operating characteristic curve (AUROC or AUC) or C-statistic, the area under the precision-recall curve (AUPRC), and other measures of accuracy. A final model is chosen from among candidate models because it maximizes the value of the metric of interest, with the implicit hope that it also maximizes its utility in a real-world setting were it to be so applied. However, this approach conflates prediction with classification---the risks of doing so which are clear, though perhaps underappreciated \cite{Harrell2019ClassificationThinking}---and hence also conflates prediction quality with decision quality.
Indeed, it is conceivable (though not likely) that even an intervention based on a perfectly accurate prediction model---for example, one that correctly identified all readmissions---could result in no net benefit, if the intervention proved ineffective for all the "high-risk" patients whose readmissions were identified. This contrasts with the setting of many classification problems (which we emphasize again are distinct from prediction problems), including, for example, image classification. In these cases, the utility function of classification is expressible in terms of the performance metric of interest alone, such as the accuracy or $F_1$ score. For a given choice of metric, the utility of an image classifier is strictly increasing in that metric, whereas in our hypothetical readmission prediction model, even perfect "classification" is not sufficient. This mismatch between predictive performance and utility is less pronounced with a less severe extent of treatment effect heterogeneity, but our hypothetical example highlights the importance of making the distinction.
To be sure, not all prediction problems decompose neatly into causal and predictive components. Some such problems are purely causal, while others can be solved with prediction alone. Models developed for risk adjustment, for example, represent an example of the latter types of problems, as in those contexts there is no link between an individual prediction and a decision. Rather, the predictions of risk adjustment are used in aggregate; for example, to obtain observed-to-expected ratios of outcomes for quality measures, or to adjust capitation payments according to case mix. \cite{MedPAC2016MedicareSystem} Another class of prediction models that can be viewed as solving purely predictive problems are those used for clinical decision support \cite{Chen2017MachineExpectations}, including the APACHE ICU mortality model \cite{Knaus1985APACHESystem}.
These decision support systems can be used in the context of an individual patient at the point of care to inform decision-making. For example, the APACHE model can be used to track the response of a patient in the intensive care unit (ICU) to treatment, or an objective measure to help decide when palliative care may be necessary. \cite{Knaus2002APACHEReflections} But it is not used in isolation to decide when to start and stop an intervention, unlike the prediction model used to enroll patients in the Transitions Program which we studied in this paper. Instead, the physician is meant to integrate APACHE model output with other information to arrive at a decision, which may incorporate other clinical data not captured in the model, resource constraints, or even family preferences. This is an significant distinction: the view of the APACHE model as solving a purely predictive problem implicitly off-loads the burden of decision-making, and thus of conceptualizing an utility function, to the physician. Without a human (or physician) in the loop, the utility function must be made more explicit in order to maximize aggregate utility, as in the framework we describe in this paper.
On the other hand, a setting where the link between prediction and decision is more explicit is that of resource allocation problems. A well-known example involves the MELD score \cite{Kamath2001ADisease}, which predicts 3-month mortality in liver failure and is used to prioritize patients for liver transplant. The prediction problem can again be classed as purely predictive, because patients with higher MELD scores (and hence higher 3-month mortality) benefit most: allocating donor livers to the patients with the highest MELD scores yields more years of life saved, in aggregate. A similar insight was employed by Kleinberg and co-authors \cite{Kleinberg2015PredictionProblems}, who investigated a different resource allocation problem---that of allocating joint replacement surgeries. A patient who receives a joint replacement will not see the full benefit until at least a year post-surgery, due to the time required for recovery and physical therapy in the interim.
Hence, patients who receive a joint replacement, but die within a year post-surgery, do not benefit from surgery. With this in mind, the problem of allocating joint replacements can be cast as a purely predictive one by developing a prediction model for mortality at 1 year post-surgery. Surgeries can then be allocated to patients in need below a certain risk threshold. Unlike with the MELD score, the authors found that risk did not appear to correlate with benefit as measured by claims for physical therapy, joint injections, and physician visits pre-surgery. If the riskiest patients would otherwise have derived the most benefit from surgery, this would constitute an example of "omitted payoff bias", requiring a different utility function that took the payoff of decreased medical need (presumably corresponding to less severe symptoms) into account in order to assign patients to surgery. \cite{Kleinberg2015PredictionProblems}
We close this subsection with an example of a prediction problem that appears purely predictive, but is in fact largely, and perhaps purely, causal. Consider the problem of identifying hospitalized patients who would benefit from a palliative care consultation, which has attracted considerable attention over the last two decades as the need for palliative care has grown. \cite{Weissman2011IdentifyingCare.} Holistic criteria have been proposed to identify these patients, spanning domains including mortality risk in the next 12 months, the presence of refractory symptoms, the patient's level of social support, and inability to carry out activities of daily living---which constitute a selection of the 13 indicators outlined in \cite{Weissman2011IdentifyingCare.}.
Recently, attempts have also been made to reduce the problem of identifying patients who might benefit from palliative care to one of modeling 12-month mortality from hospital admission, e.g., \cite{Avati2018ImprovingLearning}. Notwithstanding the causal issues involved in defining this outcome of interest retrospectively and applying it prospectively (e.g., see \cite{Einav2018PredictiveLife} for a related discussion), this approach is also problematic in that it is unlikely to reliably identify patients who will actually benefit from palliative care, because it all omits other indicators of palliative care need beyond 12-month mortality. A palliative care consultation is a complex intervention which will likely have highly heterogeneous treatment effects on quality of life---which is the real outcome of interest---and it is unclear if these effects do in fact correlate with 12-month mortality risk.
In fact, solving this proxy problem will likely identify many patients who will die in the short run, but who may not be amenable to palliative care. This includes patients in the ICU, for whom providing the full spectrum of palliative care can be challenging, due to sedation, the use of mechanical ventilation and other invasive interventions, environmental light and noise, and limits on family visitation. \cite{Cook2014DyingUnit} Conversely, this approach would also miss many patients who are at low mortality risk, but who may otherwise benefit from palliative care early in their disease course. Indeed, guidelines recommend initiating palliative care early in patients with cancer, for example---with some trials finding that initiation at as early as the time of diagnosis may be most effective. \cite{Howie2013EarlyImplications}
\subsection{New processes for predictive algorithm-driven intervention deployments within learning health systems}
Our findings could be used to retarget the Transitions Program intervention prospectively to patients who are most expected to benefit, rather than those at highest risk, resulting in larger aggregate benefit in terms of the number of readmissions prevented. Similarly, our approach could be used to retarget other population health management and QI interventions which are often deployed wholesale or using risk assessment tools. However, as we mention, these individual estimates were derived from observational data, and not from data generated via a randomized experiment---the latter type of data which represent the ideal substrate for estimating treatment effects insofar as randomization is able to mitigate the effects of confounding. \cite{Kent2018PersonalizedEffects} Furthermore, our approach requires interventional data, unlike those used to develop more traditional risk assessment tools, which are often based on retrospective data. Hence, to implement our approach, as an alternative to risk tool-driven approaches, may require rethinking how these predictive algorithm-driven interventions (or "prediction-action dyads", cf. \cite{Liu2019TheHealthcare}) are deployed within health systems, particularly in relation to existing digital infrastructure and institutional oversight processes. We outline several starting points for doing so below.
One option is to first deploy a new predictive algorithm-driven intervention as part of a simple two-arm randomized trial which compares that intervention to usual care. This represents a pilot phase, generating data that are used to derive an impactibility model, along with a payoff model if necessary for the prediction problem. Following this pilot phase, two paths are possible: 1) based on this impactibility model, the intervention could be re-targeted to the patients most expected to benefit; or 2) alternatively, carrying out another two-arm randomized trial comparing risk-based to impactibility-based targeting. In the latter choice, patients would be randomized to either of the risk or impactibility arms, and based on their covariates would either receive or not receive the intervention according to their risk or benefit estimate. Based on the results of this second trial, whichever targeting approach proved more effective could then be put into use.
Another option, if a predictive algorithm-driven intervention based on risk scores is already in use, does away with the pilot randomized trial. Instead, the first step is to derive a impactibility model from observational data, as we did in this study. Then, if the goal is to retarget this intervention to include more patients at lower risk---for example, patients between 10 and 25\% predicted risk---risk-based and impactibility-based targeting could be compared to each other in a three-arm randomized trial. The first arm consists of risk-based targeting using the current risk threshold; the second arm, impactibility-based targeting; and the third arm, risk-based targeting to patients in the 10-25\% risk subgroup. Finally, cluster-randomized trial designs based on the regression discontinuity design would allow treatment effect heterogeneity across levels of risk to be characterized, enabling the intervention to be retargeted to the groups among which it is most effective; we hope to investigate these designs in future work.
These proposed options represent major shifts from how deployments of predictive algorithm-driven interventions are usually carried out in health systems. New institutional overnight processes \cite{Faden2013AnEthics}, digital infrastructure, and statistical methodology would be required in order to realize the full potential of these approaches. Many such interventions, if deemed to create only minimal risk, fall under the umbrella of routine quality improvement (QI) studies, which exempts them from ongoing independent oversight \cite{Finkelstein2015OversightResearch}. However, incorporating randomization, as we do in the options we describe above, shifts these interventions from the category of routine QI to non-routine QI or research. These two categories of studies often require independent oversight by, for example, an institutional review board (IRB), and may need to incorporate additional ethical considerations, e.g., requiring informed consent. It is possible that a categorization of a deployment process like the ones we describe above as non-routine QI or research could impede their being carried out, as has occurred in previous QI work. \cite{Baily2006SpecialSafety} However, similar attempts at deploying QI interventions as a part of rapid-cycle randomized experiments have escaped such categorization \cite{Horwitz2019CreatingTesting}, although standards will likely vary from institution to institution.
Moreover, these new deployment processes we envision will also require new digital infrastructure to implement these models and provision their linked interventions as a part of randomized trials and to monitor their impacts in real-time (while accounting for early stopping and other biases), akin to how A/B testing is usually carried out by many Web-facing companies today \cite{Kohavi2013OnlineScale}. Another sticking point is the lack of established analogues of power analyses for causal forests and other methods. Without these analogues, it is not clear how to set the necessary sample size for a pilot randomized trial like the one we describe above, although it could possibly be determined via simulation. The investments required to establish these platforms may be considerable, and it remains unclear whether the costs in fact do outweigh the rewards, particularly when one also accounts for the opportunity costs associated with experimentation. However, the potential impact, both on patient outcomes and in terms of the knowledge generated, could be substantial, and merit further investigation.
\subsection{Limitations}
This study has several limitations. First, as this study is observational in nature, our analysis necessarily relies on certain assumptions, which, while plausible, are unverifiable. The unconfoundedness assumption that we make presumes no unmeasured confounding, and cannot be verified through inspection of the data nor via statistical tests. It is possible that our results, particularly the notional estimates of overall impact in terms of the numbers of readmissions prevented annually under the CATE-based targeting strategies, are biased. However, it appears that the magnitude of unobserved bias would have to be large to negate our results, particularly our estimates of the potential impacts of the CATE-based strategies. Furthermore, because the causal forest performs local linear estimation, the assumptions that we make with respect to the risk threshold can be considered modular in the sense that if they fail for the subgroup of patients with a predicted risk of below 25\% across both the pre- and post-implementation periods (which we anticipate would be most likely), they would still hold for the $\geq \!\!25$\% risk subgroup.
Moreover, we again note that the CATEs appeared well-calibrated in at least two aspects: the number of readmissions prevented under the risk-based targeting strategy, as estimated using the predicted CATEs, agreed with that estimated via a separate analysis; and the predicted CATEs were never larger than the predicted risk for any individual. Furthermore, this limitation becomes moot if the data are generated by a randomized experiment. We outline several approaches for how existing processes of deploying predictive algorithm-driven interventions could be redesigned to incorporate randomization in order to iteratively refine these interventions through improved targeting.
Second, these estimates of benefit must be computed with respect to an intervention, which cannot be made with historical data alone, as is often done when developing risk assessment tools. As such, the necessary data must be generated, whether through wholesale deployments of an intervention (as in many QI studies) or via randomized experiments. The latter provide a better basis for these analyses, but are costly and require additional infrastructure, and may be subject to more institutional oversight. Third, our simplifying assumption with respect to the nature of the payoffs may not have resulted in a targeting strategy that was in fact optimal, but we hope to explore in future work the feasibility of predicting payoffs using supervised machine learning in parallel with treatment effects within this framework, and observing how the resulting targeting strategies change.
Third, although we incorporated it into our analysis, the \texttt{HCUPSGDC} variable is not uniformly always available at discharge. From the point of a view of a retrospective analysis that principally seeks to characterize treatment effect heterogeneity, this does not constitute a limitation. However, this consideration may preclude the application of this impactibility model prospectively in its current form, but it is possible that the patient's problem list at discharge could be used to infer the value of this variable, as there are only 25 categories used in our analysis. Finally, our results may not be applicable to all settings, particularly hospitals and health systems without a high degree of integration. However, the framework we outline here is agnostic as to application as well as the type of causal machine learning model used, and could be applied to resource allocation and other prediction problems more generally.
\section{Conclusion}
Causal machine learning can be used to identify preventable hospital readmissions, if the requisite interventional data are available. Moreover, our results point to a mismatch between readmission risk and treatment effect, which is consistent with suggestions in prior work. In our setting, the extent of this mismatch was considerable, suggesting that many preventable readmissions may be being "left on the table" with current risk modeling methodology. Our proposed framework is also generalizable to the study of a variety of population health management and quality improvement interventions driven by predictive models, as well as of algorithm-driven interventions in a range of settings outside of healthcare.
\newpage
\section*{Acknowledgements}
The authors are immensely grateful to Colleen Plimier of the Division of Research, Kaiser Permanente Northern California, for assistance with data preparation, as well as to Dr. Tracy Lieu, also of the Division of Research, for reviewing the manuscript. In addition, the authors wish to thank Minh Nguyen, Stephen Pfohl, Scotty Fleming, and Rachael Aikens for helpful feedback on earlier versions of this work.
Mr. Marafino was supported by a predoctoral fellowship from the National Library of Medicine of the National Institutes of Health under Award Number T15LM007033, as well as by funding from a Stanford School of Medicine Dean's Fellowship. Dr. Baiocchi was also supported by grant KHS022192A from the Agency for Healthcare Research and Quality. Dr. Vincent Liu was also supported by NIH grant R35GM128672 from the National Institute of General Medical Sciences. Portions of this work were also funded by The Permanente Medical Group, Inc., and Kaiser Foundation Hospitals, Inc. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The authors have no conflicts of interest to disclose. The funders played no role in the study design, data collection, analysis, reporting of the data, writing of the report, nor the decision to submit the article for publication.
\section*{Appendix A}
\setcounter{table}{0}
\renewcommand{\thetable}{A\arabic{table}}
\begin{table}[H]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lccccc@{}}
\toprule
& Total & Pre-implementation & Post-implementation & \multicolumn{1}{l}{$p$-value} & \multicolumn{1}{l}{SMD} \\ \midrule
Hospitalizations, $n$ & 1,584,902 & 1,161,452 & 423,450 & --- & --- \\
Patients, $n$ & 753,587 & 594,053 & 266,478 & --- & --- \\
Inpatient (\%) & 82.8 (69.7-90.6) & 84.4 (73.1-90.7) & 78.5 (57.7-90.3) & $<0.0001$ & -0.151 \\
Observation (\%) & 17.2 (9.4-30.3) & 15.6 (9.3-26.9) & 21.5 (9.7-42.3) & $<0.0001$ & 0.151 \\
Inpatient stay $<24$ hours & 5.2 (3.3-6.7) & 5.1 (3.8-6.5) & 5.6 (1.8-9.1) & $<0.0001$ & 0.041 \\
Transport-in & 4.5 (1.4-8.7) & 4.5 (1.7-8.8) & 4.5 (0.4-8.4) & 0.56 & -0.001 \\
Age, mean (years) & 65.3 (62.2-69.8) & 65.1 (61.9-69.6) & 65.8 (62.8-70.4) & $<0.0001$ & 0.038 \\
Male gender (\%) & 47.5 (43.4-53.8) & 47.0 (42.4-53.5) & 48.9 (45.3-54.9 & $<0.0001$ & 0.037 \\
KFHP membership (\%) & 93.5 (75.3-97.9) & 93.9 (80.0-98.0) & 92.5 (61.7-97.6) & $<0.0001$ & -0.052 \\
Met strict membership definition (\%) & 80.0 (63.4-84.9) & 80.6 (67.6-85.5) & 78.5 (51.3-83.8) & $<0.0001$ & -0.053 \\
Met regulatory definition (\%) & 61.9 (47.2-69.7) & 63.9 (50.2-72.2) & 56.5 (38.7-66.6) & $<0.0001$ & -0.152 \\
Admission via ED (\%) & 70.4 (56.7-82.0) & 68.9 (56.0-80.3) & 74.4 (58.4-86.6) & $<0.0001$ & 0.121 \\
Charlson score, median (points) & 2.0 (2.0-3.0) & 2.0 (2.0-3.0) & 2.0 (2.0-3.0) & $<0.0001$ & 0.208 \\
Charlson score $\geq 4$ (\%) & 35.2 (29.2-40.7) & 33.2 (28.2-39.8) & 40.9 (33.0-46.2) & $<0.0001$ & 0.161 \\
COPS2, mean (points) & 45.6 (39.1-52.4) & 43.5 (38.4-51.5) & 51.2 (39.7-55.8) & $<0.0001$ & 0.159 \\
COPS2 $\geq 65$ (\%) & 26.9 (21.5-32.0) & 25.3 (21.0-31.6) & 31.1 (22.5-35.4) & $<0.0001$ & 0.129 \\
Admission LAPS2, mean (points) & 58.6 (48.0-67.6) & 57.6 (47.4-65.8) & 61.3 (50.2-72.8) & $<0.0001$ & 0.092 \\
Discharge LAPS2, mean (points) & 46.7 (42.5-50.8) & 46.3 (42.5-50.8) & 47.6 (42.3-52.9) & $<0.0001$ & 0.039 \\
LAPS2 $\geq 110$ (\%) & 12.0 (7.8-16.0) & 11.6 (7.5-15.2) & 12.9 (8.3-18.4) & $<0.0001$ & 0.039 \\
Full code at discharge (\%) & 84.4 (77.3-90.5) & 84.5 (77.7-90.5) & 83.9 (75.9-90.5) & $<0.0001$ & -0.016 \\
Length of stay, days (mean) & 4.8 (3.9-5.4) & 4.9 (3.9-5.4) & 4.7 (3.9-5.6) & $<0.0001$ & -0.034 \\
Discharge disposition (\%) & & & & & 0.082 \\
\quad To home & 72.7 (61.0-86.2) & 73.3 (63.9-85.9) & 71.0 (52.1-86.9) & $<0.0001$ & \\
\quad Home Health & 16.1 (6.9-23.3) & 15.2 (6.9-22.6) & 18.5 (7.0-34.5) & $<0.0001$ & \\
\quad Regular SNF & 9.9 (5.9-14.3) & 10.0 (6.0-15.2) & 9.5 (5.6-12.4) & $<0.0001$ & \\
\quad Custodial SNF & 1.3 (0.7-2.5) & 1.5 (0.8-2.7) & 0.9 (0.4-1.8) & $<0.0001$ & \\
Hospice referral (\%) & 2.6 (1.7-4.4) & 2.6 (1.7-4.6) & 2.7 (1.5-4.0) & $<0.0001$ & 0.007 \\ \midrule
\textbf{Outcomes}& & & & & \\
\quad Inpatient mortality (\%) & 2.8 (2.1-3.3) & 2.8 (2.1-3.3) & 2.8 (1.8-3.3) & 0.17 & -0.003 \\
\quad 30-day mortality (\%) & 6.0 (4.0-7.3) & 6.1 (4.1-7.6) & 5.9 (3.9-6.8) & $<0.0001$ & -0.006 \\
\quad Any readmission (\%) & 14.5 (12.7-17.2) & 14.3 (12.3-17.3) & 15.1 (13.3-17.0) & $<0.0001$ & 0.021 \\
\quad Any non-elective readmission (\%) & 12.4 (10.4-15.4) & 12.2 (10.2-15.5) & 13.1 (10.8-15.4) & $<0.0001$ & 0.029 \\
\quad Non-elective inpatient readmission (\%) & 10.5 (8.2-12.6) & 10.4 (8.1-12.8) & 10.8 (8.6-12.9) & $<0.0001$ & 0.012 \\
\quad Non-elective observation readmission (\%) & 2.4 (1.4-3.7) & 2.2 (1.2-3.4) & 3.0 (1.9-5.6) & $<0.0001$ & 0.049 \\
\quad 30-day post-discharge mortality (\%) & 4.0 (2.6-5.2) & 4.1 (2.7-5.4) & 3.9 (2.3-4.9) & $<0.0001$ & -0.007 \\
\quad Composite outcome (\%) & 15.2 (12.9-18.8) & 15.0 (12.9-19.1) & 15.8 (13.3-18.0) & $<0.0001$ & 0.023 \\ \bottomrule
\end{tabular}}
\caption{Characteristics of the cohort, including both index and non-index stays. Notably, comparing pre- to post-implementation, hospitalized patients were older, and tended to have higher comorbidity burden (higher COPS2) as well as a higher acuity of illness at admission (higher LAPS2). The use of observation stays also increased. These differences reflect a broader trend towards the pool of potential inpatient admissions becoming more and more ill over the decade from 2010, in large part due to the effectiveness of outpatient preventative care processes at KPNC, as well as of programs providing care outside of the hospital setting as an alternative to admission. Otherwise, care patterns did not substantially change, as evidenced by, e.g., transports-in, Kaiser Foundation Health Plan (KFHP) membership status, and discharge disposition mix, all of which had standardized mean differences (SMDs) $<0.1$. In large cohorts such as this one, SMDs can be a better guide to detecting covariate imbalances or differences between groups, owing to the effects of large sample sizes. Finally, as a consequence of increased comorbidity burden and admission acuity, and despite the implementation of the Transitions Program, rates of readmission and of the composite outcome increased from pre- to post-implementation. Abbreviations: SMD, standardized mean difference; KFHP, Kaiser Foundation Health Plan; LAPS2, Laboratory-based Acute Physiology Score, version 2; COPS2, COmorbidity Point Score, version 2; SNF, skilled nursing facility.}
\label{tab:a1}
\end{table}
\section*{Appendix B}
\setcounter{table}{0}
\renewcommand{\thetable}{B\arabic{table}}
\begin{table}[H]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lc@{}}
\toprule
Supergroup name (\texttt{HCUPSGDC}) & \begin{tabular}[c]{@{}c@{}}Clinical Classification Software (CCS) \\ category code(s)\end{tabular} \\ \midrule
Acute CVD & 109 \\
AMI & 100 \\
CAP & 122 \\
Cardiac arrest & 107 \\
CHF & 108 \\
Coma; stupor; and brain damage & 85 \\
Endocrine \& related conditions & 48-51, 53, 54, 56, 58, 200, 202, 210, 211 \\
Fluid and electrolyte disorders & 55 \\
GI bleed & 153 \\
Hematologic conditions & 59-64 \\
Highly malignant cancer & 17, 19, 27, 33, 35, 38-43 \\
Hip fracture & 226 \\
Ill-defined signs and symptoms & 250-253 \\
Less severe cancer & 11-16, 18, 20-26, 28-32, 34, 36, 37, 44-47, 207 \\
Liver and pancreatic disorders & 151, 152 \\
Miscellaneous GI conditions & 137-140, 155, 214 \\
Miscellaneous neurological conditions & 79-84, 93-95, 110-113, 216, 245, 653 \\
Miscellaneous surgical conditions & 86-89, 91, 118-121, 136, 142, 143, 167, 203, 204, 206, 208, 209, 212, 237, 238, 254, 257 \\
Other cardiac conditions & 96-99, 103-105, 114, 116, 117, 213, 217 \\
Other infectious conditions & 1, 3-9, 76-78, 90, 92, 123-126, 134, 135, 148, 197-199, 201, 246-248 \\
Renal failure (all) & 156, 157, 158 \\
Residual codes & 259 \\
Sepsis & 2 \\
Trauma & 205, 225, 227-236, 239, 240, 244 \\
UTI & 159 \\ \bottomrule
\end{tabular}}
\caption{List of Clinical Classification Software (CCS)-defined supergroups and their CCS codes used in this study. These supegroups represent levels of the covariate \texttt{HCUPSGDC}. More details on the CCS codes themselves, as well as mappings to their component ICD codes, can be found at \url{www.ahrq.gov/data/hcup}.}
\label{tab:b1}
\end{table}
\newpage
\printbibliography
\end{spacing}
\end{document}
| 8f7d583e8a9aa7a1d3b49f2fac03392c5ce3ff3e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Diabetic patients are at constant risk of developing Diabetic Retinopathy (DR) that may eventually lead to permanent vision loss if left unnoticed or untreated. In such patients, increased blood sugar, blood pressure, and cholesterol can cause small blood vessels in retina to protrude and, in due course, haemorrhage blood into retinal layers and/or vitreous humour \cite{amin2017method}. In severe conditions, scar tissues and newly proliferated fragile blood vessels blanket the retina and obstruct incoming light from falling on it. As a result, retina is unable to translate light into neural signals which results in blindness. Diabetic retinopathy advances slowly and gradually and may take years to reach proliferative stage, however, almost every diabetic patient is potentially susceptible to this complication.
Timely diagnosis is the key to appropriate prognosis. Ophthalmologists usually detect DR by examining retinal fundus and looking for any signs of microaneurysms (bulging of blood vessels), blood leakage, and/or neovascularization \cite{akram2014detection}. While the indications of advanced stages of DR are rather prominent, these symptoms remain largely discrete in early stages. Figure \ref{fig:drStages} shows progress of DR from healthy to proliferative stage in Retinal Fundus Images (RFIs) taken from EyePACS dataset\footnote{https://www.kaggle.com/c/diabetic-retinopathy-detection/data}. It can be observed from the figure that the difference between healthy and early stages of DR are very subtle and not readily discernible. Manual analysis of these images requires highly qualified and specialized ophthalmologists who may not be easily accessible in developing countries or remote areas of developed countries. Even when medical experts are available, large scale analysis of RFIs is highly time-consuming, labour-intensive and prone to human error and bias. Furthermore, manual diagnosis by clinicians is largely subjective and rarely reproducible and, therefore, inter-expert agreement for a certain diagnosis is generally very poor.
Computer-Aided Diagnosis (CAD) based on deep learning can provide easily accessible, efficient and economical solution for large-scale initial screening of many diseases including diabetic retinopathy. CAD can perform objective analysis of the given image and predict standardized and reproducible diagnosis, which is free from any bias or tiredness. It can not only help physicians by reducing their workload but can also outreach to underprivileged population and afford them the opportunity of swift and cost-effective initial screening, which may effectively prevent advancement of disease into severer stage.
\begin{figure}[h]
\centering
\addtocounter{figure}{-1}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=\textwidth]{EyePACS0r.jpeg}
\caption{(a) Healthy}
\label{fig:healthy}
\end{subfigure}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=\textwidth]{EyePACS1r.jpeg}
\caption{(b) Mild}
\label{fig:mild}
\end{subfigure}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=\textwidth]{EyePACS2r.jpeg}
\caption{(c) Moderate}
\label{fig:moderate}
\end{subfigure}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=\textwidth]{EyePACS3r.jpeg}
\caption{(d) Severe}
\label{fig:severe}
\end{subfigure}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=\textwidth]{EyePACS4r.jpeg}
\caption{(e) Proliferative}
\label{fig:prolif}
\end{subfigure}
\caption{Progression of diabetic retinopathy from healthy to proliferative stage is subtle and gradual. Images are taken from EyePACS train set.}
\label{fig:drStages}
\end{figure}
Convolutional Neural Networks (CNNs) are computer algorithms inspired by biological visual cortex. They work especially well in visual recognition tasks. CNNs have been used to perform at par with or even outperform humans in various challenging image recognition problems \cite{DBLP:journals/corr/abs-1712-00559,lee2016generalizing}. Today automated image recognition can be divided into coarse-grained classification and fine-grained classification. In former case, images are classified into high-level categories like humans, animals, vehicles and other objects in a natural scene, for example. In later case, classification is focused on low-level categories like species of dogs or models of cars etc. Fine-grained classification is particularly challenging owning to high intra-class variations and low inter-class variations. Although DR is also a fine-grained classification task, it has normally been addressed using simple coarse-grained classification algorithms.
In this work we used a combination of general and fine-grained deep CNNs to analyze RFIs and predict automated diagnosis for DR. We used two of the most popular conventional image classification architectures i.e Residual Networks \cite{he2016resnet} and Densely Connected Networks \cite{huang2017densenet}, a network search framework called NASNet \cite{zoph2018nasnet} and two recently proposed methods for fine-grained classification namely NTS-Net \cite{yang2018ntsNet} and SBS Layer \cite{recasens2018saliency}. We tried to harvest the combined potential of these two approaches by training them separately and taking their ensemble during inference. We used EyePACS and Messidor datasets for evaluation. Since previous researches have used vastly disparate experimental setups, we cannot directly compare our results with most of them. However, we performed a broad range of experiments, following the most common problem settings in the literature like normal vs abnormal, referable vs non-referable, ternary and quaternary classification in order to define benchmarks which will afford future works with an opportunity of fair comparison.
\subsection{Related Work}\label{sec:relatedWork}
Over the past decade, machine learning and deep learning have been used to detect various pathologies, segment vessels and classify DR grades using RFIs. Welikala et al. \cite{welikala2014automated}~ detected proliferative DR by identifying neovascularization. They used an ensemble of two networks trained separately on 100 different patches for each network. The patches are taken from a selected set of 60 images collected from Messidor \cite{messidor2014feedback} and a private dataset. Since the dataset had only 60 images they performed leave-one-out cross validation and achieved 0.9505 Area Under the Curve (AUC) and sensitivity of 1 with specificity of 0.95 at the optimal operating point. Wang et al. \cite{wang2017zoom} identified suspicious regions in RFIs and classified DR into normal (nDR) vs abnormal (aDR) and referable (rDR) vs non-referable (nrDR). They developed a CNN based model called Zoom-in-Network to identify important regions. To classify an image the network uses the overview of the whole images and pays particular attention to important regions. They took 182 images from EyePACS dataset and let a trained ophthalmologist draw bounding boxes around 306 lesions. On Messidor dataset they achieved 0.921 AUC, 0.905 accuracy and 0.960 sensitivity at 0.50 specificity for nDR vs aDR.
Gulshan et al. \cite{gulshan2016development} conducted a comprehensive study to distinguish rDR from nrDR grades. They trained a deep CNN on 128175 fundus images from a private dataset and tested on 9963 images from EyePACS-I and 1748 images of Messidor-2. They achieved AUC of 0.991 on EyePACS-I and 0.990 on Messidor-2. Guan et al. \cite{guan2018said} proposed that modeling each classifier after individual human grader instead of training a single classifier using average grading of all human experts improves classification performance. They trained 31 classifiers using a dataset of total 126522 images collected from EyePACS and three other clinics. The method is tested on 3547 images from EyePACS-I and Messidor-2, and achieved 0.9728 AUC, 0.9025 accuracy, and 0.8181 specificity at 0.97 sensitivity. However, it would have been more interesting if they had provided comparison of their suggested approach with ensemble of 31 networks modeled after average grading. Costa et al. \cite{costa2018end} used adversarial learning to synthesize color retinal images. However, the performance of their classifier trained on synthetic images was less than the classifier trained on real images. Aujih et al. \cite{aujih2018analysis} found that blood vessels play important role in disease classification and fundus images without blood vessels resulted in poor performance by the classifier.
The role of multiple filter sizes in learning fine-grained features was studied by Vo et al. \cite{vo2016new}. To this end they used VGG network with extra kernels and combined kernels with multiple loss networks. They achieved 0.891 AUC for rDR vs nrDR and 0.870 AUC for normal vs abnormal on Messidor dataset using 10-fold cross validation. Somkuwar et al. \cite{somkuwar2015intensity} performed classification of hard exudates by exploiting intensity features using 90 images from Messidor dataset and achieved 100\% accuracy on normal and 90\% accuracy on abnormal images. Seoud et al. \cite{seoud2016red} focused on red lesions in RFIs, like haemorrhages and microaneurysms, and detected these biomarkers using dynamic shape features in order to classify DR. They achieved 0.899 AUC and 0.916 AUC for nDR vs aDR and rDR vs nrDR, respectively on Messidor. Rakhlin et al. \cite{rakhlin2018diabetic} used around 82000 images taken from EyePACS for training and around 7000 EyePACS images and 1748 images from Messidor-2 for testing their deep learning based classifier. They achieved 0.967 AUC on Messidor and 0.923 AUC on EyePACS for binary classification. Ramachandran et al. \cite{ramachandran2018diabetic} used 485 private images and 1200 Messidor images to test a third party deep learning based classification platform, which was trained on more than 100000 images. Their validation gave them 0.980 AUC on Messidor dataset for rDR vs nrDR classification. Quellec et al. \cite{quellec2017deep} capitalized a huge private dataset of around 110000 images and around 89000 EyePACS images to train and test a classifier for rDR vs nrDR grades and achieved 0.995 AUC on EyePACS.
\vspace{-0.4 cm}
\section{Materials and Methods}
\vspace{-0.3 cm}
This section provides details on the datasets used in this work and the ensemble methodology employed to perform classification.
\subsection{Datasets}
We used EyePACS dataset published publicly by Kaggle for a competition on Diabetic Retinopathy Detection. Table \ref{tab:EyePACS} gives overview of EyePACS dataset. Although this dataset is very large in size, only about 75\% of its images are of reasonable quality that they can be graded by human experts \cite{rakhlin2018diabetic}. EyePACS is graded on a scale of 0 to 4 in accordance with International Clinical Diabetic Retinopathy (ICDR) guidelines \cite{ICDR2002}. However, low gradability of this dataset raises suspicions on the fidelity of labels provided with each image. We pruned the train set to get rid of 657 completely uninterpretable images. For testing on EyePACS we used 33423 images randomly taken from test set.
\begin{table}[h]
\caption{Overview of EyePACS Dataset. IRMA stands for IntraRetinal Microvascular Abnormalities}\label{tab:EyePACS}
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}cp{6cm}cccc@{}}
\toprule
\multirow{2}{*}{Severity Grade} & \multicolumn{1}{c}{\multirow{2}{*}{Criterion}} & \multicolumn{2}{c}{Train Set} & \multicolumn{2}{c}{Test Set} \\ \cmidrule(l){3-6}
& \multicolumn{1}{c}{} & Images & Percentage & Images & Percentage \\ \midrule
0 & No Abnormalities & 25810 & 73.48 & 39533 & 73.79 \\ \hline
1 & Microaneurysms Only & 2443 & 6.95 & 3762 & 7.02 \\ \hline
2 & More than just microaneurysms but less than Grade 3 & 5292 & 15.07 & 7861 & 14.67 \\ \hline
3 & \begin{tabular}[c]{@{}p{6cm}@{}}More than 20 intraretinal hemorrhages in each of 4 quadrants\\ OR Definite venous beading in 2+ quadrants\\ OR Prominent IRMA in 1+ quadrant\\ AND no signs of proliferative retinopathy\end{tabular} & 873 & 2.48 & 1214 & 2.27 \\ \hline
4 & \begin{tabular}[c]{@{}p{6cm}@{}}Neovascularization\\ OR Vitreous/preretinal hemorrhage\end{tabular} & 708 & 2.02 & 1206 & 2.25 \\ \hline
\multicolumn{2}{c}{Total} & 35126 & 100 & 53576 & 100 \\ \bottomrule
\end{tabular}
}
\end{table}
Messidor dataset consists of 1200 images collected at three different clinics in France. Each clinic contributed 400 images. This dataset is graded for DR on a scale of 0 to 3 following the criteria given in Table \ref{tab:Messidor}. Messidor dataset is validated by experts and is, therefore, of higher quality than EyePACS in terms of both image quality and labels.
\begin{table}[h]
\caption{Overview of Messidor Dataset}\label{tab:Messidor}
\centering
\begin{tabular}{clcc}
\hline
Severity Grade & \multicolumn{1}{c}{Criterion} & Images & Percentage \\ \hline
0 & \begin{tabular}[c]{@{}l@{}}No microaneurysms\\ AND No haemorrhages\end{tabular} & 546 & 45.50 \\ \hline
1 & \begin{tabular}[c]{@{}l@{}}Microaneurysms \textless{}= 5\\ AND No haemorrhages\end{tabular} & 153 & 12.75 \\ \hline
2 & \begin{tabular}[c]{@{}l@{}}5 \textless Microaneurysms \textless 15\\ AND 0 \textless Haemorrhages \textless 5\\ AND No Neovascularization\end{tabular} & 247 & 20.58 \\ \hline
3 & \begin{tabular}[c]{@{}l@{}}Microaneurysms \textgreater{}= 15\\ OR Haemorrhages \textgreater{}= 5\\ OR Neovascularization\end{tabular} & 254 & 21.17 \\ \hline
\multicolumn{2}{c}{Total} & 1200 & 100 \\ \hline
\end{tabular}%
\end{table}
\subsection{Methodology}
\begin{figure}[b!]
\centering
\includegraphics[width=\textwidth]{ensembleDR.png}
\caption{System Overview: Combining Coarse-Grained and Fine-Grained Classifiers}
\label{fig:systemOverview}
\end{figure}
Figure \ref{fig:systemOverview} illustrates complete pipeline of the system combining coarse-grained and find-grained classifiers.
Before feeding an image to the network, we first applied Otsu Thresholding to extract and crop retinal rim from RFI and get rid of irrelevant black background. Since the images in both datasets are taken with different cameras and under different clinical settings, they suffer from large brightness and color variations. We used adaptive histogram equalization to normalize brightness and enhance the contrast of visual artefacts which are critical for DR detection. Since the images are in RGB colour space, we first translate them into YCbCr colour space to distribute all luminosity information in Y channel and colour information in Cb and Cr channels. Adaptive histogram equalization is then applied on Y channel only and the resultant image is converted back to RGB colour space. We further normalized the images by subtracting local average colour from each pixel to highlight the foreground and help our network detect small features. Figure \ref{fig:PreProcess} shows the effects of preprocessing steps on RFIs. These pre-processed images are then used to train all five networks individually. During inference, each network gives diagnosis which are ensembled to calculate the final prediction.
\begin{figure}[t]
\centering
\addtocounter{figure}{-1}
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{original}
\caption{(a) Original Image before Preprocessing}
\label{fig:original}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{CLAHE.jpeg}
\caption{(b) After Contrast Enhancement}
\label{fig:histEq}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{subtract-local-average.jpeg}
\caption{(c) After Local Average Colour Subtraction}
\label{fig:lvcSubtr}
\end{subfigure}
\caption{Effects of preprocessing steps on retinal fundus images}
\label{fig:PreProcess}
\end{figure}
\subsubsection{Experimental Setup}
From EyePACS train set, we randomly selected 30000 images for training and rest of the 4469 images were used for validation. Test set of EyePACS was used for reporting results on this dataset. From Messidor, we used 800 images for training and 400 images from Lariboisi\`ere Hospital for testing (as done by Lam et al. \cite{carson2018automated}). We employed a broad range of hyper parameters during training. All networks are initialized with pre-trained weights and fine-tuned on ophthalmology datasets. To evaluate these models on EyePACS and Messidor datasets under similar problem settings, we first parallelized DR grades of both datasets using criteria given in Figure \ref{fig:grades}.
\begin{figure}[b]
\centering
\includegraphics[width=0.70\textwidth]{grades.png}
\caption{Conversion of EyePACS grades to Quaternary, Ternary and Binary Classification}
\label{fig:grades}
\end{figure}
\section{Results and Analysis}
From section \ref{sec:relatedWork} we observe that previous works on EyePACS and Messidor have used disparate train and test splits and different classification tasks for example Quaternary, Ternary and Binary (rDR vs nrDR and nDR vs aDR). Furthermore, different researchers use different performance metrics to evaluate their method. Therefore, in such scenario comparison of any two works in not directly possible \cite{abramoff2016improved}. However, we conducted extensive experiments to perform all four classification tasks mentioned above and report comprehensive results to allow a rough comparison with some of the published state-of-the-art results on these datasets.
\subsection{Results of Binary Classification}
As discussed above, many previous works focus primarily on binary classification as nDR vs aDR or rDR vs nrDR grading. The criteria to convert 4 or 5 grades into binary grades is given in Figure \ref{fig:grades}. For our binary classification, the number of images used for training, validation and testing from EyePACS and Messidor are given in Table \ref{tab:distNAB} and \ref{tab:distRNR}. It can be seen from the tables that there is extensive class imbalance between both classes. Table \ref{tab:allResults} provides detailed performance metrics for all classification tasks including nDR vs aDR classification. Our results are competitive to that of Wang et al. in terms of accuracy and for all other metrics we outperform them. It should be noted here that Wang et al. performed 10-fold cross validation and although their sensitivity of 96 is higher than our 89.75, it is calculated at 50\% specificity while ours is at 90\% specificity.
\begin{table}
\resizebox{\textwidth}{!}{%
\parbox[t]{.49\linewidth}{
\centering
\caption{Class Distribution for Normal vs Abnormal Classification}\label{tab:distNAB}
\resizebox{0.50\textwidth}{!}{
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{Grade}} & \multicolumn{3}{c}{EyePACS} & \multicolumn{3}{c}{Messidor} \\ \cline{2-7}
\multicolumn{1}{c}{} & Train & Validate & Test & Train & Validate & Test \\ \hline
Normal & 22668 & 2744 & 24741 & 346 & 49 & 151 \\
Abnormal & 7332 & 1725 & 8682 & 354 & 51 & 249 \\
Total & 30000 & 4469 & 33423 & 700 & 100 & 400 \\ \hline
\end{tabular}
}
}
\hfill
\parbox[t]{.49\linewidth}{
\centering
\caption{Class Distribution for Referable vs Non-Referable Classification}\label{tab:distRNR}
\resizebox{0.50\textwidth}{!}{
\begin{tabular}{lcccccc}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{Grade}} & \multicolumn{3}{c}{EyePACS} & \multicolumn{3}{c}{Messidor} \\ \cline{2-7}
\multicolumn{1}{c}{} & Train & Validate & Test & Train & Validate & Test \\ \hline
Non-Referable & 28825 & 4177 & 31937 & 453 & 65 & 181 \\
Referable & 1175 & 292 & 1486 & 247 & 35 & 219 \\
Total & 30000 & 4469 & 33423 & 700 & 100 & 400 \\ \hline
\end{tabular}
}
}
}
\end{table}
\begin{table}[]
\centering
\caption{Detailed Performance Metrics for Various Classification Settings}\label{tab:allResults}
\resizebox{\textwidth}{!}{%
\begin{tabular}{@{}lcccccccc@{}}
\toprule
\multicolumn{9}{c}{\textbf{Results of Binary (Normal vs Abnormal) Classification}} \\ \midrule
\multicolumn{1}{c}{\multirow{2}{*}{\textbf{Model}}} & \multicolumn{2}{c}{\textbf{Accuracy (\%)}} & \multicolumn{2}{c}{\textbf{AUC (\%)}} & \multicolumn{2}{c}{\textbf{Sensitivity (\%)}} & \multicolumn{2}{c}{\textbf{Specificity (\%)}} \\ \cmidrule(l){2-9}
\multicolumn{1}{c}{} & EyePACS & Messidor & EyePACS & Messidor & EyePACS & Messidor & EyePACS & Messidor \\ \midrule
NTS-Net & \textbf{88.19} & 88.00 & 92.72 & 95.20 & 88 & 88.00 & 72 & 87.51 \\
SBS Layer & 80.11 & 89.50 & 86.20 & 95.17 & 80 & 89.50 & 54 & \textbf{92.07} \\
ResNet-50 & 82.86 & 87.75 & 89.46 & 95.06 & 83 & 87.75 & 75 & 90.49 \\
DenseNet-201 & 82.66 & 87.75 & 89.69 & 95.89 & 83 & 87.75 & \textbf{77} & 88.14 \\
NASNet & 82.19 & 87.25 & 88.49 & 95.04 & 82 & 87.25 & 73 & 89.14 \\
\textbf{Ensemble} & 87.74 & \textbf{89.75} & \textbf{93.44} & \textbf{96.50} & \textbf{88} & \textbf{89.75} & 75 & 91.44 \\
Vo et. al & N/A & 87.10 & N/A & 87.00 & N/A & 88.2 & N/A & 85.7 \\
Wang et. al & N/A & 90.50 & N/A & 92.10 & N/A & 96 & N/A & 50 \\
Soud et. al & N/A & N/A & N/A & 89.90 & N/A & N/A & N/A & N/A \\ \midrule
\multicolumn{9}{c}{\textbf{Results of Binary (Referable vs Non-Referable) Classification}} \\ \midrule
NTS-Net & 94.93 & \textbf{93.25} & 99.10 & \textbf{96.56} & 95 & \textbf{93} & 75 & \textbf{94} \\
SBS Layer & \textbf{95.89} & 88.75 & \textbf{99.44} & 94.90 & \textbf{96} & 89 & 67 & 90 \\
ResNet-50 & 95.08 & 86.75 & 98.97 & 94.95 & 95 & 87 & 81 & 89 \\
DenseNet-201 & 94.70 & 89.25 & 99.05 & 95.33 & 95 & 89 & 82 & 91 \\
NASNet & 91.98 & 87.50 & 97.45 & 95.16 & 92 & 88 & \textbf{85} & 89 \\
\textbf{Ensemble} & 95.34 & 89.25 & 99.23 & 96.45 & 95 & 89 & 81 & 91 \\
Lam et. al & N/A & 74.5 & N/A & N/A & N/A & N/A & N/A & N/A \\
Vo et. al & N/A & 89.70 & N/A & 89.10 & N/A & 89.3 & N/A & 90 \\
Wang et. al & N/A & 91.10 & N/A & 95.70 & N/A & 97.8 & N/A & 50 \\
Seoud et. al & N/A & 74.5 & N/A & 91.60 & N/A & N/A & N/A & N/A \\
\midrule
\multicolumn{9}{c}{\textbf{Results of Ternary Classification}} \\ \midrule
NTS-Net & 84.43 & 84.50 & 94.89 & 94.61 & 84 & 85 & 72 & \textbf{94} \\
SBS Layer & 76.93 & 84.50 & 90.95 & 94.12 & 77 & 85 & 50 & 91 \\
ResNet-50 & 81.23 & 80.50 & 93.51 & 93.79 & 81 & 81 & 74 & 92 \\
DenseNet-201 & 79.20 & 80.25 & 92.87 & 94.25 & 79 & 80 & \textbf{77} & 93 \\
NASNet & 78.95 & 81.75 & 91.93 & 94.00 & 79 & 82 & 71 & 89 \\
\textbf{Ensemble} & \textbf{84.94} & \textbf{85.25} & \textbf{95.28} & \textbf{95.40} & \textbf{85} & \textbf{85} & 73 & 92 \\
Lam et. al & N/A & 68.8 & N/A & N/A & N/A & N/A & N/A & N/A \\ \midrule
\multicolumn{9}{c}{\textbf{Results of Quaternary Classification}} \\ \midrule
NTS-Net & 82.53 & 74.50 & 95.72 & 91.84 & 83 & 75 & \textbf{76} & \textbf{92} \\
SBS Layer & 82.00 & 65.00 & 95.69 & 88.43 & 82 & 65 & 67 & 88 \\
ResNet-50 & 81.82 & 70.25 & 95.53 & 91.31 & 82 & 70 & 71 & 89 \\
DenseNet-201 & 79.38 & 74.00 & 95.04 & 92.26 & 79 & 74 & 75 & 91 \\
NASNet & 73.73 & 71.75 & 92.06 & 90.84 & 74 & 72 & 74 & 86 \\
\textbf{Emsemble} & \textbf{83.42} & \textbf{76.25} & \textbf{96.31} & \textbf{92.99} & \textbf{83} & \textbf{76} & 73 & 91 \\
Lam et. al & N/A & 57.2 & N/A & N/A & N/A & N/A & N/A & N/A \\ \bottomrule
\end{tabular}%
}
\end{table}
Results of rDR vs nrDR classification can also be found in Table \ref{tab:allResults}. All networks performed significantly better for this task than for normal vs abnormal classification on EyePACS dataset reaching maximum accuracy around 96\% with 99.44\% AUC using SBS layer architecture. For Messidor dataset, both NTS-Net and SBS Layer stand out from traditional classifiers. NTS-Net outperforms all other methods in all metrics, whereas ensemble of all methods gives sub-optimal performance than individual fine-grained methods. This can happen when majority of classifiers used for ensembling have a skewed performance towards downside and only a few give standout results.
\subsection{Results of Multi-Class Classification}
The complexity of classification task was gradually increased from binary to ternary and quaternary classification. Table \ref{tab:dist4c} and \ref{tab:dist3c} show the class distribution in train, validation and test splits for this multi-class setting. For ternary classification we used the criterion used by \cite{carson2018automated}, as shown in Figure \ref{fig:grades}.
\begin{table}
\parbox[t]{.49\linewidth}{
\centering
\caption{Class Distribution for 4-Class Classification}\label{tab:dist4c}
\resizebox{0.50\textwidth}{!}{
\begin{tabular}{ccccccc}
\hline
\multicolumn{1}{l}{\multirow{2}{*}{Grade}} & \multicolumn{3}{c}{EyePACS} & \multicolumn{3}{c}{Messidor} \\ \cline{2-7}
\multicolumn{1}{l}{} & Train & Validate & Test & Train & \multicolumn{1}{l}{Validate} & \multicolumn{1}{l}{Test} \\ \hline
0 & 22668 & 2744 & 24741 & 346 & 49 & 151 \\
1 & 6157 & 1433 & 7196 & 107 & 16 & 30 \\
2 & 685 & 166 & 753 & 155 & 22 & 70 \\
3 & 490 & 126 & 733 & 92 & 13 & 149 \\
Total & 30000 & 4469 & 33423 & 700 & 100 & 400 \\ \hline
\end{tabular}
}
}
\hfill
\parbox[t]{.49\linewidth}{
\centering
\caption{Class Distribution for 3-Class Classification}\label{tab:dist3c}
\resizebox{0.50\textwidth}{!}{
\begin{tabular}{ccccccc}
\hline
\multicolumn{1}{l}{\multirow{2}{*}{Grade}} & \multicolumn{3}{c}{EyePACS} & \multicolumn{3}{c}{Messidor} \\ \cline{2-7}
\multicolumn{1}{l}{} & Train & Validate & Test & Train & \multicolumn{1}{l}{Validate} & \multicolumn{1}{l}{Test} \\ \hline
0 & 22668 & 2744 & 24741 & 346 & 49 & 151 \\
1 & 6157 & 1433 & 7196 & 107 & 16 & 30 \\
2 & 1175 & 292 & 1486 & 247 & 35 & 219 \\
Total & 30000 & 4469 & 33423 & 700 & 100 & 400 \\ \hline
\end{tabular}
}
}
\end{table}
Performance of individual networks and their ensemble for ternary and quaternary classification is given in Table \ref{tab:allResults}. Ensemble of all models gave better performance in this case. We also observe that the performance of NTS-Net is higher than all other individual networks. Our accuracies for both ternary and quaternary classification are superior than accuracies reported by Lam et al. \cite{carson2018automated}. Figure \ref{fig:confMats} provides a detailed overview of classification performance of ensemble.
\begin{comment}
\begin{figure}[t]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{rocNAB.png}
\caption{Normal/Abnormal}
\label{fig:referable_roc}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{rocRNR4to2.png}
\caption{Referable/Non-Referable}
\label{fig:rocRNR}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{roc3c.png}
\caption{3-ary Classification}
\label{fig:roc3c}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\textwidth]{roc4cEP.png}
\caption{4-ary Classification}
\label{fig:roc4c}
\end{subfigure}
\caption{ROC Curves for all Classification Tasks}
\label{fig:ROCs}
\end{figure}
\end{comment}
\begin{figure}[b!]
\centering
\addtocounter{figure}{-1}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMatNABEP.png}
\caption{N/AB EyePACS}
\label{fig:confMat-NAB-EP}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMatRNR4to2EP.png}
\caption{R/NR EyePACS}
\label{fig:confMat-RNR-EP}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMat3cEP.png}
\caption{3-Class EyePACS}
\label{fig:confMat-3c-EP}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{confMat4cEP.png}
\caption{4-Class EyePACS}
\label{fig:confMat-4c-EP}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMatNABM.png}
\caption{N/AB Messidor}
\label{fig:confMat-NAB-M}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMatRNR4to2M.png}
\caption{R/NR Messidor}
\label{fig:confMat-RNR-M}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMat3cM.png}
\caption{3-Class Messidor}
\label{fig:confMat-3c-M}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{ConfMat4cM.png}
\caption{4-Class Messidor}
\label{fig:confMat-4c-M}
\end{subfigure}
\caption{Confusion Matrices for EyePACS and Messidor for all Classification Tasks. N/AB refers to Normal vs Abnormal; R/NR refers to Referable vs Non-Referable}
\label{fig:confMats}
\end{figure}
\section{Conclusion}
Diabetic Retinopathy detection using retinal fundus images is a fine-grained classification task. The biomarkers of this disease on retinal images are usually very small in size, especially for early stages, and are scattered all across the image. The ratio of pathologically important region to the whole input volume is therefore minuscule. Due to this reason traditional deep CNNs usually struggle to identify regions of interest and do not learn discriminatory features well. This problem of small and distributed visual artefacts coupled with unavailability of large publicly available high quality dataset with reasonable class imbalance makes DR detection particularly challenging for deep CNN models. However, fine-grained classification networks have high potential to provide standardized and large scale initial screening of diabetic retinopathy and help in prevention and better management of this disease. These networks are equipped with specialized algorithms to discover the important region from the image and pay particular heed to learn characterizing features from those regions.
We achieved superior performance for diabetic retinopathy detection on binary, ternary and quaternary classification tasks than many previously reported results. However, due to hugely different experimental setups and choice of performance metrics, it is unfair to draw a direct comparison with any of the cited research. Nevertheless, we have provided a wide spectrum of performance metrics and detailed experimental setup for comparison by any future work.
\bibliographystyle{splncs04}
| c7cc7140ad356ea2d360ae2bb84e6a29087ba6f8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Abstract}
Researchers and managers model ecological communities to infer the
biotic and abiotic variables that shape species' ranges, habitat use,
and co-occurrence which, in turn, are used to support management decisions
and test ecological theories. Recently, species distribution models
were developed for and applied to data from ecological communities.
Model development and selection for ecological community data is difficult
because a high level of complexity is desired and achieved by including
numerous parameters, which can degrade predictive accuracy and be
challenging to interpret and communicate. Like other statistical models,
multi-species distribution models can be overparameterized. Regularization
is a model selection technique that optimizes predictive accuracy
by shrinking or eliminating model parameters. For Bayesian models,
the prior distribution automatically regularizes parameters. We propose
a tree shrinkage prior for Bayesian multi-species distributions models
that performs regularization and reduces the number of regression
coefficients associated with predictor variables. Using this prior,
the number of regression coefficients in multi-species distributions
models is reduced by estimation of unique regression coefficients
for a smaller number of guilds rather than a larger number of species.
We demonstrated our tree shrinkage prior using examples of presence-absence
data for six species of aquatic vegetation and relative abundance
data for 15 species of fish. Our results show that the tree shrinkage
prior can increase the predictive accuracy of multi-species distribution
models and enable researchers to infer the number and species composition
of guilds from ecological community data. The desire to incorporate
more detail and parameters into models must be tempered by the limitations
of the data and objectives of a study, which may include developing
models that have good predictive accuracy and output that is easier
to share with policymakers. Our tree shrinkage prior reduces model
complexity while incorporating ecological theory, expanding inference,
and can increase the predictive accuracy and interpretability of Bayesian
multi-species distribution models. \bigskip{}
\begin{doublespace}
\noindent \textbf{\textit{\emph{Key-words: }}}Bayesian statistics,
community ecology, ecological guild, joint species distribution model,
model averaging, model selection, multi-species distribution model,
prior distribution, regularization\vspace{-1cm}
\end{doublespace}
\section*{Introduction}
Over the past 20 years species distribution models (SDMs) have become
one of the most widely used quantitative methods in ecology (\citealt{elith2009species,franklin2010mapping,guisan2017habitat}).
Species distribution models are routinely fit to presence-only, presence-absence,
and abundance data to understand the biotic and abiotic variables
that influence the habitat use and geographic distribution of a species
(\citealt{aarts2012comparative,mcdonald2013location,hefleyHSDM}).
The output of SDMs are used in scientific studies to test ecological
theories and in application to delineate areas of high conservation
value which informs management decisions (e.g., Wisz et al. \citeyear{wisz2013role};
\citealt{guisan2013predicting,hefley2015use}). Regardless of the
type of data or method, the ultimate goal when using SDMs includes
making reliable inference and accurate predictions.
In many applications, a SDM is fit to data from a single species.
If the study objectives require the analysis of data from multiple
species, then a SDM is fit to data from each species and the models
are combined using stacking (Norberg et al. \textit{in press} \nocite{ecomonograph2019};
\citeauthor{zurelltesting} \textit{in press}). Although this species-by-species
approach is common it presents several challenges that limit the ability
of researchers and managers to make reliable inference and accurate
predictions (\citealt{hui2013mix,madon2013community,taylor2017joint}).
For example, if the goal of a study is to produce accurate range maps
then spatial predictions for a species that is rare will lack precision.
Imprecise spatial predictions result from parameter estimates that
have a high variance, which occurs because parameters are estimated
using limited data from only a single rare species (e.g., \citealt{hefley2015existence}).
As another example, consider the case where the goal is to infer the
influence of a predictor variable such as temperature. Fitting a SDM
for each species requires estimating a unique temperature coefficients
for each species. Even if there are only a few predictor variables
it will be difficult to summarize the regression coefficients for
a moderate number of species, which will hinder communication with
policymakers.
Recently, SDMs for multiple species have been developed. So-called
joint- or multi-species distribution models (MSDMs) enable a single
model to be fit simultaneously to data from multiple species; the
benefits include: 1) community-level inference and resulting prediction;
2) incorporation of biotic interactions; 3) more accurate predictions
of single-species' distributions due to ``borrowing of strength''
across data from co-occurring species; and 4) the potential for model
simplification.
Many types of MSDMs have been developed using statistical or machine
learning approaches (see synthesis in Norberg et al. \textit{in press}
\nocite{ecomonograph2019} and \citealt{wilkinson2019comparison}).
Hierarchical Bayesian modeling is a statistical approach that is easily
customizable and widely used to model the distribution of multiple
species (e.g., \citealt{taylor2017joint,johnson2017modeling,lany2018asymmetric,schliep2018joint,ovaskainen2017make,wilkinson2019comparison}).
Briefly, hierarchical Bayesian models are specified using conditional
probability distributions that represent the data collection process
(i.e., the ``data model''), the latent ecological process (i.e.,
the ``process model''), and prior knowledge about parameters (i.e.,
the ``parameter model''; \citealt{wikle2019spatio}, pgs. 11\textendash 13;
\citealt{hobbs2015bayesian}, ch. 6).
For presence-absence data, \citet{wilkinson2019comparison} describes
the quintessential hierarchical Bayesian MSDM. This includes the data
model
\begin{equation}
y_{i,j}|z_{i,j}\sim\text{Bern(}g^{-1}(z_{i,j}))\,,
\end{equation}
where $y_{i,j}=1$ if the $i^{\text{th}}$ site ($i=1,2,...,n$) is
occupied by the $j^{\text{th}}$ species ($j=1,2,...,J$) and $y_{i,j}=0$
if absent, ``Bern'' is a Bernoulli distribution, $g^{-1}(\cdot)$
is an inverse link function (e.g., logistic, probit), and $z_{i,j}$
is the latent process on the link scale. For MSDMs, a widely used
process model is
\begin{equation}
\mathbf{z}|\boldsymbol{\mu},\ensuremath{\boldsymbol{\Sigma}}\sim\text{N(\ensuremath{\boldsymbol{\mu}}\ensuremath{,}\ensuremath{\boldsymbol{\Sigma}})\,,}
\end{equation}
where $\mathbf{z}$ is a random vector with elements $z_{i,j}$, $\boldsymbol{\mu}$
is expected value (mean), and $\ensuremath{\boldsymbol{\Sigma}}$
is the variance-covariance matrix of a multivariate normal distribution.
The vector, $\boldsymbol{\mu}$, has elements $\mu_{i,j}$ and is
typically specified using
\begin{equation}
\mu_{i,j}=\alpha_{j}+\mathbf{x}_{i}^{'}\boldsymbol{\beta}_{j}\,,
\end{equation}
where, for the $j^{\text{th}}$ species, $\alpha_{j}$ is the intercept
and $\boldsymbol{\beta}_{j}$ are the $K$ regression coefficients
(i.e., $\boldsymbol{\beta}_{j}\equiv(\beta_{j,1},\beta_{j,2},...,\beta_{j,K})^{'}$).
The vector $\mathbf{x}_{i}$ contains $K$ measured predictor variables
at the $i^{\text{th}}$ site, which are the same for all $J$ species
(i.e., $\mathbf{x}_{i}\equiv(x_{i,1},x_{i,2},...,x_{i,K})^{'}$).
In Eq. 2, the process model is specified jointly (i.e., using a multivariate
distribution) to enable sharing of information across species.
In previous studies, the variance-covariance matrix, $\boldsymbol{\Sigma}$
in Eq. 2, accounted for residual autocorrelation due to space, time,
and/or biotic interactions. For example, within the literature on
MSDMs, a main focus of methodological developments has been on parametrizing
$\boldsymbol{\Sigma}$ to account for biotic factors, such as species
co-occurrence, that can not be explained by the measured predictor
variables $\mathbf{x}_{i}$ (e.g., \citealt{warton2008penalized,warton2015so,ovaskainen2017make,niku2017generalized,wilkinson2019comparison}).
By comparison, however, there are only a few studies that have focused
on the parameterization of $\boldsymbol{\mu}$ in Eq. 2 (e.g., \citealt{johnson2017modeling}).
Regardless of the type of data, most MSDMs use a linear equation to
model the influence of predictor variables. For example, all seven
methods compared by \citet{wilkinson2019comparison} use Eq. 3 which,
if there are $J$ species and $K$ predictor variables, will result
in $J\times K$ regression coefficients. For even a small number of
predictor variables and species, the overabundance of regression coefficients
will be challenging to summarize and interpret and may result in a
MSDM that is overparameterized.
For many Bayesian MSDMs, the parameter model or prior is specified
using simple distributions. For example, in six of the seven methods
presented by \citet{wilkinson2019comparison} priors for the regression
coefficients were specified using uniform distributions or independent
normal distributions with known hyperparameters (e.g., $\beta_{j,k}\sim\text{N}(0,10)$).
Although using simple parameter models is common practice, more sophisticated
parameter models are useful. For example, the seventh method in \citet{wilkinson2019comparison}
uses a multivariate normal distribution with unknown hyperparameters
as a parameter model which enables estimation and inference regarding
the correlation among regression coefficients. For MSDMs, the notion
of a model for the parameters is important because it can enable novel
inference and perform regularization. Regularization is a model selection
technique used to optimize predictive accuracy by controlling model
complexity (\citealt{bickel2006regularization,hooten2015guide}).
\citet{johnson2017modeling} develop a hierarchical Bayesian MSDMs,
which includes a parameter model that leverages the concept of ecological
guilds (\citealt{simberloff1991guild}). For MSDMs, the ecological
guild concept is useful because it provides a framework to incorporate
ecological theory and reduce the number of regression coefficients
associated with predictor variables. For example, the ecological guild
concept suggests that some species may respond to predictor variables
in a similar direction and magnitude as a result of similar resource
use or ecological role. In other words, species within the same guild
may be associated with similar values of the regression coefficients
in Eq. 3. If the regression coefficients for two or more species are
effectively the same, then the complexity of the MSDM can be reduced
by estimating unique regression coefficients for a smaller number
of guilds rather than a larger number of species.
If the guild membership is known, then incorporating the guild concept
into MSDMs is trivial and involves using Eq. 3 with a guild by predictor
variable interaction effect rather than a species by predictor variable
(\citealt{johnson2017modeling}). In nearly all applications, the
number and species compositions of the guilds is unknown and must
be estimated from data which can be accomplished using model-based
techniques. For example, as a heuristic imagine that the species-specific
regression coefficients for a single predictor variable in Eq. 3 were
known. In this example, the number and species composition of guilds
could be determined by classifying or clustering the regression coefficients
into homogeneous groups. The classification model would be easier
to interpret when compared to the larger number of species-specific
regression coefficients because it provides a data-driven approach
to determine which species can be modeled with a single guild-level
regression coefficient. This ad-hoc approach is a useful heuristic
because it gives insight into how a large number of regression coefficients
can be summarized using a model for the parameters. Within a hierarchical
Bayesian framework, formalizing this concept is straightforward by
specifying an appropriate parameter model.
Simple independent parameter models that are commonly used for MSDMs,
such as $\beta_{j,k}\sim\text{N}(0,10)$, do not leverage the information
in the data shared among species. The independence assumption eliminates
the possibility that information about parameters will be shared across
species. Shared information among the species, however, can be accessed
by incorporating the guild concept into the parameter model. For example,
\citet{johnson2017modeling} developed a hierarchical Bayesian MSDM
using a Dirichlet process mixture for the parameter model, which enables
regression coefficients to be estimated using data from all species
within the same guild. Briefly, the Dirichlet process mixture is a
distribution that can be used as a parameter model to cluster the
species into guilds while simultaneously fitting a MSDM to community
data.
Classification and regression trees (CART; \citealt{CART1984}) are
a machine learning approach with widespread use in ecology (\citealt{de2000classification,de2002multivariate,elith2008working}).
For ecological data, CART offer a semi-automated approach to build
predictive models that are easy to interpret. Although Bayesian variants
of CART have been available for some time (e.g., \citealt{chipman1998bayesian,denison1998bayesian}),
they are rarely used by ecologists. This is perhaps because accessible
versions of Bayesian CART software are not easily modified to accommodate
ecological data. For example, current implementations of Bayesian
CART do not include a ``data model,'' which is needed to account
for imperfect measurements of ecological processes. Conceptually,
CART can be embedded within hierarchical Bayesian models (at any level),
which would enable ecologists to exploit the benefit of both approaches
(\citealt{shaby2012embedding}).
In our work, we show how a specific type of CART can be used as a
parameter model for Bayesian MSDMs. Similar to \citet{johnson2017modeling},
our approach leverages the concept of ecological guilds but reduces
model complexity via a novel tree shrinkage prior (TSP). We illustrate
the TSP by fitting MSDMs to presence-absence and relative abundance
data collected to inform management of freshwater aquatic vegetation
and fish populations. These data are collected as part of a long-term
environmental monitoring program intended to inform management of
the Upper Mississippi River System. Managers are required to use these
data to make and justify management recommendations. This requires
understanding the predictor variables that influence the distribution
of a large number of species using models that can be easily shared
with policymakers but that are capable of making accurate predictions.
\begin{spacing}{1.9}
\section*{Materials and methods}
\end{spacing}
\begin{spacing}{1.9}
\subsection*{INCORPORATING GUILDS INTO MULTI-SPECIES DISTRIBUTION MODELS}
\end{spacing}
\noindent The mechanics of incorporating guilds into hierarchical
Bayesian MSDMs was presented by \citet{johnson2017modeling}. For
MSDMs, this involves a simple modification of Eq. 3
\begin{equation}
\mathbf{\boldsymbol{\mu}}_{i}=\mathbf{\boldsymbol{\alpha}}+(\mathbf{x}_{i}^{'}\otimes\mathbf{Z})\boldsymbol{\gamma}\,,
\end{equation}
where, $\mathbf{\boldsymbol{\mu}}_{i}\equiv(\mu_{i,1},\text{\ensuremath{\mu_{i,2}},...,\ensuremath{\mu_{i,J}}})^{'}$
is a vector that specifies the expected value for the process model
at the $i^{\text{th}}$ site for $J$ species, $\boldsymbol{\alpha}\equiv(\alpha_{1},\alpha_{2},...,\alpha_{J})^{'}$
contains intercept parameters for each species, and $\mathbf{x}_{i}\equiv(x_{i,1},x_{i,2},...,x_{i,K})^{'}$
contains $K$ predictor variables measured at the $i^{\text{th}}$
site. The matrix $\mathbf{Z}$ has dimensions $J\times G$ where the
$j^{\text{th}}$ row contains an indicator variable that links the
predictor variables to the guild-specific regression coefficients
contained within the vector $\boldsymbol{\gamma}$. The mathematical
symbol $\otimes$ denotes a Kronecker product. For example, let there
be $J=4$ species, $G=2$ guilds, and $K=2$ predictor variables,
then for the $i^{\text{th}}$ site, one configuration of Eq. 4 is
\begin{equation}
\begin{bmatrix}\mu_{i1}\\
\mu_{i2}\\
\mu_{i3}\\
\mu_{i4}
\end{bmatrix}=\begin{bmatrix}\alpha_{1}\\
\alpha_{2}\\
\alpha_{3}\\
\alpha_{4}
\end{bmatrix}+\left(\begin{bmatrix}x_{i1} & x_{i2}\end{bmatrix}\otimes\begin{bmatrix}1 & 0\\
1 & 0\\
0 & 1\\
0 & 1
\end{bmatrix}\right)\begin{bmatrix}\gamma_{1,1}\\
\gamma_{2,1}\\
\gamma_{1,2}\\
\gamma_{2,2}
\end{bmatrix}\,,
\end{equation}
where the first and second species are members of the first guild
and the third and fourth species are members of the second guild.
The guild-specific regression coefficients are $\boldsymbol{\gamma}\equiv(\gamma_{1,1},\gamma{}_{2,1},\gamma_{1,2},\gamma_{2,2})^{'}$
where the subscripts, $\gamma_{g,k}$, indicate the $g^{\text{th}}$
guild and the $k^{\text{th}}$ predictor variable.
Modifications to the example given in Eq. 5, illustrates that incorporating
the guild concept into MSDM results in a model that is flexible and
includes two special cases. For example, if $\mathbf{Z}$ is a $4\times1$
matrix with all elements equal to one (i.e., $\mathbf{Z}\equiv(1,1,1,1)^{'}$),
then all species would be in a single guild. Similarly if $\mathbf{Z}$
is a $4\times4$ identity matrix (i.e., a matrix with diagonal elements
equal to one and zero on the off diagonal elements), then each species
would be assigned to a unique guild; this results in the commonly
used MSDM with unique species-specific regression coefficients (i.e.,
Eq. 3) and demonstrates that incorporating the guild concept into
MSDMs can only improve the predictive accuracy.
If the number and species composition of the guilds was known, then
the matrix $\mathbf{Z}$ would also be known. In practice, however,
the number and species composition of the guilds are unknown, which
requires estimation of $\mathbf{Z}$. In turn, $\mathbf{Z}$ regularizes
the MSDM because $\mathbf{Z}$ shrinks and reduces the number of regression
coefficients from $J\times K$ to $G\times K$.
\begin{spacing}{1.9}
\subsection*{\vspace{0.66cm}
TREE SHRINKAGE PRIOR}
\end{spacing}
\noindent A general introduction to tree-based methods is given in
\citeauthor{james2013introduction} (2013; ch. 8) while \citet{linero2017review}
provides a technical review of Bayesian tree approaches. We use vocabulary
found in both \citeauthor{james2013introduction} (2013; ch. 8) and
\citet{linero2017review}. The idea of a TSP is not new but, to our
knowledge, has been used in only one unrelated setting (\citeauthor{guhaniyogi2018large}
\textit{in press}). For MSDMs, the idea is simple; the species within
terminal nodes of the tree represent the guilds and determines the
matrix $\mathbf{Z}$ in Eq. 4. Each terminal node is associated with
guild-level regression coefficients for the predictor variables. For
example, in what follows, we use presence-absence data for six species
of aquatic vegetation and a single predictor variable (water depth).
An example of a binary tree is given in Fig. 1 and shows which species
share the same numerical value of the regression coefficients for
the predictor variable water depth. In this example, the standard
MSDM results in six regression coefficients (i.e., one for each species).
For the predictor variable water depth, a MSDM specified using the
tree shown in Fig. 1 results in only three regression coefficients.
Recall that there are intercept parameters for each species, as a
result the binary tree in Fig. 1 reduces the number of parameters
in the MSDM from 12 to 9.
There are a variety of techniques to construct tree priors discussed
within the statistics and machine learning literature (\citealt{linero2017review}).
Most techniques rely on a tree generating stochastic process (\citealt{linero2017review}).
Binary trees, like the example in Fig. 1, are a type of stochastic
branching process (\citealt{dobrow2016introduction}, ch. 4). Specification
of a tree generating stochastic process involves specifying probability
distributions that generate a tree structure as a random variable.
For example, a simple stochastic binary tree generating process for
the six species of aquatic vegetation in Fig. 1 involves specifying
a distribution that determines the splitting of parent nodes and a
second distribution that determines the allocation of species to the
child nodes (e.g., in Fig. 1 super guild 1 is a parent node with guilds
1 and 2 as child nodes). For the aquatic vegetation example, one way
to specify a binary tree generating stochastic process would involve
using a ``splitting distribution'' that determines if a parent node
should be split into two child nodes. If a draw from the splitting
distribution results in a value of one, the parent node is split and
draws from a ``assignment distribution'' allocates the species within
the parent node to one of the two child nodes. In practice, the assignment
distribution is $\text{Bern}(0.5)$ whereas the splitting distribution
is $\text{Bern}(p_{split})$, where the hyperparameter $p_{split}$
controls the model complexity by inducing a prior on the guilds. For
example, if $p_{split}=1$ then a binary tree is generated with six
terminal nodes (guilds) that contain only a single species; when $p_{split}=0$
a tree with no splits is generated and all six species are in the
same guild. Finally, the terminal nodes of the MSDM with a TSP has
guild-level regression coefficients, $\gamma_{g,k}$, which require
the specification of priors. For tree-based methods, simple models
like $\gamma_{g,k}\sim\text{N(0,10)}$ are commonly used, (\citealt{linero2017review}),
however, more complex models have been developed (e.g., \citealt{gramacy2008bayesian}).
In sum, specifying a stochastic tree generating process results in
an implied prior on the tree structure. For example, a so-called non-informative
prior could be constructed by specifying a tree generating stochastic
process that generates all possible combinations of guilds with equal
probability (e.g., \citealt{atkinson1992generating}). For the aquatic
vegetation example such a prior would result in 63 unique guilds.
For this example, the MSDM with the TSP could be implemented using
Bayesian model averaging (\citealt{oliver1995pruning}). Model averaging
is a well-known technique among ecologists (e.g., \citealt{hobbs2015bayesian};
Dormann et al. 2018\nocite{dormann2018model}), however, this would
require fitting 63 models (i.e., a model for each unique guild). Although
we could fit a MSDM with the TSP to the aquatic vegetation data using
Bayesian model averaging, this approach would not work for a larger
number of species. For example, the fisheries data used to illustrate
the TSP contains 15 species which results in 32,767 unique guilds.
Despite advances in computational statistics, implementation of tree-based
methods using Bayesian model averaging is challenging because of the
potential for a large number trees (\citealt{chipman2001practical,hooten2015guide,BB2L},
ch. 15). In the next section, we describe alternative strategies for
model fitting that may be less familiar among ecologists.
\subsection*{\vspace{0.66cm}
MODEL FITTING}
Hierarchical Bayesian MSDMs can be fit to presence-absence, abundance,
and presence-only data using a variety of algorithms. Markov chain
Monte Carlo (MCMC) algorithms are a type of stochastic sampling approach
routinely used by ecologists and implemented in standard software
such as WinBugs, JAGS, and Stan (\citealt{BB2L}). As such, we demonstrate
the TSP using MCMC.
Compared with other commonly used ecological models, MCMC based implementation
of Bayesian trees usually requires highly specialized algorithms.
As a result, an inability to easily construct efficient algorithms
has likely slowed the adoption of Bayesian regression trees (\citealt{linero2017review}).
For example, Bayesian model averaging, used to fit Bayesian trees
to data, requires specialized algorithms (e.g., \citealt{oliver1995pruning,hernandez2018bayesian}).
As mentioned, however, this may be computationally challenging or
infeasible for the TSP when there is a large number of guilds. To
increase computational efficiency, search algorithms have been used
to identify higher probability tree structures, however, as noted
by \citet{chipman1998bayesian}, the ``procedure is a sophisticated
heuristic for finding good models rather than a fully Bayesian analysis,
which presently seems to be computationally infeasible.'' Significant
progress continues in the development of computationally efficient
implementations of Bayesian trees (\citealt{linero2017review}), however,
fitting tree-based models to data requires matching the model to computational
algorithms which can be challenging for non-experts.
An appealing implementation for ecologists is presenting by \citet{shaby2012embedding}.
The \citet{shaby2012embedding} approach enables off-the-shelf software
for tree-based methods, such as those available in R, to be embedded
within hierarchical Bayesian models. This is appealing to ecologists
because it makes the TSP accessible by eliminating the need to develop
custom software to implement the TSP. Briefly, the technique of \citet{shaby2012embedding}
is an approximate Gibbs sampler that enables machine learning algorithms
to be embedded within hierarchical Bayesian models. In what follows
we use a model-based recursive partitioning approach that is available
in the R package ``partykit'' and implemented within the \texttt{lmtree(...)}
function (\citealt{hothorn2006unbiased,zeileis2008model}; \citeauthor{zeileisparties}
2019). Model-based recursive partitioning estimates a binary tree
and consequently $\mathbf{Z}$ and $\boldsymbol{\gamma}$ in Eq. 4.
The approximate Gibbs sampler of \citet{shaby2012embedding} is similar
to an empirical Bayesian step within a Gibbs sampler because priors
for a portion of the unknown model parameters are replaced with estimates
(\citealt{casella2001empirical}). As a result, when using the \citet{shaby2012embedding}
approach, there are no distributions or hyperparameters associated
with the TSP that must be specified.
In what follows, we use standard MCMC algorithms for MSDMs, but employ
the \citet{shaby2012embedding} approach. Constructing the approximate
Gibbs sampler using the \citet{shaby2012embedding} approach for the
aquatic vegetation data example is incredibly simple and requires
only six lines of R code and the ability to sample univariate normal
and truncated normal distributions (e.g., using the R functions \texttt{rnorm(...)}
and \texttt{rtruncnorm(...)}; see section 2.3 in Appendix S1). To
assist readers implementing similar MSDMs with the TSP, we provide
tutorials with the computational details, annotated computer code,
and the necessary code to reproduce all results and figures related
to the analysis for the aquatic vegetation data and fisheries data
examples (see Appendix S1 and S2). \vspace{0.66cm}
\subsection*{STATISTICAL INFERENCE}
A benefit of Bayesian inference is that uncertainty regarding the
structure of the binary tree, and therefore the number and species
composition of the guilds, is automatically accounted for. Under the
Bayesian paradigm, samples from the posterior distribution of the
trees are obtained during the model fitting process. The posterior
distribution of the trees can be summarized to infer characteristics
of the guilds. For example, the species composition of the guilds
with the highest probability (given the data) can be obtained from
the posterior distribution of the trees.
In addition, posterior distributions of the species-level and guild-level
regression coefficients can be sampled and summarized. The species-level
regression coefficients, akin to $\boldsymbol{\beta}$ in Eq. 3, are
a derived quantity calculated using
\begin{equation}
\boldsymbol{\beta}_{j}^{(m)}=\mathbf{W}_{j}^{(m)}\boldsymbol{\gamma}^{(m)}\,,
\end{equation}
where the superscript $m$ is the current MCMC sample, $\boldsymbol{\beta}_{j}$
is vector of regression coefficients for the $j^{\text{th }}$ species,
$\mathbf{W}_{j}$ is a matrix that contains indicator variables linking
the guild-level coefficients to the species. In the presence of guild
membership uncertainty, the species-level regression coefficients
result in a posterior distribution that is a mixture of the guild-level
regression coefficients.
When using the TSP, the posterior distribution of the regression coefficients
for a species are akin to the model averaged posterior distribution
obtained from fitting and averaging multiple MSDMs with differing
species composition and numbers of guilds to the data. Similarly the
guild-level regression coefficients are akin to the regression coefficients
obtained from a single MSDM using a fixed guild structure. We recommend
using the species-level regression coefficients to make inference
because, with a finite amount of data, the species composition of
the guilds will always be uncertain. Inference using the guild-level
regression coefficients is conditional (i.e., valid for a fixed number
and species composition of the guilds) and does not fully account
for uncertainty.
\subsection*{\vspace{0.66cm}
LONG-TERM RESOURCE MONITORING DATA}
\noindent The Mississippi River is the second longest river in North
America and flows a distance of 3,734 km. The Upper Mississippi River
System includes approximately 2,100 km of the Mississippi River channel
and drains a basin covering 490,000 $\text{km}^{2}$ that contains
30 million people. The Upper Mississippi River Restoration Program
was authorized under the Water Resources Development Act of 1986 (\citet{waterlaw}).
The U.S. Army Corps of Engineers provides guidance, has overall Program
responsibility, and established a long term resource monitoring (LTRM)
element that includes collecting biological and geophysical data.
The LTRM element is implemented by the U.S. Geological Survey, Upper
Midwest Environment Sciences Center, in cooperation with the five
Upper Mississippi River System states of Illinois, Iowa, Minnesota,
Missouri, and Wisconsin (\citealt{UMRR}). The expressed goal of the
LTRM element is to ``support decision makers with the information
and understanding needed to maintain the Upper Mississippi River System
as a viable multiple-use large river ecosystem.'' The biological
data collected by LTRM contains a large number of species and exemplifies
the need for interpretable MSDMs with results that are easy to communicate
to policymakers. For example, data from 154 species, some of which
are rare or endangered, have been collected by the LTRM element since
1993; inferences from these data provide context for management recommendations
for these or collections of these species.
As part of LTRM element, data on aquatic vegetation and fish are collected
at multiple sites within six study reaches of the Upper Mississippi
River. To illustrate our method, we use aquatic vegetation and fisheries
data from navigation pool 8 collected in 2015. Complete details of
data collection procedures for the aquatic vegetation data and fisheries
data are given by \citet{vegsampling} and \citet{fishsampling} respectively.
\begin{spacing}{1.9}
\subsection*{\vspace{0.66cm}
EXAMPLE 1: PRESENCE-ABSENCE OF AQUATIC VEGETATION}
\end{spacing}
\noindent Our objective for the first example is to illustrate the
TSP using a simple MSDM. As such, this example was designed to demonstrate
the proposed methods rather than to provide scientifically valid inference.
As such, we use only a single predictor variable (water depth) and
presence-absence absence data of six species of aquatic vegetation.
At 450 unique sites, the presence or absence of a species was recorded.
For our example, we use data for broadleaf arrowhead (\textit{Sagittaria
latifolia}), Canadian waterweed (\textit{Elodea canadensis}), coontail
(\textit{Ceratophyllum demersum}), stiff arrowhead (\textit{Sagittaria
rigida}), white water lily (\textit{Nymphaea odorata}), and wild celery
(\textit{Vallisneria americana}). This resulted in a total of 2,700
observations. For our analysis, we randomly selected 225 sites and
used data from these for model fitting. We used data from the remaining
225 sites to test the predictive accuracy of the models.
For the aquatic vegetation data, we use the data model
\begin{equation}
\mathbf{y}_{i}|\ensuremath{\mathbf{\boldsymbol{\mu}}_{i}}\sim\text{Bern(\ensuremath{\Phi}(\ensuremath{\mathbf{\boldsymbol{\mu}}_{i}})})\,,
\end{equation}
where $\mathbf{y}_{i}\equiv(y_{i,1},y_{i,2}...,y_{i,6})^{'}$ is the
presence or absence of the six species at the $i^{\text{th}}$ site
($i=1,2,...,225)$. The parameter vector $\ensuremath{\mathbf{\boldsymbol{\mu}}_{i}}\equiv(\mu_{i,1},\mu_{i,2},...,\mu_{i,6})^{'}$
is mapped to a probability by applying the inverse probit link function,
$\Phi(\cdot)$, to each element. The parameter vector $\ensuremath{\mathbf{\boldsymbol{\mu}}_{i}}$
is specified using Eq. 4, but is simplified due to the single predictor
variable in this example
\begin{equation}
\mathbf{\boldsymbol{\mu}}_{i}=\mathbf{\boldsymbol{\alpha}}+x_{i}\mathbf{Z}\boldsymbol{\gamma}\,,
\end{equation}
where $\boldsymbol{\alpha}\equiv(\alpha_{1},\alpha_{2},...,\alpha_{6})^{'}$
is a vector that contains intercept parameters for each species and
$x_{i}$ is the recorded water depth at the $i^{\text{th}}$ site.
As in Eq. 4, the matrix $\mathbf{Z}$ has dimensions $J\times G$
where the $j^{\text{th}}$ row contains an indicator variable that
links the predictor variables to the guild-specific regression coefficients
contained within the vector $\boldsymbol{\gamma}\equiv(\gamma_{1,1},\gamma_{2,1},...,\gamma_{g,1})^{'}$.
For the intercept, we use the prior $\alpha_{j}\sim\text{N(0,1})$
because this results in a uniform distribution when $\alpha_{j}$
is transformed using the function $\Phi(\ensuremath{\alpha_{j}})$.
That is, using the prior $\alpha_{j}\sim\text{N(0,1})$ results in
$\Phi(\ensuremath{\alpha_{j}})\sim\text{unif(0,1})$.
We used a computationally efficient MCMC algorithm to fit the MSDM
to the aquatic vegetation data. This MCMC algorithm relies on an auxiliary
variable formulation of the binary regression model and has been used
to fit MSDMs to presence-absence data (\citealt{albert1993bayesian};
\citealt{wilkinson2019comparison}; \citealt{BB2L}, pgs. 246\textendash 253).
For each model fit, we use one chain and obtain 100,000 draws from
the posterior distribution, but retain every $10^{\text{th}}$ sample
to decrease storage requirements. Each model fit takes approximately
30 minutes using a desktop computer (see Appendix S1). We inspect
the trace plot for each parameter to ensure that the Markov chain
is sampling from the stationary distribution (i.e., the posterior
distribution). We discard the first 500 samples and use the remaining
9,500 samples for inference.
The model-based recursive partitioning approach available in the R
function \texttt{lmtree(...)} and used to implement the TSP requires
the specification of a tuning parameter (hereafter \texttt{alpha};
see Appendix S1). This tuning parameter, \texttt{alpha}, must be between
zero and one and controls the splitting of parent nodes. Similar to
the tree generating stochastic process (see Section TREE SHRINKAGE
PRIOR), when \texttt{alpha$\,=1$} the tree has terminal nodes (guilds)
that contain only a single species. When \texttt{alpha$\,=0$} the
tree does not split and all six species are in the same guild. One
way to estimate the tree tuning parameter, \texttt{alpha}, is to find
the value that results in the most accurate predictions (\citealt{hobbs2015bayesian}).
For this example, we illustrate two approaches by estimating the predictive
accuracy of our models with scores that use either in-sample or out-of-sample
data. For the out-of-sample score, we use data from the 225 sites
that were not used to fit the model to calculate the log posterior
predictive density (LPPD). For the in-sample score, we calculate the
Watanabe-Akaike information criteria (WAIC). For comparison between
scoring approaches, we report $-2\times\text{LPPD}$ as we would expect
this and WAIC to result in similar numerical values (\citealt{gelman2014understanding}).
For Bayesian models, computing in-sample scores like WAIC require
a measurement of a model's complexity. In non-hierarchical non-Bayesian
contexts, a model's complexity is the number of parameters in the
model. For example, using the MSDM applied to the aquatic vegetation
data, the number of parameters ranges from 7 (i.e., six intercept
parameters and one regression coefficient) to 12 (i.e., six intercept
parameters and six regression coefficients). Measuring the complexity
of our Bayesian MSDM with the TSP is more complicated than simply
counting the number of parameters because of the shrinkage effect
and because it is not obvious how many parameters are required to
estimate the number and species composition of the guilds. Fortunately,
when calculating Bayesian information criterion like WAIC, measure
of model complexity such as the ``effective number of parameters''
are available without any additional effort. To understanding how
the TSP controls the complexity of the MSDM, we report the effective
number of parameters using Eq. 12 from \citet{gelman2014understanding}
as \texttt{alpha} varies from zero to one.\vspace{0.66cm}
\begin{spacing}{1.9}
\subsection*{EXAMPLE 2: FISH ABUNDANCE}
\end{spacing}
Our objective for the second example is to demonstrate a different
and complex data type that necessitates a more sophisticated use of
the TSP and deeper ecological interpretation, similar to what ecologists
may encounter in practice. Here we use relative abundance data from
15 species of fish that were sampled at 83 sites, which included several
species that were present at only a small number of sites and are
considered rare or species of concern (Table 1). For example, at 79
of the 83 sites (95\%), a count of zero was recorded for river redhorse
(\textit{Moxostoma carinatum}; Table 1). At each site, numerous predictor
variables were measured, but for our analysis we use water temperature,
speed, and depth as these were most relevant to management of the
Upper Mississippi River System.
The fisheries data were collected over three time periods which included
period 1 (June 15 \textendash{} July 31), period 2 (August 1 \textendash{}
September 14), and period 3 (September 15 \textendash{} October 31).
The time periods are hypothesized to correspond to changes in habitat
use. For example, one hypothesis is that species composition of the
guilds and the regression coefficient estimates associated with water
temperature, speed, and depth may be different during each period.
In what follows, we demonstrate how to incorporate such dynamics into
MSDMs using the fisheries data.
For relative abundance data, \citet{johnson2017modeling} discuss
several commonly used MSDMs including a data model that has a marginal
distribution of
\begin{equation}
\mathbf{y}_{i}|\mathbf{z}_{i},\phi\sim\text{ZIP}(e^{\mathbf{z}_{i}},\phi)\,,
\end{equation}
where ``ZIP'' stands for zero-inflated Poisson, $e^{\mathbf{z}_{i}}$
is the expected value of the Poisson mixture component, and $\phi$
is the mixture probability (\citealt{BB2L}; pg. 320). We specify
the process model using Eq. 2 with $\boldsymbol{\Sigma\equiv}\sigma_{\varepsilon}^{2}\mathbf{I}$
and
\begin{equation}
\mathbf{\boldsymbol{\mu}}_{i}=\mathbf{\boldsymbol{\alpha}}+\left(\mathbf{x}_{i}^{'}\otimes\mathbf{Z}_{t}\right)\boldsymbol{\gamma}_{t},
\end{equation}
where the guild composition and regression coefficients vary over
the three sample periods as indicated by subscript $t$ such that
$\mathbf{Z}_{t}$ and $\boldsymbol{\gamma}_{t}$ corresponds to period
1 ($t=1$), period 2 ($t=2$), or period 3 ($t=3$). The vector $\mathbf{x}_{i}\equiv(x_{i,1},x_{i,2},x_{i,3})^{'}$
are the three predictor variables (water temperature, speed and depth)
measured at the $i^{\text{th}}$ site. In total, there are $135$
unique regression coefficients if the guilds contained only a single
species for all three periods, which would also be the case if the
standard MSDM (i.e., Eq. 3) with simple priors for the regression
coefficients was used. For the mixture probability, process model
variance, and intercepts we used the priors $\phi\sim\text{unif}(0,1)$,
$\sigma_{\varepsilon}^{2}\sim\text{IG}(2.01,1)$, and $\alpha_{j}\sim\text{N(0,10}^{3})$
respectively.
For each model fit, we use one chain and obtain 30,000 draws from
the posterior distribution, but retain every $10^{\text{th}}$ sample
to decrease storage requirements. We inspect the trace plot for each
parameter to ensure the Markov chain is sampling from the stationary
distribution. We discard the first 500 samples and use the remaining
2,500 samples for inference.
The computational cost to implement the TSP will increase as the number
of species increase. In addition, allowing the number and species
composition of the guilds to vary over the three periods increases
the computational burden because the MSDM requires three TSPs instead
of one. As a results, we do not select the value of the tree tuning
parameter parameters that produces the best predictive model because
fitting the MSDM once to the fisheries data takes about 23 hours using
a desktop computer (see Appendix S2). Instead we choose a value of
the tree tuning parameters that favors a small number of guilds, which
we anticipate will make our results easier to share with policymakers.
Similar to the aquatic vegetation example, the tree tuning parameters
could be chosen to optimize the predictive ability of the model but
this would make the annotated computer code provided in Appendix S2
difficult for many readers to use without access to high-performance
computing resources.\vspace{0.66cm}
\section*{Results}
\begin{spacing}{1.9}
\subsection*{EXAMPLE 1: PRESENCE-ABSENCE OF AQUATIC VEGETATION}
\end{spacing}
The value of the tree tuning parameter that produces the most accurate
predictions was 0.6 when scored using LPPD and 0.025 when scored using
WAIC. In what follows, we present results obtained from the MSDM with
the TSP that used a value of the tree tuning parameter of 0.025 because
this value was optimal based on WAIC, results in only a slight reduction
in predictive accuracy when scored using LPPD, and yields a relatively
large reduction in the number of parameters which simplifies statistical
inference (Fig. 2). For the six species of aquatic vegetation, the
posterior distributions for the species-level regression coefficient,
obtained from Eq. 6, shows that the presence of wild celery has no
statistical relationship with water depth while the remaining five
species of vegetation exhibit a strong negative response to increases
in water depth (Fig. 3a,b). In addition to the posterior distribution
for the regression coefficients, derived quantities from the posterior
distribution of the tree can provide insightful inference. For example,
Fig 3c shows the posterior distribution of the number of terminal
nodes (i.e., guilds) and Fig 3d shows the expected value of the posterior
distribution of species co-occurrence within the same guild, both
of which were derived from the posterior distribution of the trees.
As another example, summaries of the posterior distribution of the
tree may be of interest. For example, Fig. 3e shows the most likely
tree, which is obtained from the mode of the posterior distribution
of the trees and can be used to infer which guild structure is most
likely. For the most likely tree, figure 3f shows the posterior distribution
of the two guild-level regression coefficients. Again, caution must
be taken when making inference from the guild-level regression coefficients
because these are conditional on a single tree and do not account
for guild uncertainty. For the aquatic vegetation data, however, the
inference does not differ among the species-level and guild-level
regression coefficients because the uncertainty in the number and
species composition of the guilds is relatively low.\vspace{0.66cm}
\begin{spacing}{1.9}
\subsection*{EXAMPLE 2: FISH ABUNDANCE}
\end{spacing}
\noindent Similar to the aquatic vegetation example, we show the posterior
distribution of the number of guilds (Fig. 4a,b,c), the expected value
of species co-occurrence within the same guild (Fig. 4d,e,f), and
the posterior distribution of the regression coefficients for each
species (Figs. 5 and S1). During period 1, our results indicate that
there are most likely two guilds (Fig 4a). The posterior distributions
of the regression coefficients show that eight species (bluegill,
bullhead minnow, emerald shiner, largemouth bass, mud darter, pugnose
minnow, tadpole madtom, and weed shiner; hereafter ``guild 1 during/in
period 1'') have large decreases in relative abundance when water
temperature and depth increases (Figs. 5 and S1). These results suggest
guild 1 during period 1 is more abundant in shallower side channels
and off-channel habitats compared to the main channel, where depth
is maintained at a minimum of 2.75 m for navigation. Of the species
in guild 1 during period 1 many are nest spawners (e.g., bluegill,
bullhead minnow, largemouth bass, pugnose minnow, tadpole madtom)
and our results correspond with habitat requirements for successful
reproduction. During period 1, the remaining seven species (golden
redhorse, longnose gar, river redhorse, river shiner, shorthead redhorse,
smallmouth bass, spotfin shiner; hereafter ``guild 2 during/in period
1'') have posterior distributions of the regression coefficients
that indicate slight decrease in relative abundance when water temperature
or depth increases, however, here is a non-negligible probability
that these responses could be close to zero or even positive (Figs.
5 and S1). This suggest that during period 1 species in guild 2 are
found over a broader range of conditions when compared to species
in guild 1. Many of these species are classified as fluvial dependent
or fluvial specialists, which are either generally found in lotic
environments or require flowing water for some part of their life
history (\citealt{galat2001conserving,lifehistory}).
During period 2, our results indicate that there is most likely a
single guild that contains all 15 species of fish (Fig 4b). The posterior
distributions of regression coefficients for all species generally
have an expected value near zero, suggesting random associations with
water depth, speed, and temperature (Fig. 5 and S1). The lack of a
response to water depth, speed, and temperature during period 2 may
be a result of active foraging in which species are moving in search
of food.
During period 3, our results indicate that there are most likely two
guilds (Fig 4c). The posterior distributions of the regression coefficients
associated with depth indicate a negative response for five species
(bluegill, largemouth bass, pugnose minnow, tadpole madtom, weed shiner;
hereafter ``guild 1 during/in period 3''), which suggests an avoidance
of the main channel (Figs. 5 and S1). In addition, our results show
a negative response to water speed for the species in guild 1 during
period 3 which suggest associations with lentic habitats such as backwater
lakes. As temperatures decrease during period 3 (September 15 \textendash{}
October 31), many of the species in guild 1 are known to move into
slow-moving backwaters to minimize energy expenditures during winter
conditions. During period 3, six species (bullhead minnow, emerald
shiner, river redhorse, shorthead redhorse, smallmouth bass, spotfin
shiner; ``guild 2 during/in period 3'') had a slightly positive
association with water speed and negative association with water temperature
(Figs. 5 and S1). These results suggest that species in guild 2 during
period 3 are more likely to inhabit the main channel and side channels
during this period than off-channel habitats. During period 3, there
are four species (golden redhorse, longnose gar, mud darter, and river
shiner) whose posterior distributions of regression coefficients do
not clearly delineate guild association. These results suggest either
a high level of guild uncertainty, insufficient data, or weaker associations
with the variables examined.
In summary, the number and composition of guilds changed over the
three periods, suggesting shifts in habitat associations and species
interactions through time. Whether these shifts were related to specific
requirements for spawning, foraging, or overwintering would require
additional field research designed to answer such questions, but our
results are consistent with our understanding of fluvial dependence
and general habitat guilds.
\section*{Discussion}
A benefit of hierarchical Bayesian modeling is that statistical models
can be easily customized to match the goals of a study. In what follows,
we discuss modifications to the data and/or process models that have
been commonly used to accommodate different types and quality of data.
The modifications we discuss can be incorporated without making major
changes to the basic framework we presented. We begin by discussing
modifications to the data model. For our aquatic vegetation and fisheries
data examples we used data models that were appropriate for presence-absence
and relative abundance data, respectively. Other types of ecological
data may require the specification of different distributions for
the data. For example, a beta distribution is often used for plant
cover data, which are usually reported as percentages (\citealt{wright2017statistical};
\citeauthor{damgaard2019using} \textit{in press}). Our approach,
using the process model in Eq. 3 and TSP, could be used to develop
a MSDM for plant cover data by specifying a beta distribution as the
data model. Similarly, selecting a data model that matches key characteristics
of the observed data (e.g., the support) is a common model building
practice and applications using the TSP are straightforward (\citealt{hobbs2015bayesian,BB2L}).
Many types of ecological data are measured with error. For example,
presence-absence data may contain false-negatives (\citealt{tyre2003improving,martin2005zero,gray2013influences}).
For MSDMs, accounting for measurement error in the data is an important
component of the model building process (\citealt{beissinger2016incorporating,warton2016extending,tobler2019joint}).
For example, multi-species occupancy models have been developed to
account for false-negatives in presence-absence community data (\citealt{dorazio2005estimating}).
Model selection is challenging for multi-species occupancy models
(\citealt{broms2016model}), however, using the approaches demonstrated
in this paper, it is straightforward to implement the TSP for the
multi-species occupancy model. This would involve adding a fourth
level to the hierarchical MSDM in Eqs. 2\textendash 3 that accounts
for the possibility of false-negatives in the presence-absence data
(\citealt{BB2L}, ch. 23). Using the information provided in Appendix
S1 in concert with \citet{dorazio2012gibbs} or \citeauthor{BB2L}
(2019; ch. 23) would yield an efficient implementation that requires
a minimal amount of additional programming.
As noted in the introduction, a main focus of methodological development
within the literature on MSDMs has been on parametrizing the variance-covariance
matrix, $\boldsymbol{\Sigma}$, for the process model in Eq. 2. In
previous studies the variance-covariance matrix was specified to account
for residual autocorrelation due to space, time, or biotic interactions.
These modifications to the variance-covariance matrix can be incorporated
into a MSDM that uses the TSP. For example, latent variables parameterizations
have been used to induce dependence among species and estimate $\boldsymbol{\Sigma}$
(\citealt{walker2011random,warton2015so,hui2016boral}). As another
example, \citet{johnson2017modeling} note that spatial or temporal
autocorrelation can be accounted for by including basis functions
to estimate $\boldsymbol{\Sigma}$ (\citealt{hefley2017basis}). In
both the latent variables parameterization and basis function approach,
the matrix $\boldsymbol{\Sigma}$ is implicitly specified using a
``random effects'' or ``first-order representation'' (\citealt{hefley2017basis};
\citealt{wikle2019spatio}, pgs. 156\textendash 157).
The TSP is easy to program using the approximate Gibbs sampling approach
from \citet{shaby2012embedding} and requires only a few additional
lines of code when compared to Bayesian MSDMs that use simple priors.
The computational burden to implement the TSP, however, is considerably
larger which may increase the run-time required to fit MSDMs to data.
For example, fitting independent models for single-species or MSDM
with simple priors to data is faster because the number and species
composition of the guilds does not have to be estimate and it may
be easy to parallelize the computational procedure (\citealt{BB2L},
ch. 19; \citeauthor{hooten2018making} \textit{in press}). In general,
the TSP will become more computationally challenging to implement
as the number of species increases. For example, our implementation
which uses a model-based recursive partitioning approach available
in the R function \texttt{lmtree(...)} has a computational complexity
that increase nonlinearly with the number of species. As a result
fitting a MSDM to data for a small number of species is relatively
fast as demonstrated by our example with six species of aquatic vegetation,
however, model fitting requires much more time for a larger number
of species as demonstrated by our example with 15 species of fish.
The increase in computational time is caused by the increase in the
number of possible guilds. Based on our experience and using the computational
approaches presented in this paper, the TSP becomes infeasible to
fit to data that contains more than roughly twenty species, however,
feasibility will vary by data set and computer resources. Although
this is a current limitation of the TSP, other Bayesian MSDMs that
incorporate the guild concept face similar challenges (e.g., \citealt{johnson2017modeling})
which demonstrates the need for more efficient algorithms for ``big''
ecological communities.
Currently we suggest two approaches to alleviate the computational
burden when applying our approach to a large number of species. The
first suggestion is to choose off-the-shelf tree methods that use
multiple processors (or cores). For example, we used the R function
\texttt{lmtree(...)}which allows users to choose the number of processor
cores for tree related computations. In practice, implementation that
exploits multiple processors should reduce the time required to fit
a MSDM with the TSP to data, however, gains will vary by application.
The second approach to reduce the computational burden involves using
values of the tree hyperparameters that favor a small number of guilds.
In studies with a large number of species, choosing a value of the
tree hyperparameters that results in a small number of guilds may
reduce the predictive accuracy of the model, however, model fitting
will likely be quicker. In addition a small number of guilds may be
desirable for studies that contain a large number of species because
interpretation of the results for a larger number of guilds may be
challenging. Both techniques, using multiple cores and reducing the
number of guilds, were illustrated in our fisheries data example (see
Appendix S2). As with most computationally demanding statistical methods,
advances may soon obviate the current challenges.
Model development, implementation, and selection for data from ecological
communities is difficult because a high level of complexity is desired
which can be achieved by including numerous parameters. For example,
commonly used MSDMs are specified so that each species has a unique
regression coefficient for each predictor variable (e.g., \citealt{wilkinson2019comparison}).
As a result it is common practice to overparameterize MSDMs, which
can degrade predictive accuracy and produce results that are difficult
to interpret and communicate. We illustrated how to reduce the number
of parameters in commonly used Bayesian MSDMs by replacing simple
prior distributions for regression coefficients with a TSP. The TSP
reduces model complexity while incorporating ecological theory, expanding
inference, and can increase the predictive accuracy and interpretability
of Bayesian MSDMs.
\vspace{0.66cm}
\section*{Acknowledgements}
The U.S. Army Corps of Engineers' Upper Mississippi River Restoration
Program LTRM element is implemented by the U.S. Geological Survey,
Upper Midwest Environment Sciences Center, in cooperation with the
five Upper Mississippi River System states of Illinois, Iowa, Minnesota,
Missouri, and Wisconsin. The U.S. Army Corps of Engineers provides
guidance and has overall Program responsibility. The authors acknowledge
support for this research from USGS G17AC00289. Use of trade, product,
or firm names does not imply endorsement by the U.S. Government.\vspace{0.66cm}
\section*{Data accessibility}
The aquatic vegetation and fisheries data are publicly available from
\citet{UMRR}, but the specific subsets used in this study will be
archived in the Dryad Digital Repository. For the purpose of peer
review, the data have been submitted as supporting material as a compressed
file labeled Data.zip.
\renewcommand{\refname}{\vspace{0.1cm}\hspace{-0cm}\selectfont\large References}\setlength{\bibsep}{0pt}
\bibliographystyle{apa}
| 8675818b41e3c324f7d0a1e39d4aa67fd339d89d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{General linear response NEGF formalism for heat transfer} \label{sec:genform}
Consider a generic system exhibiting generalized displacements labeled
$x(t)$, which may represent electronic wavefunctions, collective
nuclear oscillations giving rise to phonons, EM fields, or other
oscillatory phenomena, and respond linearly to generalized forces
labeled $F(t)$. These degrees of freedom (DOFs) constitute collections
which we generically call ``components''. Each component exhibits
linear equations of motion representing its internal dynamics in
isolation and in response to external forces, and each component may
be linearly coupled to other components leading to energy transport
among them. The following sections will make clear the identities of
the components, couplings, generalized displacements, and generalized
forces in different systems of interest, like PCHT or RHT; this
section focuses on deriving relevant fully general formulas for energy
transport among generic coupled components.
Generically, in the time domain, the power radiated or absorbed by a
component may be written as $\dot{W} = \left\langle F(t)
\frac{\partial x}{\partial t} \right\rangle$. Here, $\bracket{\ldots}$
denotes a time average in the steady-state, which is equivalent to an
ensemble average due to ergodicity. As we have specified that the
internal dynamics and couplings are linear, we may equivalently work
in the frequency domain, making it easier to apply the
fluctuation--dissipation theorem~\cite{Novotny2006} and thereby
replace such ensemble averages with deterministic quantities
representing the dissipation of the system.
In the frequency domain (where we will generally suppress dependence
on angular frequency $\omega$ in the notation), we label these
generalized displacements as $\ket{x}$ and the generalized forces as
$\ket{F}$, as these quantities are vectors in a complex Hilbert space
with the standard inner product. One of the operators relevant to this
Hilbert space are the dynamical operator $\hat{Z}^{(0)}$, representing
the dynamical equations of motion for each component in isolation, and
can equivalently be seen as a generalized impedance or spring
constant; its inverse, $\hat{Y}^{(0)} = \hat{Z}^{(0)-1}$, represents a
generalized admittance for the components in isolation, such that an
external force $\ket{F^{(0)}}$ on the components \emph{in isolation}
produces a total displacement $\ket{x} = \hat{Y}^{(0)}
\ket{F^{(0)}}$. However, energy exchange among components is only
possible if couplings are present: as will become clearer later, these
couplings may act directly between components, or may act through
other components whose equations are eliminated, resulting in
effective self-couplings for the remaining components. These couplings
are generically represented by the linear operator $\Delta\hat{Z}$:
the force $\ket{F}$ on other components due to a displacement
$\ket{x}$ from equilibrium of a given component can be written as
$\ket{F} = -\Delta\hat{Z} \ket{x}$.
We generally assume this system to be reciprocal, so that in this
complex Hilbert space, $\hat{Z}^{(0)} = \hat{Z}^{(0)\top}$ and
$\Delta\hat{Z} = \Delta\hat{Z}^{\top}$ hold, where ${}^{\top}$ denotes
the transpose \emph{without} conjugation and not the Hermitian adjoint
${}^{\dagger}$, and reciprocity of related operators follows from
these relations. Additionally, causality implies that
$\hat{Y}^{(0)}(-\omega^{\star}) = \hat{Y}^{(0)\star}(\omega)$ for any
complex frequency $\omega$; passivity implies that $\operatorname{asym}(Y^{(0)})$,
where $\operatorname{asym}(\hat{A}) \equiv (\hat{A} - \hat{A}^{\dagger})/(2\operatorname{i})$ for
any operator $\hat{A}$, which represents the dissipative contribution
to the response of the uncoupled components, is Hermitian
positive-definite (in the space of its own support) for any real
positive frequency $\omega$~\cite{KrugerPRB2012}.
We now turn to the equations of motion for this system in the presence
of coupling between its different components. In particular, the total
generalized displacement can be written as the sum of the initial
displacement $\ket{x^{(0)}}$ and the response of the components
$\hat{Y}^{(0)}$, in isolation, to the total generalized force
$\ket{F}$; in order to avoid double-counting various contributions,
the total generalized force simply arises from the total displacements
through the couplings between different components $\Delta\hat{Z}$, as
we assume that no other external forces contribute to the system
dynamics, and these couplings are the only way for energy to be
transmitted among different components. Mathematically, this is
written as
\begin{equation}
\begin{split}
\ket{x} &= \ket{x^{(0)}} + \hat{Y}^{(0)} \ket{F} \\
\ket{F} &= -\Delta\hat{Z} \ket{x}
\end{split}
\end{equation}
and we formally solve this to yield
\begin{equation} \label{eq:totallinearresponse}
\ket{x} = \hat{Y} \hat{Z}^{(0)} \ket{x^{(0)}}
\end{equation}
where we define the total response $\hat{Y} = \hat{Z}^{-1}$ in terms
of the total equations of motion (generalized impedance) $\hat{Z} =
\hat{Z}^{(0)} + \Delta\hat{Z}$. We stress that the total response
$\hat{Y}$ satisfies the same reciprocity, causality, and passivity
properties as the decoupled response $\hat{Y}^{(0)}$.
At this point, we specify that the system can be partitioned into $N$
components, with each component labeled $n$ specifying a certain set
of DOFs; the operator $\hat{Z}^{(0)}$ (and also $\hat{Y}^{(0)}$ by
extension) can be written as a block-diagonal matrix as it represents
the equations of motion of each component in the absence of coupling
between components, so if $\hat{P}_{n}$ is a projection into the
subspace supported by the DOFs of component $n$, then each diagonal
block of $\hat{Z}^{(0)}$ is $\hat{Z}^{(0)}_{n} = \hat{P}_{n}
\hat{Z}^{(0)} \hat{P}_{n}$, and likewise each diagonal block of
$\hat{Y}^{(0)}$ is $\hat{Y}^{(0)}_{n} = \hat{P}_{n} \hat{Y}^{(0)}
\hat{P}_{n}$, with $\hat{Y}^{(0)}_{n} = \hat{Z}^{(0)-1}_{n}$. (We note
that $\hat{Z}$ and $\hat{Y}$ will generally not be block-diagonal with
respect to the different components.)
If each component $n$ is maintained independently at a corresponding
temperature $T_{n}$ (uniformly for all of the DOFs constituting that
component), then we may write the frequency-domain
fluctuation--dissipation theorem as
\begin{multline} \label{eq:FDT}
\hat{P}_{m} \bracket{\ket{x^{(0)} (\omega)} \bra{x^{(0)} (\omega')}}
\hat{P}_{n} = \\ \frac{2\Pi(\omega, T_{n})}{\omega}
\operatorname{asym}(\hat{Y}^{(0)}_{n} (\omega)) \delta_{mn} \times
2\pi\delta(\omega - \omega')
\end{multline}
where $\bracket{\ldots}$ represents the quantum statistical
expectation value; our use of the Planck function $\Pi(\omega, T) =
\frac{\hbar\omega}{2}\coth\left(\frac{\hbar\omega}{2k_{\mathrm{B}}
T}\right)$ implicitly assumes that all DOFs we consider, when
quantized, obey Bose statistics with no chemical potential, which is
appropriate for EM fields and for coupled mechanical oscillators under
consideration. We also point out that the use of
$\operatorname{asym}(\hat{Y}^{(0)}_{n})$, as opposed to $\operatorname{asym}((\hat{Z}^{(0)}_{n} +
\Delta\hat{Z}_{nn})^{-1})$ if $\hat{P}_{n} \Delta\hat{Z} \hat{P}_{n} =
\Delta\hat{Z}_{nn} \neq 0$, is valid because the former includes
dissipation only within component $n$, whereas the latter may
implicitly include dissipation in other coupled components that have
been eliminated. In order to compute the heat transfer from component
$m$ to component $n$, we account only for fluctuations in component
$m$, so that $\ket{x^{(0)}} = \hat{P}_{m} \ket{x^{(0)}}$ will hold,
and compute the work done on component $n$ according to $\ket{F} =
-\hat{P}_{n} \Delta\hat{Z} \ket{x}$ where $\ket{x} = \hat{Y}
\hat{Z}^{(0)} \hat{P}_{m} \ket{x^{(0)}}$. We then write the absorbed
power (energy transfer) as $\dot{W} = \int_{-\infty}^{\infty}
\int_{-\infty}^{\infty} \bracket{\trace{-\operatorname{i}\omega \hat{P}_{n}
\ket{x(\omega)} \bra{F(\omega')} \hat{P}_{n}} e^{-\operatorname{i}(\omega -
\omega')t}}~\frac{\mathrm{d}\omega~\mathrm{d}\omega'}{(2\pi)^{2}}$.
Algebraic manipulations involving the definitions of $\ket{x}$ and
$\ket{F}$, along with the fluctuation--dissipation theorem
in~\eqref{FDT} applied to $\ket{x^{(0)}}$ and the causality properties
of the relevant response quantities, yield
\begin{multline*}
\dot{W} = \frac{2}{\pi} \int_{0}^{\infty} \Pi(\omega, T_{m})
\times \\ \trace{\hat{P}_{m} \operatorname{asym}(\hat{Z}^{(0)\dagger}) \hat{P}_{m}
\hat{Y}^{\dagger} \operatorname{asym}(\hat{P}_{n} \Delta\hat{Z}) \hat{Y}
\hat{P}_{m}}~\mathrm{d}\omega
\end{multline*}
as the gross energy transfer from component $m$ at temperature $T_{m}$
to component $n$ among a collection of an arbitrary number of
thermalized components. From this, we define the NEGF energy transfer
spectrum between components $m$ and $n$ as
\begin{equation} \label{eq:Phimn}
\Phi^{(m)}_{n} = 4~\trace{\hat{P}_{m} \operatorname{asym}(\hat{Z}^{(0)\dagger})
\hat{P}_{m} \hat{Y}^{\dagger} \operatorname{asym}(\hat{P}_{n} \Delta\hat{Z})
\hat{Y} \hat{P}_{m}}
\end{equation}
independently of the temperature of each component, while the
integrated net power transfer can be written as
\begin{equation}
\dot{W}_{m \to n} = \int_{0}^{\infty} (\Pi(\omega, T_{m}) -
\Pi(\omega, T_{n}))\Phi^{(m)}_{n}~\frac{\mathrm{d}\omega}{2\pi}
\end{equation}
in terms of the component temperatures $T_{m}$ and
$T_{n}$. Furthermore, the heat transfer coefficient (thermal
conductance) between two components may be derived by replacing
$\Pi(\omega, T_{m}) - \Pi(\omega, T_{n})$ in the integrand with
$\lim_{T_{m} \to T_{n}} \frac{\Pi(\omega, T_{m}) - \Pi(\omega,
T_{n})}{T_{m} - T_{n}} = \frac{\partial}{\partial T} \Pi(\omega,
T)$. It is worth noting that reciprocity, which has not been exploited
in these derivations thus far, is required to show that
$\Phi^{(m)}_{n} = \Phi^{(n)}_{m}$ at each frequency.
The formula for $\Phi^{(m)}_{n}$ in~\eqref{Phimn} is valid for any
number of components maintained at their own uniform temperatures, and
constitutes a generalization of Landauer/Caroli formulas often used to
describe PCHT and RHT~\cite{OteyJQSRT2014, KrugerPRB2012,
VenkataramPRL2018, JinPRB2019, SegalARPC2016, ZhangPRB2018,
WangPRE2018, KlocknerPRB2018, SadasivamPRB2017, PaulyNJP2008}. The
most fruitful analogies between RHT and PCHT can be extracted from
consideration of heat transfer between two components, through direct
contact or via contact with a third component. In what follows, we
derive formulas for both situations: the first situation is most
relevant to RHT between two bodies or PCHT combined with RHT between
two bodies in direct contact, while the second situation is most
relevant to PCHT combined with RHT between two bodies via a third
intermediate body (typically a thin interface or a small atomic or
molecular junction), though it can also be applied to formally
deriving expressions for RHT between two bodies.
\subsection{Two components in direct contact}
For two components, labeled 1 or 2, in direct contact with each other,
we may write the operators $\hat{Z}^{(0)}$ and $\Delta \hat{Z}$
describing the equations of motion and couplings among these
components may be written as $2\times 2$ block matrices
\begin{align}
\hat{Z}^{(0)} &=
\begin{bmatrix}
\hat{Z}^{(0)}_{1} & 0 \\
0 & \hat{Z}^{(0)}_{2}
\end{bmatrix}
\\
\Delta \hat{Z} &=
\begin{bmatrix}
\Delta \hat{Z}_{1,1} & \Delta \hat{Z}_{1,2} \\
\Delta \hat{Z}_{2,1} & \Delta \hat{Z}_{2,2}
\end{bmatrix}
\end{align}
which in turn implies that
\begin{equation}
\hat{Y}^{(0)} =
\begin{bmatrix}
\hat{Y}^{(0)}_{1} & 0 \\
0 & \hat{Y}^{(0)}_{2}
\end{bmatrix}
\end{equation}
must also hold; the existence of nontrivial diagonal and off-diagonal
blocks in $\Delta \hat{Z}$ typically arises from couplings to other
components that are mathematically eliminated in favor of these two
components. Evaluation of the energy transfer spectrum~\eqref{Phimn}
requires inversion of these block matrices of operators, which is
saved for the appendix for the sake of brevity in this section. The
result is written as
\begin{multline}
\Phi = 4~\operatorname{Tr}\Bigg[\operatorname{asym}(\hat{Z}^{(0)\dagger}_{1})
(\hat{Z}^{(0)\dagger}_{1} + \Delta \hat{Z}^{\dagger}_{1,1})^{-1}
\times \\ (\hat{1} - \Delta \hat{Z}^{\dagger}_{2,1}
(\hat{Z}^{(0)\dagger}_{2} + \Delta \hat{Z}^{\dagger}_{2,2})^{-1}
\Delta \hat{Z}^{\dagger}_{1,2} (\hat{Z}^{(0)\dagger}_{1} + \Delta
\hat{Z}^{\dagger}_{1,1})^{-1})^{-1} \times \\ \Delta
\hat{Z}^{\dagger}_{2,1} (\hat{Z}^{(0)\dagger}_{2} + \Delta
\hat{Z}^{\dagger}_{2,2})^{-1} \times
\\ \operatorname{asym}(\hat{Z}^{(0)\dagger}_{2}) (\hat{Z}^{(0)}_{2} + \Delta
\hat{Z}_{2,2})^{-1} \Delta \hat{Z}_{2,1} \times \\ (\hat{1} -
(\hat{Z}^{(0)}_{1} + \Delta \hat{Z}_{1,1})^{-1} \Delta
\hat{Z}_{1,2} (\hat{Z}^{(0)}_{2} + \Delta \hat{Z}_{2,2})^{-1}
\Delta \hat{Z}_{2,1})^{-1} \times \\ (\hat{Z}^{(0)}_{1} + \Delta
\hat{Z}_{1,1})^{-1}\Bigg]
\end{multline}
and as $\operatorname{asym}(\hat{Z}^{(0)\dagger}_{1})$ and
$\operatorname{asym}(\hat{Z}^{(0)\dagger}_{2})$ are Hermitian positive-semidefinite
operators with well-defined Hermitian square roots, then the energy
transfer spectrum is nonnegative. Exploiting this further allows for
factorizing $\operatorname{asym}(\hat{Z}^{(0)\dagger}_{n}) =
(\operatorname{asym}(\hat{Z}^{(0)\dagger}_{n})^{1/2})^{2}$ for each component $n \in
\{1, 2\}$, and rearranging the trace allows for writing
\begin{multline} \label{eq:Phi2bodydirect}
\Phi(\omega) = 4~\Bigg\Vert
\operatorname{asym}(\hat{Z}^{(0)\dagger}_{2})^{1/2} (\hat{Z}^{(0)}_{2} + \Delta
\hat{Z}_{2,2})^{-1} \Delta \hat{Z}_{2,1} \times \\ (\hat{1} -
(\hat{Z}^{(0)}_{1} + \Delta \hat{Z}_{1,1})^{-1} \Delta \hat{Z}_{1,2}
(\hat{Z}^{(0)}_{2} + \Delta \hat{Z}_{2,2})^{-1} \Delta
\hat{Z}_{2,1})^{-1} \times \\ (\hat{Z}^{(0)}_{1} + \Delta
\hat{Z}_{1,1})^{-1} \operatorname{asym}(\hat{Z}^{(0)\dagger}_{1})^{1/2}
\Bigg\Vert_{\mathrm{F}}^{2}
\end{multline}
where $\left\Vert \hat{A} \right\Vert_{\mathrm{F}} =
\sqrt{\trace{\hat{A}^{\dagger} \hat{A}}}$ is the Frobenius norm. This
is the general NEGF formula for the energy transfer spectrum between
two components in direct contact, in terms of their individual and
mutual responses.
\subsection{Two components coupled via an intermediate component}
We now consider the case of the two components, labeled 1 or 2, which
are coupled only to a third intermediate component, labeled 3, but not
directly to each other. Mathematically, this means the operators
$\hat{Z}^{(0)}$ and $\Delta \hat{Z}$ describing the equations of
motion and couplings among these components may be written as $3\times
3$ block matrices
\begin{align}
\hat{Z}^{(0)} &=
\begin{bmatrix}
\hat{Z}^{(0)}_{1} & 0 & 0 \\
0 & \hat{Z}^{(0)}_{2} & 0 \\
0 & 0 & \hat{Z}^{(0)}_{3}
\end{bmatrix}
\\
\Delta \hat{Z} &=
\begin{bmatrix}
0 & 0 & \Delta \hat{Z}_{1,3} \\
0 & 0 & \Delta \hat{Z}_{2,3} \\
\Delta \hat{Z}_{3,1} & \Delta \hat{Z}_{3,2} & 0
\end{bmatrix}
\end{align}
which in turn implies that
\begin{equation}
\hat{Y}^{(0)} =
\begin{bmatrix}
\hat{Y}^{(0)}_{1} & 0 & 0 \\
0 & \hat{Y}^{(0)}_{2} & 0 \\
0 & 0 & \hat{Y}^{(0)}_{3}
\end{bmatrix}
\end{equation}
must also hold, where the vanishing of the components
$\Delta\hat{Z}_{1,2} = (\Delta\hat{Z}_{2,1})^{\top}$ follows from the
assumption that components 1 \& 2 have no direct coupling to each
other; we also point out that compared to the general two-component
formula, here we do not include couplings between a given component
and itself (i.e. $\Delta \hat{Z}$ has vanishing diagonal blocks), as
we assume that there are no other components which we have implicitly
eliminated. Once again leaving the details to the appendix, and again
making use of the fact that $\operatorname{asym}(\hat{Y}^{(0)}_{1})$ and
$\operatorname{asym}(\hat{Y}^{(0)}_{2})$ are Hermitian positive-semidefinite
operators to factorize the trace expression, we write~\eqref{Phimn} as
\begin{multline} \label{eq:Phi2bodyviathird}
\Phi(\omega) = 4~\Bigg\Vert \operatorname{asym}(\hat{Y}^{(0)}_{2})^{1/2} \Delta
\hat{Z}_{2,3} \times \\ (\hat{Z}^{(0)}_{3} - \Delta \hat{Z}_{3,1}
\hat{Y}^{(0)}_{1} \Delta \hat{Z}_{1,3} - \Delta \hat{Z}_{3,2}
\hat{Y}^{(0)}_{2} \Delta \hat{Z}_{2,3})^{-1} \times \\ \Delta
\hat{Z}_{3,1} \operatorname{asym}(\hat{Y}^{(0)}_{1})^{1/2}
\Bigg\Vert_{\mathrm{F}}^{2}
\end{multline}
so that all operator products may be evaluated in the space of
component 3. This is the general NEGF formula for the energy transfer
spectrum between two components in contact only with a third in terms
of their individual responses and mutual couplings.
In the previous subsection, it was noted that for heat transfer
between two components that are directly coupled, the couplings
$\Delta \hat{Z}_{mn}$ for $m, n \in \{1, 2\}$ (particularly the
diagonal blocks) often arise from mathematically eliminating another
component to which these two components are coupled, even if those are
the only couplings. At this point, we rigorously prove this
equivalence for the specific case where the two components are
physically coupled only to a third component. We start by
rewriting~\eqref{totallinearresponse} in terms of the degrees of
freedom of the three components and noting that $\ket{x^{(0)}} =
\hat{P}_{1} \ket{x^{(0)}}$ can be used when considering energy
transfer from component 1 to component 2. Explicitly, this means
writing
\begin{equation*}
\begin{bmatrix}
\hat{Z}^{(0)}_{1} & 0 & \Delta \hat{Z}_{1,3} \\
0 & \hat{Z}^{(0)}_{2} & \Delta \hat{Z}_{2,3} \\
\Delta \hat{Z}_{3,1} & \Delta \hat{Z}_{3,2} & \hat{Z}^{(0)}_{3}
\end{bmatrix}
\begin{bmatrix}
\ket{x_{1}} \\
\ket{x_{2}} \\
\ket{x_{3}}
\end{bmatrix}
=
\begin{bmatrix}
\hat{Z}^{(0)}_{1} \ket{x^{(0)}_{1}} \\
0 \\
0
\end{bmatrix}
\end{equation*}
and then eliminating $\ket{x_{3}} = -\hat{Y}^{(0)}_{3} (\Delta
\hat{Z}_{3,1} \ket{x_{1}} + \Delta \hat{Z}_{3,2} \ket{x_{2}})$. This
yields the simpler equation in terms of $2\times 2$ block matrices
\begin{multline*}
\begin{bmatrix}
\hat{Z}^{(0)}_{1} - \Delta \hat{Z}_{1,3} \hat{Y}^{(0)}_{3} \Delta
\hat{Z}_{3,1} & -\Delta \hat{Z}_{1,3} \hat{Y}^{(0)}_{3} \Delta
\hat{Z}_{3,2} \\
-\Delta \hat{Z}_{2,3} \hat{Y}^{(0)}_{3} \Delta \hat{Z}_{3,1} &
\hat{Z}^{(0)}_{2} - \Delta \hat{Z}_{2,3} \hat{Y}^{(0)}_{3} \Delta
\hat{Z}_{3,2}
\end{bmatrix}
\begin{bmatrix}
\ket{x_{1}} \\
\ket{x_{2}}
\end{bmatrix}
\\ =
\begin{bmatrix}
\hat{Z}^{(0)}_{1} \ket{x^{(0)}_{1}} \\
0
\end{bmatrix}
\end{multline*}
whence the replacements
\begin{align*}
\hat{Z}^{(0)} &\to \begin{bmatrix}
\hat{Z}^{(0)}_{1} & 0 \\
0 & \hat{Z}^{(0)}_{2}
\end{bmatrix} \\
\Delta \hat{Z} &\to \begin{bmatrix}
-\Delta \hat{Z}_{1,3} \hat{Y}^{(0)}_{3} \Delta \hat{Z}_{3,1} &
-\Delta \hat{Z}_{1,3} \hat{Y}^{(0)}_{3} \Delta \hat{Z}_{3,2} \\
-\Delta \hat{Z}_{2,3} \hat{Y}^{(0)}_{3} \Delta \hat{Z}_{3,1} &
-\Delta \hat{Z}_{2,3} \hat{Y}^{(0)}_{3} \Delta \hat{Z}_{3,2}
\end{bmatrix}
\end{align*}
may be made. Hence, the remainder of the derivation of the expression
for heat transfer is the same, as~\eqref{FDT} for component 1 and the
expression $\dot{W} = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty}
\bracket{\trace{-\operatorname{i}\omega \hat{P}_{2} \ket{x(\omega)}
\bra{F(\omega')} \hat{P}_{2}} e^{-\operatorname{i}(\omega -
\omega')t}}~\frac{\mathrm{d}\omega~\mathrm{d}\omega'}{(2\pi)^{2}}$
for the power transfer are both unchanged, thereby proving the
equivalence between the two expressions (\Eqref{Phi2bodydirect} and
\Eqref{Phi2bodyviathird}) for the general NEGF energy transfer
spectrum with these identifications in mind.
Writing the energy transfer spectrum as~\eqref{Phi2bodyviathird} can
not only clarify analogies between PCHT and RHT, but it also naturally
leads to expressions for upper bounds on the spectrum. To derive such
bounds, it will be helpful to define the operators $\hat{\Lambda}_{n}
= \Delta \hat{Z}_{n,3}^{\dagger} \operatorname{asym}(\hat{Y}^{(0)}_{n}) \Delta
\hat{Z}_{n,3}$, being the dissipation of each component $n \in \{1,
2\}$ multiplied by the corresponding couplings to component 3, and the
Green's function of component 3
\begin{equation}
\hat{Y}_{3} \equiv (\hat{Z}^{(0)}_{3} - \Delta \hat{Z}_{3,1}
\hat{Y}^{(0)}_{1} \Delta \hat{Z}_{1,3} - \Delta \hat{Z}_{3,2}
\hat{Y}^{(0)}_{2} \Delta \hat{Z}_{2,3})^{-1}
\end{equation}
which is modified from its bare value $\hat{Y}^{(0)}_{3}$ due to
couplings to components 1 \& 2. Given this, we will show that the
energy transfer spectrum can be written in the Landauer
form~\cite{KlocknerPRB2018, KlocknerPRB2017, SongAIPADV2015,
SegalARPC2016} as $\Phi = \left\lVert \hat{t}
\right\rVert_{\mathrm{F}}^{2}$ where $\hat{t} =
2\hat{\Lambda}_{2}^{1/2} \hat{Y}_{3} \hat{\Lambda}_{1}^{1/2}$. The
goal then will be to place bounds on the eigenvalues of
$\hat{t}^{\dagger} \hat{t}$ at each $\omega$. The fact that
$\hat{t}^{\dagger} \hat{t}$ is Hermitian positive-semidefinite makes
clear that its eigenvalues, called the transmission eigenvalues (as
$\hat{t}^{\dagger} \hat{t}$ is like a transmission intensity matrix),
are all nonnegative, placing a lower bound on their values. The
following will show how to derive upper bounds of unity on the
transmission eigenvalues.
The derivations thus far have actually not made use of the reciprocity
of the system, namely that $\hat{Z}^{(0)}_{3} =
\hat{Z}^{(0)\top}_{3}$, $\hat{Z}^{(0)}_{n} = \hat{Z}^{(0)\top}_{n}$,
and $\Delta \hat{Z}_{n,3} = \Delta \hat{Z}_{3,n}^{\top}$ for $n \in
\{1, 2\}$, but these reciprocity relations are needed for the upper
bounds on the transmission eigenvalues. Additionally, two further
assumptions are needed, namely that $\operatorname{asym}(\hat{Z}^{(0)}_{3}) \to 0$,
and that the block matrices $\Delta \hat{Z}_{n,3}$ for $n \in \{1,
2\}$ are purely real. These assumptions will later be justified for
the particular cases of PCHT as well as RHT.
With this, it can be seen that $\operatorname{asym}(\hat{Y}_{3}) =
\hat{Y}_{3}^{\dagger} \operatorname{asym}(\hat{Y}_{3}^{-1\dagger})
\hat{Y}_{3}$. Expanding the middle term after exploiting
$\operatorname{asym}(\hat{Z}^{(0)}_{3}) = 0$ gives $\operatorname{asym}(\hat{Y}_{3}^{-1}) =
-\hat{\Lambda}_{1} - \hat{\Lambda}_{2}$, as the real-valued and
reciprocal nature of $\Delta \hat{Z}_{n,3}$ imply $\Delta
\hat{Z}_{3,n} \operatorname{asym}(\hat{Y}^{(0)}_{n}) \Delta \hat{Z}_{n,3} = \Delta
\hat{Z}_{n,3}^{\dagger} \operatorname{asym}(\hat{Y}^{(0)}_{n}) \Delta \hat{Z}_{n,3}$
for $n \in \{1, 2\}$. This means $\operatorname{asym}(\hat{Y}_{3}^{-1\dagger}) =
-\operatorname{asym}(\hat{Y}_{3}^{-1}) = \hat{\Lambda}_{1} +
\hat{\Lambda}_{2}$. Therefore, $\operatorname{asym}(\hat{Y}_{3}) =
\hat{Y}_{3}^{\dagger} (\hat{\Lambda}_{1} + \hat{\Lambda}_{2})
\hat{Y}_{3}$. This expression may be rearranged as
$\hat{Y}_{3}^{\dagger} \hat{\Lambda}_{1} \hat{Y}_{3} +
\hat{Y}_{3}^{\dagger} \hat{\Lambda}_{2} \hat{Y}_{3} -
\operatorname{asym}(\hat{Y}_{3}) = 0$, and as $\hat{\Lambda}_{1}$ is Hermitian
positive-semidefinite, then $\hat{\Lambda}_{1}^{1/2}$ exists, so this
expression may be multiplied on the left and right by
$2\hat{\Lambda}_{1}^{1/2}$ to yield $\hat{t}^{\dagger} \hat{t} +
4\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}^{\dagger} \hat{\Lambda}_{1}
\hat{Y}_{3} \hat{\Lambda}_{1}^{1/2} - \frac{2}{\operatorname{i}}
\left(\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3} \hat{\Lambda}_{1}^{1/2} -
\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}^{\dagger} \hat{\Lambda}_{1}^{1/2}
\right) = 0$, where as a reminder, $\hat{t} = 2\hat{\Lambda}_{2}^{1/2}
\hat{Y}_{3} \hat{\Lambda}_{1}^{1/2}$. Finally, adding the identity
operator $\hat{1}$ to both sides yields $\hat{t}^{\dagger} \hat{t} +
4\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}^{\dagger} \hat{\Lambda}_{1}
\hat{Y}_{3} \hat{\Lambda}_{1}^{1/2} - \frac{2}{\operatorname{i}}
\left(\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3} \hat{\Lambda}_{1}^{1/2} -
\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}^{\dagger} \hat{\Lambda}_{1}^{1/2}
\right) + \hat{1} = \hat{1}$. This expression can be rewritten as
$\hat{t}^{\dagger} \hat{t} + (\hat{1} - \frac{2}{\operatorname{i}}
\hat{\Lambda}_{1}^{1/2} \hat{Y}_{3} \hat{\Lambda}_{1}^{1/2})^{\dagger}
(\hat{1} - \frac{2}{\operatorname{i}} \hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}
\hat{\Lambda}_{1}^{1/2}) = \hat{1}$, showing that $4\hat{t}^{\dagger}
\hat{t}$ is added another Hermitian positive-semidefinite operator to
yield the identity. Therefore, the eigenvalues of $\hat{t}^{\dagger}
\hat{t}$ can never exceed 1, matching the prior
expressions~\cite{Datta1995, PaulyNJP2008, KlocknerPRB2018,
SadasivamPRB2017}. Additionally, because the operator $\hat{1} -
\frac{2}{\operatorname{i}} \hat{\Lambda}_{1}^{1/2} \hat{Y}_{3}
\hat{\Lambda}_{1}^{1/2}$ is not the zero operator, its rank must be at
least 1, meaning at least one of its eigenvalues must be strictly
positive; in turn, at least one of the eigenvalues of
$\hat{t}^{\dagger} \hat{t}$ must be strictly less than $1$. We stress
that whenever heat transfer between two components that are directly
coupled can be physically equated to heat transfer between the same
two components with effective couplings only via a third (possibly
aggregate) component, these transmission eigenvalue bounds must hold
for that system. Additionally, we expect that even if $\Delta \hat{Z}$
were to have nonzero blocks other than $\Delta \hat{Z}_{n,3}$ (and
their transposes) for components $n \in \{1, 2\}$, which could
represent more general heat transfer between a pair of components
among a collection of $N$ components (for any integer $N \geq 3$) by
virtue of aggregating the other components into an overall third
intermediate component, similar bounds should hold in general, though
we do not prove that statement; put simply, Landauer bounds of unity
should hold for each channel even between two components connected via
a third where each of these components could in principle be connected
to many other components in turn.
\section{Applications to PCHT} \label{sec:PCHT}
The general NEGF formalism and expression for the energy transfer
spectrum~\eqref{Phimn} applies to PCHT among a collection of material
bodies, modeled as effective harmonic oscillators connected to each
other via harmonic short- or long-range couplings, each maintained at
separate uniform temperature
. Prior works have typically
focused on PCHT between two large bodies, typically leads acting as
thermal reservoirs, exchanging heat via harmonic coupling through a
third small body in between, typically a molecular junction or a thin
interfacial film; computationally, this has the benefit of allowing
most matrix evaluations to occur in the much smaller space of the
intermediate body as opposed to the larger space of one of the
leads. Given this, in what follows, we derive the equations of motion
for collective atomic oscillations effecting phonons from the
Lagrangian for three bodies, each comprising collections of coupled
oscillators with masses $m_{\alpha a}$, displacements $x_{\alpha ai}$,
and spring couplings $K_{\alpha ai, \beta bj}$ for body labels
$\alpha, \beta \in \{1, 2, 3\}$, atomic labels $a, b, c$ within each
body, and Cartesian indices $i, j, k \in \{x, y, z\}$, with sources
only in body $\alpha = 1$ denoted $x^{(0)}_{1ai}$. Note for comparison
with previous work that bodies 1 and 2, representing infinite
reservoirs (leads), are typically labeled L and R, while body 3,
representing an compact intermediate (central) device, is typically
labeled C. We emphasize that while our derivations focus on the
particular case of two bodies connected to a third in order to make
connections to past work clearer, the correspondence between abstract
linear operators and specific quantities of interest to PCHT is easily
generalized to PCHT among a collection of coupled bodies.
The Lagrangian for this system is written as
\begin{multline}
2L = \sum_{a,i} m_{1a} (\dot{x}_{1ai} - \dot{x}^{(0)}_{1ai})^{2} -
\\ \sum_{a,i,a',i'} K_{1ai,1a'i'} (x_{1ai} -
x^{(0)}_{1ai})(x_{1a'i'} - x^{(0)}_{1a'i'}) - \\ \sum_{a,i,c,k}
(K_{1ai,3ck} + K_{3ck,1ai})x_{1ai}x_{3ck} + \\ \sum_{b,j} m_{2b}
\dot{x}_{2bj}^{2} - \sum_{b,j,b',j'} K_{2bj,2b'j'} x_{2bj} x_{2b'j'}
- \\ \sum_{b,j,c,k} (K_{2bj,3ck} + K_{3ck,2bj})x_{2bj} x_{3ck} +
\\ \sum_{c,k} m_{3c} \dot{x}_{3ck}^{2} - \sum_{c,k,c',k'}
K_{3ck,3c'k'} x_{3ck} x_{3c'k'}
\end{multline}
and minimization of the action $S = \int_{-\infty}^{\infty}
L~\mathrm{d}t$ leads to the time domain classical equations of motion
\begin{multline}
m_{1a} \ddot{x}_{1ai} + \sum_{a',i'} K_{1ai,1a'i'} x_{1a'i'} +
\sum_{c,k} K_{1ai,3ck} x_{3ck} = \\ m_{1a} \ddot{x}^{(0)}_{1ai} +
\sum_{a',i'} K_{1ai,1a'i'} x^{(0)}_{1a'i'} \\ m_{2b} \ddot{x}_{2bj}
+ \sum_{b',j'} K_{2bj,2b'j'} x_{2b'j'} + \sum_{c,k} K_{2bj,3ck}
x_{3ck} = 0 \\ m_{3c} \ddot{x}_{3ck} + \sum_{c',k'} K_{3ck,3c'k'}
x_{3c'k'} + \\ \sum_{a,i} K_{3ck,1ai} x_{1ai} + \sum_{b,j} K_{3ck,2bj}
x_{2bj} =0
\end{multline}
for these displacements. In the frequency domain, these become
\begin{multline}
-\omega^{2} m_{1a} x_{1ai} + \sum_{a',i'} K_{1ai,1a'i'} x_{1a'i'} +
\sum_{c,k} K_{1ai,3ck} x_{3ck} = \\ -\omega^{2} m_{1a} x^{(0)}_{1ai} +
\sum_{a',i'} K_{1ai,1a'i'} x^{(0)}_{1a'i'} \\ -\omega^{2} m_{2b}
x_{2bj} + \sum_{b',j'} K_{2bj,2b'j'} x_{2b'j'} + \sum_{c,k}
K_{2bj,3ck} x_{3ck} = 0 \\ -\omega^{2} m_{3c} x_{3ck} + \sum_{c',k'}
K_{3ck,3c'k'} x_{3c'k'} + \\ \sum_{a,i} K_{3ck,1ai} x_{1ai} +
\sum_{b,j} K_{3ck,2bj} x_{2bj} = 0
\end{multline}
and these equations can be collected into matrix form with vectors
$x_{\alpha}$ and matrices $K_{\alpha\beta}$ and $M_{\alpha}$, upon
which the identifications $\hat{Z}^{(0)}_{\alpha} \to K_{\alpha\alpha}
- \omega^{2} M_{\alpha}$ and $\Delta \hat{Z}_{\alpha\beta} \to (1 -
\delta_{\alpha\beta})K_{\alpha\beta}$ can be made, where
$K_{\alpha\beta} = K_{\beta\alpha}^{\top}$ are real-valued, and
$M_{\alpha}$ are real-valued too; we note that the as the matrices $K$
encode spring constants which multiply \emph{differences} in atomic
positions (i.e. relative displacements) to yield forces, the diagonal
blocks $K_{\alpha\alpha}$ entering $\hat{Z}^{(0)}_{\alpha}$ should
actually include the effects of couplings to other bodies as are
present in the off-diagonal blocks $K_{\alpha\beta}$ for all $\beta
\neq \alpha$, so that all forces are balanced in the equations of
motion. With these replacements, the energy transfer spectrum becomes
\begin{multline}
\Phi = \\ 4~\operatorname{Tr}\Bigg[K_{3,1} \operatorname{asym}((K_{1,1} -
\omega^{2} M_{1})^{-1}) K_{1,3} (K_{3,3} - \omega^{2} M_{3} -
\\ K_{3,1} (K_{1,1} - \omega^{2} M_{1})^{-1} K_{1,3} - K_{3,2}
(K_{2,2} - \omega^{2} M_{2})^{-1} K_{2,3})^{-1\dagger} \times
\\ K_{3,2} \operatorname{asym}((K_{2,2} - \omega^{2} M_{2})^{-1}) K_{2,3}
(K_{3,3} - \omega^{2} M_{3} - \\ K_{3,1} (K_{1,1} - \omega^{2}
M_{1})^{-1} K_{1,3} - K_{3,2} (K_{2,2} - \omega^{2} M_{2})^{-1}
K_{2,3})^{-1}\Bigg]
\end{multline}
where the identifications
\begin{equation}
\hat{Y}^{(0)}_{\alpha} \to g_{\alpha} = (K_{\alpha\alpha} -
\omega^{2} M_{\alpha})^{-1}
\end{equation}
as the retarded Green's function of lead $\alpha \in
\{1, 2\}$ (with $g_{\alpha}^{\dagger}$ being the advanced Green's
function),
\begin{equation}
\hat{Y}_{3} \to G = (K_{3,3} - \omega^{2} M_{3} - K_{3,1} g_{1}
K_{1,3} - K_{3,2} g_{2} K_{2,3})^{-1}
\end{equation}
as the retarded Green's function of the device including connections
to the leads (with $G^{\dagger}$ being the advanced Green's function),
and
\begin{equation}
\Lambda_{\alpha} = K_{3,\alpha} \operatorname{asym}(g_{\alpha}) K_{\alpha,3}
\end{equation}
for $\alpha \in \{1, 2\}$ being the dissipation terms at the interface
of the device with each lead can immediately be made. Thus, this
general formalism does reproduce the standard Landauer
formula~\cite{MingoPRB2003, KlocknerPRB2018, KlocknerPRB2017,
SegalARPC2016}
\begin{equation}
\Phi(\omega) = 4~\trace{\Lambda_{1} G^{\dagger}
\Lambda_{2} G}
\end{equation}
for phonon heat transport between two leads across a device. Note that
while $K_{\alpha\alpha}$ and $M_{\alpha}$ are real-valued,
$g_{\alpha}$ is complex-valued because inversion of an
infinite-dimensional matrix is made finite-dimensional by considering
propagation of phonons far from the device interface to be equivalent
to energy loss (so $G$ is also complex-valued in turn); alternatively,
if the leads are large but finite, dissipation may be added
heuristically by replacing, including in the definitions of
$g_{\alpha}$, every instance of $-\omega^{2} M_{\alpha}$ with
$-\operatorname{i}\omega B_{\alpha} - \omega^{2} M_{\alpha}$ for $\alpha \in \{1,
2\}$ where the diagonal positive-definite matrices $B_{\alpha}$
represent appropriate dissipation coefficients for the
oscillators. Additionally, the assumptions underlying the derivation
of the upper bound on the transmission eigenvalues hold here, so those
derivations remain valid in this context: all of the $K$ and $M$
matrices are real-valued and reciprocal, and $\operatorname{asym}(K_{3,3} -
\omega^{2} M_{3}) = 0$ because the compact device will not have any
channels for dissipation in the absence of coupling to reservoirs
(leads). Thus, the general NEGF formalism for heat transfer in linear
response systems can be exactly mapped to the specific NEGF formalism
for linear PCHT.
Physically, the harmonic oscillators represent nuclei dressed by
inner-shell electrons, and the couplings represent chemical bonds
between these oscillators, typically computed via density functional
theory and often anisotropic. We again stress that the correspondences
$\hat{Z}^{(0)}_{\alpha} \to K_{\alpha\alpha} - \omega^{2} M_{\alpha}$
and $\Delta \hat{Z}_{\alpha\beta} \to (1 -
\delta_{\alpha\beta})K_{\alpha\beta}$ for PCHT are generally
applicable even beyond the specific case of two bodies coupled only to
a third intermediate body, which allows more general scenarios for
PCHT to be treated using~\eqref{Phimn}; moreover, these derivations do
not assume that the material bodies exhibit any particular geometry or
spatial symmetry properties.
\section{Applications to RHT} \label{sec:RHT}
\begin{figure}[h!]
\centering
\includegraphics[width=0.85\columnwidth]{photonradiationandlikephononconductionfigure.eps}
\caption{\textbf{Photon radiation.} (a) Radiation of energy in free
space between compact polarizable bodies. (b) Analogous situation
for conduction: compact phononic devices are coupled at each atom
to an infinite lattice supporting propagation of phonons
infinitely far away.}
\label{fig:photonradiationschematics}
\end{figure}
The general NEGF formalism and expression for the energy transfer
spectrum~\eqref{Phimn} also applies to RHT among a collection of
linearly polarizable bodies that can radiate EM fields. Prior works
have typically focused on RHT between two polarizable bodies, whether
spatially compact or of infinite extent, in vacuum. The connection to
the above general linear response formalism for heat transfer requires
somewhat more of a conceptual leap compared to the connection for
PCHT. In particular, components 1 \& 2 are the polarizable material
bodies in question, while component 3, rather than representing a
material body, is actually the \emph{vacuum EM field} pervading all of
space. A Lagrangian for this system can easily be written for the case
where the polarizable bodies are made of atomic harmonic oscillators,
with equilibrium positions $\vec{r}_{\alpha a}$ for body $\alpha \in
\{1, 2\}$ and atom label $a$, and with charges $q_{\alpha a}$ that
couple to EM fields; the sources are taken to be in body 1. That said,
the results are generalizable to other linear media whose response
functions are more complicated than those of harmonic oscillators, and
to cases with more than 2 material bodies present; in particular, the
use of partial bound charges associated with harmonic oscillators more
accurately describes polar dielectric media compared to metals, but
the results are generalizable to metals, semimetals, and other media
with susceptibilities that may be nonlocal, inhomogeneous, or
anisotropic.
The Lagrangian for this system is written as
\begin{multline}
2L = \sum_{a,i} m_{1a} (\dot{x}_{1ai} - \dot{x}^{(0)}_{1ai})^{2} -
\\ \sum_{a,i,a',i'} K_{1ai,1a'i'} (x_{1ai} -
x^{(0)}_{1ai})(x_{1a'i'} - x^{(0)}_{1a'i'}) + \\ 2\sum_{a,i} q_{1a}
\dot{x}_{1ai} A_{i} (\vec{r}_{1a}) + \sum_{b,j} m_{2b}
\dot{x}_{2bj}^{2} - \\ \sum_{b,j,b',j'} K_{2bj,2b'j'} x_{2bj}
x_{2b'j'} + 2\sum_{b,j} q_{2b} \dot{x}_{2bj} A_{j} (\vec{r}_{2b})
+ \\ \int \left(\frac{1}{c^{2}} \left(\frac{\partial
\vec{A}}{\partial t}\right)^{2} - (\nabla \times
\vec{A})^{2}\right)~\mathrm{d}^{3} x
\end{multline}
introducing the magnetic potential $\vec{A}$, working in the Weyl
gauge (vanishing electric potential). Minimizing the action $S =
\int_{0}^{\infty} L~\mathrm{d}t$ leads to the time domain classical
equations of motion
\begin{multline}
m_{1a} \ddot{x}_{1ai} + \sum_{a',i'} K_{1ai,1a'i'} x_{1a'i'} -
q_{1a} E_{i} (\vec{r}_{1a}) = \\ m_{1a} \ddot{x}^{(0)}_{1ai} +
\sum_{a',i'} K_{1ai,1a'i'} x^{(0)}_{1a'i'} \\ m_{2b} \ddot{x}_{2bj}
+ \sum_{b',j'} K_{2bj,2b'j'} x_{2b'j'} - q_{2b} E_{j} (\vec{r}_{2b})
= 0 \\ \left(\nabla \times (\nabla \times) + \frac{1}{c^{2}}
\frac{\partial^{2}}{\partial t^{2}} \right) \vec{E} = \\
-\frac{1}{c^{2}} \left(\sum_{a} q_{1a} \ddot{\vec{x}}_{1a}
\delta^{3} (\vec{x} - \vec{r}_{1a}) + \sum_{b} q_{2b}
\ddot{\vec{x}}_{2b} \delta^{3} (\vec{x} - \vec{r}_{2b}) \right)
\end{multline}
for the displacements $x_{1ai}$ and $x_{2bj}$ and electric field
$\vec{E}(t, \vec{x}) = -\frac{1}{c} \frac{\partial \vec{A}}{\partial
t}$, where the magnetic contribution to the Lorentz force
$\frac{q_{\alpha a}}{c} \dot{\vec{x}}_{\alpha a} \times \vec{B}$ is
dropped for each atom as it is a nonlinear term that has negligible
contribution for speeds much less than the speed of light $c$ (which
is generally true for thermal fluctuations at reasonable
temperatures). Although the third equation should initially be written
in terms of $\vec{A}$, a partial time derivative is applied to both
sides of the equation to simplify the equations in terms of
$\vec{E}$. In the frequency domain, these equations of motion become
\begin{multline}
-\omega^{2} m_{1a} x_{1ai} + \sum_{a',i'} K_{1ai,1a'i'} x_{1a'i'}
- q_{1a} E_{i} (\vec{r}_{1a}) = \\ -\omega^{2} m_{1a} x^{(0)}_{1ai}
+ \sum_{a',i'} K_{1ai,1a'i'} x^{(0)}_{1a'i'} \\
-\omega^{2} m_{2b} x_{2bj} + \sum_{b',j'} K_{2bj,2b'j'} x_{2b'j'}
- q_{2b} E_{j} (\vec{r}_{2b}) = 0 \\
\left(\frac{c^{2}}{\omega^{2}} \nabla \times (\nabla \times) -
1\right) \vec{E} = \\ \left(\sum_{a} q_{1a} \vec{x}_{1a} \delta^{3}
(\vec{x} - \vec{r}_{1a}) + \sum_{b} q_{2b} \vec{x}_{2b} \delta^{3}
(\vec{x} - \vec{r}_{2b}) \right)
\end{multline}
and these equations may again be collected into matrix form and
identified with the generic linear response operators. For polarizable
bodies $\alpha \in \{1, 2\}$, the operators
\begin{equation}
\hat{Z}^{(0)}_{\alpha} \to K_{\alpha\alpha} - \omega^{2} M_{\alpha}
\end{equation}
are the equations of motion defining the response. Meanwhile,
$\vec{E}$ is a field defined throughout all space, so matrix products
correspond to convolution integrals in space: this means the operators
\begin{align}
\hat{Z}^{(0)}_{3} &\to \frac{c^{2}}{\omega^{2}} \nabla \times (\nabla \times) - 1 \\
\hat{Y}^{(0)}_{3} &= \hat{Z}^{(0)-1}_{3} \to \mathbb{G}^{\mathrm{vac}}
\end{align}
correspond to the vacuum Maxwell partial differential operator and
associated Green's function. Finally, in the first, second, and third
equations, the coupling to the third component, i.e. the vacuum EM
field, corresponds to
\begin{equation}
\Delta \hat{Z}_{\alpha,3} \to -\sum_{a} q_{\alpha a} \delta_{ij}
\delta^{3} (\vec{x} - \vec{r}_{\alpha a})
\end{equation}
which is the convolution operator representing the charge density of
point dipoles constituting each polarizable body (with a sign flip due
to the convention chosen for the general linear response formulas):
these coupling operators are real-valued reciprocal operators, as
evinced in the equations of motion. This also means that for $\alpha
\in \{1, 2\}$, the material response operators may be written in
position space as
\begin{multline}
\Delta \hat{Z}_{3,\alpha} \hat{Y}^{(0)}_{\alpha} \Delta
\hat{Z}_{\alpha,3} \to \mathbb{V}_{\alpha ij} (\omega, \vec{x},
\vec{x}') = \\ \sum_{a,a'} q_{\alpha a} ((K_{\alpha\alpha} -
\omega^{2} M_{\alpha})^{-1})_{ai,a'j} q_{\alpha a'} \delta^{3}
(\vec{x} - \vec{r}_{\alpha a}) \delta^{3} (\vec{x}' -
\vec{r}_{\alpha a'})
\end{multline}
which is exactly the susceptibility $\mathbb{V}_{\alpha}$ of a
collection of point dipolar harmonic oscillators, while
\begin{equation}
\hat{Y}_{3} \to (\mathbb{G}^{\mathrm{vac}-1} - \mathbb{V}_{1} -
\mathbb{V}_{2})^{-1} = \mathbb{G}
\end{equation}
is exactly the Maxwell Green's function in the presence of
susceptibilities $\hat{\chi}_{\alpha}$. Thus, the heat transfer
between the two polarizable bodies can be written as
\begin{equation} \label{eq:GFPhi2bodyRHT}
\Phi(\omega) = 4~\trace{\operatorname{asym}(\mathbb{V}_{1})
\mathbb{G}^{\dagger} \operatorname{asym}(\mathbb{V}_{2}) \mathbb{G}}
\end{equation}
which exactly matches the fluctuational EM
expression~\cite{JinPRB2019}. Additionally, the assumptions underlying
the derivation of the upper bound on the transmission eigenvalues hold
here, so those derivations remain valid in this context: the coupling
operators representing the negative charge densities and real-valued
and reciprocal, and $\operatorname{asym}(\mathbb{G}^{\mathrm{vac}-1}) = 0$ comes
from the properties of Maxwell's equations, while the fact that
$\operatorname{asym}(\mathbb{G}^{\mathrm{vac}})$ does not vanish due to free space
supporting outward propagation of EM energy is irrelevant to those
particular derivations. Thus, the general NEGF formalism for heat
transfer in linear response systems can be exactly mapped to the
specific fluctuational EM formalism for linear RHT.
\begin{table*}[t]
\centering
\begin{tabular}{|l|l|l|}
\hline Heat Transfer Mechanism & Phonons & Photons \\ \hline \hline
Components 1, 2 & Infinite reservoirs (leads) & Polarizable bodies
\\ \hline
Component 3 & Compact central device & Vacuum EM field (all space)
\\ \hline $\hat{Y}^{(0)}_{\alpha}$: $\alpha \in \{1, 2\}$ &
\shortstack{Uncoupled lead mechanical Green's function} &
Susceptibilities $\mathbb{V}_{\alpha}$ \\ \hline
$\Delta \hat{Z}_{\alpha,3}$: $\alpha \in \{1, 2\}$ & Interface
lead/device harmonic couplings & All atom charges \\ \hline
$\hat{Y}_{3}$ & \shortstack{Coupled device mechanical Green's
function} & \shortstack{Maxwell Green's function
$(\mathbb{G}^{\mathrm{vac}-1} - \mathbb{V}_{1} -
\mathbb{V}_{2})^{-1}$} \\ \hline
\end{tabular}
\caption{Comparison of components and relevant linear response
quantities between PCHT and RHT within our NEGF heat-transfer formalism.}
\label{tab:comparisontab}
\end{table*}
Physically, the harmonic oscillators may represent valence electrons
or nuclei dressed by inner-shell electrons, and the couplings, namely
the effective charges, along with the effective masses and spring
constants are again computed via density functional theory. We again
stress that the correspondences $\Delta \hat{Z}_{3,\alpha}
\hat{Y}^{(0)}_{\alpha} \Delta \hat{Z}_{\alpha,3} \to
\mathbb{V}_{\alpha ij} (\omega, \vec{x}, \vec{x}')$ for RHT are
generally applicable even for more than two polarizable bodies coupled
to the vacuum EM field, which allows more general scenarios for PCHT
to be treated using~\eqref{Phimn}~\cite{PolimeridisPRB2015,
VenkataramPRL2018}. Furthermore, the derivation of Landauer-like
formulas for RHT~\eqref{GFPhi2bodyRHT} is generally applicable for
linear media even when the susceptibilities $\mathbb{V}_{\alpha}$ do
not describe harmonic oscillator response functions; our use of
harmonic oscillators was for convenience in writing a Lagrangian and
explaining salient features through physical intuition. Finally, we
emphasize that unlike previous work which has typically depended on
high-symmetry geometries and the assumption of the EM near-field
regime~\cite{TangPRAPP2019, WangPRE2018, ZhangPRB2018, JiangPRB2017,
YuNATURE2017}, these derivations are applicable to arbitrary
geometries from the near- through far-field regimes.
\section{Comparisons between PCHT and RHT} \label{sec:comparison}
Before proceeding, it is useful to summarize the comparisons between
PCHT and RHT specifically focusing on the case of two bodies
interacting through a third component (either a third body for PCHT or
the EM field for RHT), an analogy which is summarized
Table~\ref{tab:comparisontab} and illustrated schematically
in~\figref{photonradiationschematics}. While the basic formalisms are
essentially identical and both obey the same upper bounds, in what
follows we emphasize a few of the distinctions.
The typical situation considered for PCHT involves two semi-infinite
leads connected by a much smaller molecular junction or interfacial
region. As a result, when mapping $\hat{Y}^{(0)}_{\alpha} \to
g_{\alpha}$ for $\alpha \in \{1, 2\}$, even though the microscopic
oscillators have no dissipation so $\operatorname{asym}(g_{\alpha}^{-1}) \to 0$, the
fact that the leads are semi-infinite and act as thermodynamic
reservoirs means $\operatorname{asym}(g_{\alpha}) \neq 0$: this represents loss of
energy through far-field propagation of phonons into the bulks of the
leads. Meanwhile, when mapping $\hat{Y}^{(0)}_{3} \to g_{3}$ for the
junction or interfacial region, the compactness of that intermediate
body precludes dissipation through far-field propagation of phonons,
so not only is it true that $\operatorname{asym}(g_{3}^{-1}) \to 0$ but it is also
true that $\operatorname{asym}(g_{3}) \to 0$. Moreover, the smallness of the
intermediate body means that it is typically easier to evaluate the
matrix products and inverses in the space of the intermediate body
through~\eqref{Phi2bodyviathird}. The situation is flipped for RHT,
where typically energy exchange is considered between two compact
bodies via EM fields that propagate through all of space. As a result,
when mapping the response of lossless oscillators constituting each
polarizable body in the mapping $\Delta \hat{Z}_{3,\alpha}
\hat{Y}^{(0)}_{\alpha} \Delta \hat{Z}_{\alpha,3} \to \mathbb{V}_{\alpha}$ for
compact bodies $\alpha \in \{1, 2\}$, taking literally the lack of
dissipation would strictly imply that $\operatorname{asym}(\mathbb{V}_{\alpha}) \to 0$, so
heat transfer \& other fluctuational EM phenomena would not
exist. Realistically, these atomic oscillators are not perfectly
lossless but are subject to losses through scattering and propagation
of energy, which we do not consider here; this can be accounted for by
properly including reservoir DOFs in the Lagrangian and
performing some renormalization like decimation as in the phonon case
for a physically-motivated reservoir, or more typically by
phenomenologically adding an appropriate small imaginary part to some
part of $\mathbb{V}_{\alpha}$.
Meanwhile, when mapping $\hat{Y}^{(0)}_{3} \to \mathbb{G}^{\mathrm{vac}}$ through all of
space, while it is true that $\operatorname{asym}(\mathbb{G}^{\mathrm{vac}-1}) \to 0$
allows the same Landauer bounds to hold for RHT as for PCHT, the
ability of free space to support outward propagation of EM energy also
means $\operatorname{asym}(\mathbb{G}^{\mathrm{vac}}) \neq 0$. Moreover, the fact that the polarizable
bodies occupy compact regions in space (as opposed to all of space)
means that it is typically easier to evaluate the matrix products and
inverses in the spaces of the polarizable bodies
through~\eqref{Phi2bodydirect}. In particular, by using the operator
correspondences from the previous section and linking
\eqref{Phi2bodyviathird} to a special case of \eqref{Phi2bodydirect}
as above, it can be shown that \eqref{Phi2bodydirect} exactly
reproduces the T-operator formula for RHT~\cite{KrugerPRB2012}. Along
these lines, we finally note that in PCHT, the off-diagonal block of
the Green's function of component 3 in isolation connecting the
respective atoms coupled to each of the other components, which may be
denoted $P_{3(2)} g_{3} P_{3(1)}$, has a size, and therefore a maximum
rank, that scales as the surface areas of component 3 coupled with
each of the other components. For the case of RHT, the analogous
quantity is $\mathbb{P}_{2} \mathbb{G}^{\mathrm{vac}} \mathbb{P}_{1}$, where $\mathbb{P}_{\alpha}$ is the
projection operator onto the volume of body $\alpha$: this seems to
contrast with the dependence on surface area for PCHT. However, the EM
surface equivalence theorem~\cite{HarringtonJEWA1989,
RengarajanIEEE2000, ReidPRA2013, RodriguezPRB2013, OteyJQSRT2014,
SCUFF1} shows that the fields radiated by any volumetric
polarization distribution to the exterior of some fictitious bounding
surface can be exactly reproduced in that exterior region by an
equivalent surface current distribution, which therefore suggests that
the rank of $\mathbb{P}_{2} \mathbb{G}^{\mathrm{vac}} \mathbb{P}_{1}$ actually scales with the
\emph{surface} of each body, thereby producing a similar result as for
mechanical waves. The underlying physical reasons are a little
different: the general boundary conditions of EM fields at material
interfaces for radiation contrast with the specific form of coupling
of nearest-neighbor atoms for phonon propagation. That said, the
similarities can be intuitively understood as arising from the similar
physics governing mechanical wave propagation through homogeneous
media as EM wave propagation through vacuum or homogeneous media: the
spring constant matrix $K$ governing mechanical wave propagation
through a medium is essentially a discrete-space analogue of the
$\nabla \times (\nabla \times)$ operator governing EM wave
propagation, and both of these operators are then equated to double
time derivatives of the corresponding field quantities. Finally, we
note that in the concluding remarks, we connect this paper to an
accompanying manuscript that leverages this generic NEGF formalism to
generalize recent bounds on RHT~\cite{MoleskyPRB2020,
VenkataramPRL2020} to include PCHT: we point out that these bounds
rely heavily on the singular values of the off-diagonal blocks
$\mathbb{P}_{2} \mathbb{G}^{\mathrm{vac}} \mathbb{P}_{1}$ in the case of RHT, or $P_{3(2)} g_{3}
P_{3(1)}$ in the case of PCHT.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{twoC250carbynecollinearfigure.eps}
\caption{\textbf{Conduction and radiation between collinear wires.}
(a) Heat transfer coefficient $\frac{\partial P}{\partial T}$ at
room temperature ($T = 300~\mathrm{K}$) between two collinear 250
atom-long carbyne wires in vacuum, comparing the cases when heat
transfer is due purely to radiation (blue) or conduction (red)
versus both together (green). (b) Same as (a) zoomed in for $d \in
[0.1~\mathrm{nm}, 1~\mathrm{nm}]$. (c) Landauer energy transfer
spectrum $\Phi$ (independent of $T$) for $d = 0.281~\mathrm{nm}$,
clearly demonstrating the existence of nontrivial resonances.}
\label{fig:carbyne}
\end{figure*}
\section{Unifying PCHT and RHT} \label{sec:unification}
At nanometric and smaller separations, we expect that both PCHT and
RHT could exhibit comparable contributions to overall heat transfer
between two material bodies, whether through approach to direct
contact or through contact with an intermediate
junction~\cite{KimNATURE2015, KloppstechNATURE2017, CuiNATURE2017,
CuiSCIENCE2017, CuiNATURE2019, StGelaisNATURE2016}. Thus motivated,
we use this section to present a method for unifying both forms of
heat transfer in both of these scenarios. This method is based on the
retarded many-body (RMB) framework of mesoscale fluctuational
EM~\cite{VenkataramPRL2018, VenkataramPRL2017, VenkataramSCIADV2019},
allowing for accurate modeling of fluctuational EM phenomena,
including RHT, in atom-scale systems.
Each body $\alpha \in \{1, 2, 3\}$ comprises $N_{\alpha}$ atoms
labeled $a, b, c$. Each atom is centered at an equilibrium position
$\vec{r}_{\alpha a}$ and has an effective nuclear oscillator of mass
$m_{\mathrm{I}\alpha a}$ which couples to other nuclear oscillators
within the same body and which may couple to nuclear oscillators in
other bodies at interfaces: these couplings are encoded in the
matrices $K_{\mathrm{I}\alpha\alpha}$ within the same body and
$K_{\mathrm{I}\alpha\beta}$ between different bodies, where the former
has dimension $3N_{\alpha} \times 3N_{\alpha}$ while the latter has
dimension $3N_{\alpha} \times 3N_{\beta}$. The effective nuclear
oscillator in each atom is also coupled to an effective valence
electronic oscillator of mass $m_{\mathrm{e}\alpha a}$ through an
isotropic spring constant $k_{\mathrm{e}\alpha a}$. The valence
electronic oscillators couple as point charges to the vacuum EM field
via the charge $q_{\mathrm{e}\alpha a}$; these electrons along with
the inner electrons screen the nuclei, so we model the nuclear
oscillators as having no direct coupling to the EM field. The
displacements of the effective valence electronic oscillators are
labeled $x_{\mathrm{e}\alpha ai}$, while those of the nuclear
oscillators are labeled $x_{\mathrm{I}\alpha ai}$, for Cartesian
direction $i$. We collect the displacements into
$3N_{\alpha}$-dimensional vectors $x_{\mathrm{e}\alpha}$ and
$x_{\mathrm{I}\alpha}$, and the masses, charges, and valence
electronic spring couplings into diagonal $3N_{\alpha} \times
3N_{\alpha}$ matrices $M_{\mathrm{e}\alpha}$, $M_{\mathrm{I}\alpha}$,
$Q_{\mathrm{e}\alpha}$, and $K_{\mathrm{e}\alpha}$. Additionally, the
electric field in vacuum must be evaluated at each equilibrium
position $\vec{r}_{\alpha a}$ when entering the equations of motion
for the effective valence electronic oscillators, so we collect the
$N_{\alpha}$ Cartesian vectors $\vec{E}(\vec{r}_{\alpha a})$ into the
$3N_{\alpha}$-dimensional vector $e_{\mathrm{e}\alpha}$.
For two bodies coming into direct conductive contact (with no third
intermediate material body present) and interacting via the vacuum EM
field, we may use the above matrix notation to
write the equations of motion as
\begin{multline} \label{eq:PCHTRHTequationsofmotion}
(K_{\mathrm{e}\alpha} - \omega^{2}
M_{\mathrm{e}\alpha})x_{\mathrm{e}\alpha} - K_{\mathrm{e}\alpha}
x_{\mathrm{I}\alpha} - Q_{\mathrm{e}\alpha} e_{\mathrm{e}\alpha}
\\ = ((K_{\mathrm{e}\alpha} - \omega^{2}
M_{\mathrm{e}\alpha})x_{\mathrm{e}\alpha} - K_{\mathrm{e}\alpha}
x_{\mathrm{I}\alpha})\delta_{\alpha, 1} \\ (K_{\mathrm{e}\alpha} -
\omega^{2} M_{\mathrm{I}\alpha}) x_{\mathrm{I}\alpha} + \sum_{\beta}
K_{\mathrm{I}\alpha\beta} x_{\mathrm{I}\beta} -
K_{\mathrm{e}\alpha} x_{\mathrm{e}\alpha} = 0
\\ \left(\frac{c^{2}}{\omega^{2}} \nabla \times (\nabla \times) -
1\right) \vec{E} = \\ \sum_{\alpha,a} q_{\mathrm{e}\alpha a}
\vec{x}_{\mathrm{e}\alpha a} \delta^{3} (\vec{x} - \vec{r}_{\alpha
a})
\end{multline}
for each $\alpha, \beta \in \{1, 2\}$, $a \in \{1, \ldots,
N_{\alpha}\}$, and $i \in \{x, y, z\}$, for sources in body 1. We may
then formally solve the final equation and eliminate
$e_{\mathrm{e}\alpha}$ in favor of $x_{\mathrm{e}\alpha}$ and
$x_{\mathrm{I}\alpha}$, yielding the equations of motion
\begin{multline} \label{eq:PCHTRHTreducedequations}
(K_{\mathrm{e}\alpha} - \omega^{2}
M_{\mathrm{e}\alpha})x_{\mathrm{e}\alpha} - K_{\mathrm{e}\alpha}
x_{\mathrm{I}\alpha} - \sum_{\beta} Q_{\mathrm{e}\alpha}
G^{\mathrm{vac}}_{\alpha\beta} Q_{\mathrm{e}\beta}
x_{\mathrm{e}\beta} \\ = ((K_{\mathrm{e}\alpha} - \omega^{2}
M_{\mathrm{e}\alpha})x^{(0)}_{\mathrm{e}\alpha} -
K_{\mathrm{e}\alpha} x^{(0)}_{\mathrm{I}\alpha})\delta_{\alpha, 1}
\\ (K_{\mathrm{e}\alpha} - \omega^{2} M_{\mathrm{I}\alpha})
x_{\mathrm{I}\alpha} + \sum_{\beta} K_{\mathrm{I}\alpha,\beta}
x_{\mathrm{I}\beta} - K_{\mathrm{e}\alpha} x_{\mathrm{e}\alpha} = 0
\end{multline}
where $G^{\mathrm{vac}}_{\alpha\beta}$ is the $3N_{\alpha} \times
3N_{\beta}$ matrix whose elements are
$G^{\mathrm{vac}}_{ij}(\omega, \vec{r}_{\alpha a},
\vec{r}_{\beta b})$ for each pair of atomic coordinates. Hence, we
identify the relevant operators as $2\times 2$ block matrices
\begin{equation} \label{eq:PCHTRHToperators}
\begin{split}
\hat{Z}^{(0)}_{\alpha} &\to \begin{bmatrix}
K_{\mathrm{e}\alpha} - \omega^{2} M_{\mathrm{e}\alpha} &
-K_{\mathrm{e}\alpha} \\
-K_{\mathrm{e}\alpha} & K_{\mathrm{e}\alpha} +
K_{\mathrm{I}\alpha\alpha} - \omega^{2} M_{\mathrm{I}\alpha}
\end{bmatrix} \\
\Delta \hat{Z}_{\alpha\beta} &\to \begin{bmatrix}
-Q_{\mathrm{e}\alpha} G^{\mathrm{vac}}_{\alpha\beta}
Q_{\mathrm{e}\beta} & 0 \\
0 & K_{\mathrm{I}\alpha\beta} (1 - \delta_{\alpha\beta})
\end{bmatrix}
\end{split}
\end{equation}
where the top row and left column blocks represent the effective
valence electronic DOFs, while the bottom row and right
column blocks represent the effective nuclear degrees of
freedom. Strictly speaking, the matrices $-\omega^{2}
M_{\mathrm{e}\alpha}$ and $-\omega^{2} M_{\mathrm{I}\alpha}$ should
respectively be replaced by $-\operatorname{i}\omega B_{\mathrm{e}\alpha} -
\omega^{2} M_{\mathrm{e}\alpha}$ and $-\operatorname{i}\omega B_{\mathrm{I}\alpha}
- \omega^{2} M_{\mathrm{I}\alpha}$ in order to account for nonzero
dissipation, though the dissipation matrices $B_{\mathrm{e}\alpha}$
and $B_{\mathrm{I}\alpha}$ may be taken to be infinitesimal; also,
once again, the diagonal blocks $K_{\mathrm{I}\alpha\alpha}$ entering
$\hat{Z}^{(0)}_{\alpha}$ should actually include the effects of
couplings to nuclear oscillators in other bodies as are present in the
off-diagonal blocks $K_{\mathrm{I}\alpha\beta}$ for all $\beta \neq
\alpha$. With details explained in~\cite{VenkataramPRL2018,
VenkataramSCIADV2019}, the RMB oscillator matrix parameters
$Q_{\mathrm{e}\alpha}$, $M_{\mathrm{e}\alpha}$,
$M_{\mathrm{I}\alpha}$, $K_{\mathrm{e}\alpha}$, and
$K_{\mathrm{I}\alpha\alpha}$ (the latter initially excluding couplings
to nuclear oscillators in other bodies) along with the equilibrium
atomic positions are all computed using density functional theory
(DFT) for each body in isolation, while the matrices
$B_{\mathrm{e}\alpha}$ and $B_{\mathrm{I}\alpha}$ are assigned
phenomenological values. These $2\times 2$ block matrices can then be
used in place of $\hat{Z}^{(0)}_{\alpha}$ and $\Delta
\hat{Z}_{\alpha\beta}$ in the formula for two components with general
couplings~\eqref{Phi2bodydirect} to find the combined heat transfer
including PCHT and RHT: the couplings among valence electronic and
nuclear DOFs through EM fields means that PCHT and RHT
contributions are not separable, but in fact affect each
other~\cite{WangPRE2018, ZhangPRB2018, KlocknerPRB2017}.
For two bodies whose nuclear coordinates are coupled only to a third
intermediate body, which also has nuclear and valence electronic
DOFs, in which all electronic coordinates are coupled to
the EM field, the formalism is similar to above. In particular, the
formulas in~\eqref{PCHTRHTequationsofmotion} still hold for all bodies
$\alpha, \beta \in \{1, 2, 3\}$, although $K_{\mathrm{I}1,3}$ and
$K_{\mathrm{I}2,3}$ and their transposes are the only nonzero
off-diagonal blocks of $K_{\mathrm{I}}$. With that caveat in mind,
this further means that~\eqref{PCHTRHTreducedequations} and the
correspondences in~\eqref{PCHTRHToperators} holds as well for all
bodies $\alpha, \beta \in \{1, 2, 3\}$. That said, the fact that
$\Delta \hat{Z}_{\alpha\beta}$ has nonzero blocks for all $(\alpha,
\beta)$ means that~\eqref{Phi2bodyviathird} cannot be used. Instead,
the more general formula~\eqref{Phimn} for the energy transfer
spectrum must be used, plugging the $2\times 2$ block matrices
in~\eqref{PCHTRHToperators} into the overall $3\times 3$ block
matrices
\begin{align}
\hat{Z}^{(0)} &=
\begin{bmatrix}
\hat{Z}^{(0)}_{1} & 0 & 0 \\
0 & \hat{Z}^{(0)}_{2} & 0 \\
0 & 0 & \hat{Z}^{(0)}_{3}
\end{bmatrix}
\\
\Delta \hat{Z} &=
\begin{bmatrix}
\Delta \hat{Z}_{1,1} & \Delta \hat{Z}_{1,2} & \Delta \hat{Z}_{1,3}
\\
\Delta \hat{Z}_{2,1} & \Delta \hat{Z}_{2,2} & \Delta \hat{Z}_{2,3}
\\
\Delta \hat{Z}_{3,1} & \Delta \hat{Z}_{3,2} & \Delta \hat{Z}_{3,3}
\end{bmatrix}
\end{align}
to evaluate~\eqref{Phimn}.
These formulas for the energy transfer spectrum and associated linear
response operators are thus the application of the general NEGF
formalism for combined PCHT and RHT. In contrast to the derivations of
pure RHT which ultimately do not depend on the form of the
susceptibilities $\mathbb{V}_{\alpha}$ as long as it is linear, these
particular derivations do depend on the harmonicity of the material
models, though they may be generalizable through a more complicated
formalism. However, beyond that approximation as well as the
assumptions regarding material dissipation, these formulas are
independent of specific geometries and material properties, and can be
evaluated in the EM near- or far-field regimes. Additionally, we point
out that unlike previous works which have cast formulas for combined
electronic conduction and RHT in a more complicated (Meir--Wingreen)
form rather than the typical Landauer/Caroli form~\cite{ZhangPRB2018,
TangPRAPP2019, WangPRE2018} as electrons and photons obey different
quantum statistics, no such complication arises here because phonons
and photons obey the same statistics.
We apply this unified formalism to an illustrative model of heat
transfer between two collinear 250 atom-long atomically thin wires,
taken to be made of carbon (i.e. carbyne wires), and particularly
compute the heat transfer coefficient $\frac{\partial P}{\partial T}$
at room temperature ($T = 300~\mathrm{K}$). Specifically, we compute
the heat transfer coefficient $\frac{\partial
P_{\mathrm{both}}}{\partial T}$ by calculating the Landauer energy
transfer spectrum $\Phi_{\mathrm{both}}$ arising from
plugging~\eqref{PCHTRHToperators} as written into~\eqref{Phimn},
$\frac{\partial P_{\mathrm{rad}}}{\partial T}$ by computing
$\Phi_{\mathrm{rad}}$ arising from plugging~\eqref{PCHTRHToperators}
with $K_{\mathrm{I}\alpha\beta} = 0$ for $\beta \neq \alpha$ (so
$K_{\mathrm{I}\alpha\alpha}$ refers only to the spring constant
matrices among nuclei for each body in isolation) into~\eqref{Phimn},
and $\frac{\partial P_{\mathrm{cond}}}{\partial T}$ by computing
$\Phi_{\mathrm{cond}}$ arising from plugging~\eqref{PCHTRHToperators}
with $G^{\mathrm{vac}} = 0$ for all pairs of electronic oscillators
into~\eqref{Phimn}; in all cases, $\Phi$ refers to $\Phi^{(1)}_{2}$.
Within each body, as described above, the charges, masses, and spring
constants are all taken from DFT evaluated for each body in isolation,
the matrix elements of the Maxwell Green's function $\mathbb{G}^{\mathrm{vac}}$ are
evaluated in a Gaussian basis to mitigate short-range EM
divergences~\cite{VenkataramPRL2018, HermannCR2017,
VenkataramPRL2017}, and the dissipation matrices are chosen such
that $B_{\mathrm{e}} = \gamma_{\mathrm{e}} M_{\mathrm{e}}$ \&
$B_{\mathrm{I}} = \gamma_{\mathrm{I}} M_{\mathrm{I}}$ hold with
$\gamma_{\mathrm{e}} = 10^{11}~\mathrm{s}^{-1}$ \&
$\gamma_{\mathrm{I}} = 10^{13}~\mathrm{s}^{-1}$; the damping rates are
chosen phenomenologically to be large enough to allow reasonably
coarse frequency sampling, but small compared to the characteristic
frequencies of the relevant polaritons. For computational simplicity,
these properties are not recomputed as functions of the separation
between the bodies, but while we expect such recomputation to yield
significantly different results due to the greater probability of
supporting longer-wavelength collective electronic and phononic waves
when the wires are in proximity, such recomputation could in principle
be performed consistently with this formalism. Likewise, for
computational simplicity, the off-diagonal blocks of $K_{\mathrm{I}}$
for each body (including both electronic and nuclear oscillator
coordinates) have only the couplings between each end atom nearest to
the other molecule be nonzero, and these are modeled via the Morse
potential, but this could be further generalized in future work. The
Morse potential spring constant for a bond of length $r$ compared to
equilibrium length $a_{0}$ is computed as $k(r) = -\frac{1}{r - a_{0}}
\frac{\partial U_{\mathrm{Morse}}}{\partial r}$, where the potential
energy $U_{\mathrm{Morse}}(r) = U_{\mathrm{min}} (1 -
e^{-\sqrt{k_{0}/(2U_{\mathrm{min}})}(r - a_{0})})^{2}$ exhibits a
harmonic well of depth $U_{\mathrm{min}}$ and curvature defined by the
equilibrium spring constant $k(a_{0}) = k_{0}$, all of which are
empirical parameters, and exponentially decays as $r \gg a_{0}$.
As can be seen in~\figref{carbyne}(a, b), many interesting features
arise from the coupling of conductive and radiative processes. The
exponential decay of the Morse potential with distance means that for
$d > 0.4~\mathrm{nm}$, conduction ceases to have any meaningful effect
on the heat transfer, and the total heat transfer aligns with that of
pure radiation. However, for decreasing $d \leq 0.4~\mathrm{nm}$, not
only does conduction become more significant, but the total heat
transfer including both radiative and conductive processes falls
\emph{below} the corresponding individual cases, and only rises above
both for $d < 0.2~\mathrm{nm}$ before all three powers
saturate. Therefore, this unified formalism is clearly necessary for
subnanometric separations, as the total power including both PCHT and
RHT is not simply the sum of the individual contributions (as has been
found in related systems involving electronic
conduction~\cite{ZhangPRB2018}), but behaves in a much more
complicated way.
In~\figref{carbyne}(c), the Landauer energy transfer spectra $\Phi$
make clear that for small enough $d$ where conduction is nontrivial
(plotted for $d = 0.281~\mathrm{nm}$), the conduction spectrum only
has nontrivial contributions at lower frequencies $\omega <
10^{14}~\mathrm{rad/s}$. Meanwhile, the total spectrum rises above the
radiation spectrum for larger $\omega$ but falls below for smaller
$\omega$: the latter is more relevant given the exponential decay of
$\frac{\partial \Pi(\omega, T)}{\partial T}$ with $\omega$, leading to
$P_{\mathrm{both}} < P_{\mathrm{rad}}$ there. Ultimately, this occurs
due to the confluence of EM screening as captured by the Gaussian
basis functions along with shifts in the response due to conductive
coupling between nuclei of the two different wires: not only does this
shift the frequencies of resonances in the Landauer energy transfer
spectra, but it can also suppress the resulting amplitudes. This
therefore makes clear that the existence of situations where
$P_{\mathrm{both}}$ (or its derivative with respect to $T$) falls
between or below $P_{\mathrm{cond}}$ or $P_{\mathrm{rad}}$ is not
simply a fluke arising from a particular choice of $T$: $\Phi$ is
independent of $T$, yet the spectrum $\Phi_{\mathrm{both}}$, far from
being a simple case of superimposing $\Phi_{\mathrm{rad}}$ on
$\Phi_{\mathrm{cond}}$, shows a delicate interplay among radiative and
conductive effects in creating new hybrid resonances.
Our calculations are meant to be
qualitatively illustrative of the complexities of heat transfer when
both conduction and radiation contribute: they are not meant to be
quantitatively predictive given the practical limitations in
recomputing relevant oscillator parameters at each separation, but we
stress that these limitations are not fundamental to the formalism we
have presented.
\section{Concluding Remarks} \label{sec:conclusion}
We have demonstrated a general NEGF formulation of heat transfer
applicable to a wide variety of bosonic systems. This NEGF framework
is general enough to explain the salient features of PCHT and RHT
separately, show how upper bounds on PCHT can be generalized and then
applied to RHT, and demonstrate how to unify PCHT and RHT in
situations when both are strongly coupled and relevant. The latter is
particularly relevant at atomistic scales or separations, when
continuum material models begin to fail and the net heat transfer is
no longer simply the sum of individual radiative or phononic
contributions. We stress that our approach is general enough to treat
semiclassical heat transfer through other massless bosonic
excitations, not just photons or phonons. Moreover, while our analysis
of combined PCHT and RHT focused on effective valence electronic and
nuclear response as being represented by coupled harmonic oscillators,
more complicated linear response models could be considered as well,
which we leave for future work. We expect this framework to pave the
way for future works investigating the conjunction of PCHT and RHT in
complex geometries, particularly at separations where each is relevant
and where recent experiments have raised questions about where each
form of heat transfer is dominant~\cite{KimNATURE2015,
KloppstechNATURE2017, CuiNATURE2017, ChiloyanNATURE2015}.
In an accompanying manuscript, we generalize bounds previously derived
for RHT~\cite{MoleskyPRB2020, VenkataramPRL2020} using the generic
NEGF formalism for heat transfer in linear systems presented in this
paper. We particularly apply such bounds to PCHT, showing that
channel-based bounds on PCHT can be much tighter than the Landauer
limits of unity~\cite{BurklePRB2015, KlocknerPRB2017,
KlocknerPRB2018}.
\emph{Acknowledgments}.---The authors thank Sean Molesky for the
helpful comments and suggestions. This work was supported by the
National Science Foundation under Grants No. DMR-1454836, DMR 1420541,
DGE 1148900, the Cornell Center for Materials Research MRSEC (award
no. DMR1719875), the Defense Advanced Research Projects Agency (DARPA)
under agreement HR00111820046, and the Spanish Ministry of Economy and
Competitiveness (MINECO) (Contract No. FIS2017-84057-P). The views,
opinions and/or findings expressed herein are those of the authors and
should not be interpreted as representing the official views or
policies of any institution.
| bb8f34bd30366c295cba90bb20fb901a88eb84dc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this note, we examine the efficacy of a recently developed approach to the recovery of nonlinear water
waves from pressure measurements, by applying it to the celebrated extreme Stokes wave.
The reconstruction of the water wave surface profile from bottom pressure measurements is an important
issue for marine engineering applications, but corresponds to a difficult mathematical problem. From the
viewpoint of the qualitative modelling of water waves, the pressure plays an important role in establishing
various properties of travelling waves \citep{CS2010,hen11,lyons16}. Measuring the surface of water waves
directly is extremely difficult and costly, particularly in the ocean, so a commonly employed
alternative is to calculate the free-surface profile of water waves by way of the so-called pressure transfer
function, which recovers the free-surface elevation using measurements from submerged pressure transducers.
The key to the success of this approach is in the derivation of a suitable candidate for the pressure
transfer function, an issue which is the subject of a large body of experimental and theoretical research
\citep{BD,TH}. Until recent years, most studies were based on a linear transfer function \citep{ES}, primarily
due to the intractability of the nonlinear governing equations, and even then for irrotational waves
(cf. \cite{HT} for linear, and weakly nonlinear, recovery formulae for waves with vorticity). However,
the inadequacy of linear formulae for waves of even moderate amplitude is well known \citep{BD,TH},
hence the importance of taking into account nonlinear effects.
In \citep{Cp,D2}, various nonlinear nonlocal equations relating the dynamic pressure on the bed and the
wave profile were obtained without approximation from the governing equations using different approaches,
however these formulae are quite entangled and not very transparent. This is a significant impediment in
the pursuit of practical applications, and alternative approaches have focussed on the reconstruction of
nonlinear water waves in approximate intermediate and shallow water regimes \citep{BL}. Recently, new
exact tractable relations were derived which allow a straightforward numerical procedure for deriving the
free surface from the pressure at the bed \citep{CC}. This surface reconstruction procedure requires the
numerical resolution of an ordinary differential equation, and in \citep{clam13} it was shown that it is
possible to solve this ODE analytically by recasting it in terms of an implicit functional equation. The
approach of \citep{clam13,CC} is quite unique in that it works directly with nonlinear waves in the physical
plane, as opposed to transforming to a conformal plane, and instead obtains recovery formulae through the
manipulation of holomorphic mappings of the physical variables. The directness of this approach offers a
number of advantages, and we demonstrate the efficacy and immediate applicability of this approach, both
from a theoretical and practical viewpoint, by applying it to the wave-profile recovery problem for Stokes
extreme wave, or wave of greatest height.
The wave of greatest height was postulated to exist by \citet{sto1880}, who used formal
arguments to obtain it as the limit of smooth periodic travelling waves. Remarkably, the existence of the
extreme wave solution with a stagnation point at the crest, along with many important qualitative features (such as convexity, the existence
of an angular crest with inner angle $2\pi/3$) were rigorously proven only in recent decades (cf.
\citep{BT,con2016,tol} for surveys of the rigorous analysis of Stokes waves). Also, physical properties
of the extreme Stokes wave are not obviously transparent: although it is the Stokes wave of greatest height,
it is not the most energetic, or fastest, Stokes wave \citep{SF}. While the rigorous mathematical analysis
of nonlinear Stokes waves is already technically challenging for regular (smooth) waves, this is particularly
so for the Stokes wave of greatest height, which is impervious to standard analytical methods due to the
presence of a stagnation point at the wave-crest. A commonly employed approach for the extreme Stokes wave
is to use continuation arguments which assume that the extreme wave is a uniform limit of smooth (`almost
extreme') Stokes waves \citep{BT,const12,lyons16,tol}. Due to the direct nature of the pressure surface
recovery approach developed in \citep{clam13}, we do not need to employ convoluted continuation arguments
in our analysis. This illustrates the robustness and applicability of the reconstruction method of
\citep{clam13}, which does not invoke intermediate conformal changes of variables, and which accordingly
offers potential for further development towards capturing more physically complex scenarios, such as the
recovery of nonlinear rotational water waves. {{For extreme and near-extreme waves, the advantage of working
in the physical plane, instead of the conformal one, is outlined in \citep{Clamond2018}; the
present paper provides further support to this claim.}}
\section{Hypothesis and notations}\label{secnot}
We consider two-dimensional periodic travelling waves propagating on the surface of an incompressible
inviscid fluid under the restoring force of gravity. The wave is periodic\footnote{
Obviously, aperiodic waves are obtained letting $k\to0$. Thus, discussing periodic waves (for
simplicity) is not a limitation for our purpose here.} in the $x$-direction, with
period $L$, and the wavenumber is given by $k=2\pi/L$. If $(X,Y)$ denotes horizontal and vertical coordinates
in a fixed reference frame, with $(U,V)$ the respective velocity field, then the travelling wave hypothesis
implies a functional dependence on the variables of the form $(X-ct,Y)$: the Galilean change of variables
\[
x=X-ct, \quad y=Y, \quad u=U-c, \quad v=V,
\]
transforms to a reference frame where the fluid motion is steady, with the resulting velocity field in
the frame moving with the wave given by $(u,v)$. The wave phase velocity $c$ is
an arbitrary constant as long as we have not defined what we mean by ``fixed frame of reference''.
If it is the frame where the horizontal velocity has zero mean at bottom (which corresponds to Stokes' first definition
of the wave speed), the wave phase velocity $c$ is such that $\left<\/U_\text{b}\/\right>=0$ in the fixed
frame, that is $\left<\/u_\text{b}\/\right>=-c$ in the frame of reference moving with the wave. Here $\left<\boldsymbol{\cdot}
\right>$ is the Eulerian average operator over one wavelength and, in what
follows, the subscripts `$\text{s}$' and `$\text{b}$' denote the restrictions to the free surface and
at the bed, respectively. Note that $c>0$ if the wave travels toward the increasing $x$-direction; a
choice that can always be made without lost of generality. Many other definitions of $c$ can of course
be made, with resulting consequences for the Bernoulli constant, and the gauge condition for the velocity
potential, for instance (cf. \citep{Clamond2017} for discussions on this, and related, matters).
The horizontal bottom is located at $y=-\depth$, and the free surface is described by $y=\eta(x)$: since
these boundaries are impermeable and steady we have $\sur{v}\/=\/\eta_x\/\sur{u}$ (with $\eta_x\/=\/\mathrm{d}
\eta/\mathrm{d} x$) and $\bot{v}\/=\/0$. The mean water level is located at $y=0$, which implies that
$\left<\eta\right>=0$ for the ($2\pi/k$)-periodic wave profile $\eta$, that is
\begin{align} \label{defmean}
\left<\,\eta\,\right>\, \stackrel{\text{\tiny{def}}}{=}\ \frac{k}{2\,\pi}\int_{-{\pi/k}}^{\,{\pi/k}}\eta(x)\ \mathrm{d}\/x\ =\ 0.
\end{align}
The local spin or rotation of an infinitesimal fluid element is measured by the vorticity, which is
expressed by $\omega=u_y-v_x$ for two-dimensional motion. An inherent property of inviscid flows is
that the vorticity of a given fluid element is preserved by the resultant fluid motion \citep{con2011},
and accordingly a fluid mass which is initially irrotational will remain so for all further times.
Therefore assuming irrotational motion, $u_y=v_x$, allied to the incompressibility condition $\nabla
\cdot (u,v)=0$, enables us to define a velocity potential, $\phi(x,y)$, and stream function, $\psi
(x,y)$, respectively, such that
$u\/=\/\phi_x\/=\/\psi_y$ and $v\/=\/\phi_y\/=\/-\/\psi_x$.
It follows that the complex potential, defined by $f(z)\/=\/\phi(x,y)\/+\/\mathrm{i}\/\psi(x,y)$, and the
complex velocity, $w\/=\mathrm{d} f/\mathrm{d} z\/=\/u\/-\/\mathrm{i}\/v$, are holomorphic functions of the complex abscissa
$z=x+\mathrm{i}\/y$ (where $\mathrm{i}^2=-1$). The stream function $\psi\/=\/\Im\{f\}$ is constant at $y=-\depth$ and
at $y=\eta$ since these boundaries are impermeable and steady, hence $f(z)$ represents a conformal
mapping from the fluid domain to a fixed rectangle, the so-called conformal plane, once the non-stagnation
condition
\begin{equation}\label{stag}
u(x,y) \leqslant \epsilon<0
\end{equation}
holds throughout the fluid domain.
The equation of motion is given by the Euler equation which, for irrotational fluid, can be expressed
in terms of the following Bernoulli condition
\begin{align} \label{bernbase}
2\,p\ +\ 2\,g\,y\ + \ u^2\ +\ v^2\ =\ B, \qquad\
x \in \mathds{R},\ -\depth \leqslant y \leqslant \eta,
\end{align}
where $B$ is the Bernoulli constant, and $p=p(x,y)$ is the pressure (relative to the atmospheric
pressure) divided by the (constant) density. At the free surface, surface tension is neglected and
the relative normalised pressure is zero, $\sur{p}\/=\/0$. From (\ref{defmean}) and (\ref{bernbase}),
we get
\begin{equation}\label{mt}
B\ =\,\left<\,u_\text{s}^{\,2}\,+\,v_\text{s}^{\,2}\,\right>\,=\,\left<\,u_\text{b}^{\,2}
\, \right>\,>\ 0,
\end{equation}
where the second equality in (\ref{mt}) follows from an application of the divergence theorem applied
to the vector field $(-2uv,u^2-v^2)$, cf. \citep{CC}. Note that $B=c^2$ in deep water ($\depth\to\infty$),
and for solitary waves ($k\to0$), since in these settings $(u,v)\to (-c,0)$ as $y\to -\infty$, and
$|x|\rightarrow \infty$, respectively.
Finally we observe that relations (\ref{bernbase}) and (\ref{mt}) yield
\begin{equation} \label{defpmean}
\left<\,\mathfrak{p}\,\right>\,=\ g\,\depth,
\end{equation}
where $\mathfrak{p}(x)\/\stackrel{\text{\tiny{def}}}{=}\/p(x,-\depth)=\bot{p}(x)$ is the normalised relative pressure
at the bed. This is a useful formula for determining the mean fluid depth $\depth$ from pressure
measurements at the flat bed.
\section{Stokes extreme wave}
In this paper, by `Stokes wave' we mean a periodic, irrotational water wave, with a strictly monotone
symmetric free-surface profile --- increasing from trough-to-crest and decreasing from crest-to-trough
--- which moves with a constant speed and with a constant form over a body of water on a flat bed. For
such waves, the functions $u$, $\eta$, $p$ are even functions and $v$ is an odd function with respect
to the $x$-variable; we assume the wave crest is located at $x=0$, and the wave troughs are located at
$x=\pm\pi/k$. It was shown by \citet{sto1880}, using formal methods, that regular wave solutions exist
for which the strict inequality \eqref{stag} holds throughout the fluid domain. Given this inequality,
it can be rigorously proven that a continuously differentiable free-surface must {\em a priori\/} be a
symmetric, real-analytic graph of a function: surveys of rigorous analytical results for Stokes waves
can be found in \citep{BT,con2011,con2016,tol}.
Stokes also posited, using heuristic arguments, that a limiting wave solution exists for which \eqref{stag}
fails and the equality $u=0$ is attained precisely at the wave crest, with the associated breakdown in
regularity of the wave-profile at the crest manifested in the form of an angular point\footnote{
{In the conformal plane, the curve $\eta(\phi)$ presents a cusp at the crest where
$\eta\propto|\phi|^{2/3}$ locally, while in the physical plane the curve $\eta(x)$ presents an angular
point where $\eta\propto|x|$ locally.}} which has a containing angle of $2\pi/3$. The existence of Stokes
waves is rigorously established by way of highly-technical bifurcation theory methods and, indeed,
it was only in recent decades that a number of fundamental questions concerning the existence of an
extreme Stokes wave, and properties of the associated wave profile, were rigorously settled (cf.
\citep{BT,con2016,tol}). For this limiting wave `of greatest height', or extreme Stokes wave, it can be
shown that the free-surface is real-analytic everywhere except for the stagnation point at the crest,
where it is merely continuous. We note that the wave crest is an `apparent' stagnation point, with
fluid particles only coming to rest there instantaneously, cf. \cite{const12}.
All technical complications arising in the rigorous analysis of extreme Stokes waves essentially stem from
the lack of regularity at the stagnation point, with the primary impediment being a breakdown in the
conformality of the mapping $f(z)$ due to the failure of condition \eqref{stag} at the wave-crest.
Nevertheless, a detailed asymptotic description of the wave profile, and velocity field, in the
neighbourhood of the wave-crest is possible. This work originated with \citet{sto1880}, who postulated
that the wave of maximum amplitude has an included crest angle of $2\pi/3$, the crest being a stagnation
point which, in the conformal plane, has a power $2/3$ singularity. It was shown by \citet{grant73}
that higher-order expansions in the vicinity of the crest involve non-algebraic powers (see also
\citep{nor74}); this work was placed on a rigorous footing in \citep{AF,mc87}.
Performing the same local analysis (around the crest at $x = 0$ where $\eta= a$) in the physical plane,
one obtains after some algebra
\begin{align}
\eta\, &=\ a\ -\ 3^{-1/2}\,|\/x\/|\ +\ \beta\,\kappa^{\nu-1}\,|\/x\/|^{\nu}\ +\ O\!\left(|x|^{2\nu-1}\right),
\label{etamaxloc}\\
w^2\, &=\ g\,(\/a\/+\/\mathrm{i}\/z\/)\ +\ \beta\,\kappa^{\nu-1}\left[\,3^{1+\nu}\,4^{-\nu}\,
(1+\nu+\nu^2)\,\right]^{1\over2}g\,(\/a\/+\/\mathrm{i}\/z\/)^\nu\ +\ O\!\left((a\/+\/\mathrm{i}\/z)^{2\nu-1}\right),
\label{w2maxloc}
\end{align}
where $\beta$ is a strictly positive (dimensionless) constant \citep{mc87}, $\kappa$ is a (freely
choosable) characteristic wavenumber (e.g. $\kappa=k$ or $\kappa=g/B$) and $\nu\approx2.20401861065003478$
is the smallest positive root\footnote{The next positive root of \eqref{traneq}
is $\nu_2\approx5.36>2\nu-1$. There are an infinite number of real positive roots $\nu_j$
of \eqref{traneq} such that $\nu_j\sim3j-{\textstyle{1\over2}}-\left.\sqrt{3}\right/2\pi j$ as $j\to\infty$. There
are also two complex roots, approximately $-{\textstyle{1\over2}}\pm\mathrm{i}\times1.0714$, that do not
appear in the extreme wave (as they would yield an unbounded free surface); they do play a role
in near extreme waves, however \citep{LHF77}.} of the transcendental equation \citep{grant73}
\begin{equation}\label{traneq}
\sqrt{3}\,\tan\!\left({\textstyle{1\over3}}\nu{\pi}\right)\,=\ -\/1\ -\ 2\,\nu^{-1}.
\end{equation}
In order to derive expression \eqref{w2maxloc}, we use (cf. eq. (3.3) in \citep{CC}) the relation
\begin{equation}\label{wCC}
w_\text{s}^{\,2}\ =\ 2\,g\,(a-\eta)\,(1-\mathrm{i}\eta_x)\,/\,(1+\mathrm{i}\eta_x),
\end{equation} and taking $\eta$ of the form \eqref{etamaxloc},
for some $\nu>1$, equation \eqref{wCC} yields
\begin{equation}
\frac{w_\text{s}^{\,2}}{g}\ =\ \frac{|\/x\/|}{\sqrt{3}}\ +\ \mathrm{i}\,x\ +\ \beta\,|\/\kappa\/x\/|^{\nu-1}
\frac{(3\nu-2)\,|x|\,-\,\mathrm{i}\,(\nu+2)\,\sqrt{3}\,x}{2}\ +\ O\!\left(|x|^{2\nu-1}\right),\label{w2smaxloc}
\end{equation}
that can be conveniently rewritten
\begin{equation}
\frac{w_\text{s}^{\,2}}{g\,(a+\mathrm{i}\/z_\text{s})}\ =\ 1\ -\ 3\,\beta\,|\/\kappa\/x\/|^{\nu-1}
\frac{\sqrt{3}\,+\,\mathrm{i}\,(2\nu+1)\,\operatorname{sgn}(x)}{4}\ +\ O\!\left(|x|^{2\nu-2}\right).\label{w2smaxlocbis}
\end{equation}
Then, seeking a complex velocity of the form $w^2=g(a+\mathrm{i} z)+A(a+\mathrm{i} z)^\nu+O\!\left((a+
\mathrm{i} z)^{2\nu-1}\right)$, comparison with \eqref{w2smaxlocbis} leads to $A$
as defined by \eqref{w2maxloc} and the relation \eqref{traneq}.
Note the simplicity of this approximation,
in particular the leading approximation for $w^2\approx g\/(\/a\/+\/\mathrm{i}\/z\/)$, which yields
at once (from the Bernoulli equation) the following approximation for the pressure near the wave crest:
\begin{equation}\label{prescrest}
p(x,y)\ \approx\ g\,(a-y)\ -\ {\textstyle{1\over2}}\,g\,\sqrt{\,x^2\,+\,(a-y)^2\,}.
\end{equation}
This relation will play an important role in section \ref{secew} below. To the best of our knowledge,
relations \eqref{w2maxloc} and \eqref{prescrest} have not appeared before in the literature.
A study of the qualitative behaviour of the velocity field for the extreme Stokes near the wave crest
has been explored in \cite{const12}, where it is shown that the velocity field exhibits the limiting
behaviours
\refstepcounter{equation}
\[
\lim_{x\to 0} \frac{u^2(x,\eta(x))}{g\,|x|}\,=\,\frac{\sqrt3}{2}, \qquad \lim_{x\to 0}
\frac{v^2(x,\eta(x))}{g\,|x|}\,=\,\frac{1}{2\,\sqrt3 },
\eqno{(\theequation{\mathit{a},\mathit{b}})}\label{limvel}
\]
which accord with the limits given by \eqref{w2maxloc} and \eqref{w2smaxloc}.
On the crest line $x=0$ (i.e., for $\Re\{z\}=0$ and $\Im\{z\}\leqslant a$)
we have $\Re\{w^2\}=u^2\geqslant0$ and $\Im\{w^2\}=0$, the relation \eqref{w2maxloc} then implies
at once that $\beta$ is necessarily positive. This result was first obtained by \citet{mc87} with
much more convoluted arguments. However, \citet{mc87} also proved that $\beta\neq0$, the possibility
$\beta=0$ being not easily ruled out here (that would require more than just a local analysis around
the crest).
From the inequality $\beta>0$, the semi-derivatives at the crest are given by
\refstepcounter{equation}
\[
\lim_{x\to0^{\pm}} \eta_{x}\,=\,\mp\,3^{-1/2}, \qquad
\lim_{x\to0^{\pm}} \eta_{xx}\,=\,0^{+}, \qquad
\lim_{x\to0^{\pm}} \eta_{xxx}\,=\,\pm\,\infty,
\eqno{(\theequation{\mathit{a},\mathit{b},\mathit{c}})}\label{limeta}
\]
in comparison with the corresponding semi-derivatives in the conformal plane (where, for example,
$\eta_\phi$ is infinite at the crest since $\eta\propto|\phi|^{2/3}$ locally).
\section{Wave surface recovery: regular waves}\label{secCC}
In this section, we briefly recall the salient features of the surface profile recovery procedures
presented in \citep{clam13,CC}. The function $\mathfrak{P}$ defined by
\begin{equation}\label{P}
{\mathfrak P}(z)\ \stackrel{\text{\tiny{def}}}{=}\ {\textstyle{1\over2}}\,B\ +\ g\,\depth\ -\ {\textstyle{1\over2}}\,w^2(z)\ =\ {\textstyle{1\over2}}\,B\
+\ g\,\depth\ -\ {\textstyle{1\over2}}\,(u^2-v^2)\ +\ \mathrm{i}\,u\,v,
\end{equation}
is holomorphic, and its restriction at the flat bed coincides with the real-valued function
$\mathfrak{p}(x)$, since
\(
\bot{\mathfrak{P}}\/=\/\mathfrak{P}(x-\mathrm{i}\depth)\/=\/g\/d\/+\/{\textstyle{1\over2}}\/(\/B\/-\/u_\text{b}^{\,2}\/)
\/=\/\mathfrak{p}(x).
\)
Accordingly ${\mathfrak P}(z)\/=\/{\mathfrak p}(z+\mathrm{i}\depth)$ and, since $p$ is not a harmonic function
\citep{con2011,CS2010,hen11}, the flat bed is the only location where $\mathfrak p$ coincides with
$ \operatorname{Re}\mathfrak{P}$. It can be shown (cf. \citep{clam13,CC}) that
\begin{equation} \label{cde}
g\,\eta\,(1\/-\/\mathrm{i}\/\eta_x)\ +\ \mathrm{i}\,B\,\eta_x\ =\ (\,\sur{\mathfrak{P}}\,
-\, g\,\depth\,)\,(1\/+\/\mathrm{i}\/\eta_x),
\end{equation}
with $\sur{\mathfrak P}\/=\/\mathfrak{P}(\sur{z})\/=\/\mathfrak{P}(x+\mathrm{i}\eta(x))$.
The real and imaginary parts of (\ref{cde}) give the two equations
\begin{align}
g\,\eta\ &=\ \operatorname{Re}\{\mathfrak{P}_\text{s}\}\ -\ g\,\depth\ -\ \eta_x\,
\operatorname{Im}\{\mathfrak{P}_\text{s}\}\ =\ {\textstyle{1\over2}}\,(\,B\,-\,u_\text{s}^{\,2}\,
-\,v_\text{s}^{\,2}\,), \label{e1} \\
(B\, -\, g\,\eta)\,\eta_x\ &=\ (\,\operatorname{Re}\{\mathfrak{P}_\text{s}\}\,
-\,g\,\depth\,)\,\eta_x\ +\ \operatorname{Im}\{\mathfrak{P}_\text{s}\}\ =\
{\textstyle{1\over2}}\,(\,B\,+\,u_\text{s}^{\,2}\,+\,v_\text{s}^{\,2}\,)\,\eta_x.\label{e2}
\end{align}
where the second equalities are obtained using (\ref{P}) and $\sur{v}\/=\/\eta_x\/\sur{u}$.
From the viewpoint of determining the wave amplitude $a$, for regular Stokes waves we have
$\eta_x(0)=0$ at the crest $z=\mathrm{i}\/\eta(0)=\mathrm{i}\/a$, and so \eqref{e1} reduces to a simple
implicit equation which can be iteratively solved to find the amplitude $a$
as solution of the equation
\begin{equation}\label{defamp}
a\ =\ g^{-1}\,\Re\{\mathfrak{P}(\mathrm{i} a)\}\ -\ d.
\end{equation}
The convergence of fixed point iterations of \eqref{defamp} was studied in \citep{CC}
considering the functional iterations (convergent for all regular waves)
\begin{equation}\label{defiteramp}
y\, =\, \hat{F}(y)\, \stackrel{\text{\tiny{def}}}{=}\, y\, +\, g^{-1}\,p(0,y)\,
=\, g^{-1}\,\Re\{\mathfrak{P}(\mathrm{i} y)\}\, -\, d\,
=\, {\textstyle{1\over2}}\,B\,g^{-1}\, -\, {\textstyle{1\over2}}\,g^{-1}\,\Re\!\left\{w(\mathrm{i} y)^2\right\}.
\end{equation}
With the wave amplitude $a$ so determined, in \citep{CC} it was shown that $\eta$ can
subsequently be obtained by re-expressing (\ref{e2}) as the ordinary differential equation
\begin{equation}\label{r2}
\eta_x\ =\ \operatorname{Im}\{\mathfrak{P}_\text{s}\}\ /\ \left(\,B\, -\, g\,\eta\,
-\, \operatorname{Re}\{\mathfrak{P}_\text{s}\}\ +\ g\,\depth\,\right),
\end{equation}
with initial data $\eta(0)=a$, which can iteratively be solved. In \cite{CC} it was rigorously proven
that these recovery procedures converge.
An alternative, more direct approach to the wave profile recovery problem was instigated in
\citep{clam13} by considering the holomorphic function $\mathfrak{Q}$ defined by
\begin{align} \label{defQfun}
\mathfrak{Q}(z)\ \stackrel{\text{\tiny{def}}}{=}\ \int_{z_0}^z\left[\,\mathfrak{P}(z')\,-\,g\,\depth\,\right]\mathrm{d}\/z'
\ =\ \int_{z_0}^z{\textstyle{1\over2}}\left[\,B\,-\,w(z')^2\,\right] \mathrm{d}\/z',
\end{align}
where $z_0$ is an arbitrary constant. Taking, as in \citep{clam13}, $z_0=\mathrm{i} a$ at the origin
of the free surface, and integrating along the surface, leads to the expression
\begin{equation} \label{defQfunsurfeta}
{\mathfrak{Q}}_\text{s}(x)\ =\ \int_0^xg\,\eta(x')\,\mathrm{d}\/x'\ +\
\mathrm{i}\left[\,\eta(x)\,-\,a\,\right]\left[\,B\,-\,{\textstyle{1\over2}}\,g\,a\,-\,{\textstyle{1\over2}}\,g\,\eta(x)\,\right].
\end{equation}
The imaginary part of (\ref{defQfunsurfeta}) yields an implicit relation defining $\eta$ of the form
\begin{equation}\label{reletaimp}
\eta\ =\ g^{-1}\left[\,B\, -\, \sqrt{\,(\/B\/-\/g\/a\/)^2\ -\ 2\,g\,
\operatorname{Im}{\mathfrak{Q}}_\text{s}\}\,}\,\right],
\end{equation}
which is a local algebraic (i.e., neither differential nor integral) equation for the free
surface $\eta$, and it is easily shown that (\ref{reletaimp}) is an exact implicit solution
of (\ref{r2}). An explicit expression for $\eta$ can be obtained numerically via fixed point
iterations, whose convergence is determined as follows. Let be the functional iterations
\begin{equation}\label{defitF}
y\ =\ F(y)\ \stackrel{\text{\tiny{def}}}{=}\ g^{-1}\left[\,B\, -\, \sqrt{\,(\/B\/-\/g\/a\/)^2\, -\, 2\,g\,
\operatorname{Im}\{{\mathfrak{Q}}(x+\mathrm{i} y)\}\,}\,\right], \qquad
y\in[-\depth,\eta(x)],
\end{equation}
for a known holomorphic function ${\mathfrak{Q}}$ at a fixed abscissa $x$ and for a given $B$.
Iterations of (\ref{defitF}) converge if we can show that $F$ is a contraction, that is, if
$-1<F_y<1$. The definition of $F$ gives
\begin{align}\label{Fy}
F_y\, =\, \frac{\left[\,\operatorname{Im}\{{\mathfrak{Q}}\}\,\right]_y}
{\sqrt{\,(\/B\/-\/g\/a\/)^2\, -\, 2\,g\,\operatorname{Im}\{{\mathfrak{Q}}\}\/}}\,
=\, \frac{p\, +\, g\,y\, +\, v^2}{B\, -\, g\,y}\, =\,
\frac{p\, +\, g\,y\, +\, v^2}{2\,p\,+\,g\,y\,+\,u^2\,+\,v^2},
\end{align}
where we have used definitions (\ref{P}), (\ref{defQfun}) and the Bernoulli equation (\ref{bernbase}).
The lower bound $F_y>-1$ yields the inequality $p+v^2+B>0$ that is obviously always satisfied.
The upper bound for contraction condition $F_y<1$ yields the inequality $p+u^2>0$
that is always satisfied when \eqref{stag} holds. Therefore, for regular Stokes waves, $F$ is a
contraction everywhere in the bulk and at the free surface and the
iterations (\ref{defitF}) converge to an unique solution.
Note that, since $F_y$ is independent of $a$ (and so independent of the choice of $z_0$ in the
definition of $\mathfrak{Q}$), the convergence proof for the iterations (\ref{defitF}) is valid
wherever $z_0$ is chosen in the fluid.
Furthermore, the analysis above for the wave profile recovery
procedure is valid for all regular waves, periodic or not, with one or more crests and troughs per
period, symmetric or not. It is not contingent on the waves being Stokes waves.
\section{Wave surface recovery: extreme waves}\label{secew}
The above considerations may breakdown for the extreme waves, where \eqref{stag}
fails to hold at the stagnation point at a wavecrest. In this case the wavecrest located at
$(0,\eta(0))=(0,a)$ takes an angular form with containing angle $2\pi/3$, with an asymptotic description
provided by \eqref{etamaxloc}, and so in particular $\eta_x(0^\pm)\/\neq\/0$. The analysis of the
previous section (see also \citep{clam13,CC}) then needs to be refined.
\subsection{Crest determination}
The analysis of the functional iterations (\ref{defitF}) assumes that the constants
$a$ and $B$ are known. This is fine for a mathematical proof, but not so for practical applications
because $a$ and $B$ are generally unknown {\em a priori}. Since the crest of an extreme wave is a
stagnation point where $u=v=p=0$, it follows directly from \eqref{bernbase} that $B\/=\/2\/g\/a$,
so only $a$ remains to be determined before studying the iterations \eqref{defitF}.
The convergence of fixed-point iterations of \eqref{defamp} determining the wave amplitude for extreme
waves, where $\eta_x\neq0$ at the crest, was not addressed in \citep{CC}, however we remark that formula
\eqref{defamp} is still applicable in this regime since $ \operatorname{Im}\{\mathfrak{P}_\text{s}\}=0$
along the crest-line (where $v=v_y=0$): fixed-point iterations determined by relation \eqref{defiteramp}
converge if $-1<\hat{F}_y<1$. At the crest of an extreme wave, employing the asymptotic relation
\eqref{prescrest} (which is valid in a neighbourhood of the crest) we can compute $\hat{F}_y=1+g^{-1}
p_y(0,y)=1/2$, so it follows that the fixed point iterations defined by \eqref{defiteramp} converge. \\
\subsection{Remarks}
{\it i}- On the crest line of a Stokes wave the inequality $-g\leqslant p_y<0$ holds: although the
second inequality
is physically quite intuitive, the rigorous proof of this fact for nonlinear waves relies on a fine technical
analysis of harmonic functions for regular waves \citep{CS2010,hen11}, with further complications needing to
be addressed in the extreme wave setting \cite{lyons16}. These bounds are sharp with respect to the analytical
methods being employed. In this context, it is interesting that the asymptotic approximation \eqref{prescrest}
yields that $p_y=-g/2<0$ at the crest of an extreme wave: one might be tempted to conjecture as to whether
the inequality $-g\leqslant p_y\leqslant-g/2$ holds on the crest line of extreme waves and, perhaps, of
regular Stokes waves which converge to the extreme wave by the standard limiting process \cite{const12,tol}.
As a caveat to the latter postulation, we note that many physical properties of the extreme Stokes wave are not
obviously transparent; for instance, although it is the Stokes wave of greatest height, it is not the most
energetic, fastest, or impulsive of the Stokes waves \citep{SF}.
{\it ii}- A further consequence of the relation $p_y=-g/2$ at the crest of an extreme wave is that, since
the relation $u\partial_x\left[ v(x,\eta(x)) \right]=-p_y-g$ holds along the wave profile $y=\eta(x)$ for
all Stokes waves, we must have $\partial_x\left[ v(x,\eta(x)) \right]$ blowing up as $x\to 0$ since $u$
vanishes at the wave crest. It follows that $u_y=v_x$ must also blow-up the wave-crest: this can also be
seen directly from setting $-p_y-g=-g/2$ in the second component of the Euler equation.
{\it iii}- A useful relation involving the wave profile, particularly in the context of Stokes extreme wave,
is obtained by multiplying (\ref{e2}) by $\eta_x$ and adding (\ref{e1}), from which one gets (after
some elementary algebra):
\begin{align}
g\,\eta\ &=\ \operatorname{Re}\{\mathfrak{P}_\text{s}\}\ -\ g\,\depth\ -\
(B\, -\,2\,g\,\eta)\,\eta_x^{\,2}\,\left(\,1+\,\eta_x^{\,2}\,\right)^{-1}.\label{e5}
\end{align}
At the crest of an extreme wave we have $\eta_x\neq0$ is finite, since the angular crest is not a cusp,
and $B-2g\eta=0$, and so \eqref{e5} reduces to relation \eqref{defamp} at the crest.
\subsection{Surface determination}
The convergence of the functional iterations (\ref{defitF}) is proven above (see \citep{clam13}
for more details) everywhere except at an angular crest where $F_y(a)=1$: this corresponds to a
so-called {\em positively neutral\/} fixed point \citep{bair97}. To determine if the iterations
converge where $F_y=1$, we can consider the second derivative of $F$:
\begin{equation}\label{defFyy}
F_{yy}\ =\ \frac{g\,F_y\,+\,p_y\, +\, g\, +\,2\,v\,v_y}{2\,p\,+\,g\,y\,+\,u^2\,+\,v^2}\
=\ \frac{g\,F_y\,+\,v\,v_y\,-\,u\,v_x}{B\,-\,g\,y}\
=\ \frac{g\,F_y\,+\,\operatorname{Im}\{\/w\,w_z\/\}}{B\,-\,g\,y}.
\end{equation}
In order to estimate $F_{yy}$ at angular crests, we use the expansions \eqref{etamaxloc} and
\eqref{w2maxloc}, to find that $F_{yy}=3/2a>0$ at the angular crests. Hence, the fixed point
$y=a$ is monotonously semi-stable from below \citep{bair97}, which means that the iterations are convergent
when the fixed point $y=a$ is approached from the fluid side $y<a$, and divergent when approached
from above the surface $y>a$.
For practical applications, this results suggests that basic fixed-point iterations of
\eqref{reletaimp} convergence very slowly in the vicinity of an angular crest.
More sophisticated iterations (such as Newton and Levenberg--Marquardt methods) should then
be used instead. It is not our purpose here to investigate these possibilities, however.
\subsection{Trough determination}
At the wave trough, where $\eta=-b\leqslant0$ and so $a+b$ is the total wave height, for all Stokes waves
including the extreme wave, we have $\eta_x(\pm \pi/k)=0$. Equation \eqref{e1} then reduces to the
implicit equation
\begin{equation} \label{r1t}
b\ =\ \depth\ -\ g^{-1}\,\operatorname{Re}\{\mathfrak{P}(-\mathrm{i}\/b)\}.
\end{equation}
In \citep{clam13}, presented numerical evidence suggests that iterations of
(\ref{r1t}) converge to $b$, however a mathematically rigorous proof of this result is still lacking.
Indeed, as shown in \citep{clam13}, convergence is guaranteed if $-2g<p_y<0$ on the trough line
$y\in[-d,-b]$. The upper bound is ensured by the inequality $p_y\leqslant-g$ on the trough line, which
can be rigorously established for nonlinear Stokes waves of all amplitudes \cite{con06,CS2010,hen11,lyons16}
by way of maximum principles, using the fact that the pressure distribution is a super-harmonic function
with respect to the physical variables, cf. \cite{con2011}. However, a rigorous proof of the lower bound
for $p_y$ is currently lacking, being unachievable by this approach. This is not a problem for extreme
waves since $b$ can be computed without relying on (\ref{r1t}). Indeed, $a$ being given by \eqref{defamp},
the Bernoulli constant is $B=2ga$ and $b$ is obtained directly by computing \eqref{reletaimp} at a trough.
\section{Conclusion}\label{secconclu}
In this note, we examined the efficacy of exact relations for the recovery of nonlinear free surface
wave profiles from pressure measurements at the bed, derived in \citep{clam13,CC}, in particular
their applicability to the reconstruction of the extreme Stokes wave. The extreme Stokes wave possesses
intrinsic mathematical complications even beyond the level exhibited by regular nonlinear waves, due to
the presence of a stagnation point at the wave-crest. There is a breakdown in regularity at the crest of
an extreme Stokes wave, whereby the profile is no-longer differentiable, but merely continuous:
the sharp wave-crest has an inner angle $2\pi/3$.
The relations introduced in \citep{clam13} employ a relatively simple local formulation without derivatives
and integrations, which allows for great scope in their practical application. The directness of the
surface reconstruction approach initiated in \citep{clam13} is due to the fact that physical variables,
and the physical plane, are used throughout: intermediate conformal transformations are not invoked.
The holomorphic function $\mathfrak{Q}$ allows a reformulation of the problem into a simplified, but
still exact, form.
The convergence of the involved fixed-point iterations, required for obtaining the surface explicitly,
was examined for the case of limiting extreme waves. A local analysis suggests (semi) convergence at
angular crests, thereby offering an insight into the applicability of the reconstruction method in
such an extreme case. Practical reconstructions of regular waves are demonstrated
in \citep{clam13,CC} via numerical examples, so they do not need to be reproduced here. An illustration
for the extreme wave would require data (experimental or numerical) for the pressure at the seabed,
that are not available. Beside the possible reconstruction of the extreme wave, one of our goals here
is to illustrate the usefulness of working in the physical plane, since we could gain theoretical
insights with only elementary mathematical tools, in addition to a practically useful method.
It is hoped that the demonstrated robustness of this wave surface reconstruction approach will prove
useful in treating other water wave problems, for instance the recovery of fully nonlinear rotational
water wave profiles from pressure measurements, an objective which has hitherto proven unattainable.
This will be the subject of future investigations.
\section*{Acknowledgements}
D. Clamond acknowledges the support of the Universit\'e C\^ote d'Azur under the research grant
{\em Paul Montel Chair\/} 2019.
D. Henry acknowledges the support of the Science Foundation Ireland (SFI) under the research
grant 13/CDA/2117.
\addcontentsline{toc}{section}{References}
\bibliographystyle{abbrvnat}
| 1df0667d96a3626cdd690db917658b2b6fcece66 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Acknowledgements}
We would like to thank Alex Acero, Anmol Walia, Arjun Rangarajan, Barry-John Theobald, Bhagyashree Shekawat, Christopher Klein, Dhivya Piraviperumal, Hong Yu, Jianpeng Cheng, Jiarui Lu, John Giannandrea, John Keesling, Katy Linsky, Lanqing Wang, Ryan Templin, Robert Daland, Shruti Bhargava, Xi Chen and Yvonne Xiao for discussions and for their help and feedback. We also want to thank all the anonymous reviewers for the helpful feedback and suggestions on this work.
\bibliographystyle{acl_natbib}
\section{Architecture Design}
\label{sec:method}
For a given utterance, we first detect and label entities
using the NER model and generate the top-$l$ candidate hypotheses using beam search.
The EL model consists of two stages: (i) candidate retrieval and (ii) joint linking and re-ranking.
In the retrieval stage, for each NER hypothesis, we construct a structured search query
and retrieve the top-$c$ candidates from the retrieval engine.
In the ranking stage, we use a neural network to rank these
candidate entity links within each NER hypothesis while simultaneously using rich signals (entity popularity,
similarity between entity embeddings, the relation across multiple entities in one utterance, etc.) from these entity links as additional features to re-rank the
NER hypotheses from the previous step, thus jointly addressing both the NER and EL tasks.
\input{sections/s3.1-ner.tex}
\input{sections/s3.2-el.tex}
\input{sections/s3.3-application.tex}
\subsection{Improvement on Other Language Understanding Tasks}
\label{subsec:method_application}
We also explore the impact of our NEU feature encoding on two tasks: a domain classifier and a domain-specific shallow semantic parser.
\subsubsection{Domain Classification}
\label{subsubsec:method_intent}
Domain classification identifies which domain a user's request falls into: sports, weather, music, etc., and is usually done by posing the task as sequence classification: our baseline uses word embeddings and gazetteer features as inputs to an RNN, in a manner similar to \newcite{chen2019active}.
Consider a specific token $t$. Let $a$ be the number of alternatives used from the NER model in the domain classifier
(which we treat as a hyperparameter), $p_{i}$ represent the (scalar) sequence level confidence score $p_{seq} ( \mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf \theta )$ of the
$i$\textsuperscript{th} NER alternative defined in Section~\ref{sec:ner}, $c_i$ represent an integer for the entity type that NER hypothesis $i$
assigns to the token $t$, and $\mathbf{o}(.)$ represent a function converting an integer into its corresponding one-hot
vector. Then the additional NER feature vector $\mathbf{f_{r}}$ concatenated to the input vector fed into token $t$ as part of the domain classifier can be written as:
\begin{equation} \label{eq:dc_ner}
\mathbf{f_{r}} = \bigoplus_{i=1}^{i=a} p_i\mathbf{o}(c_i).
\end{equation}
Likewise, for the featurization that uses both NER and EL, let $a$ be the number of alternatives used from the NER+EL system in the domain classifier (also a hyperparameter); these $a$ alternatives are now sorted by the scores from the EL hypotheses, as opposed to the sequence level confidence scores from NER. Let $s_i$ be the $i$\textsuperscript{th} re-ranked alternative's cosine similarity score between the mention and knowledge base entity name as output by the EL model. $p_i$ and $c_i$ are consistent with our earlier notation, except that they now correspond to the $i$\textsuperscript{th} NER alternative after re-ranking. Then the additional NER+EL feature vector $\mathbf{f_{u}}$ concatenated to the input fed into token $t$ as part of the domain classifier can be written as:
\begin{equation} \label{eq:dc_rer}
\mathbf{f_{u}} = \bigoplus_{i=1}^{i=a} p_i\mathbf{o}(c_i) \oplus s_i\mathbf{o}(c_i).
\end{equation}
\subsubsection{Semantic Parsing}
\label{subsubsec:method_parse}
Our virtual assistant also uses domain-specific shallow semantic parsers, running after domain classification, responsible both for identifying the correct intent that the user expects (such as the ``play'' intent associated with a song) and for assigning semantic labels to each of the tokens in a user's utterance (such as the word ``score" and ``game'' respectively being tagged as tokens related to a sports statistic and sports event respectively in the utterance ``What's the score of yesterday's Warriors game?''). Each semantic parser is structured as a multi-task sequence classification (for the intent) and sequence tagging (for the token-level semantic labelling) task, with our production baseline using word embeddings and gazetteer features as inputs into an RNN similar to our domain classifier.
Here, $\mathbf{f_{r}}$ and $\mathbf{f_{u}}$ are featurized as described above.
Note that in contrast to the NEU system, the semantic parser uses a domain-specific ontology, to enable each domain to work independently and to not be encumbered by the need to align ontologies.
\subsection{NER}
\label{sec:ner}
For the NER task, following \newcite{lample-etal-2016-neural}
we use a combination of character and word level features.
They are extracted by a bi-directional LSTM (biLSTM) \cite{hochreiter1997long}, and then concatenated with pre-trained GloVe word embeddings \footnote{We also tried more recent contextual
embeddings such as BERT \cite{devlin2019bert}, and empirically observed very little difference in performance when compared to GloVe. So we adopt GloVE, which is substantially more efficient in terms of inference time required.}
\cite{pennington-etal-2014-glove} to pass through another biLSTM and fed into a CRF model to produce
the final label prediction based on a score $s( \mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf{\theta} )$ that jointly optimizes the
probability of labels for the tokens and the transition score for the entire sequence
$\mathbf{\tilde{y_{i}}}=(y_1,\dots,y_T)$ given the input $\mathbf{x}$:
\begin{equation*}
s( \mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf{\theta} ) = \sum_{t=0}^{T} \left( \psi_{t, \mathbf \theta}(y_t) +
\phi_{t,t+1} (y_t, y_{t+1}) \right),
\end{equation*}
\noindent where $\psi_{t,\theta}$ is the biLSTM prediction score from the label $y_t$ of the $t$\textsuperscript{th} token, and $\phi(j, k)$ is the transition score
from label $j$ to label $k$.
During training, we maximize the probability of the correct label sequence $p_{seq}$, which is defined as
\begin{equation*}
p_{seq} ( \mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf \theta ) = \frac{\exp({s( \mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf \theta )) }}{\sum_{\mathbf{\tilde{y_{j}}} \in S} \exp{(s( \mathbf{\tilde{y_{j}}}, \mathbf{x}; \mathbf \theta ))}},
\end{equation*}
\noindent where $\mathbf{\tilde{y_{i}}}$ is the label sequence for hypothesis $i$, and $S$ is the set of all possible label sequences.
During inference, we generate up to 5 NER alternatives for each utterance using beam search.
We also calculate a mention level confidence $p_{men}$ for each
entity mention $ \mathbf{m_k}$. $p_{men}$ is computed by aggregating the sequence level confidence for all the prediction sequences that
share the same mention sub-path $\mathbf{m_k}$:
\begin{equation*}
p_{men} (\mathbf{m_k}, \mathbf x ; \mathbf \theta) = \frac{ \sum_{ \mathbf{\tilde{y_{i}}} \in S_{\mathbf{m_i}}} \exp({s (\mathbf{\tilde{y_{i}}}, \mathbf{x}; \mathbf \theta )}) }{\sum_{ \mathbf{\tilde{y_j}} \in S} \exp(s (\mathbf{\tilde{y_{j}}}, \mathbf{x}; \mathbf \theta ))},
\end{equation*}
\noindent where ${S_{m_i}}$ is the set of prediction sequences that all have $ \mathbf{m_k}$ as the prediction for the corresponding tokens.
Both $p_{seq}$ and $p_{men}$ are computed by dynamic programming, and serve as informative features in the EL model.
\subsection{Qualitative Analysis}
\label{sec:qualitative}
We provide a few examples to showcase the effectiveness of our NEU system.
Firstly, the EL model is able to link noisy entity mentions to the corresponding entity canonical name in the knowledge
base. For instance, when the transcribed utterance is ``play Carla Cabello'', the EL model is able to resolve the
mention ``Carla Carbello'' to the correct artist name ``Camila Cabello''.
Secondly, the EL model is able to recover from errors made by the NER system by leveraging the knowledge base to disambiguate entity mentions. The reranking is especially powerful when the utterance contains little context of the entity for the NER model to leverage. For example, for ``Doctor Strange'', the top NER hypothesis labels the full utterance as a generic ``Person'' type, and after reranking, EL model is able to leverage the popularity information (``Doctor Strange'' is a movie that was recently released and has a high popularity in our knowledge base) and correctly label the utterance as ``movieTitle''.
Reranking is also effective when the entity mentions are noisy, which will cause mismatches for the gazetteer features
that NER uses. For ``play Avengers Age of Ultra'', the top NER hypothesis only predicts ``Avengers'' as ``movieTitle'', while after reranking, the EL model is able to recover the full span ``Avengers Age of Ultra'' as a ``movieTitle'', and resolve it to ``Avengers: Age of Ultron'', the correct canonical title.
The entity relations from the knowledge base are helpful for entity disambiguation. When the user refers to a sports
team with the name ``Giants'', they could be asking for either ``New York Giants'', a National Football League (NFL) team, or ``San Francisco Giants'', a Major League Baseball team. When there are multiple sports team mentions in an utterance, the EL model leverages a relation feature from the knowledge base indicating whether the teams are from the same sports league (as the user is more likely to mention two teams from the same league and the same sport). Knowing entity relations, the EL model is able to link the mention ``Giants'' in ``Cowboys versus Giants'' to the NFL team, knowing that ``Cowboys'' is referring to ``Dallas Cowboys''.
\section{Conclusion}
\label{sec:concl}
In this work, we have proposed a Named Entity Understanding framework that jointly identifies and resolves entities present in an utterance when a user interacts with a voice assistant. Our proposed architecture consists of two modules: NER and EL, with the EL serving the additional task of possibly correcting the recognized entities from NER by leveraging rich signals from entity links in the knowledge base while simultaneously linking these entities to the knowledge base. With several design strategies in our system targeted towards noisy natural language utterances, we have shown that our framework is robust to speech transcription and user errors that occur frequently in spoken dialog systems. We have also shown that featurizing the output of NEU and feeding these features into other language understanding tasks substantially improves the accuracy of these models.
\section{System Evaluation}
\label{sec:expt_res}
\input{sections/s5.1-ner_el.tex}
\input{sections/s5.2-case_study.tex}
\input{sections/s5.3-application.tex}
\section{Introduction}
\label{sec:intro}
Understanding named entities correctly when interacting with virtual assistants (e.g. ``Call Jon'', ``Play Adele
hello'', ``Score for Warrior Kings game'') is crucial for a satisfying user experience. However, NER and EL methods that work well
on written text often perform poorly in such applications:
utterances are relatively short (with just $5$ tokens, on average), so there is not much
context to help disambiguate; speech recognizers make errors (``Play Bohemian raspberry'' for ``Play Bohemian Rhapsody"); users also make
mistakes (``Cristiano Nando'' for ``Cristiano Ronaldo''); non-canonical forms of names are frequent (``Shaq'' for
``Shaquille O'Neal''); and users often mention new entities unknown to the system.
In order to address these issues we propose a novel Named Entity Understanding (NEU) system that combines and optimizes NER and EL for noisy spoken
natural language utterances. We pass multiple NER hypotheses to EL for reranking, enabling NER to benefit from EL by
including information from the knowledge base (KB).
We also design a retrieval engine tuned for spoken utterances for retrieving candidates from the KB. The retrieval engine, along with other techniques devised to address fuzzy entity mentions, lets the EL model be more robust to partial mentions, variation in named entities, use of aliases, as well as human and speech transcription errors.
Finally, we demonstrate that our framework can also empower other natural language understanding tasks, such as domain classification (a sentence classification task) and semantic parsing.
\section{Datasets and Training Methodology}
\label{sec:expt_setup}
To create our datasets, we randomly sampled around $600k$ unique anonymous English transcripts
(machine transcribed utterances), and annotated them with NER and EL labels.
Utterances are subject to Apple’s baseline privacy practices with respect to Siri requests, including that such requests
are not associated with a user’s Apple ID, email address, or other data Apple may have from a user’s use of other Apple
services, and have been filtered as described in Section \ref{sec:ethics}.
We then split the annotated data into 80/10/10 for train, development and test sets.
For both the NER and EL tasks, we report our results on test sets sampled from
the ``music'', ``sports'' and ``movie \& TV'' domains. These are popular domains in the usage and have a high
percentage of named entities: with an average of 0.6, 1.1 and 0.7 entities for each utterance in the 3 domains respectively.
To evaluate model performance specifically on noisy user inputs,
we select queries from the test sets that are marked as
containing speech transcription or user errors by the annotators and report results on this ``noisy" subset, which constitutes 13.5\%, 12.7\% data for movie\&TV and music domain respectively when an entity exists. \footnote{Sports domain does not have the annotation for noisy data available when this experiment was conducted.}
To evaluate the relation feature, we also look at the ``related" subset where a valid relation exists in the utterance. This subset consists 13.4\% and 5.3\% of data for the music and sports domain with at least one entity.
\footnote{Our KB does not have relation information for movie\&TV domain.}
We first train the NER model described in Section \ref{sec:ner}.
Next, for every example in our training dataset, we run inference on the trained NER model and generate the
top-5 NER hypotheses using beam search.
Following this, we retrieve the top 25 candidates for each of these hypotheses using our search engine
combined with the ground truth NER and EL labels and fed to the EL model for training.
To measure the NER model performance, we use the standard NER F1 metric used for the CoNLL-2003 shared task \cite{tjong-kim-sang-de-meulder-2003-introduction}.
To measure the quality of the top-5 NER hypotheses, we compute the oracle top-5 F1 score by comparing and choosing
the best alternative hypothesis among the $5$ and calculate its F1 score, for each test utterance. In this manner, we also know the upper
bound that EL can reach from reranking NER hypotheses.
As described in section \ref{subsec:el}, the EL model is optimized to perform two tasks simultaneously:
entity linking and reranking of NER hypotheses.
Hence to evaluate the performance of the EL model, we use two metrics: reranked NER-F1 score and the
EL-F1 score.
The reranked NER F1 score is measured on the NER predictions according to the top EL hypothesis, and is defined in the same way as the previous NER task.
To evaluate entity linking quality, we adopt a strict F1 metric similar to the one used for NER.
Besides entity boundary and entity type, the resolved entity also needs to be correct for the entity prediction
to be counted as a true positive.
\label{sec:hyperparams}
For NER model training, we use standard mini-batch gradient descent using the Adam optimizer with an initial
learning rate of 0.001, a scheduled learning rate decay of 0.99, LSTM with a hidden
layer of size 350 and a batch size of 256.
We apply a dropout of 0.5 to the embedding and biLSTM layers, and include token level
gazetteer features \cite{ratinov-roth-2009-design} to boost performance in recognizing common entities.
We linearly project these gazetteer features and concatenate the projection with the 200 dimensional word embeddings
and 100 dimensional character embeddings which are then fed into the biLSTM followed by the CRF.
For EL, the character CNN model we use has two layers, each with 100 convolution kernels of size 3, 4, and 5. Character embedding are 25 dimensional and trained end to end with the entity linking task. The MLP for embedding similarity takes the concatenation of two name embeddings, as well as their element-wise sum, difference, minimum, maximum, and multiplication. It has two hidden layers of size 1024 and 256, with output dimension 64. Similarity features of mentions in the prediction are averaged, while the other features like NER confidence and entity popularity are concatenated to the representation. The final MLP for scoring has two hidden layers, with size 256 and 64. We train the model on 4 GPUs with synchronous SGD, and for each gradient step we send a batch of 100 examples to each GPU.
\section{Related Work}
\label{sec:rel_work}
There have been a few attempts to explore NER on the output of a speech pipeline
\cite{ghannay2018end,abujabal2018neural,coucke2018snips}. Among these, our NER model is closest to
\newcite{abujabal2018neural} and \newcite{coucke2018snips}; however, unlike the former, we use a richer set of features rather than phonemes as input, and unlike the latter, we are able to use a deep model because of the large volume of data available.
EL has been well explored in
the context of clean \cite{martins-etal-2019-joint,kolitsas-etal-2018-end,luo2015joint} and noisy text inputs
\cite{eshel-etal-2017-named,guo-etal-2013-link,liu-etal-2013-entity}, but as with NER, there have been only a few
efforts to explore EL in the context of transcribed speech \cite{benton-dredze-2015-entity,gao-2017-support}, although crucially, both these works assume gold standard NER and focus purely on the EL component.
Traditionally, a pipelined architecture of NER followed by EL has been used to address the entity linking task
\cite{lin-etal-2012-entity,derczynski2015analysis,bontcheva2017crowdsourcing,bowden-etal-2018-slugnerds}. Since these approaches rely only on the best NER hypothesis, errors from NER propagate to the EL step.
To alleviate this, joint models have been proposed: \newcite{sil-yates} proposed an NER+EL model which re-ranks candidate mentions and entity links produced by their
base model. Our work differs in that we use a high precision NER system, while they use a large number of heuristically
obtained Noun Phrase (NP) chunks and word n-grams as input to the EL stage. \newcite{luo2015joint} jointly train an NER
and EL system using a probabilistic graphical model. However, these systems are trained and tested on clean text and do
not address the noise problems we are concerned with.
\section{Ethical Considerations}
\label{sec:ethics}
We randomly sampled transcripts from Siri
production datasets over a period of months, and we believe it to be a representative sample
of usage in the domains described. In accordance with Apple’s privacy practices with respect to Siri requests, Siri
utterances are not associated with a user’s Apple ID, email address, or other data Apple may have from a user’s use of
other Apple services. In addition to Siri’s baseline privacy guarantees, we filtered the sampled utterances to remove
transcripts that were too long, contained rare words, or contained references to
contacts before providing the dataset to our annotators.
\subsection{Joint Linking and Re-ranking}
\label{subsec:el}
\begin{table*}[b]
\begin{equation}
s(\mathbf{y}) = \textbf{MLP} (
\mathbf{f}_{\text{utter}} \oplus \mathbf{f}_{\text{NER}} \oplus \mathbf{f}_{\text{CR}}
\bigoplus_{j=1}^{k} [\textbf{MLP}(\textbf{CNN}(\mathbf{m}_j), \textbf{CNN}(\mathbf{e}_j)) \oplus \\ \textbf{CNN}(\mathbf{m}_j) \oplus \textbf{CNN}(\mathbf{e}_j)]
)
\label{eq-nen-score}
\end{equation}
\end{table*}
The entity linking system follows the NER model and consists of two steps: candidate retrieval,
and joint linking and re-ranking.
To build the candidate retrieval engine, we first index the list of entities in our knowledge base, which can be updated daily to capture new entities and change of their popularity.
To construct the index, we iterate through the flattened list of entities and construct
token-level unigram, bigram and trigram terms from the surface form of each entity.
Apart from using the original entity names, we also use common aliases, harvested from usage logs, for popular entities (e.g. LOTR as an alias for ``Lord of the Rings'') to make the retrieval engine more
robust to commonly occurring variations.
Next, we create an inverted index which maps the unique list of n-gram terms to the list of entities that these n-grams are part of, also known as posting lists.
Further, to capture cross-entity relationships in the knowledge base (such as relationships between an artist and a song or two sports teams belonging to the same league),
we assign a pointer\footnote{Each entity in our knowledge base consists of metadata (for example, a song entry in our knowledge base would contain metadata such as the music artist, album, year the song was released in etc.) that we leverage to automatically construct these relationship pointers.} for each entity in the knowledge base to its related entities and this relational information is leveraged by the EL model for entity disambiguation (described in \ref{sec:qualitative}).
We then compute the tf-idf score
for all the n-gram terms present in the entities and store them in the inverted index.
For each hypothesis predicted by the NER model we query the retrieval engine with the corresponding
text.
We first send the query through a high-precision seq-to-seq correction model \cite{schmaltz-etal-2017-adapting,ge-etal-2019-automatic}
trained using common errors observed in usage.
Next, we construct n-gram features from the corrected query in a similar way to the indexing phase and
retrieve all entities matching these n-gram features in our inverted index.
Additionally, we use synonyms derived from usage for each term in the query to expand our
search criteria: for example, our synonym list for ``Friend" contains ``Friends", which matches the TV show name which
would have been missed if only the original term was used.
For each entity retrieved, we get the tf-idf score for the terms present in the query chunk from the
inverted index.
We then aggregate the tf-idf scores of all the terms present in the query for this entity and linearly combine
this aggregate score with other attributes such as popularity (i.e. prior usage probability) of the entity to generate a final
score for all retrieved entity candidates for this query.
Finally, we perform an efficient sort across all the entity candidates based on this score and return a top-$c$
(in our case $c$ = 25) list filtered by the entity type detected by the NER model for that hypothesis.
These entity candidates coupled with the original NER hypothesis are sent to the ranker model described below
for joint linking and re-ranking.
Following the candidate retrieval step,
we introduce a neural model to rerank the candidate entities, aggregating features from both the NER model and the candidate retrieval engine.
The EL model scores each entity linking hypothesis separately. An entity linking hypothesis consists of a prediction from the NER model (which consists of named entity chunks in the input utterance and their types), and the candidate retrieval results for each chunk. Formally, we define an entity linking hypothesis $\mathbf{y}$ with $k$ entity predictions as:
\begin{equation*}
\mathbf{y} = \{ \mathbf{f}_{\text{utter}}, \mathbf{f}_{\text{NER}}, \mathbf{f}_{\text{CR}}, \{j \in \{1 \dots k\}:
( \mathbf{m}_j, \mathbf{e}_j )
\} \}
\end{equation*}
\noindent where $\mathbf{m}_j$ is the $j$-th mention in the utterance, and $\mathbf{e}_j$ is the entity name associated with this mention from the knowledge base. $\mathbf{f}_{\text{utter}}$, $\mathbf{f}_{\text{NER}}$, $\mathbf{f}_{\text{CR}}$ are features derived from the original utterance text, the NER model and the candidate retrieval system respectively. In our system, $\mathbf{f}_{\text{utter}}$ is a representation of the utterance from averaging the pre-trained word embeddings for the tokens in the utterance. Intuitively, having a dense representation of the full utterance can help the EL model better leverage signals from the utterance context. $\mathbf{f}_{\text{NER}}$ includes the type of each mention, as well as the sequence and mention confidence computed by the NER model. $\mathbf{f}_{\text{CR}}$ includes popularity, and whether a relation exists between the retrieved entities in $\mathbf{y}$.
To be robust to noise, the EL model adopts a pair of CNNs to compare each entity mention $\mathbf{m}_j$ and its
corresponding knowledge base entity name $\mathbf{e}_j$. The CNN learns a name embedding with one-dimensional
convolution on the character sequence, and the kernel parameters are shared between the CNN used for user mention and the one used for the canonical name. A character-based text representation model is better at handling mis-transcriptions or mis-pronounced entity names. While a noisy entity name may be far from the canonical name in the word embedding space when they are semantically different, they are usually close to each other in the character embedding space due to similar spellings.
To model the similarity between CNN name embeddings of $\mathbf{m}_j$ and $\mathbf{e}_j$, we use the standard cosine
similarity as a baseline, we experiment with an MLP that takes the concatenated name embeddings as input.
We are able to model more expressive interactions between the two name embeddings with the MLP, and in turn better handle errors.
Finally, we concatenate the similarity features with other features as input to another MLP that computes
the final score for $\mathbf{y}$. Formally, the scoring function is defined in Equation \ref{eq-nen-score}, where $\oplus$ means concatenation.
In our data, the number of entity mentions in an utterance averages less than 3. We pad the entity feature sequence
to length 5, which provides a good coverage. In the scoring model above, we use a simple concatenation to
aggregate the embedding similarities of multiple entity mentions which empirically performs as well as sequence models like
LSTM, while being much cheaper in computation.
To train the EL model, we use the standard max-margin loss for ranking tasks. If for the $i$-th example, we denote the ground truth as $\mathbf{y^*}_i$ and an incorrect prediction as $\mathbf{\hat{y}}_i$, and the scoring function $s(\cdot)$ is as defined in Equation \ref{eq-nen-score}, the loss function is
\begin{equation}
\mathcal{L} = \frac{1}{N}\sum_{i = 1}^{N} [\gamma(\mathbf{\hat{y}}_i, \mathbf{y^*}_i) + s(\mathbf{\hat{y}}_i) -
s(\mathbf{y^*}_i)]_+.
\label{eq-nen-loss}
\end{equation}
The max-margin loss
encourages the ground truth score to be at least a margin $\gamma$ higher than the score of an incorrect
prediction. The margin is defined as a function of the ground truth and the incorrect prediction, thus adaptive to the
quality of prediction. A larger margin is needed when the incorrect prediction is further away from the ground
truth. For our reranking task, we set a smaller margin when only the resolved entities are incorrect but the NER result
is correct, and a larger margin when the NER result is wrong. This adaptive margin helps rerank NER hypotheses even when the model cannot rank the linking results correctly. During training, we uniformly
sample the negative predictions from the candidates retrieved by the retrieval engine.
\section{Architecture Diagram}
\label{app:arch_diag}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{figure/NEU_arch_diagram.pdf}
\caption{NEU system architecture.}
\label{fig:neu_architecture}
\end{figure}
\subsection{Results}
\label{subsec:el-results}
We present F1 scores in different domains of the NER and EL model in Table \ref{tb:nen-best}. Since the EL model takes 5 NER hypotheses as input, it also acts as a re-ranker of the NER model, and we show substantial improvements on top-1 NER F1 score consistently over all test sets.
\begin{table}[h]
\begin{center}
\begin{tabular}{|@{\hskip3pt}l@{\hskip3pt}|r|c|l@{\hskip2pt}|}
\hline
& \multicolumn{1}{c|}{\textbf{\begin{tabular}[c|]{@{}c@{}}NER F1\\ top-1/top-5\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c|]{@{}c@{}}reranked\\ NER F1\end{tabular}}} & \multicolumn{1}{@{\hskip4pt}l@{\hskip2pt}|}{\textbf{EL F1}} \\ \hline
\textbf{movie\&TV} & 78.76 / 96.83 & 81.62 & 79.67 \\
\textbf{music} & 84.27 / 97.26 & 87.40 & 84.95 \\
\textbf{sports} & 92.97 / 99.15 & 93.48 & 91.13 \\ \hline
\end{tabular}
\caption{Results for the best model setting. NER F1 are reported on the top-1 and top-5 NER prediction from the NER model that provides features for EL. Reranked NER F1 and EL F1 are reported on top-1 prediction from the best EL model selected by development sets.}
\label{tb:nen-best}
\end{center}
\end{table}
\begin{table*}[!htb]
\begin{minipage}{.265\linewidth}
\centering
\begin{tabular}{|l|r|}
\hline
\hline
\multicolumn{1}{|c|}{\textbf{}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}{\bf (a)}\\+ MLP \end{tabular}}} \\ \hline
\textbf{movie\&TV} & +3.58 \\
\textbf{(noisy)} & +9.67 \\ \hline
\textbf{music} & +2.05 \\
\textbf{(noisy)} & +10.03 \\ \hline
\end{tabular}
\end{minipage}%
\begin{minipage}{.27\linewidth}
\centering
\begin{tabular}{|l|r|}
\hline
& \textbf{ (b)\, \, \, \,} \\
& \textbf{+ Relation} \\ \hline
\textbf{music} & +0.86 \\
\textbf{(related)} & +1.97 \\ \hline
\textbf{sports} & +0.07 \\
\textbf{(related)} & +0.81 \\ \hline
\end{tabular}
\end{minipage} %
\begin{minipage}{.4\linewidth}
\centering
\begin{tabular}{|l|r|r|}
\hline
\multicolumn{1}{|c|}{\textbf{}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}{\bf (c)}\\+ Utterance \\ Embedding\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}{\bf (d)}\\+ Log-scale \\ Popularity\end{tabular}}} \\ \hline
\textbf{movie\&TV} & +0.25 & +0.27 \\
\textbf{music} & +0.39 & +0.02 \\
\textbf{sports} & -0.07 & +0.08 \\ \hline
\end{tabular}
\end{minipage} %
\caption{EL mean F1 relative \% improvements, reported on 10 runs average.}
\label{tb:nen-improve}
\end{table*}
In Table \ref{tb:nen-improve}, we show improvements achieved by several specific model design choices and features on
entity linking performance. Table~\ref{tb:nen-improve}(a) shows the MLP similarity substantially improves entity linking accuracy with its capacity to
model text variations, especially on utterances with noisy entity mentions. The relation feature is powerful for disambiguating entities with similar names, and we show a considerable improvement in EL F1 on the subset of utterances that have related entities in Table~\ref{tb:nen-improve}(b). Table~\ref{tb:nen-improve}(c) shows utterance embeddings brought
improvements in the music, and media \& TV domains. The improvement brought by log-scale popularity feature is the largest for the movie \& TV domain as shown in Table~\ref{tb:nen-improve}(d), where the popularity distribution has extremely long tails compared to other
domains.
| b510358ae5d869139468a8f96a6e88f79e0fb36d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Functional registration or time-warping, refers to the process of aligning two or more functions or curves in time and is often a pre-processing step necessary for appropriately analyzing such functions.
By registering functions before performing statistical analysis on such functions, we can account for arbitrary reparameterizations of functions.
This can lead to incorrect analyses of many functional data applications, such as the well-known growth rate curve analysis.
Function registration is heavily relied upon in image and shape registration where analysis results can vastly differ depending on whether the data has been properly registered or not.
More details on the importance of functional data registration can be found in \cite{FSDA, marron2015}, and \cite{ramsay2005}.
Functional data registration is also referred to as phase-amplitude separation, because the underlying goal of the procedure is to effectively distinguish phase ($x$-axis) from amplitude ($y$-axis) variation. For example, in tracking the migration paths of birds or hurricanes, functional data registration would allow one to isolate the statistical variation in the paths, from the statistical variation in the speed of traversing the paths.
The registration of a pair of functions results in an optimal warping function, which enables one function to be aligned to the other. The resulting aligned functions characterize the amplitude variability, while the warping function captures the phase variability between the two functions.
Early approaches to functional data alignment fall short in four major ways.
(1) They lack the ability to characterize the uncertainty of the optimal warping function.
(2) They assume the observed functions are naturally smooth and lack measurement error.
(3) They do not consider more than one optimal alignment between a pair of functions.
Finally, (4) Most methods do not use a proper distance as the objective function for registration.
Traditional approaches to functional data alignment, which are demonstrated in \cite{srivastava-etal-JASA:2011,ramsay2005, sangalli-etal2:2010}, and \cite{kneip-ramsay:2008}, are not flexible enough to characterize the uncertainty of the resulting optimal alignment solution.
These approaches make the common assumption that the functions to be aligned are smooth, and thus they break down when a function exhibits too much noise.
These methods usually rely on a derivative which is extremely noisy when the original functions are noisy, presenting additional computational challenges.
Recently, Bayesian frameworks have been proposed that can characterize the uncertainty of the warping function solution \citep{telesca2008,kurtek2017,lu2017, cheng2016}.
The clear difference between these approaches is in how they specify a prior on the space of the warping function; as it lives on a nonlinear infinite dimensional manifold.
The more recent Bayesian approaches by \cite{cheng2016, lu2017}, and \cite{kurtek2017} rely on the square root velocity function (SRVF) representation of the warping function space, introduced by \cite{srivastava2011} and \cite{srivastava-etal-JASA:2011}, to simplify the complicated geometry. We will also take advantage of this transformation and provide more details in Section \ref{sec:review}. Although these approaches demonstrate how a Bayesian framework can naturally handle uncertainty quantification of the warping function through posterior inference, few have account for measurement error in the observed data and none allow for a multimodal posterior, i.e. multiple optimal warpings.
Only the most recently proposed approach \cite{matuk2019}, has addressed the need to account for observed measurement error.
There approach is for sparsely sampled functions where they propose a data-driven prior for inference and solely look at sparsely sampled functions and do not treat the dense problem.
Previous approaches, Bayesian or otherwise, have assumed the observed functions to be aligned are naturally smooth. Although smoothness is a convenient assumption, it can be an invalid one in many practical applications.
Recent work \cite{panaretos:2016}, utilized a fPCA construction to avoid overfitting when the data has additive amplitude variation.
Wrobel \cite{wrobel:2019}, recently extended the alignment to non-Gaussian data. Both of these methods do not address the estimation of the additive noise process that we will address in this paper.
Lastly, no current approaches consider the challenging case of multiple optimal alignments between a pair of functions; or briefly mention a possible solution without application.
In this paper, we propose a hierarchical Bayesian approach to the registration of a pair of noisy functions on $\mathbb{R}^1$ and demonstrate its advantages over previously proposed registration algorithms.
We propose a new framework for functional alignment which relies on a proper distance metric, is robust to noise, capable of characterizing the uncertainty of the warping function, and can account for multiple optimal alignments.
The novelty of our new framework is the following:
Most notably, we employ parallel MCMC chains to capture multi-modal posterior distributions of our warping functions; to reflect the case when there are multiple optimal alignments.
Additionally, we favor the geometric Hamiltonian Monte Carlo algorithm of \cite{beskos2017} ($\infty$-HMC), which is nontrivial to apply to problem of functional data alignment.
The implementation of $\infty$-HMC within our MCMC sampler can appropriately account for the complex geometry of the target space and more efficiently sample the warping function.
Lastly, our approach accounts for appropriate propagates measurement error, a challenge which has only recently been addressed by \cite{matuk2019}.
This paper is organized in the following way. Section \ref{sec:review} reviews pairwise functional registration in $\mathbb{R}^1$, the general challenges associated with it, and introduces the square-root velocity function representation and proper distance metric relied upon throughout this paper.
Section \ref{sec:pairwise} specifies our hierarchical Bayesian model for registering a pair of noisy functions on $\mathbb{R}^1$, including our MCMC algorithm and multi-chain approach to the challenge of multiple alignments.
In Section \ref{sec:examples}, we evaluate the approach on simulated functional data as well as two real data sets: a SONAR dataset that is naturally noisy and iPhone-collected accelerometer data with multiple possible alignments.
Code for the method is found in the \verb+fdasrvf+ Matlab package on GitHub\footnote{\url{https://github.com/jdtuck/fdasrvf_MATLAB}}.
Finally, we discuss the impact of this approach, extensions to multiple pairwise alignments, nontrivial extensions to functions on more complex geometries, and future work in Section \ref{sec:disc}.
\section{A review of function registration in $\mathbb{R}^1$}\label{sec:review}
Following the notation of \cite{tucker2013} and without loss of generality, let $f$ be a real-valued and absolutely continuous function on domain $[0, 1]$.
Note that in practice, $f$ is observed as discrete and interpolation can be used to more easily perform the requisite calculations.
Let $\mathbb{F}$ denote the set of all such functions and $\Gamma$ denote the set of boundary-preserving diffeomorphisms: $\Gamma = \{\gamma : [0,1]\mapsto [0,1] \mid \gamma(0)=0,\gamma(1)=1\}$ such that the mapping $ [0,1]\mapsto [0,1]$ is bijective (invertible) and differentiable.
From this point on, we will refer to $\Gamma$ as the set of warping functions.
Then, for any $f\in \mathbb{F}$ and any $\gamma \in \Gamma,$ $f \circ \gamma$ denotes the time-warping of $f$ by $\Gamma$.
For simplicity, consider the pairwise alignment problem where we wish to align functions $f_1,f_2 \in \mathbb{F}.$ A simplified registration problem can be formulated by finding a warping, $\gamma^*,$ that minimizes the cost of translating $f_2$ to $f_1$, for chosen cost function
$$\gamma^* = \text{argmin}_{\gamma} \| f_1 - f_2 \circ \gamma\|^*, $$
where $\| \|^*$ is our distance metric yet to be chosen.\footnote{Although we are proposing a more robust approach, we are introducing a simplified registration problem to motivate the use of a proper distance metric.}
There are a few main challenges that all alignment approaches face and handle differently.
One challenge of function alignment, as pointed out by \cite{srivastava-etal-JASA:2011} and \cite{tucker2013}, is finding a cost function to do the alignment that is both symmetric, that is, aligning $f_1$ to $f_2$ is the same as aligning $f_2$ to $f_1$, and positive-definite, so that the metric is always non-negative and zero if and only if $f_1$ and $f_2$ are the same function after alignment.
A natural choice is the usual $\mathbb{L}^2$ norm, denoted by $\|\cdot\|^2$, but it does not satisfy the symmetry requirement.
More precisely,
$$\text{argmin}_{\gamma}\|f_1 \circ \gamma - f_2 \|^2 \neq \text{argmin}_{\gamma}\|f_1 - f_2 \circ \gamma \|^2. $$
Second, there can be the issue of degeneracy in which case $\gamma^*$ is so distorted it can align functions which should not be aligned; giving a false impression that these functions are \textit{close} in the $\mathbb{L}^2$ sense when indeed they are not.
Lastly, the $\mathbb{L}^2$ norm is not invariant under warping. That is,
$$\| f_1 - f_2 \|^2 \neq \| f_1 \circ \gamma - f_2 \circ \gamma\|^2.$$
This means that two functions which are similar in an $\mathbb{L}^2$ sense, could be vastly different under the same warping, and vice versa.
To overcome these three challenges, we follow \cite{tucker2013}, \cite{cheng2016} and \cite{lu2017} and take advantage of the square-root velocity function (SRVF) representation introduced by \cite{srivastava2011}.
A function $f\in \mathbb{F}$ can be represented as a SRVF via the mapping:
$$q:[0,1]\mapsto \mathbb{R},\ \ q(f(t))=\hbox{sign}(\dot{f}(t))\sqrt{|\dot{f}(t)|} \hbox{ for any } f\in \mathbb{F}.$$
There is an equivalency, up to a constant, between the SRVF of $f$ and $f$ itself and it is given by $f(t) = f(0) + \int_0^t q(f(t))|q(f(t))| ds$, where $|\cdot|$ is the absolute value. Moreover, the SRVF of $f \circ \gamma$ is $(q \circ \gamma)\sqrt{\dot{\gamma}}$.
The SRVF transformation lends itself naturally to the Fisher-Rao (FR) metric. It can be shown that the FR distance between two functions is equivalent to the $\mathbb{L}^2$ distance between their respective SRVF transformations, i.e., $\| f_1 - f_2 \|_{FR} = \|q(f_1) - q(f_2) \|^2$ (see \cite{srivastava-etal-JASA:2011}). More importantly, the FR metric is phase invariant under warping, i.e., $\| f_1 \circ \gamma - f_2 \circ \gamma \|_{FR} = \|f_1 - f_2 \|_{FR}$ and \cite{srivastava2011} shows that the FR metric is symmetric under warping as well, i.e., $\| f_1 \circ \gamma - f_2 \|_{FR} = \| f_1 - f_2 \circ \gamma \|_{FR}$.
Following this framework and without loss of generality, we will align the SRVF of $f_2$ to the SRVF of $f_1$, and then map the aligned functions back to the original space $\mathbb{F}$.
In other words, we seek to estimate the warping function that minimizes $||q_1- (q_2,\gamma)||^2$, where $(q_2,\gamma)=(q_2\circ\gamma)\sqrt{\dot{\gamma}}$ and $q_j \doteq q(f_j(t))$ for $j=1,2$.
Additional challenges in function alignment come when placing the problem in a Bayesian framework, namely selecting a prior distribution for $\gamma \in \Gamma$. Recall that $\Gamma$ is the space of diffeomorphic function mappings from $[0,1] \mapsto [0,1]$.
Both optimization and Bayesian inference over this space is difficult, in particular, because this space is nonlinear and infinite dimensional.
For example, the sum or scalar product of functions in $\Gamma$ is not necessarily still contained in $\Gamma$.
To overcome this difficulty, we follow the approach of \cite{lu2017,tucker2013} which transforms $\gamma$ to its corresponding SRVF representation, and exploits the Riemannian-geometric structure of this transformation.
Which ultimately allows us to utilize more traditional inference or optimization algorithms that rely on the linearity of the underlying search space.
To understand how this transformation works, it is helpful to think of the set of warping functions $\Gamma$ as the space of univariate cumulative distribution functions for random variables on $[0,1]$. Then, for each $\gamma \in \Gamma$ there is an associated density function.
Now, let $\psi=\sqrt{\dot{\gamma}}$ be the corresponding SRVF of $\gamma$ and $\Psi^+=\{\psi: [0,1]\mapsto \mathbb{R}^+\mid||\psi||^2=1\}$ \citep{bhattachayya1943} be the space of square-root densities (SRD).
The SRD space is the positive orthant of the unit sphere in the Hilbert space $\mathbb{L}^2([0,1])$ denoted by $\Psi\{\psi: [0,1]\mapsto \mathbb{R}\mid||\psi||^2=1\}$.
While this space is still infinite and nonlinear, it is much more simply defined. Since the SRVF is a bijective mapping and $\gamma(0)=0,$ we can reconstruct $\gamma$ by the inverse mapping $q^{-1}(\psi)(t)=\int_0^t\psi^2(s)ds.$
We can further simplify, and even linearize $\Psi^+$ by mapping from the top half of the unit sphere onto a tangent space at $\psi$ defined as
$$T_{\psi}(\Psi)=\{g\in \mathbb{L}^2\mid\int_0^1g(s)\psi(s) ds=0\}.$$
For simplicity, we typically take $\psi = 1$, i.e., the identity function.
Geometric details on $T_1(\Psi)$ can be found in \cite{srivastava2007} and \cite{kurtek2012,tucker2013}.
The exponential map and its inverse can be used to map between $\Psi$ and $T_1(\Psi):$
$$\begin{aligned}
&\exp_1: & T_1(\Psi)\mapsto \Psi &\hspace{4mm} \exp_1(g)=\cos(||g||)+\sin(||g||)\frac{g}{||g||},~~ g\in T_1(\Psi)\\
&\exp^{-1}_1: & \Psi\mapsto T_1(\Psi) &\hspace{4mm} \exp^{-1}_1(\psi)=\frac{\theta}{\sin(\theta)}(\psi-\cos(\theta))
,~\theta=d(1,\psi) \hbox{ and } \psi\in\Psi,
\end{aligned}$$
where $d(1,\psi) = \cos^{-1}\left(\int_0^1 \psi(t) dt \right)$.
Mapping $\gamma$ onto the tangent space $T_1(\Psi)$ (summarized in Figure \ref{fig:mappings}) gives a convenient representation of $\gamma$ in the parametric vector space, and thus allows for a straight-forward prior specification to be placed on $T_1(\Psi)$ (discussed further in \ref{sec:modelspec}).
Intuitively, this linearization is akin to mapping the points on the top half of the sphere to the tangent plane at the the north pole.
Moreover, because $T_1(\Psi)$ is the space of square integrable functions with mean zero, one can parameterize this space of functions with a basis representation.
\begin{figure}
\centering
\includegraphics[scale=0.3]{mappings.pdf}
\caption{A summary of the mapping from $\Gamma$ to $T_1(\Psi)$ where the space $\Gamma$ is a nonlinear manifold and $\Psi$ is the unit sphere in the Hilbert space $\mathbb{L}^2([0,1])$.}\label{fig:mappings}
\end{figure}
Finally, through all of these transformations, the SRVF of $f_2 \circ \gamma$, denoted as $\mathcal{G}(g)$ can be written as
$$\begin{aligned}
\mathcal{G}(g(t))&:=(q_2,\gamma)(t)= q_2(\gamma(t))\sqrt{|\dot{\gamma(t)}|}&\\
&=q_2\bigg(\int_0^t\psi^2(s)ds\bigg)\psi(t)=q_2\bigg(\int_0^t\exp^2(g)(s)ds\bigg)\exp(g)(t),&&\\
\end{aligned}$$
where $g$ lives on the tangent space of $\Psi^+$, which is again linear.
\section{Pairwise registration in $\mathbb{R}^1$}\label{sec:pairwise}
First, we will detail our proposed approach in the simple case of pairwise alignment of functions $y_1$ and $y_2$ on $\mathbb{R}^1$.
The functions are observed at a set of discretized time points $[t] = \{t_1,\dots,t_N\}$.
In this section, we will fully specify our proposed Bayesian hierarchical framework for this pairwise alignment setting.
Our Bayesian approach looks for optimal warping functions that warp one function to another using the $\mathbb{L}^2$ distance of the SRVF representations of the two functions, as defined in the previous section. Furthermore, by mapping the space of warping functions to the linear tangent space $T_1(\Psi)$ on the infinite dimensional half sphere we can perform the Bayesian inference over a parametrized, linear space.
\subsection{Model specification}\label{sec:modelspec}
Let $y_1$ and $y_2$ be noisy observations of $f_1, f_2 \in \mathbb{F}$, respectively.
Then, at the first level of our hierarchical model we have,
\textit{Level 1.}
$$y_1([t])=f_1([t])+\epsilon_1([t]),\ \epsilon_1([t]) \sim MVN(0_N,\sigma_1^2 I_N)$$
and
$$y_2([t])=f_2([t])+\epsilon_2([t]),\ \epsilon_2([t]) \sim MVN(0_N,\sigma_2^2 I_N),$$
where the observed noise processes $\epsilon_1, \epsilon_2$ are are assumed to follow a Gaussian white noise distribution with variances $\sigma_1^2$ and $\sigma_2^2$, respectfully, for each point in time.
Here we assume no temporal correlation between noise parameters.
Using the SRVF representation described in Section \ref{sec:review}, we aim to align $q_2(\gamma([t]))=q(f_2(\gamma([t]))$ to $q_1([t])=q(f_1([t]))$.
Using the transformations from the previous section, we can equate $q_2(\gamma([t]))$ to $\mathcal{G}(g([t]))$ where $g$ is the projection of the square root density of $\gamma$ onto the tangent space of the infinite half sphere and so we model $q_1([t])-\mathcal{G}(g([t]))$ using a zero-mean multivariate Gaussian distribution following \cite{lu2017}. It should be noted that $q_2(\gamma([t])) = q_2(\int_0^{[t]}\psi(s)\,ds)\psi([t])$, where $\int_0^{[t]}\psi(s)\,ds$ denotes the $N$-dimensional vector \\ $\{\int_0^{t_1}\psi(s)\,ds,\dots,\int_0^{t_N}\psi(s)\,ds\}$.
Thus, at the second level we have,
\textit{Level 2.}
$$q_1([t])-\mathcal{G}(g([t])) \sim MVN(0_N,\sigma^2 I_{N}),$$
with negative log-likelihood
\begin{equation}\label{eq:phi}
\Phi(g) := -\log p(q_1,q_2; g)\propto\bigg(\frac{1}{\sigma^2}\bigg)^{N}\exp\bigg\{-\frac{1}{2\sigma^2}\sum_{i=1}^{N}\big(q_1(t_i)-(q_2,\gamma)(t_i)\big)\bigg\}.
\end{equation}
Following the approach of \cite{lu2017}, we will assign a zero-mean Gaussian process prior to the sampled tangent space $T_1(\Psi),$
$$g \sim GP(0,\mathcal{C}),$$
with $\mathcal{C}$ being a positive, self-adjoint and trace-class operator on $\mathbb{R}^1.$
Zero-mean multivariate normal priors are assigned to mean functions $f_1([t])$ and $f_2([t])$ with structured kernel functions $K_1$ and $K_2$, i.e.
$$f_1([t])\sim MVN(0_N,K_1) \hbox{ and } f_2([t]) \sim MVN(0_N,K_2).$$
In this work, we specify the squared exponential kernel function $K([t])=s^2\exp(-(d/2l)^2)$ for both $K_1$ and $K_2$ where $d$ is the computed distance matrix of $[t].$
Conditionally-conjugate inverse-gamma priors are specified for all covariance parameters $\sigma^2,\ \sigma_1^2,\ \sigma_2^2$ with the prior for $\sigma^2$ less informative than prior for $\sigma_1^2$ and $\sigma_2^2$ to help with potential identifiability issues. Hyperparameters $s_1^2, s_2^2$ and $l_1,l_2$ are assigned inverse-gamma and uniform priors respectively.
The full hierarchical posterior can now be written as
$$p(g,\theta | y_1,y_2) = \left[ \prod_{j=1}^2p(y_j|f_j,\sigma_j^2) \right] \exp \left[ -\Phi(g) \right]\pi(g,\theta), $$
where $\theta \doteq \{f_1,f_2,\sigma^2,\sigma_1^2,\sigma_2^2,\ s_1^2,\ s_2^2,\ l_1,\ l_2\}$, $\pi(g,\theta)$ is the prior distribution over parameters $g$ and $\theta$, and $p(y_j|f_j,\sigma_j^2)=\prod_{i=1}^Np(y_j(t_i)|f_j(t_i),\sigma_j^2)$ for $j=1,2$ is a product of univariate Gaussian distributions.
\subsection{MCMC sampling}
A Metropolis within Gibbs sampler is used to sample from the complete posterior distribution $p(g,\theta|y_1,y_2).$
For practical implementation, functions $f_1, f_2$ are discretized and we specify a basis representation for $g$, i.e., $g=B\mathbf{v}$ where $\bf{v}$ and $B$ are the set of $n_v$ basis coefficients and matrix of basis functions, respectively. The reason for this specification is as follows.
Recall that $g$ lives on $T_1(\Psi)$, the tangent space to the top half of the infinite-dimensional unit sphere at the point $\psi \in \Psi$. For simplicity, we take $\psi$ to be the identity element so that $T_1$ becomes the space of all $\mathbb{L}^2$ integrable functions on $[0,1]$ that have zero mean with respect to unit weight, i.e., the uniform density on the unit interval. Even though $g$ is still infinite-dimensional we can sample it using the Karhunen-Lo\`{e}ve expansion where we specify the covariance operator $C$ via its eigenpairs $(b_k,\lambda_k^2)$, where the set $\{b_k(\cdot)\}$ forms an orthonormal basis for the functions space $T_1(\Psi)$. We therefore sample independent variates $v_k\sim\mathcal{N}(0,\lambda_k^2)$ and set
$$ g = \sum_{k=1}^M v_k b_k, $$
where $M$ is the number of basis functions. This is similar to the method proposed in \cite{lu2017}.
We can then use a Fourier series type representation to parameterize $C$ in terms of a finite collection of basis coefficients, $\mathbf{v}$, and discretized orthogonal basis vectors represented by columns of $B$.
The eigenvalues, $\lambda_k$, are set to $\sigma_g^2/k_2$, where $\sigma_g$ is assumed to be known or provided.
The choice of orthogonal basis functions are plentiful, but this paper uses the Fourier series and Legendre polynomial expansions. Both sets form a complete basis over the space of $\mathbb{L}^2$ integrable functions on a compact interval.
Moreover, for the application at hand, both basis functions (other than the constant) are orthogonal to the unit weight; thus making it easy to satisfy the tangent space property.
The Fourier basis is typically used for periodic functions and so convergence can be hindered if the function values are different at the domain end points.
Legendre polynomials do not have this limitation, but they are not as widely known.
The smoother the function to be approximated is, the faster the decay of the basis coefficients.
This means we need fewer basis coefficients to represent the underlying function.
In fact if the function is infinitely differentiable, the decay of the basis coefficients can be spectral, i.e., exponentially decaying. This applies to both Fourier and polynomial based expansions (see \cite{hesthaven2007}).
The basis specification for $C$ thus allows us to more efficiently explore the posterior space $g|y_1,y_2,\theta$.
To update $g,$ we wish to sample from the conditional posterior $p(g|y_1,y_2,\theta, B)$ $\propto \exp\big(-\Phi(g)\big)$.
To account for the complex geometry of the target space which ${g|\cdot}$ lies on and to allow for our sampler, we will exploit Hamiltonian dynamics.
Specifically we use the $\infty$-HMC algorithm described in \cite{beskos2017} to sample ${g|\cdot}$ within our MCMC sampling procedure. $\infty$-HMC is a modification of the standard Hamiltonian Monte Carlo (HMC algorithm developed to be well-defined on the infinite-dimensional Hilbert Space and thus is ideal for sampling from $T_1(\Psi)$.
HMC is a natural choice for more complex geometries, as it allows for efficient sampling along posterior contours.
As \cite{beskos2017} describes, HMC is an improvement over the preconditioned Crank-Nicholson method.
Moreover, as we will show, in cases where the target density is multi-modal and separated by low probability regions, a parallel chain implementation with random starting points is needed to fully explore the state space of ${g|\cdot}$ (details given in Section \ref{sec:multalign}).
The algorithm, detailed in Algorithm \ref{alg:infhmc} utilizes an auxiliary variable.
As suggested in \cite{beskos2011} and implemented in \cite{beskos2017}, we choose to specify the auxiliary variable $v,$ interpreted as the \textit{velocity} of $g$.
To implement Algorithm \ref{alg:infhmc}, and to update $g$, it is necessary to define the location-specific preconditioner matrix $\mathcal{K}(g)$ as the covariance of a local Gaussian approximation $N(m(g),\mathcal{K}(g))$ to the posterior.
We define this pre-conditioner through its inverse:
$$\mathcal{K}(g)^{-1} = \mathcal{C}^{-1} + \beta \mathcal{H}(g),$$
where $\mathcal{H}(g)$ is chosen as the Gauss-Newton Hessian (GNH), i.e.
$$\mathcal{H}(g) = \langle \bigtriangledown \mathcal{G} (g),\ \mathcal{C}^{-1}\bigtriangledown \mathcal{G} (g)\rangle.$$
We can then calculate the natural gradient $\eta$:
$$\eta (g)= -\mathcal{K}(g) \big[ \bigtriangledown \Phi (g)-\beta \mathcal{H} (g) g\big], $$
to calculate the Hamiltonian flow, $\Xi^t.$
As the exact analytic expression of $\Xi^t$ is often not available, $\infty$-HMC defines the flows $\Xi_1^t,\ \Xi_2^t$ of a split Hamiltonian system to numerically approximate $\Xi^t:$
$$\Xi_1^t(g,v)=\big(g,v+\frac{t}{2}\eta(g)\big)\ \hbox{ and }\ \Xi_2^t(g,v) = \big(g\cos t+v\sin t,-g\sin t+v\cos t\big).$$
The leapfrog map $\Psi_h(g,v): (u_0, v_0) \mapsto (g_h,v_h)$ is the composition of three sub-steps:
$$\Psi_h(g,v)=\Xi_1^{h/2} \circ \Xi_2^h \circ \Xi_1^{h/2},$$
and the exact Hamiltonian flow $\Xi^T$ is then approximated by
$$\Psi_h^{I}(g,v)=\Psi_h^{\lfloor T/h \rfloor},$$ a concatenation of $I=\lfloor T/h \rfloor$ Verlet steps.
$\infty$-HMC requires user chosen time-step $h$ and leap-frog step $T.$ $\bigtriangledown\Phi(g)$ is the directional derivative of $\Phi$, defined in \eqref{eq:phi}, with respect to $g$ which is provided in the \ref{sec:deriv}.
Compared to the \textit{Z-mixture pCN algorithm} \citep{cotter2013} used in \cite{lu2017} to update $\bf{v}$, we find that using $\infty$-HMC algorithm to update $g$ is significantly more efficient.
We favor $\infty$-HMC over Z-mixture pCN because we are sampling from an infinite dimensional space. \cite{beskos2017} demonstrates that random walk-like algorithms, like the Z-mixture pCN, are not efficient and can break down on these high dimensional spaces.
Additionally, the gradient information $\infty$-HMC relies upon helps inform the sampler and has been shown to increase overall MCMC sampling performance for hierarchical models, especially when sampling on infinite dimensional spaces \citep{beskos2017, BDA}.
For broader application and sampling efficiency of the prior on the space of diffeomorphisms we chose to implement $\infty$-HMC for these reasons.
We found that the Z-mixture pCN algorithm, even when implemented via parallel chains with random starts, was not able to capture the expected bimodal posterior in our simulated data example.
The posterior was biased towards one mode only as shown in the posterior samples and achieved higher SSE. Figures demonstrating these comparisons are given in \ref{sec:mcmcdiag}.
The additional tuning parameters, specifically the leap frog and time step sizes ($T$ and $h$ in Algorithm~\ref{alg:infhmc} respectively), available using $\infty$-HMC allowed for the posterior samples to more easily jump between the multiple possible target modes.
\begin{algorithm}
\caption{Sampling Algorithm for $g$ via $\infty$-HMC (\cite{beskos2011})}\label{alg:infhmc}
\begin{enumerate}
\item Given current $g$, propose $g'= P_g\{\Psi_h^{I}(g,v))\}$ where $P_g\{\cdot\}$ is the projection of $(g',v')=\Psi_h^{I}(g,v)$ onto the $g$ argument and $v$ (velocity) is an auxiliary variable sampled from $N(0,\mathcal{C}).$\\
\item Accept $g'$ with probability $1 \wedge \exp\{-\Delta H(g,v)\},$
$$\Delta H(g,v)\equiv H(g',v')-H(g,v),\ \ H(g,v)=\Phi(g)+\frac{1}{2}\langle g,\mathcal{C}^{-1}g\rangle+\frac{1}{2}\langle v,\mathcal{C}v\rangle.$$
\end{enumerate}
\end{algorithm}
Mean functions $f_1, f_2 \in \mathbb{R}^1$ are sampled from their respective conditional posteriors
$$p(f_k|y_k,\theta)\propto p(y_k|f_k,\sigma_k^2)p(q_k([t])-\mathcal{G}(g([t]))|y_1,y_2,\theta)\pi(f_k),\ k=1,2~,$$
via Metropolis-Hastings with specified proposals
$$f_k'([t]) \sim MVN(f_k([t]),\Sigma_k^*),\ k=1,2,$$
and squared exponential correlation structures specified for $\Sigma_1^*$ and $\Sigma_2^*.$
Variance parameters $\sigma^2,\ \sigma_1^2,\ \sigma_2^2$ and hyperparameters $s_1^2$ and $s_2^2$ can be Gibbs sampled from their respective inverse-gamma posteriors. Hyperparameters $l_1$ and $l_2$ can be sampled via Metropolis-Hastings with Gaussian proposal distributions and acceptance ratios:
$$\alpha_{l_k}=\frac{p(f_k|y_k,l'_k)\pi(l_k)}{p(f_k|y_k,l_k)\pi(l'_k)},\ k=1,2.$$
\subsection{MCMC for multiple optimal alignments}\label{sec:multalign}
Due to the high efficiency of the $\infty$-HMC algorithm, a single posterior chain generated using $\infty$-HMC would not be able to jump between modes and thus would be unable to capture multi-modal distributions.
As highlighted by \cite{nishimura2017}, if the target distribution is multi-modal with modes separated by regions of low probability density; it is nearly impossible for any HMC algorithm to transition between modes due to the conservation of energy property of Hamiltonian dynamics.
To obtain a multi-modal posterior distribution for $g|y_1,y_2,\theta,$ we run $K$ parallel chains with initial values for the basis coefficients of $g$ chosen randomly from a standard Normal distribution.
The idea is to run separate chains at enough randomly generated starting points to land in the vicinity of the multiple modes; so that we don't have to rely on the HMC algorithm to cross the low-probability regions. Of course, we have no idea where those modes are in general, so it is important to explore the initial starting space.
In this case the starting space is the space of coefficients, $\mathbf{v}$, of the linear expansion of $g$. Furthermore, under the assumption that each chain reaches their target sampling density after a thorough burn-in period, we can treat each chain as providing independent samples; and thus simply pool them together. For example, if each of the ten chains produce 10k samples, we can pool them together to obtain 100k samples.
Alternatively, we can think of this approach as a type of parallel MCMC with a constant temperature parameter and no mixing.
It is important to highlight that there could be more than one best possible solution. Having a posterior that can capture the total variability of the warping function sample all possible solutions, rather than simple pick one as a local optima, is an advantage of our method. However in practice, it may be necessary to choose between multiple possible solutions. This determination is application dependent; a good metric is to choose the best warping function that produces the smallest amplitude distance $||q_1 - (q_2,\gamma))||^2$ between the two SRVFs $q_1$ and $q_2$. Another suggestion is one sees see more of the posterior favoring one alignment over the other and this solution could be chosen.
\section{Examples}\label{sec:examples}
To demonstrate our proposed method, we will compare our flexible Bayesian approach with the state-of-the-art Dynamic Programming (DP) method of \cite{FSDA}. In short, the DP algorithm solves the optimal warping function by proposing successively better piecewise linear pathways on a unit square grid; where better is defined in terms of the Fisher-Rao metric.
First, we will assess our method's performance on simulated noisy data for which multiple optimal alignments exist. Next, we will apply our method to two real datasets: a SONAR dataset and an iPhone movement dataset. The SONAR data is notably noisy and thus difficult to register. The iPhone data is less noisy but has multiple optimal alignments that previous methods discussed in Section \ref{sec:intro} do not account for.
Before we move on to the examples, we need to address the issue of computing means or averages in non-Euclidean geometries, e.g., the positive orthant of infinite-dimensional unit sphere. For problems of this type, we typically use the Karcher or Fr\'{e}chet mean.
This is obtained by minimizing the average squared distance between all pairs of sample points, $x_1,\dots,x_N \in \Psi$ i.e.
$$m = \text{arg} \min_{p \in \Psi} \sum_{i=1}^N d^2(p,x_i), $$
where $d$ is the metric or geodesic distance measure on $\Psi$, the infinite-dimensional positive orthant of the unit sphere, given by
$$d(\psi_1,\psi_2) = \cos^{-1}\left(\int_0^1 \psi_1(s)\psi_2(s) ds \right).$$
Note that in the case of Euclidean geometry, the Karcher mean is simply the average. The computation of the Karcher median is analogous. We use the method presented in \cite{XieKurtek:2016}, which is an extension of Algorithm 2 in \cite{tucker2013}. Now, in the multiple alignment case, our posterior will be multi-modal. In order to find the centroid of these individual modes, we do the following. We begin by computing all the pairwise distances under the metric $d$, between all posterior samples. We then use a clustering algorithm to separate the samples into their respective clusters. Then, for each cluster the Karcher mean or median can be computed as in the unimodal case.
\subsection{Simulated data} \label{sec:simex}
Consider the misaligned and noisy observations $y_1$ and $y_2$ shown in Figure \ref{fig:simf1f2} simulated by adding Gaussian white noise with $\sigma_1^2=\sigma_2^2=0.001$ to the true mean functions $f_1$ and $f_2$. Without loss of generality, we wish to align $y_2$ to $y_1.$
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{bimodal_simex.png}
\includegraphics[width=0.48\textwidth]{bimodal_f1f2posteriors.png}
\caption{Simulations $y_1$ and $y_2$ (left) and their respective posteriors $f_1|y_1$ and $f_2|y_2$ (right) plotted against the true functions.}
\label{fig:simf1f2}
\end{figure}
The right panel in Figure \ref{fig:simf1f2} shows the estimated posteriors for $f_1$ and $f_2$ given our Bayesian approach.
Figure \ref{fig:simwarping} compares the dynamic programming (DP) solution with our Bayesian approach.
On the left, we see the optimal warping function for $y_2$ in blue, found by DP ($\gamma_{DP}$), which reflects some of the noise in the data and only identifies one possible warping function.
The corresponding warped solution $y_2 \circ \gamma_{DP}$ is also shown in blue on the right.
Also the warped solution of $\hat{f}_2 \circ \gamma_{DP}$ is shown in green, where $\hat{f}_2$ smoothed version of $y_2$ and DP is performed on a smoothed versions of $y_1$ and $y_2$.
The smoothing in this case was performed using Gaussian process regression.
In our Bayesian approach, we ran 8 MCMC chains each with 20,000 iterations, a burn-in of 5,000 iterations without thinning, for a total of 4,000 effective samples.
The proposal parameters were chosen such that acceptance rates for each chain were between 0.2 and 0.4 (see Figure \ref{fig:simex_chains} for MCMC chains).
Furthermore, we used $n_v=10$ Fourier basis functions to approximate the function $g$ on the tangent space $T_1(\Psi)$ resulting in a dimensionality of 20 (two for each basis function). The left figure of Figure \ref{fig:simwarping} compares the posterior samples of $\gamma|y_1,y_2$, with credible intervals and posterior mode to the DP solution. The right figure of Figure \ref{fig:simwarping} compares the posterior median warped solution to the DP warped solutions. From this we can see that the Bayesian approach gives us three possible solutions: (1) $f_2$ is aligned to the leftmost peak of $f_1$, (2) $f_2$ is aligned to the rightmost peak of $f_1$, and (3) $f_2$ is aligned to both peaks of $f_1$. Note that the DP method only gives us one out of the possible three solutions.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{bimodal_gammaall.png}
\includegraphics[width=0.48\textwidth]{bimodal_f2warped_med.png}
\caption{Comparison of the posterior distribution $\gamma|y_1,y_2$ to the DP solution $\gamma_{DP}$ applied to the simulated data, plotting the two identified modes and respective 95\% credible regions of the posterior $\gamma|y_1,y_2$ (left). Comparison of the DP warping of $y_2$ and smoothed $y_2$ to the median of the warping $f_2 \circ \gamma|y_1,y_2$ (right).}
\label{fig:simwarping}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{bimodal_gamma_colored.png}
\includegraphics[width=0.49\textwidth]{bimodal_f2warped_colored.png}
\caption{$\gamma|y_1,y_2$ (left) and $f_2 \circ \gamma|y_1,y_2$ (right) from the 8 different MCMC chains, colored by the 3 posterior modes found.}
\label{fig:simcolored}
\end{figure}
Further exploration of the Bayesian approach is shown in Figure \ref{fig:simcolored}. Here, we break down the posteriors $\gamma|y_1,y_2$ and $f_2 \circ \gamma|y_1,y_2$ from the 8 different MCMC chains, colored by the 3 posterior modes found. It is clear that 3 unique alignments were captured in the posteriors of $\gamma|y_1,y_2$ and $(f_2\circ \gamma)|y_1,y_2$ among the 8 different MCMC chains; identified by three distinct colors. The left shows the three posterior modes and respective 95\% credible regions, characterizing the uncertainty of each possible warping and the uncertainty of $(f_2\circ \gamma)|y_1,y_2$.
This is exhibited by the width of the color-coded samples on the right.
To give a statement on computational cost our implementation of the MCMC algorithm is done in MATLAB; and one chain for $3.2e4$ iterations takes 24 seconds on a 8 Core Mac Pro with 32GB of RAM.
The dynamic programming algorithm is computationally $O(N^2)$, where $N$ is the dimension of the grid in the search space for the minimal energy, and for this simulated example $N=101$.
The code provided the \verb+fdasrvf+ MATLAB package for DP is written in C and runs in 0.2 seconds.
We executed this simulation for 25 replicates and computed the point-wise sum-of-squared error (SSE) for the median warping from our method and the warping function from DP performed using the raw data, $y_j,~j=1,2$; also from performing DP on the smoothed version of $y_j$.
Figure~\ref{fig:multsim_SSE} presents a box plot for the SSE for all three methods.
The SSE of the Bayesian method is similar to performing DP on the smoothed data.
It should be noted that there is a larger spread in the inter-quartile region of the Bayesian method, and this is due to the fact that it is capturing multiple possible solutions that have different SSE values.
Both of the DP solutions only find one solution, and this uncertainty (e.g., more than one registration) is not characterized in these methods.
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{multsim_SSE.png}
\caption{Boxplot of the sum-of-squared errors for the Bayesian method compared to DP on the noisy data $y_j,~j=1,2$ (DP(y)) and DP on the smoothed noisy data (DP(f)).}
\label{fig:multsim_SSE}
\end{figure}
Figure~\ref{fig:multsim_ex} presents four replicates from the simulated data with the median from the Bayesian posterior along with the posterior samples.
The median of the posterior of $y_1$ and $y_2$ is shown respectively as $f_1$ and $f_2$.
Additionally, the DP solutions for the noisy data (green) and the smoothed noisy data (blue).
As expected the median of the Bayesian solution and the DP on the smoothed noisy data match nicely.
However, the benefit of the Bayesian solution is shown in the posterior, where the multiple possible registrations are captured (with associated uncertainties) as the example as $f_1$ has two peaks and $f_2$ has one peak.
\begin{figure}[H]
\centering
\begin{minipage}{\textwidth}
\centering
\includegraphics[width=.48\textwidth]{multsim_1.png}\quad
\includegraphics[width=.48\textwidth]{multsim_3.png}\\
\includegraphics[width=.48\textwidth]{multsim_4.png}\quad
\includegraphics[width=.48\textwidth]{multsim_7.png}
\end{minipage}
\caption{Four replicates from the simulated data showing the Bayesian alignment with the posterior in gray, also shown is the DP solution on both the noisy and smoothed noisy data.}
\label{fig:multsim_ex}
\end{figure}
\subsection{SONAR data} \label{sec:sonar}
Next, we apply our alignment approach to naturally noisy SONAR data collected at the Naval Surface Warfare Center Panama City Division's (NSWC PCD) test pond.
For a description of the pond and measurement setup, the reader is referred to \cite{art:kargl}.
In summary, the idea is to use the acoustic signals, generated from SONAR data, to discriminate and classify underwater ``targets'', e.g., unexploded ordnance. More precisely, acoustic signals were generated from the raw SONAR data to construct ``target strength" as a function of frequency and aspect angle.
Due to the relatively small separation distances between the targets in the measurement setup and orientation, and uncertainty in the placement, the scattered fields from the targets overlap leading to misaligned signatures.
This could lead to erroneous identification and classification of the target.
For this example, we took two one-dimensional misaligned signatures, each with 1102 data samples, from a target that was a small notched aluminum cylinder.
In our Bayesian approach, we ran 8 MCMC chains with 4,000 iterations each and a burn-in period of 2,000 steps without thinning; for a total of 16,000 samples.
The proposal, prior, and hyperprior parameters were chosen such that acceptance rates were between 0.2 and 0.4 (see Figure \ref{fig:sonar_chains} for MCMC chains).
We additionally specified $n_v=4$ Fourier basis functions to approximate $g$, as opposed to 8 in the previous example.
We observed that reducing the number of basis functions $n_v$, allowed our HMC algorithm to more easily explore the posterior space of $g$; likely by allowing the inference to focus entirely on the lower frequency basis elements.
Lastly, due to the noisiness of this data we only sampled $20\%$ of the data, instead of the full set of 1102 data points, in order to obtain smooth estimates of $f_1$ and $f_2$ and avoid overfitting.
The posteriors $f_1|y_1$ and $f_2|y_2$ of the acoustic signatures from the SONAR data are shown in Figure \ref{fig:sonar_f1f2post} on the left (gray) and right (in orange and yellow). On the right, we compare the DP aligned solution (in blue) and the Bayesian aligned solution (in gray). Note that our Bayesian approach offers multiple possible alignments, one of which is similar to the DP approach, but another which more accurately matches the maximum peaks of $f_1$ and $f_2$.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{sonar_f1f2.png}
\includegraphics[width=0.48\textwidth]{sonar_f2warped.png}
\caption{Posteriors $f_1|y_1$ and $f_2|y_2$ shown in gray compared to the raw data $y_1$ and $y_2$ (left). The posterior of the warped function $(f_2 \circ \gamma)|y_1,y_2$ captures more possible alignments than the single DP warped solution in blue (right).}\label{fig:sonar_f1f2post}
\end{figure}
Figure \ref{fig:sonar_gampost} on the right panel compares the MAP estimates to the DP solution. Figure \ref{fig:sonar_gampost} on the left compares the DP solution to the posterior distribution of $\gamma|y_1,y_2$ and shows the posterior median and $95\%$ credible region of $\gamma|y_1, y_2$.
The credible region quantifies the uncertainty around the estimated warping function $\gamma$ and thus the warped function $(f_2 \circ \gamma)$.
Note that the credible region obtained from the Bayesian approach encapsulates the DP solution, which itself is not able to adequately account for the noise in the data.
Subsequently, the uncertainty can be pushed through any later analyses for any quantity of interest that may depend on the aligned or warped functions.
For example, any statistical moment calculations performed on the aligned functions, e.g., principal component analysis, can be modified to account for the uncertainty in the warping functions.\footnote{The most straightforward way to do this would be in Monte Carlo type fashion - compute the statistics for your quantity of interest (QoI) for each posterior sample $\gamma|y_1,y_2$, resulting in a histogram for the desired QoI.}
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{sonar_gammaall.png}
\includegraphics[width=0.49\textwidth]{sonar_f2warped_med.png}
\caption{Comparison of the posterior distribution $\gamma|y_1,y_2$ to the DP solution $\gamma_{DP}$ applied to the SONAR data, plotting the two identified modes and respective 95\% credible regions of the posterior $\gamma|y_1,y_2$ (left). Comparison of the DP warping of $y_2$ and smoothed $y_2$ to the median of the warping $f_2 \circ \gamma|y_1,y_2$ (right).}
\label{fig:sonar_gampost}
\end{figure}
\subsection{iPhone data} \label{sec:iphone}
This data set consists of aerobic actions of subjects, such as biking, running, walking, etc., recorded using the Inertial Measurement Unit (IMU) on an Apple iPhone 4 smartphone; which these days is like using a cassette tape to listen to music.
The IMU included a 3D accelerometer, gyroscope, and magnetometer.
Each sample was taken at 60Hz, and manually trimmed to 500 samples (every 8.33s) to eliminate starting and stopping movements.
For more information on the data set the reader is referred to \cite{mccall-reddy-shah}.
We chose to demonstrate our method on two of forty-five functional samples from the walking accelerometer data in the $x$-direction. In this context, we are specifically interested in the information contained in the separated phase and amplitude components, rather than the resulting alignment.
For example we are only interested in the amplitude variability of the acceleration in the $x$-direction. Comparing the functions after warping would allow us to consider this variability.
The two examples we are aligning, shown in the right figure of Figure \ref{fig:iphone_f1f2post}, clearly have multiple possible alignments of $y_2$ to $y_1$ that we would like to also capture in the posterior $\gamma|y_1,y_2.$
To perform the Bayesian inference, we again ran 8 MCMC chains with 5,000 iterations each, a burn-in period of 1,000 steps, and thinning every other sample, for a total of 16,000 effective samples.
The proposal, prior and hyperprior parameters were chosen such that acceptance rates were between 0.2 and 0.4 (see Figure \ref{fig:iphone_chains} for MCMC chains).
We additionally specified $n_v=8$ Legendre basis functions to approximate $g.$ We specify Legendre polynomials here since the data is non-periodic in nature.
The posteriors $f_1|y_1$ and $f_2|y_2$ can be seen in the left panel of Figure \ref{fig:iphone_f1f2post} along with $y_1$ and $y_2$ and the medians are shown in the right panel of Figure \ref{fig:iphone_gampost}.
This image also compares the DP aligned solution $(y_2 \circ \gamma_{DP})$ to the posterior alignment of $f_2|y_2$, $(f_2 \circ \gamma) |y_1,y_2,$ from our Bayesian approach. It is clear that our Bayesian approach can capture multiple possible alignments, including alignments similar to the DP solution. For example, in addition to the alignment captured by the DP algorithm, the Bayesian approach also accounted for an alignment of $f_2$ to the leftmost peak of $f_1$.
It is interesting to note here that the method didn't match $f_2$ to the rightmost peak of $f_1$. This would be caused by an extremely distorted warping function and the Fisher Rao metric natively guards against that by penalizing too high of a gradient.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{iphone_f1f2_legendre.png}
\includegraphics[width=0.48\textwidth]{iphone_f2warped_legendre.png}
\caption{Posteriors $f_1|y_1$ and $f_2|y_2$ shown in gray compared to the raw data $y_1$ and $y_2$ (left). The posterior of the warped function $(f_2 \circ \gamma)|y_1,y_2$ captures a more accurate alignment than the single DP warped solution in blue (right).}
\label{fig:iphone_f1f2post}
\end{figure}
Figure \ref{fig:iphone_gampost} on the left compares the DP solution to the posterior distribution of $\gamma|y_1,y_2$ and shows the posterior median and $95\%$ credible region of $\gamma|y_1, y_2$.
The credible region quantifies the uncertainty around the estimated warping function $\gamma$ and warped function $(f_2 \circ \gamma)$.
The DP solution, obtained by trying to align the noisy data without prior data smoothing, is not built to account for the multiple possible alignments.
\begin{figure}[H]
\centering
\includegraphics[width=0.48\textwidth]{iphone_gammaall_legendre.png}
\includegraphics[width=0.48\textwidth]{iphone_f2warped_med.png}
\caption{Comparison of the posterior distribution $\gamma|y_1,y_2$ to the DP solution $\gamma_{DP}$ applied to the iPhone data, plotting the two identified modes and respective 95\% credible regions of the posterior $\gamma|y_1,y_2$ (left). Comparison of the DP warping of $y_2$ to the median of the warping $f_2 \circ \gamma|y_1,y_2$ (right).}
\label{fig:iphone_gampost}
\end{figure}
\section{Discussion}\label{sec:disc}
We have proposed a new flexible Bayesian approach to functional alignment that handles measurement error and accounts for multiple optimal alignments in the posterior of the warping function $\gamma.$
We have demonstrated its advantages over the state-of-the-art Dynamic Programming method that has been used in recent alignment literature \cite{FSDA,srivastava2011,tucker2013} using both simulated and real datasets.
Unlike DP, a Bayesian approach can characterize the uncertainty in the warping function solution $\gamma$, with Bayesian credible intervals.
The hierarchical structure of our Bayesian method allows us to estimate measurement error observed in $y_1$ and $y_2$ and extract the mean functions $f_1$ and $f_2$ to be aligned, respectively.
Accounting for measurement error directly in the method itself avoids any need for smoothing of the data typically done prior to applying DP in practice.
Additionally, by running parallel MCMC chains, the posterior of the warping function $\gamma$ is able to capture all possible alignments and visualize the posterior modes and respective credible regions for one or more distinct alignments.
Although existing Bayesian methods can account for uncertainty in the warping function, they are not able to find more than one possible alignment.
This is because they focus on converging efficiently to one solution, and do not account for measurement uncertainty.
Without accounting for all possible alignments, we are missing information; and not accounting for all uncertainty in $\gamma$.
For future work, we look to extend the method to the multiple alignment problem.
This will not be trivial in the case of multi-modal posteriors and multi-function alignment, as the solution will not be as simple as running multiple chains.
The extension in our case would be the following: given $I$ functions you would construct $I$ pairwise alignments in Level 1 and Level 2 and then model the mean or template function using a Gaussian process prior and update.
This processes would be somewhat similar to what was proposed in \cite{lu2017}.
However, there lies the additional difficulty in sampling a model which can give multiple solutions. For some cases, each pairwise alignment to the template would produce multiple possible gamma functions.
One would either need to determine which mode (e.g., \textit{optimal}) warping function to use or the uncertainty would need to be propagated to the sampling of the mean function.
As noted in Cheng et al., \cite{cheng2016}, this template is only identifiable up to an equivalence class of warpings.
The issue is more pronounced in the case of multiple alignments.
The use of wormhole MCMC (\cite{lan2014}) or something similar would need to be explored and possibly used in this situation.
Additionally, we can extend this from curves to trajectories that lie on Riemannian manifolds $\mathcal{M}$.
In this case, one has to account for the non-zero curvature of the space and in particular, the calculation of the gradient in the Fisher Rao metric.
\section*{Acknowledgment}
This paper describes objective technical results and analysis. Any subjective views or opinions that might be expressed in the paper do not necessarily represent the views of the U.S. Department of Energy or the United States Government. This work was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories; a multi-mission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-NA0003525.
| af1c714722d537ee3ee9e05a7d9e8103ef922eca | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:preliminaries}
\subsection{Notations}
The~set of Young diagrams will be denoted by $\mathbb{Y}$; the~set of Young
diagrams with $n$ boxes will be denoted by $\mathbb{Y}_n$. The~set $\mathbb{Y}$
has a~structure of an~oriented graph, called \emph{Young graph}; a~pair
$\mu\nearrow\lambda$ forms an~oriented edge in this graph if the~Young diagram
$\lambda$ can~be created from the~Young diagram $\mu$ by addition of a~single
box.
We will draw Young diagrams and tableaux in the~French convention with the
Cartesian~coordinate system $Oxy$, cf.~\cref{fig:RSK,fig:tableauFrench}. We
index the~rows and the~columns of tableaux by \emph{non-negative} integers from
$\mathbb{N}_0=\{0,1,2,\dots\}$. In particular, if $\Box$ is a~box of a~tableau, we
identify it with the~Cartesian~coordinates of its \emph{lower-left corner}:
$\Box=(x,y)\in \mathbb{N}_0\times \mathbb{N}_0$. For a~tableau $\mathcal{T}$ we denote by
$\mathcal{T}_{x,y}$ its entry which lies in the~intersection of the~row $y\in\mathbb{N}_0$
and the~column $x\in\mathbb{N}_0$.
The position of the box $s$ in the tableau $\mathcal{T}$
will be denoted by $\Pos_s(\mathcal{T})\in\mathbb{N}_0\times \mathbb{N}_0$.
Also the~rows of any Young diagram $\lambda=(\lambda_0,\lambda_1,\dots)$ are
indexed by the~elements of~$\mathbb{N}_0$; in particular the~length of the~bottom row
of $\lambda$ is denoted by~$\lambda_0$.
\subsection{Schensted row insertion}
\begin{figure}[t]
\centering
\hfill
\subfloat[]{
\begin{tikzpicture}[scale=0.75]
\clip (-0.1,-0.5) rectangle (4.1,4.5);
\draw (0,0) rectangle +(1,1);
\node at (0.5,0.5) {16};
\draw (1,0) rectangle +(1,1);
\node at (1.5,0.5) {37};
\draw (2,0) rectangle +(1,1);
\node at (2.5,0.5) {41};
\draw (3,0) rectangle +(1,1);
\node at (3.5,0.5) {82};
\draw (0,1) rectangle +(1,1);
\node at (0.5,1.5) {23};
\draw (1,1) rectangle +(1,1);
\node at (1.5,1.5) {53};
\draw (2,1) rectangle +(1,1);
\node at (2.5,1.5) {70};
\draw (0,2) rectangle +(1,1);
\node at (0.5,2.5) {74};
\draw (1,2) rectangle +(1,1);
\node at (1.5,2.5) {99};
\end{tikzpicture}
\label{fig:RSKa}
}
\hfill
\subfloat[]{
\begin{tikzpicture}[scale=0.75]
\clip (-1.5,-0.5) rectangle (4.1,4.5);
\fill[blue!10] (1,0) rectangle +(1,1);
\fill[blue!10] (1,1) rectangle +(1,1);
\fill[blue!10] (0,2) rectangle +(1,1);
\fill[blue!10] (0,3) rectangle +(1,1);
\draw (0,0) rectangle +(1,1);
\node at (0.5,0.5) {16};
\draw (1,0) rectangle +(1,1);
\node at (1.5,0.5) {37};
\draw (2,0) rectangle +(1,1);
\node at (2.5,0.5) {41};
\draw (3,0) rectangle +(1,1);
\node at (3.5,0.5) {82};
\draw (0,1) rectangle +(1,1);
\node at (0.5,1.5) {23};
\draw (1,1) rectangle +(1,1);
\node at (1.5,1.5) {53};
\draw (2,1) rectangle +(1,1);
\node at (2.5,1.5) {70};
\draw (0,2) rectangle +(1,1);
\node at (0.5,2.5) {74};
\draw (1,2) rectangle +(1,1);
\node at (1.5,2.5) {99};
\draw[->] (-0.3,-0.45) to[bend left=60] (-0.3,0.45);
\draw[->] (0,1) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\draw[->] (0,2) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\draw[->] (0,3) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\tiny
\node[] at (-0.85,0) {18};
\node[] at (-0.85,1) {37};
\node[] at (-0.85,2) {53};
\node[] at (-0.85,3) {74};
\end{tikzpicture}
\label{fig:RSKb}
}
\hfill
\subfloat[]{
\begin{tikzpicture}[scale=0.75]
\clip (-1.5,-0.5) rectangle (4.1,4.5);
\fill[blue!10] (1,0) rectangle +(1,1);
\fill[blue!10] (1,1) rectangle +(1,1);
\fill[blue!10] (0,2) rectangle +(1,1);
\fill[blue!10] (0,3) rectangle +(1,1);
\draw (0,0) rectangle +(1,1);
\node at (0.5,0.5) {16};
\draw (1,0) rectangle +(1,1);
\node at (1.5,0.5) {18};
\draw (2,0) rectangle +(1,1);
\node at (2.5,0.5) {41};
\draw (3,0) rectangle +(1,1);
\node at (3.5,0.5) {82};
\draw (0,1) rectangle +(1,1);
\node at (0.5,1.5) {23};
\draw (1,1) rectangle +(1,1);
\node at (1.5,1.5) {37};
\draw (2,1) rectangle +(1,1);
\node at (2.5,1.5) {70};
\draw (0,2) rectangle +(1,1);
\node at (0.5,2.5) {53};
\draw (1,2) rectangle +(1,1);
\node at (1.5,2.5) {99};
\draw (0,3) rectangle +(1,1);
\node at (0.5,3.5) {74};
\draw[->] (-0.3,-0.45) to[bend left=60] (-0.3,0.45);
\draw[->] (0,1) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\draw[->] (0,2) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\draw[->] (0,3) +(-0.3,-0.45) to[bend left=60] +(-0.3,0.45);
\tiny
\node[] at (-0.85,0) {18};
\node[] at (-0.85,1) {37};
\node[] at (-0.85,2) {53};
\node[] at (-0.85,3) {74};
\end{tikzpicture}
\label{fig:RSKc}
}
\caption{\protect\subref{fig:RSKa} The original tableau $\mathcal{T}$.
\protect\subref{fig:RSKb} We consider the Schensted row insertion of the number
$18$ to the~tableau~$\mathcal{T}$. The~highlighted boxes form the corresponding
bumping route. The small numbers on the left (next
to the arrows) indicate the~inserted/bumped numbers. \protect\subref{fig:RSKc} The
output $\mathcal{T}\leftarrow 18$ of the Schensted insertion. } \label{fig:RSK}
\end{figure}
\newcommand{a}{a}
\emph{The Schensted row insertion} is an algorithm which takes as an input a
tableau $\mathcal{T}$ and some number $a$. The number $a$ is inserted
into the first row (that is,~the bottom row, the row with the index~$0$) of
$\mathcal{T}$ to the~leftmost box which contains an entry which is strictly bigger
than $a$.
In the case when the row contains no entries which
are bigger than~$a$, the~number~$a$ is inserted into the leftmost
empty box in this row and the algorithm terminates.
If, however, the number $a$ is inserted into a box which was not empty,
the~previous content $a'$ of the box is \emph{bumped} into the second row
(that is, the~row with the index $1$). This means that the algorithm is
iterated but this time the number $a'$ is inserted into the second row to
the leftmost box which contains a number bigger than $a'$.
We repeat these steps of row insertion and bumping
until some number is inserted into a previously empty box.
This process is illustrated on \cref{fig:RSKb,fig:RSKc}. The outcome of
the Schensted insertion is defined as the~result of the~aforementioned procedure;
it will be denoted by $\mathcal{T} \leftarrow a$.
\medskip
Note that this procedure is well defined also in the~setup when $\mathcal{T}$ is
an~\emph{infinite} tableau (see \cref{fig:tableauFrench} for an~example), even
if the~above procedure does not terminate after a~finite number of steps.
\subsection{Robinson--Schensted--Knuth algorithm}
For the purposes of this article we consider a simplified version of \emph{the
Robinson--Schensted--Knuth algorithm}; for this reason we should rather call it
\emph{the Robinson--Schensted algorithm}. Nevertheless, we use the first name
because of its well-known acronym RSK. The RSK algorithm associates to a finite
sequence $w=(w_1,\dots,w_\ell)$ a~pair of tableaux: \emph{the insertion tableau
$P(w)$} and \emph{the recording tableau~$Q(w)$}.
The insertion tableau
\begin{equation}
\label{eq:insertion}
P(w) = \Big( \big( (\emptyset \leftarrow w_1) \leftarrow w_2 \big)
\leftarrow \cdots \Big) \leftarrow w_\ell
\end{equation}
is defined as the result of the iterative Schensted row insertion applied to the
entries of the sequence $w$, starting from \emph{the empty tableau $\emptyset$}.
The recording tableau $Q(w)$ is defined as the standard Young tableau of the
same shape as $P(w)$ in which each entry is equal to the number of
the~iteration of \eqref{eq:insertion} in which the given box of~$P(w)$ stopped
being empty; in other words the entries of $Q(w)$ give the order in which the
entries of the~insertion tableau were filled.
The tableaux $P(w)$ and $Q(w)$ have the same shape; we will denote this common
shape by $\RSK(w)$ and call it \emph{the RSK shape associated to $w$}.
The RSK algorithm is of great importance in algebraic combinatorics, especially in
the context of the representation theory \cite{Fulton1997}.
\subsection{Plancherel measure, Plancherel growth process}
\label{sec:pgp}
Let $\Sym{n}$ denote the symmetric group of order $n$. We will view each
permutation $\pi \in\Sym{n}$ as a sequence $\pi=(\pi_1,\dots,\pi_n)$ which has
no repeated entries, and such that $\pi_1,\dots,\pi_n\in\{1,\dots,n\}$. The
restriction of RSK to the symmetric group is a bijection which to a given
permutation from $\Sym{n}$ associates a pair $(P,Q)$ of standard Young tableaux
of the same shape, consisting of $n$ boxes. A~fruitful area of study concerns
the RSK algorithm applied to a uniformly random permutation from $\Sym{n}$,
especially asymptotically in the limit $n\to\infty$, see \cite{Romik2015a} and
the references therein.
\emph{The Plancherel measure} on $\mathbb{Y}_{n}$, denoted $\Plancherel_n$, is
defined as the probability distribution of the random Young diagram $\RSK(w)$
for a uniformly random permutation $w\in\Sym{n}$.
\medskip
An \emph{infinite standard Young tableau}
\cite[Section~2.2]{Kerov1999}
is a filling of the boxes
in a subset of the upper-right quarterplane with positive integers, such that
each row and each column is increasing, and each positive integer is used
exactly once.
There is a natural bijection between the set of infinite standard Young tableaux and the set of infinite paths in the Young graph
\begin{equation}
\label{eq:PGP0}
\lambda^{(0)} \nearrow \lambda^{(1)} \nearrow \cdots \qquad
\text{with } \lambda^{(0)}=\emptyset;
\end{equation}
this bijection is given by setting $\lambda^{(n)}$ to be the set of boxes of a
given infinite standard Young tableau which are $\leq n$.
If
$w=(w_1,w_2,\dots)$ is an \emph{infinite} sequence, the recording tableau $Q(w)$
is defined as the infinite standard Young tableau in which each non-empty entry
is equal to the number of the iteration in the infinite sequence of Schensted row insertions
\[ \big( (\emptyset \leftarrow w_1) \leftarrow w_2 \big)
\leftarrow \cdots
\]
in which the corresponding box stopped being empty, see \cite[Section~1.2.4]{RomikSniady-AnnPro}. Under the aforementioned bijection, the recording tableau $Q(w)$ corresponds to the sequence \eqref{eq:PGP0}
with
\[ \lambda^{(n)}= \RSK( w_1,\dots, w_n). \]
Let $\xi=(\xi_1,\xi_2,\dots)$ be an infinite sequence of independent, identically
distributed random variables with the uniform distribution $U(0,1)$ on the unit
interval $[0,1]$. \emph{The Plan\-che\-rel measure on the set of infinite standard
Young tableaux} is defined as the probability distribution of $Q(\xi)$.
Any sequence with the same
probability distribution as \eqref{eq:PGP0} with
\begin{equation}
\label{eq:lambda-rsk}
\lambda^{(n)}= \RSK( \xi_1,\dots, \xi_n)
\end{equation}
will be called \emph{the Plancherel growth process} \cite{Kerov1999}. It turns
out that the Plancherel growth process is a Markov chain \cite[Sections~2.2 and
2.4]{Kerov1999}. For a~more systematic introduction to this topic we recommend
the monograph \cite[Section~1.19]{Romik2015a}.
\subfile{figures-Poisson/big_simulation/random_SYT.tex}
\subsection{Bumping route}
The~\emph{bumping route} consists of the~boxes the~entries of which were
changed by the~action of Schensted insertion,
including the~last, newly created box, see \cref{fig:RSKb,fig:RSKc}.
The~bumping route will be denoted by
$\mathcal{T}{\leftsquigarrow a}$ or by $\mathcal{T}_{\leftsquigarrow a}$ depending on
current typographic needs. In any row $y\in\mathbb{N}_0$ there is at most one box from
the~bumping route $\mathcal{T}{\leftsquigarrow a}$; we denote by
$\mathcal{T}_{\leftsquigarrow a}(y)$ its $x$-coordinate. We leave
$\mathcal{T}_{\leftsquigarrow a}(y)$ undefined if such a~box does not exist. In
this way
\begin{equation}
\label{eq:bumping_points}
\mathcal{T}{\leftsquigarrow a}~=
\Big\{ \big( \mathcal{T}_{\leftsquigarrow a}(y),y\big) : y\in\mathbb{N}_0\Big\}.
\end{equation}
For example, for the~tableau $\mathcal{T}$ from \cref{fig:RSKa} and $a=18$ we have
\[ \mathcal{T}_{\leftsquigarrow a}(y) =
\begin{cases}
1 & \text{for } y\in\{0,1\}, \\
0 & \text{for } y\in\{2,3\}.
\end{cases}
\]
The~bumping route $\mathcal{T}{\leftsquigarrow a}$ can~be visualized either as a
collection of its boxes or as a~plot of the~function
\[ x(y) = \mathcal{T}_{\leftsquigarrow a}(\lfloor y \rfloor), \qquad y\in\mathbb{R}_+, \]
cf.~the~thick red line on \cref{fig:tableauFrench}.
\subsection{Bumping routes for infinite tableaux}
Any bumping route which corresponds to an~insertion to a~\emph{finite} tableau
is, well, also finite. This is disadvantageous when one aims at the
asymptotics of such a~bumping route in a~row of index $y$ in the~limit
$y\to\infty$. For such problems it would be preferable to work in a~setup in
which the~bumping routes are infinite; we present the~details in the~following.
Let us fix the~value of an~integer $m\in\mathbb{N}_0$.
Now, for an~integer $n\geq m$ we consider a~real number $0<\alpha_n<1$ and a
finite sequence $\xi=(\xi_1,\dots,\xi_n)$ of independent, identically distributed
random variables with the~uniform distribution $U(0,1)$ on the~unit interval
$[0,1]$. In order to remove some randomness from this picture we will condition
the~choice of $\xi$ in such a~way that there are exactly $m$ entries of $\xi$ which
are smaller than~$\alpha_n$; heuristically this situation is similar to a
scenario without conditioning, for the~choice of
\begin{equation}
\label{eq:alpha}
\alpha_n=\frac{m}{n}.
\end{equation}
We will study the~bumping route
\begin{equation}
\label{eq:bumping-1}
P(\xi_1,\dots,\xi_n) \leftsquigarrow \alpha_n
\end{equation}
in the~limit as $n\to\infty$ and $m$ is fixed.
Without loss of generality we may assume that the~entries of the~sequence~$\xi$
are all different. Let $\pi\in\Sym{n}$ be the~unique permutation which encodes
the~relative order of the~entries in the~sequence $\xi$, that is
\[ \left( \pi_i < \pi_j \right) \iff \left( X_i < X_j \right) \]
for any $1\leq i,j\leq n$. Since the~algorithm behind the Robinson--Schensted--Knuth
correspondence depends only on the~relative order of the~involved numbers and
not their exact values, it follows that the~bumping route \eqref{eq:bumping-1}
coincides with the~bumping route
\begin{equation}
\label{eq:bumping-2}
P(\pi_1,\dots,\pi_n) \leftsquigarrow m+\nicefrac{1}{2}.
\end{equation}
The~probability distribution of $\pi$ is the~uniform measure on $\Sym{n}$; it
follows that the~probability distribution of the~tableau $P(\pi_1,\dots,\pi_n)$
which appears in \eqref{eq:bumping-2} is the~Plancherel measure $\Plancherel_n$
on the~set of standard Young tableaux with $n$ boxes. Since such a
Plancherel-distributed random tableau with $n$~boxes can~be viewed as a
truncation of an~infinite standard Young tableau~$\mathcal{T}$ with the Plancherel
distribution, the~bumping routes \eqref{eq:bumping-1} and \eqref{eq:bumping-2}
can~be viewed as truncations of the~infinite bumping route
\begin{equation}
\label{eq:bumping-3}
\mathcal{T}\leftsquigarrow m + \nicefrac{1}{2},
\end{equation}
see \cref{fig:tableauFrench} for an~example.
\subsection{The~main problem: asymptotics of infinite bumping routes}
The~aim of the~current paper is to investigate the~asymptotics of the
\emph{infinite} bumping route \eqref{eq:bumping-3} in the~limit $m\to\infty$.
Heuristically, this corresponds to investigation of the~asymptotics of
the~\emph{finite} bumping routes \eqref{eq:bumping-1} in the~simplified setting
when we do not condition over some additional properties of $\xi$, in the~scaling
in which $\alpha_n$ does not tend to zero too fast (so that $\lim_{n\to\infty}
\alpha_n n =\infty$, cf.~\eqref{eq:alpha}), but on the~other hand $\alpha_n$
should tend to zero fast enough so that the~bumping route is long enough that
our asymptotic questions are well defined. We will not pursue in this direction
and we will stick to the investigation of the~\emph{infinite} bumping route~\eqref{eq:bumping-3}.
\medskip
Even though Romik and {the~last named author} \cite{RomikSniadyBumping} considered the
asymptotics of \emph{finite} bumping routes, their discussion is nevertheless
applicable in our context. It shows that in the~\emph{balanced scaling} when we
focus on the~part of the~bumping route with the~Cartesian~coordinates $(x,y)$ of
magnitude $x,y=O \! \left( \sqrt{m} \right)$, the~shape of the~bumping route
(scaled down by the~factor $\frac{1}{\sqrt{m}}$) converges in probability
towards an~explicit curve, which we refer to as \emph{the~limit bumping curve}, see
\cref{fig:simulated-linear} for an~illustration.
\subfile{figures-Poisson/multiple_bumping_routes/multiple_routes-linear.tex}
In the~current paper we go beyond the~scaling used by Romik and {the~last named author} and
investigate the~part of the~bumping route with the~Cartesian~coordinates of order
$x=O(1)$ and $y\gg \sqrt{m}$. This part of the~bumping curves was not visible
on \cref{fig:simulated-linear}; in order to reveal it one can~use the
semi-logarithmic plot, cf.~\cref{fig:simulated-log,fig:simulated-log-stretch-2}.
\subsection{The~naive hyperbola}
\label{sec:naive}
The~first step in this direction would be to stretch the
validity of the~results of Romik and {the~last named author} \cite{RomikSniadyBumping} beyond
their limitations and to expect that the~limit bumping curve describes the
asymptotics of the~bumping routes also in this new scaling.
This would correspond to the~investigation of the~asymptotics of the
(non-rescaled) limit bumping curve $\big(x(y), y\big)$ in the~regime
$y\to\infty$. The~latter analysis was performed by {the~first named author}
\cite{Marciniak2020}; one of his results is that
\[ \lim_{y\to\infty} x(y) y = 2;\]
in other words, for $y\to \infty$ the~\emph{non-rescaled} limit bumping curve
can~be approximated by the~hyperbola~$x y = 2$ while its \emph{rescaled}
counterpart which we consider in the~current paper by the~hyperbola
\begin{equation}
\label{eq:hyperbola}
x y = 2m
\end{equation}
which is shown on \cref{fig:simulated-linear} as the~dashed line. At the very end of
\cref{sec:naive-hyperbola} we will discuss the~extent to which this naive
approach manages to confront the~reality.
\subfile{figures-Poisson/multiple_bumping_routes/multiple_routes-log.tex}
\subfile{figures-Poisson/multiple_bumping_routes/multiple_routes-log-stretch-2.tex}
\subsection{In which row a~bumping route reaches a~given column?}
\label{sec:in-which-row}
Let us fix some (preferably infinite) standard Young tableau $\mathcal{T}$.
The~bumping route in each step jumps to the~next row, directly up or to the
left to the~original column; in other words
\[ \mathcal{T}_{\leftsquigarrow m+\nicefrac{1}{2}} (0) \geq
\mathcal{T}_{\leftsquigarrow m+\nicefrac{1}{2}}(1) \geq \cdots \]
is a~weakly decreasing sequence of non-negative integers.
For $x,m\in\mathbb{N}_0$ we denote by
\begin{equation}
\label{eq:Y}
Y_x^{[m]} = Y_x = \min\left\{ y\in\mathbb{N}_0 :
\mathcal{T}_{\leftsquigarrow m+\nicefrac{1}{2}}(y) \leq x \right\}
\end{equation}
the~index of the~first row in which the~bumping route
$\mathcal{T} {\leftsquigarrow m+\nicefrac{1}{2}}$ reaches the~column with
the index $x$ (or less, if the~bumping route skips the~column $x$ completely).
For example, for the~tableau $\mathcal{T}$ from \cref{fig:tableauFrench} we have
\[ Y_0^{[3]}= 4, \quad Y_1^{[3]}= 2, \quad Y_2^{[3]}= 1, \quad
Y_3^{[3]}=Y_4^{[3]}=\cdots= 0.
\]
If such a~row does not exist we set $Y_x=\infty$; the~following result shows
that we do not have to worry about such a~scenario.
\begin{proposition}
\label{prop:is-finite}
For a~random infinite standard Young tableau $\mathcal{T}$ with the~Plancherel
distribution
\[ Y_x^{[m]} < \infty \qquad \text{for all } x,m\in\mathbb{N}_0\]
holds true almost surely.
\end{proposition}
The~proof is postponed to \cref{sec:proof-prop:is-finite}.
For a~sketch of the~proof of an~equivalent result see the~work of Vershik
\russian%
{}%
{\cite[Predlozhenie 4]{Vershik2020}}
who uses different methods.
\medskip
\begin{theorem}[The~main result]
\label{thm:poisson-point}
Assume that $\mathcal{T}$ is an~infinite standard
Young tableau with the~Plancherel distribution. With the~above notations,
the~random set
\[ \left( \frac{2 m}{Y^{[m]}_0}, \frac{2 m}{Y^{[m]}_1}, \dots \right) \]
converges in distribution, as $m\to\infty$, to the Poisson point process on
$\mathbb{R}_+$ with the~unit intensity.
\end{theorem}
The~proof is postponed to \cref{sec:proof:thm:poisson-point}.
\begin{remark}
\label{rem:poisson}
The~Poisson point process \cite[Section 4]{Kingman1993}
\begin{equation}
\label{eq:Poisson-point-process}
\left( 0< \xi_0 < \xi_1 < \cdots \right)
\end{equation}
on $\mathbb{R}_+$ can~be viewed concretely as the~sequence of partial sums
\begin{align*}
\xi_0 &= \psi_0,\\
\xi_1 &= \psi_0+\psi_1, \\
\xi_2 &= \psi_0+\psi_1+\psi_2, \\
\vdots
\end{align*}
for a~sequence $(\psi_i)$ of independent, identically distributed random
variables with the~exponential distribution $\operatorname{Exp}(1)$. Thus a
concrete way to express the~convergence in \cref{thm:poisson-point} is to say
that for each $l\in\mathbb{N}_0$ the~joint distribution of the~\emph{finite} tuple of random
variables
\[ \left( \frac{2 m}{Y^{[m]}_0},\dots, \frac{2 m}{Y^{[m]}_l} \right) \]
converges, as $m\to\infty$, to the~joint distribution of the~sequence of partial sums
\[ \left( \psi_0,\;\;\; \psi_0+\psi_1,\;\;\; \dots, \;\;\; \psi_0+\psi_1+\cdots+\psi_l \right).\]
\end{remark}
\begin{corollary}
\label{cor:first-column}
For each $x\in\mathbb{N}_0$ the~random variable $\frac{Y_x^{[m]}}{2m}$
converges in distribution, as $m\to\infty$, to the~reciprocal of the Erlang distribution $\operatorname{Erlang}(x+1,1)$.
\end{corollary}
In particular, for $x=0$ it follows that the~random variable
$\frac{Y_0^{[m]}}{2m}$ which measures the~(rescaled) number of steps of the
bumping route to reach the~leftmost column converges in
distribution, as $m\to\infty$, to the~Fréchet distribution of shape parameter
$\alpha=1$:
\begin{equation}
\label{eq:frechet}
\lim_{m \to\infty} \mathbb{P}\left( \frac{Y_0^{[m]}}{2m} \leq u \right) =
e^{- \frac{1}{u} } \qquad \text{for any $u\in\mathbb{R}_+$.}
\end{equation}
The~Fréchet distribution has a~heavy tail; in particular its first moment is
infinite which provides a~theoretical explanation for a~bad time complexity
of some of our Monte Carlo simulations.
Equivalently, the~random variable $e^{ -\frac{2m}{Y_0}}$ converges in
distribution, as \mbox{$m\to\infty$}, to the~uniform distribution $U(0,1)$ on the~unit
interval. \cref{fig:cdf} presents the~results of Monte Carlo simulations which
illustrate this result.
\subfile{figures-Poisson/CDF/cdf.tex}
\subsection{Projective convention for drawing Young diagrams}
\label{sec:naive-hyperbola}
Usually in order to draw a~Young diagram we use the~French convention and the
$Oxy$ coordinate system, cf.~\cref{fig:tableauFrench}.
For our purposes it will be more convenient to change the~parametrization of
the~coordinate $y$ by setting
\[ z= z(y)= \frac{2 m}{y}.\]
This convention allows us to show an~infinite number of
rows of a~given tableau on a~finite piece of paper,
cf.~\cref{fig:tableauProjection}. We will refer to this
way of drawing Young tableaux as \emph{the~projective convention}; it is somewhat
reminiscent of the~\emph{English convention} in the~sense that the~numbers in
the~tableau increase along the~columns \emph{from top to bottom}.
\medskip
\subfile{figures-Poisson/multiple_bumping_routes/multiple_routes-projective.tex}
In the~projective convention the~bumping route can~be seen as the~plot of the~function
\begin{equation}
\label{eq:projective-coordinates}
x^{\projective}_{\mathcal{T},m}(z)=
\mathcal{T}_{\leftsquigarrow m+\nicefrac{1}{2}}\left(\left\lfloor \frac{2m}{z} \right\rfloor\right)
\qquad \text{for } z\in\mathbb{R}_+
\end{equation}
shown on \cref{fig:tableauProjection} as the~thick red line.
\medskip
With these notations \cref{thm:poisson-point} allows the~following convenient
reformulation.
\begin{theorem}[The~main result, reformulated]
\label{thm:main-poisson}
Let $\mathcal{T}$ be a~random infinite standard Young
tableau with the~Plancherel distribution. For $m\to\infty$ the~stochastic
process
\begin{equation}\label{eq:projective-process}
\left\{ x^{\projective}_{\mathcal{T},m}(z),\; z>0\right\}
\end{equation}
converges
in distribution to the~standard Poisson counting process \mbox{$\{ N(z),\; z>0 \}$}
with the~unit intensity.
\end{theorem}
For an~illustration see \cref{fig:projective}.
\begin{remark}
In \cref{thm:main-poisson} above, the~convergence in distribution for
stochastic processes is understood as follows: for any \emph{finite} collection
$z_1,\dots,z_l>0$ we claim that the~joint distribution of the~tuple of random
variables
\[ \left( x^{\projective}_{\mathcal{T},m}(z_1), \dots,
x^{\projective}_{\mathcal{T},m}(z_l)
\right)\]
converges in the~weak topology of probability measures, as $m\to\infty$, to
the~joint distribution of the~tuple of random variables
\[ \big( N(z_1), \dots, N(z_l) \big).\]
\end{remark}
\begin{proof}[Proof of \cref{thm:main-poisson}]
The process \eqref{eq:projective-process} is a counting process.
By the definition \eqref{eq:projective-coordinates}, the~time of its $k$-th jump (for an integer $k\geq 1$)
\[ \inf\left\{ z>0:\; x^{\projective}_{\mathcal{T},m}(z) = k \right\} = \frac{2m}{Y_{k-1}^{[m]}} \]
is directly related to the number of the row in which the bumping route
reaches the column with the index $k-1$. By \cref{thm:poisson-point} the joint distribution of the times of the jumps converges to the Poisson point process;
it follows therefore that \eqref{eq:projective-process} converges to the Poisson counting process, as required.
\end{proof}
The~plot of the~mean~value of the~standard Poisson process $ z \mapsto
\mathbb{E} N(z) $ is the~straight line $x=z$ which is shown on
\cref{fig:projective} as the~dashed line. Somewhat surprisingly it coincides
with the~hyperbola~\eqref{eq:hyperbola} shown in the~projective coordinate
system; a~posteriori this gives some justification to the~naive discussion from
\cref{sec:naive}.
\subsection{The~main result with the~right-to-left approach}
\cref{thm:poisson-point} was formulated in a~compact way which may
obscure the~true nature of this result. Our criticism is focused on the
\emph{left-to-right approach} from \cref{rem:poisson} which might give a~false
impression that the~underlying mechanism for generating
the~random variable~$\frac{2m}{Y_{x+1}^{[m]}}$
describing the~`time of arrival' of the~bumping
route to the~column number $x+1$ is based on generating first the~random
variable~$\frac{2m}{Y_x^{[m]}}$ related to the~previous column (that is
the~column directly to the~left), and adding some `waiting time' for the
transition. In fact, such a~mechanism is not possible without the time travel
because the~chronological order of the events is opposite: the~bumping route first
visits the~column $x+1$ and \emph{then} lands in the~column $x$. In the
following we shall present an~alternative, \emph{right-to-left} viewpoint which
explains better the~true nature of \cref{thm:poisson-point}.
\medskip
For the~Poisson point process \eqref{eq:Poisson-point-process} and an~integer
$l\geq 1$ we consider the~collection of random variables
\begin{equation}
\label{eq:ratios}
\xi_l, R_0, R_1, \dots, R_{l-1}
\end{equation}
which consists of $\xi_l$ and the~ratios
\[ R_i := \frac{\xi_{i+1}}{\xi_i}\]
of consecutive entries of $(\xi_i)$.
Then \eqref{eq:ratios} are independent random variables with the~distributions
that can~be found easily. This observation can~be used to \emph{define}
$\xi_0,\dots,\xi_l$ from the~Poisson point process by setting
\[ \xi_i = \xi_l\ \frac{1}{R_{l-1}} \frac{1}{R_{l-2}} \cdots \frac{1}{R_i}
\qquad \text{for } 0\leq i \leq l.\]
With this in mind we may reformulate \cref{thm:poisson-point} as follows.
\begin{theorem}[The~main result, reformulated]
\label{thm:multiplicative} For any integer $l\geq 0$ the~joint distribution
of the~tuple of random variables
\begin{align}
\label{eq:chrono-plus}
& \bigg( \frac{Y_l^{[m]}}{2m}, & \hspace*{2\arraycolsep} &\frac{Y_{l-1}^{[m]}}{2m}, &\hspace*{2\arraycolsep}
& \frac{Y_{l-2}^{[m]}}{2m}, & \hspace*{2\arraycolsep} & \dots, & \hspace*{2\arraycolsep} &
\frac{Y_{0}^{[m]}}{2m} \bigg) \\
\intertext{converges, as $m\to\infty$, to the~joint distribution of the~random variables}
\label{eq:funny-products}
& \bigg( \frac{1}{\xi_l}, & & \frac{1}{\xi_l} R_{l-1}, & & \frac{1}{\xi_l} R_{l-1} R_{l-2}, & &
\dots, & & \frac{1}{\xi_l} R_{l-1} \cdots R_0 \bigg),
\end{align}
where $\xi_l,R_{l-1},\dots,R_0$ are independent random variables, the
distribution of $\xi_l$ is equal to $\operatorname{Erlang}(l+1,1)$,
and for each
$i\geq 0$ the~distribution of the~ratio $R_i$ is supported on $[1,\infty)$ with
the~power law
\begin{equation}
\label{eq:power-law}
\mathbb{P}(R_i > u) = \frac{1}{u^{i+1}} \qquad \text{for } u\geq 1.
\end{equation}
\end{theorem}
The~order of the~random variables in \eqref{eq:chrono-plus} reflects the
chronological order of the~events, from left to right. Heuristically,
\eqref{eq:funny-products} states that the~transition of the~bumping route from
the column $x+1$ to the column $x$ gives a~\emph{multiplicative} factor~$R_x$ to
the~total waiting time, with the~factors $R_0,R_1,\dots$ independent.
It is more common in mathematical and physical models that the~total waiting
time for some event arises as a~\emph{sum} of some independent summands, so the
\emph{multiplicative} structure in \cref{thm:multiplicative} comes as a~small
surprise. We believe that this phenomenon can~be explained heuristically as
follows: when we study the~transition of the~bumping route from row $y$ to the
next row $y+1$, the~probability of the~transition from column $x+1$ to column
$x$ seems asymptotically to be equal to
\[ \frac{x+1}{y}+o\left(\frac{1}{y}\right)
\qquad\text{for fixed value of $x$, and for }y\to\infty.\]
This kind of decay would explain both the~multiplicative structure (`if
a~bumping route arrives to a~given column very late, it will stay in this column
even longer') as well as the~power law \eqref{eq:power-law}. We are tempted
therefore to state the~following conjecture which might explain the
aforementioned transition probabilities of the~bumping routes.
\begin{conjecture}
For a~Plancherel-distributed random infinite standard Young tableau $\mathcal{T}$
\begin{align*}
\mathbb{P}\left\{ \mathcal{T}_{x-1,y+1} < \mathcal{T}_{x,y} \right\} &= \frac{x}{y} + o\left(\frac{1}{y}\right) &
& \text{for fixed $x\geq 1$ and $y\to\infty$},
\\
\mathbb{P}\left\{ \mathcal{T}_{x-2,y+1} < \mathcal{T}_{x,y} \right\} &= o\left(\frac{1}{y}\right) &
& \text{for fixed $x\geq 2$ and $y\to\infty$}.
\end{align*}
Furthermore, for each $x\in\{1,2,\dots\}$
the~set of points
\begin{equation}
\label{eq:mysterious-set}
\left\{ \log \frac{y}{c} : y\in\{c,c+1,\dots\} \text{ and } \mathcal{T}_{x-1,y+1} < \mathcal{T}_{x,y} \right\}
\end{equation}
converges, as $c\to\infty$, to Poisson point process on $\mathbb{R}_+$ with the
constant intensity equal to $x$.
\end{conjecture}
Numerical experiments are not conclusive and indicate interesting clustering
phenomena~for the~random set \eqref{eq:mysterious-set}.
\subsection{Asymptotics of fixed $m$}
The~previous results concerned the~bumping routes $\mathcal{T} \leftsquigarrow
m+\frac{1}{2}$ in the~limit $m\to\infty$ as the~inserted number tends to
infinity. In the~following we concentrate on another class of asymptotic
problems which concern the~fixed value of $m$.
The~following result shows that \eqref{eq:frechet} gives asymptotically a~very
good approximation for the~distribution tail of $Y_0^{[m]}$ in the
scaling when $m$ is fixed and the~number of the~row $y\to\infty$ tends to
infinity.
\begin{proposition}
\label{prop:m=1}
For each integer $m\geq 1$
\[ \lim_{y\to\infty} y\ \mathbb{P}\left\{ Y_0^{[m]} \geq y\right\} = 2m. \]
\end{proposition}
This result is illustrated on \cref{fig:cdf} in the~behavior of each of the
cumulative distribution functions in the~neighborhood of $u=1$.
The~proof is postponed to \cref{sec:proof:prop:m=1}.
\begin{question}
What can~we say about the~other columns, that is
the~tail asymptotics of $\mathbb{P}\left\{ Y_x^{[m]} \geq y\right\}$
for fixed values of $x\in\mathbb{N}_0$ and
$m\geq 1$, in~the~limit $y\to\infty$?
\end{question}
\subsection{More open problems}
Let $\mathcal{T}$ be a~random Plancherel-distributed infinite standard Young
tableau. We consider the~\emph{bumping tree} \cite{Duzhin2019}
which is defined as the~collection
of all possible bumping routes for this tableau
\[ \left( \mathcal{T}\leftsquigarrow m + \nicefrac{1}{2} \: : \: m\in \mathbb{N}_0 \right) ,\]
which can~be visualized, for example, as on~\cref{fig:all-bumping}.
Computer simulations suggest that the~set of boxes which
can~be reached by \emph{some} bumping route for a~given tableau $\mathcal{T}$ is
relatively `small'. It would be interesting to state this vague observation in
a~meaningful way. We conjecture that the~pictures such as
\cref{fig:all-bumping} which use the~logarithmic scale for the~$y$ coordinate
converge (in the~scaling when $x=O(1)$ is bounded and $y\to\infty$) to some
meaningful particle jump-and-coalescence process.
\subfile{figures-Poisson/zlepianie/zlep-log.tex}
\subsection{Overview of the paper. Sketch of the proof of \cref{thm:poisson-point}}
As we already mentioned, the detailed proof of \cref{thm:poisson-point} is
postponed to \cref{sec:proof:thm:poisson-point}. In the following we present an
overview of the paper and a rough sketch of the proof.
\subsubsection{Trajectory of infinity. Lazy parametrization of the bumping route}
\label{sec:intro-toi}
Without loss of generality we may assume that the Plancherel-distributed infinite
tableau $\mathcal{T}$ from the statement of \cref{thm:poisson-point} is of the form $\mathcal{T}=Q(\xi_1,\xi_2,\dots)$ for a sequence
$\xi_1,\xi_2,\dots$ of independent, identically distributed random variables with
the uniform distribution $U(0,1)$.
We will iteratively apply Schensted row insertion to
the~entries of the~infinite sequence
\begin{equation}
\label{eq:inftyinfty}
\xi_1,\dots,\xi_m, \infty, \xi_{m+1}, \xi_{m+2}, \dots
\end{equation}which is the~initial sequence $\xi$ with our favorite symbol $\infty$ inserted at the
position~$m+1$. At step $m+1$ the symbol $\infty$ is inserted at the end of the bottom
row; as further elements of the sequence \eqref{eq:inftyinfty} are inserted, the
symbol~$\infty$ stays put or is being bumped to the next row, higher and higher.
In \cref{prop:lazy-traj-correspondence} we will show that the trajectory of
$\infty$ in this infinite sequence of Schensted row insertions
\begin{equation}
\label{eq:inftyrowinsertions}
\Big( \big( P(\xi_1,\dots,\xi_m, \infty) \leftarrow \xi_{m+1} \big) \leftarrow \xi_{m+2} \Big) \leftarrow \cdots
\end{equation}
coincides with the bumping route $\mathcal{T}\leftsquigarrow m + \nicefrac{1}{2} $.
Thus our main problem is equivalent to studying the time evolution of the
position of $\infty$ in the infinite sequence of row insertions
\eqref{eq:inftyrowinsertions}. This time evolution also provides a convenient alternative
parametrization of the bumping route, called \emph{lazy parametrization}.
\subsubsection{Augmented Young diagrams}
\label{sec:intro-augmented}
\newcommand{\mytableau}[1]{\mathcal{T}^{(#1)}}
For $t\geq m$ we consider the insertion tableau
\begin{equation}
\label{eq:mytableau}
\mytableau{t}= P(\xi_1,\dots,\xi_m, \infty,\xi_{m+1},\dots,\xi_t)
\end{equation}
which appears at an intermediate step in \eqref{eq:inftyrowinsertions} after some
finite number of row insertions was performed. By removing the information about
the entries of the tableau $\mytableau{t}$ we obtain the \emph{shape} of $\mytableau{t}$, denoted by
$\operatorname{sh} \mytableau{t}$, which is a Young diagram with $t+1$ boxes. In the
following we will explain how to modify the notion of \emph{the shape of a tableau} so
that it better fits our needs.
Let us remove from the tableau
$\mytableau{t}$ the numbers $\xi_1,\dots,\xi_t$ and let us keep only the information about
the position of the box which contains the symbol~$\infty$. The resulting object, called
\emph{augmented Young diagram}
(see \cref{fig:augmented-yd} for an illustration),
can be regarded as a pair $\Lambda^{(t)}=(\lambda,\Box)$ which consists of:
\begin{itemize}
\item
the Young diagram $\lambda$ with $t$ boxes which keeps track of the positions of the boxes with the entries $\xi_i, i \in \{1, \dots, t\}$, in $\mytableau{t}$;
\item the outer corner $\Box$ of $\lambda$ which is the position of the box with $\infty$ in~$\mytableau{t}$.
\end{itemize}
We will say that $\operatorname{sh}^* \mytableau{t}=\Lambda^{(t)}$ is the \emph{augmented shape} of $\mytableau{t}$.
The set of augmented Young diagrams, denoted $\diagrams^*$, has a structure of an
oriented graph which is directly related to Schensted row insertion, as follows.
For a pair of augmented Young diagrams $\Lambda,\widetilde{\Lambda}\in\diagrams^*$ we say that $\Lambda\nearrow
\widetilde{\Lambda}$ if there exists a tableau $\mathcal{T}$ (which contains exactly one
entry equal to $\infty$) such that $\Lambda=\operatorname{sh}^* \mathcal{T}$ and there exists some
number $x$ such that $\widetilde{\Lambda}=\operatorname{sh}^* (\mathcal{T}\leftarrow
x)$, see \cref{fig:augmented-young-graph} and
\cref{sec:augmented-Young-graph} for more details.
With these notations the time evolution of the position of $\infty$ in the
sequence of row insertions \eqref{eq:inftyrowinsertions} can be extracted from
the sequence of the corresponding augmented shapes
\begin{equation}
\label{eq:aPgpro}
\Lambda^{(m)} \nearrow \Lambda^{(m+1)} \nearrow \cdots .
\end{equation}
\subsubsection{Augmented Plancherel growth processes}
The random sequence~\eqref{eq:aPgpro} is called \emph{the augmented Plancherel
growth process initiated at time $m$}; in~\cref{sec:aPgp} we will show that it is
a Markov chain with dynamics closely related to the usual (i.e., non-augmented)
Plancherel growth process.
Since we have a freedom of choosing the value of the integer $m\in\mathbb{N}_0$,
we get a whole family of augmented Plancherel growth processes.
It turns out that the transition probabilities for these Markov chains
do not depend on the value of $m$.
Our strategy is to use the Markov property of
augmented Plancherel growth processes combined with the following two pieces of information.
\begin{itemize}
\item \label{item:component-A}
\emph{Probability distribution at a given time $t$.} In \cref{prop:distribution-fixed-time} we give an asymptotic
description of the probability distribution of
$\Lambda^{(t)}$
in the scaling when $m,t\to\infty$ in such a way that $t=\Theta(m^2)$.
\item \label{item:component-B} \emph{Averaged transition probabilities.}
In \cref{prop:transition-augmented} we give an asymptotic description of the
transition probabilities for the augmented Plancherel growth processes
between two moments of time $n$ and $n'$ (with $n<n'$) in the scaling when
$n,n'\to\infty$.
\end{itemize}
Thanks to these results we will prove \cref{thm:lazy-poisson} which gives an
asymptotic description of the probability distribution of the trajectory of the
symbol $\infty$ or, equivalently, the bumping route in the lazy parametrization.
Finally, in \cref{sec:removing-laziness} we explain how to translate this result
to the non-lazy parametrization of the bumping route in which the boxes of the
bumping route are parametrized by the index of the row; this completes the proof
of \cref{thm:poisson-point}.
\medskip
The main difficulty lies in the proofs of the aforementioned
\cref{prop:distribution-fixed-time} and \cref{prop:transition-augmented};
in the following we sketch their proofs.
\subsubsection{Probability distribution of the augmented Plancherel growth process at a
given time.}
\label{sec:at-given-time}
In order to prove the aforementioned \cref{prop:distribution-fixed-time} we need
to understand the probability distribution of the augmented shape of the
insertion tableau $\mytableau{t}$ given by \eqref{eq:mytableau} in the scaling
when $m=O \big( \sqrt{t} \big)$.
Thanks to some symmetries of the RSK algorithm,
the tableau $\mytableau{t}$ is equal to the transpose of the insertion tableau
\begin{equation}
\label{eq:transposed}
P(\underbrace{\xi_t, \xi_{t-1}, \dots,\xi_{m+1}}_{\text{$t-m$ entries}}, \infty,\xi_{m},\dots,\xi_1)
\end{equation}
which corresponds to the original sequence read backwards.
Since the probability distribution of the sequence $\xi$ is invariant under permutations,
the augmented shape of the tableau \eqref{eq:transposed} can be viewed as
the value at time~$t$ of the augmented Plancherel growth process initiated at time $m':=t-m$.
The remaining difficulty is therefore to understand the probability distribution
of the augmented Plancherel growth process initiated at time $m'$, after
additional $m$ steps of Schensted row insertion were performed. We are interested in the asymptotic setting when
$m'\to\infty$ and the number of additional steps $m=O\big( \sqrt{m'} \big)$ is
relatively small. This is precisely the setting which was considered in our
recent paper about the Poisson limit theorem for the Plancherel growth process
\cite{MMS-Poisson2020v2}. We summarize these results in
\cref{sec:application-thm-poisson}; based on them we
prove in \cref{prop:distribution-fixed-time-A} that the index of the row of the
symbol $\infty$ in the tableau \eqref{eq:transposed} is asymptotically given by
the Poisson distribution.
By taking the transpose of the augmented Young diagrams we recover \cref{prop:distribution-fixed-time}, as desired.
\subsubsection{Averaged transition probabilities}
We will sketch the proof of the aforementioned \cref{prop:transition-augmented}
which concerns an augmented Plancherel growth process
\begin{equation}
\label{eq:pgp-atcolumnk}
\Lambda^{(n)} \nearrow \Lambda^{(n+1)} \nearrow \cdots
\end{equation}
for which the initial probability distribution at time $n$ is given by
$\Lambda^{(n)}= \big(\lambda^{(n)},\Box^{(n)} \big)$, where $\lambda^{(n)}$ is a random
Young diagram with $n$ boxes distributed (approximately) according to the
Plancherel measure and $\Box^{(n)}$ is its outer corner located in the column
with the fixed index $k$. Our goal is to describe the probability distribution of
this augmented Plancherel growth process at some later time $n'$, asymptotically as $n,n'\to\infty$.
Our first step in this direction is to approximate the probability distribution
of the Markov process \eqref{eq:pgp-atcolumnk} by a certain linear combination
(with real, positive and negative, coefficients) of the probability distributions
of augmented Plancherel growth processes initiated at time $m$. This linear combination is
taken over the values of $m$ which are of order $O\!\left( \sqrt{n} \right)$.
Finding such a linear combination required the results which we discussed above
in \cref{sec:at-given-time}, namely a~good understanding of the probability
distribution at time $n$ of the augmented Plancherel growth process initiated at some specified time $m=O\!\left( \sqrt{n} \right)$.
The probability distribution of $\Lambda^{(n')}$ is then approximately equal to
the aforementioned linear combination of the laws (this time evaluated at
time~$n'$) of the augmented Plancherel growth processes initiated at some specific
times~$m$. This linear combination is straightforward to analyze because for each
individual summand the results from \cref{sec:at-given-time} are applicable. This
completes the sketch of the proof of \cref{prop:transition-augmented}.
\section{Growth of the~bottom rows}
\label{sec:application-thm-poisson}
In the~current section we will gather some results and some notations from our
recent paper \cite[Section~2]{MMS-Poisson2020v2} which will be necessary for
the~purposes of the~current work.
\subsection{Total variation distance}
Suppose that $\mu$ and $\nu$ are (signed) measures on the~same discrete set $S$. Such
measures can~be identified with real-valued functions on $S$.
We define the~\emph{total
variation distance} between the~measures $\mu$ and $\nu$
\begin{equation}
\label{eq:TVD1}
\delta( \mu, \nu) := \frac{1}{2} \left\| \mu-\nu \right\|_{\ell^1}
\end{equation}
as half of their $\ell^1$ distance as functions.
If $X$ and $Y$ are two random variables with values in the~same discrete set
$S$, we define their total variation distance $\delta(X,Y)$ as the~total
variation distance between their probability distributions
(which are probability measures on $S$).
Usually in the~literature the~total variation distance is defined only for
probability measures. In such a~setup the~total variation distance can~be expressed as
\begin{equation}
\label{eq:TVD2}
\delta( \mu, \nu) = \max_{X\subset S} \left| \mu(X) - \nu(X) \right| =
\left\| (\mu-\nu)^+ \right\|_{\ell^1}.
\end{equation}
In the~current paper we will occasionally use the~notion of the~total variation
distance also for signed measures for which \eqref{eq:TVD1} and \eqref{eq:TVD2}
are \emph{not} equivalent.
\subsection{Growth of rows in Plancherel growth process}
\label{sec:plancherel-growth-process-independent}
Let \mbox{$\lambda^{(0)}\nearrow \lambda^{(1)} \nearrow \cdots$} be the~Plancherel growth
process. For integers $n\geq 1$ and $r\in \mathbb{N}_0$ we denote by $E^{(n)}_r$ the
random event which occurs if the~unique box of
the skew diagram $\lambda^{(n)} / \lambda^{(n-1)}$
is located in the~row with the~index $r$.
The following result was proved by Okounkov \cite[Proposition 2]{Okounkov2000},
see also \cite[Proposition~2.7]{MMS-Poisson2020v2} for an alternative proof.
\begin{proposition}
\label{prop:asymptotic-probability}
For each $r\in \mathbb{N}_0$
\[
\lim_{n\to\infty} \sqrt{n}\ \mathbb{P}\left( E_r^{(n)} \right) =1 .\]
\end{proposition}
\medskip
Let us fix an~integer $k\in\mathbb{N}_0$. We define $\mathcal{N}=\{0,1,\dots,k,\infty\}$.
For $n\geq 1$ we define the~random variable
$R^{(n)}\in \mathcal{N}$
which is given by
\[ R^{(n)} =
\begin{cases}
r & \text{if the~event $E^{(n)}_r$ occurs for $0\leq r\leq k$},\\
\infty & \text{if the~event $E^{(n)}_r$ occurs for some $r>k$}.
\end{cases}
\]
\medskip
\newcommand{n}{n}
Let $\ell=\ell(n)$ be a~sequence of non-negative integers such that
\[
\ell=O \! \left(\sqrt{n} \right).
\]
For a~given integer $n\geq (k+1)^2$ we focus on the
specific part of the~Plancherel growth process
\begin{equation}
\label{eq:focus}
\lambda^{(n)}\nearrow \cdots \nearrow\lambda^{(n+\ell)}.
\end{equation}
We will encode some partial information about the~growths of the~rows as well
as about the~final Young diagram in \eqref{eq:focus}
by the~random vector
\begin{equation}
\label{eq:vn0}
V^{(n)}=
\left(
R^{(n+1)}, \: \dots , \: R^{(n+\ell)}, \:
\lambda^{\left( n+\ell\right)} \right) \in \mathcal{N}^{\ell}
\times \mathbb{Y}.
\end{equation}
We also consider the~random vector
\begin{equation}
\label{eq:ovV-def}
\overline{V}^{(n)}=
\left(
\overline{R}^{(n+1)}, \: \dots , \: \overline{R}^{(n+\ell)}, \:
\overline{\lambda}^{\left( n+\ell\right)} \right) \in \mathcal{N}^{\ell}
\times \mathbb{Y}
\end{equation}
which is defined as a~sequence of independent random variables; the~random
variables $\overline{R}^{(n+1)},\dots,\overline{R}^{(n+\ell)}$ have the
same distribution given by
\begin{align*}
\mathbb{P}\left\{ \overline{R}^{(n+i)}= r \right\} &= \frac{1}{\sqrt{n}}
& \text{for }r\in\{0,\dots,k\}, \; 1\leq i\leq \ell,
\\
\mathbb{P}\left\{ \overline{R}^{(n+i)}= \infty \right\} &= 1- \frac{k+1}{\sqrt{n}}
\end{align*}
and $\overline{\lambda}^{\left( n+\ell\right)}$ is distributed according to the
Plancherel measure $\Plancherel_{n+\ell}$; in particular the random
variables ${\lambda}^{\left( n+\ell\right)}$ and $\overline{\lambda}^{\left(
n+\ell\right)}$ have the same distribution.
\medskip
Heuristically, the~following result states that when the Plancherel growth process
is in an~advanced stage and we observe a~relatively small number of its
additional steps, the~growths of the~bottom rows occur approximately like
independent random variables. Additionally, these growths do not affect too
much the~final shape of the~Young diagram.
\begin{theorem}[{\cite[Theorem~2.2]{MMS-Poisson2020v2}}]
\label{thm:independent-explicit}
With the~above notations, for each fixed $k\in\mathbb{N}_0$ the~total variation
distance between $V^{(n)}$ and $\overline{V}^{(n)}$ converges to
zero, as $n\to\infty$; more specifically
\[ \delta\left(V^{(n)}, \overline{V}^{(n)}\right) =
o\left(\frac{\ell}{\sqrt{n}}\right). \]
\end{theorem}
\section{Augmented Plancherel growth process}
\label{sec:augmented-plancherel}
In this section we will introduce our main tool: \emph{the~augmented Plancherel
growth process} which keeps track of the~position of a~very large number in
the~insertion tableau when new random numbers are inserted.
\subsection{Lazy parametrization of bumping routes}
Our first step towards the~proof of \cref{thm:poisson-point} is to introduce
a~more convenient parametrization of the~bumping routes. In
\eqref{eq:bumping_points} we used $y$, the~number of the~row, as the~variable
which parametrizes the~bumping route. In the~current section we will introduce
the~\emph{lazy parametrization}.
\medskip
Let us fix a~(finite or infinite) standard Young tableau $\mathcal{T}$
and an~integer $m\in\mathbb{N}_0$. For a~given integer $t\geq m$ we denote by
\[ \Box^{\lazy}_{\mathcal{T},m}(t) =
\left(x^{\lazy}_{\mathcal{T},m}(t), y^{\lazy}_{\mathcal{T},m}(t) \right)\]
the~coordinates of the~first box in the~bumping route $\mathcal{T} \leftsquigarrow
m+\nicefrac{1}{2}$ which contains an~entry of $\mathcal{T}$ which is bigger than $t$.
If such a box does not exists, this means that the bumping route is finite, and
all boxes of the tableau $\mathcal{T}$ which belong to the bumping route are $\leq t$. If this is the case we define
$ \Box^{\lazy}_{\mathcal{T},m}(t)$ to be the last box of the bumping route, i.e.~the box of the bumping route which lies outside of $\mathcal{T}$.
We will refer to
\begin{equation}
\label{eq:lazy}
t\mapsto \big(x^{\lazy}_{\mathcal{T},m}(t), y^{\lazy}_{\mathcal{T},m}(t) \big)
\end{equation}
as the~\emph{lazy parametrization of the~bumping route}.
For example, for the~infinite tableau $\mathcal{T}$ from \cref{fig:tableauFrench}
and $m=3$ the~usual parametrization of the~bumping route is given by
\begin{align*} \mathcal{T}_{\leftsquigarrow m+\nicefrac{1}{2}}(y) &=
\begin{cases}
3 & \text{for } y=0, \\
2 & \text{for } y=1, \\
1 & \text{for } y=2, \\
1 & \text{for } y=3, \\
0 & \text{for } y\geq 4,
\end{cases}
\\
\intertext{while its lazy counterpart is given by}
\Box^{\lazy}_{\mathcal{T},m}(t)=\left(x^{\lazy}_{\mathcal{T},m}(t), y^{\lazy}_{\mathcal{T},m}(t) \right) & =
\begin{cases}
(3,0) & \text{for } t\in\{3,4,5\}, \\
(2,1) & \text{for } t\in\{6,7,8\}, \\
(1,2) & \text{for } t=9, \\
(1,3) & \text{for } t\in\{10,11\}, \\
(0,4) & \text{for } t=12, \\
(0,5) & \text{for } t=13, \\
(0,6) & \text{for } t\in\{14,15,16\}, \\
\;\;\;\vdots &
\end{cases}
\end{align*}
Clearly, the~set of values of the~function \eqref{eq:lazy} coincides with the
bumping route understood in the~traditional way \eqref{eq:bumping_points}.
We denote by $\mathcal{T}|_{\leq t}$ the outcome of keeping only these boxes of
$\mathcal{T}$ which are at most $t$. Note that the element of the bumping route
\begin{equation}
\label{eq:bumping-explicit}
\Box^{\lazy}_{\mathcal{T},m}(t) =
\operatorname{sh} \big( \mathcal{T}|_{\leq t} \leftarrow m+\nicefrac{1}{2} \big)\; / \; \operatorname{sh} \left( \mathcal{T}|_{\leq t} \right)
\end{equation}
is the unique box of the difference of two Young diagrams on the right-hand side.
\subsection{Trajectory of $\infty$}
\label{sec:trajectory}
Let $\xi=(\xi_1,\xi_2,\dots)$ be a~sequence of independent, identically distributed
random variables with the~uniform distribution $U(0,1)$ on the~unit interval
$[0,1]$ and let $m\geq 0$ be a~fixed integer.
We will iteratively apply Schensted row insertion to
the~entries of the~infinite sequence
\[ \xi_1,\dots,\xi_m, \infty, \xi_{m+1}, \xi_{m+2}, \dots\]
which is the~initial sequence $\xi$ with our favorite symbol $\infty$
inserted at position~$m+1$.
(The~Readers who are afraid of infinity may replace
it by any number which is strictly bigger than~all of the~entries of the
sequence $\xi$.)
Our goal is to investigate the~position of the~box containing $\infty$ as a
function of the~number of iterations. More specifically, for an~integer $t\geq m$ we
define
\begin{equation}
\label{eq:trajectory}
\Box^{\traj}_m(t)=
\Pos_\infty \! \big( P(\xi_1,\dots,\xi_m, \infty, \xi_{m+1}, \dots, \xi_t ) \big)
\end{equation}
to be the~position of the~box containing $\infty$ in the~appropriate insertion
tableau.
This problem was formulated by Duzhin \cite{Duzhin2019}; the~first
asymptotic results in the~scaling in which $m\to\infty$ and $t=O(m)$ were found
by the~first named author~\cite{Marciniak2020}. In the~current paper we go beyond this
scaling and consider \mbox{$m\to\infty$} and $t=O \! \left(m^2\right)$;
the~answer for this problem is essentially contained in
\cref{thm:lazy-poisson}.
The~following result shows a~direct link between the~above
problem and the~asymptotics of bumping routes. This result also shows an
interesting link between the~papers \cite{RomikSniadyBumping} and
\cite{Marciniak2020}.
\begin{proposition}
\label{prop:lazy-traj-correspondence}
Let $\xi_1,\xi_2,\dots$ be a~(non-random or random) sequence and
$\mathcal{T}=Q(\xi_1,\xi_2,\dots)$ be the~corresponding recording tableau.
Then for each $m\in\mathbb{N}$ the~bumping route in the~lazy parametrization
coincides with the~trajectory of $\infty$ as defined in
\eqref{eq:trajectory}:
\begin{equation}
\label{eq:link}
\Box^{\lazy}_{\mathcal{T},m}(t) = \Box^{\traj}_m(t) \qquad \text{for each integer } t\geq m.
\end{equation}
\end{proposition}
We will provide two proofs of \cref{prop:lazy-traj-correspondence}. The first one
is based on the following classic result of Sch\"utzengerger.
\begin{fact}[{\cite{Schuetzenberger1963}}]
\label{fact}
For any permutation $\sigma$ the
insertion tableau $P(\sigma)$ and the recording
tableau $Q(\sigma^{-1})$, which corresponds to the inverse of $\sigma$,
are equal.
\end{fact}
\begin{proof}[The first proof of \cref{prop:lazy-traj-correspondence}.]
Let $\pi=(\pi_1, \dots, \pi_t)\in\Sym{t}$
be the \emph{permutation generated by the sequence
$(\xi_1, \dots,\xi_t)$}, that is the unique
permutation such that for any choice of indices
$i<j$ the condition $\pi_i<\pi_j$ holds true if
and only if $\xi_i \leq \xi_j$.
Let $\pi^{-1}=(\pi_1^{-1}, \dots,\pi^{-1}_t)$
be the inverse of $\pi$.
Since RSK depends only on the relative order of entries,
the restricted tableau $\mathcal{T}|_{\leq t}$ is equal to
\begin{equation}
\label{eq:Marciniak-equality}
\mathcal{T}|_{\leq t} =Q(\xi_1,\dots, \xi_t)=Q(\pi)=P(\pi^{-1}).
\end{equation}
By \eqref{eq:bumping-explicit}, \eqref{eq:Marciniak-equality} and \cref{fact}
it follows that
\begin{align*}
\Box^{\lazy}_{\mathcal{T},m}(t)&=
\operatorname{sh} \left( P(\pi^{-1}) \leftarrow m+\nicefrac{1}{2} \right)\; /\; \operatorname{sh} P(\pi^{-1}) \\
&=
\operatorname{sh} P\left(\pi^{-1}_1,\dots,\pi^{-1}_t,m+\nicefrac{1}{2} \right)
\; /\;
\operatorname{sh} P\left(\pi^{-1}_1,\dots,\pi^{-1}_t\right) \\
&=\Pos_{t+1}\left(Q\left(\pi_1^{-1}, \dots, \pi_t^{-1}, m+\nicefrac{1}{2}\right)\right)\\
&\stackrel{\text{\cref{fact}}}{=}\Pos_{t+1} \left(P(\pi_1, \dots, \pi_{m}, t+1, \pi_{m+1}, \dots, \pi_t)\right) \\
&=\Pos_{\infty} \left( P(\xi_1, \dots, \xi_{m}, \infty, \xi_{m+1}, \dots, \xi_t) \right) \\
&=\Box^{\traj}_m(t)
\end{align*}
since the permutation $(\pi_1, \dots, \pi_{m}, t+1, \pi_{m+1},
\dots, \pi_t)$ is the inverse of the permutation generated by the sequence
$(\pi_1^{-1}, \dots, \pi_t^{-1}, m+\nicefrac{1}{2})$.
\end{proof}
The above proof has an advantage of being short and abstract.
The following alternative proof highlights the `dynamic' aspects
of the bumping routes and the trajectory of infinity.
\begin{proof}[The second proof of \cref{prop:lazy-traj-correspondence}]
We use induction over the~variable $t$.
\medskip
The~induction base $t=m$ is quite easy: $\Box^{\lazy}_{\mathcal{T},m}(m)$ is
the~leftmost box in the~bottom row of $\mathcal{T}$ which contains a~number which is
bigger than~$m$. This box is the~first to the~right of the~last box in the~bottom
row in the~tableau $Q(\xi_1, \ldots, \xi_m)$. On the other hand, since this
recording tableau has the~same shape as the insertion tableau $P(\xi_1, \ldots,
\xi_m)$, it follows that $\Box^{\traj}_m(m) = \Box^{\lazy}_{\mathcal{T},m}(m)$ and
the proof of the induction base is completed.
\medskip
We start with an~observation that $\infty$ is bumped in the~process of
calculating the~row insertion
\begin{equation}
\label{eq:indution}
P(\xi_1,\dots,\xi_m, \infty, \xi_{m+1}, \dots, \xi_t ) \leftarrow \xi_{t+1}
\end{equation}
if and only if the~position of $\infty$ at time $t$,
that is $\Box^{\traj}_m(t)$,
is the~unique box which belongs to the skew diagram
\[
\RSK(\xi_1,\dots, \xi_{t+1}) /
\RSK(\xi_1,\dots, \xi_t ). \]
The~latter condition holds true if and only if
the~entry of $\mathcal{T}$ located in the~box
$ \Box^{\traj}_m(t)$ fulfills
\[ \mathcal{T}_{\Box^{\traj}_m(t)} = t+1.\]
\medskip
In order to make the~induction step we assume that the~equality
\eqref{eq:link} holds true for some $t\geq m$. There are the~following two
cases.
\smallskip
\emph{Case 1. Assume that
the~entry of $\mathcal{T}$ located in the~box
$\Box^{\lazy}_{\mathcal{T},m}(t)$ is strictly bigger than~$t+1$.}
In this case the~lazy bumping route stays put and
\[ \Box^{\lazy}_{\mathcal{T},m}(t+1) = \Box^{\lazy}_{\mathcal{T},m}(t).\]
By the~induction hypothesis, the~entry of $\mathcal{T}$ located in the~box $
\Box^{\traj}_m(t)= \Box^{\lazy}_{\mathcal{T},m}(t)$ is bigger than~$t+1$. By
the~previous discussion, $\infty$ is not bumped in the~process of
calculating the~row insertion \eqref{eq:indution} hence
\[ \Box^{\traj}_m(t+1) = \Box^{\traj}_m(t) \]
and the~inductive step holds true.
\smallskip
\emph{Case 2. Assume that
the~entry of $\mathcal{T}$ located in the~box $\Box^{\lazy}_{\mathcal{T},m}(t)$ is
equal to $t+1$.} In this case the~lazy bumping route moves to the~next row.
It~follows that $\Box^{\lazy}_{\mathcal{T},m}(t+1)$ is the~leftmost box of $\mathcal{T}$ in the
row above $\Box^{\lazy}_{\mathcal{T},m}(t)$ which contains a~number which is
bigger than~$\mathcal{T}_{\Box^{\lazy}_{\mathcal{T},m}(t)} = t+1$.
By the~induction hypothesis, $\mathcal{T}_{\Box^{\traj}_m(t)} =
\mathcal{T}_{\Box^{\lazy}_{\mathcal{T},m}(t)} = t+1$, so $\infty$ is bumped in the
process of calculating the~row insertion \eqref{eq:indution} to the~next row~$r$.
The~box $\Box^{\traj}_m(t+1)$ is the~first to the~right of the~last box in
the~row~$r$ in $\RSK(\xi_1, \dots, \xi_t, \xi_{t+1})$. Clearly, this is the~box
in the~row $r$ of $\mathcal{T}$ which has the~least entry among those which are
bigger than~$t+1$, so it is the~same as~$\Box^{\lazy}_{\mathcal{T},m}(t+1)$.
\end{proof}
\subsection{Augmented Young diagrams. Augmented shape of a tableau}
\label{sec:augmented-Young-diagrams}
For the motivations and heuristics behind the notion of augmented Young diagrams see \cref{sec:intro-augmented}.
A pair $\Lambda=(\lambda,\Box)$ will be called an \emph{augmented Young diagram} if $\lambda$ is a~Young diagram and $\Box$ is one of its outer corners,
see \cref{fig:tableau-after}. We will say that $\lambda$ is \emph{the regular part of $\Lambda$} and that $\Box$ is \emph{the special box of $\Lambda$}.
The~set of augmented Young diagrams will be denoted
by $\diagrams^*$ and for $n\in\mathbb{N}_0$
we will denote by $\diagrams^*_n$ the set of
augmented Young diagrams~$(\lambda,\Box)$ with the~additional
property that $\lambda$ has $n$ boxes
(which we will shortly denote by $|\lambda|=n$).
\medskip
Suppose $\mathcal{T}$ is a tableau with the property that exactly one of its entries is equal to $\infty$. We define the \emph{augmented shape of $\mathcal{T}$}
\[ \operatorname{sh}^* \mathcal{T} = \left( \operatorname{sh} \left( \mathcal{T} \setminus \{ \infty\} \right), \Pos_{\infty} \mathcal{T} \right) \in \diagrams^*,\]
as the pair which consists of (a) the shape of $\mathcal{T}$ after removal of the box with~$\infty$, and (b) the location of the box with~$\infty$ in~$\mathcal{T}$,
see \cref{fig:augmented-yd}.
\subsection{Augmented Young graph}
\label{sec:augmented-Young-graph}
The~set $\diagrams^*$ can~be equipped with a~structure of an~oriented graph,
called \emph{augmented Young graph}.
We declare that a~pair
$\Lambda\nearrow \widetilde{\Lambda}$ forms an~oriented edge
(with $\Lambda=(\lambda,\Box)\in\diagrams^*$ and $\widetilde{\Lambda}=(\widetilde{\lambda},\widetilde{\Box})\in\diagrams^*$)
if the~following two conditions hold true:
\begin{equation}
\label{eq:augmented-edge}
\lambda\nearrow\widetilde{\lambda} \ \ \text{and} \ \
\widetilde{\Box} = \begin{cases}
\text{the~outer corner of $\widetilde{\lambda}$} & \\ \quad \text{which is in the~row above $\Box$}
& \text{if }\widetilde{\lambda}/\lambda= \left\{ \Box \right\}, \\[1.5ex]
\Box & \text{otherwise,}
\end{cases}
\end{equation}
see \cref{fig:augmented-yd} for an~illustration. If $\Lambda\nearrow
\widetilde{\Lambda}$ (with $\Lambda=(\lambda,\Box)\in\diagrams^*$ and
$\widetilde{\Lambda}=(\widetilde{\lambda},\widetilde{\Box})\in\diagrams^*$) are
such that $\Box\neq \widetilde{\Box}$ (which corresponds to the first case on the
right-hand side of \eqref{eq:augmented-edge}), we will say that \emph{the
edge $\Lambda\nearrow\widetilde{\Lambda}$ is a bump}.
\subfile{figures-Poisson/augmented-Young-diagram.tex}
The above definition was specifically tailored so that the following simple lemma holds true.
\begin{lemma}
\label{lem:why-the-bump}
Assume that $\mathcal{T}$ is a tableau which has exactly one entry equal to $\infty$ and let $x$ be some finite number. Then
\[ \left( \operatorname{sh}^* \mathcal{T} \right) \nearrow \left( \operatorname{sh}^* \! \left( \mathcal{T} \leftarrow x\right) \right). \]
\end{lemma}
\begin{proof}
Let $\mathcal{T}':=\mathcal{T}/\{\infty\}$ be the tableau $\mathcal{T}$ with the box containing $\infty$ removed.
Denote $(\lambda,\Box)=\operatorname{sh}^* \mathcal{T}$ and
$(\widetilde{\lambda},\widetilde{\Box})=
\operatorname{sh}^* \! \left( \mathcal{T} \leftarrow x\right)$;
their regular parts
\begin{align*}
\lambda &=\operatorname{sh} \mathcal{T}' , \\
\widetilde{\lambda} &=\operatorname{sh} \left( \mathcal{T}' \leftarrow x \right)
\end{align*}
clearly fulfill $\lambda\nearrow\widetilde{\lambda}$.
The position $\widetilde{\Box}$ of $\infty$ in $\mathcal{T}\leftarrow x$ is either:
\begin{itemize}
\item in the row immediately
above the position $\Box$ of $\infty$ in $\mathcal{T}$ (this happens exactly if
$\infty$ was bumped in the insertion $\mathcal{T}\leftarrow x$; equivalently if $\widetilde{\lambda}/\lambda=\{
\Box\}$), or
\item the same as the position $\Box$ of $\infty$ in $\mathcal{T}$ (this happens
exactly when $\infty$ was not bumped; equivalently if
$\widetilde{\lambda}/\lambda\neq\{ \Box\}$).
\end{itemize}
Clearly these two cases correspond to the second condition in
\eqref{eq:augmented-edge} which completes the proof.
\end{proof}
\newcommand{m}{m}
\subsection{Lifting of paths}
\label{sec:lifting}
We consider the~\emph{`covering map'} $p:\diagrams^*\to\mathbb{Y}$ given by taking
the regular part
\[ \diagrams^* \ni (\lambda, \Box) \xmapsto{p} \lambda\in \mathbb{Y}. \]
\begin{lemma}
\label{lem:lifting}
For any $\Lambda^{(m)}\in\diagrams^*$ and
any path in the~Young graph
\begin{equation}
\label{eq:path}
\lambda^{(m)}\nearrow \lambda^{(m+1)} \nearrow \cdots \in \mathbb{Y}
\end{equation}
with a~specified initial element $\lambda^{(m)}=p\big( \Lambda^{(m)} \big) $
there exists the~unique \emph{lifted path}
\[ \Lambda^{(m)}\nearrow \Lambda^{(m+1)} \nearrow \cdots \in \diagrams^* \]
in the~augmented Young graph
with the~specified initial element $\Lambda^{(m)}$, and
such that
$\lambda^{(t)}=p \big( \Lambda^{(t)} \big) $
holds true for each $t\in\{m,m+1,\dots\}$.
\end{lemma}
\begin{proof}
From \eqref{eq:augmented-edge} it follows that for each
$(\lambda,\Box)\in\diagrams^*$ and each $\widetilde{\lambda}$ such that
$\lambda\nearrow \widetilde{\lambda}$ there exists a unique
$\widetilde{\Box}$ such that
$(\lambda,\Box)\nearrow(\widetilde{\lambda},\widetilde{\Box})$. This shows
that, given $\Lambda^{(i)}$, the value of $\Lambda^{(i+1)}$ is determined
uniquely. This observation implies that the lemma can be proved by a
straightforward induction.
\end{proof}
\subfile{figures-Poisson/augmented-young-graph.tex}
\subsection{Augmented Plancherel growth process}
\label{sec:aPgp}
We keep the~notations from the beginning of \cref{sec:trajectory}, i.e., we
assume that $\xi=(\xi_1,\xi_2,\dots)$ is a~sequence of independent, identically
distributed random variables with the~uniform distribution $U(0,1)$ on the~unit
interval $[0,1]$ and $m\geq 0$ is a~fixed integer. We consider a~path in
the~augmented Young graph
\begin{equation}
\label{eq:Markov}
\Lambda_m^{(m)}\nearrow \Lambda_m^{(m+1)} \nearrow \cdots
\end{equation}
given by
\[ \Lambda_m^{(t)} = \operatorname{sh}^* P(\xi_1,\dots,\xi_m,\infty,\xi_{m+1},\dots,\xi_t) \qquad \text{for any integer $t\geq m$}\]
(\cref{lem:why-the-bump} shows that \eqref{eq:Markov} is indeed a path in $\diagrams^*$).
We will call \eqref{eq:Markov} \emph{the augmented Plancherel growth process initiated at time~$m$}.
The coordinates of the special box of $\Lambda_m^{(t)}=\big( \lambda^{(t)},\Box_m^{(t)}\big)$ will be denoted by
\[
\Box_m^{(t)}=\big( x_m^{(t)},y_m^{(t)} \big).
\]
\begin{theorem}
The augmented Plancherel growth process initiated at time~$m$ is a Markov chain with the transition probabilities given for any $t\geq m$ by
\begin{multline}
\label{eq:transition-probability-augmented}
\PPcond{ \Lambda_m^{(t+1)} =\widetilde{\Lambda} }{ \Lambda_m^{(t)} = \Lambda } =\\
\begin{cases}
\PPcond{ \lambda^{(t+1)}=\widetilde{\lambda} }{ \lambda^{(t)}=\lambda~}
& \text{if } \Lambda \nearrow \widetilde{\Lambda},
\\
0 & \text{otherwise}
\end{cases}
\end{multline}
for any $\Lambda,\widetilde{\Lambda}\in\diagrams^*$, where $\lambda$ is the
regular part of $\Lambda$ and $\widetilde{\lambda}$ is the regular part of
$\widetilde{\Lambda}$. These transition probabilities do not depend on the choice
of $m$. The conditional probability on the right-hand side is the transition
probability for the Plancherel growth process
$\lambda^{(0)}\nearrow\lambda^{(1)}\nearrow\cdots$.
\end{theorem}
\begin{proof}
The~path \eqref{eq:Markov}
is the unique lifting (cf.~\cref{lem:lifting}) of the~sequence of the~regular parts
\begin{equation}
\label{eq:lambdas}
\lambda^{(m)}\nearrow \lambda^{(m+1)} \nearrow \cdots
\end{equation}
with the~initial condition that the special box $\Box_m^{(m)}$ is the outer
corner of $\lambda^{(m)}$ which is located in the bottom row.
It follows that for any augmented Young diagrams $\Sigma_{m},\dots,\Sigma_{t+1}\in\diagrams^*$ with the regular parts $\sigma_m,\dots,\sigma_{t+1}\in \mathbb{Y}$
\begin{multline}
\label{eq:probability-prefix}
\mathbb{P}\left( \Lambda_m^{(m)} = \Sigma_{m}, \; \dots, \; \Lambda_m^{(t+1)}= \Sigma_{t+1} \right) = \\
= \begin{cases}
\mathbb{P}\left( \lambda^{(m)} = \sigma_m, \; \dots ,\; \lambda^{(t+1)}=\sigma_{t+1} \right) & \text{if } \Sigma_{m}\nearrow \cdots \nearrow \Sigma_{t+1} \text{ and} \\ & \text{\quad the special box of $\Sigma_m$} \\ & \quad \text{is in the bottom row}, \\[1.5ex]
0 & \text{otherwise}.
\end{cases}
\end{multline}
The sequence of the regular parts \eqref{eq:lambdas}
forms the~usual Plancherel growth process (with the~first
$m$ entries truncated) hence it is a~Markov chain (the proof that the usual
Plancherel growth process is a Markov chain can be found in \cite[Sections~2.2
and 2.4]{Kerov1999}).
It follows that the probability on the top of the
right-hand side of \eqref{eq:probability-prefix} can be written in the product
form in terms of the probability distribution of $\lambda^{(m)}$ and the
transition probabilities for the Plancherel growth process.
We compare \eqref{eq:probability-prefix} with its counterpart for $t:=t-1$;
this shows that the conditional probability
\[\PPcond{\Lambda_{m}^{(t+1)}=\Sigma_{t+1}}{\Lambda_m^{(m)} = \Sigma_{m}, \; \dots, \; \Lambda_m^{(t)}= \Sigma_{t}} \]
is equal to the right-hand side of \eqref{eq:transition-probability-augmented} for $\widetilde{\Lambda}:=\Sigma_{t+1}$ and $\Lambda:=\Sigma_t$.
In particular, this conditional probability does not depend on the values of $\Sigma_{m},\dots,\Sigma_{t-1}$ and the Markov property follows.
\end{proof}
The special box in the augmented Plancherel growth process can be thought of as a
\emph{test particle} which provides some information about the local behavior of the
usual Plancherel growth process. From this point of view it is reminiscent of the
\emph{second class particle} in the theory of interacting particle systems or
\emph{jeu de taquin trajectory} for infinite tableaux \cite{RomikSniady-AnnPro}.
\subsection{Probability distribution of the~augmented Plancherel growth process}
\cref{prop:distribution-fixed-time-A} and \cref{prop:distribution-fixed-time}
below provide information about the~probability distribution of the~augmented
Plancherel growth process at time $t$ for $t\to\infty$ in two distinct
asymptotic regimes: very soon after the~augmented Plancherel process was
initiated (that is when
$t=m+O(\sqrt{m})$, cf.~\cref{prop:distribution-fixed-time-A}) and after a
very long time after the~augmented Plancherel process was initiated
(that is~when $t=\Theta(m^2) \gg m$, cf.~\cref{prop:distribution-fixed-time}).
\begin{proposition}
\label{prop:distribution-fixed-time-A} Let $z>0$ be a~fixed positive number and
let $t=t(m)$ be a~sequence of positive integers such that $t(m) \geq m$ and
with the~property that
\[ \lim_{m\to\infty} \frac{t-m}{\sqrt{t}} = z. \]
Let $\Lambda_m^{(m)}\nearrow \Lambda_m^{(m+1)} \nearrow \cdots$
be the augmented Plancherel growth process initiated at time~$m$.
We denote $\Lambda_m^{(t)}=\big(\lambda^{(t)}, \Box_m^{(t)}\big)$;
let $\Box_m^{(t)}=\big(x_m^{(t)}, y_m^{(t)}\big)$ be the
coordinates of the special box of $\Lambda_m^{(t)}$.
\begin{enumerate}[label=\emph{\alph*}),itemsep=1ex]
\item \label{item:distribution-fixed-time-A-a}
The~probability distribution
of $y_{m}^{(t)}$ converges, as $m\to\infty$, to the Poisson distribution $\operatorname{Pois}(z)$
with parameter $z$.
\item \label{item:distribution-fixed-time-A-b}
For each $k\in\mathbb{N}_0$ the~total
variation distance between
\begin{itemize}
\item the~conditional probability distribution of $\lambda^{(t)}$ under
the~condition that $y_{m}^{(t)}=k$, and \item the~Plancherel measure
$\Plancherel_t$
\end{itemize}
converges to $0$, as $m\to\infty$.
\item \label{item:distribution-fixed-time-A-c}
The~total variation distance between
\begin{itemize}
\item
the~probability distribution of the~random vector
\begin{equation}
\label{eq:ymlambda}
\left( \lambda^{(t)}, y_{m}^{(t)} \right) \in \mathbb{Y} \times \mathbb{N}_0
\end{equation}
and
\item
the~product measure
\[ \Plancherel_t \times \operatorname{Pois}(z) \]
\end{itemize}
converges to $0$, as $m\to\infty$.
\end{enumerate}
\end{proposition}
Let us fix an~integer $k\geq 0$.
We use the~notations from
\cref{sec:plancherel-growth-process-independent} for $n:=m$ and $\ell=t-m$ so
that $n+\ell=t$; we assume that $m$ is big enough so that $m\geq (k+1)^2$. Our general strategy is to read the~required information
from the~vector $V^{(n)}$ given by \eqref{eq:vn0} and to apply
\cref{thm:independent-explicit}. Before the proof of \cref{prop:distribution-fixed-time-A} we start with the following auxiliary result.
\medskip
For $s\geq m$ we define the random variable $y_{m}^{(s)}\!\downarrow_\mathcal{N} \: \in \! \mathcal{N}$ by
\begin{align*} y_{m}^{(s)}\!\downarrow_\mathcal{N} &=
\begin{cases}
y_{m}^{(s)} & \text{if } y_{m}^{(s)}\in\{0,1,\dots,k\}, \\
\infty & \text{otherwise.}
\end{cases}\\
\intertext{We also define the random variable $F_{s}\in\{0,\dots,k\}$ by}
F_{s} &=
\begin{cases}
y_{m}^{(s)}\!\downarrow_\mathcal{N} & \text{if } y_{m}^{(s)}\!\downarrow_\mathcal{N} \neq \infty,\\
\text{arbitrary element of $\{0,\dots,k\}$} & \text{otherwise}.
\end{cases}
\end{align*}
\begin{lemma}
\label{lem:past}
For each $s\geq m$ the value of $F_s$ can~be
expressed as an~explicit function of the~entries of the~sequence $R$ related to the~past, that is
\[
R^{(m+1)}, \: \dots, \: R^{(s)}.
\]
\smallskip
For any integer $p\in\{0,\dots,k\}$
the equality $y_m^{(s)}= p$ holds true if and only if
there are exactly $p$ values of the~index
$u\in\{m+1,\dots,s\}$ with the~property that
\begin{equation}
\label{eq:Bernoulli}
R^{(u)}= F_{u-1}.
\end{equation}
The inequality $y_m^{(s)}>k$ holds if and only if there are at least $k+1$ values of the index \mbox{$u
\in\{m+1,\dots,s\}$} with this property.
\end{lemma}
\begin{proof}
There are \emph{exactly} $y_m^{(s)}$ edges which are bumps in the path
\begin{equation}\label{eq:golden-path}
\Lambda^{(m)} \nearrow \cdots \nearrow \Lambda^{(s)}
\end{equation}
because each bump increases the $y$-coordinate of the special box by $1$.
Note that an edge $\Lambda^{(u-1)}\nearrow \Lambda^{(u)}$ in this path is a bump if and only if
\begin{equation}\label{eq:event}
\text{the event $E^{(u)}_r$ occurs for the row $r=y_m^{(u-1)}$.}
\end{equation}
If $y_m^{(s)}\leq k$ then for any $u\in\{m+1,\dots,s\}$ the equality $F_u=y_m^{(u)}$
holds true; furthermore the event \eqref{eq:event} occurs if and only if
$R^{(u)}=y_m^{(u-1)}$. It~follows that there are \emph{exactly} $y_m^{(s)}$ values of the~index \mbox{$u
\in\{m+1,\dots,s\}$} such that
\eqref{eq:Bernoulli} holds true.
On the~other hand, if $y_m^{(s)}>k$ we can apply the above reasoning to the
truncation of the path \eqref{eq:golden-path} until after the $(k+1)$-st bump occurs. It follows that in this case there are \emph{at least} $k+1$
values of the~index $u\in\{m+1,\dots,s\}$ with the~property~\eqref{eq:Bernoulli}.
In this way we proved the second part of the lemma.
\medskip
By the second part of the lemma, the~value of $y_m^{(s)}\!\downarrow_\mathcal{N}$ can~be
expressed as an~explicit function of both (i) the~previous values
\begin{equation}
\label{eq:previous}
y_m^{(m)}\!\downarrow_\mathcal{N},\: \dots, \: y_m^{(s-1)}\!\downarrow_\mathcal{N},
\end{equation}
\emph{and} (ii) the~entries of the~sequence $R$ related to the~past, that is
\begin{equation}
\label{eq:previousR}
R^{(m+1)}, \: \dots, \: R^{(s)}.
\end{equation}
By iteratively applying this observation to the~previous values
\eqref{eq:previous} it is possible to express the~value of
$y_m^{(s)}\!\downarrow_\mathcal{N}$ \emph{purely}
in terms of \eqref{eq:previousR}.
Also the~value of
\[F_{s} = F_{s}\left(
R^{(m+1)}, \: \dots, \: R^{(s)}\right)\]
can~be expressed as a~function of the~entries of the~sequence $R$ related to~the~past, as required.
\end{proof}
\newcommand{A}{A}
\begin{proof}[Proof of \cref{prop:distribution-fixed-time-A}]
\cref{lem:past} shows that the~event $y_{m}^{(t)}=k$ can~be expressed in
terms of the~vector $V^{(n)}$ given by \eqref{eq:vn0}. We apply
\cref{thm:independent-explicit}; it follows that the~probability $\mathbb{P}\left\{
y_{m}^{(t)}=k \right\}$ is equal, up to an~additive error term $o(1)$,
to~the~probability that there are exactly
$k$ values of the~index \mbox{$u\in\{m+1,\dots,t\}$} with the~property that
\begin{equation}
\label{eq:Bernoulli2}
\overline{R}^{(u)}= F_{u-1}\left(
\overline{R}^{(m+1)},\dots,\overline{R}^{(u-1)}\right).
\end{equation}
We denote by $A_u$ the random event that the equality \eqref{eq:Bernoulli2} holds true.
\medskip
Let $i_1<\dots<i_{l}$ be an increasing sequence of integers from the set $\{m+1, \dots, t\}$ for $l\geq 1$.
We will show that
\begin{equation}
\label{eq:independence-day}
\mathbb{P}\left( A_{i_1} \cap \cdots \cap A_{i_{l}} \right) = \frac{1}{\sqrt{m}}\ \mathbb{P}\left( A_{i_1} \cap \cdots \cap A_{i_{l-1}} \right).
\end{equation}
Indeed, by \cref{lem:past}, the event $A_{i_1} \cap \cdots \cap A_{i_{l-1}}$ is a disjoint finite union of some random events
of the form
\[ B_{r_{m+1},\dots,r_{j}}=
\left\{ \overline{R}^{(m+1)}=r_{m+1},\; \overline{R}^{(m+2)}=r_{m+2},\; \dots ,\; \overline{R}^{(j)}=r_{j} \right\} \]
over some choices of $r_{m+1},r_{m+2},\dots,r_{j}\in \mathcal{N}$, where $j:=i_l-1$.
Since the random variables $\big( \overline{R}^{(i)} \big)$ are independent, it follows that
\begin{equation}
\label{eq:problem0101}
\mathbb{P} \left( B_{r_{m+1},\dots,r_{j}} \cap A_{i_l} \right) = \frac{1}{\sqrt{m}}\ \mathbb{P} \left( B_{r_{m+1},\dots,r_{j}} \right).
\end{equation}
By summing over the appropriate values of $r_{m+1},\dots,r_{j}\in
\mathcal{N}$ the equality~\eqref{eq:independence-day} follows.
By iterating \eqref{eq:independence-day} it follows that the events $A_{m+1},\dots, A_{t}$ are independent
and each has equal probability $\frac{1}{\sqrt{m}}$.
\medskip
By the Poisson limit theorem \cite[Theorem~3.6.1]{Durrett2010}
the~probability of $k$ successes in $\ell$ Bernoulli
trials as above converges to the~probability of the~atom $k$ in the Poisson
distribution with the~intensity parameter equal to
\[ \lim_{m \to\infty} \frac{\ell}{\sqrt{m}} =
\lim_{m \to\infty} \frac{t-m}{\sqrt{t}}=
z\]
which concludes
the~proof of part \ref{item:distribution-fixed-time-A-a}.
\medskip
The~above discussion also shows that the~conditional probability distribution
considered in point \ref{item:distribution-fixed-time-A-b} is equal to the
conditional probability distribution of the~last coordinate $\lambda^{(t)}$ of
the~vector $V^{(n)}$ under certain condition which is expressed in terms of
the~coordinates $R^{(m+1)},\dots,R^{(t)}$.
By~\cref{thm:independent-explicit} this conditional probability distribution is in
the~distance $o(1)$ (with respect to the~total variation distance) to its
counterpart for the~random vector $\overline{V}^{(n)}$. The~latter
conditional probability distribution, due to the~independence of the
coordinates of $\overline{V}^{(n)}$, is equal to the~Plancherel measure
$\Plancherel_t$, which concludes the~proof
of~\ref{item:distribution-fixed-time-A-b}.
\medskip
Part \ref{item:distribution-fixed-time-A-c} is a~direct consequence of parts
\ref{item:distribution-fixed-time-A-a} and
\ref{item:distribution-fixed-time-A-b}.
\end{proof}
For an augmented Young diagram $\Lambda=\big(\lambda, (x,y) \big)$ we define its
\emph{transpose} $\Lambda^T=\big( \lambda^T, (y,x) \big)$.
\begin{lemma}
\label{lem:transpose}
For any integers $m,m'\geq 0$ the probability distributions at time $t=m+m'$ of the augmented Plancherel growth processes initiated at times $m$ and $m'$ respectively are related by
\[ \Lambda_m^{(t)} \overset{d}{=} \left[ \Lambda_{m'}^{(t)} \right]^T. \]
\end{lemma}
\begin{proof}
Without loss of generality we may assume that the random variables
$\xi_1,\dots,\xi_t$ are \emph{distinct} real numbers. An application of
Greene's theorem \cite[Theorem~3.1]{Greene1974} shows that the insertion tableaux
which correspond to a given sequence of distinct numbers and this sequence
read backwards
\begin{multline*} P(\xi_1,\dots,\xi_m,\infty,\xi_{m+1},\dots,\xi_t) = \\
\left[ P( \xi_t,\xi_{t-1},\dots,\xi_{m+1}, \infty, \xi_{m},\xi_{m-1}, \dots, \xi_1 ) \right]^T
\end{multline*}
are transposes of one another. It follows that also the augmented shapes are transposes of one another:
\begin{multline*}
\Lambda_m^{(t)}=
\operatorname{sh}^* P(\xi_1,\dots,\xi_m,\infty,\xi_{m+1},\dots,\xi_t) = \\
\Big[ \operatorname{sh}^* P( \underbrace{ \xi_t,\xi_{t-1},\dots,\xi_{m+1}}_{\text{$m'$ entries}}, \infty, \xi_{m},\xi_{m-1}, \dots, \xi_1 ) \Big]^T.
\end{multline*}
Since the sequence $(\xi_i)$ and its any permutation $\left(\xi_{\sigma(i)} \right)$ have the same distributions, the right-hand side has
the same probability distribution as $\left[ \Lambda_{m'}^{(t)} \right]^T$, as
required.
\end{proof}
\begin{proposition}
\label{prop:distribution-fixed-time}
Let $z>0$ be a~fixed real number.
Let $t=t(m)$ be a~sequence of positive integers such that $t(m) \geq m$ and
with the~property that
\[ \lim_{m\to\infty} \frac{m}{\sqrt{t}} = z. \]
Let $\Lambda_m^{(m)}\nearrow \Lambda_m^{(m+1)} \nearrow \cdots$ be the augmented Plancherel growth process initiated at time~$m$.
We denote $\Lambda_m^{(t)}=\big(\lambda^{(t)},
\Box_m^{(t)}\big)$; let $\Box_m^{(t)}=\big(x_m^{(t)}, y_m^{(t)}\big)$ be the
coordinates of the special box at time $t$.
The~total variation distance between
\begin{itemize}
\item
the~probability distribution of the~random vector
\begin{equation}
\label{eq:xmlambda}
\left( x_m^{(t)}, \lambda^{(t)} \right) \in \mathbb{N}_0\times \mathbb{Y}
\end{equation}
and
\item
the~product measure
\[ \operatorname{Pois}(z) \times \Plancherel_t \]
\end{itemize}
converges to $0$, as $m\to\infty$.
\end{proposition}
\begin{proof}
By \cref{lem:transpose} the probability distribution of \eqref{eq:xmlambda} coincides with the probability distribution of \begin{equation}
\label{eq:magictranspose}
\left( y_{m'}^{(t)}, \big[ \lambda^{(t)} \big]^T \right)
\end{equation} for $m':=t-m$.
The random vector \eqref{eq:magictranspose} can be viewed as the image of the vector $ \big( y_{m'}^{(t)}, \lambda^{(t)} \big)$ under the bijection
\[\operatorname{id}\times T: (y,\lambda)\mapsto (y,\lambda^T).\]
By \cref{prop:distribution-fixed-time-A} it follows that
the total variation distance
between \eqref{eq:magictranspose} and the push-forward measure
\[ \left(\operatorname{id}\times T \right)\left( \operatorname{Pois}(z) \times \Plancherel_t \right) = \operatorname{Pois}(z) \times \Plancherel_t \]
converges to zero as $m\to\infty$;
the last equality holds since the Plancherel measure is invariant under transposition.
\end{proof}
\subsection{Lazy version of \cref{prop:m=1}. Proof of \cref{prop:is-finite}}
\label{sec:proof-prop:is-finite}
In \cref{sec:in-which-row} we parametrized the~shape of the bumping route by
the~sequence $Y_0,Y_1,\dots$ which gives \emph{the~number of the~row} in which the
bumping route reaches a~specified column, cf.~\eqref{eq:Y}.
With the help of \cref{prop:lazy-traj-correspondence} we can define
the~lazy counterpart of these quantities:
for $x,m\in\mathbb{N}_0$ we denote by
\[ T_x^{[m]}=T_x= \min\left\{ t : x_m^{(t)}\leq x \right\} \]
the~\emph{time} it takes for the~bumping route (in the~lazy parametrization) to
reach the~specified column.
The~following result is the~lazy version of \cref{prop:m=1}.
\begin{lemma}
\label{lem:lazy:m=1}
For each integer $m\geq 1$
\[ \lim_{u\to\infty}
\sqrt{u} \ \mathbb{P}\left\{ T_0^{[m]} > u\right\} = m.
\]
\end{lemma}
\begin{proof}
By \cref{lem:transpose}, for any $u\in \mathbb{N}_0$
\[ \mathbb{P}\left\{ T_0^{[m]} > u\right\}=
\mathbb{P} \left\{ x_m^{(u)} \geq 1 \right\} =
\mathbb{P}\left\{ y^{(u)}_{u-m} \geq 1 \right\}.\]
\medskip
In the~special case $m=1$ the~proof is particularly easy: the~right-hand
side is equal to $\mathbb{P} \! \left( E^{(u)}_0 \right)$ and
\cref{prop:asymptotic-probability} provides the~necessary asymptotics.
\medskip
For the~general case $m\geq 1$ we use the~notations from
\cref{sec:plancherel-growth-process-independent} for $k=0$, and $n=u-m$,
and $\ell=m$. The~event $y^{(u)}_{u-m} \geq 1$ occurs if and only if at
least one of the~numbers
$R^{(n+1)},\dots,R^{(n+\ell)}$ is equal to $0$.
We apply \cref{thm:independent-explicit}; it follows that the~probability
of the~latter event is equal, up to an~additive error term of the~order
$o \! \left(\frac{m}{\sqrt{u-m}}\right) = o \! \left(\frac{1}{\sqrt{u}}\right)$,
to the~probability that in $m$~Bernoulli
trials with success probability $\frac{1}{\sqrt{n}}$ there is at least
one success. In~this way we proved that
\[ \mathbb{P}\left\{ y^{(u)}_{u-m} \geq 1 \right\} =
\frac{m}{\sqrt{u}} + o\left( \frac{1}{\sqrt{u}} \right),\]
as desired.
\end{proof}
\begin{proof}[Proof of \cref{prop:is-finite}]
Since $Y_0^{[m]}\geq Y_1^{[m]}\geq \cdots $ is a~weakly decreasing sequence, it is enough
to consider the~case $x=0$. We apply \cref{lem:lazy:m=1} in the~limit
$u\to\infty$. It follows that the~probability that the~bumping route
$\mathcal{T}\leftsquigarrow m+\nicefrac{1}{2}$ does not reach the~column with the
index $0$ is equal to
\[ \lim_{u \to\infty} \mathbb{P}\left\{ T_0^{[m]} \geq u\right\} = 0,
\]
as required.
\end{proof}
\section{Transition probabilities for the~augmented Plancherel growth process}
Our main result in this section is \cref{thm:lazy-poisson}. It will be the~key
tool for proving the~main results of the~current paper.
\subsection{Approximating Bernoulli distributions by linear combinations of
Poisson distributions}
\newcommand{\nu}{\nu}
The~following \cref{lem:poisson-binom} is a~technical result which will be
necessary later in the~proof of \cref{prop:transition-augmented}. Roughly
speaking, it gives a~positive answer to the~following question: \emph{for a
given value of $k\in\mathbb{N}_0$, can the point measure $\delta_k$ be approximated by
a~linear combination of the Poisson distributions in some explicit, constructive
way?}
A~naive approach to this problem would be to consider a~scalar multiple
of the Poisson distribution $e^z \operatorname{Pois}(z)$ which corresponds to the
sequence of weights
\[ \mathbb{N}_0\ni m \mapsto \frac{1}{m!} z^m\]
and then to consider its $k$-th derivative with respect to the~parameter $z$
for~\mbox{$z=0$}. This is not exactly a~solution to the~original question (the
derivative is not a~linear combination), but since the~derivative can~be
approximated by the~forward difference operator, this naive approach gives a
hint that an~expression such as \eqref{eq:defmukph} in the~special case $p=1$
might be, in fact, a~good answer.
\begin{lemma}
\label{lem:poisson-binom}
Let us fix an~integer $k\geq 0$ and a~real number $0\leq p\leq 1$.
For~each~$h>0$ the~linear combination of the Poisson distributions
\begin{equation}
\label{eq:defmukph}
\nu_{k,p,h}:= \frac{1}{\left( e^h -1 \right)^k} \sum_{0\leq j \leq k}
(-1)^{k-j} \binom{k}{j} e^{j h} \operatorname{Pois}(p j h)
\end{equation}
is a~probability measure on $\mathbb{N}_0$.
As $h\to 0$, the~measure $\nu_{k,p,h}$ converges (in the~sense of total
variation distance) to the~binomial distribution $\operatorname{Binom}(k,p)$.
\end{lemma}
\begin{proof}
\emph{The~special case $p=1$.}
For a~function $f$ on the~real line
we consider its \emph{forward difference} function $\Delta[f]$
given by
\[ \Delta[f] (x) = f(x+1)- f(x).\]
It follows that the~iterated forward difference is given by
\[ \Delta^k[f] (x) = \sum_{0\leq j\leq k} (-1)^j
\binom{k}{j}
f(x+k-j).\]
A~priori, $\nu_{k,1,h}$ is a~signed measure with the~total mass equal to
\begin{equation}
\label{eq:total-mass}
\frac{1}{\left( e^h -1 \right)^k} \sum_{0\leq j \leq k}
(-1)^{k-j} \binom{k}{j} e^{j h} =
\frac{1}{\left( e^h -1 \right)^k} \Delta^k \left[e^{hx}\right](0).
\end{equation}
The~right-hand side of \eqref{eq:total-mass} is equal to $1$,
since the~forward difference of an~exponential function is again an~exponential:
\[
\Delta \left[e^{hx}\right] = \left( e^h -1 \right) e^{h x}.
\]
The~atom of $\nu_{k,1,h}$ at an~integer $m\geq 0$ is equal to
\begin{multline*}
\nu_{k,1,h}(m)=
\frac{1}{\left( e^h -1 \right)^k m!} \sum_{0\leq j \leq k} (-1)^{k-j} \binom{k}{j} (j h)^m = \\
\frac{h^m}{\left( e^h -1 \right)^k m!} \Delta^k \left[ x^m \right](0).
\end{multline*}
Note that the~monomial $x^m$ can~be expressed in terms of the~falling
factorials~$x^{\underline{p}}$ with the~coefficients given by the Stirling numbers
of the~second kind:
\[ x^m = \sum_{0\leq p\leq m} \stirling{m}{p} x^{\underline{p}} \: , \]
hence
\[ \Delta^k \left[ x^m\right] = \sum_{p} \stirling{m}{p} \Delta^k \left[ x^{\underline{p}} \right] =
\sum_{p \geq k} \stirling{m}{p} p^{\underline{k}}\ x^{\underline{p-k}}.
\]
When we evaluate the~above expression at $x=0$, there is only one non-zero
summand
\[ \Delta^k \left[ x^m\right] (0) =
\stirling{m}{k} k!.
\]
Thus
\[
\nu_{k,1,h}(m)=
\frac{h^m k!}{\left( e^h -1 \right)^k m!}
\stirling{m}{k}\geq 0,\]
and the~above expression is non-zero only for $m\geq k$.
All in all, $\nu_{k,1,h}$ is a~probability measure on $\mathbb{N}_0$, as required.
It follows that the~total variation distance between $\operatorname{Binom}(k,1)=\delta_k$ and
$\nu_{k,1,h}$ is equal to
\[ \sum_{i=0}^\infty \left[ \delta_k(i) - \nu_{k,1,h}(i) \right]^+ =
1 - \nu_{k,1,h}(k) =
1 - \frac{h^k}{\left( e^h -1 \right)^k} \xrightarrow{h\to 0} 0, \]
as required.
\medskip
\emph{The~general case.}
For a~signed measure $\mu$ which is supported on $\mathbb{N}_0$ and $0\leq p\leq 1$ we define
the~signed measure $C_p[\mu]$ on $\mathbb{N}_0$ by
\[ C_p[\mu](k) = \sum_{j\geq k} \mu(j) \binom{j}{k} p^k (1-p)^{j-k}.\]
In the~case when $\mu$ is a~probability measure, $C_p[\mu]$ has a~natural
interpretation as the~probability distribution of a~compound binomial random
variable $\operatorname{Binom}(M,p)$, where $M$ is a~random variable with
the probability distribution given by $\mu$.
It is easy to check that for any $0\leq q \leq 1$ the~image of a~binomial distribution
\[ C_p[ \operatorname{Binom}(n,q) ] = \operatorname{Binom}(n, pq) \]
is again a~binomial distribution, and for any $\lambda\geq 0$ the~image of a
Poisson distribution
\[ C_p [ \operatorname{Pois}(\lambda) ] = \operatorname{Pois}(p\lambda)\]
is again a~Poisson distribution.
Since $C_p$ is a~linear map, by the~very definition \eqref{eq:defmukph} it
follows that
\begin{equation}
\label{eq:cp}
C_p [ \nu_{k,1,h} ] = \nu_{k,p,h};
\end{equation}
in particular the~latter is a~probability measure, as required. By considering
the~limit $h\to 0$ of \eqref{eq:cp} we get
\[ \lim_{h\to 0} \nu_{k,p,h}
= C_p \left[ \operatorname{Binom}(k,1) \right]
= \operatorname{Binom}(k,p)
\]
in the~sense of~total variation distance,
as required.
\end{proof}
\subsection{The inclusion $\diagrams^*\subset \mathbb{N}_0 \times \mathbb{Y}$ }
\label{sec:inclusion-x}
We will extend the meaning of the
notations from \cref{sec:augmented-Young-diagrams} to a larger set.
The~map
\begin{equation}
\label{eq:augmented-x}
\diagrams^*\ni (\lambda,\Box) \mapsto (x_\Box, \lambda~) \in \mathbb{N}_0 \times \mathbb{Y},
\end{equation}
where $\Box=(x_\Box,y_\Box)$, allows us to identify $\diagrams^*$ with a~subset of
$\mathbb{N}_0 \times \mathbb{Y}$.
For a pair $(x,\lambda)\in\mathbb{N}_0\times\mathbb{Y}$
we will say that $\lambda$ is its \emph{regular part}.
We define the~edges in this larger set \mbox{$\mathbb{N}_0\times \mathbb{Y}\supset \diagrams^*$} as follows:
we declare that $(x,\lambda)\nearrow (\widetilde{x},\widetilde{\lambda})$
if the~following two conditions hold true:
\begin{equation}
\label{eq:augmented-edge-B}
\lambda\nearrow\widetilde{\lambda} \quad \text{and} \quad
\widetilde{x} = \begin{cases}
\max\left\{ \widetilde{\lambda}_i :
\widetilde{\lambda}_i \leq x \right\} & \text{if the~unique box of $\widetilde{\lambda}/\lambda$} \\[-1ex]
& \quad \text{is located in the column $x$},\\[1ex]
x & \text{otherwise.}
\end{cases}
\end{equation}
In this way the oriented graph $\diagrams^*$ is a subgraph of $\mathbb{N}_0\times\mathbb{Y}$.
An~analogous lifting
property as in \cref{lem:lifting} remains valid if we assume that the~initial element $\Lambda^{(m)}\in\mathbb{N}_0\times \mathbb{Y}$
and the~elements of the~lifted path
\[ \Lambda^{(m)}\nearrow \Lambda^{(m+1)} \nearrow \cdots \in \mathbb{N}_0\times \mathbb{Y} \]
are allowed to be taken from this larger oriented graph.
With these definitions the transition probabilities
\eqref{eq:transition-probability-augmented} also make sense if
$\Lambda,\widetilde{\Lambda}\in\mathbb{N}_0\times \mathbb{Y}$ are taken from this larger
oriented graph and can be used to define Markov chains valued in $\mathbb{N}_0\times
\mathbb{Y}$.
\subsection{Transition probabilities for augmented Plancherel growth processes}
\label{sec:augmented-plancherel-growth-proccess}
For the~purposes of the~current section we will view $\diagrams^*$
as a subset of $\mathbb{N}_0\times \mathbb{Y}$, cf.~\eqref{eq:augmented-x}. In this way the augmented Plancherel growth process
initiated at time $m$, cf.~\eqref{eq:Markov}, can be viewed as the aforementioned Markov chain
\begin{equation}
\label{eq:augmented-process-x}
\Big( \big( x_m^{(t)}, \lambda^{(t)} \big) \Big)_{t \geq m}
\end{equation}
valued in $\mathbb{N}_0\times\mathbb{Y}$.
Let us fix some integer $n\in\mathbb{N}_0$. For each integer $m\in\{0,\dots,n\}$
we may remove some initial entries of the~sequence
\eqref{eq:augmented-process-x}
and consider the~Markov chain
\begin{equation}
\label{eq:Markov-truncated}
\Big( \big( x_m^{(t)}, \lambda^{(t)} \big) \Big)_{t \geq n}
\end{equation}
which is indexed by the~time parameter $t\geq n$.
In this way we obtain a~whole family of
Markov chains \eqref{eq:Markov-truncated} indexed by an~integer
$m\in\{0,\dots,n\}$ which have the~same transition
probabilities \eqref{eq:transition-probability-augmented}.
The latter encourages us to consider a~general class of Markov chains
\begin{equation}
\label{eq:Markov-truncated2}
\Big( \big( x^{(t)}, \lambda^{(t)} \big) \Big)_{t \geq n}
\end{equation}
valued in $\mathbb{N}_0\times \mathbb{Y}\supset \diagrams^*$, for which the~transition
probabilities are given by \eqref{eq:transition-probability-augmented} and for
which the~initial probability distribution of $ \big( x^{(n)},
\lambda^{(n)}\big)$ can~be arbitrary. We will refer to each such a~Markov
chain as \emph{augmented Plancherel growth process}.
\begin{proposition}
\label{prop:transition-augmented}
Let an~integer $k\in\mathbb{N}_0$ and a~real number $0<p<1$ be fixed,
and let $n'=n'(n)$ be a~sequence of integers such that $n'\geq n$ and
\[ \lim_{n\to\infty} \sqrt{\frac{n}{n'}}= p.\]
For a~given integer $n\geq 0$ let \eqref{eq:Markov-truncated2} be an~augmented
Plancherel growth process with the~initial probability distribution at time $n$
given by
\[ \delta_k \times \Plancherel_n .\]
Then the~total variation
distance
\begin{equation}
\label{eq:tvd}
\delta\left\{ \left(x^{(n')},\lambda^{(n')} \right),
\; \operatorname{Binom}\left(k, p \right) \times \Plancherel_{n'}\right\}
\end{equation}
converges to $0$, as $n\to\infty$.
\end{proposition}
\begin{proof}
Let $\epsilon>0$ be given. By \cref{lem:poisson-binom} there exists some $h>0$
with the~property that for each $q \in \{1, p\}$ the~total variation distance
between the~measure $\nu_{k,q,h}$ defined in~\eqref{eq:defmukph}
and the~binomial
distribution $\operatorname{Binom}(k,q)$ is bounded from above by $\epsilon$.
\medskip
Let $T$ be a~map defined on the set of
probability measures on $\mathbb{N}_0 \times \mathbb{Y}_n$
in the~following way.
For a~probability measure $\mu$ on $\mathbb{N}_0\times \mathbb{Y}_n$ consider
the~augmented Plancherel growth process \eqref{eq:Markov-truncated2} with the
initial probability distribution at time $n$ given by $\mu$ and
define $T \mu$ to be the~probability measure on $\mathbb{N}_0\times \mathbb{Y}_{n'}$ which gives the
probability distribution of $\big(x^{(n')},\lambda^{(n')} \big)$ at time~$n'$.
It is easy to extend the~map $T$ so that it becomes a~linear map between
the~vector space of \emph{signed} measures on $\mathbb{N}_0\times \mathbb{Y}_n$ and the
vector space of \emph{signed} measures on $\mathbb{N}_0\times \mathbb{Y}_{n'}$. We equip
both vector spaces with a~metric which corresponds to the~total variation
distance. Then $T$ is a~contraction because of
Markovianity of the~augmented Plancherel growth process.
\medskip
For $m\in\{0,\dots,n\}$ and $t\geq n$ we denote by $\mu_m(t)$ the~probability
measure on $\mathbb{N}_0\times \mathbb{Y}$, defined by the~probability distribution at time $t$
of the~augmented Plancherel growth process $\big(x_m^{(t)},\lambda^{(t)}
\big)$ initiated at time $m$. For the~aforementioned value of $h>0$ we consider the
signed measure on $\mathbb{N}_0\times \mathbb{Y}_t$ given by
the~linear combination
\[ \mathbb{P}(t):=\frac{1}{\left( e^h -1 \right)^k} \sum_{0\leq j \leq k}
(-1)^{k-j} \binom{k}{j} e^{j h} \mu_{\left\lfloor j h \sqrt{n} \right\rfloor}(t) \]
(which is well-defined for sufficiently big values of $n$ which assure that $k
h \sqrt{n}< n\leq t$).
We apply \cref{prop:distribution-fixed-time};
it follows that for any $j \in \{0, \dots, k\}$
the~total variation distance between
$\mu_{\left\lfloor j h \sqrt{n} \right\rfloor}(n)$
and the~product measure
\[
\operatorname{Pois}(jh) \times \Plancherel_n
\]
converges to $0$, as $n\to\infty$; it follows that
the~total
variation distance between $\mathbb{P}(n)$ and the~product measure
\begin{equation}
\label{eq:almost-delta}
\nu_{k,1,h} \times \Plancherel_{n}
\end{equation}
converges to $0$, as $n\to\infty$.
On the~other hand, the~value of $h>0$ was
selected in such a~way that the~total variation distance between the
probability measure \eqref{eq:almost-delta} and the~product measure
\begin{equation}
\label{eq:deltatimesplan}
\delta_k \times \Plancherel_n
\end{equation}
is smaller than~$\epsilon$. In this way we proved that
\[ \limsup_{n\to\infty} \, \delta\Big\{\, \mathbb{P}(n), \ \delta_k \times \Plancherel_n \Big\} \leq \epsilon.\]
An~analogous reasoning shows that
\[ \limsup_{n\to\infty} \, \delta\Big\{\, \mathbb{P}(n'), \ \operatorname{Binom}(k,p) \times \Plancherel_{n'} \Big\} \leq \epsilon.\]
The~image of $\mathbb{P}(n)$ under the~map $T$ can~be calculated by linearity of $T$:
\[ \mathbb{P}(n') = T \mathbb{P}(n).\]
By the triangle inequality and the~observation that the~map $T$ is a~contraction,
\begin{multline*} \eqref{eq:tvd} \leq
\delta\left\{ \left(x^{(n')},\lambda^{(n')} \right),
\; \mathbb{P}(n') \right\} + \epsilon \leq \\
\delta\left\{ \left(x^{(n)},\lambda^{(n)} \right),
\; \mathbb{P}(n) \right\} + \epsilon \leq 2\epsilon
\end{multline*}
holds true for sufficiently big values of $n$, as required.
\end{proof}
\subsection{Bumping route in the~lazy parametrization converges to the Poisson
process}
Let $\left( N(t) : t\geq 0 \right)$ denote the Poisson counting process which is independent
from the Plancherel growth process $\lambda^{(0)}\nearrow\lambda^{(1)}\nearrow
\cdots$. The following result is the lazy version of \cref{thm:main-poisson}.
\begin{theorem}
\label{thm:lazy-poisson}
Let $l\geq 1$ be a~fixed integer, and $z_1>\cdots> z_l$ be a~fixed sequence
of positive real numbers.
Let $\Lambda_m^{(m)}\nearrow \Lambda_m^{(m+1)} \nearrow \cdots$ be the augmented Plancherel growth process initiated at time~$m$.
We denote $\Lambda_m^{(t)}=\big(\lambda^{(t)},
\Box_m^{(t)}\big)$; let $\Box_m^{(t)}=\big(x_m^{(t)}, y_m^{(t)}\big)$ be the
coordinates of the special box at time $t$.
For each $1\leq i\leq l$ let $t_i=t_i(m)$ be a~sequence of positive
integers such that
\[ \lim_{m\to\infty} \frac{m}{\sqrt{t_i}} = z_i.\]
We assume that $t_1\leq \cdots \leq t_l$.
Then the~total variation distance between
\begin{itemize}
\item the~probability distribution of the~vector
\begin{equation}
\label{eq:RANDOM-A}
\left( x_m^{(t_1)}, \dots , x_m^{(t_l)}, \lambda^{(t_l)} \right),
\end{equation}
and
\item the~probability distribution of the~vector
\begin{equation}
\label{eq:RANDOM-B}
\left( N(z_1), \dots,
N(z_l), \lambda^{(t_l)} \right)
\end{equation}
\end{itemize}
converges to $0$, as $m\to\infty$.
\end{theorem}
\begin{proof}
We will perform the proof by induction over $l$.
Its main idea is that the~collection of the~random vectors
\eqref{eq:RANDOM-A} over $l\in\{1,2,\dots\}$ forms a~Markov chain;
the~same holds true for the~analogous collection of
the~random vectors~\eqref{eq:RANDOM-B}.
We will compare their initial probability
distributions (thanks to \cref{prop:distribution-fixed-time}) and --- in
a~very specific sense --- we will compare the~kernels of these Markov chains
(with \cref{prop:transition-augmented}). We present the~details below.
\medskip
The~induction base $l=1$ coincides with
\cref{prop:distribution-fixed-time}.
\medskip
We will prove now the~induction step. We start with the~probability
distribution of the~vector \eqref{eq:RANDOM-A} (with the~substitution
$l:=l+1$). Markovianity of the augmented Plancherel growth process implies
that this probability distribution is given by
\begin{multline}
\label{eq:tvda}
\mathbb{P}\left\{ \left( x_m^{(t_1)}, \dots , x_m^{(t_{l+1})}, \lambda^{(t_{l+1})} \right) =
(x_1,\dots,x_{l+1},\lambda) \right\} =\\
\shoveleft{ \sum_{\mu\in\mathbb{Y}_{t_l}}
\mathbb{P}\left\{ \left( x_m^{(t_1)}, \dots , x_m^{(t_{l})}, \lambda^{(t_{l})} \right) =
(x_1,\dots,x_{l},\mu) \right\} \times }\\
\times \PPcondCurly{ \left( x_m^{(t_{l+1})}, \lambda^{(t_{l+1})} \right) =
\left( x_{l+1}, \lambda\right) }{
\left( x_m^{(t_{l})}, \lambda^{(t_{l})} \right) =
\left( x_{l}, \mu \right) }
\end{multline}
for any $x_1,\dots,x_{l+1}\in\mathbb{N}_0$ and $\lambda\in\mathbb{Y}$.
We define the~probability measure $\mathbb{Q}$
on $\mathbb{N}_0^{l+1}\times \mathbb{Y}$ which to
a~tuple $\left( x_1,\dots,x_{l+1},\lambda~\right)$ assigns the~probability
\begin{multline}
\label{eq:tvdb}
\mathbb{Q}\left( x_1,\dots,x_{l+1},\lambda~\right) :=\\
\shoveleft{
\sum_{\mu\in\mathbb{Y}_{t_l}}
\mathbb{P}\left\{ \left( N(z_1),\dots, N(z_l) \right) =
(x_1,\dots,x_{l}) \right\} \times
\Plancherel_{t_l}(\mu) \times} \\
\times \PPcondCurly{ \left( x_m^{(t_{l+1})}, \lambda^{(t_{l+1})} \right) =
\left( x_{l+1}, \lambda\right) }{
\left( x_m^{(t_{l})}, \lambda^{(t_{l})} \right) =
\left( x_{l}, \mu \right) }
. \end{multline}
In the~light of the~general definition~\eqref{eq:Markov-truncated2} of the
augmented Plancherel growth process, the~measures \eqref{eq:tvda} and
\eqref{eq:tvdb} on $\mathbb{N}_0^{l+1}\times \mathbb{Y}$ can~be viewed as applications
of the~same Markov kernel (which correspond to the last factors on the right-hand side of \eqref{eq:tvda} and \eqref{eq:tvdb})
\[
\PPcondCurly{ \left( x_m^{(t_{l+1})}, \lambda^{(t_{l+1})} \right) =
\left( x_{l+1}, \lambda\right) }{
\left( x_m^{(t_{l})}, \lambda^{(t_{l})} \right) =
\left( x_{l}, \mu \right) },
\quad
\mu\in\mathbb{Y}_{t_l},
\]
to two specific initial probability distributions.
Since such an~application of a~Markov kernel is a~contraction (with respect to
the~total variation distance), we proved in this way that the~total variation
distance between \eqref{eq:tvda} and \eqref{eq:tvdb} is bounded from above by
the~total variation distance between the~initial distributions, that is the~random
vectors \eqref{eq:RANDOM-A} and~\eqref{eq:RANDOM-B}.
By the~inductive hypothesis
the total variation distance between the measures $\mathbb{P}$ and $\mathbb{Q}$
converges to zero as $m\to\infty$.
The remaining difficulty is to understand the asymptotic behavior of the measure $\mathbb{Q}$.
\smallskip
Observe that the~sum on the~right hand side of \eqref{eq:tvdb}
\begin{multline}
\sum_{\mu\in\mathbb{Y}_{t_l}} \Plancherel_{t_l}(\mu) \times \\
\times \PPcondCurly{ \left( x_m^{(t_{l+1})}, \lambda^{(t_{l+1})} \right) =
\left( x_{l+1}, \lambda\right) }{
\left( x_m^{(t_{l})}, \lambda^{(t_{l})} \right) =
\left( x_{l}, \mu \right) } = \\[1ex]
= \PPcondCurly{ \left(x^{(n')},\lambda^{(n')}\right) =
\left(x_{l+1}, \lambda\right) }{ \left(x^{(n)},\lambda^{(n)}\right) \overset{d}{=} \delta_k \times \Plancherel_n }
\end{multline}
is the~probability distribution of the~random vector
$\big(x^{(n')},\lambda^{(n')}\big)$ which appears in
\cref{prop:transition-augmented} with $n'=t_{l+1}$, and $n=t_l$, and $p=
\frac{z_{l+1}}{z_{l}}$, and $k=x_l$.
Therefore we proved that the~measure $\mathbb{Q}$ is in
an~$o(1)$-neighborhood of the~following probability measure
\begin{multline}
\label{eq:prob-Q}
\mathbb{Q}'\left( x_1,\dots,x_{l+1},\lambda~\right) :=\\
{
\mathbb{P}\left\{ \big( N(z_1),\dots, N(z_l) \big) =
(x_1,\dots,x_{l}) \right\} \times} \\
\times \Plancherel_{n'}(\lambda) \
{\operatorname{Binom}\left( x_l, \frac{z_{l+1}}{z_l} \right)(x_{l+1})}.
\end{multline}
It is easy to check that
\[
\PPcond{N(z_{l+1})=x_{l+1}}{N(z_{l})=x_{l}}
= \operatorname{Binom}\left( x_l, \frac{z_{l+1}}{z_l} \right)(x_{l+1}).
\]
Hence the~probability of the~binomial distribution
which appears as the~last factor
on the~right-hand side of \eqref{eq:prob-Q}
can~be interpreted as the~conditional probability distribution of the Poisson
process in the~past, given its value in the~future.
We show that the Poisson counting process with the~reversed time is also a~Markov process.
Since the Poisson counting process has independent increments, the probability of the event
\[
\big( N(z_1),\dots, N(z_l) \big) = (x_1,\dots,x_l)
\]
can be written as a product; an analogous observation is valid for
$l:=l+1$. Due to cancellations of the factors which contribute to the numerator
and the denominator, the following conditional probability can be simplified:
\begin{multline*}
\PPcond{N(z_{l+1})=x_{l+1}}{\big( N(z_1),\dots, N(z_l) \big) = (x_1,\dots,x_l)}
= \\[1ex]
= \frac{ \mathbb{P}\left\{ \big( N(z_1),\dots, N(z_{l+1}) \big) = (x_1,\dots,x_{l+1}) \right\}
}{
\mathbb{P}\left\{ \big( N(z_1),\dots, N(z_{l}) \big) = (x_1,\dots,x_{l}) \right\}
}
= \\[1ex]
= \frac{\mathbb{P}\big( N(z_{l})=x_{l} \: \wedge \: N(z_{l+1})=x_{l+1} \big) }{ \mathbb{P}\big( N(z_{l})=x_{l} \big)} = \\[1ex] =
\PPcond{N(z_{l+1})=x_{l+1}}{N(z_{l})=x_{l}}.
\end{multline*}
By combining the above observations with \eqref{eq:prob-Q} it follows that
\begin{multline*}
\mathbb{Q}'\left( x_1,\dots,x_{l+1},\lambda \right)=\\
\mathbb{P}\left\{ \left( N(z_1),\dots, N(z_{l+1}) \right) =
(x_1,\dots,x_{l+1}) \right\} \ \Plancherel_{t_{l+1}}(\lambda)
\end{multline*}
is the~probability distribution of \eqref{eq:RANDOM-B} (with the~obvious
substitution $l:=l+1$) which completes the~inductive step.
\end{proof}
\subsection{Lazy version of \cref{rem:poisson}}
The~special case $l=0$ of the~following result seems to be closely related to a
very recent work of Azangulov and Ovechkin \cite{Azangulov2020} who used
different methods.
\begin{proposition}
\label{prop:lazy-remark}
Let $(\psi_i)$ be a~sequence of independent, identically distributed random
variables with the~exponential distribution $\operatorname{Exp}(1)$.
For each $l\in\mathbb{N}_0$ the~joint distribution of the~finite tuple of random
variables
\begin{equation}
\label{eq:vector}
\left( \frac{m}{\sqrt{T^{[m]}_0}},\dots, \frac{m}{\sqrt{T^{[m]}_l}} \right)
\end{equation}
converges, as $m\to\infty$, to the~joint distribution of the~sequence of partial sums
\[ \left( \psi_0,\;\;\; \psi_0+\psi_1,\;\;\; \dots, \;\;\; \psi_0+\psi_1+\cdots+\psi_l \right).\]
\end{proposition}
\begin{proof}
For any $s_0,\dots,s_l>0$ the~cumulative distribution function of the~random
vector \eqref{eq:vector}
\begin{multline}
\label{eq:CDF}
\mathbb{P}\left( \frac{m}{\sqrt{T^{[m]}_0}} < s_0, \quad \dots, \quad
\frac{m}{\sqrt{T^{[m]}_l}} < s_l
\right) =\\
\mathbb{P}\left( x_m^{(t_0)} > 0, \quad x_m^{(t_1)} > 1, \quad \dots, \quad x_m^{(t_l)} > l \right)
\end{multline}
can~be expressed directly in terms of the~cumulative distribution of
the~random vector
$\left( x_m^{(t_0)}, \dots, x_m^{(t_l)} \right) $
with
\[ t_i = t_i(m)= \left\lfloor \left(\frac{m}{s_i}\right)^2\right\rfloor. \]
\cref{thm:lazy-poisson} shows that the~right-hand side of \eqref{eq:CDF}
converges to
\begin{multline*} \mathbb{P}\Big( N(s_0)>0, \quad N(s_1)>1, \quad \ldots, \;\;\; N(s_l)>l \Big) = \\
\mathbb{P}\Big( \psi_0 \leq s_0, \quad \psi_0+\psi_1 \leq s_1, \quad \dots, \quad \psi_0+\cdots+\psi_l\leq s_l \Big),
\end{multline*}
where
\[
\psi_i=
\inf\Big\{ t : N(t)\geq i+1 \Big\} - \inf\Big\{ t : N(t)\geq i \Big\},
\ \ \ i \in \mathbb{N}_0,
\]
denote the~time between the~jumps of the Poisson process. Since
$(\psi_0,\psi_1,\dots)$ form a~sequence of independent random variables with
the~exponential distribution, this concludes the~proof.
\end{proof}
\subsection{Conjectural generalization}
We revisit \cref{sec:trajectory} with some changes. This time let
\[
\xi=(\ldots,\xi_{-2},\xi_{-1},\xi_{0},\xi_1,\dots)
\]
be a~\emph{doubly} infinite
sequence of independent, identically distributed random variables with the
uniform distribution $U(0,1)$ on the~unit interval $[0,1]$.
Let us fix $m\in\mathbb{R}_+$. For $s,t\in\mathbb{R}_+$ we define
\begin{multline*} \big( x_m(s,t), y_m(s,t) \big) = \\
\Pos_\infty \left( P\left(\xi_{-\left\lfloor m s \right\rfloor},\dots,\xi_{-2},\xi_{-1}, \infty, \xi_{1}, \xi_{2}, \dots, \xi_{\left\lfloor \frac{m^2}{t^2} \right\rfloor} \right) \right).
\end{multline*}
Let $\mathcal{N}$ denote the~Poisson point process with the~uniform unit
intensity on~$\mathbb{R}_+^2$. For $s,t\in\mathbb{R}_+$ we denote by
\[ \mathcal{N}_{s,t}=\mathcal{N}\left( [0,s] \times [0,t] \right) \]
the~number of sampled points in the~specified rectangle.
\begin{conjecture}
The~random function
\begin{align}
\label{eq:2dPoisson}
\mathbb{R}_+^2 \ni (s,t) &\mapsto x_m(s,t) \\
\intertext{converges in distribution to Poisson point process}
\label{eq:2dPoisson-true}
\mathbb{R}_+^2 \ni (s,t) &\mapsto \mathcal{N}_{s,t}
\end{align}
in the~limit as $m\to\infty$.
\end{conjecture}
Note that the~results of the~current paper show the~convergence of
the marginals which correspond to (a) fixed value of $s$ and all values of $t>0$
(cf.~\cref{thm:lazy-poisson}), or (b) fixed value of $t$ and all values of
$s>0$ (this is a~corollary from the~proof of
\cref{prop:distribution-fixed-time}).
It is a~bit discouraging that the~contour curves obtained in computer
experiments (see \cref{fig:2dPoisson}) \emph{do not seem} to be counting the
number of points from some set which belong to a~specified rectangle,
see \cref{fig:2dPoissont} for comparison.
On the
other hand, maybe the~value of $m$ used in our experiments was not big enough
to reveal the~asymptotic behavior of these curves.
\subfile{figures-Poisson/2D-Poisson/2D-Poisson.tex}
\section{Removing laziness}
\label{sec:removing-laziness}
Most of the~considerations above concerned the~lazy parametrization of the
bumping routes. In this section we will show how to pass to the~parametrization
by the~row number and, in this way, to prove the~remaining claims from
\cref{sec:preliminaries} (that is \cref{thm:poisson-point,prop:m=1}).
\subsection{Proof of \cref{prop:m=1}}
\label{sec:proof:prop:m=1}
\newcommand{y}{y}
Our general strategy in this proof is to use \cref{lem:lazy:m=1} and to use the
observation that a Plancherel-distributed random Young diagram with $n$ boxes has
approximately $2\sqrt{n}$ columns in the scaling when $n\to\infty$.
\begin{proof}[Proof of \cref{prop:m=1}]
We denote by $c^{(n)}$ the~number of rows (or, equivalently, the~length of the
leftmost column) of the~Young diagram $\lambda^{(n)}$. Our~proof will be based
on an~observation (recall~\cref{prop:lazy-traj-correspondence}) that
\[ Y_0^{[m]} = c^{\left( T_0^{[m]} \right)}. \]
Let $\epsilon>0$ be fixed.
Since $c^{(n)}$ has the~same distribution as the
length of the~bottom row of a~Plancherel-distributed random Young diagram with
$n$ boxes, the~large deviation results \cite{Deuschel1999,Seppaelaeinen1998}
show that there exists a~constant $C_\epsilon>0$ such that
\begin{multline}
\label{eq:LD}
\mathbb{P}\left( \sup_{n\geq n_0} \left| \frac{c^{(n)}}{\sqrt{n}}-2 \right| > \epsilon\right) \leq
\sum_{n\geq n_0} \mathbb{P}\left( \left| \frac{c^{(n)}}{\sqrt{n}}-2 \right| > \epsilon\right) \leq \\
\sum_{n\geq n_0} e^{-C_\epsilon \sqrt{n}}
= O \! \left( e^{-C_\epsilon \sqrt{n_0}} \right) \leq
o\left( \frac{1}{n_0} \right)
\end{multline}
in the~limit as $n_0 \to\infty$.
\medskip
Consider an arbitrary integer $y\geq 1$. Assume that (i) the~event on the~left-hand side of \eqref{eq:LD} \emph{does not} hold true for
$n_0:=y$, \emph{and} (ii) $Y_0^{[m]}\geq y$.
Since $T_0^{[m]}\geq Y_0^{[m]}\geq
y$ it follows that
\[
\left| \frac{Y^{[m]}_0}{\sqrt{T_0^{[m]}}}-2 \right|
= \left| \frac{c^{\left(T_0^{[m]}\right)}}{\sqrt{T_0^{[m]}}}-2 \right| \leq \epsilon \]
hence
\begin{equation}
\label{eq:CV}
T_0^{[m]} \geq \left( \frac{y}{2+\epsilon} \right)^2.
\end{equation}
By considering two possibilities: either the~event on the~left-hand side
of~\eqref{eq:LD} holds true for $n_0:=y$ or not, it follows that
\[ \mathbb{P}\left\{ Y_0^{[m]} \geq y\right\} \leq
o\left( \frac{1}{y} \right) +
\mathbb{P}\left\{ T_0^{[m]} \geq \left( \frac{y}{2+\epsilon} \right)^2 \right\}.
\]
\cref{lem:lazy:m=1} implies therefore that
\[ \mathbb{P}\left\{ Y_0^{[m]} \geq y\right\} \leq
\frac{(2+\epsilon) m}{y} +
o\left( \frac{1}{y} \right)
\]
which completes the~proof of the~upper bound.
\medskip
For the~lower bound, assume that (i) the~event on the~left-hand side of
\eqref{eq:LD} \emph{does not} hold true for
\[n_0:= \left\lceil \left( \frac{y}{2-\epsilon}\right)^2 \right\rceil\]
\emph{and} (ii) $T_0^{[m]}\geq n_0$.
In an~analogous way as in the~proof of \eqref{eq:CV} it follows that
\[ Y_0^{[m]} \geq (2-\epsilon) \sqrt{T_0^{[m]}} \geq
(2-\epsilon) \sqrt{n_0} \geq y .\]
By considering two possibilities: either the~event on the~left-hand side
of~\eqref{eq:LD} holds true or not, it follows that that
\[ \mathbb{P}\left\{ T_0^{[m]} \geq n_0 \right\} \leq o\left( \frac{1}{y} \right) +
\mathbb{P}\left\{ Y_0^{[m]} \geq y \right\}.
\]
\cref{lem:lazy:m=1} implies therefore that
\[ \mathbb{P}\left\{ Y_0^{[m]} \geq y\right\} \geq
\frac{(2-\epsilon) m}{y} +
o\left( \frac{1}{y} \right)
\]
which completes the~proof of the~lower bound.
\end{proof}
\subsection{Lazy parametrization versus row parametrization}
\begin{proposition}
\label{prop:lazy-nonlazy}
For each $x\in\mathbb{N}_0$
\[ \lim_{m \to\infty} \frac{Y^{[m]}_x }{\sqrt{T^{[m]}_x} } = 2 \]
holds true in probability.
\end{proposition}
Regretfully, the~ideas used in the~proof of \cref{prop:m=1}
(cf.~\cref{sec:proof:prop:m=1} above) are not directly applicable for the~proof
of \cref{prop:lazy-nonlazy} when $x\geq 1$ because we are not aware of suitable
large deviation results for the~lower tail of the~distribution of a~specific row a
Plancherel-distributed Young diagram, other than~the~bottom row.
\medskip
Our general strategy in this proof is to study the length $\mu^{(t)}_x$ of the
column with the fixed index $x$ in the Plancherel growth process $\lambda^{(t)}$,
as $t\to\infty$. Since we are unable to get asymptotic \emph{uniform} bounds for
\begin{equation}
\label{eq:myratio}
\left| \frac{\mu_x^{(t)}}{\sqrt{t}} - 2 \right|
\end{equation}
over \emph{all} integers $t$ such that $\frac{t}{m^2}$ belongs to some compact
subset of $(0,\infty)$ in the limit $m\to\infty$, as a substitute we consider a
\emph{finite} subset of $(0,\infty)$ of the form
\[
\left\{ c (1+\epsilon),\; \dots,\; c(1+\epsilon)^l \right\}
\]
for arbitrarily small values of $c,\epsilon>0$ and arbitrarily large integer
$l\geq 0$ and prove the appropriate bounds for the integers $t_i(m)$ for which
\scalebox{0.95}{$\displaystyle \frac{t_i}{m^2}$} are approximately elements of this finite set.
We will use monotonicity in order to get some information about \eqref{eq:myratio} also for the integers $t$ which are between the numbers $\{ t_i(m) \}$.
\begin{proof}
Let $\epsilon>0$ be fixed. Let $\delta>0$ be arbitrary. By
\cref{prop:lazy-remark} the~law of the~random variable
$\frac{m}{\sqrt{T^{[m]}_x} }$ converges to the Erlang distribution which is
supported on $\mathbb{R}_+$ and has no atom in $0$. Let $W$ be a~random variable with
the~latter probability distribution; in this way the law of $\frac{T_x^{[m]}}{m^2}$ converges to the law of $W^{-2}$.
Let $c>0$ be a~sufficiently small number
such that
\[ \mathbb{P}\left( W^{-2} < c \right) < \delta. \]
Now, let $l\in\mathbb{N}_0$ be a~sufficiently big integer so that
\[ \mathbb{P}\left( c (1+\epsilon)^l < W^{-2} \right) < \delta. \]
We define
\[ t_i = t_i (m) = \left\lfloor m^2 c (1+\epsilon)^i \right\rfloor \qquad \text{for } i\in\{0,\dots,l\}.\]
With these notations there exists some $m_1$ with the~property that for each $m\geq m_1$
\begin{equation}
\label{eq:goodlimits}
\mathbb{P}\left( t_0 < T_x^{[m]} \leq t_l \right) > 1-2\delta.
\end{equation}
\medskip
Let $\mu^{(n)}=\big[ \lambda^{(n)}\big]^T$ be the~transpose of
$\lambda^{(n)}$; in this way $\mu^{(n)}_x$ is the~number of the~boxes of
$\mathcal{T}$ which are in the~column $x$ and contain an~entry $\leq n$. The
probability distribution of $\mu^{(n)}$ is also given by the Plancherel measure.
The monograph of Romik \cite[Theorem~1.22]{Romik2015a} contains that proof that
\[ \frac{\mu^{(t_i(m))}_x}{\sqrt{t_i(m)}} \xrightarrow{\mathbb{P}} 2\]
holds true in the special case of the bottom row $i=0$;
it is quite straightforward to check that this proof is also valid for
each $i\in\{0,\dots,l\}$, for the details see \cite[proof of Lemma 2.5]{MMS-Poisson2020v2}.
Hence
there exists some $m_2$ with the~property that for each $m\geq m_2$
the~probability of the~event
\begin{equation}
\label{eq:happy-event}
\left| \frac{\mu^{(t_i(m))}_x}{\sqrt{t_i(m)}} -2 \right| < \epsilon
\quad \text{holds \emph{for each} $i\in\{0,\dots,l\}$}
\end{equation}
is at least $1-\delta$.
\medskip
Let us consider an~elementary event $\mathcal{T}$ with the~property that the~event
considered in \eqref{eq:goodlimits} occurred, that is~$t_0 < T_x^{[m]} \leq t_l$,
\emph{and} the~event \eqref{eq:happy-event} occurred. Since $t_0\leq \cdots
\leq t_l$ form a~weakly increasing sequence, there exists an~index
$j=j(\mathcal{T}) \in\{0,\dots,l-1\}$ such that
\[
t_{j} < T_x^{[m]} \leq t_{j+1}.
\]
It follows that
\[ \mu_x^{(t_j)} < Y^{[m]}_x \leq \mu_x^{(t_{j+1})} \]
hence
\begin{multline}
\label{eq:goodbound}
(2-\epsilon) \frac{1}{\sqrt{1+\epsilon}+o(1)}
<
\frac{\mu_x^{(t_{j} )}}{\sqrt{t_{j+1}}}
\leq \frac{Y^{[m]}_x }{\sqrt{T^{[m]}_x} } \leq \\
\frac{\mu_x^{(t_{j+1})}}{\sqrt{t_j}} <
(2+\epsilon) \left( \sqrt{1+\epsilon} + o(1) \right).
\end{multline}
In this way we proved that for each $m\geq \max(m_1,m_2)$ the~probability of
the~event \eqref{eq:goodbound} is at least $1-3\delta$, as required.
\end{proof}
\subsection{Proof of \cref{thm:poisson-point}}
\label{sec:proof:thm:poisson-point}
\begin{proof}
For each integer $l\geq 0$ \cref{prop:lazy-remark} gives the asymptotics of
the joint probability distribution of the random variables
$T_0^{[m]},\dots,T_l^{[m]}$ which concern the shape of the bumping route in
the lazy parametrization. On the other hand, \cref{prop:lazy-nonlazy}
allows us to express asymptotically these random variables by their non-lazy counterparts
$Y_0^{[m]},\dots,Y_l^{[m]}$.
The discussion from \cref{rem:poisson} completes the proof.
\end{proof}
\section{Acknowledgments}
We thank Iskander Azangulov, Maciej Dołęga, Vadim Gorin, Piet Groeneboom, Adam
Jakubowski, Grigory Ovechkin, Timo Sepp\"{a}l\"{a}inen, and Anatoly Vershik for
discussions and bibliographic suggestions.
\section{Declarations}
\textbf{Funding.} Research supported by Narodowe Centrum Nauki, grant number 2017/26/A/ST1/00189.
Mikołaj Marciniak was additionally supported by
Narodowe Centrum Badań i Rozwoju, grant number POWR.03.05.00-00-Z302/17-00.
\textbf{Conflicts of interest/Competing interests:} not applicable.
\textbf{Availability of data and material (data transparency):} not applicable.
\textbf{Code availability:} not applicable.
\printbibliography
\end{document}
| 7178f27a96aa8875a02a585a7c813a8dad69cd72 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
When discretizing the Poisson equation with Lagrange finite elements,
flux equilibrated error estimators can be employed to build polynomial-degree-robust
(or $p$-robust for short) a posteriori error estimators
\cite{Brae_Pill_Sch_p_rob_09,Ern_Voh_p_rob_15}. This property, which is particularly important
for $hp$-adaptivity (see for instance~\cite{Dan_Ern_Sme_Voh_guar_red_18} and the references therein), means that
the local a posteriori error estimator is, up to data oscillation, a lower bound
of the local approximation error, up to a constant that is independent of the
polynomial degree (the constant can depend on the shape-regularity of the mesh).
It turns out that one of the cornerstones of $p$-robust local efficiency
is a $p$-robust $\HH(\ddiv)$-stability result
of a discrete minimization problem posed in a single mesh tetrahedron.
More precisely, let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\emptyset\subseteq \FF \subseteq \FF_K$ be a (sub)set of its faces.
Then there is a constant $C$ such that for every polynomial degree $p\ge0$ and all polynomial data
$r_K \in \mathcal P_p(K)$ and $r_F \in \mathcal P_p(F)$ for all
$F\in \FF$, such that $(r_K,1)_K=\sum_{F\in\FF} (r_F,1)_F$ if $\FF=\FF_K$
(detailed notation is explained below), one has
\begin{equation}
\label{stable_minimization_Hdiv}
\min_{\substack{
\vv_p \in \RT_p(K)
\\
\div \vv_p = r_K
\\
\vv_p \cdot \nn_K|_F = r_F \; \forall F \in \FF
}}
\|\vv_p\|_{0,K}
\leq
C
\min_{\substack{
\vv \in \HH(\ddiv,K)
\\
\div \vv = r_K
\\
\vv \cdot \nn_K|_F = r_F \; \forall F \in \FF
}}
\|\vv\|_{0,K}.
\end{equation}
This result is shown in~\cite[Lemma~A.3]{Ern_Voh_p_rob_3D_20},
and its proof relies on~\cite[Theorem~7.1]{Demk_Gop_Sch_ext_III_12} and~\cite[Proposition~4.2]{Cost_McInt_Bog_Poinc_10}. Importantly, the constant $C$ in~\eqref{stable_minimization_Hdiv} only depends on the shape-regularity of $K$, that is,
the ratio of its diameter to the diameter of its largest inscribed ball.
Notice that the converse bound of~\eqref{stable_minimization_Hdiv} trivially
holds with constant $1$. The stability result stated in~\eqref{stable_minimization_Hdiv}
is remarkable since it states that the minimizer from the discrete minimization set
performs as well as the minimizer from the continuous minimization set, up to a
$p$-robust constant.
The main contribution of the present work is to establish the counterpart of
\eqref{stable_minimization_Hdiv} for the N\'ed\'elec finite elements of order $p\ge0$ and
the Sobolev space $\HH(\ccurl)$.
As in the $\HH(\ddiv)$ case, our discrete stability result
relies on two key technical tools: a
stable polynomial-preserving lifting of volume data from
\cite[Proposition~4.2]{Cost_McInt_Bog_Poinc_10}, and
stable polynomial-preserving liftings of boundary data from
\cite{Demk_Gop_Sch_ext_I_09,Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12}.
Our main result, Theorem~\ref{theorem_stability_tetrahedra} below,
may appear as a somewhat expected consequence of these lifting operators, but our
motivation here is to provide all the mathematical
details of the proofs, which turn out to be nontrivial and in particular
more complex than in~\cite[Lemma~A.3]{Ern_Voh_p_rob_3D_20}. In particular
the notion of tangential traces in $\HH(\ccurl)$ is somewhat delicate, and
we employ a slightly different definition compared to
\cite{Demk_Gop_Sch_ext_I_09,Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12}.
Theorem~\ref{theorem_stability_tetrahedra} is to be used as a building block
in the construction of a $p$-robust a posteriori error estimator for
curl-curl problems. This construction will be analyzed in a forthcoming work.
The remainder of this paper is organized as follows.
We introduce basic notions in Section~\ref{sec:preliminaries_K} so as to state our main result, Theorem~\ref{theorem_stability_tetrahedra}. Then
Section~\ref{sec:proof} presents its proof.
\section{Statement of the main result}
\label{sec:preliminaries_K}
\subsection{Tetrahedron}
Let $K \subset \mathbb R^3$ be an arbitrary tetrahedron. We assume that $K$
is non-degenerate, i.e., the volume of $K$ is positive. We employ the notation
\begin{equation*}
h_K \eq \max_{\xx,\yy \in \overline{K}} |\xx-\yy|,
\qquad
\rho_K
\eq
\max \left \{ d \geq 0
\; \left | \;
\exists \xx \in K; \; B\left (\xx,\frac{d}{2} \right ) \subset \overline{K}
\right . \right \},
\end{equation*}
for the diameter of $K$ and the diameter of the largest closed ball contained
in $\overline{K}$. Then $\kappa_K \eq h_K / \rho_K$ is the so-called
shape-regularity parameter of $K$. Let $\FF_K$ be the set of faces of $K$,
and for every face $F \in \FF_K$, we denote by $\nn_F$ the unit vector normal to $F$ pointing outward $K$.
\subsection{Lebesgue and Sobolev spaces}
The space of square-integrable scalar-valued (resp. vector-valued)
functions on $K$ is denoted by $L^2(K)$ (resp. $\LL^2(K)$),
and we use the notation $(\cdot,\cdot)_K$ and $\|\cdot\|_{0,K}$ for, respectively, the inner product and the associated norm of both $L^2(K)$ and $\LL^2(K)$. $H^1(K)$ is the usual Sobolev space of scalar-valued functions
with weak gradient in $\LL^2(K)$, and $\HH^1(K)$ is the space of vector-valued functions
having all their components in $H^1(K)$.
If $F \in \FF_K$ is a face of $K$, then $\LL^2(F)$ is the set of vector-valued functions
that are square-integrable with respect to the surfacic measure of $F$.
For all $\ww \in \HH^1(K)$, we define the tangential component of $\ww$ on $F$ as
\begin{equation}
\label{eq:def_pi_tau_F}
\ppi^\ttau_F(\ww) \eq \ww|_F - (\ww|_F \cdot \nn_F) \nn_F \in \LL^2(F).
\end{equation}
More generally, if $\FF \subseteq \FF_K$ is a nonempty (sub)set of the faces of $K$,
we employ the notation $\Gamma_\FF \subseteq \partial K$ for the corresponding part of
the boundary of $K$, and $\LL^2(\Gamma_{\FF})$ is the associate Lebesgue space of
square-integrable functions over $\Gamma_{\FF}$.
\subsection{N\'ed\'elec and Raviart--Thomas polynomial spaces}
For any polynomial degree $p \geq 0$, the notation $\PP_p(K)$ stands for the space
of vector-valued polynomials such that all their components belong to $\mathcal P_p(K)$
which is composed of the restriction to $K$ of real-valued polynomials of total
degree at most $p$.
Following~\cite{Ned_mix_R_3_80,Ra_Tho_MFE_77}, we define the
polynomial spaces of N\'ed\'elec and Raviart--Thomas functions as follows:
\begin{equation*}
\NN_p(K) \eq \PP_p(K) + \xx \times \PP_p(K)
\quad \text{ and } \quad
\RT_p(K) \eq \PP_p(K) + \xx \mathcal P_p(K).
\end{equation*}
Let $\FF \subseteq \FF_K$ be a nonempty (sub)set of the faces of $K$.
On $\Gamma_\FF$, we define the (piecewise) polynomial space composed
of the tangential traces of the N\'ed\'elec polynomials
\begin{equation}
\label{eq_tr_K}
\NN_p^\ttau(\Gamma_\FF) \eq
\left \{
\ww_\FF \in \LL^2(\Gamma_\FF) \; | \;
\exists \vv_p \in \NN_p(K);
\ww_F \eq (\ww_\FF)|_F = \ppi^\ttau_F (\vv_p) \quad \forall F \in \FF
\right \}.
\end{equation}
Note that $\ww_\FF \in \NN_p^\ttau(\Gamma_\FF)$ if and only if
$\ww_F\in \NN_p^\ttau(\Gamma_{\{F\}})$ for all $F\in\FF$ and whenever $\FF$ contains two or more faces, $|\FF|\ge2$,
for every pair $(F_-,F_+)$ of distinct faces in $\FF$, the
compatibility condition $(\ww_{F_+})|_\edge \cdot \ttau_\edge =
(\ww_{F_-})|_\edge \cdot \ttau_\edge$ holds true on their common
edge $\edge \eq F_+\cap F_-$, i.e., the tangential trace is continuous along $\edge$.
For all $\ww_\FF \in \NN_p^\ttau(\Gamma_\FF)$, we define its surface curl as
\begin{equation}
\label{eq_scurl_curl_el}
\scurl_F (\ww_F) \eq (\curl \vv_p)|_F \cdot \nn_F \qquad \forall F \in \FF,
\end{equation}
where $\vv_p$ is any element of $\NN_p(K)$ such that $\ww_F = \ppi_F^{\ttau}(\vv_p)$
for all $F \in \FF$. This function is well-defined independently of the choice of $\vv_p$.
\subsection{Weak tangential traces for fields in $\HH(\ccurl,K)$ by integration by parts}
Let $\HH(\ccurl,K) \eq \left \{
\vv \in \LL^2(K) \; | \; \curl \vv \in \LL^2(K)
\right \}$ denote the Sobolev space composed of square-integrable vector-valued fields
with square-integrable curl. We equip this space with the norm
$\|\vv\|_{\ccurl,K}^2 \eq \|\vv\|_{0,K}^2 + \ell_K^2\|\curl \vv\|_{0,K}^2$,
where $\ell_K$ is a length scale associated with $K$, e.g., $\ell_K\eq h_K$
(the choice of $\ell_K$ is irrelevant in what follows).
For any field $\vv \in \HH^1(K)$, its tangential trace on a face $F\in\FF_K$
can be defined by using~\eqref{eq:def_pi_tau_F}.
This notion of (tangential) trace is defined (almost everywhere)
on $F$ without invoking test functions. The situation for a field
in $\HH(\ccurl,K)$ is more delicate.
The tangential trace over the whole boundary of $K$
can be defined by duality, but it is not straightforward to define
the tangential trace on a part of the boundary of $K$.
While it is possible to use restriction operators
\cite{Demk_Gop_Sch_ext_I_09,Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12},
we prefer a somewhat more direct
definition based on integration by parts. This approach is also more convenient
when manipulating (curl-preserving) covariant Piola transformations
(see, \eg, \cite[Section~7.2]{Ern_Guermond_FEs_I_2020} and Section~\ref{sec_st_2} below),
which is of importance, \eg, when mapping tetrahedra of a mesh to a reference tetrahedron.
In this work, we consider the following definition of the tangential trace on a (sub)set
$\Gamma_\FF \subseteq \partial K$.
\begin{definition}[Tangential trace by integration by parts]
\label{definition_partial_trace}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\FF \subseteq \FF_K$ be a nonempty (sub)set of its faces.
Let $\rr_\FF \in \NN_p^\ttau(\Gamma_\FF)$ as well as $\vv \in \HH(\ccurl,K)$.
We will employ the notation ``$\vv|^\ttau_\FF = \rr_\FF$'' to say that
\begin{equation*}
(\curl \vv,\pphi)_K - (\vv,\curl \pphi)_K = \sum_{F \in \FF} (\rr_F,\pphi \times \nn_F)_{F}
\quad
\forall \pphi \in \HH^1_{\ttau,\FF^{\mathrm{c}}}(K),
\end{equation*}
where
\begin{equation*}
\HH_{\ttau,\FF^{\mathrm{c}}}^1(K) \eq
\left \{
\ww \in \HH^1(K) \; | \; \ppi^\ttau_F(\ww) = \boldsymbol 0
\quad
\forall F \in \FF^{\mathrm{c}} \eq \FF_K \setminus \FF
\right \}.
\end{equation*}
Whenever $\vv \in \HH^1(K)$, $\vv|^\ttau_\FF = \rr_\FF$ if and only if
$\ppi^\ttau_F (\vv) = \rr_F$ for all $F \in \FF$.
\end{definition}
\subsection{Main result}
\label{sec_main_result}
We are now ready to state our main result. The proof is given in Section~\ref{sec:proof}.
\begin{theorem}[Stability of $\HH(\ccurl)$ discrete minimization in a tetrahedron]
\label{theorem_stability_tetrahedra}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\emptyset\subseteq \FF \subseteq \FF_K$ be a (sub)set of its faces.
Then, for every polynomial degree $p \geq 0$, for all $\rr_K \in \RT_p(K)$
such that $\div \rr_K = 0$, and, if $\emptyset\ne\FF$, for all
$\rr_\FF \in \NN_p^\ttau(\Gamma_\FF)$ such that
$\rr_K \cdot \nn_{F} = \scurl_F (\rr_F)$ for all $F \in \FF$, the following holds:
\begin{equation}
\label{eq_minimization_element_K}
\min_{\substack{
\vv_p \in \NN_p(K)
\\
\curl \vv_p = \rr_K
\\
\vv_{p}|^\ttau_\FF = \rr_\FF
}}
\|\vv_p\|_{0,K}
\le C_{\mathrm{st},K}
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
\\
\vv|^\ttau_\FF = \rr_\FF
}}
\|\vv\|_{0,K},
\end{equation}
where the condition on the tangential trace in the minimizing sets is null if $\emptyset=\FF$.
Both minimizers in~\eqref{eq_minimization_element_K} are uniquely defined and the constant
$C_{\mathrm{st},K}$ only depends on the shape-regularity parameter $\kappa_K$ of $K$,
so that it is in particular independent of $p$.
\end{theorem}
\section{Proof of the main result}
\label{sec:proof}
The discrete minimization set in~\eqref{eq_minimization_element_K}, which is a subset of
the continuous minimization set, is nonempty owing to classical properties of
the N\'ed\'elec polynomials and the compatibility conditions imposed on the data
$\rr_K$ and $\rr_\FF$. This implies the existence and uniqueness of both
minimizers owing to standard convexity arguments.
The proof of the bound~\eqref{eq_minimization_element_K} proceeds in three steps.
Fist we establish in Section~\ref{sec_st_1} the bound for minimization problems without trace constraints.
This first stability result crucially relies on~\cite{Cost_McInt_Bog_Poinc_10}
and is established directly on the given tetrahedron $K \subset \mathbb R^3$.
Then we establish in Section~\ref{sec_st_2} the bound for minimization problems without curl constraints.
This second stability result crucially relies on the results of~\cite{Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12}. Since the notion of
tangential trace employed therein slightly differs from the present one,
we first establish in Section~\ref{sec_aux} some auxiliary results on tangential traces and then prove the stability result by first working on the reference tetrahedron in $\mathbb R^3$
and then by mapping the fields defined on the given tetrahedron $K \subset \mathbb R^3$
to fields defined on the reference tetrahedron.
In all cases, the existence and uniqueness of the minimizers follows by the same arguments
as above.
Finally, in Section~\ref{sec_st_3} we combine both results so as to prove Theorem~\ref{theorem_stability_tetrahedra}.
To simplify the notation we write
$A\lesssim B$ for two nonnegative numbers $A$ and $B$
if there exists a constant $C$ that only depends on the
shape-regularity parameter $\kappa_K$ of $K$ but is independent of $p$
such that $A \leq C B$. The value of $C$ can change at each occurrence.
\subsection{Step 1: Minimization without trace constraints} \label{sec_st_1}
\begin{lemma}[Minimization without trace constraint]
\label{lemma_lifting_curl}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron.
Let $\rr_{K} \in \RT_p(K)$ be such that $\div \rr_{K} = 0$. The following holds:
\begin{equation}
\label{eq_minimization_zero_face}
\min_{\substack{
\vv_p \in \NN_p(K)
\\
\curl \vv_p = \rr_{K}
}}
\|\vv_p\|_{0,K}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K}
}
\|\vv\|_{0,K}.
\end{equation}
\end{lemma}
\begin{proof}
1) Let us first show that
\begin{equation*}
\|\rr_K\|_{-1,K} \leq \min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
}}
\|\vv\|_{0,K}.
\end{equation*}
Indeed, for every $\vv \in \HH(\ccurl,K)$ such
that $\curl \vv = \rr_K$, we have
\begin{align*}
\|\rr_K\|_{-1,K}
&=
\sup_{\substack{
\pphi \in \HH^1_0(K)
\\
|\pphi|_{1,K} = 1
}}
(\rr_K,\pphi)_{K}
=
\sup_{\substack{
\pphi \in \HH^1_0(K)
\\
|\pphi|_{1,K} = 1
}}
(\curl \vv,\pphi)_{K}
\\
&=
\sup_{\substack{
\pphi \in \HH^1_0(K)
\\
|\pphi|_{1,K} = 1
}}
(\vv,\curl \pphi)_{K}
\leq
\|\vv\|_{0,K}
\left (
\sup_{\substack{
\pphi \in \HH^1_0(K)
\\
|\pphi|_{1,K} = 1
}}
\|\curl \pphi\|_{0,K}
\right )
\leq
\|\vv\|_{0,K},
\end{align*}
since $\|\curl \pphi\|_{0,K} \leq |\pphi|_{1,K}$ for all $\pphi \in \HH^1_0(K)$.
The claim follows by taking the minimum (which exists owing to standard convexity arguments)
over all $\vv \in \HH(\ccurl,K)$ such that $\curl \vv = \rr_K$.
\\
2) Since $\div \rr_{K} = 0$, \cite[Proposition~4.2]{Cost_McInt_Bog_Poinc_10} ensures
the existence of an element $\ww_p \in \NN_p(K)$ such that $\curl \ww_p = \rr_{K}$ and
\begin{equation*}
\|\ww_p\|_{0,K} \lesssim \|\rr_{K}\|_{-1,K}.
\end{equation*}
We can conclude using 1) since
\begin{equation*}
\min_{\substack{
\vv_p \in \NN_p(K)
\\
\curl \vv_p = \rr_{K}
}}
\|\vv_p\|_{0,K}
\leq
\|\ww_p\|_{0,K}
\lesssim
\|\rr_{K}\|_{-1,K}
\leq
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_{K}
}}
\|\vv\|_{0,K}.
\end{equation*}
This proves~\eqref{eq_minimization_zero_face}.
\end{proof}
\subsection{Auxiliary results on the tangential component} \label{sec_aux}
We first establish a density result concerning the space composed of
$\HH(\ccurl,K)$ functions with vanishing tangential trace on $\Gamma_\FF$
in the sense of Definition~\ref{definition_partial_trace}. We consider the subspace
\begin{equation}
\label{eq_definition_HH_Gamma_ccurl}
\HH_{\Gamma_{\FF}}(\ccurl,K)
\eq
\left \{
\vv \in \HH(\ccurl,K) \; | \; \vv|^\ttau_\FF = \boldsymbol 0
\right \},
\end{equation}
equipped with the $\|\cdot\|_{\ccurl,K}$-norm defined above.
\begin{lemma}[Density]
\label{lemma_density}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\FF \subseteq \FF_K$ be a nonempty (sub)set of its faces.
The space $\CC^\infty_{\Gamma_\FF}(\overline{K})
\eq
\left \{
\vv \in \CC^\infty(\overline{K}) \; | \;
\vv|_{\Gamma_\FF} = \boldsymbol 0
\right \}$
is dense in $\HH_{\Gamma_\FF}(\ccurl,K)$.
\end{lemma}
\begin{proof}
Recalling~\cite[Remark 3.1]{Fer_Gil_Maxw_BC_97}, if
$\ww \in \HH^{-1/2}(\partial K)$, we can define its restriction
$\ww|_{\Gamma_\FF} \in (\HH^{1/2}_{00}(\Gamma_{\FF}))'$ by setting
\begin{equation} \label{eq_rest}
\langle \ww|_{\Gamma_\FF},\pphi \rangle
\eq
\langle \ww,\widetilde \pphi \rangle_{\partial K}
\quad \forall \pphi \in \HH^{1/2}_{00}(\Gamma_{\FF}),
\end{equation}
where $\widetilde \pphi \in \HH^{1/2}(\partial K)$ denotes the zero-extension of $\pphi$
to $\partial K$. Following~\cite{Fer_Gil_Maxw_BC_97}, we then introduce the space
\begin{equation*}
\VV_{\Gamma_\FF}(K) \eq \left \{
\vv \in \HH(\ccurl,K) \; | \;
(\vv \times \nn)|_{\Gamma_\FF} = \boldsymbol 0
\right \}.
\end{equation*}
Proposition 3.6 of~\cite{Fer_Gil_Maxw_BC_97} states that
$\CC^\infty_{\Gamma_\FF}(\overline{K})$ is dense in $\VV_{\Gamma_\FF}(K)$.
Thus, it remains to show that $\HH_{\Gamma_\FF}(\ccurl,K) \subset \VV_{\Gamma_\FF}(K)$.
Let $\vv \in \HH_{\Gamma_\FF}(\ccurl,K)$. For all $\ttheta \in \HH^{1/2}_{00}(\Gamma_{\FF})$,
we have $\widetilde \ttheta \in \HH^{1/2}(\partial K)$, and there exists
$\pphi \in \HH^1(K)$ such that $\widetilde \ttheta = \pphi|_{\partial K}$.
In addition, since $\widetilde \ttheta|_{\partial K \setminus \Gamma_{\FF}} = \mathbf 0$,
we have $\pphi \in \HH^1_{\FF^{\mathrm c}}(K)$, and in particular
$\pphi \in \HH^1_{\ttau,\FF^{\mathrm c}}(K)$. Then using~\eqref{eq_rest},
integration by parts, and Definition~\ref{definition_partial_trace}, we have
\begin{equation*}
\langle (\vv \times \nn)|_{\Gamma_\FF}, \ttheta \rangle
=
\langle \vv \times \nn, \widetilde \ttheta \rangle_{\partial K}
=
\langle \vv \times \nn, \pphi|_{\partial K} \rangle_{\partial K}
=
(\vv,\curl \pphi)_K - (\curl \vv,\pphi)_K = 0,
\end{equation*}
since $\vv \in \HH_{\Gamma_\FF}(\ccurl,K)$. Hence $(\vv \times \nn)|_{\Gamma_\FF} = \boldsymbol 0$,
and therefore $\vv \in \VV_{\Gamma_\FF}(K)$.
\end{proof}
Since we are going to invoke key lifting results established
in~\cite{Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12},
we now recall the main notation employed
therein (see~\cite[Section~2]{Demk_Gop_Sch_ext_II_09}). Let
\begin{equation*}
\trc_K^\ttau: \HH(\ccurl,K) \to \HH^{-1/2}(\partial K)
\end{equation*}
be the usual tangential trace operator obtained as in
Definition~\ref{definition_partial_trace}
with $\FF \eq \FF_K$ and let us equip the image space
\begin{equation*}
\XX^{-1/2}(\partial K) \eq \trc_K^\ttau(\HH(\ccurl,K))
\end{equation*}
with the quotient norm
\begin{equation}
\label{eq_puotient_norm_XX}
\|\ww\|_{\XX^{-1/2}(\partial K)}
\eq
\inf_{\substack{
\vv \in \HH(\ccurl,K)
\\
\trc_K^\ttau(\vv) = \ww
}}
\|\vv\|_{\ccurl,K}.
\end{equation}
For each face $F \in \FF_K$, there exists a Hilbert function space $\XX^{-1/2}(F)$
and a (linear and continuous) ``restriction'' operator
$\RRR_F: \XX^{-1/2}(\partial K) \to \XX^{-1/2}(F)$
that coincides with the usual pointwise restriction for smooth functions.
In particular, we have
\begin{equation} \label{eq:trace_DGS_NN}
\RRR_F(\trc_K^\ttau(\vv_p)) = \ppi^\ttau_F(\vv_p) \qquad \forall \vv_{p} \in \NN_p(K),
\end{equation}
with the tangential trace operator defined in~\eqref{eq:def_pi_tau_F}.
We have thus introduced two notions of ``local traces'' for $\HH(\ccurl,K)$
functions. On the one hand, Definition~\ref{definition_partial_trace} defines
an equality for traces on $\Gamma_{\FF}$ based on integration by parts. On the other hand,
the restriction operators $\RRR_F$ provide another notion of trace on any
face $F \in \FF$. The following result provides a connection
between these two notions.
\begin{lemma}[Trace restriction]
\label{lemma_trace_restriction}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\FF \subseteq \FF_K$ be a nonempty (sub)set of its faces.
For all $\rr_\FF \in \NN^\ttau_p(\Gamma_\FF)$ and all $\pphi \in \HH(\ccurl,K)$, if
$\pphi|^\ttau_\FF = \rr_\FF$ according to Definition~\ref{definition_partial_trace}, then
\begin{equation*}
\RRR_F (\trc^\ttau_K (\pphi)) = \rr_F \qquad \forall F \in \FF.
\end{equation*}
\end{lemma}
\begin{proof}
Let $\rr_\FF \in \NN^\ttau_p(\Gamma_\FF)$.
Recalling definition~\eqref{eq_tr_K} of $\NN_p^\ttau(\Gamma_\FF)$ and the
last line of Definition~\ref{definition_partial_trace}, there exists $\vv_p \in \NN_p(K)$
such that $\vv_{p}|^\ttau_\FF = \rr_\FF$. Consider an arbitrary function
$\pphi \in \HH(\ccurl,K)$ satisfying $\pphi|^\ttau_\FF = \rr_\FF$
and set $\widetilde \pphi \eq \pphi - \vv_p \in \HH(\ccurl,K)$. By linearity
we have $\widetilde \pphi|^\ttau_\FF = \mathbf 0$. Using again the fact that
$\vv_p$ is smooth (recall that it is a polynomial), we also have
\begin{equation*}
\RRR_F (\trc_K^\ttau (\vv_p)) = \rr_F \qquad \forall F \in \FF.
\end{equation*}
Thus, by linearity, it remains to show that $\RRR_F(\trc_K^\ttau (\widetilde \pphi)) = \mathbf 0$
for all $F \in \FF$. Recalling~\eqref{eq_definition_HH_Gamma_ccurl}, the identity
$\widetilde \pphi|^\ttau_\FF = \mathbf 0$ means that $\widetilde \pphi \in \HH_{\Gamma_\FF}(\ccurl,K)$.
By Lemma~\ref{lemma_density}, there exists a sequence
$(\widetilde \pphi_m)_{m\in\mathbb N} \subset \CC^\infty_{\Gamma_\FF}(\overline{K})$
that converges to $\widetilde \pphi$ in $\HH_{\Gamma_\FF}(\ccurl,K)$. Now consider a face $F \in \FF$.
Since each function $\widetilde \pphi_m$ is smooth, we easily see that
$\|\RRR_F(\trc^\ttau_K (\widetilde \pphi_m))\|_{\XX^{-1/2}(F)} = 0$.
Then, since the map
$\HH(\ccurl,K) \ni \vv \longmapsto \|\RRR_F(\trc_K^\ttau (\vv))\|_{\XX^{-1/2}(F)} \in \mathbb R$
is continuous, we have
\begin{equation*}
\|\RRR_F(\trc_K^\ttau (\widetilde \pphi))\|_{\XX^{-1/2}(F)} =
\lim_{m \to +\infty}
\|\RRR_F(\trc_K^\ttau (\widetilde \pphi_m))\|_{\XX^{-1/2}(F)} = 0,
\end{equation*}
so that $\RRR_F (\trc_K^\ttau (\widetilde \pphi)) = 0$, which concludes the proof.
\end{proof}
\subsection{Step 2: Minimization without curl constraints}
\label{sec_st_2}
To avoid subtle issues concerning the equivalence of norms, we first establish
the stability result concerning minimization without curl constraints on
the reference tetrahedron $\hK\subset \mathbb R^3$ with vertices $(1,0,0)$,
$(0,1,0)$, $(0,0,1)$, and $(0,0,0)$.
\begin{lemma}[Curl-free minimization, reference tetrahedron]
\label{lemma_lifting_boundary}
Let $\hK\subset \mathbb R^3$ be the reference tetrahedron and let
$\hFF \subseteq \FF_{\hK}$ be a nonempty (sub)set
of its faces. Then, for every polynomial degree $p \geq 0$ and
for all $\hrr_\hFF \in \NN_p^\ttau(\Gamma_\hFF)$ such that
$\scurl_{\hF} (\hrr_{\hF}) = 0$ for all $\hF \in \hFF$, the following holds:
\begin{equation}\label{eq:min_ref_curl_free}
\min_{\substack{
\vv_p \in \NN_p(\hK)
\\
\curl \vv_p = \mathbf 0
\\
\vv_p|^\ttau_{\hFF} = \hrr_\hFF
}}
\|\vv_p\|_{0,\hK}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,\hK)
\\
\curl \vv = \mathbf 0
\\
\vv|^\ttau_{\hFF} = \hrr_\hFF
}}
\|\vv\|_{0,\hK}.
\end{equation}
\end{lemma}
\begin{proof}
The proof proceeds in two steps. \\
1) Using a key lifting result that is a direct consequence
of~\cite{Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12}, let us first establish that
\begin{equation}
\label{eq_stability_dgs}
\min_{\substack{
\vv_p \in \NN_p(\hK)
\\
\curl \vv_p = \mathbf 0
\\
\RRR_{\hF}(\trc_\hK^\ttau (\vv_p)) = \hrr_{\hF} \; \forall \hF \in \hFF
}}
\|\vv_p\|_{0,\hK}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,\hK)
\\
\curl \vv = \mathbf 0
\\
\RRR_{\hF}(\trc_\hK^\ttau (\vv)) = \hrr_{\hF} \; \forall \hF \in \hFF
}}
\|\vv\|_{0,\hK}.
\end{equation}
Let us denote respectively by $\vv_p^\star \in \NN_p(\hK)$ and $\vv^\star \in \HH(\ccurl,\hK)$
the discrete and continuous minimizers.
Let us define $\ww^\star \eq \trc_{\hK}^\ttau (\vv^\star) \in \XX^{-1/2}(\partial \hK)$.
Since $\curl \vv^\star = \mathbf 0$, we have $\|\vv^\star\|_{\ccurl,\hK} = \|\vv^\star\|_{0,\hK}$,
and the definition~\eqref{eq_puotient_norm_XX} of the quotient norm
of $\XX^{-1/2}(\partial K)$ implies that
\begin{equation} \label{eq_wwstar}
\|\ww^\star\|_{\XX^{-1/2}(\partial \hK)} \leq \|\vv^\star\|_{0,\hK}.
\end{equation}
Since $\RRR_{\hF}(\trc_\hK^\ttau (\vv^{\star})) = \hrr_{\hF}$, we have
$\RRR_{\hF} (\ww^\star )= \hrr_{\hF}$ for all $\hF \in \hFF$.
We assume that the faces $\FF_{\hK}$ of $\hK$ are numbered as $\hF_1,\dots,\hF_4$
in such a way that the
$n \eq |\hFF|$ first faces are the elements of $\hFF$. We introduce a ``partial lifting''
$\tvv_p \in \NN_p(\hK)$ of $\ww^\star$ using
\cite[Equation (7.1)]{Demk_Gop_Sch_ext_II_09} but taking only
the $n$ first summands. Then, one sees from
\cite[Proof of Theorem 7.2]{Demk_Gop_Sch_ext_II_09} that
\begin{equation} \label{eq_st_DGS}
\|\tvv_p\|_{\ccurl,\hK} \lesssim \|\ww^\star\|_{\XX^{-1/2}(\partial \hK)}
\end{equation}
and $\RRR_{\hF}(\trc_\hK^\ttau (\tvv_p)) = \hrr_{\hF}$ for all $\hF \in \hFF$.
Thus, relying on~\eqref{eq:trace_DGS_NN} we have $\ppi^\ttau_\hF (\tvv_p) = \hrr_\hF$
for all $\hF \in \hFF$, and we notice that the last line of
Definition~\ref{definition_partial_trace} also equivalently gives
$\tvv_p|^\ttau_\hFF = \hrr_\hFF$.
We must now check that $\curl \tvv_p = \mathbf 0$. This is possible since the
$\HH(\ccurl,\hK)$ and $\HH(\ddiv,\hK)$ trace liftings introduced
in~\cite{Demk_Gop_Sch_ext_II_09,Demk_Gop_Sch_ext_III_12}
commute in appropriate sense. Specifically, recalling that
$\scurl_{\hF} (\hrr_{\hF}) = 0$ for all $\hF \in \hFF$, using the identity
$\scurl_{\hF} (\ppi^\ttau_\hF (\tvv_p)) = \curl \tvv_p \cdot \nn_{\hF}$
valid for all $\hF \in \FF_{\hK}$ (recall that $\nn_{\hF}$ conventionally
points outward $\hK$), see~\eqref{eq_scurl_curl_el}, and with the help of
Theorem 3.1 and Propositions 4.1, 5.1, and 6.1 of~\cite{Demk_Gop_Sch_ext_III_12},
one shows by induction on the summands that $\curl \tvv_p = \mathbf 0$.
Now, since $\tvv_p$ belongs to the discrete minimization set and
using~\eqref{eq_st_DGS} and~\eqref{eq_wwstar}, \eqref{eq_stability_dgs} follows from
\begin{equation*}
\|\vv_p^\star\|_{0,\hK}
\leq
\|\tvv_p\|_{0,\hK}
=
\|\tvv_p\|_{\ccurl,\hK}
\lesssim \|\ww^\star\|_{\XX^{-1/2}(\partial \hK)}
\leq
\|\vv^\star\|_{0,\hK}.
\end{equation*}
2) Let us now establish~\eqref{eq:min_ref_curl_free}.
We first invoke Lemma~\ref{lemma_trace_restriction}. If $\vv \in \HH(\ccurl,\hK)$
satisfies $\vv|^\ttau_\hFF = \hrr_\hFF$, it follows that
$\RRR_{\hF}(\trc_\hK^\ttau (\vv)) = \hrr_{\hF}$ for all $\hF \in \hFF$. As a result, we have
\begin{equation*}
\min_{\substack{
\vv \in \HH(\ccurl,\hK)
\\
\curl \vv = \mathbf 0
\\
\RRR_{\hF}(\trc_\hK^\ttau (\vv)) = \hrr_{\hF} \; \forall \hF \in \hFF
}}
\|\vv\|_{0,\hK}
\leq
\min_{\substack{
\vv \in \HH(\ccurl,\hK)
\\
\curl \vv = \mathbf 0
\\
\vv|^\ttau_\hFF = \hrr_\hFF
}}
\|\vv\|_{0,\hK},
\end{equation*}
the minimization set of the left-hand side being (possibly) larger.
Invoking~\eqref{eq_stability_dgs} then gives
\begin{equation*}
\min_{\substack{
\vv_p \in \NN_p(\hK)
\\
\curl \vv_p = \mathbf 0
\\
\RRR_{\hF}(\trc_\hK^\ttau (\vv_p)) = \hrr_{\hF} \; \forall \hF \in \hFF
}}
\|\vv_p\|_{0,\hK}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,\hK)
\\
\curl \vv = \mathbf 0
\\
\vv|^\ttau_\hFF = \hrr_\hFF
}}
\|\vv\|_{0,\hK},
\end{equation*}
and we conclude the proof by observing that
\begin{equation*}
\min_{\substack{
\vv_p \in \NN_p(\hK)
\\
\curl \vv_p = \mathbf 0
\\
\vv_p|^\ttau_\hFF = \hrr_\hFF
}}
\|\vv_p\|_{0,\hK}
=
\min_{\substack{
\vv_p \in \NN_p(\hK)
\\
\curl \vv_p = \mathbf 0
\\
\RRR_{\hF}(\trc_\hK^\ttau (\vv_p)) = \hrr_{\hF} \; \forall \hF \in \hFF
}}
\|\vv_p\|_{0,\hK},
\end{equation*}
the two notions of local trace being equivalent for the discrete functions in $\NN_p(\hK)$.
\end{proof}
To establish the counterpart of Lemma~\ref{lemma_lifting_boundary} in a generic non-degenerate tetrahedron
$K \subset \mathbb R^3$, we are going to invoke the covariant Piola mapping
(see, e.g., \cite[Section~7.2]{Ern_Guermond_FEs_I_2020}).
Consider any invertible affine geometric mapping $\TTT: \mathbb R^3 \to \mathbb R^3$
such that $\Kout = \TTT(\Kin)$. Let $\JJJ_{\TTT}$ be the (constant) Jacobian
matrix of $\TTT$ (we do not require that $\det \JJJ_{\TTT}$ is positive, and in any
case we have $|\det \JJJ_{\TTT}|=|K|/|\hK|$).
The affine mapping $\TTT$ can be identified by specifying
the image of each vertex of $\Kin$. The covariant Piola mapping
$\ppsi^{\mathrm{c}}_{\TTT}:\HH(\ccurl,\Kout) \to \HH(\ccurl,\Kin)$
is defined as follows:
\begin{equation} \label{eq:Piola_c}
\hvv \eq \ppsi^{\mathrm{c}}_{\TTT}(\vv)
=
\left (\JJJ_{\TTT}\right )^{T} \left (\vv \circ \TTT \right ).
\end{equation}
It is well-known that
$\ppsi^{\mathrm{c}}_{\TTT}$ maps bijectively $\NN_p(\Kout)$ to $\NN_p(\Kin)$
for any polynomial degree $p\ge0$. Moreover, for all $\vv\in \HH(\ccurl,\Kout)$, we have
\begin{equation}\label{eq:curl_free}
\curl \vv = \mathbf 0 \; \Longleftrightarrow \curl \hvv = \mathbf 0,
\end{equation}
as well as
the following $\LL^2$-stability properties:
\begin{equation}
\label{eq_stab_piola_L2}
\frac{\rho_{\Kout}}{h_{\Kin}} \|\vv\|_{0,\Kout}
\leq
|\det \JJJ_{\TTT}|^{\frac12} \|\ppsi_{\TTT}^{\mathrm c}(\vv)\|_{0,\Kin}
\leq
\frac{h_{\Kout}}{\rho_{\Kin}} \|\vv\|_{0,\Kout}.
\end{equation}
Finally the covariant Piola mapping preserves tangential traces. This implies
in particular that for all $F\in\FF_{K}$, setting $\hF \eq \TTT^{-1}(F)$, we have
for all $\vv \in \HH^1(K)$
\begin{equation} \label{equiv_trace_H1}
\ppi^{\ttau}_F(\vv)=\mathbf 0 \; \Longleftrightarrow \; \ppi^{\ttau}_{\hF}(\hvv)=\mathbf 0.
\end{equation}
Finally, for all $\vv\in \HH(\ccurl,\Kout)$, for every nonempty (sub)set
$\FF \subseteq \FF_K$, and for all $\rr_\FF \in \NN_p^{\ttau}(\Gamma_\FF)$, we have
\begin{equation} \label{equiv_trace_Hcurl}
\vv|^\ttau_{\FF}=\rr_\FF \; \Longleftrightarrow \; \hvv|^\ttau_\hFF=\hrr_\hFF,
\end{equation}
where $\hFF\eq \TTT^{-1}(\FF)$ and $\hrr_{\hFF} \in \NN_p^{\ttau}(\Gamma_\hFF)$
is defined such that $\hrr_\hF \eq (\hrr_\hFF)|_{\hF} \eq \ppi^\ttau_{\hF} (\hvv_p)$ for all
$\hF\in\hFF$, where $\hvv_p\eq \ppsi^{\mathrm{c}}_{\TTT}(\vv_p)$ and
$\vv_p$ is any function in $\NN_p^{\ttau}(K)$ such that
$\rr_F \eq (\rr_\FF)|_{F} \eq \ppi^\ttau_{F} (\vv_p)$ for all $F\in\FF$.
The equivalence~\eqref{equiv_trace_Hcurl} is established by using
Definition~\ref{definition_partial_trace}, the properties of
the covariant Piola mapping, and the fact that
$\pphi \in \HH_{\ttau,\FF^{\mathrm{c}}}^1(K)$ if and only if
$\ppsi^{\mathrm{c}}_{\TTT}(\pphi) \in \HH_{\ttau,\hFF^{\mathrm{c}}}^1(\hK)$,
which follows from~\eqref{equiv_trace_H1}.
\begin{lemma}[Curl-free minimization, generic tetrahedron]
\label{lemma_lifting_boundary_K}
Let $K \subset \mathbb R^3$ be a non-degenerate tetrahedron and let
$\FF \subseteq \FF_K$ be a nonempty (sub)set of its faces.
Then, for every polynomial degree $p \geq 0$ and for all
$\rr_\FF \in \NN_p^\ttau(\Gamma_\FF)$ such that
$\scurl_F (\rr_F)=0$ for all $F \in \FF$, the following holds:
\begin{equation}
\label{eq_minimization_element_K_face}
\min_{\substack{
\vv_p \in \NN_p(K)
\\
\curl \vv_p = \mathbf 0
\\
\vv_{p}|^\ttau_\FF = \rr_\FF
}}
\|\vv_p\|_{0,K}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \mathbf 0
\\
\vv|^\ttau_\FF = \rr_\FF
}}
\|\vv\|_{0,K}.
\end{equation}
\end{lemma}
\begin{proof}
Consider an invertible affine mapping $\TTT: \hK \to K$ and denote
$\ppsi^{\mathrm{c}}_{\TTT}$ the associated Piola mapping defined in~\eqref{eq:Piola_c}.
Let us set
\begin{alignat*}{2}
{\boldsymbol V}(\hK)&\eq\{\hvv\in \HH(\ccurl,\hK)\ | \ \curl\hvv=\mathbf 0,\ \hvv|^\ttau_\hFF=\hrr_\hFF\}, &\qquad {\boldsymbol V}_p(\hK)&\eq {\boldsymbol V}(\hK) \cap \NN_p(\hK),\\
{\boldsymbol V}(K)&\eq\{\vv\in \HH(\ccurl,K)\ | \ \curl\vv=\mathbf 0,\ \vv|^\ttau_\FF=\rr_\FF\}, &\qquad {\boldsymbol V}_p(K)&\eq {\boldsymbol V}(K) \cap \NN_p(K),
\end{alignat*}
where $\hrr_\hFF$ is defined from $\rr_\FF$ as above.
Owing to~\eqref{eq:curl_free} and~\eqref{equiv_trace_Hcurl}, we infer that
\begin{equation} \label{eq:identities_piola_V}
\ppsi_{\TTT}^{\mathrm c}({\boldsymbol V}(K))={\boldsymbol V}(\hK), \qquad
\ppsi_{\TTT}^{\mathrm c}({\boldsymbol V}_p(K))={\boldsymbol V}_p(\hK).
\end{equation}
One readily checks that $\hrr_\hFF$ satisfies the assumptions of
Lemma~\ref{lemma_lifting_boundary}, so that
\begin{equation*}
\min_{\hvv_p \in{\boldsymbol V}_p(\hK)} \|\hvv_p\|_{0,\hK}
\lesssim
\min_{\hvv \in{\boldsymbol V}(\hK)} \|\hvv\|_{0,\hK}.
\end{equation*}
Invoking the stability properties~\eqref{eq_stab_piola_L2} and the
identities~\eqref{eq:identities_piola_V}, we conclude that
\begin{align*}
\min_{\vv_p \in{\boldsymbol V}_p(K)} \|\vv_p\|_{0,K} &\le
\frac{h_{\Kin}}{\rho_{\Kout}} |\det \JJJ_{\TTT}|^{\frac12}
\min_{\hvv_p \in{\boldsymbol V}_p(\hK)} \|\hvv_p\|_{0,\hK} \\
&\lesssim \frac{h_{\Kin}}{\rho_{\Kout}} |\det \JJJ_{\TTT}|^{\frac12}
\min_{\hvv \in{\boldsymbol V}(\hK)} \|\hvv\|_{0,\hK}
\le \frac{h_{\Kin}}{\rho_{\Kout}} \frac{h_{\Kout}}{\rho_{\Kin}}
\min_{\vv \in{\boldsymbol V}(K)} \|\vv\|_{0,K}.
\end{align*}
This completes the proof.
\end{proof}
\subsection{Step 3: Conclusion of the proof}
\label{sec_st_3}
We are now ready to conclude the proof of Theorem~\ref{theorem_stability_tetrahedra}.
We first apply Lemma~\ref{lemma_lifting_curl} on the tetrahedron $K$ and infer that there exists
$\xxi_p \in \NN_p(K)$ such that $\curl \xxi_p = \rr_K$ and
\begin{equation*}
\|\xxi_p\|_{0,K} \lesssim \min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
}} \|\vv\|_{0,K}.
\end{equation*}
Then, we define $\trr_\FF \in \NN_p^\ttau(\Gamma_\FF)$ by
setting $\trr_F \eq \rr_F - \ppi^\ttau_{F} (\xxi_p)$
for all $F \in \FF$.
Since $\scurl _F(\ppi^\ttau_{F} (\xxi_p)) = \curl \xxi_p \cdot \nn_{F} = \rr_K \cdot \nn_{F}$,
we see that $\scurl_F (\widetilde{\rr}_F) = 0$ for all $F \in \FF$. It follows from
Lemma~\ref{lemma_lifting_boundary_K} that there exists
$\widetilde{\xxi}_p \in \NN_p(K)$ such that $\curl \widetilde{\xxi}_p = \mathbf 0$,
$\widetilde{\xxi}_p|^\ttau_\FF = \widetilde{\rr}_\FF$, and
\begin{equation*}
\|\widetilde{\xxi}_p\|_{0,K}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \mathbf 0
\\
\vv|^\ttau_\FF = \widetilde{\rr}_\FF
}}
\|\vv\|_{0,K}.
\end{equation*}
We then define $\ww_p \eq \xxi_p + \widetilde{\xxi}_p \in \NN_p(K)$. We observe that
$\ww_p$ belongs to the discrete minimization set of~\eqref{eq_minimization_element_K}.
Thus we have
\begin{equation*}
\min_{\substack{
\vv_p \in \NN_p(K)
\\
\curl \vv_p = \rr_K
\\
\vv_p|^\ttau_\FF = \rr_\FF
}}
\|\vv_p\|_{0,K}
\leq
\|\ww_p\|_{0,K}
\leq
\|\xxi_p\|_{0,K} + \|\widetilde{\xxi}_p\|_{0,K}.
\end{equation*}
Finally we observe that
\begin{equation*}
\|\xxi_p\|_{0,K}
\lesssim
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
}}
\|\vv\|_{0,K}
\leq
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
\\
\vv|^\tau_\FF = \rr_\FF
}}
\|\vv\|_{0,K},
\end{equation*}
and
\begin{align*}
\|\widetilde{\xxi}_p\|_{0,K}
& \lesssim
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \mathbf 0
\\
\vv|^\tau_\FF = \widetilde{\rr}_\FF
}}
\|\vv\|_{0,K}
=
\min_{\substack{
\widetilde{\vv} \in \HH(\ccurl,K)
\\
\curl \widetilde{\vv} = \curl \xxi_p
\\
\widetilde{\vv}|^\tau_\FF = \trr_\FF + \xxi_p|^\ttau_\FF
}}
\|\widetilde{\vv} - \xxi_p\|_{0,K}
\\
&
=
\min_{\substack{
\widetilde{\vv} \in \HH(\ccurl,K)
\\
\curl \widetilde{\vv} = \rr_K
\\
\tvv|^\tau_\FF = \rr_\FF
}}
\|\widetilde{\vv} - \xxi_p\|_{0,K}
\leq
\|\xxi_p\|_{0,K} +
\min_{\substack{
\vv \in \HH(\ccurl,K)
\\
\curl \vv = \rr_K
\\
\vv|^\tau_\FF = \rr_\FF
}}
\|\vv\|_{0,K}.
\end{align*}
\bibliographystyle{siam}
| b5b564c71818f18fa7319b67aadf1ff4f6bb2536 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Dissipative quantum systems are of a large importance as
they pave a way for manipulating quantum matter for preparation of pure \cite{diehl2008quantum} as well as highly entangled
\cite{lin2013dissipative} states, and implementation of quantum computations \cite{verstraete2009quantum}.
One particular example of dissipative quantum systems realized with cold atoms are open systems which can
exchange particles with a reservoir \cite{witthaut2008dissipation,prosen2010exact,barmettler2011controllable,brantut2012conduction,barontini2013controlling,
krinner2015observation,lebrat2018band},
so, neither the energy nor the particle number is preserved. When an open system is coupled with two reservoirs with different chemical potentials,
it realizes an atomtronic analogue \cite{Pepino10} of semiconductor devices \cite{brantut2012conduction,ivanov2013bosonic,krinner2015observation,lebrat2018band,kolovsky2018landauer}.
Alternatively a quantum lattice system can be coupled with the environment only in a single lattice site as experimentally demonstrated in \cite{labouvie2016bistability}.
Here we consider
decay of Bose particles from a triple quantum well with the central well coupled to environment,
i.e. the bosonic particles are drained into a reservoir \cite{Kordas15}.
Despite its simplicity the three site-model BH
does not allow for exact analytic solution \cite{nemoto00, franzosi01, franzosi03, Kordas15,Bychek19a}.
In this paper we employ pseudoclassical approach \cite{mahmud2005quantum,
mossmann2006semiclassical,graefe2007semiclassical,
trimborn2008exact, kolovsky2009bloch,zibold2010classical,bychek2018noon,Bychek19, Bychek20}
which allows us to cast the problem into a form of coupled driven nonlinear oscillators.
From the pure classical perspective this
system supports a non-decaying solution with equal intensities but opposite phases on the edge sites. Such a solution has a zero
amplitude at the central site, and, therefore, is not directly coupled to the reservoir.
This kind of localized solution, existing in the
system despite the loss channels is allowed, is known as a bound state in the continuum (BIC) \cite{Hsu16}.
In particular, the BIC to be considered in the present work is analogous to that supported by a pair of side-defects coupled to
photonic crystalline waveguides \cite{Bulgakov11, Bulgakov11a}. From the quantum mechanical perspective the discussed BIC is
the antisymmetric Bose-Einstein condensate (BEC) with particles occupying only the edge sites. The central problem to be addressed in the work is the
account of the inter-particle interaction. We shall examine the stability of non-linear BIC in the classical
regime and investigate the link between the classical and quantum solutions. It shall be demonstrated that even a classically stable nonlinear BIC
undergoes a slow rate decay due to quantum fluctuations. Thus, the antisymmetric BEC is always a metastable state.
We shall show that the quantum fluctuations can be accurately
described within the pseudoclassical framework by the stochastic force emerging in the
nonlinear coupled oscillator model.
\section{Master equation and pseudoclassical approach}
\begin{figure*}[t]
\includegraphics[width=0.7\textwidth,height=0.36\textwidth,trim={0.0cm 0.0cm 0.0cm 0.0cm},clip]{Fig1.eps}
\caption{Sketch of the system.} \label{fig1}
\end{figure*}
We consider a linear trimer of three coupled potential wells. The trimer is initially occupied by $N_0$ bosons which can tunnel between the wells as shown
in Fig. \ref{fig1}. The tunnelling dynamics is controlled by the Bose-Hubbard Hamiltonian
\begin{align}\label{Hamiltonian}
&\widehat{{\cal H}}=
-\frac{J}{2}\sum_{\ell=1}^{2}\left(\hat{a}_{\ell+1}^{\dagger}\hat{a}_{\ell} +{\rm h.c.}\right) +\frac{U}{2}\sum_{\ell=1}^3\hat{n}_{\ell}
(\hat{n}_{\ell}-1),
\end{align}
where $\hat{a}_{\ell}^{\dagger}(\hat{a}_{\ell})$ is the creation (annihilation) operator at the $\ell_{\rm th}$ site, $\hat{n}_{\ell}$ is the number operator
at the $\ell_{\rm th}$ site, $J$ is the interwell tunnelling rate and $U$ is the interaction constant.
By now the BH model has grown to one of the seminal model in physics of cold atoms which scopes quantum phase transitions \cite{greiner2002quantum},
the effects of Josephson oscillations \cite{Gati07, Esteve08}, atomic Bloch oscillations \cite{meinert2014interaction,fujiwara2019transport},
refill dynamics in the presence of induced losses, \cite{labouvie2015negative, labouvie2016bistability}, spontaneous breaking of the symmetry \cite{trenkwalder2016quantum},
and quantized current in the
engineered transport channel \cite{brantut2012conduction,krinner2015observation,lebrat2018band} to mention a few results relevant to
the present paper. Here, following \cite{Kordas15}, we assume that the central well is attached to an atom sink as shown in Fig. \ref{fig1}.
Then the system dynamics is described by the density matrix $\widehat{{\cal R}}$ which obeys the master equation
\begin{equation}\label{master}
\frac{\partial \widehat{{\cal R}}}{\partial t}=-i[\widehat{{\cal H}}, \widehat{{\cal R}}]+
\widehat{{\cal L}}(\widehat{\cal R})
\end{equation}
with the loss operator of a Lindblad form
\begin{equation}\label{Liouvillian}
\widehat{{\cal L}}(\widehat{\cal R})=-\frac{\gamma}{2}
\left(\hat{a}_{2}^{\dagger}\hat{a}_{2}\widehat{\cal R }-2\hat{a}_{2}\widehat{\cal R }\hat{a}_{2}^{\dagger}
+\widehat{\cal R }\hat{a}_{2}^{\dagger}\hat{a}_{2} \right),
\end{equation}
where $\gamma$ is the loss rate.
The pseudoclassical approach is introduced by replacing each operator $\widehat{A}$ by its Weyl
symbol which is a function on the phase space \cite{McDonald88},
\begin{equation}
{\rm symb}[\widehat{A}]=A( a,{a}^*),
\end{equation}
with $a, a^*$ as the complex conjugated canonical variables
defined as the Weyl symbols of the annihilation and creation operators
\begin{equation}\label{conjugated}
{\rm symb}[\hat{a}]=a, \ {\rm symb}[\hat{a}^{\dagger}]=a^{*},
\end{equation}
where we omitted the subindex $\ell$ for simplicity. The Weyl symbols of an operator product of
two operators are computed via the Moyal star product of the Weyl symbols of the two operators
\begin{equation}\label{Moyal}
A\star
B=A\exp\left[\frac{\hbar}{2}\left(\frac{\partial^{\leftarrow}}{\partial
a}\frac{\partial^{\rightarrow}}{\partial a^*}-
\frac{\partial^{\leftarrow}}{\partial
a^*}\frac{\partial^{\rightarrow}}{\partial a} \right)\right]B.
\end{equation}
For instance, it is easy to see from Eq. (\ref{Moyal}) that the Weil symbol of the number operator is
\begin{equation}\label{number}
{\rm symb}[\hat{n}]=a^*\star a=|a|^2-\frac{1}{2}.
\end{equation}
The figure of merit in the pseudoclassical approach is the Weyl symbol of the density matrix known as the Wigner function
\begin{equation}
{\cal W}={\rm symb}[\widehat{{\cal R}}].
\end{equation}
Applying Eq. (\ref{Moyal}) to the master equation, Eq. (\ref{master}) one finds that the Wigner function obeys the following equation
\cite{Vogel88, Bortman95}
\begin{align}
& \frac{\partial {\cal W}}{\partial t}=-i\sum_{\ell=1}^{3}\left[U\left (1-|\alpha_{\ell}^2|\right)\left(\alpha_{\ell}\frac{\partial
{\cal W}}{\partial \alpha_{\ell}}
-\alpha_{\ell}^*\frac{\partial
{\cal W}}{\partial \alpha_{\ell}^*}\right)
-\frac{U}{4}\left(\frac{\partial^3 \alpha_{\ell}^*{\cal W}}{\partial \alpha_{\ell}\partial{\alpha_{\ell}^*}^2}
-\frac{\partial^3 \alpha_{\ell}{\cal W}} {\partial \alpha_{\ell}^2\partial\alpha_{\ell}^*}
\right)
\right] \nonumber\\
& -i\frac{J}{2}\sum_{\ell=1}^{2}\left(
\alpha_{\ell+1} \frac{\partial {\cal W}}{\partial \alpha_{\ell}}
+\alpha_{\ell} \frac{\partial {\cal W}}{\partial \alpha_{\ell+1}}
-\alpha_{\ell+1}^* \frac{\partial {\cal W}}{\partial \alpha_{\ell}^*}
-\alpha_{\ell}^* \frac{\partial {\cal W}}{\partial \alpha_{\ell+1}^*}
\right)
\nonumber \\
& +\frac{\gamma}{2}\left(\alpha_2\frac{\partial {\cal W}}{\partial \alpha_2} +2{\cal W} + \alpha_2^*\frac{\partial {\cal W}}{\partial \alpha_2^*}\right)
+\frac{\gamma}{2}\frac{\partial^2 {\cal W}}{\partial \alpha_2
\partial \alpha_2^*}.
\label{FP_W}
\end{align}
The above equation contains third order derivatives which do no allow to interpret it
as a Fokker-Plank equation with a positive definite or positive semidefinite diffusion matrix \cite{Vogel88}.
The pseudoclassical limit of Eq. (\ref{FP_W}) is obtained by setting $N_0\rightarrow \infty$ while keeping $g=UN_0={\rm Const}$.
In what follows the quantity $g$ will be referred to as the macroscopic interaction constant. Let us apply the following substitution
\begin{equation}\label{substitution}
\alpha_{\ell}=\sqrt{N_0}a_{\ell}, \ \ \alpha^*_{\ell}=\sqrt{N_0}a^*_{\ell} .
\end{equation}
Then Eq. (\ref{FP_W}) transforms to
\begin{align}
& \frac{\partial {\cal W}}{\partial t}=-i\sum_{\ell=1}^{3}\left[g\left (\frac{1}{N_0}-|a_{\ell}^2|\right)\left(a_{\ell}\frac{\partial
{\cal W}}{\partial a_{\ell}} -a_{\ell}^*\frac{\partial \cal {\cal W}}{\partial a_{\ell}^*}\right) \right]
\nonumber\\
& -i\frac{J}{2}\sum_{\ell=1}^{2}\left(
a_{\ell+1} \frac{\partial {\cal W}}{\partial a_{\ell}}
+a_{\ell} \frac{\partial {\cal W}}{\partial a_{\ell+1}}
-a_{\ell+1}^* \frac{\partial {\cal W}}{\partial a_{\ell}^*}
-a_{\ell}^* \frac{\partial {\cal W}}{\partial a_{\ell+1}^*} \right)
\nonumber \\
& +\frac{\gamma}{2}\left(a_2\frac{\partial {\cal W}}{\partial a_2} + 2{\cal W}+a_2^*\frac{\partial {\cal W}}{\partial a_2^*}\right)
+\frac{\gamma}{2N_0}\frac{\partial^2 {\cal W}}{\partial a_2
\partial a_2^*}+{\cal O}(N_0^{-2}).
\label{FP_W_classical}
\end{align}
Neglecting ${\cal O}(N_0^{-2})$ term we arrive at a true Fokker-Plank equation where the first term in the third line can be
viewed as dissipation while the second term in the same line is diffusion \cite{Bychek19, Bychek20}.
The dynamics under the Fokker-Planck equation, Eq. (\ref{FP_W_classical}) can be unravelled into a set of dissipative Langevin equations
\begin{align}
& id{a}_1=\left(-\frac{J}{2}a_2 +g|a_1|^2 a_1\right)dt, \nonumber \\
& id{a}_2=\left[-\frac{J}{2}\left(a_1+a_3\right) +g|a_2|^2 a_2 -i\frac{\gamma}{2} a_2\right]dt
+\sqrt{\frac{\gamma}{2N_0}}d\xi, \nonumber \\
& id{a}_3=\left(-\frac{J}{2}a_2 +g|a_3|^2 a_3\right)dt, \label{Langevin}
\end{align}
where $d\xi$ is the complex white noise
\begin{equation}\label{noise}
\overline{d\xi}=0, \ \overline{d\xi^{*} d\xi}=dt, \ \overline{d\xi d\xi}=0.
\end{equation}
%
Notice that compared to Eq. (\ref{FP_W}) in Eq. (\ref{Langevin}) we omitted the "self-energy" term proportional to $g/N_0$. This can be done as the oscillating factor $\exp(-igt/N_0)$ can be absorbed into the noise Eq. (\ref{noise}) without changing its correlation properties.
Let us assume for a moment that there is no noise term in Eq. (\ref{Langevin}). Then Eq. (\ref{Langevin}) has a antisymmetric solution decoupled from the lossy site
\begin{equation}
{\bm a}_{\rm BIC}(t)=
e^{-igIt}
\left(
\begin{array}{c}
\sqrt{I} \\
0 \\
-\sqrt{I}
\end{array}
\right) \label{BIC}
\end{equation}
where intensity $I$ can be linked to the mean population of the edge sites $\overline{n}_{1,2}=IN_0+{1}/{2}$. By examination of
Eq.~(\ref{Langevin}) one immediately identifies the three factors affecting the decay dynamics of this state:
(i) The stability of the BIC. If Eq.(\ref{BIC}) is unstable, it can be destroyed by
small perturbations.
(ii) The initial condition for solving Eq. (\ref{Langevin}). In more detail, we expect that the decay rate is dependent
on how close the initial condition is to the symmetry protected BIC, Eq. (\ref{BIC}).
Moreover in establishing quantum to classical correspondence one can not deal with a single trajectory but with
the ensemble of trajectories whose initial conditions are determined by the initial many-body quantum state of the system \cite{kolovsky2009bloch}.
(iii) The noise term in Eq. (\ref{Langevin}) inversely proportional to $\sqrt{N_0}$. The
noise can perturb even an intrinsically stable state driving it out of equilibrium.
Notice
that even though the reservoir does not supply particles into the system a stochastic driving force is still present in Eq. (\ref{Langevin}).
Physically, this intrinsic noise is nothing but the quantum fluctuations arising from the noncommutativity
of creation and annihilation operators. The noise term is important for the correct application of the pseudoclassical approach.
For example, in the paradigm problem of the decaying quantum oscillator it ensures that the oscillator does not decay
below its ground energy.
In the next section we discuss each of these factors in more detail.
\section{Decay of the antisymmetric state}
\subsection{Stability analysis}
First we analyze the stability of the solution ${\bm a}_{\rm BIC}(t)$ for $g\ne0$.
Using the standard stability analysis \cite{lichtenberg2013regular}
the stability of
this solution can be examined by analyzing the matrix
\begin{equation}\label{M}
\widehat{M}=
\left(
\begin{array}{cccccc}
gI & -J/2 & 0 & gI & 0 & 0 \\
-J/2 & -i\gamma/2-gI & -J/2 & 0 & 0 & 0 \\
0 & -J/2 & gI & 0 & 0 & gI \\
-gI & 0 & 0 & gI & J/2 & 0 \\
0 & 0 & 0 &J/2 & -i\gamma/2+gI & J/2 \\
0 & 0 & -gI & 0 & J/2 & gI \\
\end{array}
\right).
\end{equation}
If the imaginary part of all eigenvalues of $ \widehat{M}$ is non-positive, the BIC solution is stable.
Fig. 1 (a) shows the imaginary parts of the eigenvalues as the function of $gI$ for $J=1$ and $\gamma=0.4$. It is seen that
in this case the stability threshold corresponds to $gI=0.2$. Above the threshold any tiny imbalance in the population
of the edge sites will leads to excitation of the symmetric modes and the BIC looses its intensity.
This process is exponential in time resulting in a rapid drop
of intensity at the initial stage. However, once the stability threshold is crossed the solution ${\bm a}_{\rm BIC}(t)$ stabilizes
at a certain value of intensity $I_{\rm st}$,
\begin{equation}\label{thres}
I_{\rm st}=I_{\rm st}(\gamma,g) .
\end{equation}
In what follows we shall refer to Eq.~(\ref{thres}) as the stabilization level. We mention that, as expected, the stabilization level is
approximately inversely proportional to $g$ yet it is always smaller than the stability threshold deduced from Eq.~(\ref{M}).
The described scenario is exemplified by thin solid lines in Fig.~2 (d, e) where, to provoke the symmetry breaking, we introduced a
tiny population imbalance $\sim10^{-3}$ in the initial BIC state. The exponential decrease of intensity for $g$ above the stability threshold
and the effect of stabilization is clearly seen in the figure.
\begin{figure*}[t]
\includegraphics[width=1\textwidth,height=0.72\textwidth,trim={3.0cm 0.0cm 0.0cm 0.0cm},clip]{Fig2.eps}
\caption{(a) Imaginary parts $\Im\left\{\lambda_n \right\}$ of eigenvalues of matrix Eq. (\ref{M}) for $\gamma=0.4$.
(b, c) Initial conditions for the antisymmetric BEC in the space of populations of the first, $n_1$ and the third, $n_3$ sites for (b) $N_0=20$ and (c) $N_0=100$.
(d, e) Decay dynamics of the antisymmetric state with (d) $N_0=20$ and (e) $N_0=100$.
Thick dashed lines show the total number of particle against time for antisymmetric BEC initial conditions shown in subplot (d, c).
Thin solid lines show the decay of the classical
BIC state.} \label{fig2}
\end{figure*}
\subsection{Quantum ensemble}
Before simulating the decay of truly quantum states we have to introduce an ensemble classical initial condition
corresponding to a quantum state loaded into the system.
This, can be done by using the Husimi ${\cal Q}$-function,
\begin{equation}\label{Husimi}
{\cal Q}({\bm \alpha})=\frac{1}{\pi^3}\langle{\bm \alpha}| \widehat{\cal R}|{\bm \alpha}\rangle,
\end{equation}
where $|{\bm \alpha}\rangle$ is the Glauber coherent state,
\begin{equation}\label{coherent}
|{\bm \alpha}\rangle=e^{-\frac{|\alpha_1|^2+|\alpha_2|^2+
|\alpha_3|^2}{2}}e^{\alpha_1\hat{a}_1^{\dagger}+\alpha_2\hat{a}_2^{\dagger}+\alpha_3\hat{a}_3^{\dagger}}|{\rm vac} \rangle
\end{equation}
with ${\bm \alpha}=\left\{ \alpha_1, \alpha_2, \alpha_3\right\}$. At first, for the initial state we choose an antisymmetric $N$-particle BEC.
However, for the future convenience below we
present a single formula for both symmetric, $|\Psi^{(+)}_{\rm BEC}\rangle$ and antisymmetric $|\Psi^{(-)}_{\rm BEC}\rangle$
condensates
\begin{equation}\label{BEC}
|\Psi^{(\pm)}_{\rm BEC}\rangle=\frac{1}{\sqrt{2^NN!}}\left({\hat{a}_1^\dag \pm \hat{a}_3^\dag}\right)^N |{\rm vac} \rangle.
\end{equation}
After applying the Husimi transformation Eq. (\ref{Husimi}) one finds
\begin{equation}\label{BECH}
{\cal Q}^{(\pm)}_{\rm BEC}({\bm \alpha})=\frac{|\alpha_1\pm\alpha_3|^{2N}}{\pi^3 2^N(N!)}e^{-|\alpha_1|^2-|\alpha_2|^2-
|\alpha_3|^2}.
\end{equation}
By applying the acception-rejection method \cite{kolovsky2009bloch} we generate the ensembles of the initial conditions
according to the distribution function (\ref{BECH}) for $N_0=20$ and $N_0=100$, see panels (b) and (c) in Fig.~2, and
then simulate the system dynamics. The result is depicted by the thick dashed lines in panels (d) and (e). Remarkably,
in the unstable cases ($g>0.4$) we get the same fraction of bosons which is left in the system as it is predicted by
the pure classical stability analysis given in the previous subsection. In the stable case $g=0.4$, however, we observe essential
deviations. These can be understood by noticing that every initial condition ${\bf a}(t=0)$ from the quantum ensemble is a
superposition of the system linear eigenmodes,
\begin{equation}
{\bm b}_{1}=
\left(
\begin{array}{c}
\frac{1}{\sqrt{2}} \\
0 \\
\frac{-1}{\sqrt{2}}
\end{array}
\right), \
{\bm b}_{2}=
\left(
\begin{array}{c}
\frac{1}{{2}} \\
\frac{1}{\sqrt{2}} \\
\frac{1}{{2}}
\end{array}
\right), \
{\bm b}_{3}=
\left(
\begin{array}{c}
\frac{1}{{2}} \\
\frac{-1}{\sqrt{2}} \\
\frac{1}{{2}}
\end{array}
\right) \label{eigenmodes}
\end{equation}
where the first eigenmode obviously corresponds to the symmetry protected BIC while the other two modes are coupled to the reservoir and decay within the characteristic time $2\pi/\gamma$. Thus, the solid and dashed lines in Fig.~2(d-e) may coincide only in the limit $N_0\rightarrow\infty$ where the quantum ensemble shrinks to the single point.
\subsection{The Role of the Noise}
Now let us return to the Langevin dynamics governed by Eq. (\ref{Langevin}) where one could expect that even a stable
antysimmetric BIC state Eq. (\ref{BIC}) is subject to decay. To test this conjecture we solve numerically both the Langeven equation and
the exact master equation, Eq. (\ref{master}). The results are shown in Fig.~\ref{fig3}. In Fig. \ref{fig3}(a) we depict the exact quantum
solution for the total population. One can see that unlike the classical solutions in Fig. \ref{fig2} the populations now continues to
decay even after crossing the stabilization level. This decay is still exponential, however, with much smaller rate. In Fig. \ref{fig3}(b)
we compare the population dynamics for the first and the second sites obtained by the pseudoclassical and quantum approaches. One can see that
the two results are in a good agreement. This supports our conjecture that the noise destroys the classical symmetry protected BIC.
\begin{figure*}[t]
\includegraphics[width=0.8\textwidth,height=0.48\textwidth,trim={0.0cm 0.0cm 0.0cm 1.4cm},clip]{Fig3.eps}
\caption{Decay of the antisymmetric condensate for $N_0=20$.
(a) Logarithmic plot of the full population against time obtained by solving the master equation Eq.
(\ref{master}). The thin dash lines show the stabilization levels from Fig. \ref{fig2} (d).
(b) Populations of the first and the second sites against time. Here and in the panels (c, d) thin lines are quantum simulations while thick grey lines are
the results computed with the pseudoclassical approach, $g=0.8$.
(c) Normalized eigenvalues of the single particle density matrix Eq. (\ref{single_density}), $g=0.8$.
(c) Correlation function Eq. (\ref{corfun}), $g=0.8$. } \label{fig3}
\end{figure*}
To look at the decay dynamics in more detail we compute the single particle density matrix $\hat{\rho}$, whose matrix elements are defined as
\begin{equation}\label{single_density}
\rho_{\ell,\ell'}={\rm Tr}\left(\hat{a}_{\ell}^{\dagger}\hat{a}_{\ell'}\widehat{\cal R }\right).
\end{equation}
In the pseudoclassical framework this matrix corresponds to the correlation functions
\begin{equation}
\rho_{\ell,\ell'}=\langle a_{\ell}^*a_{\ell'}\rangle-\frac{1}{2}\delta_{\ell,\ell'},
\end{equation}
where the pointy brackets designate the ensemble average over Langevin trajectories. The single particle density matrix allows us to test whether the quantum state remains a BEC during the decay \cite{Mueller06}. Namely, if all but one eigenvalues are zero the system
is a condensate state. Another related quantity is the correlation function \cite{Kordas15},
\begin{equation}\label{corfun}
C(t)=\frac{\rho_{1,3}}{\sqrt{ \rho_{1,1}\rho_{3,3}}}
=\frac{\langle a_1^* a_3 \rangle}{\sqrt{\langle |a_1|^2 \rangle \langle |a_3|^2 \rangle}}.
\end{equation}
If $C(t)=-1$ the system is an antisymmetric condensate and if $C(t)=1$, then the condensate is symmetric. In Fig. \ref{fig3} (c) we show the
normalized eigenvalues of $\hat{\rho}$, while in Fig. \ref{fig3} (d) we plotted the correlation functions, Eq. (\ref{corfun}). Both subplots are consistent with the decay dynamics described above: First, the system is rapidly departs from antisymmetric BEC. Then,
after the stabilization level is crossed the system recoheres into antisymmetric BEC again and slowly decays as the metastable
antisymmetric solution.
\subsection{Symmetric BEC}
To see the full picture in this section we also present results on the decay of symmetric BEC state, Eq. (\ref{BEC}). One can see from Fig. \ref{fig4}
that in the course of evolution the symmetric BEC state rapidly drops intensity and decoheres into a fractional condensate
with two non-zero eigenvalues of the single particle density matrix. Eventually, only a tiny fraction of
the initial population survives after the system transits into a pure antisymmetric BEC well below the stability threshold.
Again we see a good coincidence between the quantum and pseudoclassical results.
\begin{figure*}[t]
\includegraphics[width=0.8\textwidth,height=0.48\textwidth,trim={0.0cm 0.0cm 0.0cm 1.8cm},clip]{Fig4.eps}
\caption{Decay of symmetric condensate for $g=0.8$ and $N_0=20$.
(a) The full population against time as obtained by solving the master equation Eq. (\ref{master}).
(b) Populations of the first and the second wells against time; thin line -- quantum simulations, thick grey lines -- pseudoclassical approach.
(c) Normalized eigenvalues of the single particle density matrix Eq. (\ref{single_density}).
(d) Correlation function Eq. (\ref{corfun}).} \label{fig4}
\end{figure*}%
\begin{figure*}[t]
\includegraphics[width=1\textwidth,height=0.8\textwidth,trim={0.0cm 0.0cm 0.0cm 0cm},clip]{Fig5.eps}
\caption{Decay of fractional condensates, $N_0=20, \ T=2\pi \ {\rm (p.d.u.)}$; thick grey lines show the result by the pseudoclassical
approach, thin blue - lines the solutions of the master equations. (a-c) Dynamics of the full population. The insets show
the ensembles of initial conditions. (d-f) Normalized eigenvalues of the single particle density matrix for (d) Fock state, (e) symmetric NOON state, and (f) antisymmetric NOON state. (g-i) Correlation function, Eq. (\ref{corfun}) for (g) Fock state, (h) symmetric NOON state, and (i) antisymmetric NOON state.} \label{fig5}
\end{figure*}
\section{Decay of fractional condensates}
Next we examine the decay dynamics of fractional condensate states. The fractional condensate states are defined as those having more than one non-zero eigenvalues of the single particle density matrix \cite{Mueller06}. The most obvious example of a fractional condensate is a Fock state
\begin{equation}\label{Fock}
|\Psi_{\rm Fock}\rangle=\frac{1}{(N/2)!}\left({\hat{a}_1^\dag\hat{a}_3^\dag}\right)^{(N/2)} |{\rm vac} \rangle,
\end{equation}
where $N_0/2$ particles occupy the first site and the rest $N_0/2$ particle the third site. Directly applying Eq. (\ref{Husimi}) one finds
\begin{equation}\label{FockH}
{\cal Q}_{\rm Fock}({\bm \alpha})=\frac{|\alpha_1|^N|\alpha_3|^N}{\pi^3 [(N/2)!]^2}e^{-|\alpha_1|^2-|\alpha_2|^2-
|\alpha_3|^2}.
\end{equation}
Another, less trivial, example is the (anti-)symmetric NOON state which is a Shr\"odinger cat state of $N$ bosons in two wells,
\begin{equation}\label{NOON}
|\Psi^{(\pm)}_{\rm NOON}\rangle=\frac{1}{\sqrt{2(N!)}}\left[(\hat{a}_1^\dag)^N\pm(\hat{a}_3^\dag)^N\right] |{\rm vac} \rangle.
\end{equation}
This state has the following ${\cal Q}$-function
\begin{equation}\label{NOONH}
{\cal Q}^{(\pm)}_{\rm NOON}({\bm \alpha})=
\frac{|\alpha_1|^{2N}+|\alpha_3|^{2N}\pm(\alpha_1\alpha_3^*)^N\pm(\alpha_1^*\alpha_3)^N}{\pi^3 2(N!)}e^{-|\alpha_1|^2-|\alpha_2|^2-
|\alpha_3|^2}.
\end{equation}
In Fig. \ref{fig5} we show the simulation results by both pure quantum and pseudoclassical approaches.
One can see that despite the profound difference between the Fock state and the (anti-)symmetric NOON states clearly
seen in insets in Fig. \ref{fig5} (a-c), the decay dynamics is essentially identical. In all cases we see a rapid decay
of a fractional state below the stability threshold after which the system recoheres to the antisymmetric BEC having lost
the major part of the initial population. As before we see a good accuracy of the pseudoclassical approach.
\section{Summary and Conclusion}
We have examined the decay dynamics of quantum states with a definite number of bosons in three well open Bose-Hubbard model.
It is demonstrate that the stability of the quantum state can be predicted from the classical perspective. The decay
scenarios are drastically different depending on whether the solution is stable in the pseudoclassical limit.
In particular it is shown that in the pseudoclassical regime the antisymmetric BEC state is mapped to
a symmetry protected bound state in the continuum (BIC). The BIC is only stable below a certain intensity threshold.
Above the classical stability threshold the antisymmetric BEC rapidly decays and decoheres due to inter-particle interactions.
Once the population has dropped below the threshold, however, the system recoheres to the antisymmetric BEC which decays at much slower rate due
to the quantum fluctuations. It is demonstrated that the quantum fluctuations can be accurately described
in the pseudoclassical framework by introducing a stochastic force with amplitude inversely proportional to the square root of the initial number
of particles.
The pseudoclassical approach has been applied to several types of initial states with initial population only at the edge sites.
Besides the antisymmetric BEC we have studied the decay of symmetric BEC, Fock, (anti-)symmetric NOON states. In all cases the
initial bosonic cloud rapidly losses population to the reservoir and the decay can only slow down well below the
classical stability threshold when the system recoheres to the metastable antisymmetric BEC. In all cases we observed
a good coincidence between the numerical data obtained by the pseudoclassical approach and direct quantum simulations.
Recently, we have seen a surge of interest to decay dynamics
of two-photon states \cite{Crespi15, Chen18, Zhang20, Poddubny20}. We believe that
the approach presented here provides the key to understanding the decay dynamics in the other
solvable limit, namely, pseudoclassical regime.
Finally, we would like to outline the future fork ensuing from the present paper.
It is remains a question whether the asymptotic law of below threshold decay can be derived from the Langevin
equations, Eq. (\ref{Langevin}) or the corresponding Fokker-Plank equation.
We speculate that this problem may pose an interesting topic for future research.
This work has been supported by Russian Science
Foundation through grant N19-12-00167. We appreciate discussions with A.F. Sadreev and E.N. Bulgakov.
We are also grateful to G.P. Fedorov for his critical reading of the manuscript.
| eccb32bdd948556eab6e7bb693add54440dcb30e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the past decade, progress in deep neural networks has resulted in the advancements in various computer vision tasks, such as image classification~\cite{alexnet, vgg, multicolumndnn, stochasticdepth}, object detection~\cite{frcnn, ssd}, and segmentation~\cite{maskrcnn}. The big success of deep neural networks is mainly contributed to the well-designed cells and sophisticated architectures. For example, VGGNet~\cite{vgg} suggested the use of smaller convolutional filters and stacked a series of convolution layers for achieving higher performance, ResNet~\cite{resnet} introduced the residual blocks to benefit the training of deeper neural networks, and DenseNet~\cite{densenet} designed the densely connected blocks to stack features from different depths. Besides the efforts on the initial architecture design, extensive experiments are often required to determine the weights and hyperparameters of the deep neural network.
To automatically and efficiently search for neural networks of desireable properties (\emph{e.g.}, model size and FLOPs) from a predefined search space, a number of Neural Architecture Search (NAS) algorithms~\cite{darts, yang2019evaluation, li2020random, yu2019evaluating, nasbench101, nasbench201} have been recently proposed. Wherein, Evolutionary Algorithm (EA) based methods~\cite{amoebanet} maintain a set of architectures and generate new architectures using genetic operations like mutation and crossover. Reinforcement Learning (RL) based methods~\cite{zoph_rl_iclr, nasnet} sample architectures from the search space and train the controllers accordingly. The differentiable based methods~\cite{darts, fbnet, snas, metaarchitecturesearch} optimize the shared weights and architecture parameters, which significantly reduces the demand for computation resources and makes the search process efficient.
\begin{figure*}[t]
\centering
\iffalse
\subfigure[]{\label{fig_resnet}\includegraphics[height=5cm]{figs/fig_demo_multipathnetwork.png}}
\subfigure[]{\label{fig_resnet_hourglass}\includegraphics[height=5cm]{figs/fig_hourglass.png}}
\hspace{20pt}
\subfigure[]{\label{fig_critical}\includegraphics[height=5cm]{figs/fig_critical.png}}
\subfigure[]{\label{fig_minimal_supernet}\includegraphics[height=5cm]{figs/fig_demo_criticallayersearch.png}}
\subfigure[]{\label{fig_non_critical_supernet}\includegraphics[height=5cm]{figs/fig_demo_noncriticallayersearch.png}}
\subfigure[]{\label{fig_searched}\includegraphics[height=5cm]{figs/fig_demo_searched.png}}
\subfigure{\label{fig_search_space}\includegraphics[width=0.7\linewidth]{figs/fig_demo_notations.png}}
\fi
\subfigure{\label{fig_search_space}\includegraphics[width=0.9\linewidth]{overall.png}}
\caption{Blocks in the residual network are either ``vital'' or ``non-vital'', and they form the neck or bulb parts in the hourglass network. Two-stage search scheme speed up architecture search by 9$\times$ and the resource constrained search further accelerates architecture search by 72$\times$.}
\label{fig_HourNAS}
\vspace{-10pt}
\end{figure*}
These methods have made tremendous efforts to greatly accelerate the search process of NAS. Nevertheless, given the huge computation cost on the large-scale dataset~\cite{fbnet,sposnas,resourceconstrainednas,amc,onceforall}, \emph{e.g.}, 9 GPU days for NAS on the ImageNet benchmark, most methods execute NAS in a compromised way. The architecture is first searched on a smaller dataset (\emph{e.g.}, CIFAR-10 ~\cite{cifar}), and then the network weight of the derived architecture is trained on the large dataset. An obvious disadvantage of this concession is that the performance of the selected architecture on the CIFAR-10 may not be well extended to the ImageNet benchmark~\cite{undersandingrobustfyingnas}. We tend to use the minimal number of parameters without discrimination to construct an architecture that would achieve the maximal accuracy. But just as the popular 80-20 rule\footnote{{\color{black} The 80/20 rule (a.k.a Pareto principle, the law of vital few) is an aphorism that states, for many events, roughly 80\% of the effects come from 20\% of the causes.}} goes, only a few parts could be critical to the architecture's success, and we need to give them the most focus while balancing the parameter volume for other parts.
In this paper, we propose HourNAS for an accurate and efficient architecture search on the large-scale ImageNet dataset. Blocks in an architecture are not created equally. Given all the possible paths for the information flow from the input to the output of the network, blocks shared by these paths are \emph{vital}, just as the neck of an hourglass to regulate the flow. We identify these vital blocks and make them the priority in the architecture search. The other non-vital blocks may not be critical to the accuracy of the architecture, but they often occupy the major volume of the architecture (like the bulb of an hourglass) and will carve up the resource budget left by the vital blocks. Instead of directly working in a large search space flooding with architectures that obviously do not comply with the constraints during deployment, we develop a space proposal method to screen out and optimize the most promising architecture candidates under the constraints. By treating the architecture search through an hourglass lens, the limited computation resource can be well allocated across vital and non-vital blocks. We design toy experiments on residual networks to illustrate the varied influence of vital and non-vital blocks on the network accuracy. The resulting HourNAS costs only about 3 hours~(0.1 GPU days) on the entire ImageNet dataset to achieve a 77.0\% Top-1 accuracy, which outperforms the state-of-the-art methods.
\section{Related Works}\label{sec:relatedworks}
This section reviews the methods for neural architecture search algorithms. Then, we discuss layer equality, \emph{i.e.}, the importance of different blocks in deep neural networks.
\iffalse
{\noindent \bf Efficient Network Design}
In early age of deep learning, most widely used neural architectures are manually designed by human experts, \emph{e.g.}, VGG~\cite{vgg}, ResNet~\cite{resnet}, ~\cite{densenet}, and \emph{etc}. With the increasing demand for real-time applications on mobile or embedded devices, designing efficient neural networks has attracted more research interest. MobileNet serials~\cite{mobilenet, mobilenetv2} utilize the depthwise separable convolution and inverted residual block to construct efficient networks. ShuffleNet serials~\cite{shufflenet,shufflenetv2} employ pointwise group convolution and channel shuffle module to design computation-efficient models for mobile devices. Despite their strong performance, manually designed networks require a huge effort of human experts.
\fi
{\noindent \bf Neural Architecture Search.} To automate the design of neural models, neural architecture search (NAS) was introduced to discover efficient architectures with competitive performance. Reinforcement learning (RL) and evolution algorithm~(EA) were widely adopted in NAS~\cite{nasnet,zoph_rl_iclr,nsganet,amoebanet,mnasnet,sposnas,autogan}. However, these methods were highly computationally demanding. Many works have been devoted to improving the efficiency of NAS from different perspectives, \emph{e.g.}, by adopting the strategy of weight sharing~\cite{enas} or progressive search~\cite{pnas}. Differentiable based NAS attracted great interest as it drastically reduces the searching cost to several days or even hours~\cite{fbnet,darts,setnas,snas,autoreid,autodeeplab}. For example, DARTS~\cite{darts} adopted the continuous architecture representation to allow efficient architecture search via gradient descent. {\color{black} Meanwhile, as discusses in TuNAS~\cite{tunas}, gradient-based NAS methods consistently performed better than random search, showing the power of searching for excellent architectures. However, the most efficient gradient-based NAS methods still took dozens of days for directly searching on target large-scale dataset (\eg, ImageNet).} Thus, an efficient method for directly searching deep neural architectures on large-scale datasets and search spaces is urgently required.
{\noindent \bf Layer Equality.} Most of existing NAS methods treated all layers with the same importance during the search. However, convolution neural networks are always over-parameterized and the impact on the final accuracy of each layer is totally different. Zhang \emph{et al.}~\cite{arealllayerscreatedequal} re-initialized and re-randomized the pre-trained networks, and found that some layers are robust to the disturbance. For some intermediate layers, the re-initialization and re-randomization steps did not have negative consequences on the accuracy. Veit \emph{et al.}~\cite{resnetensembles} decomposed residual networks and found that skipping some layers does not decrease the accuracy significantly. Ding \emph{et al.}~\cite{globalsparsemomentumsgd} pruned different layers separately and found some layers are more important to the final accuracy. It is obvious that the layers are not created equal, and some layers are more important than others.
In this paper, we analyze the causes of the inequality phenomenon in the residual network and exploit this property in neural architecture search to improve its efficiency.
\section{Hourglass Neural Architecture Search}
In this section, we revisit the neural architecture search from an hourglass way. The vital few blocks should be searched with a higher priority to guarantee the potential accuracy of the architecture. The non-vital blocks that occupy the major volume are then searched in an extremely fast way by focusing on the discovered space proposals.
\subsection{Vital Blocks: the Neck of Hourglass}\label{sec_notequal}
In this paper, we focus on the serial-structure NAS SuperNet~\cite{mnasnet, efficientnet, fbnet, proxylessnas}, as it is hardware-friendly and capable of achieving superior performance. Before we illustrate vital blocks in a general NAS, we first take ResNet~\cite{resnet} as an example for the analysis. ResNet is one of the most popular manually designed architectures. It is established by stacking a series of residual blocks, and the residual block is defined as,
\begin{equation}
\mathbf{y} = \mathcal{F}(\mathbf{x}, \mathbf{w}) + \mathbf{x},
\label{eq_resnet}
\end{equation}
where $\mathbf{x}$ is the input feature map, $\mathcal{F}$ denotes the transformation (\emph{e.g.}, convolution and batch normalization for vision tasks) and $\mathbf{w}$ stands for the trainable parameters.\footnote{As for the downsample blocks (reduce the feature map size) and the channel expansion blocks (increase the channel number), we follow~\cite{resnetensembles} and use $\mathbf{y} = \mathcal{F}(\mathbf{x}, \mathbf{w})$ to express.} From the information flow perspective, there are two paths to transmit the information from the node $\mathbf{x}$ to the node $\mathbf{y}$, \emph{i.e.}, the shortcut connection and the transformation $\mathcal{F}$. If there are $m$ residual blocks in a network, there will be $2^m$ different paths for the information propagation in total. A general neural network $\mathcal{N}$ based on the residual blocks~\cite{resnet, mnasnet, fbnet} can therefore be approximated as the ensemble of a number of paths~\cite{resnetensembles} $\{\mathcal{P}_1, \dots, \mathcal{P}_n\}$, \emph{i.e.}, $\mathcal{N}(X) \approx \sum_{i=1}^{n} \mathcal{P}_i (X)$,
where each path $\mathcal{P}_i$ is set up by a series of blocks, $X$ is the input data, and $n$ is the number of all the paths.
It is worth noticing that there are a few blocks existing in all the possible paths, \emph{e.g.}, the gray blocks in Fig.~\ref{fig_HourNAS}. These self-contained blocks do not participate in forming any residual blocks, but they are \emph{vital}, because of their appearance in every path from the input to the output of the network. On the other hand, the green and blue blocks in Fig.~\ref{fig_HourNAS} are a part of the residual blocks $\mathbf{y} = \mathcal{F}(\mathbf{x}, \mathbf{w}) + \mathbf{x}$, where the information can be transmitted through the plain transformation $\mathcal{F}(\mathbf{x}, \mathbf{w})$ or the shortcut connection $\mathbf{x}$ to the next block, {\color{black} so they are not that vital}.
{\color{black}
\noindent \textbf{Identify and Examine Vital Blocks.}} Given the paths $\{\mathcal{P}_1, \dots, \mathcal{P}_n\}$ in a general residual network $\mathcal{N}$, the vital blocks shared by all the paths can be identified through $\hat{\mathcal{P}} = \mathcal{P}_1 \cap \dots \cap \mathcal{P}_n$, where $\mathcal{P}_i \cap \mathcal{P}_j$ denotes the intersection set of those blocks in paths $\mathcal{P}_i$ and $\mathcal{P}_j$. In the popular residual networks, such as ResNet~\cite{resnet} and FBNet~\cite{fbnet}, the vital blocks are exactly the first convolution layer, the last fully connected layer, the downsampling blocks, and the channel expansion blocks. These vital blocks are critical to the accuracy of the whole architecture, as they exist in all paths and act as the neck of the hourglass to control the information flow. In contrast, the other blocks would always find substitutes for themselves to keep the information flow, and they thus play a secondary role in the whole architecture.
We further take mobile architectures as an example to illustrate the different influence of vital and non-vital blocks on the network accuracy. A random mask function $\mathcal{M}(\mathbf{y}, p)$ is introduced to destroy the output of blocks in the pretrained MnasNet~\cite{mnasnet} and MobileNetV2~\cite{mobilenetv2}, where $\mathbf{y}$ is the output feature map, and $0\le p\le 1$ is the probability. In particular, every channel is reset to 0 with a probability $p$.
\begin{figure}
\centering
\subfigure[MnasNet]{\includegraphics[width=0.48\linewidth]{mnasnet_importance.pdf}}
\subfigure[MobileNetV2]{\includegraphics[width=0.48\linewidth]{mobilenet_importance.pdf}}
\caption{The diagram of block importance by using the MnasNet and MobileNetV2 pretrained models.}
\label{fig_block_importance}
\vspace{-15pt}
\end{figure}
Fig.~\ref{fig_block_importance} shows the accuracy change of MnasNet~\cite{mnasnet} and MobileNetV2~\cite{mobilenetv2} resulting from the feature distortion on different blocks. {\color{black} The $blocki\_j$ denotes the $j$-th block in the $i$-th stage. As discussed above, the first block in every stage is the vital block}. We set $p = \{0.3, 0.6, 1.0\}$ to gradually increase the degree of feature distortion. Each time we only manipulate one block while keeping others unchanged in the network. For the non-vital blocks (\emph{e.g.}, {\color{black}block3\_2 and block3\_3 in both MnasNet and MobileNetV2}), even if all the channels are reset to zero (\emph{i.e.}, p=1.0), the network does not undergo a significant accuracy degradation. However, a small portion (p=0.3) of channels that are masked out for those vital blocks ({\color{black}\emph{e.g.}, block1\_1 and block2\_1}) will lead to an obvious accuracy drop.
\iffalse
{\color{black}
\noindent \textbf {Theoretical Analysis of Vital blocks.} In this part, we try to analyze why some layers play a more vital role by analyzing the rank. The rank in the feature represents the richness of information, and also represents the distinctiveness~\cite{yang2017breaking}. We start by considering two transformations,
\begin{equation}
\begin{aligned}
y_1 &= \mathcal{F}(w \otimes x),\\
y_2 &= \mathcal{F}(w \otimes x) + x,
\end{aligned}
\end{equation}
where $\mathcal{F}$ is the non-linear transformation, $\otimes$ is the matrix multiplication and $x$, $y$ are the input and output, respectively. The ranks of the outputs are:
\begin{equation}
\begin{aligned}
rank(y_1) &\le \min\{rank(w), rank(x)\}, \\
rank(y_2) &\le \min\{rank(w), rank(x)\} + rank(x).
\end{aligned}
\end{equation}
Adding the identity mapping makes $y_2$ capable of containing much more information, and the high rank feature maps are critical to the final performance~\cite{hrank}. The non-vital blocks and the vital blocks act like $y_2$ and $y_1$, respectively. For example, removing the learned transformation $\mathcal{F}$ of these transformations does not affect much to the whole network as Fig.~\ref{fig_block_importance} shows because the information could be preserved from the shortcut. For the transformation $y_1 = \mathcal{F}(w \otimes x)$, this is how downsampling blocks transform features, the quality of the transformation $\mathcal{F}$ itself determines the information preservation.
}
\fi
{\color{black} \noindent \textbf{Revisit Neural Architecture Search.}} The goal of NAS is to search for the architecture of a higher accuracy under the constraints of available computational resources. In other words, NAS can be regarded as a problem of resource allocation. Vital blocks are potentially the most important and need to be put as the priority. As a result, more resources are better to be first allocated to the vital blocks, and the remaining resources are used for the non-vital blocks design. This therefore naturally motivates us to develop a two-stage search algorithm. During the first stage, we construct the minimal SuperNet $\mathcal{S}_{vital}$ {\color{black} by stacking all the vital layers and} search the vital blocks. {\color{black} The weights and architecture parameters are optimized alternatively in a differentiable way~\cite{fbnet}.} In the second stage, we fix the derived architecture of those vital blocks, and allocate the computational resources to search for the non-vital blocks.
\subsection{Non-Vital Blocks: the Bulb of Hourglass}
Non-vital blocks are often composed of a large number of parameters. They look like the bulb of the hourglass to determine the whole volume size. If the computational resources are unlimited to deploy the neural network, then we can make the network wider and deeper for achieving a higher performance~\cite{resnet, efficientnet}. However, the searched architectures are to be deployed on mobile devices which have demanding constraints of the resources, \emph{e.g.}, model size, and FLOPs. Without investigating these constraints, it would be ineffective to directly sample the architecture from a large search space. For example, if the sampled architectures cannot fully use the available computation resource, the resulting models might perform poorer than expected, which has been analyzed by a number of multi-objective NAS works~\cite{nsganet, cars, lemonade, mnasnet} (the Pareto front). Otherwise, if the sampled architectures overwhelm the use of computation resources, they would be difficult to be deployed. To tackle the dilemma, we introduce an efficient sampling operation to {\color{black} avoid wasting too much time on search unimportant operations}, and a space proposal strategy to put more attention on {\color{black} architectures that meet the target computational resources}.
\subsection{Space Proposal for Efficient Resource Constrained Search}
\label{sec_space_proposal}
A general differentiable neural architecture search algorithm can be formulated as a bilevel optimization problem~\cite{darts}:
\begin{equation}
\begin{aligned}
\theta^* &= \mathop{\mathrm{\arg\min}}_{\theta} \mathcal{H}_{val}(w^*(\theta), \theta),\\
\textnormal{s.t.} &~~ w^*(\theta) = \mathop{\mathrm{\arg\min}}_{w} \mathcal{H}_{train}(w, \theta),
\end{aligned}
\end{equation}
where $\mathcal{H}$ is the cross-entropy loss function, and $\theta$ denotes the architecture parameters. If the accuracy is the only objective to be considered for searching, a complex architecture would be preferred for achieving a highest accuracy (see Sec.~\ref{sec_ablation}). However, if the obtained architectures are to be deployed on mobile devices, we may always have the computational resource constraints from the environment . Thus, neural architecture search that considers the multiple objectives can be formulated as,
\begin{equation}
\begin{aligned}
\theta^* &= \mathop{\mathrm{\arg\min}}_{\theta} \mathcal{H}_{val}(w^*(\theta), \theta) + \alpha \times \mathcal{T}(\theta),\\
\textnormal{s.t.} &~~ w^*(\theta) = \mathop{\mathrm{\arg\min}}_{w} \mathcal{H}_{train}(w, \theta),
\end{aligned}
\end{equation}
where $\mathcal{T}(\theta)$ is the regularization term that encourages the produced architecture to satisfy the target computational resource constraints. Assuming the constraints~(targets) on computational resources (\eg, model size, FLOPs) are $T_{i \in \{1, \dots, n\}}$, where $n$ is the number of objectives, an efficient and controllable way is to initialize architectures that satisfy $T$. Thus, we introduce the concept of space proposal. The space proposal is a subspace of the large search space, and all the sampled architectures from the space proposal satisfy the target resources. As a result, the search phase would not waste resources on optimizing useless architectures. In addition, the space proposal ensures ``what you set is what you get''. Similar to gradient-based NAS, the space proposal is represented by the architecture parameters.
We take the FLOPs as an example to describe how to optimize a space proposal. Suppose $\theta$ represents the trainable architecture parameters of the NAS SuperNet and the size is $L\times O$, where $L$ is the maximum depth and $O$ is the number of candidate operations {\color{black} in each layer}. A number of methods $\mathcal{G}$ are capable of sampling architectures $A_{\theta}$ from architecture parameters $\theta$,
\begin{equation}
A_{\theta} = \mathcal{G}(\theta),~~\sum_{o} A_{l,o} = 1,
\label{eq_sample}
\end{equation}
where $\mathcal{G}$ is usually specified as softmax~\cite{darts}, Gumbel-softmax~\cite{fbnet}, Gumbel-Max~\cite{gdas,data}, \emph{etc}. The $A_{l,o}$ is the $o$-th operation in the $l$-th layer.
The FLOPs table $F$ of the SuperNet $\mathcal{S}$ is of size $L \times O$, where $F_{l, o}$ denotes the FLOPs of the $o$-th operation in the $l$-th SuperBlock. The FLOPs for sampled architecture $A_\theta$ is calculated as $\mathcal{R}_{F}(A_\theta) = \textnormal{sum} (A_\theta \odot F)$, where $\odot$ is the element-wise product. Assuming the target FLOPs is $T_{F}$, the optimization is formulated as,
\begin{equation}
\theta^F = \mathop{\mathrm{\arg\min}}_{\theta} | \mathcal{R}_{F}(\mathcal{G}(\theta)) - T_{F} | / M_{F},
\label{eq_clock_loss_flops}
\end{equation}
where $M_{F}$ is a constant scalar denotes the maximum FLOPs of the sampled architectures, and this term is used for normalizing the objective to $[0, 1]$. We extend Eqn~\ref{eq_clock_loss_flops} to $n$ different objectives. The targets for $n$ objectives are $T_{i \in \{1, \dots, n\}}$, and the optimization is defined as,
\begin{equation}
\theta^T = \mathop{\mathrm{\arg\min}}_{\theta} \frac{1}{n}\sum_{i=1}^n | \mathcal{R}_i (\mathcal{G}(\theta)) - T_i | / M_i,
\label{eq_clock_loss}
\end{equation}
which $\mathcal{R}_i (\mathcal{G}(\theta))$ is the resource demand of the architecture sampled by $\mathcal{G}(\theta)$ on the $i$-th objective.
This optimization problem is easily to be solved in a few seconds. The solution $\theta^T$ can be regarded as a space proposal under the constraints $T$, and the structure $A_{\theta^T}$ sampled from $\theta^{T}$ by Eqn~\ref{eq_sample} would be more easily to satisfiy target resources $T$. Instead of relying on a single optimal solution $\theta^T$, we turn to an ensemble way to start from different random initializations and derive a series of space proposals $\Theta^{T} = \{ \theta^{T}_1, \cdots, \theta^{T}_m \}$, where $m$ is the number of space proposals. The orthogonal constraint is also introduced to further increase the diversity of different space proposals, which is formulated as,
\
\begin{equation}
\begin{aligned}
\theta_1, \dots, \theta_m = \mathop{\arg\min}_{\theta_1, \dots, \theta_m} (\frac{1}{nm}&\sum_{j=1}^m \sum_{i=1}^n |\mathcal{R}_i(\mathcal{G}(\theta_j)) - T_i| / M_i + \\ &\beta \times \sum (|O - I|)),
\end{aligned}
\label{eq_regularization_loss}
\end{equation}
where $O = \Theta\Theta^T$ is an $m\times m$ matrix and element $O_{i, j}$ denotes the inner product of $\theta_i$ and $\theta_j$. The term $\sum(|O-I|)$ regularizes $m$ space proposals to be orthogonal, which indicates that the architectures sampled from different space proposals are different. A uniformly initialized auxiliary parameter $\Pi$ of size $m$ is then introduced to sample space proposals,
\
\begin{equation}
\pi = \mathcal{P}(\Pi),~~\sum_i \pi_i = 1,
\label{eq_pi}
\end{equation}
where $\mathcal{P}$ could be softmax, Gumbel-softmax or Gumbel-Max, $\pi$ is the sampled vector from $\Pi$ that used for combine the architectures sampled from $m$ space proposals, and the ensembled architecture $A_{\Theta}$ that used for updating the SuperNet is defined as,
\
\begin{equation}
A_{\Theta} = \sum_{i} \pi_i \cdot A_{\theta_i} = \sum_{i} \pi_{i} \cdot \mathcal{G}(\theta_i),
\end{equation}
where $A_{\Theta}$ is utilized for updating network parameter $w$ on train set $\mathcal{D}_{train}$ and architecture parameter $\Pi, \Theta$ on validation set $\mathcal{D}_{val}$, respectively. Every space proposal $\theta_i$ optimizes towards the good architectures in the space proposal and $\Pi$ optimizes towards better proposals. The NAS framework by using the space proposal strategy is summarized as,
\
\begin{equation}
\begin{aligned}
\Pi^*, \Theta^* &= \mathop{\mathrm{\arg\min}}_{\Pi, \Theta} \mathcal{H}_{val}(w^*(\Pi, \Theta), \Pi, \Theta) + \alpha \times \mathcal{T}(\Pi, \Theta), \\
\textnormal{s.t.} &~~ w^*(\Pi, \Theta) = \mathop{\mathrm{\arg\min}}_{w} \mathcal{H}_{train}(w, \Pi, \Theta),
\end{aligned}
\end{equation}
where $\mathcal{T}$ (Eqn~\ref{eq_clock_loss}) is the regularization on space proposal parameters (Eqn~\ref{eq_regularization_loss}), and $\alpha$ is the slope of the multi-objective loss term.
\subsection{Overall Search Algorithm}
{\color{black} Based on the proposed method, we summarize the overall search algorithm in Alg.~\ref{alg_fnas}.}
\begin{algorithm}[h]
\caption{The searching algorithm of HourNAS.}
\begin{algorithmic}[1]
\REQUIRE The NAS supernet $\mathcal{S}$, the computational targets $T_{i \in \{1, \dots, n\}}$, the train set $D_{train}$ and validation set $D_{val}$, the searching epochs for vital blocks $E_{vital}$ and non-vital blocks $E_{n-vital}$, the number of space proposals $m$, iterations $I_{sp}$ for training space proposals.
\STATE \textbf{// Search Vital Blocks}
\STATE Constructing the minimal SuperNet $\mathcal{S}_{vital}$ {\color{black} by stacking all the vital layers} and the architecture parameter $\theta_{vital}$.
\FOR{e = 1 to $E_{vital}$}
\FOR{data and target pair $(X_{tr}, Y_{tr})$ in $D_{train}$}
\STATE Sample network $A$ from $\theta_{vital}$, calculate loss and update network parameters.
\ENDFOR
\FOR{data and target pair $(X_{val}, Y_{val})$ in $D_{val}$}
\STATE Sample network $A$ from $\theta_{vital}$, calculate loss and update $\theta_{vital}$.
\ENDFOR
\ENDFOR
\STATE The operations with the highest importance are selected to form the vital layers.
\STATE \textbf{// Optimize $m$ space proposals}
\STATE According to the computational targets $T$, HourNAS optimizes $m$ proposals $\Theta^{T} = \{ \theta^{T}_1, \cdots, \theta^{T}_m \}$ for $I_{sp}$ iterations (Eqn~\ref{eq_regularization_loss}), and construct the proposal sampler $\pi$~(Eqn~\ref{eq_pi}).
\STATE \textbf{// Search Non-Vital Blocks}
\FOR{e = 1 to $E_{n-vital}$}
\FOR{data and target pair $(X_{tr}, Y_{tr})$ in $D_{train}$}
\STATE Sample network $A$ from $\pi$ and $\Theta$, calculate loss and update network parameters.
\ENDFOR
\FOR{data and target pair $(X_{val}, Y_{val})$ in $D_{val}$}
\STATE Sample network $A$ from $\pi$ and $\Theta$, calculate loss and update $\pi$ and $\Theta$.
\ENDFOR
\ENDFOR
\STATE Fix operations by selecting the space proposal and operations with the highest probability.
\ENSURE The architecture $A$ which satisfies the computational targets $T_{i \in \{1, \dots, n\}}$.
\end{algorithmic}
\label{alg_fnas}
\end{algorithm}
\section{Experiments}\label{sec:exp}
In this section, we first describe the experimental settings and then extensively evaluate the proposed HourNAS on several popular NAS search spaces~\cite{fbnet, mnasnet, efficientnet} on ImageNet.
\subsection{Experimental Settings}
We use the HourNAS to search on the complete ImageNet train set. The subset $\mathcal{D}_{tr}$ takes $80\%$ of the train set to train network parameters and the rest $\mathcal{D}_{val}$ is used to update architecture parameters. We search on three popular search spaces, \ie, FBNet~\cite{fbnet}, MnasNet~\cite{mnasnet}, and EfficientNet~\cite{efficientnet}. For any of our searched architecture, the training strategy is the same as the corresponding baseline. {\color{black} We use the NVIDIA V100 GPU to measure the search time and compare with previous works fairly. The V100 GPU is also adopted by a number of literatures, for example, PDARTS~\cite{pdarts}, FBNet~\cite{fbnet}.}
The HourNAS first searches the vital blocks for one epoch~(about 1 hour). Then, HourNAS optimizes multiple space proposals according to the computational targets and searches the non-vital blocks for one epoch~(about 2 hours). The whole searching process requires only one V100 GPU within 3 hours. Extending the search time will not further improve the accuracy, because the distribution of architecture parameters is stable. For competing methods like MnasNet~\cite{mnasnet}, 3 GPU hours could only train one sampled architecture and the RL controller has no difference with random search. We utilize the Gumbel-Max~\cite{astarsampling, data, gdas} to sample operations according to the learned importance, which avoids wasting searching time on undesired operations. Gumbel-Max samples an operation according to the learned probability distribution (\emph{i.e.}, importance). The sampling frequencies of those poor operations tend to be relatively low, so that we can effectively reduce the time spent on them. The Gumbel-Max sampling accelerates every iteration by around $O$ times, where $O$ is the number of candidate operations in every layer. During searching, we follow FBNet~\cite{fbnet} and add the temperature $\tau$ (Eqn~\ref{eq_sample}) for sharpening the distribution progressively. The temperature $\tau$ starts from 5.0 and multiply 0.9999 at every iteration. Slope parameter $\alpha$, $\beta$ are emperically set to 5.0 and 1e-2, respectively. Learning rates for optimizing weights and architecture parameters are 0.1 and 0.01, respectively. Adam~\cite{adam} optimizer is used to update architecture parameters.
\begin{table*}
\caption{Overall comparison on the ILSVRC2012 dataset.}
\small
\centering
\begin{tabular}{c|c|c|c|c|c|c|c}
\toprule
\multirow{2}{*}{{Model}} & \multirow{2}{*}{{Type}} & Search & Search Cost & Params & FLOPS & Top-1 & Top-5 \\
& & Dataset & (GPU days) & (M) & (M) & (\%) & (\%) \\
\midrule
ResNet50~\cite{resnet} & manual & - & - & 25.6 & 4100 & 75.3 & 92.2 \\
MobileNetV1~\cite{mobilenet} & manual & - & - & 4.2 & 575 & 70.6 & 89.5 \\
MobileNetV2~\cite{mobilenetv2} & manual & - & - & 3.4 & 300 & 72.0 & 91.0 \\
MobileNetV2~(1.4$\times$) & manual & - & - & 6.9 & 585 & 74.7 & 92.5 \\
ShuffleNetV2~\cite{shufflenetv2} & manual & - & - & - & 299 & 72.6 & - \\
ShuffleNetV2~(1.5$\times$) & manual & - & - & 3.5 & 299 & 72.6 & - \\
\midrule
FPNASNet~\cite{fpnas} & auto & CIFAR-10 & 0.8 & 3.4 & 300 & 73.3 & - \\
SNAS~(mild)~\cite{snas} & auto & CIFAR-10 & 1.5 & 4.3 & 522 & 72.7 & 90.8\\
AmoebaNet-A~\cite{amoebanet} & auto & CIFAR-10 & 3150 & 5.1 & 555 & 74.5 & 92.0 \\
PDARTS~\cite{pdarts} & auto & CIFAR-10 & 0.3 & 4.9 & 557 & 75.6 & 92.6 \\
NASNet-A~\cite{nasnet} & auto & CIFAR-10 & 1800 & 5.3 & 564 & 74.0 & 91.3 \\
GDAS~\cite{gdas} & auto & CIFAR-10 & 0.2 & 5.3 & 581 & 74.0 & 91.5 \\
PNAS~\cite{pnas} & auto & CIFAR-10 & 225 & 5.1 & 588 & 74.2 & 91.9 \\
CARS-I~\cite{cars} & auto & CIFAR-10 & 0.4 & 5.1 & 591 & 75.2 & 92.5 \\
DARTS~\cite{darts} & auto & CIFAR-10 & 4 & 4.9 & 595 & 73.1 & 91.0 \\
MdeNAS~\cite{mdenas} & auto & CIFAR-10 & 0.2 & 6.1 & - & 74.5 & 92.1 \\
RCNet~\cite{resourceconstrainednas} & auto & ImageNet & 8 & 3.4 & 294 & 72.2 & 91.0 \\
SPOSNAS~\cite{sposnas} & auto & ImageNet & 13 & 5.3 & 465 & 74.8 & - \\
ProxylessNAS~\cite{proxylessnas} & auto & ImageNet & 8.3 & 7.1 & 465 & 75.1 & 92.5 \\
\midrule
FBNet-B~\cite{fbnet} & auto & ImageNet & 9 & 4.8 & 295 & 74.1 & - \\
FBNet-C~\cite{fbnet} & auto & ImageNet & 9 & 5.5 & 375 & 74.9 & - \\
\textbf{HourNAS-FBNetSS-A} & auto & ImageNet & \textbf{0.1} & 4.8 & 298 & \textbf{74.1} & \textbf{91.8} \\
\textbf{HourNAS-FBNetSS-B} & auto & ImageNet & \textbf{0.1} & 5.5 & 406 & \textbf{75.0} & \textbf{92.2} \\
\midrule
\textbf{HourNAS-EFBNetSS-C} & auto & ImageNet & \textbf{0.1} & 4.8 & 296 & \textbf{74.1} & \textbf{91.6} \\
\textbf{HourNAS-EFBNetSS-D} & auto & ImageNet & \textbf{0.1} & 5.5 & 394 & \textbf{75.3} & \textbf{92.3} \\
\midrule
MnasNet-A1~\cite{mnasnet} & auto & ImageNet & 3800 & 3.9 & 312 & 75.2 & 92.5 \\
\textbf{HourNAS-MnasNetSS-E} & auto & ImageNet & \textbf{0.1} & 3.8 & 313 & \textbf{75.7} & \textbf{92.8} \\
\midrule
EfficientNet-B0~\cite{efficientnet} & auto & ImageNet & - & 5.3 & 390 & 76.8 & - \\
\textbf{HourNAS-EfficientNetSS-F} & auto & ImageNet & \textbf{0.1} & 5.3 & 383 & \textbf{77.0} & \textbf{93.5} \\
\bottomrule
\end{tabular}
\label{tab_imagenet}
\end{table*}
\subsection{Comparison with State-of-the-arts}
\textbf{FBNet Search Space (FBNetSS).} We first evaluate our HourNAS on the popular FBNet search space (FBNetSS), {\color{black} which contains $9^{22} \approx 1\times 10^{21}$ architectures}. HourNAS first searches the vital blocks and the results show that operations with the expansion ratio of six have significantly higher probabilities than other operations. We choose the operations with the highest probabilities to form the vital blocks. This result is in line with our intuition that the complex operations have the greatest feature extraction ability on the large dataset, \emph{i.e.}, ImageNet.
After fixing the operations of the vital blocks, we are interested in finding the appropriate proposal number $m$. We set the computational resources the same as FBNet-B and search architectures using $m$ different space proposals. We first visualize the distribution of sampled architectures from the space proposal by gradually decreasing the temperature $\tau$. Each space proposal is optimized for 1000 iterations in total (Eqn~\ref{eq_clock_loss}). The computational targets are set to 4.8M parameters (x-axis) and 300M FLOPs (y-axis), which are consistent with FBNet-B~\cite{fbnet}. As shown in Fig.~\ref{fig_vis_space_proposal}, with the decrease of temperature of $\tau$, the sampled architectures satisfy the computational targets more precisely.
\iffalse
\begin{figure}[H]
\centering
\subfigure[$\tau=5.0$]{\includegraphics[width=0.2\linewidth]{figs/proposals/5.png}}
\hspace{10pt}
\subfigure[$\tau=1.0$]{\includegraphics[width=0.2\linewidth]{figs/proposals/1.png}}
\hspace{10pt}
\subfigure[$\tau=0.5$]{\includegraphics[width=0.2\linewidth]{figs/proposals/05.png}}
\hspace{10pt}
\subfigure[$\tau=0.1$]{\includegraphics[width=0.2\linewidth]{figs/proposals/01.png}}
\caption{The distribution of 10,000 architectures sampled from optimized space proposal under different temperatures $\tau$.}
\label{fig_vis_space_proposal}
\end{figure}
\fi
\begin{figure}
\centering
\subfigure[$\tau=5.0$]{\includegraphics[width=0.24\linewidth]{5.png}}
\subfigure[$\tau=1.0$]{\includegraphics[width=0.24\linewidth]{1.png}}
\subfigure[$\tau=0.5$]{\includegraphics[width=0.24\linewidth]{05.png}}
\subfigure[$\tau=0.1$]{\includegraphics[width=0.24\linewidth]{01.png}}
\caption{The distribution of 10,000 architectures sampled from optimized space proposal under different temperatures $\tau$.}
\label{fig_vis_space_proposal}
\vspace{-10pt}
\end{figure}
The optimization step for constructing several space proposals takes only a few seconds, which is an efficient solution for controllable multi-objective neural architecture search. We enumerate $m= \{1, 2, 4, 8, 16\}$ and the operations with the highest probability according to $\Pi$ and $\Theta$ are selected after searching. The architectures are evaluated on the CIFAR-10 dataset to determine the appropriate space proposal number $m$ for finding superior architectures. {\color{black} In retraining, we integrate CutOut~\cite{cutout} to make networks generalize better.} As shown in Tab.~\ref{tab_appropriate_m}, $m=8$ could result in a well-performed architecture and we use $m=8$ in the following experiments to achieve a better trade-off between searching costs and performance.
\begin{table}[H]
\caption{Comparison of image classifiers on CIFAR-10 dataset.}
\small
\centering
\begin{tabular}{l|c|c|c}
\toprule
\multirow{2}{*}{{Model}} & Test Error & {Params} & Search Cost \\
& (\%) & (M) & (GPU days) \\
\midrule
HourNAS~(m=1) & 3.86 & 2.8 & 0.1 \\
HourNAS~(m=2) & 3.54 & 2.8 & 0.1 \\
HourNAS~(m=4) & 3.41 & 2.7 & 0.1 \\
HourNAS~(m=8) & 3.39 & 2.8 & 0.1 \\
HourNAS~(m=16) & 3.37 & 2.8 & 0.1 \\
\bottomrule
\end{tabular}
\label{tab_appropriate_m}
\vspace{-15pt}
\end{table}
We search for models on the ImageNet dataset, \ie, HourNAS-FBNetSS-A and HourNAS-FBNetSS-B. To fairly compare with FBNet, HourNAS-FBNetSS-A has the same model size and FLOPs as FBNet-B, and HourNAS-FBNetSS-B has the same computational requirements as FBNet-C. {\color{black} We train the networks for 350 epochs in total with a batch size of 512. Learning rate starts from 0.25 and the weight decay is 1e-5. Label smoothing and learning rate warmup strategies are also used. The activation after convolution is the ReLU function.} The training strategy is the same as FBNet~\cite{fbnet} without using bells and whistles. As shown in Tab.~\ref{tab_imagenet}, the HourNAS-FBNetSS-A and HourNAS-FBNetSS-B achieve competitive accuracies with FBNet-B and FBNet-C, and the search time is drastically reduced by two orders of magnitude. We search HourNAS-FBNetSS-A for three times using different random seeds, and the standard deviation of Top-1 accuracy is 0.1\%.
\textbf{Enlarged FBNet Search Space (EFBNetSS).} MixConv~\cite{mixconv} indicates that a larger kernel size leads to better performance. To understand the impact of search space and to further verify the effectiveness of HourNAS, we slightly enlarge the search space of FBNet. We add the blocks with kernel size $k=7$ and remove the blocks with group $g=2$. {\color{black} This modification results in a search space containing $1\times 10^{22}$ architectures, which is 10 times larger than the original one.} The multi-objectives are the same as HourNAS-FBNetSS-A and HourNAS-FBNetSS-B. We list the searched architectures in Tab.~\ref{tab_imagenet}. The HourNAS-EFBNetSS-C achieves the same Top-1 accuracy with HourNAS-FBNetSS-A and HourNAS-EFBNetSS-D surpasses HourNAS-FBNetSS-B by 0.3\% Top-1 accuracy. The larger kernel size $k=7$ ensures that the architectures are capable of perceiving the characteristics of a larger area.
\begin{table*}
\caption{The results of FBNet-Max and EfficientNet-Max on ILSVRC2012 dataset.}
\centering
\small
\begin{tabular}{c|c|c|c|c}
\toprule
{Model} & Params (M) & FLOPS (M) & Top-1 (\%) & Top-5 (\%) \\
\midrule
FBNet-Max & 5.7 & 583 & 75.7 & 92.8 \\
\midrule
EfficientNet-Max & 5.8 & 738 & {78.3} & {94.0} \\
\bottomrule
\end{tabular}
\label{tab_full}
\vspace{10pt}
\caption{Comparisons of searching with and without vital block priori on ILSVRC2012 dataset. The search spaces are original (upper) and enlarged (lower) FBNet search space, respectively.}
\centering
\small
\begin{tabular}{c|c|c|c|c|c|c|c}
\toprule
\multirow{2}{*}{{Model}} & \multirow{2}{*}{{Type}} & Search & Search Cost & Params & FLOPS & Top-1 & Top-5 \\
& & Dataset & (GPU days) & (M) & (M) & (\%) & (\%) \\
\midrule
\textbf{HourNAS-FBNetSS-A} & auto & ImageNet & \textbf{0.1} & 4.8 & 298 & \textbf{74.1} & \textbf{91.8} \\
{HourNAS-FBNetSS-G}~(w/o vital block priori) & auto & ImageNet & {0.2} & 4.7 & 297 & 73.2 & 91.4 \\
\midrule
\textbf{HourNAS-EFBNetSS-D} & auto & ImageNet & \textbf{0.1} & 4.8 & 296 & \textbf{74.1} & \textbf{91.6} \\
{HourNAS-EFBNetSS-H}~(w/o vital block priori) & auto & ImageNet & {0.2} & 4.8 & 299 & {73.5} & {91.3} \\
\bottomrule
\end{tabular}
\label{tab_vital_priori}
\vspace{10pt}
\caption{The results comparison on ILSVRC2012 dataset. }
\centering
\small
\begin{tabular}{c|c|c|c|c|c|c|c}
\toprule
\multirow{2}{*}{{Model}} & \multirow{2}{*}{{Type}} & Search & Search Cost & Params & FLOPS & Top-1 & Top-5 \\
& & Dataset & (GPU days) & (M) & (M) & (\%) & (\%) \\
\midrule
{HourNAS-FBNetSS-A} & auto & ImageNet & {0.1} & 4.8 & 298 & {74.1} & {91.8} \\
{HourNAS-FBNetSS-I} & auto & ImageNet & 1.0 & 4.8 & 318 & 74.2 & 91.8 \\
\bottomrule
\end{tabular}
\label{tab_gumbel}
\vspace{-15pt}
\end{table*}
\textbf{MnasNet Search Space (MnasNetSS).} We further apply our proposed HourNAS to the search space of MnasNet~\cite{mnasnet}. {\color{black} The search space contains $2.5 \times 10^{23}$ architectures in total and is larger than FBNet search space.} We select MnasNet-A1 as the baseline and use its number of the parameters and FLOPs as two objectives to optimize 8 space proposals. The discovered HourNAS-MnasNetSS-E achieves a Top-1 accuracy of 75.7\% on the ILSVRC2012 dataset, which surpasses MnasNet-A1 by 0.5\%.
\textbf{EfficientNet Search Space (EfficientNetSS).} To compare with the state-of-the-art architecture EfficientNet-B0~\cite{efficientnet}, we also use HourNAS to search on the same search space as EfficientNet, which {\color{black} contains $4 \times 10^{18}$ architectures.} Targeting at EfficientNet-B0, we use its model size and FLOPs as two objectives to regularize space proposals and name the searched architecture as HourNAS-EfficientNetSS-F. {\color{black} Same as EfficientNet\footnote{https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet}, we use the Swish~\cite{searchactivation} activation and Exponential Moving Average~(EMA) in fully training.} Note that the AutoAugment~\cite{autoaugment} is not used. The result in Tab.~\ref{tab_imagenet} shows HourNAS-EfficientNetSS-F surpasses EfficientNet-B0 by 0.2\% Top-1 accuracy with similar number of parameters and FLOPs.
\subsection{Ablation Study}\label{sec_ablation}
If we do not restrict the computational resources of the sampled architectures in searching, the most complex block achieves the highest probability after searching for enough time. As shown in Tab.~\ref{tab_full}, we train the most complex architectures in both FBNet and EfficientNet search spaces, namely FBNet-Max and EfficientNet-Max. These two models obtain 75.7\%, and 78.3\% Top-1 accuracies, respectively. However, the computational resource requirements of these structures are relatively high. Therefore, the neural architecture search~(NAS) could be regarded as the problem of computational resource allocation given the resource constraints.
\textbf{The Impact of Vital Block Priori.} In order to investigate the impact of the vital block priori, we directly search architectures without using the vital block information. All the blocks in the SuperNet $S$ are treated equally in searching. We use the Gumbel-Max sampling and space proposal strategy to search architectures under the same predefined computational resources.
We use the previously described original and enlarged FBNet search spaces. We optimize 8 space proposals and it takes 6 hours for searching, which is twice of the counterpart that utilize the vital block priori. As shown in Tab.~\ref{tab_vital_priori}, the Top-1 accuracies of the discovered models (HourNAS-FBNetSS-G, HourNAS-EFBNetSS-H) drop by 0.9\% and 0.6\% on the ImageNet validation set, respectively. The searched vital blocks of HourNAS-FBNetSS-A uses 0.9M parameters and 130M FLOPs, and the HourNAS-FBNetSS-G uses 0.5M parameters and 55M FLOPs. The vital blocks in HourNAS-FBNetSS-G are not as expressive as HourNAS-FBNetSS-A, which results in worse performance. The results show the necessity of the vital block priori. Searching the vital blocks with higher priority is helpful in finding high-quality architectures. Therefore, we use a two-stage search method to allocate resources to vital blocks first, which can allocate resources more effectively, so as to complete the architecture search in a short time. The architectures are provided in the supplementary file, under same computational resources constraints, inclining more resources on the vital blocks gains more performance profit.
\textbf{The Impact of Gumbel-Max Sampling.} As discussed in Sec.~\ref{sec_space_proposal}, there are several design choices for the sampling methods. To find out the impact of the Gumbel-Max sampling method, here we instead use the Gumbel softmax~\cite{fbnet} to optimize architecture parameters and network parameters. The search space and target resource constraints are the same as HourNAS-FBNetSS-A. The search process takes around 1 GPU day and the finalized architecture is denoted as HourNAS-FBNetSS-I. As shown in Tab.~\ref{tab_gumbel}, HourNAS-FBNetSS-I outperforms HourNAS-FBNetSS-A by 0.1\% Top-1 accuracy with much less searching cost, which demonstrate that Gumbel-Max is an efficient strategy for optimizing the SuperNet with almost no less of accuracy.
\section{Conclusions}\label{sec:con}
This paper investigates an efficient algorithm to search deep neural architectures on the large-scale dataset (\emph{i.e.}, ImageNet) directly. To reduce the complexity of the huge search space, we present an Hourglass-based search framework, namely HourNAS. The entire search space is divided into ``vital'' and ``non-vital'' parts accordingly. By gradually search the components in each part, the search cost can be reduced significantly. Since the ``vital'' parts are more important for the performance of the obtained neural network, the optimization on this part can ensure accuracy. By exploiting the proposed approach, we can directly search architectures on the ImageNet dataset that achieves a 77.0\% Top-1 accuracy using only 3 hours (\emph{i.e.}, about 0.1 GPU days), which outperforms the state-of-the-art methods in both terms of search speed and accuracy.
\clearpage
\newpage
{
\bibliographystyle{ieee_fullname}
| 303302cbb57810530fc0b2c7da30f615a2188f80 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{The ALICE Collaboration}
\begingroup
\small
\begin{flushleft}
S.~Acharya\Irefn{org141}\And
D.~Adamov\'{a}\Irefn{org95}\And
A.~Adler\Irefn{org74}\And
J.~Adolfsson\Irefn{org81}\And
M.M.~Aggarwal\Irefn{org100}\And
G.~Aglieri Rinella\Irefn{org34}\And
M.~Agnello\Irefn{org30}\And
N.~Agrawal\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
Z.~Ahammed\Irefn{org141}\And
S.~Ahmad\Irefn{org16}\And
S.U.~Ahn\Irefn{org76}\And
Z.~Akbar\Irefn{org51}\And
A.~Akindinov\Irefn{org92}\And
M.~Al-Turany\Irefn{org107}\And
S.N.~Alam\Irefn{org40}\textsuperscript{,}\Irefn{org141}\And
D.S.D.~Albuquerque\Irefn{org122}\And
D.~Aleksandrov\Irefn{org88}\And
B.~Alessandro\Irefn{org59}\And
H.M.~Alfanda\Irefn{org6}\And
R.~Alfaro Molina\Irefn{org71}\And
B.~Ali\Irefn{org16}\And
Y.~Ali\Irefn{org14}\And
A.~Alici\Irefn{org10}\textsuperscript{,}\Irefn{org26}\textsuperscript{,}\Irefn{org54}\And
N.~Alizadehvandchali\Irefn{org125}\And
A.~Alkin\Irefn{org2}\textsuperscript{,}\Irefn{org34}\And
J.~Alme\Irefn{org21}\And
T.~Alt\Irefn{org68}\And
L.~Altenkamper\Irefn{org21}\And
I.~Altsybeev\Irefn{org113}\And
M.N.~Anaam\Irefn{org6}\And
C.~Andrei\Irefn{org48}\And
D.~Andreou\Irefn{org34}\And
A.~Andronic\Irefn{org144}\And
M.~Angeletti\Irefn{org34}\And
V.~Anguelov\Irefn{org104}\And
C.~Anson\Irefn{org15}\And
T.~Anti\v{c}i\'{c}\Irefn{org108}\And
F.~Antinori\Irefn{org57}\And
P.~Antonioli\Irefn{org54}\And
N.~Apadula\Irefn{org80}\And
L.~Aphecetche\Irefn{org115}\And
H.~Appelsh\"{a}user\Irefn{org68}\And
S.~Arcelli\Irefn{org26}\And
R.~Arnaldi\Irefn{org59}\And
M.~Arratia\Irefn{org80}\And
I.C.~Arsene\Irefn{org20}\And
M.~Arslandok\Irefn{org104}\And
A.~Augustinus\Irefn{org34}\And
R.~Averbeck\Irefn{org107}\And
S.~Aziz\Irefn{org78}\And
M.D.~Azmi\Irefn{org16}\And
A.~Badal\`{a}\Irefn{org56}\And
Y.W.~Baek\Irefn{org41}\And
S.~Bagnasco\Irefn{org59}\And
X.~Bai\Irefn{org107}\And
R.~Bailhache\Irefn{org68}\And
R.~Bala\Irefn{org101}\And
A.~Balbino\Irefn{org30}\And
A.~Baldisseri\Irefn{org137}\And
M.~Ball\Irefn{org43}\And
S.~Balouza\Irefn{org105}\And
D.~Banerjee\Irefn{org3}\And
R.~Barbera\Irefn{org27}\And
L.~Barioglio\Irefn{org25}\And
G.G.~Barnaf\"{o}ldi\Irefn{org145}\And
L.S.~Barnby\Irefn{org94}\And
V.~Barret\Irefn{org134}\And
P.~Bartalini\Irefn{org6}\And
C.~Bartels\Irefn{org127}\And
K.~Barth\Irefn{org34}\And
E.~Bartsch\Irefn{org68}\And
F.~Baruffaldi\Irefn{org28}\And
N.~Bastid\Irefn{org134}\And
S.~Basu\Irefn{org143}\And
G.~Batigne\Irefn{org115}\And
B.~Batyunya\Irefn{org75}\And
D.~Bauri\Irefn{org49}\And
J.L.~Bazo~Alba\Irefn{org112}\And
I.G.~Bearden\Irefn{org89}\And
C.~Beattie\Irefn{org146}\And
C.~Bedda\Irefn{org63}\And
N.K.~Behera\Irefn{org61}\And
I.~Belikov\Irefn{org136}\And
A.D.C.~Bell Hechavarria\Irefn{org144}\And
F.~Bellini\Irefn{org34}\And
R.~Bellwied\Irefn{org125}\And
V.~Belyaev\Irefn{org93}\And
G.~Bencedi\Irefn{org145}\And
S.~Beole\Irefn{org25}\And
A.~Bercuci\Irefn{org48}\And
Y.~Berdnikov\Irefn{org98}\And
D.~Berenyi\Irefn{org145}\And
R.A.~Bertens\Irefn{org130}\And
D.~Berzano\Irefn{org59}\And
M.G.~Besoiu\Irefn{org67}\And
L.~Betev\Irefn{org34}\And
A.~Bhasin\Irefn{org101}\And
I.R.~Bhat\Irefn{org101}\And
M.A.~Bhat\Irefn{org3}\And
H.~Bhatt\Irefn{org49}\And
B.~Bhattacharjee\Irefn{org42}\And
A.~Bianchi\Irefn{org25}\And
L.~Bianchi\Irefn{org25}\And
N.~Bianchi\Irefn{org52}\And
J.~Biel\v{c}\'{\i}k\Irefn{org37}\And
J.~Biel\v{c}\'{\i}kov\'{a}\Irefn{org95}\And
A.~Bilandzic\Irefn{org105}\And
G.~Biro\Irefn{org145}\And
R.~Biswas\Irefn{org3}\And
S.~Biswas\Irefn{org3}\And
J.T.~Blair\Irefn{org119}\And
D.~Blau\Irefn{org88}\And
C.~Blume\Irefn{org68}\And
G.~Boca\Irefn{org139}\And
F.~Bock\Irefn{org96}\And
A.~Bogdanov\Irefn{org93}\And
S.~Boi\Irefn{org23}\And
J.~Bok\Irefn{org61}\And
L.~Boldizs\'{a}r\Irefn{org145}\And
A.~Bolozdynya\Irefn{org93}\And
M.~Bombara\Irefn{org38}\And
G.~Bonomi\Irefn{org140}\And
H.~Borel\Irefn{org137}\And
A.~Borissov\Irefn{org93}\And
H.~Bossi\Irefn{org146}\And
E.~Botta\Irefn{org25}\And
L.~Bratrud\Irefn{org68}\And
P.~Braun-Munzinger\Irefn{org107}\And
M.~Bregant\Irefn{org121}\And
M.~Broz\Irefn{org37}\And
E.~Bruna\Irefn{org59}\And
G.E.~Bruno\Irefn{org33}\textsuperscript{,}\Irefn{org106}\And
M.D.~Buckland\Irefn{org127}\And
D.~Budnikov\Irefn{org109}\And
H.~Buesching\Irefn{org68}\And
S.~Bufalino\Irefn{org30}\And
O.~Bugnon\Irefn{org115}\And
P.~Buhler\Irefn{org114}\And
P.~Buncic\Irefn{org34}\And
Z.~Buthelezi\Irefn{org72}\textsuperscript{,}\Irefn{org131}\And
J.B.~Butt\Irefn{org14}\And
S.A.~Bysiak\Irefn{org118}\And
D.~Caffarri\Irefn{org90}\And
A.~Caliva\Irefn{org107}\And
E.~Calvo Villar\Irefn{org112}\And
J.M.M.~Camacho\Irefn{org120}\And
R.S.~Camacho\Irefn{org45}\And
P.~Camerini\Irefn{org24}\And
F.D.M.~Canedo\Irefn{org121}\And
A.A.~Capon\Irefn{org114}\And
F.~Carnesecchi\Irefn{org26}\And
R.~Caron\Irefn{org137}\And
J.~Castillo Castellanos\Irefn{org137}\And
A.J.~Castro\Irefn{org130}\And
E.A.R.~Casula\Irefn{org55}\And
F.~Catalano\Irefn{org30}\And
C.~Ceballos Sanchez\Irefn{org75}\And
P.~Chakraborty\Irefn{org49}\And
S.~Chandra\Irefn{org141}\And
W.~Chang\Irefn{org6}\And
S.~Chapeland\Irefn{org34}\And
M.~Chartier\Irefn{org127}\And
S.~Chattopadhyay\Irefn{org141}\And
S.~Chattopadhyay\Irefn{org110}\And
A.~Chauvin\Irefn{org23}\And
C.~Cheshkov\Irefn{org135}\And
B.~Cheynis\Irefn{org135}\And
V.~Chibante Barroso\Irefn{org34}\And
D.D.~Chinellato\Irefn{org122}\And
S.~Cho\Irefn{org61}\And
P.~Chochula\Irefn{org34}\And
T.~Chowdhury\Irefn{org134}\And
P.~Christakoglou\Irefn{org90}\And
C.H.~Christensen\Irefn{org89}\And
P.~Christiansen\Irefn{org81}\And
T.~Chujo\Irefn{org133}\And
C.~Cicalo\Irefn{org55}\And
L.~Cifarelli\Irefn{org10}\textsuperscript{,}\Irefn{org26}\And
L.D.~Cilladi\Irefn{org25}\And
F.~Cindolo\Irefn{org54}\And
M.R.~Ciupek\Irefn{org107}\And
G.~Clai\Irefn{org54}\Aref{orgI}\And
J.~Cleymans\Irefn{org124}\And
F.~Colamaria\Irefn{org53}\And
D.~Colella\Irefn{org53}\And
A.~Collu\Irefn{org80}\And
M.~Colocci\Irefn{org26}\And
M.~Concas\Irefn{org59}\Aref{orgII}\And
G.~Conesa Balbastre\Irefn{org79}\And
Z.~Conesa del Valle\Irefn{org78}\And
G.~Contin\Irefn{org24}\textsuperscript{,}\Irefn{org60}\And
J.G.~Contreras\Irefn{org37}\And
T.M.~Cormier\Irefn{org96}\And
Y.~Corrales Morales\Irefn{org25}\And
P.~Cortese\Irefn{org31}\And
M.R.~Cosentino\Irefn{org123}\And
F.~Costa\Irefn{org34}\And
S.~Costanza\Irefn{org139}\And
P.~Crochet\Irefn{org134}\And
E.~Cuautle\Irefn{org69}\And
P.~Cui\Irefn{org6}\And
L.~Cunqueiro\Irefn{org96}\And
D.~Dabrowski\Irefn{org142}\And
T.~Dahms\Irefn{org105}\And
A.~Dainese\Irefn{org57}\And
F.P.A.~Damas\Irefn{org115}\textsuperscript{,}\Irefn{org137}\And
M.C.~Danisch\Irefn{org104}\And
A.~Danu\Irefn{org67}\And
D.~Das\Irefn{org110}\And
I.~Das\Irefn{org110}\And
P.~Das\Irefn{org86}\And
P.~Das\Irefn{org3}\And
S.~Das\Irefn{org3}\And
A.~Dash\Irefn{org86}\And
S.~Dash\Irefn{org49}\And
S.~De\Irefn{org86}\And
A.~De Caro\Irefn{org29}\And
G.~de Cataldo\Irefn{org53}\And
J.~de Cuveland\Irefn{org39}\And
A.~De Falco\Irefn{org23}\And
D.~De Gruttola\Irefn{org10}\And
N.~De Marco\Irefn{org59}\And
S.~De Pasquale\Irefn{org29}\And
S.~Deb\Irefn{org50}\And
H.F.~Degenhardt\Irefn{org121}\And
K.R.~Deja\Irefn{org142}\And
A.~Deloff\Irefn{org85}\And
S.~Delsanto\Irefn{org25}\textsuperscript{,}\Irefn{org131}\And
W.~Deng\Irefn{org6}\And
P.~Dhankher\Irefn{org49}\And
D.~Di Bari\Irefn{org33}\And
A.~Di Mauro\Irefn{org34}\And
R.A.~Diaz\Irefn{org8}\And
T.~Dietel\Irefn{org124}\And
P.~Dillenseger\Irefn{org68}\And
Y.~Ding\Irefn{org6}\And
R.~Divi\`{a}\Irefn{org34}\And
D.U.~Dixit\Irefn{org19}\And
{\O}.~Djuvsland\Irefn{org21}\And
U.~Dmitrieva\Irefn{org62}\And
A.~Dobrin\Irefn{org67}\And
B.~D\"{o}nigus\Irefn{org68}\And
O.~Dordic\Irefn{org20}\And
A.K.~Dubey\Irefn{org141}\And
A.~Dubla\Irefn{org90}\textsuperscript{,}\Irefn{org107}\And
S.~Dudi\Irefn{org100}\And
M.~Dukhishyam\Irefn{org86}\And
P.~Dupieux\Irefn{org134}\And
R.J.~Ehlers\Irefn{org96}\And
V.N.~Eikeland\Irefn{org21}\And
D.~Elia\Irefn{org53}\And
B.~Erazmus\Irefn{org115}\And
F.~Erhardt\Irefn{org99}\And
A.~Erokhin\Irefn{org113}\And
M.R.~Ersdal\Irefn{org21}\And
B.~Espagnon\Irefn{org78}\And
G.~Eulisse\Irefn{org34}\And
D.~Evans\Irefn{org111}\And
S.~Evdokimov\Irefn{org91}\And
L.~Fabbietti\Irefn{org105}\And
M.~Faggin\Irefn{org28}\And
J.~Faivre\Irefn{org79}\And
F.~Fan\Irefn{org6}\And
A.~Fantoni\Irefn{org52}\And
M.~Fasel\Irefn{org96}\And
P.~Fecchio\Irefn{org30}\And
A.~Feliciello\Irefn{org59}\And
G.~Feofilov\Irefn{org113}\And
A.~Fern\'{a}ndez T\'{e}llez\Irefn{org45}\And
A.~Ferrero\Irefn{org137}\And
A.~Ferretti\Irefn{org25}\And
A.~Festanti\Irefn{org34}\And
V.J.G.~Feuillard\Irefn{org104}\And
J.~Figiel\Irefn{org118}\And
S.~Filchagin\Irefn{org109}\And
D.~Finogeev\Irefn{org62}\And
F.M.~Fionda\Irefn{org21}\And
G.~Fiorenza\Irefn{org53}\And
F.~Flor\Irefn{org125}\And
A.N.~Flores\Irefn{org119}\And
S.~Foertsch\Irefn{org72}\And
P.~Foka\Irefn{org107}\And
S.~Fokin\Irefn{org88}\And
E.~Fragiacomo\Irefn{org60}\And
U.~Frankenfeld\Irefn{org107}\And
U.~Fuchs\Irefn{org34}\And
C.~Furget\Irefn{org79}\And
A.~Furs\Irefn{org62}\And
M.~Fusco Girard\Irefn{org29}\And
J.J.~Gaardh{\o}je\Irefn{org89}\And
M.~Gagliardi\Irefn{org25}\And
A.M.~Gago\Irefn{org112}\And
A.~Gal\Irefn{org136}\And
C.D.~Galvan\Irefn{org120}\And
P.~Ganoti\Irefn{org84}\And
C.~Garabatos\Irefn{org107}\And
J.R.A.~Garcia\Irefn{org45}\And
E.~Garcia-Solis\Irefn{org11}\And
K.~Garg\Irefn{org115}\And
C.~Gargiulo\Irefn{org34}\And
A.~Garibli\Irefn{org87}\And
K.~Garner\Irefn{org144}\And
P.~Gasik\Irefn{org105}\textsuperscript{,}\Irefn{org107}\And
E.F.~Gauger\Irefn{org119}\And
M.B.~Gay Ducati\Irefn{org70}\And
M.~Germain\Irefn{org115}\And
J.~Ghosh\Irefn{org110}\And
P.~Ghosh\Irefn{org141}\And
S.K.~Ghosh\Irefn{org3}\And
M.~Giacalone\Irefn{org26}\And
P.~Gianotti\Irefn{org52}\And
P.~Giubellino\Irefn{org59}\textsuperscript{,}\Irefn{org107}\And
P.~Giubilato\Irefn{org28}\And
A.M.C.~Glaenzer\Irefn{org137}\And
P.~Gl\"{a}ssel\Irefn{org104}\And
A.~Gomez Ramirez\Irefn{org74}\And
V.~Gonzalez\Irefn{org107}\textsuperscript{,}\Irefn{org143}\And
\mbox{L.H.~Gonz\'{a}lez-Trueba}\Irefn{org71}\And
S.~Gorbunov\Irefn{org39}\And
L.~G\"{o}rlich\Irefn{org118}\And
A.~Goswami\Irefn{org49}\And
S.~Gotovac\Irefn{org35}\And
V.~Grabski\Irefn{org71}\And
L.K.~Graczykowski\Irefn{org142}\And
K.L.~Graham\Irefn{org111}\And
L.~Greiner\Irefn{org80}\And
A.~Grelli\Irefn{org63}\And
C.~Grigoras\Irefn{org34}\And
V.~Grigoriev\Irefn{org93}\And
A.~Grigoryan\Irefn{org1}\And
S.~Grigoryan\Irefn{org75}\And
O.S.~Groettvik\Irefn{org21}\And
F.~Grosa\Irefn{org30}\textsuperscript{,}\Irefn{org59}\And
J.F.~Grosse-Oetringhaus\Irefn{org34}\And
R.~Grosso\Irefn{org107}\And
R.~Guernane\Irefn{org79}\And
M.~Guittiere\Irefn{org115}\And
K.~Gulbrandsen\Irefn{org89}\And
T.~Gunji\Irefn{org132}\And
A.~Gupta\Irefn{org101}\And
R.~Gupta\Irefn{org101}\And
I.B.~Guzman\Irefn{org45}\And
R.~Haake\Irefn{org146}\And
M.K.~Habib\Irefn{org107}\And
C.~Hadjidakis\Irefn{org78}\And
H.~Hamagaki\Irefn{org82}\And
G.~Hamar\Irefn{org145}\And
M.~Hamid\Irefn{org6}\And
R.~Hannigan\Irefn{org119}\And
M.R.~Haque\Irefn{org63}\textsuperscript{,}\Irefn{org86}\And
A.~Harlenderova\Irefn{org107}\And
J.W.~Harris\Irefn{org146}\And
A.~Harton\Irefn{org11}\And
J.A.~Hasenbichler\Irefn{org34}\And
H.~Hassan\Irefn{org96}\And
Q.U.~Hassan\Irefn{org14}\And
D.~Hatzifotiadou\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
P.~Hauer\Irefn{org43}\And
L.B.~Havener\Irefn{org146}\And
S.~Hayashi\Irefn{org132}\And
S.T.~Heckel\Irefn{org105}\And
E.~Hellb\"{a}r\Irefn{org68}\And
H.~Helstrup\Irefn{org36}\And
A.~Herghelegiu\Irefn{org48}\And
T.~Herman\Irefn{org37}\And
E.G.~Hernandez\Irefn{org45}\And
G.~Herrera Corral\Irefn{org9}\And
F.~Herrmann\Irefn{org144}\And
K.F.~Hetland\Irefn{org36}\And
H.~Hillemanns\Irefn{org34}\And
C.~Hills\Irefn{org127}\And
B.~Hippolyte\Irefn{org136}\And
B.~Hohlweger\Irefn{org105}\And
J.~Honermann\Irefn{org144}\And
D.~Horak\Irefn{org37}\And
A.~Hornung\Irefn{org68}\And
S.~Hornung\Irefn{org107}\And
R.~Hosokawa\Irefn{org15}\textsuperscript{,}\Irefn{org133}\And
P.~Hristov\Irefn{org34}\And
C.~Huang\Irefn{org78}\And
C.~Hughes\Irefn{org130}\And
P.~Huhn\Irefn{org68}\And
T.J.~Humanic\Irefn{org97}\And
H.~Hushnud\Irefn{org110}\And
L.A.~Husova\Irefn{org144}\And
N.~Hussain\Irefn{org42}\And
S.A.~Hussain\Irefn{org14}\And
D.~Hutter\Irefn{org39}\And
J.P.~Iddon\Irefn{org34}\textsuperscript{,}\Irefn{org127}\And
R.~Ilkaev\Irefn{org109}\And
H.~Ilyas\Irefn{org14}\And
M.~Inaba\Irefn{org133}\And
G.M.~Innocenti\Irefn{org34}\And
M.~Ippolitov\Irefn{org88}\And
A.~Isakov\Irefn{org95}\And
M.S.~Islam\Irefn{org110}\And
M.~Ivanov\Irefn{org107}\And
V.~Ivanov\Irefn{org98}\And
V.~Izucheev\Irefn{org91}\And
B.~Jacak\Irefn{org80}\And
N.~Jacazio\Irefn{org34}\textsuperscript{,}\Irefn{org54}\And
P.M.~Jacobs\Irefn{org80}\And
S.~Jadlovska\Irefn{org117}\And
J.~Jadlovsky\Irefn{org117}\And
S.~Jaelani\Irefn{org63}\And
C.~Jahnke\Irefn{org121}\And
M.J.~Jakubowska\Irefn{org142}\And
M.A.~Janik\Irefn{org142}\And
T.~Janson\Irefn{org74}\And
M.~Jercic\Irefn{org99}\And
O.~Jevons\Irefn{org111}\And
M.~Jin\Irefn{org125}\And
F.~Jonas\Irefn{org96}\textsuperscript{,}\Irefn{org144}\And
P.G.~Jones\Irefn{org111}\And
J.~Jung\Irefn{org68}\And
M.~Jung\Irefn{org68}\And
A.~Jusko\Irefn{org111}\And
P.~Kalinak\Irefn{org64}\And
A.~Kalweit\Irefn{org34}\And
V.~Kaplin\Irefn{org93}\And
S.~Kar\Irefn{org6}\And
A.~Karasu Uysal\Irefn{org77}\And
D.~Karatovic\Irefn{org99}\And
O.~Karavichev\Irefn{org62}\And
T.~Karavicheva\Irefn{org62}\And
P.~Karczmarczyk\Irefn{org142}\And
E.~Karpechev\Irefn{org62}\And
A.~Kazantsev\Irefn{org88}\And
U.~Kebschull\Irefn{org74}\And
R.~Keidel\Irefn{org47}\And
M.~Keil\Irefn{org34}\And
B.~Ketzer\Irefn{org43}\And
Z.~Khabanova\Irefn{org90}\And
A.M.~Khan\Irefn{org6}\And
S.~Khan\Irefn{org16}\And
A.~Khanzadeev\Irefn{org98}\And
Y.~Kharlov\Irefn{org91}\And
A.~Khatun\Irefn{org16}\And
A.~Khuntia\Irefn{org118}\And
B.~Kileng\Irefn{org36}\And
B.~Kim\Irefn{org61}\And
B.~Kim\Irefn{org133}\And
D.~Kim\Irefn{org147}\And
D.J.~Kim\Irefn{org126}\And
E.J.~Kim\Irefn{org73}\And
H.~Kim\Irefn{org17}\And
J.~Kim\Irefn{org147}\And
J.S.~Kim\Irefn{org41}\And
J.~Kim\Irefn{org104}\And
J.~Kim\Irefn{org147}\And
J.~Kim\Irefn{org73}\And
M.~Kim\Irefn{org104}\And
S.~Kim\Irefn{org18}\And
T.~Kim\Irefn{org147}\And
T.~Kim\Irefn{org147}\And
S.~Kirsch\Irefn{org68}\And
I.~Kisel\Irefn{org39}\And
S.~Kiselev\Irefn{org92}\And
A.~Kisiel\Irefn{org142}\And
J.L.~Klay\Irefn{org5}\And
C.~Klein\Irefn{org68}\And
J.~Klein\Irefn{org34}\textsuperscript{,}\Irefn{org59}\And
S.~Klein\Irefn{org80}\And
C.~Klein-B\"{o}sing\Irefn{org144}\And
M.~Kleiner\Irefn{org68}\And
A.~Kluge\Irefn{org34}\And
M.L.~Knichel\Irefn{org34}\And
A.G.~Knospe\Irefn{org125}\And
C.~Kobdaj\Irefn{org116}\And
M.K.~K\"{o}hler\Irefn{org104}\And
T.~Kollegger\Irefn{org107}\And
A.~Kondratyev\Irefn{org75}\And
N.~Kondratyeva\Irefn{org93}\And
E.~Kondratyuk\Irefn{org91}\And
J.~Konig\Irefn{org68}\And
S.A.~Konigstorfer\Irefn{org105}\And
P.J.~Konopka\Irefn{org34}\And
G.~Kornakov\Irefn{org142}\And
L.~Koska\Irefn{org117}\And
O.~Kovalenko\Irefn{org85}\And
V.~Kovalenko\Irefn{org113}\And
M.~Kowalski\Irefn{org118}\And
I.~Kr\'{a}lik\Irefn{org64}\And
A.~Krav\v{c}\'{a}kov\'{a}\Irefn{org38}\And
L.~Kreis\Irefn{org107}\And
M.~Krivda\Irefn{org64}\textsuperscript{,}\Irefn{org111}\And
F.~Krizek\Irefn{org95}\And
K.~Krizkova~Gajdosova\Irefn{org37}\And
M.~Kr\"uger\Irefn{org68}\And
E.~Kryshen\Irefn{org98}\And
M.~Krzewicki\Irefn{org39}\And
A.M.~Kubera\Irefn{org97}\And
V.~Ku\v{c}era\Irefn{org34}\textsuperscript{,}\Irefn{org61}\And
C.~Kuhn\Irefn{org136}\And
P.G.~Kuijer\Irefn{org90}\And
L.~Kumar\Irefn{org100}\And
S.~Kundu\Irefn{org86}\And
P.~Kurashvili\Irefn{org85}\And
A.~Kurepin\Irefn{org62}\And
A.B.~Kurepin\Irefn{org62}\And
A.~Kuryakin\Irefn{org109}\And
S.~Kushpil\Irefn{org95}\And
J.~Kvapil\Irefn{org111}\And
M.J.~Kweon\Irefn{org61}\And
J.Y.~Kwon\Irefn{org61}\And
Y.~Kwon\Irefn{org147}\And
S.L.~La Pointe\Irefn{org39}\And
P.~La Rocca\Irefn{org27}\And
Y.S.~Lai\Irefn{org80}\And
M.~Lamanna\Irefn{org34}\And
R.~Langoy\Irefn{org129}\And
K.~Lapidus\Irefn{org34}\And
A.~Lardeux\Irefn{org20}\And
P.~Larionov\Irefn{org52}\And
E.~Laudi\Irefn{org34}\And
R.~Lavicka\Irefn{org37}\And
T.~Lazareva\Irefn{org113}\And
R.~Lea\Irefn{org24}\And
L.~Leardini\Irefn{org104}\And
J.~Lee\Irefn{org133}\And
S.~Lee\Irefn{org147}\And
S.~Lehner\Irefn{org114}\And
J.~Lehrbach\Irefn{org39}\And
R.C.~Lemmon\Irefn{org94}\And
I.~Le\'{o}n Monz\'{o}n\Irefn{org120}\And
E.D.~Lesser\Irefn{org19}\And
M.~Lettrich\Irefn{org34}\And
P.~L\'{e}vai\Irefn{org145}\And
X.~Li\Irefn{org12}\And
X.L.~Li\Irefn{org6}\And
J.~Lien\Irefn{org129}\And
R.~Lietava\Irefn{org111}\And
B.~Lim\Irefn{org17}\And
V.~Lindenstruth\Irefn{org39}\And
A.~Lindner\Irefn{org48}\And
C.~Lippmann\Irefn{org107}\And
M.A.~Lisa\Irefn{org97}\And
A.~Liu\Irefn{org19}\And
J.~Liu\Irefn{org127}\And
S.~Liu\Irefn{org97}\And
W.J.~Llope\Irefn{org143}\And
I.M.~Lofnes\Irefn{org21}\And
V.~Loginov\Irefn{org93}\And
C.~Loizides\Irefn{org96}\And
P.~Loncar\Irefn{org35}\And
J.A.~Lopez\Irefn{org104}\And
X.~Lopez\Irefn{org134}\And
E.~L\'{o}pez Torres\Irefn{org8}\And
J.R.~Luhder\Irefn{org144}\And
M.~Lunardon\Irefn{org28}\And
G.~Luparello\Irefn{org60}\And
Y.G.~Ma\Irefn{org40}\And
A.~Maevskaya\Irefn{org62}\And
M.~Mager\Irefn{org34}\And
S.M.~Mahmood\Irefn{org20}\And
T.~Mahmoud\Irefn{org43}\And
A.~Maire\Irefn{org136}\And
R.D.~Majka\Irefn{org146}\Aref{org*}\And
M.~Malaev\Irefn{org98}\And
Q.W.~Malik\Irefn{org20}\And
L.~Malinina\Irefn{org75}\Aref{orgIII}\And
D.~Mal'Kevich\Irefn{org92}\And
P.~Malzacher\Irefn{org107}\And
G.~Mandaglio\Irefn{org32}\textsuperscript{,}\Irefn{org56}\And
V.~Manko\Irefn{org88}\And
F.~Manso\Irefn{org134}\And
V.~Manzari\Irefn{org53}\And
Y.~Mao\Irefn{org6}\And
M.~Marchisone\Irefn{org135}\And
J.~Mare\v{s}\Irefn{org66}\And
G.V.~Margagliotti\Irefn{org24}\And
A.~Margotti\Irefn{org54}\And
A.~Mar\'{\i}n\Irefn{org107}\And
C.~Markert\Irefn{org119}\And
M.~Marquard\Irefn{org68}\And
C.D.~Martin\Irefn{org24}\And
N.A.~Martin\Irefn{org104}\And
P.~Martinengo\Irefn{org34}\And
J.L.~Martinez\Irefn{org125}\And
M.I.~Mart\'{\i}nez\Irefn{org45}\And
G.~Mart\'{\i}nez Garc\'{\i}a\Irefn{org115}\And
S.~Masciocchi\Irefn{org107}\And
M.~Masera\Irefn{org25}\And
A.~Masoni\Irefn{org55}\And
L.~Massacrier\Irefn{org78}\And
E.~Masson\Irefn{org115}\And
A.~Mastroserio\Irefn{org53}\textsuperscript{,}\Irefn{org138}\And
A.M.~Mathis\Irefn{org105}\And
O.~Matonoha\Irefn{org81}\And
P.F.T.~Matuoka\Irefn{org121}\And
A.~Matyja\Irefn{org118}\And
C.~Mayer\Irefn{org118}\And
F.~Mazzaschi\Irefn{org25}\And
M.~Mazzilli\Irefn{org53}\And
M.A.~Mazzoni\Irefn{org58}\And
A.F.~Mechler\Irefn{org68}\And
F.~Meddi\Irefn{org22}\And
Y.~Melikyan\Irefn{org62}\textsuperscript{,}\Irefn{org93}\And
A.~Menchaca-Rocha\Irefn{org71}\And
C.~Mengke\Irefn{org6}\And
E.~Meninno\Irefn{org29}\textsuperscript{,}\Irefn{org114}\And
A.S.~Menon\Irefn{org125}\And
M.~Meres\Irefn{org13}\And
S.~Mhlanga\Irefn{org124}\And
Y.~Miake\Irefn{org133}\And
L.~Micheletti\Irefn{org25}\And
L.C.~Migliorin\Irefn{org135}\And
D.L.~Mihaylov\Irefn{org105}\And
K.~Mikhaylov\Irefn{org75}\textsuperscript{,}\Irefn{org92}\And
A.N.~Mishra\Irefn{org69}\And
D.~Mi\'{s}kowiec\Irefn{org107}\And
A.~Modak\Irefn{org3}\And
N.~Mohammadi\Irefn{org34}\And
A.P.~Mohanty\Irefn{org63}\And
B.~Mohanty\Irefn{org86}\And
M.~Mohisin Khan\Irefn{org16}\Aref{orgIV}\And
Z.~Moravcova\Irefn{org89}\And
C.~Mordasini\Irefn{org105}\And
D.A.~Moreira De Godoy\Irefn{org144}\And
L.A.P.~Moreno\Irefn{org45}\And
I.~Morozov\Irefn{org62}\And
A.~Morsch\Irefn{org34}\And
T.~Mrnjavac\Irefn{org34}\And
V.~Muccifora\Irefn{org52}\And
E.~Mudnic\Irefn{org35}\And
D.~M{\"u}hlheim\Irefn{org144}\And
S.~Muhuri\Irefn{org141}\And
J.D.~Mulligan\Irefn{org80}\And
A.~Mulliri\Irefn{org23}\textsuperscript{,}\Irefn{org55}\And
M.G.~Munhoz\Irefn{org121}\And
R.H.~Munzer\Irefn{org68}\And
H.~Murakami\Irefn{org132}\And
S.~Murray\Irefn{org124}\And
L.~Musa\Irefn{org34}\And
J.~Musinsky\Irefn{org64}\And
C.J.~Myers\Irefn{org125}\And
J.W.~Myrcha\Irefn{org142}\And
B.~Naik\Irefn{org49}\And
R.~Nair\Irefn{org85}\And
B.K.~Nandi\Irefn{org49}\And
R.~Nania\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
E.~Nappi\Irefn{org53}\And
M.U.~Naru\Irefn{org14}\And
A.F.~Nassirpour\Irefn{org81}\And
C.~Nattrass\Irefn{org130}\And
R.~Nayak\Irefn{org49}\And
T.K.~Nayak\Irefn{org86}\And
S.~Nazarenko\Irefn{org109}\And
A.~Neagu\Irefn{org20}\And
R.A.~Negrao De Oliveira\Irefn{org68}\And
L.~Nellen\Irefn{org69}\And
S.V.~Nesbo\Irefn{org36}\And
G.~Neskovic\Irefn{org39}\And
D.~Nesterov\Irefn{org113}\And
L.T.~Neumann\Irefn{org142}\And
B.S.~Nielsen\Irefn{org89}\And
S.~Nikolaev\Irefn{org88}\And
S.~Nikulin\Irefn{org88}\And
V.~Nikulin\Irefn{org98}\And
F.~Noferini\Irefn{org10}\textsuperscript{,}\Irefn{org54}\And
P.~Nomokonov\Irefn{org75}\And
J.~Norman\Irefn{org79}\textsuperscript{,}\Irefn{org127}\And
N.~Novitzky\Irefn{org133}\And
P.~Nowakowski\Irefn{org142}\And
A.~Nyanin\Irefn{org88}\And
J.~Nystrand\Irefn{org21}\And
M.~Ogino\Irefn{org82}\And
A.~Ohlson\Irefn{org81}\textsuperscript{,}\Irefn{org104}\And
J.~Oleniacz\Irefn{org142}\And
A.C.~Oliveira Da Silva\Irefn{org130}\And
M.H.~Oliver\Irefn{org146}\And
C.~Oppedisano\Irefn{org59}\And
A.~Ortiz Velasquez\Irefn{org69}\And
A.~Oskarsson\Irefn{org81}\And
J.~Otwinowski\Irefn{org118}\And
K.~Oyama\Irefn{org82}\And
Y.~Pachmayer\Irefn{org104}\And
V.~Pacik\Irefn{org89}\And
S.~Padhan\Irefn{org49}\And
D.~Pagano\Irefn{org140}\And
G.~Pai\'{c}\Irefn{org69}\And
J.~Pan\Irefn{org143}\And
S.~Panebianco\Irefn{org137}\And
P.~Pareek\Irefn{org50}\textsuperscript{,}\Irefn{org141}\And
J.~Park\Irefn{org61}\And
J.E.~Parkkila\Irefn{org126}\And
S.~Parmar\Irefn{org100}\And
S.P.~Pathak\Irefn{org125}\And
B.~Paul\Irefn{org23}\And
J.~Pazzini\Irefn{org140}\And
H.~Pei\Irefn{org6}\And
T.~Peitzmann\Irefn{org63}\And
X.~Peng\Irefn{org6}\And
L.G.~Pereira\Irefn{org70}\And
H.~Pereira Da Costa\Irefn{org137}\And
D.~Peresunko\Irefn{org88}\And
G.M.~Perez\Irefn{org8}\And
S.~Perrin\Irefn{org137}\And
Y.~Pestov\Irefn{org4}\And
V.~Petr\'{a}\v{c}ek\Irefn{org37}\And
M.~Petrovici\Irefn{org48}\And
R.P.~Pezzi\Irefn{org70}\And
S.~Piano\Irefn{org60}\And
M.~Pikna\Irefn{org13}\And
P.~Pillot\Irefn{org115}\And
O.~Pinazza\Irefn{org34}\textsuperscript{,}\Irefn{org54}\And
L.~Pinsky\Irefn{org125}\And
C.~Pinto\Irefn{org27}\And
S.~Pisano\Irefn{org10}\textsuperscript{,}\Irefn{org52}\And
D.~Pistone\Irefn{org56}\And
M.~P\l osko\'{n}\Irefn{org80}\And
M.~Planinic\Irefn{org99}\And
F.~Pliquett\Irefn{org68}\And
M.G.~Poghosyan\Irefn{org96}\And
B.~Polichtchouk\Irefn{org91}\And
N.~Poljak\Irefn{org99}\And
A.~Pop\Irefn{org48}\And
S.~Porteboeuf-Houssais\Irefn{org134}\And
V.~Pozdniakov\Irefn{org75}\And
S.K.~Prasad\Irefn{org3}\And
R.~Preghenella\Irefn{org54}\And
F.~Prino\Irefn{org59}\And
C.A.~Pruneau\Irefn{org143}\And
I.~Pshenichnov\Irefn{org62}\And
M.~Puccio\Irefn{org34}\And
J.~Putschke\Irefn{org143}\And
S.~Qiu\Irefn{org90}\And
L.~Quaglia\Irefn{org25}\And
R.E.~Quishpe\Irefn{org125}\And
S.~Ragoni\Irefn{org111}\And
S.~Raha\Irefn{org3}\And
S.~Rajput\Irefn{org101}\And
J.~Rak\Irefn{org126}\And
A.~Rakotozafindrabe\Irefn{org137}\And
L.~Ramello\Irefn{org31}\And
F.~Rami\Irefn{org136}\And
S.A.R.~Ramirez\Irefn{org45}\And
R.~Raniwala\Irefn{org102}\And
S.~Raniwala\Irefn{org102}\And
S.S.~R\"{a}s\"{a}nen\Irefn{org44}\And
R.~Rath\Irefn{org50}\And
V.~Ratza\Irefn{org43}\And
I.~Ravasenga\Irefn{org90}\And
K.F.~Read\Irefn{org96}\textsuperscript{,}\Irefn{org130}\And
A.R.~Redelbach\Irefn{org39}\And
K.~Redlich\Irefn{org85}\Aref{orgV}\And
A.~Rehman\Irefn{org21}\And
P.~Reichelt\Irefn{org68}\And
F.~Reidt\Irefn{org34}\And
X.~Ren\Irefn{org6}\And
R.~Renfordt\Irefn{org68}\And
Z.~Rescakova\Irefn{org38}\And
K.~Reygers\Irefn{org104}\And
A.~Riabov\Irefn{org98}\And
V.~Riabov\Irefn{org98}\And
T.~Richert\Irefn{org81}\textsuperscript{,}\Irefn{org89}\And
M.~Richter\Irefn{org20}\And
P.~Riedler\Irefn{org34}\And
W.~Riegler\Irefn{org34}\And
F.~Riggi\Irefn{org27}\And
C.~Ristea\Irefn{org67}\And
S.P.~Rode\Irefn{org50}\And
M.~Rodr\'{i}guez Cahuantzi\Irefn{org45}\And
K.~R{\o}ed\Irefn{org20}\And
R.~Rogalev\Irefn{org91}\And
E.~Rogochaya\Irefn{org75}\And
D.~Rohr\Irefn{org34}\And
D.~R\"ohrich\Irefn{org21}\And
P.F.~Rojas\Irefn{org45}\And
P.S.~Rokita\Irefn{org142}\And
F.~Ronchetti\Irefn{org52}\And
A.~Rosano\Irefn{org56}\And
E.D.~Rosas\Irefn{org69}\And
K.~Roslon\Irefn{org142}\And
A.~Rossi\Irefn{org28}\textsuperscript{,}\Irefn{org57}\And
A.~Rotondi\Irefn{org139}\And
A.~Roy\Irefn{org50}\And
P.~Roy\Irefn{org110}\And
O.V.~Rueda\Irefn{org81}\And
R.~Rui\Irefn{org24}\And
B.~Rumyantsev\Irefn{org75}\And
A.~Rustamov\Irefn{org87}\And
E.~Ryabinkin\Irefn{org88}\And
Y.~Ryabov\Irefn{org98}\And
A.~Rybicki\Irefn{org118}\And
H.~Rytkonen\Irefn{org126}\And
O.A.M.~Saarimaki\Irefn{org44}\And
R.~Sadek\Irefn{org115}\And
S.~Sadhu\Irefn{org141}\And
S.~Sadovsky\Irefn{org91}\And
K.~\v{S}afa\v{r}\'{\i}k\Irefn{org37}\And
S.K.~Saha\Irefn{org141}\And
B.~Sahoo\Irefn{org49}\And
P.~Sahoo\Irefn{org49}\And
R.~Sahoo\Irefn{org50}\And
S.~Sahoo\Irefn{org65}\And
P.K.~Sahu\Irefn{org65}\And
J.~Saini\Irefn{org141}\And
S.~Sakai\Irefn{org133}\And
S.~Sambyal\Irefn{org101}\And
V.~Samsonov\Irefn{org93}\textsuperscript{,}\Irefn{org98}\And
D.~Sarkar\Irefn{org143}\And
N.~Sarkar\Irefn{org141}\And
P.~Sarma\Irefn{org42}\And
V.M.~Sarti\Irefn{org105}\And
M.H.P.~Sas\Irefn{org63}\And
E.~Scapparone\Irefn{org54}\And
J.~Schambach\Irefn{org119}\And
H.S.~Scheid\Irefn{org68}\And
C.~Schiaua\Irefn{org48}\And
R.~Schicker\Irefn{org104}\And
A.~Schmah\Irefn{org104}\And
C.~Schmidt\Irefn{org107}\And
H.R.~Schmidt\Irefn{org103}\And
M.O.~Schmidt\Irefn{org104}\And
M.~Schmidt\Irefn{org103}\And
N.V.~Schmidt\Irefn{org68}\textsuperscript{,}\Irefn{org96}\And
A.R.~Schmier\Irefn{org130}\And
J.~Schukraft\Irefn{org89}\And
Y.~Schutz\Irefn{org136}\And
K.~Schwarz\Irefn{org107}\And
K.~Schweda\Irefn{org107}\And
G.~Scioli\Irefn{org26}\And
E.~Scomparin\Irefn{org59}\And
J.E.~Seger\Irefn{org15}\And
Y.~Sekiguchi\Irefn{org132}\And
D.~Sekihata\Irefn{org132}\And
I.~Selyuzhenkov\Irefn{org93}\textsuperscript{,}\Irefn{org107}\And
S.~Senyukov\Irefn{org136}\And
D.~Serebryakov\Irefn{org62}\And
A.~Sevcenco\Irefn{org67}\And
A.~Shabanov\Irefn{org62}\And
A.~Shabetai\Irefn{org115}\And
R.~Shahoyan\Irefn{org34}\And
W.~Shaikh\Irefn{org110}\And
A.~Shangaraev\Irefn{org91}\And
A.~Sharma\Irefn{org100}\And
A.~Sharma\Irefn{org101}\And
H.~Sharma\Irefn{org118}\And
M.~Sharma\Irefn{org101}\And
N.~Sharma\Irefn{org100}\And
S.~Sharma\Irefn{org101}\And
O.~Sheibani\Irefn{org125}\And
K.~Shigaki\Irefn{org46}\And
M.~Shimomura\Irefn{org83}\And
S.~Shirinkin\Irefn{org92}\And
Q.~Shou\Irefn{org40}\And
Y.~Sibiriak\Irefn{org88}\And
S.~Siddhanta\Irefn{org55}\And
T.~Siemiarczuk\Irefn{org85}\And
D.~Silvermyr\Irefn{org81}\And
G.~Simatovic\Irefn{org90}\And
G.~Simonetti\Irefn{org34}\And
B.~Singh\Irefn{org105}\And
R.~Singh\Irefn{org86}\And
R.~Singh\Irefn{org101}\And
R.~Singh\Irefn{org50}\And
V.K.~Singh\Irefn{org141}\And
V.~Singhal\Irefn{org141}\And
T.~Sinha\Irefn{org110}\And
B.~Sitar\Irefn{org13}\And
M.~Sitta\Irefn{org31}\And
T.B.~Skaali\Irefn{org20}\And
M.~Slupecki\Irefn{org44}\And
N.~Smirnov\Irefn{org146}\And
R.J.M.~Snellings\Irefn{org63}\And
C.~Soncco\Irefn{org112}\And
J.~Song\Irefn{org125}\And
A.~Songmoolnak\Irefn{org116}\And
F.~Soramel\Irefn{org28}\And
S.~Sorensen\Irefn{org130}\And
I.~Sputowska\Irefn{org118}\And
J.~Stachel\Irefn{org104}\And
I.~Stan\Irefn{org67}\And
P.J.~Steffanic\Irefn{org130}\And
E.~Stenlund\Irefn{org81}\And
S.F.~Stiefelmaier\Irefn{org104}\And
D.~Stocco\Irefn{org115}\And
M.M.~Storetvedt\Irefn{org36}\And
L.D.~Stritto\Irefn{org29}\And
A.A.P.~Suaide\Irefn{org121}\And
T.~Sugitate\Irefn{org46}\And
C.~Suire\Irefn{org78}\And
M.~Suleymanov\Irefn{org14}\And
M.~Suljic\Irefn{org34}\And
R.~Sultanov\Irefn{org92}\And
M.~\v{S}umbera\Irefn{org95}\And
V.~Sumberia\Irefn{org101}\And
S.~Sumowidagdo\Irefn{org51}\And
S.~Swain\Irefn{org65}\And
A.~Szabo\Irefn{org13}\And
I.~Szarka\Irefn{org13}\And
U.~Tabassam\Irefn{org14}\And
S.F.~Taghavi\Irefn{org105}\And
G.~Taillepied\Irefn{org134}\And
J.~Takahashi\Irefn{org122}\And
G.J.~Tambave\Irefn{org21}\And
S.~Tang\Irefn{org6}\textsuperscript{,}\Irefn{org134}\And
M.~Tarhini\Irefn{org115}\And
M.G.~Tarzila\Irefn{org48}\And
A.~Tauro\Irefn{org34}\And
G.~Tejeda Mu\~{n}oz\Irefn{org45}\And
A.~Telesca\Irefn{org34}\And
L.~Terlizzi\Irefn{org25}\And
C.~Terrevoli\Irefn{org125}\And
D.~Thakur\Irefn{org50}\And
S.~Thakur\Irefn{org141}\And
D.~Thomas\Irefn{org119}\And
F.~Thoresen\Irefn{org89}\And
R.~Tieulent\Irefn{org135}\And
A.~Tikhonov\Irefn{org62}\And
A.R.~Timmins\Irefn{org125}\And
A.~Toia\Irefn{org68}\And
N.~Topilskaya\Irefn{org62}\And
M.~Toppi\Irefn{org52}\And
F.~Torales-Acosta\Irefn{org19}\And
S.R.~Torres\Irefn{org37}\And
A.~Trifir\'{o}\Irefn{org32}\textsuperscript{,}\Irefn{org56}\And
S.~Tripathy\Irefn{org50}\textsuperscript{,}\Irefn{org69}\And
T.~Tripathy\Irefn{org49}\And
S.~Trogolo\Irefn{org28}\And
G.~Trombetta\Irefn{org33}\And
L.~Tropp\Irefn{org38}\And
V.~Trubnikov\Irefn{org2}\And
W.H.~Trzaska\Irefn{org126}\And
T.P.~Trzcinski\Irefn{org142}\And
B.A.~Trzeciak\Irefn{org37}\textsuperscript{,}\Irefn{org63}\And
A.~Tumkin\Irefn{org109}\And
R.~Turrisi\Irefn{org57}\And
T.S.~Tveter\Irefn{org20}\And
K.~Ullaland\Irefn{org21}\And
E.N.~Umaka\Irefn{org125}\And
A.~Uras\Irefn{org135}\And
G.L.~Usai\Irefn{org23}\And
M.~Vala\Irefn{org38}\And
N.~Valle\Irefn{org139}\And
S.~Vallero\Irefn{org59}\And
N.~van der Kolk\Irefn{org63}\And
L.V.R.~van Doremalen\Irefn{org63}\And
M.~van Leeuwen\Irefn{org63}\And
P.~Vande Vyvre\Irefn{org34}\And
D.~Varga\Irefn{org145}\And
Z.~Varga\Irefn{org145}\And
M.~Varga-Kofarago\Irefn{org145}\And
A.~Vargas\Irefn{org45}\And
M.~Vasileiou\Irefn{org84}\And
A.~Vasiliev\Irefn{org88}\And
O.~V\'azquez Doce\Irefn{org105}\And
V.~Vechernin\Irefn{org113}\And
E.~Vercellin\Irefn{org25}\And
S.~Vergara Lim\'on\Irefn{org45}\And
L.~Vermunt\Irefn{org63}\And
R.~Vernet\Irefn{org7}\And
R.~V\'ertesi\Irefn{org145}\And
L.~Vickovic\Irefn{org35}\And
Z.~Vilakazi\Irefn{org131}\And
O.~Villalobos Baillie\Irefn{org111}\And
G.~Vino\Irefn{org53}\And
A.~Vinogradov\Irefn{org88}\And
T.~Virgili\Irefn{org29}\And
V.~Vislavicius\Irefn{org89}\And
A.~Vodopyanov\Irefn{org75}\And
B.~Volkel\Irefn{org34}\And
M.A.~V\"{o}lkl\Irefn{org103}\And
K.~Voloshin\Irefn{org92}\And
S.A.~Voloshin\Irefn{org143}\And
G.~Volpe\Irefn{org33}\And
B.~von Haller\Irefn{org34}\And
I.~Vorobyev\Irefn{org105}\And
D.~Voscek\Irefn{org117}\And
J.~Vrl\'{a}kov\'{a}\Irefn{org38}\And
B.~Wagner\Irefn{org21}\And
M.~Weber\Irefn{org114}\And
S.G.~Weber\Irefn{org144}\And
A.~Wegrzynek\Irefn{org34}\And
S.C.~Wenzel\Irefn{org34}\And
J.P.~Wessels\Irefn{org144}\And
J.~Wiechula\Irefn{org68}\And
J.~Wikne\Irefn{org20}\And
G.~Wilk\Irefn{org85}\And
J.~Wilkinson\Irefn{org10}\And
G.A.~Willems\Irefn{org144}\And
E.~Willsher\Irefn{org111}\And
B.~Windelband\Irefn{org104}\And
M.~Winn\Irefn{org137}\And
W.E.~Witt\Irefn{org130}\And
J.R.~Wright\Irefn{org119}\And
Y.~Wu\Irefn{org128}\And
R.~Xu\Irefn{org6}\And
S.~Yalcin\Irefn{org77}\And
Y.~Yamaguchi\Irefn{org46}\And
K.~Yamakawa\Irefn{org46}\And
S.~Yang\Irefn{org21}\And
S.~Yano\Irefn{org137}\And
Z.~Yin\Irefn{org6}\And
H.~Yokoyama\Irefn{org63}\And
I.-K.~Yoo\Irefn{org17}\And
J.H.~Yoon\Irefn{org61}\And
S.~Yuan\Irefn{org21}\And
A.~Yuncu\Irefn{org104}\And
V.~Yurchenko\Irefn{org2}\And
V.~Zaccolo\Irefn{org24}\And
A.~Zaman\Irefn{org14}\And
C.~Zampolli\Irefn{org34}\And
H.J.C.~Zanoli\Irefn{org63}\And
N.~Zardoshti\Irefn{org34}\And
A.~Zarochentsev\Irefn{org113}\And
P.~Z\'{a}vada\Irefn{org66}\And
N.~Zaviyalov\Irefn{org109}\And
H.~Zbroszczyk\Irefn{org142}\And
M.~Zhalov\Irefn{org98}\And
S.~Zhang\Irefn{org40}\And
X.~Zhang\Irefn{org6}\And
Z.~Zhang\Irefn{org6}\And
V.~Zherebchevskii\Irefn{org113}\And
Y.~Zhi\Irefn{org12}\And
D.~Zhou\Irefn{org6}\And
Y.~Zhou\Irefn{org89}\And
Z.~Zhou\Irefn{org21}\And
J.~Zhu\Irefn{org6}\textsuperscript{,}\Irefn{org107}\And
Y.~Zhu\Irefn{org6}\And
A.~Zichichi\Irefn{org10}\textsuperscript{,}\Irefn{org26}\And
G.~Zinovjev\Irefn{org2}\And
N.~Zurlo\Irefn{org140}\And
\renewcommand\labelenumi{\textsuperscript{\theenumi}~}
\section*{Affiliation notes}
\renewcommand\theenumi{\roman{enumi}}
\begin{Authlist}
\item \Adef{org*}Deceased
\item \Adef{orgI}Italian National Agency for New Technologies, Energy and Sustainable Economic Development (ENEA), Bologna, Italy
\item \Adef{orgII}Dipartimento DET del Politecnico di Torino, Turin, Italy
\item \Adef{orgIII}M.V. Lomonosov Moscow State University, D.V. Skobeltsyn Institute of Nuclear, Physics, Moscow, Russia
\item \Adef{orgIV}Department of Applied Physics, Aligarh Muslim University, Aligarh, India
\item \Adef{orgV}Institute of Theoretical Physics, University of Wroclaw, Poland
\end{Authlist}
\section*{Collaboration Institutes}
\renewcommand\theenumi{\arabic{enumi}~}
\begin{Authlist}
\item \Idef{org1}A.I. Alikhanyan National Science Laboratory (Yerevan Physics Institute) Foundation, Yerevan, Armenia
\item \Idef{org2}Bogolyubov Institute for Theoretical Physics, National Academy of Sciences of Ukraine, Kiev, Ukraine
\item \Idef{org3}Bose Institute, Department of Physics and Centre for Astroparticle Physics and Space Science (CAPSS), Kolkata, India
\item \Idef{org4}Budker Institute for Nuclear Physics, Novosibirsk, Russia
\item \Idef{org5}California Polytechnic State University, San Luis Obispo, California, United States
\item \Idef{org6}Central China Normal University, Wuhan, China
\item \Idef{org7}Centre de Calcul de l'IN2P3, Villeurbanne, Lyon, France
\item \Idef{org8}Centro de Aplicaciones Tecnol\'{o}gicas y Desarrollo Nuclear (CEADEN), Havana, Cuba
\item \Idef{org9}Centro de Investigaci\'{o}n y de Estudios Avanzados (CINVESTAV), Mexico City and M\'{e}rida, Mexico
\item \Idef{org10}Centro Fermi - Museo Storico della Fisica e Centro Studi e Ricerche ``Enrico Fermi', Rome, Italy
\item \Idef{org11}Chicago State University, Chicago, Illinois, United States
\item \Idef{org12}China Institute of Atomic Energy, Beijing, China
\item \Idef{org13}Comenius University Bratislava, Faculty of Mathematics, Physics and Informatics, Bratislava, Slovakia
\item \Idef{org14}COMSATS University Islamabad, Islamabad, Pakistan
\item \Idef{org15}Creighton University, Omaha, Nebraska, United States
\item \Idef{org16}Department of Physics, Aligarh Muslim University, Aligarh, India
\item \Idef{org17}Department of Physics, Pusan National University, Pusan, Republic of Korea
\item \Idef{org18}Department of Physics, Sejong University, Seoul, Republic of Korea
\item \Idef{org19}Department of Physics, University of California, Berkeley, California, United States
\item \Idef{org20}Department of Physics, University of Oslo, Oslo, Norway
\item \Idef{org21}Department of Physics and Technology, University of Bergen, Bergen, Norway
\item \Idef{org22}Dipartimento di Fisica dell'Universit\`{a} 'La Sapienza' and Sezione INFN, Rome, Italy
\item \Idef{org23}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Cagliari, Italy
\item \Idef{org24}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Trieste, Italy
\item \Idef{org25}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Turin, Italy
\item \Idef{org26}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Bologna, Italy
\item \Idef{org27}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Catania, Italy
\item \Idef{org28}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Padova, Italy
\item \Idef{org29}Dipartimento di Fisica `E.R.~Caianiello' dell'Universit\`{a} and Gruppo Collegato INFN, Salerno, Italy
\item \Idef{org30}Dipartimento DISAT del Politecnico and Sezione INFN, Turin, Italy
\item \Idef{org31}Dipartimento di Scienze e Innovazione Tecnologica dell'Universit\`{a} del Piemonte Orientale and INFN Sezione di Torino, Alessandria, Italy
\item \Idef{org32}Dipartimento di Scienze MIFT, Universit\`{a} di Messina, Messina, Italy
\item \Idef{org33}Dipartimento Interateneo di Fisica `M.~Merlin' and Sezione INFN, Bari, Italy
\item \Idef{org34}European Organization for Nuclear Research (CERN), Geneva, Switzerland
\item \Idef{org35}Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split, Split, Croatia
\item \Idef{org36}Faculty of Engineering and Science, Western Norway University of Applied Sciences, Bergen, Norway
\item \Idef{org37}Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czech Republic
\item \Idef{org38}Faculty of Science, P.J.~\v{S}af\'{a}rik University, Ko\v{s}ice, Slovakia
\item \Idef{org39}Frankfurt Institute for Advanced Studies, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org40}Fudan University, Shanghai, China
\item \Idef{org41}Gangneung-Wonju National University, Gangneung, Republic of Korea
\item \Idef{org42}Gauhati University, Department of Physics, Guwahati, India
\item \Idef{org43}Helmholtz-Institut f\"{u}r Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universit\"{a}t Bonn, Bonn, Germany
\item \Idef{org44}Helsinki Institute of Physics (HIP), Helsinki, Finland
\item \Idef{org45}High Energy Physics Group, Universidad Aut\'{o}noma de Puebla, Puebla, Mexico
\item \Idef{org46}Hiroshima University, Hiroshima, Japan
\item \Idef{org47}Hochschule Worms, Zentrum f\"{u}r Technologietransfer und Telekommunikation (ZTT), Worms, Germany
\item \Idef{org48}Horia Hulubei National Institute of Physics and Nuclear Engineering, Bucharest, Romania
\item \Idef{org49}Indian Institute of Technology Bombay (IIT), Mumbai, India
\item \Idef{org50}Indian Institute of Technology Indore, Indore, India
\item \Idef{org51}Indonesian Institute of Sciences, Jakarta, Indonesia
\item \Idef{org52}INFN, Laboratori Nazionali di Frascati, Frascati, Italy
\item \Idef{org53}INFN, Sezione di Bari, Bari, Italy
\item \Idef{org54}INFN, Sezione di Bologna, Bologna, Italy
\item \Idef{org55}INFN, Sezione di Cagliari, Cagliari, Italy
\item \Idef{org56}INFN, Sezione di Catania, Catania, Italy
\item \Idef{org57}INFN, Sezione di Padova, Padova, Italy
\item \Idef{org58}INFN, Sezione di Roma, Rome, Italy
\item \Idef{org59}INFN, Sezione di Torino, Turin, Italy
\item \Idef{org60}INFN, Sezione di Trieste, Trieste, Italy
\item \Idef{org61}Inha University, Incheon, Republic of Korea
\item \Idef{org62}Institute for Nuclear Research, Academy of Sciences, Moscow, Russia
\item \Idef{org63}Institute for Subatomic Physics, Utrecht University/Nikhef, Utrecht, Netherlands
\item \Idef{org64}Institute of Experimental Physics, Slovak Academy of Sciences, Ko\v{s}ice, Slovakia
\item \Idef{org65}Institute of Physics, Homi Bhabha National Institute, Bhubaneswar, India
\item \Idef{org66}Institute of Physics of the Czech Academy of Sciences, Prague, Czech Republic
\item \Idef{org67}Institute of Space Science (ISS), Bucharest, Romania
\item \Idef{org68}Institut f\"{u}r Kernphysik, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org69}Instituto de Ciencias Nucleares, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org70}Instituto de F\'{i}sica, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil
\item \Idef{org71}Instituto de F\'{\i}sica, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org72}iThemba LABS, National Research Foundation, Somerset West, South Africa
\item \Idef{org73}Jeonbuk National University, Jeonju, Republic of Korea
\item \Idef{org74}Johann-Wolfgang-Goethe Universit\"{a}t Frankfurt Institut f\"{u}r Informatik, Fachbereich Informatik und Mathematik, Frankfurt, Germany
\item \Idef{org75}Joint Institute for Nuclear Research (JINR), Dubna, Russia
\item \Idef{org76}Korea Institute of Science and Technology Information, Daejeon, Republic of Korea
\item \Idef{org77}KTO Karatay University, Konya, Turkey
\item \Idef{org78}Laboratoire de Physique des 2 Infinis, Ir\`{e}ne Joliot-Curie, Orsay, France
\item \Idef{org79}Laboratoire de Physique Subatomique et de Cosmologie, Universit\'{e} Grenoble-Alpes, CNRS-IN2P3, Grenoble, France
\item \Idef{org80}Lawrence Berkeley National Laboratory, Berkeley, California, United States
\item \Idef{org81}Lund University Department of Physics, Division of Particle Physics, Lund, Sweden
\item \Idef{org82}Nagasaki Institute of Applied Science, Nagasaki, Japan
\item \Idef{org83}Nara Women{'}s University (NWU), Nara, Japan
\item \Idef{org84}National and Kapodistrian University of Athens, School of Science, Department of Physics , Athens, Greece
\item \Idef{org85}National Centre for Nuclear Research, Warsaw, Poland
\item \Idef{org86}National Institute of Science Education and Research, Homi Bhabha National Institute, Jatni, India
\item \Idef{org87}National Nuclear Research Center, Baku, Azerbaijan
\item \Idef{org88}National Research Centre Kurchatov Institute, Moscow, Russia
\item \Idef{org89}Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark
\item \Idef{org90}Nikhef, National institute for subatomic physics, Amsterdam, Netherlands
\item \Idef{org91}NRC Kurchatov Institute IHEP, Protvino, Russia
\item \Idef{org92}NRC \guillemotleft Kurchatov\guillemotright~Institute - ITEP, Moscow, Russia
\item \Idef{org93}NRNU Moscow Engineering Physics Institute, Moscow, Russia
\item \Idef{org94}Nuclear Physics Group, STFC Daresbury Laboratory, Daresbury, United Kingdom
\item \Idef{org95}Nuclear Physics Institute of the Czech Academy of Sciences, \v{R}e\v{z} u Prahy, Czech Republic
\item \Idef{org96}Oak Ridge National Laboratory, Oak Ridge, Tennessee, United States
\item \Idef{org97}Ohio State University, Columbus, Ohio, United States
\item \Idef{org98}Petersburg Nuclear Physics Institute, Gatchina, Russia
\item \Idef{org99}Physics department, Faculty of science, University of Zagreb, Zagreb, Croatia
\item \Idef{org100}Physics Department, Panjab University, Chandigarh, India
\item \Idef{org101}Physics Department, University of Jammu, Jammu, India
\item \Idef{org102}Physics Department, University of Rajasthan, Jaipur, India
\item \Idef{org103}Physikalisches Institut, Eberhard-Karls-Universit\"{a}t T\"{u}bingen, T\"{u}bingen, Germany
\item \Idef{org104}Physikalisches Institut, Ruprecht-Karls-Universit\"{a}t Heidelberg, Heidelberg, Germany
\item \Idef{org105}Physik Department, Technische Universit\"{a}t M\"{u}nchen, Munich, Germany
\item \Idef{org106}Politecnico di Bari, Bari, Italy
\item \Idef{org107}Research Division and ExtreMe Matter Institute EMMI, GSI Helmholtzzentrum f\"ur Schwerionenforschung GmbH, Darmstadt, Germany
\item \Idef{org108}Rudjer Bo\v{s}kovi\'{c} Institute, Zagreb, Croatia
\item \Idef{org109}Russian Federal Nuclear Center (VNIIEF), Sarov, Russia
\item \Idef{org110}Saha Institute of Nuclear Physics, Homi Bhabha National Institute, Kolkata, India
\item \Idef{org111}School of Physics and Astronomy, University of Birmingham, Birmingham, United Kingdom
\item \Idef{org112}Secci\'{o}n F\'{\i}sica, Departamento de Ciencias, Pontificia Universidad Cat\'{o}lica del Per\'{u}, Lima, Peru
\item \Idef{org113}St. Petersburg State University, St. Petersburg, Russia
\item \Idef{org114}Stefan Meyer Institut f\"{u}r Subatomare Physik (SMI), Vienna, Austria
\item \Idef{org115}SUBATECH, IMT Atlantique, Universit\'{e} de Nantes, CNRS-IN2P3, Nantes, France
\item \Idef{org116}Suranaree University of Technology, Nakhon Ratchasima, Thailand
\item \Idef{org117}Technical University of Ko\v{s}ice, Ko\v{s}ice, Slovakia
\item \Idef{org118}The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland
\item \Idef{org119}The University of Texas at Austin, Austin, Texas, United States
\item \Idef{org120}Universidad Aut\'{o}noma de Sinaloa, Culiac\'{a}n, Mexico
\item \Idef{org121}Universidade de S\~{a}o Paulo (USP), S\~{a}o Paulo, Brazil
\item \Idef{org122}Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil
\item \Idef{org123}Universidade Federal do ABC, Santo Andre, Brazil
\item \Idef{org124}University of Cape Town, Cape Town, South Africa
\item \Idef{org125}University of Houston, Houston, Texas, United States
\item \Idef{org126}University of Jyv\"{a}skyl\"{a}, Jyv\"{a}skyl\"{a}, Finland
\item \Idef{org127}University of Liverpool, Liverpool, United Kingdom
\item \Idef{org128}University of Science and Technology of China, Hefei, China
\item \Idef{org129}University of South-Eastern Norway, Tonsberg, Norway
\item \Idef{org130}University of Tennessee, Knoxville, Tennessee, United States
\item \Idef{org131}University of the Witwatersrand, Johannesburg, South Africa
\item \Idef{org132}University of Tokyo, Tokyo, Japan
\item \Idef{org133}University of Tsukuba, Tsukuba, Japan
\item \Idef{org134}Universit\'{e} Clermont Auvergne, CNRS/IN2P3, LPC, Clermont-Ferrand, France
\item \Idef{org135}Universit\'{e} de Lyon, Universit\'{e} Lyon 1, CNRS/IN2P3, IPN-Lyon, Villeurbanne, Lyon, France
\item \Idef{org136}Universit\'{e} de Strasbourg, CNRS, IPHC UMR 7178, F-67000 Strasbourg, France, Strasbourg, France
\item \Idef{org137}Universit\'{e} Paris-Saclay Centre d'Etudes de Saclay (CEA), IRFU, D\'{e}partment de Physique Nucl\'{e}aire (DPhN), Saclay, France
\item \Idef{org138}Universit\`{a} degli Studi di Foggia, Foggia, Italy
\item \Idef{org139}Universit\`{a} degli Studi di Pavia, Pavia, Italy
\item \Idef{org140}Universit\`{a} di Brescia, Brescia, Italy
\item \Idef{org141}Variable Energy Cyclotron Centre, Homi Bhabha National Institute, Kolkata, India
\item \Idef{org142}Warsaw University of Technology, Warsaw, Poland
\item \Idef{org143}Wayne State University, Detroit, Michigan, United States
\item \Idef{org144}Westf\"{a}lische Wilhelms-Universit\"{a}t M\"{u}nster, Institut f\"{u}r Kernphysik, M\"{u}nster, Germany
\item \Idef{org145}Wigner Research Centre for Physics, Budapest, Hungary
\item \Idef{org146}Yale University, New Haven, Connecticut, United States
\item \Idef{org147}Yonsei University, Seoul, Republic of Korea
\end{Authlist}
\endgroup
\section*{Acknowledgements}
\input{fa_2020-05-11.tex}
\end{acknowledgement}
\bibliographystyle{utphys}
| f31d10ccaeba351f11a938af3c1c18377a876273 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The numerous applications of 3D vision, such as augmented reality, autonomous driving and robotic locomotion, aided by the ever increasing computational power of modern machines have caused a steep increase of interest in 3D vision and its applications. The research in the field has also been predominantly motivated by the wide scale access to good quality and low cost 3D scanners, which has opened many new avenues.
However, despite of all the recent advancements of algorithms for 3D meshes and point clouds, there is still room for enhancements to match the performance of the systems designed for 2D images on similar tasks. Nearly all the state-of-the-art methods in the various fields of computer vision are designed for 2D images and there are also immense differences in the sizes of training datasets for 2D images compared to any of their 3D counterparts. This points to a need for the development of bridging techniques, which can make efficient uses of both 2D images and 3D point clouds or meshes. A technique that can fuse information from both 2D and 3D domains can benefit from the matured 2D vision as well as the contemporary advances in 3D vision. The complementary nature of 2D and 3D data can lead to an improved performance and efficiency in many potential applications, e.g., face recognition and verification, identification of objects or different regions of interest from coloured images in 3D maps, use of the 3D model of an object to locate it in an image in the presence of severe perspective distortion \citep{laga20183d}, as well as localization of a 2D images in a 3D point cloud.
To move towards the goal of multi-domain information fusion from 2D and 3D domains, this paper proposes a novel approach to learn a framework to directly match feature descriptors extracted from 3D point clouds with those extracted from 2D images. This concept is used to localize images in point clouds directly generated from 3D scanners. To localize images, the developed framework is used to match 3D points of the point clouds to their corresponding pixels in 2D images with the help of multi-domain descriptor matching. The matched points between the images and the point clouds are then used to estimate the six degrees-of-freedom (DOF) position and orientation (pose) of images in the 3D world.
6-DOF pose estimation is an important research topic due to its numerous applications such as augmented reality, place recognition, robotic grasping, navigation, robotic pose estimation as well as simultaneous localization and mapping (SLAM). Current techniques for camera pose estimation can be classified into two major categories: \textbf{(i)} Regression networks based methods and \textbf{(ii)} Features based methods.
The regression networks based methods (e.g. \cite{posenet,kendall2017geometric,walch2017image}), use deep neural networks to estimate the pose of the camera and consequently have high requirements for computational resources such as powerful GPUs and require a lot of training data \citep{xin2019review} from different viewpoints to ensure that the poses of query cameras are sufficiently close to the training ones \citep{sattler2019understanding}. Moreover they scale poorly with the increase in the size of 3D models and usually run into the problems of non-convergence for end-to-end training \citep{brachmann2018learning}.
Features based method use hand-crafted approaches or deep learning methods to extract local or global features from images. An essential step in the state-of-the-art techniques in this category is the use of the Structure from Motion (SfM) pipeline \citep{colmapsfm,han2019image} to create sparse 3D models from images (e.g. \citep{li2010location,sattler2015hyperpoints,sattler2017efficient}). Structure from Motion pipelines provide a one-to-one correspondence between the points in the generated sparse point cloud and the pixels of the 2D images that were used to create the model. Several works use this information to localize images with respect to the 3D point cloud generated by SfM. However, model creation with SfM is a computationally expensive process which may be very time consuming depending on the number of images used and the quality of the point cloud to be generated. Models generated with SfM have especially poor quality for or miss out entirely on texture-less regions. Moreover, dependency on SfM generated point clouds renders such techniques futile for scenarios where point clouds have been obtained from 3D scanners. Now-a-days, high quality and user friendly 3D scanners (e.g. LIDAR, Microsoft Kinect, Matterport scanners and Faro 3D scanners) are available which can effectively render dense point clouds of large areas without the use of SfM. These point clouds are of better quality, not only because of their higher point density but also due to the reason that they can effectively capture bland surfaces that SfM based techniques tend to miss.
To be able to directly localize 2D images in the point clouds generated from any 3D scanners, we propose a novel concept to directly match feature descriptors extracted from 3D point clouds with descriptors extracted from 2D images. The matched descriptors can then be used for 6-DOF camera localization of the image in the 3D point cloud.
Figure \ref{fig:result} shows a section of the dense point cloud of the Shop Facade dataset \citep{posenet} along with a query image and the localization results of the proposed technique.
To match the feature descriptors from 2D and 3D domain, we generate a dataset of corresponding 3D and 2D descriptors to train a two-stage classifier called `Descriptor-Matcher'. For localization of a 2D image in a point cloud, we first use image feature extraction techniques such as Scale Invariant Feature Transform (SIFT) \citep{lowesift} to extract keypoints and their descriptors from the 2D image. Similarly, we use techniques designed for point clouds \citep{guo20143d} such as 3D-SIFT key-points \citep{lowesift,pcl}, 3D-Harris key-points \citep{harris,laga20183d} and Rotation Invariant Features Transform (RIFT) descriptors \citep{rift} to extract 3D keypoints and descriptors from the point cloud.
The Descriptor-Matcher is then used to find the matching pairs of 3D and 2D descriptors and their corresponding keypoints. This results in a list of coordinates in the point cloud and their corresponding pixels in the query image. The resulting one-to-one matches are then used in a robust algorithm for 6-DOF pose estimation to find the location and orientation of the query camera in the 3D point cloud. Figure \ref{fig:test_chart} shows the steps involved in our technique to estimate the camera pose for a query image.
A preliminary version of this work appeared in \cite{my2d3d}. To the best of our knowledge, \cite{my2d3d} was the first work: \textbf{(i)} to match directly 3D descriptors extracted from dense point clouds with the 2D descriptors from RGB images, and \textbf{(ii)} to use direct matching of 2D and 3D descriptors to localize camera pose with 6-DOF in dense 3D point clouds. This work extends \cite{my2d3d} by improving all the elements of the proposed technique, including dataset generation, Descriptor-Matcher and pose estimation. Additionally, we propose a reliable method to create a large collection of 2D and 3D points with matching locations in images and point cloud, respectively. We also present more details of the proposed method and also evaluate the proposed technique on more challenging datasets compared to \cite{my2d3d}.
The rest of this paper is organized as follows. Section \ref{related_work} discusses the various categories of the techniques in the literature for camera localization or geo-registration of images. The details of the proposed technique are presented in Section \ref{technique}. Section \ref{experiments} reports the experimental setup and a detailed evaluation of our results.
Finally, the paper is concluded in Section \ref{conclusion}.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1\linewidth]{Figures2.pdf}
\end{center}
\caption{A block diagram of the test pipeline of the proposed technique. We extract 3D key-points and descriptors from the dense 3D Point Cloud. 2D key-points and their corresponding descriptors are extracted from the 2D RGB Query Image. Then our proposed `Descriptor-Matcher' algorithm directly matches the 2D descriptors with the 3D descriptors to generate correspondence between points in 2D image and 3D point cloud. This is then used with a robust pose estimation algorithm to estimate the 6-DOF pose of the query image in the 3D point cloud.}
\label{fig:test_chart}
\end{figure*}
\section{Related Work}\label{related_work}
The numerous applications of camera pose estimation and image localization render it as an interesting and active field of research. Traditionally, there are two distinct approaches to estimate the position and orientation of a camera \citep{xin2019review}: \textbf{(i)} Features based methods and \textbf{(ii)} Network based pose regression methods. The proposed method forms a new category of possible approaches: \textbf{(iii)} Direct 2D-3D descriptor matching based methods.
\subsection{Features based methods} Features based methods extract convolutional or hand-crafted features from 2D images and use them in different manners to localize the images. However, many of these methods only estimate approximate locations and use number of inliers found with RANSAC \citep{ransac} as a criterion for image registration e.g., a query image is considered as registered if the RANSAC stage of the method can find 12 inliers among the features for the query image. This is partly due to the fact that some datasets for image localization do not provide the ground truth position and orientation information for the query cameras. However, inlier count is not a reliable criterion and it does not represent the actual performance of any given method. Feature based methods can further be classified into two types of methods: Image retrieval based methods and SfM based methods.
\subsubsection{Image retrieval based methods }
Image retrieval based methods involve the use of a large geo-tagged database of images. To localize a 2D query image, these methods use different types of features to retrieve images similar to the query image. The average of the retrieved database images can be used as the predicted location of the query image. Alternatively, the poses of the retrieved images can be used to triangulate the pose of the query camera \citep{chen2011city,zamir2010accurate}. These methods cannot be used for the localization of 2D images in point clouds. Also, many times the images in the dataset are not sufficiently close to the query image which results in significant errors in pose estimation.
\subsubsection{SfM-based methods}
SfM-based methods produce better pose estimates than image retrieval based methods \citep{sattler2017efficient}. These methods use the SfM pipeline \citep{colmapsfm}. SfM first extracts and matches features such as SIFT, SURF or ORB, from the set of training images. Then the matched features between the different 2D images are used to create a 3D model on an arbitrary scale. Each point in the 3D model is created by triangulation of points from multiple images.
\cite{irschara2009structure} used image retrieval to extract 2D images that were similar to the query 2D image from the training database. The extracted images were used with the SfM model to improve the accuracy of the estimated camera poses.
\cite{li2010location} compared the features extracted from the query 2D images with the 2D features corresponding to the points in the SfM model to localize the images.
\cite{sattler2015hyperpoints} improved the localization process with a visual vocabulary of 16 million words created from the features of the database images and their locations in the SfM point cloud.
Later, \cite{sattler2017efficient} further extended their work with the help of a prioritized matching system to improve the estimated poses for the query images.
However, these methods can only work with point clouds generated with SfM based pipelines \citep{piasco2018survey}. This is mainly due to the reason that the sparse point clouds generated with SfM save the corresponding information of the points and features from 2D images that were used to create the sparse SfM point cloud. SfM based methods rely on this information for pose estimation. Also some works use element-wise mean (or any other suitable function) of the feature descriptors of the points that were used to create a 3D point in the SfM point cloud as a 3D descriptor of that 3D point. Such an approximation of 3D features is dependent on the inherent information in the point cloud generated with SfM, which is not available if the point cloud had to be generated from a 3D scanner.
Structure from Motion is a computationally expensive and time consuming process. The point clouds generated with SfM are sparse and very noisy \citep{feng20192d3d}. SfM models have especially poor quality at bland or texture-less regions and may miss out such areas altogether. Moreover, it requires multiple images of the same area from many different angles for good results, which is practically a difficult requirement, especially for large areas or buildings. Also the generated models are created on an arbitrary scale and it is not possible to determine the exact size of the model without extra information from other sources \citep{feng20192d3d}. The availability of high quality 3D scanners has made it possible to capture large scale point clouds in an efficient manner without the need to capture thousands of images to be used for the SfM pipeline. Moreover, LIDAR and other 3D scanners are now becoming an essential part of robots, particularly for locomotion and grasping. Therefore, it is essential to develop methods that can estimate 6-DOF poses for cameras in point clouds captured from any scanner.
\subsection{Network-based pose regression methods}
These methods use deep neural networks to estimate the position and orientation of query images through pose regression. However, \cite{sattler2019understanding} showed that these methods can only produce good results when the poses of the query images are sufficiently close to the training images.
\cite{posenet} proposed a convolutional neural network, called PoseNet, which was trained to regress the 6-DOF camera pose of the query image.
\cite{kendall2017geometric} later improved the loss function in Posenet while \cite{walch2017image} improved the network architecture to reduce the errors in pose estimations.
\cite{brachmann2017dsac} introduced the concept of differentiable RANSAC. Instead of direct pose regression through a neural network, they created a differential version of RANSAC \citep{ransac} to estimate the location and orientation of the query cameras. Specifically, model of \cite{brachmann2017dsac} was composed of two CNNs, one to predict scene coordinates and the other CNN to select a camera pose from a pool of hypotheses generated from the output of the first CNN. However, their method was prone to over-fitting and did not always converge during end-to-end-optimization especially for outdoor scenes \citep{brachmann2018learning}.
\cite{radwan2018vlocnet++} used a deep learning based architecture to simultaneously carry out camera pose estimation, semantic segmentation and odometry estimation by exploiting the inter-dependencies of these tasks for assisting each other.
\cite{brachmann2018learning} modified \citep{brachmann2017dsac} with a fully convolutional network for scene coordinate regression as the only learnable component in their system. Soft inlier count was used to test the pose hypotheses. Although it improved the localization accuracy, the system still failed to produce results for datasets with large scale.
\subsection{Direct 2D-3D descriptor matching based methods}
Our preliminary work \citep{my2d3d} was the first technique to estimate the 6-DOF pose for query cameras by directly matching features extracted from 2D images and 3D point clouds. \cite{feng20192d3d} trained a deep convolutional network with triplet loss to estimate descriptors for patches extracted from images and point clouds. The estimated patches were used in an exhaustive feature matching algorithm for pose estimation. However, their system achieved a limited localization capability.
This paper improves the concept of \citep{my2d3d} with an unconstrained method to extract pairs of corresponding 2D and 3D points for training, and improves the structure of Descriptor-Matcher. It also introduces a better pose estimation strategy. Our technique directly extracts 3D key-points and feature descriptors from point clouds which, in contrast to SfM based approaches, enables us to estimate poses in point clouds generated from any 3D scanner. Moreover, the capability of our method to work with dense point clouds allows it to use better quality feature descriptors, which additionally helps in the pose estimation of the query images.
\section{Proposed Technique}\label{technique}
The direct matching of 2D and 3D descriptors in our technique provides a way to localize the position and orientation of 2D query images in point clouds which is not constrained by the method of point cloud generation, the type of 3D scanners used or the indoor and outdoor nature of the environment.
At the core of our technique is a classifier, called `Descriptor-Matcher', which is used to match the feature descriptors from 2D and 3D domain. Descriptor-Matcher is composed of two stages: coarse and fine. To train the Descriptor-Matcher, we need a dataset of 3D points and their corresponding pixel locations in the training images. It is impractical to manually create a dataset of matching 2D and 3D points, especially for large scales. To overcome this issue, we propose a method to automatically collect a large number of matching 3D and 2D points and their corresponding descriptors from a point cloud and a given set of 2D training images. The trained Descriptor-Matcher can then be used to localize a 2D query image in the point cloud.
\subsection{Extraction of Key-points and Feature Descriptors}
\label{technique:keypoint_extraction}
Let us assume a dataset with $N$ number of training images with known ground truth poses.
Let $PC_d$ be the point cloud in which the query images need to be localized:
\begin{equation}
\label{e1}
PC_{d}= \bigcup_{i=1}^{Np}(x_i,y_i,z_i)
\end{equation}
where $\bigcup$ is the union operator, $Np$ is the number of points in the point cloud and $(x_i,y_i,z_i)$ are the Cartesian coordinates of the $i^{th}$ point of the point cloud. Nowadays, most of the 3D sensors can produce RGB images with ground truth camera positions relative to the generated point clouds. In case such information is not available, Multi View Stereo \citep{colmapmvs} can be used to generate ground truth camera poses for the training images as explained in \citep{my2d3d}.
To create the dataset for training, we first use methods designed for point clouds to extract 3D keypoints and feature descriptors from the point cloud $PC_d$ \citep{pcl,guo20143d}. This results in a set of 3D keypoints:
\begin{equation}
\label{e2}
keys3D= \bigcup_{j=1}^{N3}(x_j,y_j,z_j)
\end{equation}
and their corresponding 3D descriptors:
\begin{equation}
\label{e3}
desc3D= \bigcup_{j=1}^{N3}(u_j^1,u_j^2,u_j^3,...u_j^{p})
\end{equation}
where $(u_j^1,u_j^2,u_j^3,...u_j^{p})$ is the $p$ dimensional 3D descriptor of the $j^{th}$ 3D keypoint with Cartesian coordinates $(x_j,y_j,z_j)$, and $N3$ is the number of the detected keypoints in the dense point cloud.
Similarly, we use keypoint and descriptor extraction methods specific for images to get a set of 2D keypoints $keys2d$ and feature descriptors $desc2d$ for each training image:
\begin{equation}
\label{e4}
keys2D_n= \bigcup_{k=1}^{N2_n}(\widetilde{x}_{k,n},\widetilde{y}_{k,n})
\end{equation}
\begin{equation}
\label{e5}
desc2D_n= \bigcup_{k=1}^{N2_n}(v_{k,n}^1,v_{k,n}^2,v_{k,n}^3,...v_{k,n}^{q})
\end{equation}
where $N2_n$ is the number of the detected keypoints in $n^{th}$ image. $\widetilde{x}_{k,n}$ and $\widetilde{y}_{k,n}$ are the horizontal and vertical pixel coordinates, respectively, of the $k^{th}$ keypoint in the $n^{th}$ training image and $(v_{k,n}^1,v_{k,n}^2,v_{k,n}^3,...v_{k,n}^{q})$ is the $q$ dimensional 2D descriptor of that keypoint.
\subsection{Dataset Generation}
\label{technique:dataset}
\subsubsection{Back ray Tracing}
To calculate the distance between 2D and 3D keypoints, we need to either project the 3D keypoints onto the 2D image or the 2D keypoints onto the 3D point cloud.
Projecting the 2D keypoints will involve the voxelization of the point cloud which is a computationally expensive process compared to its alternative (i.e., back ray tracing) for the task at hand. Therefore, for each image in the training set, we use the intrinsic and extrinsic matrix of the camera to trace back rays from the 3D key points $keys_{3d}$ on the image. Mathematically, back ray tracing can be represented through the following equation:
\begin{equation}
\label{e6}
\begin{aligned}
\begin{bmatrix} \hat{s}_{j,n}\, \hat{x}_{j,n},\,\hat{s}_{j,n}\, \hat{y}_{j,n},\,\hat{s}_{j,n} \end{bmatrix}^{tr} = \\ \begin{bmatrix}K_n\end{bmatrix} \! \begin{bmatrix}R_n|T_n\end{bmatrix} \! \begin{bmatrix}x_j,y_j,z_j,1\end{bmatrix}^{tr}
\\ \forall j=1,2,3,...,N3
\end{aligned}
\end{equation}
Where $K_n$ is the intrinsic matrix of the camera, $R_n$ is the rotation matrix and $T_n$ is the translation vector for the conversion of points from world coordinates to camera coordinates for the $n^{th}$ image. The $tr$ superscript denotes the transpose of a matrix. The $\hat{x}_{j,n}$ and $\hat{y}_{j,n}$ are the horizontal and vertical pixel coordinates, respectively, of the projected 3D keypoint on the $n^{th}$ image and $\hat{s}_{j,n}$ is the depth of the 3D keypoint from the camera. We keep only those projected keypoints that are within the boundaries of the image and have a positive depth value so that they are in front of the camera:
\begin{equation}\label{e7}
\begin{aligned}
\Theta_n= \bigcup(\hat{x}_{j,n}, & \hat{y}_{j,n},\hat{s}_{j,n}) \ s.t. \ (0<\hat{x}_{j,n}<img\_w_n)\ \\ and & \ (0<\hat{y}_{j,n}<img\_h_n)\ and\ (\hat{s}_{j,n}>0) \\ & \quad \qquad \qquad \qquad \forall \ j=1,2,3,...,N3
\end{aligned}
\end{equation}
Where $img\_w_n$ and $img\_h_n$ are the width and height of the $n^{th}$ image, respectively. However, as keypoints are extremely sparse, it is possible that points from the areas of the point cloud that are not visible in the image, get projected on the image. For example, if the camera is looking at a wall at a right angle and there is another wall behind the first one, then it is possible that the keypoints detected on the back wall get projected on the image. Theses false projections can create problems during the training of the Descriptor-Matcher as they result in the matching of locations between images and point clouds which are not the same and increase the number of false positives. Figure \ref{fig:depth_block_filter} shows examples of false projections of 3D keypoints from locations in point cloud which are not visible in the 2D image.
\subsubsection{Depth Block Filtering}
\label{technique:dataset:DBF}
To overcome this problem, we devised a strategy based on the 3D points of the point cloud $PC_d$, called Depth Block Filtering.
Similar to the 3D keypoints, we back trace the points in $PC_d$ to the image with the help of the intrinsic matrix of the camera, $K_n$, and the augmented matrix created by appending the rotation matrix $R_n$ with the translation vector $T_n$ to get the pixel coordinates of the 3D points:
\begin{equation}
\label{e8}
\begin{aligned}
\begin{bmatrix} \hat{s}_{i,n}\,\hat{x}_{i,n},\,\hat{s}_{i,n}\,\hat{y}_{i,n},\,\hat{s}_{i,n} \end{bmatrix} ^{tr}\!=\\ \begin{bmatrix}K_n\end{bmatrix} \! \begin{bmatrix}R_n|T_n\end{bmatrix} \! \begin{bmatrix}x_i,y_i,z_i,1\end{bmatrix}^{tr}
\\ \forall i=1,2,3,...,Np
\end{aligned}
\end{equation}
where $\hat{x}_{i,n}$ and $\hat{y}_{i,n}$ are the horizontal and vertical pixel coordinates, respectively, of the projected points of the 3D point cloud on the $n^{th}$ image and $\hat{s}_{i,n}$ is the depth of the projected point from the camera.
We filter out all those points that are outside the boundaries of the image and behind the camera i.e., those with a negative depth value $\hat{s}_{i,n}$.
\begin{equation}\label{e9}
\begin{aligned}
\Gamma_n= \bigcup(\hat{x}_{i,n}, & \hat{y}_{i,n},\hat{s}_{i,n}) \ s.t. \ (0<\hat{x}_{i,n}<img\_w_n)\ \\ and & \ (0<\hat{y}_{i,n}<img\_h_n)\ and\ (\hat{s}_{i,n}>0) \\ & \quad \qquad \qquad\qquad \forall \ i=1,2,3,...,Np
\end{aligned}
\end{equation}
We convert the projected pixel values $\hat{x_i}$ and $\hat{y_i}$ in $\Gamma_n$ to whole numbers. Any duplicates in the resulting set that have the same values for both $\hat{x_i}$ and $\hat{y_i}$ are removed by keeping only those elements of $\Gamma_n$ which are closest to the position of the camera i.e., among the duplicate pixel locations, we keep the triple with the minimum value of $\hat{s_i}$. The depth values $\hat{s_i}$ in $\Gamma_n$ are then used to create a depth map for the training image.
Next, we use the generated depth map to create blocks of size $\tau \times \tau$ pixels around the locations of the triples in the set of projected 3D keypoints $\Theta_n$.
Finally, we filter out all the triples in $\Theta_n$ whose depth values $\hat{s_j}$ are greater than a threshold $\varphi$ in their respective depth block, where $\varphi$ and $\tau$ are constants. This strategy helps to cater not only for wrong projections but also for any holes in the point cloud as well. Figure \ref{fig:depth_block_filter} shows examples of the results from depth block filtering.
\begin{figure*}[htbp]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{DBF1.png}
\captionsetup{justification=centering}
\caption{}
\label{fig:3a}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{DBF2.png}
\captionsetup{justification=centering}
\caption{}
\label{fig:3b}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{DBF3.png}
\captionsetup{justification=centering}
\caption{}
\label{fig:3c}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{DBF4.png}
\captionsetup{justification=centering}
\caption{}
\label{fig:3d}
\end{subfigure}
\caption{Examples of results from Depth Block Filtering (DBF). Due to the sparse nature of 3D keypoints, even the points occluded from the camera get projected by back ray tracing e.g., keypoints occluded by the fridge in (a) and (b) and points from around the shop in (c) and (d). Depth Block Filtering effectively removes the occluded keypoints and only retains the keypoints in the direct line of sight of the camera. \textbf{Left column:} Snapshot of sections from point clouds with 3D keypoints (red + green) that get projected on a training image. \textbf{Right column:} Image with the projected locations of 3D keypoints. \textbf{Red points:} Keypoints removed by DBF as occluded points. \textbf{Green points:} Keypoints retained by DBF. See Section \ref{technique:dataset:DBF} for details of DBF.}
\label{fig:depth_block_filter}
\end{figure*}
\subsubsection{Creating 2D-3D Correspondences}
We then calculate the Euclidean distances between the remaining projected points $(\hat{x}_{j,n},\hat{y}_{j,n})$ in the set of projections of 3D keypoints $\Theta_n$ and the 2D keypoints $keys2D_n$ on the image. The pairs of 2D and 3D keypoints whose projections are within a specified distance $\alpha$, measured in the units of pixels, are considered as the matching pairs. In the case that there are more than one 2D keypoints within the specified distance of the projection of a 3D keypoint or multiple 3D keypoints' projections are close to a 2D keypoint, only the pair with the smallest distance is considered for the dataset, i.e., we treat the matching of keypoints as a one-to-one correspondence for each image.
Let $\zeta_n$ be the set of the pair of indices, corresponding to the points of $N3$ and $N2_n$, that are within an error threshold $\alpha$:
\begin{equation}\label{e10}
\begin{aligned}
\zeta_n= \bigcup(j,k) \quad \forall \left || (\hat{x_j}-\widetilde{x}_{k,n}),(\hat{y_j}-\widetilde{y}_{k,n}) \right || < \alpha \\ \quad where \quad j=1,2,3,...,N3, \\ k=1,2,3,...,N2
\end{aligned}
\end{equation}
We use $\zeta_n$ to create a matching set of 3D and 2D points by retrieving the keypoints from $keys3D$ and $keys2D_n$ according to the indexes. This process is repeated for each training image and then the resulting sets of matching keypoints are concatenated to create a dataset of matching 3D and 2D points for the whole training set.
To obtain a dataset of corresponding 3D and 2D descriptors for the matching points, we retrieve the corresponding descriptors for the 3D key-points from $desc3D$ and the 2D descriptors from $desc2D$.
It is to be noted that in the resulting dataset, one 3D descriptor can correspond to multiple 2D descriptors as one 3D keypoint can appear in multiple images of the same place taken from different poses.
\subsection{Training}
\label{technique:train}
The generated dataset of corresponding 2D and 3D features is used to train the Descriptor-Matcher. The Descriptor-Matcher is composed of two stages: a coarse matcher and a fine matcher, connected in series (see Figure \ref{fig:test_chart}). The two stages allow the Descriptor-Matcher to maintain both a high precision and computational efficiency at the same time, as explained in Section \ref{technique:test}.
We evaluated several different classifiers for the coarse and fine stages including multi-layer fully connected Neural Networks \cite{khan2018guide}, Support Vector Machines with different kernels, Nearest Neighbour Classifiers with dictionary learning and Discriminant Analysis. However, we empirically found that Classification Tree \citep{treeref} performs best for the coarse stage of the matcher, while for the fine stage only Random Forest classifier \citep{randomforest} produces suitable results. Moreover, these classifiers have better robustness, speed, generalization capability and ability to handle over-fitting for the task at hand.
We treat the problem of matching the descriptors from images and point cloud as a binary classification problem. At the input of the Descriptor-Matcher, we provide 2D and 3D descriptors, concatenated in the form of a $(p\! +\! q)\! \times \! 1$ vector. A positive result at the output indicates that the concatenated descriptors are from the same location in the image and the point cloud, while a negative result shows a mismatch.
To split nodes in the classification tree of coarse matcher, as well as in the trees in Random Forest, we used Gini's Diversity Index to measure the impurity of nodes. The impurity criterion for our case is defined as:
\begin{equation}\label{eq:gdi_binary}
Gini's\ index=\frac{2\times r^+ \times r^-}{(r^++r^-)^2}
\end{equation}
where $r^+$ and $r^-$ are the number of positive and negative samples, respectively, at any given node of a tree. Consequently, an impure node will have a positive value for the splitting criterion, while a node with only positive or only negative samples will have Gini's diversity index equal to zero.
To train the classifiers, the corresponding descriptors of the matching 3D and 2D keypoints in the generated dataset (Section \ref{technique:dataset}) were concatenated to create positive training samples. For the negative samples, we concatenated the 2D and 3D descriptors of the non-matching points in the dataset. We define non-matching points as those pairs whose corresponding locations in the point cloud are
at least a distance $\beta$ apart. However, with the increase in the scale of the point cloud, it is possible that there are regions with similar appearance at different locations in the point cloud. This is particularly a concern for indoor scenarios, where there are comparatively more bald regions and repeated structures. To overcome this problem we also ensured that the Euclidean distance between the descriptors of non-matching points is greater than a threshold $\gamma$. We only used a randomly selected subset of the generated negative samples for training due to the large number of possible one-to-one correspondences between non-matching pairs. We optimized the fine-matcher for maximum precision to minimize the number of false positives in the final matches.
\subsection{Testing}
\label{technique:test}
At test time, we first extract key-points and descriptors from the 2D query image that needs to be localized in the point cloud. Similarly, 3D keypoints and descriptors are extracted from the point cloud, if not already available. Then we concatenate the 2D descriptors of the query image and the 3D descriptors of the point cloud in a one-to-one fashion and use the two stage Descriptor-Matcher to find the matching and non-matching pairs. The pairs of descriptors are first tested by the coarse matcher and only the pairs with positive results for matches are passed to the fine-matcher. As the coarse matcher is composed of a single classification tree, it greatly reduces the number of pairs which need to be tested by the fine matcher, at only a small computational cost, thus greatly decreasing the prediction time of the algorithm.
Then, we retrieve the corresponding keypoints for the descriptor pairs positively matched by the fine-matcher and use them as matching locations from the image and the point cloud for pose estimation.
However, just like any classifier, the output of the Descriptor-Matcher contains some false positives and the matched 3D and 2D points are not completely free of wrong matches. We use two conditions to improve the descriptor matching.
\textbf{First}, we use the probability scores for a positive match (which are in the range $ 0-1$) as the confidence value that a 2D descriptor and a 3D descriptor represent the same locations in the quey image and the point cloud, respectively.
For a positive match, the predicted confidence value must be greater than $0.5$. \textbf{Second}, we also use two-way matching to improve the reliability of the matches. Specifically, based on the highest prediction confidence, we find the closest matching 3D descriptor for each 2D descriptor, and the closest matching 2D descriptor for each 3D descriptor. For a 2D descriptor to be treated as a corresponding pair for a 3D descriptor, the confidence value of the predicted match must be greater than the confidence value of that specific 2D descriptor matching with any other 3D descriptor and vice versa.
To further filter out the false positives we use the MLESAC \citep{mlesacmatlab} algorithm along with the P3P \citep{p3pmatlab} algorithm to find the best set of points for the estimation of the position and orientation of the camera for the query image.
MLESAC is an improved and more generalized algorithm compared to the traditional RANSAC \citep{ransac}. MLESAC generates the tentative solutions in the same manner as RANSAC. However, it produces a better final solution as, in addition to maximizing the number of inliers in the solution, it uses maximum likelihood estimation based on a Gaussian distribution of noise to minimize the re-projection error between the 2D image and 3D points \citep{mlesacmatlab}.
Figure \ref{fig:test_chart} shows a block diagram of the steps involved to localize a 2D query image in a 3D point cloud. The predicted pose of the camera can further be refined with the application of the R1PPnP pose estimation algorithm \citep{zhou2018re} on the points identified as inliers by the MLESAC algorithm to estimate the final position and viewing direction of the query camera in the point cloud. The predicted pose information can be used to extract the section of the point cloud that corresponds to the query image for visualization purposes (Figure \ref{fig:result}).
\section{Experiments and Analysis}\label{experiments}
To evaluate the performance of our technique, we carried out extensive experiments on a number of publicly available datasets. These include Shop Facade \citep{posenet}, Old Hospital \citep{posenet}, Trinity Great Court \citep{kendall2017geometric}, King's College \citep{posenet} and St. Mary Church \citep{posenet}
datasets from the Cambridge Landmarks Database for \textbf{outdoor scenarios}.
To test the localization capability of the proposed technique for \textbf{indoor cases}, we used the Kitchen \citep{valentin2016learning} and Living Room \citep{valentin2016learning} Datasets from the Stanford RGB Localization Database and Baidu Indoor Localization Dataset \citep{idl}.
In our experiments, SIFT key-points and descriptors \citep{lowesift} were extracted from the 2D query images. For the point clouds, we used 3D SIFT key-points \citep{pcl} and 3D RIFT descriptors \citep{rift} to extract keypoints and feature descriptors, respectively. 3D SIFT is a key-point extraction method designed for point clouds, which is inspired from the 2D SIFT key-point extraction algorithm \citep{lowesift} for 2D images. The original 2D SIFT algorithm was adapted for 3D point clouds by substituting the function of intensity of pixels in an image with the principal curvature of points in a point cloud \citep{pcl}.
We set the maximum distance between the image keypoints and the projected 3D keypoints $\alpha=5\ pixels$ for the generation of positive samples. For the formation of one-to-one pairs for negative samples, we set the minimum distance between the 3D keypoints, $\beta$, to 0.5 and the minimum Euclidean distance between the 3D descriptors, $\gamma$ to $0.3$.
To optimize the parameters for coarse and fine matchers, we divided the generated dataset of corresponding 2D and 3D descriptors (see Section \ref{technique:dataset}) into training and validation sets in the ratio of $8\! :\! 2$. Based on the performance on the validation set, we used grid search \citep{gridsearch} to fine tune the classification cost and the maximum number of splits in a tree for both coarse and fine matchers, as well as the number of trees in the Random Forest.
Finally, the complete dataset of matching 2D and 3D features was used to train the Descriptor-Matcher with the optimized parameters.
\subsection{Evaluation Metrics}
We use the positional and the rotational errors in the predicted poses of the query images to evaluate the performance of our technique. As such, the positional error is calculated as the Euclidean distance between the ground truth and the predicted positions of the query camera in the 3D point cloud:
\begin{equation}
\label{pos_err}
position\_error\! =\! \left || (x_g-x_e),(y_g-y_e),(z_g-z_e) \right |\! |
\end{equation}
where $(x_g,y_g,z_g)$ and $(x_e,y_e,z_e)$ are the ground truth and estimated positions of the query image's camera in the 3D point cloud, respectively.
For the estimation of the rotational error, we calculated the minimum angle between the viewing directions of the predicted and the ground truth cameras for the 2D query image. If the predicted rotation matrix is represented by $R_e$ and the ground truth rotation matrix is $R_{g}$, then the rotational error $\phi$ in degrees can be calculated as follows:
\begin{equation}
\label{eq:R_err}
\phi = \frac{\pi}{180} \times cos^{-1}(\frac{trace(R_g \times R_e^{tr})-1}{2})
\end{equation}
\begin{sidewaystable*}[htbp]%
\caption{Median and percentile localization errors for the outdoor datasets: Shop Facade, Old Hospital, St. Mary Church, King's College and Great Court Datasets. P stands for percentile, e.g., P 25\% means the maximum error for 25\% of the data when errors are sorted in the ascending order. m stands for metres}
\centering
\begin{tabular}{|l|l|c|c|c|c|c|}
\specialrule{.15em}{.0em}{.0em}
\textbf{Outdoor Datasets $\downarrow$} & \textbf{Errors $\downarrow$ \textbackslash{} Metrics $\rightarrow$} & \textbf{Median} & \textbf{P 25\%} & \textbf{P 50\%} & \textbf{P 75\%} & \textbf{P 90\%} \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Shop Facade}} & \textbf{Position Error (m)} & 0.0860 m & 0.0307 m & 0.0860 m & 0.4258 m & 3.3890 m \\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.7792\degree & 0.2776\degree & 0.7792\degree & 4.2890\degree & 33.6595\degree \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Old Hospital}} & \textbf{Position Error (m)} & 0.1295 m & 0.0649 m & 0.1295 m & 0.2561 m & 3.2744 m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.2210\degree & 0.1312\degree & 0.2210\degree & 0.6153\degree & 5.6263\degree \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{St. Mary Church}} & \textbf{Position Error (m)} & 0.1479 m & 0.0689 m & 0.1479 m & 0.9518 m & 13.3455 m \\\cline{2-7}
& \textbf{Angle Error (degrees)} &0.4671\degree & 0.1814\degree & 0.4671\degree & 3.5267\degree & 35.3109\degree\\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{King's College}} & \textbf{Position Error (m)} & 0.0877 m & 0.0533 m & 0.0877 m & 0.1332 m & 0.2108 m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.1476\degree & 0.0880\degree & 0.1476\degree & 0.2613\degree & 0.4694\degree \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Great Court}} & \textbf{Position Error (m)} & 0.5098 m & 0.2520 m & 0.5098 m & 1.5401 m & 14.7130
m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.3526\degree & 0.1489\degree & 0.3526\degree & 1.3117\degree & 12.8685\degree \\\specialrule{.15em}{.0em}{.0em}
\end{tabular}
\label{tab:shop}
\label{tab:hospital}
\label{tab:church}
\label{tab:king}
\label{tab:court}
\label{tab:street}
\end{sidewaystable*}
\subsection{Outdoor Datasets}
We used the datasets from the Cambridge Landmarks Database \citep{posenet,kendall2017geometric} to test the localization performance of the proposed method in the outdoor scenarios.
The images in these datasets were captured at different times under different lighting and weather conditions and contain a lot of urban clutter which increases the challenges for precise localization of camera poses.
As the Cambridge Landmarks Database does not contain dense point clouds, we used the COLMAP's Multi View Stereo Pipeline \citep{colmapmvs} to generate dense point clouds for these datasets.
Table \ref{tab:shop} reports the median errors for the estimated positions and orientations of the cameras for the query images for the outdoor datasets. We also report the percentile errors for the 25\%, 50\%, 75\% and 90\% of the query images in these datasets for the estimated camera poses.
\subsubsection{Shop Facade Dataset}
The Shop Facade Dataset \citep{posenet} from the Cambridge Landmarks Database is composed of the images of the intersection of two streets in Cambridge. The images mainly focus on the shops at the intersection. It covers an area of more than $900\ m^2$. It contains a total of 334 images with 103 images in the query set. We used the standard train-test split as defined by the authors of the dataset. Our proposed technique was able to localize all the images in the query set.
\subsubsection{Old Hospital Dataset}
The Old Hospital dataset \citep{posenet} contains 1077 images. There are 182 query images in the train-test split defined by the dataset's authors. The dataset covers an area of $2000\ m^2$. This dataset suffers particularly from the challenges of repetitive patterns, as well as high symmetry due to similar constructions on both sides of the centre of the building. The localization results are shown in Table \ref{tab:hospital}.
\subsubsection{St. Mary Church Dataset}
The St. Mary Church Dataset \citep{posenet} is composed of 2017 images of the Great St. Mary Church in Cambridge. It encompasses an area of $4800\ m^2$. Many of the images contain occlusions caused by pedestrians and other urban clutter due to which it becomes challenging to localize images in the 3D point cloud. We used the query set defined by the authors of the dataset which contains $530$ images for the evaluation of our technique, while the remaining 2D images were used to train the Descriptor-Matcher. Our technique successfully localized all the images in the query set.
\subsubsection{King's College Dataset}
This dataset covers the location and building of the King's College, which is one of the constituent colleges of the University of Cambridge \citep{posenet}. It consists of 1563 images captured with the camera of a smart phone. King's College covers an area of more than $5600\ m^2$. We used the train-query split of images defined by the authors of the dataset. There are 343 images in the query set. Table \ref{tab:king} shows the results of our technique on the dataset.
\subsubsection{Trinity Great Court Dataset}
Trinity Great Court is the main court (courtyard) of Trinity College, Cambridge. It is one of the largest enclosed courtyards in Europe with an area of $8000\ m^2$. The train and test sets consist of $1532$ and $760$ number of images, respectively \citep{kendall2017geometric}. The proposed technique successfully localized all the images in the dataset.
\subsection{Indoor Datasets}
Most of the pose estimation techniques in the literature have either been evaluated only for outdoor scenarios or for very small indoor scenes (e.g. tested for a maximum volume of $6m^3$ on Seven Scenes Database \citep{sevenscenes} ).
Indoor localization of images is a more challenging problem compared to the outdoor settings \citep{idl}. Due to similar items (e.g., furniture) and the same construction patterns (e.g., cubicle shape of rooms, similar patterns on floor or ceiling, stairs), it is possible that different regions of the building look very similar, which greatly increases the possibility of wrong matches and incorrect camera pose estimation. On the other hand, indoor localization is a more useful application of image localization as many of the traditional localization methods, such as GPS based localization, do not work properly in indoor settings. To evaluate the performance of our technique on practical indoor localization scenarios, we used the Kitchen and Living Room Datasets from the Stanford RGB Localization Database \citep{valentin2016learning} and Baidu Indoor Localization dataset \citep{idl}.
Table \ref{tab:idl} shows the median positional and rotational errors along with the percentile errors for the intervals of 25\%, 50\%, 75\%, and 90\% data on the indoor datasets.
\subsubsection{Kitchen Dataset}
This dataset contains RGB images and a 3D model of a Kitchen in an apartment. It is part of the Stanford RGB Localization Database \citep{valentin2016learning}. It covers a total volume of $33m^3$. The 2D images and the 3D point cloud were created with a Structure.io\footnote{https://structure.io/structure-sensor} 3D sensor coupled with an iPad. Both sensors were calibrated and temporally synced. We randomly selected $20\%$ of the images for the query set while used the remaining images for training the Descriptor-Matcher. Our technique successfully localized all the query images in the 3D model with high accuracy. We were able to estimate the 6-DOF poses of more than 90\% of the images with errors less than 4 cm and withing a degree, as shown in Table \ref{tab:idl}.
\begin{sidewaystable*}[htbp]
\caption{Median and percentile localization errors for the Kitchen, Living Room and Baidu Indoor Localization Datasets. P stands for percentile, e.g., P 25\% means the maximum error for 25\% of the data when errors are sorted in the ascending order. m stands for metres}
\centering
\begin{tabular}{|l|l|c|c|c|c|c|}
\specialrule{.15em}{.0em}{.0em}
\textbf{Indoor Datasets $\downarrow$} & \textbf{Errors $\downarrow$ \textbackslash{} Metrics $\rightarrow$} & \textbf{Median} & \textbf{P 25\%} & \textbf{P 50\%} & \textbf{P 75\%} & \textbf{P 90\%} \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Kitchen}} & \textbf{Position Error (m)} & 0.0117 m & 0.0066 m & 0.0117 m & 0.0223 m & 0.0395 m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.2128\degree & 0.1448\degree & 0.2128\degree & 0.3845\degree & 0.6399\degree \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Living Room}} & \textbf{Position Error (m)} & 0.0101 m & 0.0059 m & 0.0101 m & 0.0180 m & 0.0470 m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 0.3254\degree & 0.2011\degree & 0.3254\degree & 0.4877\degree & 1.6218\degree \\\specialrule{.15em}{.0em}{.0em}
\multirow{2}[0]{*}{\textbf{Baidu Indoor Localization}} & \textbf{Position Error (m)} & 0.6930 m & 0.2020 m & 0.6930 m & 10.44 m & 24.55 m\\\cline{2-7}
& \textbf{Angle Error (degrees)} & 2.30\degree & 0.6105\degree & 2.30\degree & 13.72\degree & 22.05\degree \\\specialrule{.15em}{.0em}{.0em}
\end{tabular}
\label{tab:idl}
\end{sidewaystable*}
\subsubsection{Living Room Dataset}
The Living Room Dataset is part of the Stanford RGB Localization Database \citep{valentin2016learning} and comprises the 3D model and 2D images of a living area in an apartment with a total volume of $30m^3$. This dataset was also captured with a combination of Structure.io sensor and iPad cameras. For the quantitative evaluation on this dataset, $20\%$ of the images were randomly held out for query set, while the remaining images were used for the training set. We were able to localize all the images in the query set with more than 90\% localizations within a 5 cm error (Table \ref{tab:idl}).
\begin{figure*}[htbp]
\centering
\begin{subfigure}[htbp]{1\textwidth}
\centering
\includegraphics[width=\textwidth]{PICIDLCropped.png}
\caption{Top view}
\label{fig:idl1}
\end{subfigure}
\hfill
\begin{subfigure}[htbp]{1\textwidth}
\centering
\includegraphics[width=\textwidth]{PICIDL2Cropped.png}
\caption{Side view}
\label{fig:idl2}
\end{subfigure}
\caption{Different views of the point cloud of the shopping mall from Baidu Indoor Localization Dataset \citep{idl}. The large scale and repetitive structures in the dataset make it extremely challenging to localize 2D images in the point cloud compared to other indoor datasets. }
\label{fig:idl}
\end{figure*}
\subsubsection{Baidu Indoor Localization Dataset}
The Baidu Indoor Localization Dataset \citep{idl} is composed of a 3D point cloud of a multi-storey shopping mall created with the scans from a LIDAR scanner and a database of images with ground truth information for camera positions and orientations. The 3D point cloud contains more than 67 million points. Figure \ref{fig:idl} shows different views of the point cloud of the mall. The images in the dataset are captured with different cameras and smart phones and at different times which increases the complexity of the dataset. Moreover, many of the images contain occlusions due to the persons shopping in the mall.
We randomly selected $10\%$ of the images for the query set, while the remaining images were used for the training set. Our technique successfully localized 78.2\% of the images in the query set. Despite the challenges of the dataset, we achieved a median pose estimation error of 0.69 m and 2.3 degrees as shown in Table \ref{tab:idl}.
\begin{sidewaystable*}[htbp]
\caption{Median errors for position and orientation estimation of our technique compared to other approaches on Cambridge Landmarks outdoor datasets. Pos stands for median positional error in meters and Ang stands for rotational error in degrees. NA: Results not available. Best results are shown in \textbf{bold} and second best results are \underline{underlined}.}
\centering
\begin{tabular}{|A|R|R|R|R|R|R|R|R|R|R|R|R|R|R|}
\hline
\textbf{Methods$\rightarrow$} & \multicolumn{2}{V|}{\textbf{PoseNet \citep{posenet} ICCV'15}} & \multicolumn{2}{V|}{\textbf{Geom. Loss Net \citep{kendall2017geometric} CVPR'17}} & \multicolumn{2}{V|}{\textbf{VLocNet \citep{valada2018deep} ICRA'18}} & \multicolumn{2}{V|}{\textbf{DSAC \citep{brachmann2017dsac} CVPR'17}} & \multicolumn{2}{V|}{\textbf{Active Search \citep{sattler2017efficient} TPAMI'17}} & \multicolumn{2}{V|}{\textbf{DSAC++ \citep{brachmann2018learning} CVPR'18}} &\multicolumn{2}{V|}{\textbf{Ours}} \\
\hline
\textbf{Methods' Type$\rightarrow$} & \multicolumn{2}{V|}{\textbf{Network-based}} & \multicolumn{2}{V|}{\textbf{Network-based}} & \multicolumn{2}{V|}{\textbf{Network-based}} & \multicolumn{2}{V|}{\textbf{Network + RANSAC}} & \multicolumn{2}{V|}{\textbf{SfM-based}} & \multicolumn{2}{V|}{\textbf{Network + RANSAC}} &\multicolumn{2}{V|}{\textbf{2D to 3D Descriptors Matching}} \\
\hline
\textbf{Datasets$\downarrow$} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang} & \textbf{Pos (m)} & \textbf{Ang}\\
\hline
\textbf{Shop Facade} & 1.46 & 4.04\degree & 0.88 & 3.78\degree & 0.593 & 3.529\degree & \underline{0.09} & \textbf{0.4\degree} & 0.12 & \textbf{0.4\degree} & \underline{0.09} & \textbf{0.4\degree} & \textbf{0.086} & \underline{0.779\degree} \\
\hline
\textbf{Old Hospital} & 2.31 & 2.69\degree & 3.2 & 3.29\degree & 1.075 & 2.411\degree & 0.33 & 0.6\degree & 0.44 & 1\degree & \underline{0.24} & \underline{0.5\degree} & \textbf{0.129} & \textbf{0.221\degree} \\
\hline
\textbf{King's College} & 1.92 & 2.70\degree & 0.88 & 1.04\degree & 0.836 & 1.419\degree & 0.30 & 0.5\degree & 0.42 & 0.6\degree & \underline{0.23} & \underline{0.4\degree} & \textbf{0.087} & \textbf{0.147\degree} \\
\hline
\textbf{Great Court } & NA & NA & 6.83 & 3.47\degree & NA & NA & 2.80 & 1.5\degree & NA & NA & \underline{0.66} & \underline{0.4\degree} & \textbf{0.509} & \textbf{0.352\degree} \\
\hline
\textbf{St. Mary Church} & 2.65 & 4.24\degree & 1.57 & 3.32\degree & 0.631 & 3.906\degree & 0.55 & 1.6\degree & \underline{0.19} & \underline{0.5\degree} & 0.20 & 0.7\degree & \textbf{0.147} & \textbf{0.467\degree} \\
\hline
\end{tabular}
\label{tab:comparison}
\end{sidewaystable*}
\subsection{Comparison with Other Approaches}
The proposed technique for image localization in point clouds is based on a novel concept of directly matching the descriptors extracted from images and point clouds. The SfM based techniques are based on the matching of 2D features and require the point cloud to be generated from SfM pipeline. On the other hand, our technique can work with point clouds created with any 3D scanner or generated with any method and has no dependency on any information from SfM.
The network based regression methods train directly to regress the poses of the images. However, this causes the networks to over-fit on the ground truth poses. Therefore at test time, they only produce good results for the images with poses close to the training ones \citep{sattler2019understanding}. In our technique, training of the Descriptor-Matcher is carried out on the feature descriptors with no information of the poses or the location of the points. This ensures that the Descriptor-Matcher does not over-fit on the training poses, rather it learns to find a mapping between the 2D and 3D descriptors. The final pose estimation is based on the principles of geometry which produce better pose estimates compared to end-to-end trained methods \citep{sattler2019understanding}. Therefore, the proposed method benefits from the non-reliance on SfM pipeline like the network based methods, as well as the ability to use geometry based pose estimation similar to the SfM based methods.
For quantitative analysis, we provide a comparison of our results with the
state-of-the-art methods
for camera pose localization on the common datasets between the various approaches. We compare the median values of errors in the estimation of camera position and rotation on the outdoor datasets of Cambridge Landmarks Database \citep{posenet}, as mostly only the median errors are reported by other methods. The results of the compared methods are not available for Baidu Indoor Localization Dataset or Stanford RGB Localization Database.
Our proposed method achieved competitive or superior localization accuracy to the state-of-the-art methods for 6-DOF pose estimation on all the datasets.
Table \ref{tab:comparison} shows the median values for position and rotational errors of our technique compared to other prominent methods.
\section{Conclusion}\label{conclusion}
This paper proposed a novel method to directly match descriptors from 2D images with those from 3D point clouds, where the descriptors are extracted using 2D and 3D feature extraction techniques, respectively.
We have shown that direct matching of 2D and 3D feature descriptors is an unconstrained and reliable method for 6-DOF pose estimation in point clouds.
Our approach results in a pipeline which is much simpler compared to SfM or network based methods.
Extensive quantitative evaluation has demonstrated that the proposed method achieved competitive performance with the state-of-the-art methods in the field.
Moreover, unlike SfM based techniques, the proposed method can be used with point clouds generated with any type of 3D scanners and can work in both indoor and outdoor scenarios.
\section*{Acknowledgements}
This work was supported by the SIRF scholarship from the University of Western Australia (UWA) and by the Australian Research Council under Grant DP150100294.
\balance
| 983cc979cfd8d73df087f4fba0f116b2f63f5744 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Molecular communication (MC) uses molecules as carriers of information. Molecular communication
is envisioned to be applicable in a wide range of engineering and medical applications, especially as a means to realize communications among engineered biological nanomachines
(see \cite{pierobon2010physical, nakano2012molecular, guo2015molecular, survey:medicine}). In diffusion-based MC, molecules are released into the medium by molecular transmitters. Information is coded by the transmitter in the type, number, or release time of the molecules. The released molecules randomly diffuse in the medium, with some reaching the molecular receivers. The receiver decodes the message sent by the transmitter based on the number of sensed molecules (see \cite{kuran2011modulation, arjmandi2013diffusion, srinivas2012molecular} for some examples of modulation schemes).
Broadly speaking, two main types of diffusion-based MC exists: microscale MC and macroscale MC. In microscale MC, a small number of molecules are released into the medium by transmitters. The movement of each molecule is random and follows Brownian motion. On the other hand, in macroscale MC, a large number of molecules are released into the medium (the number of released molecules is in the order of moles). Macroscale MC is of interest as an alternative to electromagnetic wave-based systems in certain media \cite{guo2015molecular, farsad2016comprehensive}.
In macroscale MC, instead of focusing on the location of individual molecules, by the law of large numbers, the \emph{average concentration} of molecules is described by Fick's law of diffusion, which is a \emph{deterministic} differential equation. Everything is described by deterministic equations until this point. However, various sources of noise or imperfection could be considered for molecular transmitters and receivers \cite{farsad2016comprehensive, gohari2016information}. In particular,
the literature on MC usually associates a measurement noise to the receivers. The variance of this measurement noise depends on the concentration level of molecules at the receiver \cite{einolghozati2012collective}. In particular, a commonly used measurement noise is the ``Poisson particle counting noise" where the measurement follows a Poisson distribution whose mean is proportional to the concentration of molecules at the time of measurement.
\textbf{Motivation:} One of the unique features of MC with no parallel in classical wireless communication is \emph{chemical reactions}: different types of molecules can react with each other and form new types of molecules in the medium. This feature of MC poses a challenge in macroscale MC since equations describing chemical reaction with diffusing particles are \emph{nonlinear} partial differential equations with no closed-form solutions\cite{noclosedform:1,noclosedform:2,noclosedform:3}.
In other words, there is no analytical closed-form formula that describes the variations in the concentration of molecules over time and space for given boundary conditions, inputs to the medium, and system parameters such as the diffusion and reaction rate constants. These equations are commonly solved by numerical methods. While numerical techniques (such as the finite difference method and the finite element method) can provide accurate solutions, they require extensive computational resources. More importantly, these solutions are not generalizable: assume that we have solved the reaction-diffusion equations for one particular release pattern of molecules by the transmitters into the medium. Next, if we make some changes to the release pattern of molecules, we should start all over again and solve the reaction-diffusion equations again for the new release pattern. Numerical methods provide little insight into how the solution would change if the input or parameters of the medium change. This has a major drawback for the design of coding and modulation schemes, where we wish to optimize and find the best possible release pattern. A tractable (even if approximate) model for the solution of the reaction-diffusion systems is required to design codes or modulation schemes. Only with such a model can one formally the optimality of certain waveforms for a given communication problem.
\textbf{Our contribution:}
We apply the perturbation theory to design modulation schemes among manufactured devices in macroscale MC systems involving chemical reactions.
Perturbation theory is a set of
mathematical methods for finding approximate solutions. It was originally proposed to calculate the motions of planets in the solar system, which is an intractable problem (as the so-called three-body problem shows). The theory has been broadly applied in different scientific areas such as physics and mechanics for solving algebraic equations with no analytical solutions.
While the perturbation theory has been used in the study of differential equations in general and
in the context of fiber optics communication systems \cite{pert-opt-cite}, to the best of our knowledge it has not been hitherto utilized in molecular communication.\footnote{ Note that \cite{singularpert} uses the idea of \emph{
singular perturbation} which should not be confused with the \emph{perturbation theory}.} Just as non-linear functions can be locally approximated by the first terms of their Taylor series, perturbation theory shows that non-linear reaction-diffusion equations can be also approximated by their lower order terms. Our main contribution is a proposal to design coding and modulation schemes based on the lower order terms, which admit closed-form solutions. In other words, we provide an analytically tractable (approximate) model for studying chemical reactions in molecular communication. The accuracy of the model can be increasingly improved by including more terms from the Taylor series.
We concretely illustrate the perturbation method by considering the following reaction-diffusion system in which molecules of type $\mathtt{A}$ and $\mathtt{B}$ react with each other in the medium and produce molecules of type $\mathtt{C}$:
\begin{align}\label{ex1:reactionN1}
\mathtt{A} + \mathtt{B}
\underset{\lambda_2}{\stackrel{\lambda_1}{\rightleftharpoons}} \mathtt{C}.
\end{align}
Forward and backward reaction rates $\lambda_1$ and $\lambda_2$ represent the speed at which the chemical reaction takes place in the forward and backward directions. This reaction-diffusion system is chosen because it is one of the simplest chemical reactions with no closed-form solution. This basic reaction appears in many contexts, e.g. $\text{base}+\text{acid}\rightarrow \text{salt}$, or $NH_3+{H_2O}\rightarrow {NH_4OH}$, $\text{ADP}+\text{P}\rightarrow \text{ATP}$, or
$
\text{lipoprotein} + \text{PRAP1}
\underset{\lambda_2}{\stackrel{\lambda_1}{\rightleftharpoons}}\text{PRAP1-lipoprotein}$
in biology.
It is used in \cite{farahnak2018medium:j:16} to design a communication strategy between engineered devices. It may be also used in chemical computing, \emph{e.g.} to implement a binary AND gate (as the product molecules are produced only when both reactants are present).
The application considered in this paper is a communication system involving three engineered devices, two transmitters, and one receiver: the two transmitters release molecules of type $\mathtt A$ and $\mathtt B$ respectively, and the receiver senses the product molecules of type $\mathtt C$.
The goal is to compute the best transmission waveform for each transmitter so that the receiver has minimum error probability to decode the messages of both transmitters (a multiple access channel setting).
In this setting, we find the optimal waveforms that minimize the error probability in the low rate reaction regime (or in the small transmission time interval regime). In particular, we show that instead of using a continuous release waveform, it is optimal for the transmitters to release molecules at
two time instances.
To illustrate the broad applicability of our technique, we also discuss other reaction-diffusion systems (different from \eqref{ex1:reactionN1}) in the Appendix \ref{ex2,3}. These examples cover scenarios where the chemical reaction is either facilitating or impeding effective communication.
\textbf{Related works:} The math literature on nonlinear partial differential equations (PDEs) studies different aspects of these equations, such as the existence and uniqueness of the solution, semianalytical solutions, numerical solutions, etc. Since most PDEs do not have closed-form solutions,
showing the existence and in some cases, uniqueness of solution are important topics in the theory of PDEs. Unlike ordinary differential equations (ODE) which have a straightforward theorem for the existence and uniqueness of a solution, there is no general method for PDEs to show the existence and uniqueness of the solution. For instance, the existence and smoothness of solutions for the Navier-Stokes equations that describe the motion of a fluid in space are fundamental open problems in physics.
Semianalytical techniques are series expansion methods in which the solution is expressed as an infinite series of explicit functions. Semianalytical techniques include the Adomian decomposition method, Lyapunov artificial small parameter method, homotopy perturbation method, and perturbation methods \cite{liao2003beyond,liao2012homotopy}.
There are many numerical methods for solving PDEs which are classified in terms of complexity and stability of the solution. The finite element method, finite difference method, spectral finite element method, meshfree finite element method, discontinuous Galerkin finite element method are some examples of numerical techniques for solving PDEs \cite{numericalmethode:1,noclosedform:2}. For example, in the finite difference method, we partition the domain using a mesh and
approximate derivatives with finite differences computed over the mesh. One of the challenges of using this method is determining the appropriate mesh size to guarantee the stability of the solution.\footnote{ Our own experiment with this method indicates that choosing the appropriate mesh size is particularly challenging when we have a reaction-diffusion system involving molecules that have very different diffusion constants (especially when one diffusion constant is more than ten times another diffusion constant).}
There are some prior works in the literature on chemical physics that utilize the perturbation theory. For instance, the average survival time of diffusion-influenced chemical reactions is studied in \cite{pagitsas1992perturbation:e:1:1}. In \cite{kryvohuz2014nonlinear:e:1:2} some classes of chemical kinetics (describe by ODEs) have been solved by the perturbation theory. However, the setup and approach in these problems differ from the one encountered in MC.
Due to their evident importance to MC, chemical reactions have been the subject of various studies. A molecular communication system involves transmitters, receivers, and the molecular media; chemical reactions may occur in each of these individual building blocks. Transmitters could use chemical reactions to produce signaling molecules \cite{bi2019chemical:j:9}. This is a chemical reaction inside the transmitter unit. On the receiver side, one might have ligand (or other types of) receptors on the surface of the receiver which react with incoming molecules. These receptors have been the subject of various studies, e.g., see \cite{chou2015impact:e:2:2:3,Ligand0, Ligand1, Ligand2, Ligand3,kuscu2018modeling:e:2:3:2, ahmadzadeh2016comprehensive:e:2:3:1,chou2014molecular:e:2:2:4}.
Finally, chemical reactions have also been considered in the communication medium. This aspect of chemical reactions is of interest in this work.
It is pointed out that chemical reactions could be used to suppress signaling molecules and thereby reduce
intersymbol interference (ISI) and the signal-dependent measurement noise
\cite{noclosedform:1,cho2017effective:j:12}, or to amplify the signal \cite{nakano2011repeater:F:22}. Complicated propagation patterns can be obtained by using chemical reactions\cite{nakano2015molecular:F:19}, and negative molecular signals could be implemented using chemical reactions \cite{farsad2016molecular:F:16,wang2014transmit:F:20,mosayebi2017type:F:21}. Notably, chemical reactions are shown to be beneficial for coding and
modulation design \cite{farsad2016molecular:F:16,farahnak2018medium:j:16,noclosedform:2,nakano2015molecular:F:19,wang2014transmit:F:20,mosayebi2017type:F:21}. Since solving the reaction-diffusion equations is intractable, these works either use numerical simulations to back up the presented ideas or else simplify the reaction-diffusion process via idealized assumptions about chemical reactions. Authors in \cite{cao2019chemical:j:ro} provide an iterative numerical algorithm for solving reaction-diffusion equations.
In \cite{farsad2017novel:e:2:2:2}, a neural network is used as the decoder in a medium with chemical reactions.
This paper is organized as follows: in Section \ref{sec::generalperturbation}, we illustrate the perturbation method for solving reaction-diffusion equations through an example.
In Section \ref{Sec:Mod} we design modulation schemes for our example in Section \ref{sec::generalperturbation}. In Section \ref{sec:Val}, we validate our model through numerical simulation. Finally concluding remarks and future work are given in Section \ref{Conclusion and Future Work}.
\textbf{Notations and Remarks:}
For functions $u_i(x,t)$ and $v_i(x,t)$, we define index convolution as follows:
\begin{align}\label{dicreteconv}
(u_{0:i-1}*^d v_{0:i-1})(x,t)=\sum_{j=0}^{i-1} u_{j}(x,t)v_{i-1-j}(x,t).
\end{align}
For two functions $u(x,t)$ and $v(x,t)$, convolution in both time and space is denoted by $**$ and defined as follows:
\begin{align}(u**v)(x,t)=\int_{x'}\int_{t'}u(x,t)v(x-x',t-t')dx'dt'.\label{eqnconvts}
\end{align}
The concentration of molecules of type $\mathtt{A}$ at location $x$ and time $t$ is denoted by $[\mathtt{A}](x,t)$.
Similarly, we use $[\mathtt{B}](x,t)$ and $[\mathtt{C}](x,t)$ to denote the concentration of molecules of types $\mathtt{B}$ and $\mathtt{C}$ respectively. We write $[\mathtt{A}](x,t)$ as a series in the perturbation method and use
$[\mathtt{A}]_{i}(x,t)$ as the coefficient for the $i$th term in the expansion.
For simplicity, we sometimes show a function without its arguments, e.g. we write $[\mathtt{A}]_{i}$ or $[\mathtt{A}]$ instead of $[\mathtt{A}]_{i}(x,t)$ or $[\mathtt{A}](x,t)$. Finally, all chemical reactions are considered in the time interval $[0,T]$ for some $T>0$.
\section{Solving Reaction-Diffusion Equations via the Perturbation Method}\label{sec::generalperturbation}
The perturbation method provides an explicit analytical solution in the form of an infinite series of functions. One can approximate the solution by taking a finite number of terms from this series.
In this paper, we use the perturbation method to obtain an analytical model for the reaction-diffusion equations, which can be used for molecular communication.
Our working example in the paper is as follows: consider two molecular transmitters $\mathsf \mathsf T_{\mathtt {A}}$ and $\mathsf \mathsf T_{\mathtt {B}}$, which release molecules $\mathtt{A}$ and $\mathtt{B}$, respectively. Molecules of type $\mathtt{A}$ and $\mathtt{B}$ react with each other in the medium and produce molecules of type $\mathtt{C}$ as follows:
\begin{align}\label{ex1:reaction}
\mathtt{A} + \mathtt{B}
\underset{\lambda_2}{\stackrel{\lambda_1}{\rightleftharpoons}} \mathtt{C}
\end{align}
We assume a molecular receiver $\mathsf \mathsf R_{\mathtt {C}}$, which measures the concentration of molecules of type $\mathtt{C}$ at its location in order to detect the messages of the two transmitters.
For simplicity, assume that the transmitters and receiver are placed on a one-dimensional line, with $\mathsf T_{\mathtt {A}}$, $\mathsf T_{\mathtt {B}}$, and $\mathsf R_{\mathtt {C}}$ being located at $x=0$, $x=d_\mathtt{B}$, and $x=d_R$ respectively. This assumption is just for the simplicity of exposition; the problem could be solved in two or three dimensions and we also provide simulation results in dimensions two and three in Section \ref{sec:Val}. We assume that the transmitters and receivers are small in size and view them as points on the real line; the transmitter and receiver are considered to be transparent, not impeding the diffusion of the released molecules (assumption of point sources is adopted for simplicity and is common in the MC literature, e.g., \cite{farahnak2018medium:j:16,noclosedform:2,farsad2016molecular:F:16}). This is depicted in Fig. \ref{fig:ex1}. Further details of the communication model are described in Section \ref{Sec:Mod} where we design a modulation scheme. In the rest of this section, we are merely interested in solving the reaction-diffusion equation describing reaction \eqref{ex1:reaction}.
\begin{figure}
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.25]{Slide11.png}
\caption{The system model}
\label{fig:ex1}
\vspace{-2.5em}
\end{figure}
Let $\gamma=\lambda_2/\lambda_1$ be the ratio of the backward and forward reaction rates, and set $\lambda=\lambda_1,\lambda_2=\gamma\lambda$ for some $\lambda\geq 0$.
The following equation describes the system dynamic:
\begin{align}
&\frac{\partial[\mathtt{A}]}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]}{{\partial x}^{2}}-\lambda~[\mathtt{A}][\mathtt{B}]+\gamma \lambda [\mathtt{C}]+f_\mathtt{A}(x,t),\label{A:ex1eq}\\
& \frac{\partial[\mathtt{B}]}{\partial t}=D_\mathtt{B}~\frac{\partial^{2}[\mathtt{B}]}{{\partial x}^{2}}-\lambda~[\mathtt{A}][\mathtt{B}]+\gamma \lambda [\mathtt{C}]+f_\mathtt{B}(x,t),\label{B:ex1eq}\\
& \frac{\partial[\mathtt{C}]}{\partial t}=D_\mathtt{C}~\frac{\partial^{2}[\mathtt{C}]}{{\partial x}^{2}}+\lambda~[\mathtt{A}][\mathtt{B}]-\gamma \lambda [\mathtt{C}]\label{C:ex1eq},
\end{align}
where $f_\mathtt{A}(x,t),f_\mathtt{B}(x,t)$ are the released concentration of molecules (called the transmission waveforms), and $D_\mathtt{A}, D_\mathtt{B}, D_\mathtt{C}$ are the diffusion constants. The initial and boundary conditions are set to zero. We are interested in the solution for time $t\in[0,T]$.
For the special case of $\lambda=0$, \emph{i.e.,} when there is no reaction, the above system of equations is linear and tractable.
Consider a solution of these equations in the form,
$[\mathtt{A}](x,t)=\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i}(x,t)$,
$[\mathtt{B}](x,t)=\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i}(x,t)$ and $[\mathtt{C}](x,t)=\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{C}]_{i}(x,t)$ for some Taylor series coefficients
$[\mathtt{A}]_{i}(x,t)$, $[\mathtt{B}]_{i}(x,t)$ and $[\mathtt{C}]_{i}(x,t)$.
By substituting these expressions in \eqref{A:ex1eq}, \eqref{B:ex1eq} and \eqref{C:ex1eq}, we obtain
\begin{align}\label{New1h}
&\frac{\partial}{\partial t}(\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i})
=D_\mathtt{A}~\frac{\partial^{2}}{{\partial x}^{2}}(\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{A}]_{i})-\lambda (\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i})(\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{B}]_{i})+\gamma\lambda \sum_{i=0}^{\infty}\lambda^{i}[\mathtt{C}]_{i} +f_\mathtt{A}(x,t),
\end{align}
\begin{align}\label{New2h:b}
&\frac{\partial}{\partial t}(\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i})
=D_\mathtt{B}~\frac{\partial^{2}}{{\partial x}^{2}}(\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{B}]_{i})-\lambda (\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i})(\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i})+\gamma\lambda \sum_{i=0}^{\infty}\lambda^{i}[\mathtt{C}]_{i}+
f_\mathtt{B}(x,t),
\end{align}
\begin{align}\label{New3h}
&\frac{\partial}{\partial t}(\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{C}]_{i})
=D_\mathtt{C}~\frac{\partial^{2}}{{\partial x}^{2}}(\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{C}]_{i})+\lambda (\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i})(\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i})-\gamma\lambda \sum_{i=0}^{\infty}\lambda^{i}[\mathtt{C}]_{i}
.
\end{align}
Equations \eqref{New1h}, \eqref{New2h:b}, and \eqref{New3h} could be viewed as power series in $\lambda$ for fixed values of $x$ and $t$. Matching the coefficients of $\lambda^i$ on both sides of the equation, we obtain the following:
\begin{align}\label{zerosol}
\frac{\partial [\mathtt{A}]_{0}}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]_{0}}{{\partial x}^{2}}+f_\mathtt{A}(x,t),~~
\frac{\partial [\mathtt{B}]_{0}}{\partial t}=D_\mathtt{B}~\frac{\partial^{2} [\mathtt{B}]_{0}}{{\partial x}^{2}}+f_\mathtt{B}(x,t),~~
\frac{\partial [\mathtt{C}]_{0}}{\partial t}=D_\mathtt{C}~\frac{\partial^{2} [\mathtt{C}]_{0}}{{\partial x}^{2}}.
\end{align}
For $i\geq 1$ we obtain
\label{othersolA}
\begin{align}
&\frac{\partial[\mathtt{A}]_{i}}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1},\label{othersolA} \\
&\frac{\partial[\mathtt{B}]_{i}}{\partial t}=D_\mathtt{B}~\frac{\partial^{2}[\mathtt{B}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1},\label{othersolB} \\
&\frac{\partial[\mathtt{C}]_{i}}{\partial t}=D_\mathtt{C}~\frac{\partial^{2}[\mathtt{B}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}-\gamma[\mathtt{C}]_{i-1},\label{othersolC}
\end{align}
where we used the index convolution defined in \eqref{dicreteconv}.
The functions $[\mathtt{A}]_{i}(x,t)$, $[\mathtt{B}]_{i}(x,t)$ and $[\mathtt{C}]_{i}(x,t)$ could be find recursively by first computing $[\mathtt{A}]_{0}(x,t)$, $[\mathtt{B}]_{0}(x,t)$ and $[\mathtt{C}]_{0}(x,t)$ from \eqref{zerosol}, and then using \eqref{othersolA}-\eqref{othersolC} to compute $[\mathtt{A}]_{i}(x,t)$, $[\mathtt{B}]_{i}(x,t)$ and $[\mathtt{C}]_{i}(x,t)$ from $[\mathtt{A}]_{j}(x,t)$, $[\mathtt{B}]_{j}(x,t)$ and $[\mathtt{C}]_{j}(x,t)$ for $j\leq i-1$.
\subsection{Limitation on the simulation time interval and a remedy}
The perturbation method provides recursive equations to find the functions $[\mathtt{A}]_{i}(x,t)$ and $[\mathtt{B}]_{i}(x,t)$.
However, one still needs to show that the power series $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i}(x,t)$ and $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i}(x,t)$ are convergent to functions that satisfy the original reaction-diffusion differential equation. Perturbation theory does not provide a general recipe for showing this convergence, and it should be done on a
case-by-case basis. We show in Appendix \ref{AppA0 } that the series is convergent as long as $\lambda\leq \mathcal{O}(1/T)$, \emph{i.e.,} the radius of convergence of the series is proportional with the inverse of
the total time period. In other words, we need $\lambda$ to be sufficiently small (or for $T$ to be sufficiently small). Provided that $T$ is sufficiently small, Appendix \ref{AppA0 } shows that the solution given by the perturbation method is equal to the true solution (satisfies the reaction-diffusion equations).
Next, we propose an approach to tackle the above shortcoming if we wish to compute the solution for larger values of $T$: we can divide the time interval $T$ into $k$ subintervals of size $T/k$. Since the radius of convergence is of order $\mathcal{O}(\frac 1T)$, reducing the size of the interval by a factor of $k$ increases the radius of convergence by a multiplicative factor of $k$. Thus, we can first compute the solution in the interval $[0,T/k]$ and use the value at $T/k$ to compute the solution in the interval $[T/k,2T/k]$, etc.
\subsection{Solving equations \eqref{zerosol}-\eqref{othersolC}}
In order to solve \eqref{zerosol}, let $\phi_\mathtt{A}(x,t)$ be the solution of the following equation with diffusion coefficient $D_\mathtt{A}$:
\begin{align}\label{greenA}
\frac{\partial \phi_\mathtt{A}}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}\phi_\mathtt{A}}{{\partial x}^{2}}+\delta(x)\delta(t).
\end{align}
We have
\begin{align}
\phi_\mathtt{A}(x,t)=\frac{1}{\sqrt{4\pi D_\mathtt{A} t}}\exp(-\frac{x^2}{4 D_\mathtt{A} t}),~ x\in \mathbb{R},~t\geq 0.
\end{align}
Similarly, we define $\phi_\mathtt{B}(x,t)$ and $\phi_\mathtt{C}(x,t)$ as the solutions of
\begin{align}\label{greenBC}
\frac{\partial \phi_\mathtt{B}}{\partial t}=D_\mathtt{B}~\frac{\partial^{2}\phi_\mathtt{B}}{{\partial x}^{2}}+\delta(x)\delta(t), \qquad
\frac{\partial \phi_\mathtt{C}}{\partial t}=D_\mathtt{C}~\frac{\partial^{2}\phi_\mathtt{C}}{{\partial x}^{2}}+\delta(x)\delta(t).
\end{align}
Then,
the solutions of \eqref{zerosol},
\eqref{othersolA}-\eqref{othersolC} are as follows:
\begin{align}\label{ex1:A0,B0}
&[\mathtt{A}]_{0}=\phi_\mathtt{A}**f_{\mathtt{A}},~~ [\mathtt{B}]_{0}=\phi_\mathtt{B}**f_{\mathtt{B}},~~[\mathtt{C}]_{0}=0,
\end{align}
and for $i\geq 1$,
\begin{align}
&[\mathtt{A}]_{i}=-\phi_\mathtt{A}**([\mathtt{A}]_{0:i-1}*^d[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1}),\label{recu-eq1}\\
&[\mathtt{B}]_{i}=-\phi_\mathtt{B}**([\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1}),\label{recu-eq2}\\
&[\mathtt{C}]_{i}=+\phi_\mathtt{C}**([\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}-\gamma[\mathtt{C}]_{i-1}),\label{recu-eq3}
\end{align}
where $**$ stands for the two-dimensional convolution (see \eqref{eqnconvts}).
In general, the perturbation method does not yield a closed-form analytical solution for the reaction-diffusion equation because it is not possible (in general) to find a closed-form solution for the recursive equations \eqref{recu-eq1}-\eqref{recu-eq3}. However, the following example indicates a special instance where a closed-form solution can be found.
\begin{example}\label{exampleA1}
Consider the special choice of $f_{\mathtt{A}}(x,t)=f_{\mathtt{B}}(x,t)=1 ,x\in \mathbb{R}^{d}, t\geq 0 $, $D_{{\mathtt{A}}}=D_{{\mathtt{B}}}$, and $\gamma=0$. While these input signals are not interesting, the convolutions in \eqref{recu-eq1}-\eqref{recu-eq3} can be explicitly worked out for this choice of input signals.
Using symmetry, we have $[\mathtt{A}]_i(x,t)=[\mathtt{B}]_i(x,t)$ for $i\geq 0$. The zero and first terms of perturbation can be computed as follows:
\begin{align}
&[\mathtt{A}]_{0}(x,t)=\phi**f_{ \mathtt{A}}=\int_{0}^{t}\int_{x\in \mathbb{R}^{d}}\phi(x-x',t-t')f_{ \mathtt{A}}(x',t')dx'dt'=t.
\end{align}
We prove, by induction, that
$[\mathtt{A}]_{i}(x,t)=(-1)^{i}\alpha_{i}t^{2i+1}$ where $\alpha_{i}=\sum_{j=0}^{i-1} \alpha_{j}\alpha_{i-1-j}/(2i+1), i\geq 1, $ and $\alpha_0=1$. We have:
\begin{align}
[\mathtt{A}]_{i}&=-\phi**\sum_{j=0}^{i-1}[\mathtt{A}]_{j}[\mathtt{A}]_{i-1-j}\nonumber\\ &=-\int_{0}^{t}\int_{x\in \mathbb{R}^{d}}\phi(x-x',t-t')\sum_{j=0}^{i-1}(-1)^{j}\alpha_{j}t'^{2j+1}(-1)^{i-1-j}\alpha_{i-1-j}t'^{2i-2j-1} dx'dt'\nonumber\\&=(-1)^{i}\frac{\sum_{j=0}^{i-1} \alpha_{j}\alpha_{i-1-j}}{2i+1}t^{2i+1}.
\end{align}
By induction, one can prove that $(1/3)^{i}\leq \alpha_{i}\leq (1/2)^i$. The solution of reaction-diffusion for this particular example is as follows: $[\mathtt{A}](x,t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{A}]_i=\sum_{i=0}^{\infty}\lambda^{i}(-1)^{i}\alpha_{i}t^{2i+1}$.
\end{example}
\subsection{Interpretation in terms of Picard's iterative process}
An alternative way to understand the recursive equations \eqref{ex1:A0,B0}-\eqref{recu-eq3} is through
Picard's iterative process. In this process, one converts the differential equation to an integral form and uses it to identify the Picard iteration operator.
Consider the reaction-diffusion equations in \eqref{A:ex1eq}-\eqref{C:ex1eq}. To convert this system of
differential equations to an integral form, we utilize
the fact that the solution of the differential equation $
\frac{\partial[\mathtt{u}]}{\partial t}=D_\mathtt{u}~\frac{\partial^{2}[\mathtt{u}]}{{\partial x}^{2}}+g_{u}$
with zero initial and boundary condition is $u=\phi_{u}**g_{u}$ where $\phi_{u}$ is the solution of the differential equation for an impulse input (impulse response). Thus, from \eqref{A:ex1eq}-\eqref{C:ex1eq} we obtain
\begin{align}
&[\mathtt{A}](x,t)=\phi_{\mathtt A}**(-\lambda[\mathtt{A}][\mathtt{B}]+\gamma\lambda[\mathtt{C}]+f_{\mathtt{A}}(x,t)),\\
&[\mathtt{B}](x,t)=\phi_{\mathtt B}**(-\lambda[\mathtt{A}][\mathtt{B}]+\gamma\lambda[\mathtt{C}]+f_{\mathtt{B}}(x,t)),\\
&[\mathtt{C}](x,t)=\phi_{\mathtt C}**(\lambda[\mathtt{A}][\mathtt{B}]-\gamma\lambda[\mathtt{C}]).
\end{align}
Therefore, the Picard iteration operator is obtained as follows:
\begin{align}
&\mathcal{\tau}:\mathcal{C}\times \mathcal{C}\times \mathcal{C} \mapsto \mathcal{C}\times \mathcal{C}\times \mathcal{C}\nonumber\\
&\mathcal{\tau}(\mathtt{a},\mathtt{b},\mathtt{c})=(\phi_\mathtt{A}**(-\lambda\mathtt{a}\mathtt{b}+\lambda\gamma\mathtt{c}+f_\mathtt{A}),\phi_\mathtt{B}**(-\lambda\mathtt{a}\mathtt{b}+\lambda\gamma\mathtt{c}+f_\mathtt{B}),\phi_\mathtt{C}**(\lambda\mathtt{a}\mathtt{b}-\gamma\lambda\mathtt{c})),
\end{align}
where $\mathcal{C}$ is the space of continuous functions. The solution of the reaction-diffusion system is a fixed point of the Picard iteration operator. Picard's iterative process suggests the following recursive algorithm to find a fixed point of the operator $\mathcal{\tau}$:
\begin{align}
(\mathtt{a}^{(n+1)},\mathtt{b}^{(n+1)},\mathtt{c}^{(n+1)})=\mathcal{\tau}(\mathtt{a}^{(n)},\mathtt{b}^{(n)},\mathtt{c}^{(n)}),~~~(\mathtt{a}^{(-1)},\mathtt{b}^{(-1)},\mathtt{c}^{(-1)})=(0,0,0),~~n\geq-1\label{banchiter},
\end{align}
where $(\mathtt{a}^{(n)},\mathtt{b}^{(n)},\mathtt{c}^{(n)})$ is the triple of functions that we obtain at the $n$th iteration. Picard's iterative process relates to the perturbation method as follows: using induction and comparing \eqref{recu-eq1}-\eqref{recu-eq3} with the recursive equations given in \eqref{banchiter}, one obtains
\begin{align}
\mathtt{a}^{(n)}&=\sum_{i=0}^{n}\lambda^{i}
[\mathtt{A}]_{i}(x,t)&n=0,1\\
\mathtt{a}^{(n)}&=\sum_{i=0}^{n}\lambda^{i}
[\mathtt{A}]_{i}(x,t)+\mathcal{O}(\lambda^{n+1})&n\geq 2.
\end{align}
A similar statement holds for $\mathtt{b}^{(n)}$ and $\mathtt{c}^{(n)}$, showing that their first $n$ lower terms match the first $n$ lower terms of the Taylor series expansion in the perturbation method.
\subsection{Approximate Solutions}
In practice, we can approximate the true concentrations of molecules by computing the first $n$ terms in the Taylor series expansion:
\begin{align}
[\mathtt{A}](x,t)\approx \sum_{i=0}^{n}\lambda^{i}
[\mathtt{A}]_{i}(x,t)
\end{align}
for some fixed $n$. One can stop at the $n$-term of the Taylor expansion when the contribution of the $(n+1)$th term is negligible compared to the first $n$ terms. While $n$th term can be explicitly found in Example \ref{exampleA1}, for general reaction-diffusion equations, there is no (non-recursive) explicit expression for the terms in the Taylor series expansion.
In Section \ref{sec:Val}, we illustrate the sufficiency of choosing $n=1$ or $n=2$ when the length of the time interval $T$ is in the order of seconds. In the following, we provide a theoretical justification for choosing $n=1$ when the time interval $T$ is small: consider the reaction-diffusion equations
\begin{align}
&\frac{\partial[\mathtt{A}]}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]}{{\partial x}^{2}}-\lambda~[\mathtt{A}][\mathtt{B}]+\gamma \lambda [\mathtt{C}]+f_\mathtt{A}(x,t),\\
& \frac{\partial[\mathtt{B}]}{\partial t}=D_\mathtt{B}~\frac{\partial^{2}[\mathtt{B}]}{{\partial x}^{2}}-\lambda~[\mathtt{A}][\mathtt{B}]+\gamma \lambda [\mathtt{C}]+f_\mathtt{B}(x,t),\\
& \frac{\partial[\mathtt{C}]}{\partial t}=D_\mathtt{C}~\frac{\partial^{2}[\mathtt{C}]}{{\partial x}^{2}}+\lambda~[\mathtt{A}][\mathtt{B}]-\gamma \lambda [\mathtt{C}],
\end{align}
along with the initial conditions $
[\mathtt{A}](x,0)=g_{\mathtt{A}}(x), [\mathtt{B}](x,0)=g_{\mathtt{B}}(x)$ and $ [\mathtt{C}](x,0)=g_{\mathtt{C}}(x)$.
The one-step finite difference iteration (in the finite difference method) for a small value of $t$ is as follows:
\begin{align}
[\mathtt{A}]_{FDM}(x,t)&=[\mathtt{A}](x,0)+t\frac{\partial[\mathtt{A}]}{\partial t}(x,0)
\nonumber\\&=g_{\mathtt{A}}(x)+t\left[ D_{\mathtt{A}}g_{\mathtt{A}}''(x)- \lambda g_{\mathtt{A}}(x)g_{\mathtt{B}}(x)+ \gamma\lambda g_{\mathtt{C}}(x)+ f_{\mathtt{A}}(x,0)\right]
\nonumber\\&=g_{\mathtt{A}}(x)+\mathcal{O}(t).\end{align}
Let $[\mathtt{A}]^{(1)}(x,t)=[\mathtt{A}]_{0}(x,t)+\lambda [\mathtt{A}]_{1}(x,t)$ be the approximate solution of the perturbation method up to the order one terms. In Appendix \ref{AppA:3} we show the following equivalence with the finite difference solution:
\begin{theorem}\label{thmN1}
The first-order approximation $[\mathtt{A}]^{(1)}(x,t)=[\mathtt{A}]_{0}(x,t)+\lambda [\mathtt{A}]_{1}(x,t)$ and the finite difference method are equivalent for small values of $t$, \emph{i.e.,} $ [\mathtt{A}]^{(1)}(x,t)= [\mathtt{A}]_{FDM}(x,t)+\mathcal{O}(t^2)$.
A similar statement holds for molecules of type $\mathtt{B}$ and $\mathtt{C}$.
\end{theorem}
As illustrated in Section
\ref{sec:Val}, the time steps do not necessarily need to be very small for the perturbation solution to closely track the true solution (and can be in the order of seconds). Therefore, the perturbation method can be understood as a generalization of the finite difference method that allows for longer time steps.
\section{Modulation Design}\label{Sec:Mod}
In Section \ref{sec::generalperturbation}, a closed-form solution for the reaction-diffusion equation \eqref{ex1:reaction} was obtained. In this section, we use this solution to find the optimal release pattern of molecules by the transmitters in order to minimize the error probability of the communication system.
Suppose transmitter $ \mathsf T_{\mathtt {A}}$ intends to send a message bit $M_\mathtt{A}\in\{0,1\}$ to the receiver. The transmitter $ \mathsf T_{\mathtt {A}}$ releases molecules of type $[\mathtt {A}]$ into the medium according to $f_\mathtt{A}^{0}(x,t)=a_0(t)\delta(x), 0\leq t\leq T$ if $M_\mathtt{A}=0$, and according to $f_\mathtt{A}^{1}(x,t)=a_1(t)\delta(x), 0\leq t\leq T$ if
$M_\mathtt{A}=1$.
In other words, the waveform $a_{0}(t)$ is the released concentration of molecules of type $\mathtt A$ when $M_\mathtt{A}=0$, and $a_{1}(t)$
is the released density when $M_\mathtt{A}=1$. Similarly, suppose that transmitter $ \mathsf T_{\mathtt {B}}$
encodes message $M_\mathtt{B}\in\{0,1\}$ by $f_{\mathtt{B}}^{j}(x,t)=b_{j}(t)
\delta(x-d_{\mathtt{B}}), 0\leq t\leq T, j=0,1$, where the waveforms $b_{0}(t)$ and $b_{1}(t)$ are defined similarly for transmitter $\mathsf T_{\mathtt {B}}$. The total amount of released molecules of types $\mathtt A$ and $\mathtt B$ during the transmission period $T$ is assumed to be at most $s_{\mathtt A}$ and $s_{\mathtt B}$ respectively, \emph{i.e.,}
\begin{align}\int_{0}^T a_i(t)dt\leq s_{\mathtt A},~\int_{0}^T b_i(t)dt\leq s_{\mathtt B}, \quad i=0,1.\label{eqn:New6}
\end{align}
The input messages
${M}_\mathtt{A}, {M}_\mathtt{B}$ are assumed to be uniform Bernoulli random variables.
Receiver $\mathsf R_{\mathtt {C}}$
samples the density of molecules of type $\mathtt{C}$ at its location at the end of the time slot at time $t=T$. The receiver is assumed to be transparent and suffers from the particle counting noise. More specifically, the receiver is assumed to get a sample from $\mathsf{Poisson}\big(V\cdot [\mathtt{C}](d_{R},T)\big)$ where $[\mathtt{C}](d_{R},T)$ is the density of molecules of type $\mathtt C$ at time $T$ and location $d_{R}$, and $V$ is a constant.
Using this observation, $\mathsf R_{\mathtt {C}}$
outputs its estimate of the transmitted message pair $(\hat{M}_\mathtt{A},\hat{M}_\mathtt{B})\in \{00,01,10,11\}$. The probability of error is
\begin{equation}
Pr(e)=Pr\{ (M_{\mathtt{A}}, M_{\mathtt{B}})\neq (\hat{M}_\mathtt{A},\hat{M}_\mathtt{B})\}.\label{eqnErrorProb1}
\end{equation}
Having transmitted messages
$a_{m_{\mathtt{A}}}(t), b_{m_{\mathtt{B}}}(t)$
for a message pair $(m_{\mathtt{A}}, m_{\mathtt{B}})$, let $\rho_{m_{\mathtt{A}}, m_{\mathtt{B}}}$ denote the density of molecules of type $\mathtt{C}$ at the time of sampling. The receiver aims to recover $\hat m_{\mathtt{A}}, \hat m_{\mathtt{B}}$ using a sample $\mathsf{Poisson}(V\cdot \rho_{m_{\mathtt{A}}, m_{\mathtt{B}}})$. Our goal is to design nonnegative signals $a_{i}(t), b_{j}(t)$ satisfying \eqref{eqn:New6}
to minimize the error probability. The key utility of the technique given in Section \ref{sec::generalperturbation} is that it provides a closed-form expression for
$\rho_{m_{\mathtt{A}}, m_{\mathtt{B}}}$ in terms of $a_i(t)$ and $b_j(t)$.
An order $n$ approximation to the reaction-diffusion is when we consider the reaction-diffusion equations up to the terms of order $\lambda^n$ in the Taylor expansion. We say that waveforms $a_i(t)$ and $b_j(t)$ are
order $n$ optimal waveforms if they minimize the error probability for the approximate reaction-diffusion equations up to order $n$. For low reaction rates, we are mainly interested in order one optimal waveforms. In the following theorem, we show that for low reaction rates, one possible optimal strategy is for
$ \mathsf T_{\mathtt {A}},\mathsf T_{\mathtt {B}}$ to release molecules in at most two time instances in a bursty fashion. The location of these instantaneous releases is determined by the message it wants to transmit. In other words, waveform $a_{i}(t)$ is the sum of two Dirac's delta functions.
\begin{theorem}\label{th::1}
Consider approximating the solution of the reaction-diffusion equations with the perturbation solution up to the
first-order term. Consider the problem of choosing waveforms $a_i(t), b_i(t)$ to minimize the error probability. Then, restricting minimization to waveforms $a_i(t), b_i(t)$ with the following structure does not change the minimum value of error probability:
\begin{align}
&a_0(t)=\hat{a}_{01}\delta(t-t_1^{[a_0]})+\hat{a}_{02}\delta(t-t_2^{[a_0]}),~
a_1(t)=\hat{a}_{11}\delta(t-t_1^{[
a_1]})+\hat{a}_{12}\delta(t-t_2^{[a_1]}),\\
& b_0(t)=\hat{b}_{01}\delta(t-t_1^
{[b_0]})+\hat{b}_{02}\delta(t-t_2^
{[b_0]}),~~
b_1(t)=\hat{b}_{11}\delta(t-t_1^{[b_1]})+\hat{b}_{12}\delta(t-t_2^{[b_1]}),
\end{align}
for some non-negative constants $\hat{a}_{ij},\hat{b}_{ij}$ satisfying
$\sum_{j}\hat{a}_{ij}\leq s_{\mathtt A}$ and
$\sum_{j}\hat{b}_{ij}\leq s_{\mathtt B}$ for $i=0,1$,
and for some
$t_j^{[a_i]},t_j^{[b_i]}\in[0,T]$.
\end{theorem}
\begin{proof}
By substituting $f_\mathtt{A}^{i}(x,t),f_\mathtt{B}^{i}(x,t), i=0,1$ in \eqref{ex1:A0,B0} and \eqref{recu-eq1}-\eqref{recu-eq3}, one obtains an explicit formula for (the first-order approximation of) the concentration of molecules $\mathtt{C}$ in terms of $f_\mathtt{A}^{i}(x,t),f_\mathtt{B}^{i}(x,t), i=0,1$. In particular, the concentration $[\mathtt{C}](d_{R},T)$ at receiver's location at the sampling time $T$ is as follows:
\begin{align}
[\mathtt{C}](d_
{R},T)=\lambda\int_{0}^{T}
\int_{-\infty}^{+\infty}\int_{0}^{t'}
\int_{0}^{t'}&
\phi_{\mathtt{C}}(d_R-x',T-t')
\phi_{\mathtt{A}}(x',t'-t_1)
\phi_{\mathtt{B}}(x'-d_{\mathtt{B}},t'-t_2)\nonumber\\&
a_{M_{\mathtt{A}}}(t_1)
b_{M_{\mathtt{B}}}(t_2)
dt_{1}dt_{2}dx'dt'.
\end{align}
Observe that
$[\mathtt{C}](d_
{R},T)$ has a \emph{bilinear} form with respect to $a_{M_{\mathtt{A}}}(t)$ and $b_{M_{\mathtt{B}}}(t)$ . In the other words, if $a_{M_{\mathtt{A}}}(t)$ is kept fixed then $[\mathtt{C}](d_
{R},T)$ is linear with respect to $b_{M_{\mathtt{B}}}(t)$ and vice versa.
Let
\begin{equation}\label{densities::ex1:final}
\rho=\left\{ {\begin{array}{*{20}{llll}}
\rho_{00}=[\mathtt{C}](d_{R},T)& & \text{if}&(M_{\mathtt{A}},M_{\mathtt{B}})=(0,0)\\
\rho_{01}=[\mathtt{C}](d_{R},T)& & \text{if}&(M_{\mathtt{A}},M_{\mathtt{B}})=(0,1)\\
\rho_{10}=[\mathtt{C}](d_{R},T)& & \text{if}&(M_{\mathtt{A}},M_{\mathtt{B}})=(1,0)\\ \rho_{11}=[\mathtt{C}](d_{R},T)& & \text{if}&(M_{\mathtt{A}},M_{\mathtt{B}})=(1,1)
\\
\end{array}}\right.
\end{equation}
Using \eqref{densities::ex1:final} we can compute $
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$ for any given $a_{0}(t), a_{1}(t), b_0(t)$ and $b_1(t)$. The receiver aims to recover $\hat m_{\mathtt{A}}, \hat m_{\mathtt{B}}$ using a sample $\mathsf{Poisson}(V\rho_{m_{\mathtt{A}}, m_{\mathtt{B}}})$. Minimizing the error probability (given in \eqref{eqnErrorProb1}) is equivalent to solving a Poisson hypothesis testing problem (see Appendix \ref{AppB} for a review). Our goal is to design nonnegative signals $a_{i}(t), b_{j}(t)$ satisfying \eqref{eqn:New6}
to minimize the error probability of this Poisson hypothesis testing problem.
Note that the tuple $
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$ forms a \emph{sufficient statistic} for computing the error probability. In other words, if we change the signals $a_{i}(t), b_{j}(t)$ in such a way that the four numbers $
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$ are unchanged, the error probability remains unchanged.
Take signals $a_i(t)$ and $b_j(t)$ for $i,j\in\{0,1\}$ that minimize the error probability. We claim that we can find $\hat a_0(t)$ of the form
$\hat a_0(t)=\hat{a}_{01}\delta(t-t_1^{[a_0]})+\hat{a}_{02}\delta(t-t_2^{[a_0]})$
for some $\hat{a}_{01}$ and $\hat{a}_{02}$
such that if we replace the $(a_0(t), a_1(t), b_0(t), b_1(t))$ by $(\hat a_0(t), a_1(t), b_0(t), b_1(t)),$ the corresponding values of $
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$ remain unchanged. Moreover, $\hat a_0(t)$ satisfies the power constraint \eqref{eqn:New6}. Since the error probability depends on the waveforms only through the values of
$
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$, we deduce that replacing $a_0(t)$ with $\hat a_0(t)$ does not change the error probability. Since signals $a_i(t)$ and $b_j(t)$ for $i,j\in\{0,1\}$ minimized the error probability, we deduce that using $\hat a_0(t)$ would also achieve the minimum possible probability of error.
By a similar argument we can reduce the support of waveforms $a_1(t)$, $b_0(t)$, and $b_1(t)$ one by one while fixing the other waveforms. This completes the proof.
It remains to show that changing $a_0(t)$ by $\hat{a}_0(t)$ while preserving the values of $
(\rho_{00},\rho_{01},\rho_{10},\rho_{11})$ is possible.
Replacing $a_0(t)$ by $\hat{a}_0(t)$ while fixing $a_1(t)$, $b_0(t)$, and $b_1(t)$ may only change $\rho_{00}$ and $\rho_{01}$. The values of $\rho_{10}$ and $\rho_{11}$ are only functions of $a_1(t)$, $b_0(t)$, and $b_1(t)$ which we are fixing. Moreover, $\rho_{00}$ and $\rho_{01}$ are linear functions of $a_0(t)$ when we fix $b_0(t)$ and $b_1(t)$.
In other words, one can find functions $f_1(t)$ and $f_2(t)$ such that
\begin{equation}
\rho_{00}=\int_{t\in[0,T]}a_0(t)f_1(t)dt,\qquad \rho_{01}=\int_{t\in[0,T]}a_0(t)f_2(t)dt.\label{eqnDeff1}
\end{equation}
The power constraint \eqref{eqn:New6} is also a linear constraint on $a_0(t)$. We would like to change $a_0(t)$ to $\hat a_0(t)$ in such a way that two linear constraints corresponding to $\rho_{00}$ and $\rho_{01}$ are preserved, and the power of $\hat a_0(t)$ is less than or equal to the power of $a_0(t)$. Existence of $\hat{a}_0(t)$ with these properties follows from Lemma \ref{SupportLemma} given in Appendix \ref{AppC} (with the choice of $f(t)=1$, $n=2$, and $f_1(t)$ and $f_2(t)$ given in \eqref{eqnDeff1}), once we view $a_0(t)$ as an unnormalized probability distribution.
\end{proof}
In order to determine parameters in the statement of Theorem \ref{th::1} we need to optimize over constants $\hat{a}_{ij}$, $\hat{b}_{ij}$, $t_j^{[a_i]},t_j^{[b_i]}$.
From the first-order equation,
$\rho_{ij}$ (the concentration when the first transmitter sends bit $i\in\{0,1\}$ and the second transmitter sends the bit $j\in\{0,1\}$) equals
\begin{equation}
\begin{array}{*{20}{l}}
\rho_{00}=\sum_{i,j=1}^{2}\hat{a}_{0i}\hat{b}_{0j}G(t_{i}^{[a_0]},t_{j}^{[b_0]})
\\
\rho_{01}=\sum_{i,j=1}^{2}\hat{a}_{0i}\hat{b}_{1j}G(t_{i}^{[a_0]},t_{j}^{[b_1]})\\
\rho_{10}=\sum_{i,j=1}^{2}\hat{a}_{1i}\hat{b}_{0j}G(t_{i}^{[a_1]},t_{j}^{[b_0]})\\
\rho_{11}=\sum_{i,j=1}^{2}\hat{a}_{1i}\hat{b}_{1j}G(t_{i}^{[a_1]},t_{j}^{[b_1]}),
\end{array}
\end{equation}
where
\begin{equation}
G(t_i,t_j)=\lambda\phi_{\mathtt{C}}(x,t)**\big(\phi_{\mathtt{A}}(x,t-t_i)
\phi_{\mathtt{B}}(x-d_{\mathtt{B}},t-t_j)\big)
|_{t=t_s,x=d_R}.
\end{equation}
To recover the messages of the transmitters, the receiver needs to solve a Poisson hypothesis testing problem (see Appendix \ref{AppB}) with four hypotheses: $\rho=V\rho_{00}$ or $\rho=V\rho_{01}$ or $\rho=V\rho_{10}$ or $\rho=V\rho_{11}$ for a constant $V$. The error probability of this hypothesis testing problem should be minimized with respect to $\hat{a}_{ij}$, $\hat{b}_{ij}$, $t_j^{[a_i]},t_j^{[b_i]}$.
As an example, consider the following parameters in Table\ref{tabj}:
\begin{table}[H]
\centering
\caption{Parameters}
\label{tabj}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(D_{\mathtt{A}},D_{\mathtt{B}},D_{\mathtt{C}})[m^{2}s^{-1}]$} &\multirow{2}{*} {$10^{-9}\times(1,1,1)$}\\
&\\
\hline
\multirow{2}{*} {$\lambda[molecules^{-1}.m^{3}.s^{-1}]$} & \multirow{2}{*}{$ 10^{-23}$}\\
&\\
\hline
\multirow{2}{*}{$V[m^{3}]$} &\multirow{2}{*} {$10^{-11}$}\\
&\\
\hline
\multirow{2}{*}{$T[s]$ }& \multirow{2}{*}{$3$} \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{B}} [m]$} & \multirow{2}{*}{$5\times 10^{-5}\times(2,0,0) $ } \\
&\\
\hline
\multirow{2}{*}{$d_{R}[m]$} & \multirow{2}{*}{$5\times 10^{-5}\times(1,0,0)$} \\
&\\
\hline
\multirow{2}{*}{$s_{\mathtt{A}}[molecules.m^{-3}],s_{\mathtt{B}}[molecules.m^{-3}])$} & \multirow{2}{*}{$10^7\times(1,1)$} \\
&\\
\hline
\end{tabular}
\end{table}
Simulation results yield the optimal values for $\hat{a}_{ij}$, $\hat{b}_{ij}$, $t_j^{[a_i]},t_j^{[b_i]}$ as in Table\ref{tab78}
\begin{table}[H]
\centering
\caption{Values of unknown parameters.}
\label{tab78}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(\hat{a}_{01},\hat{a}_{02})$} &\multirow{2}{*} {$(1.13\times 10^6,8.82\times 10^6)$}\\
&\\
\hline
\multirow{2}{*} {$(\hat{a}_{11},\hat{a}_{12})$} & \multirow{2}{*}{$ (2.17\times 10^6,1.07\times 10^6)$}\\
&\\
\hline
\multirow{2}{*}{$(\hat{b}_{01},\hat{b}_{02})$} &\multirow{2}{*} {$(0.97\times 10^6,3.2\times 10^6)$}\\
&\\
\hline
\multirow{2}{*}{$(\hat{b}_{11},\hat{b}_{12})$ }& \multirow{2}{*}{$(3.52\times 10^6,6.46\times 10^6)$} \\
&\\
\hline
\multirow{2}{*}{$(t_{1}^{[a_0]},t_{2}^{[a_0]})$} & \multirow{2}{*}{$(0,2) $ } \\
&\\
\hline
\multirow{2}{*}{$(t_{1}^{[a_1]},t_{2}^{[a_1]})$} & \multirow{2}{*}{$(0,1) $ } \\
&\\
\hline
\multirow{2}{*}{$(t_{1}^{[b_0]},t_{2}^{[b_0]})$} & \multirow{2}{*}{$(1,2)$} \\
&\\
\hline
\multirow{2}{*}{$t_{1}^{[b_1]},t_{2}^{[b_1]}$} & \multirow{2}{*}{$(1,2)$} \\
&\\
\hline
\multirow{2}{*}{$p_{e}$} & \multirow{2}{*}{$2.6\times 10^{-5}$} \\
&\\
\hline
\end{tabular}
\end{table}
Figs. \ref{probabilty1} and \ref{probabilty2} show the optimal error probability (as defined in \eqref{eqnErrorProb1}) as a function of $s=s_{\mathtt{A}}=s_{\mathtt{B}}$ for the above parameters.
Fig. \ref{pulsevsdelta} compares the error probabilities of optimal waveforms (sum of two delta functions) with a pulse release pattern, \emph{i.e.,} when we consider four waveforms $(f_{\mathtt{A}}^{0}(t)=a_{0}\mathbbm{1}_{0\leq t\leq T},f_{\mathtt{A}}^{1}(t)=a_{1}\mathbbm{1}_{0\leq t\leq T},f_{\mathtt{B}}^{0}(t)=b_{0}\mathbbm{1}_{0\leq t\leq T},f_{\mathtt{B}}^{1}(t)=b_{1}\mathbbm{1}_{0\leq t\leq T})$ as input signals to reaction diffusion equations. The error probability of the pulse waveforms is obtained by minimizing the error probability over pulse amplitudes $a_0, a_1\leq s_{\mathtt{A}}/T$ and $b_0, b_1\leq s_{\mathtt{B}}/T$.
\begin{figure}[H]
\centering
\includegraphics[scale=.75]{errorvsS5.png}
\caption{Error probability of the optimal waveforms for
$10^5\leq s \leq 4\times 10^6$ [molecules.$m^{-3}$].}
\label{probabilty1}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=.75]{errorvsS3.png}
\caption{Error Probability of the optimal waveforms
for
$4\times 10^6\leq s \leq 1.4\times 10^7$
[molecules.$m^{-3}$].}
\label{probabilty2}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.75]{logpulsevsdelta.png}
\caption{Comparison between optimal wave-form and pulse family.}
\label{pulsevsdelta}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.75]{accurace1.png}
\caption{First-order approximation for the first set parameters }
\label{fig:1e-22,unequal,Nx=64,first}
\end{figure}
\section{Simulation}
\label{sec:Val}
In this section, we validate our method through numerical simulation. We need to specify the diffusion coefficients, reaction rate $\lambda$, input signals $f_{\mathtt A}$ and $f_{\mathtt B}$ as well as the locations of the transmitters. Previous works in the literature commonly take the number of released molecules to be either in the order of $10^{-14}$ moles (which is $10^9$ molecules) \cite{arjmandi2019diffusive,farahnak2018medium:j:16,cao2019chemical:j:ro}, or in the order of $10^{-3}$ moles \cite{farsad2016molecular:F:16}. The reaction rates in these works are generally in the order of $1$ to $10^8$ $\text{mole}^{-1}.m^{3}.s^{-1}$ or equivalently $10^{-23}$ to $10^{-15}$ $\text{molecules}^{-1}.m^{3}.s^{-1}$ \cite{cao2019chemical:j:ro,farsad2016molecular:F:16,bi2019chemical:j:9,chang2005physical}.
We run three simulations for three choices of parameters for the reaction given in \eqref{ex1:reaction}. The exact solutions are obtained by the finite difference method (FDM). The first simulation borrows its system parameters from \cite{cao2019chemical:j:ro} as in Table \ref{tab1}:
\begin{table}
\centering
\caption{First set of parameters.}
\label{tab1}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(D_{\mathtt{A}},D_{\mathtt{B}},D_{\mathtt{C}})[m^{2}s^{-1}]$} &\multirow{2}{*} {$10^{-10}\times(10,7,1)$}\\
&\\
\hline
\multirow{2}{*} {$\lambda[molecules^{-1}.m.s^{-1}]$} & \multirow{2}{*}{$ 10^{-22}$}\\
&\\
\hline
\multirow{2}{*} {$\gamma[s^{-1}]$} & \multirow{2}{*}{$0$}\\
&\\
\hline
\multirow{2}{*}{$f_{\mathtt{A}}(x,t)$} &\multirow{2}{*} {$5\times 10^{8}\delta(x)\delta(t)$}\\
&\\
\hline
\multirow{2}{*}{$f_{\mathtt{B}}(x,t)$ }& \multirow{2}{*}{$2.4\times 10^{9}\delta(x-d_{\mathtt{B}})\delta(t)$} \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{A}} [m]$} & \multirow{2}{*}{$0 $ } \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{B}} [m]$} & \multirow{2}{*}{$ 10^{-4} $ } \\
&\\
\hline
\multirow{2}{*}{$d_{R}[m]$} & \multirow{2}{*}{$5\times 10^{-5}$} \\
&\\
\hline
\end{tabular}
\end{table}
In the second simulation we use the same parameters as in the first simulation, but make the three diffusion constants be the same $(D_{\mathtt{A}}=D_{\mathtt{B}}=D_{\mathtt{C}}=10^{-9}[m^2/s])$.
Figs. \ref{fig:1e-22,unequal,Nx=64,first} and \ref{fig:1e-22,equal,Nx=120,first} are plotted based on the first and second set of system parameters respectively. They depict the concentration of the molecule of type $\mathtt{C}$ at the receiver's location calculated by the first-order term of perturbation method along with the true solution (computed via FDM).
It can be seen that for these system parameters, the first-order term of the perturbation series is closely following the true solution.
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.75]{jomong5.png}
\caption{First-order approximation for the second set parameters}
\label{fig:1e-22,equal,Nx=120,first}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.75]{jomong67new.png}
\caption{First-order and second-order approximate solutions for $\lambda=10^{-15}$.}
\label{fig:1e-15,equal,Nx=120,first,T=15}
\end{figure}
Using the same parameters as in the second simulation, one can observe that the first-order term of perturbation series is still accurate as long as $\lambda\leq 10^{-16} [\text{molecules}^{-1}.m.s^{-1}]$. Fig. \ref{fig:1e-15,equal,Nx=120,first,T=15} plots the curves for $\lambda=10^{-15}$. As one can observe while the first-order term of the perturbation series does not track the exact solution for $t>5s$, while the second-order term tracks the true solution in $t<15s$.
For $t>15s$, the perturbation solution up to the second-order term does not track the exact solution well, hence to obtain an accurate solution we have to consider third or higher-order terms in the perturbation solution. This observation is consistent with the theoretical analysis given in Appendix \ref{AppA0 }, where the convergence rate of perturbation solution depends on the length of the time interval $T$. We also remark that in many prior works on chemical reactions in MC, the observation time $T$ is rather small: \cite{arjmandi2019diffusive} uses $T\leq 0.1 s$, \cite{bi2019chemical:j:9} uses $T\leq 4 s$; \cite{cao2019chemical:j:ro} uses $T\leq 10 s$, while in other cases $T$ is in taken of order $\mu s$ \cite{farahnak2018medium:j:16,mosayebi2017type:F:21}.
Fig. \ref{relative1} fixes $T=10[s]$ and plots the relative error of the first-order approximation solution of $[\mathtt{C}]$ for the second set of parameters in terms of $\lambda$. Fig. \ref{relative21} draws a similar curve for the second-order approximation of the perturbation solution.
These figures confirm (the theoretical result) that for sufficiently small reaction rates, the relative error is negligible. For instance, to have a relative error of at most $0.01$, the first-order solution can be used for
$\lambda<3\times10^{-16}$ (Fig. \ref{relative1}) while the second-order solution is valid until $\lambda<3\times10^{-15}$ (Fig. \ref{relative21}).
Next, let us fix the relative error to be $0.05$ and define the permissible time interval as
\begin{align}
T_{\max}=\max\left\{T:\left|\frac{[\mathtt{C}]_{FDM}(t)-[\mathtt{C}]_{PER-Order-1}(t)}{[\mathtt{C}]_{FDM}(t)}\right|\leq .05, \qquad \forall t\in [0,T]\right\}.\label{TmaxDef}\end{align}
In other words, $T_{\max}$ is the maximum simulation time interval for which the relative error is less than 5 percent.
Fig. \ref{maxt2} shows the $T_{\max}$ for the first-order approximation of $[\mathtt{C}]$ in terms of $\lambda$.
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{relativeerror1.png}
\caption{Relative error for the first-order approximation.}
\label{relative1}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{relativeeerror21.png}
\caption{Relative error for up to second-order approximation.}
\label{relative21}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{maxt2.png}
\caption{Permissible simulation time interval in terms of $\lambda$.}
\label{maxt2}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{jomong9.png}
\caption{First-order approximation curve for the third set parameters.}
\label{fig:.1,equal,Na=1micromole,first}
\end{figure}
In some applications, a significantly larger number of molecules (in the order of moles) are released into the medium. In the third simulation, we consider this case and take the following parameters in Table \ref{tab2}:
\begin{table}
\centering
\caption{Third set of parameters}
\label{tab2}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(D_{\mathtt{A}},D_{\mathtt{B}},D_{\mathtt{C}})[m^{2}.s^{-1}]$} &\multirow{2}{*} {$10^{-9}\times(1,1,1)$}\\
&\\
\hline
\multirow{2}{*} {$\lambda[molecules^{-1}.m.s^{-1}]$} & \multirow{2}{*}{$ 10^{-24}/6.02214$}\\
&\\
\hline
\multirow{2}{*} {$\gamma[s^{-1}]$} & \multirow{2}{*}{$0$}\\
&\\
\hline
\multirow{2}{*}{$f_{\mathtt{A}}(x,t)$} &\multirow{2}{*} {$6.02214\times 10^{17}\delta(x)\delta(t)$}\\
&\\
\hline
\multirow{2}{*}{$f_{\mathtt{B}}(x,t)$ }& \multirow{2}{*}{$6.02214\times 10^{17}\delta(x-d_{\mathtt{B}})\delta(t)$} \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{A}} [m]$} & \multirow{2}{*}{$0 $ } \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{B}} [m]$} & \multirow{2}{*}{$ 10^{-4} $ } \\
&\\
\hline
\multirow{2}{*}{$d_{R}[m]$} & \multirow{2}{*}{$5\times 10^{-5}$} \\
&\\
\hline
\end{tabular}
\end{table}
Fig. \ref{fig:.1,equal,Na=1micromole,first} shows that the first term of the perturbation series is close to the exact solution.
\textbf{The effect of dimension:}
While we assumed the transmitter and receivers to lie on a one-dimensional line, the problem can be solved in higher dimensions in a similar manner. In Figs. \ref{dim2} and \ref{dim3}, we compare the perturbation method (up to the first-order approximation) and the true solution in two and three dimensions. The parameters are the same as the second set of parameters with $\lambda[{molecules}^{-1}.m^{3}.s^{-1}]=10^{-23}$. As we see our solution is still accurate. We observe that as the dimension of the medium increases, concentrations decay faster. This is due to the extra degrees of freedom in the dispersion of molecules.
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.75]{dim2.png}
\caption{2d medium, first-order approximation}
\label{dim2}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{dim3.png}
\caption{3d medium, first-order approximation}
\label{dim3}
\end{figure}
\textbf{The effect of diffusion coefficients:}
In Fig. \ref{DconC} we plot $[\mathtt{C}]$ for different values of $D_\mathtt{C}$, for the second set of parameters. Observe that $[\mathtt{C}]$ is a decreasing function of $D_\mathtt{C}$. Fig. \ref{maxt} plots $T_{\max}$ (as defined in \eqref{TmaxDef}) versus $D_\mathtt{C}$ for the same parameters, showing that increasing $D_\mathtt{C}$ decreases $T_{\max}$.
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{effectonctwobox.png}
\caption{Effect of $D_\mathtt{C}$ on $[\mathtt{C}]$ }
\label{DconC}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.75]{maxt.png}
\caption{Impact of diffusion coefficient on the permissible simulation time interval}
\label{maxt}
\end{figure}
\section{Conclusion and Future Work}\label{Conclusion and Future Work}
We addressed the difficulty of designing modulation due to the lack of existence of closed-form solutions for the reaction-diffusion equation by providing an approximate solution to the reaction-diffusion equations. We observed that for many choices of system parameters our solution is accurate. Also,
we observed that the accuracy of our solution depends on time observation $T$. Using the proposed model,
we designed optimal waveforms for a multiple-access setting. In the Appendix\ref{ex2,3} of this paper two more examples are considered (Example II and Example III).
In Example II, we consider a communication system with one transmitter and one receiver. The transmitter can send molecules of type $\mathtt{A}$ or $\mathtt{B}$, while the receiver can only measure molecules of type $\mathtt{A}$. The medium has molecules of a different type $\mathtt{C}$ that can react with both $\mathtt{A}$ and $\mathtt{B}$. The transmitter releases molecules of type $\mathtt{B}$ in order to ``clean up" the medium of molecules of type $\mathtt{C}$ so that molecules of type $\mathtt{A}$ can reach the receiver without being dissolved through reaction with $\mathtt{C}$ as they travel from the transmitter to the receiver. In Example III, we consider a two-way communication system with two transceiver nodes. In other words, each node aims to both send and receiver information from the other node. Transmitters use different molecule types and these molecules can react with each other. The chemical reaction in this scenario is destructive because it reduces the concentration of molecules at the two transceivers.
Many other settings could be studied and left for future work. Firstly, we only considered transparent receivers in this work. The literature on MC also considers absorbing or ligand/reactive receivers. Since the perturbation method is broadly applicable to non-linear differential equations, the proposed framework can be also applied to absorbing or reactive receivers with more work. An absorbing receiver adds a boundary condition to the reaction-diffusion equations (the concentration of absorbed molecules being zero at the receiver's location). On the other hand, the reactive receiver adds an ordinary differential equation to the system's equations, for receptors on the surface of the receiver.
Further, one could study the design of optimal waveforms when taking multiple samples in the transmission time slot (instead of just one sample, as in this paper) or study optimal waveform design for channels with ISI. These studies are possible and left as future work. Finally, there are other existing approaches in the literature for obtaining semianalytic solutions of chemical reaction-diffusion equations. Studying these approaches which might have a higher radius of convergence is also left as future work.
\appendices
\section{Convergence Analysis}\label{AppA0 }
To show that the power series $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i}(x,t)$, $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{B}]_{i}(x,t)$ and $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{C}]_{i}(x,t)$ are convergent to functions that satisfy the original reaction-diffusion differential equation,
we consider time $t\in[0,T]$ for some fixed $T$ and prove that
\begin{align}\label{eqnNewN14}
&\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i}(x,t),\quad \sum_{i=0}^{\infty}\lambda^{i}
\frac{\partial}{\partial x} [\mathtt{A}]_{i}(x,t),\quad
\sum_{i=0}^{\infty}\lambda^{i}
\frac{\partial^2}{\partial x^2}[\mathtt{A}]_{i}(x,t),\quad\sum_{i=0}^{\infty}\lambda^{i}
\frac{\partial}{\partial t}[\mathtt{A}]_{i}(x,t)
\end{align}
uniformly converge over all $\lambda, x, t$. Similarly, we prove that the corresponding power series for molecules of types $\mathtt B$ and $\mathtt C$ also uniformly converge.
Uniform convergence of the series given in equations \eqref{eqnNewN14}
follows from Lemma \ref{LemmaBounds} given below. For a function $f_{\mathtt{A}}(x,t)$ defined for $x\in\mathbb{R}$ and $t\in[0,T]$, we define
\begin{align}&\lVert f_{\mathtt{A}}\lVert_{\infty}=\sup_{x,t\in[0,T]}|f_{\mathtt{A}}(x,t)|,
\lVert f_{\mathtt{A}}(x,0)\lVert_{\infty}=\sup_{x}|f_{\mathtt{A}}(x,0)|,
\lVert f_{\mathtt{A}}(x,t)\lVert_{1}=\int_{0}^{T}\int_{x}|f_{\mathtt{A}}(x,t)| dx dt.
\end{align}
\begin{lem}\label{LemmaBounds}
Let
\begin{align}
&M_0=\max\{T\lVert f_{\mathtt{A}}\lVert_{\infty},T\lVert f_{\mathtt{B}}\lVert_{\infty}\},
N_0=\max\big\{T\lVert \frac{\partial f_{\mathtt{A}}}{\partial t}\lVert_{\infty}+\lVert f_{\mathtt{A}}(x,0)\lVert_{\infty},
T\lVert \frac{\partial f_{\mathtt{B}}}{\partial t}\lVert_{\infty}+\lVert f_{\mathtt{B}}(x,0)\lVert_{\infty}\big\},\nonumber\\
&G_0=\max\{\sqrt{\frac{4T}{\pi D_\mathtt{A}}}\lVert f_{\mathtt{A}}\lVert_{\infty},\sqrt{\frac{4T}{\pi D_\mathtt{B}}}\lVert f_{\mathtt{B}}\lVert_{\infty}\},~
H_0=\max\{\sqrt{\frac{4T}{\pi D_\mathtt{A}}}\lVert \frac{\partial f_{\mathtt{A}}}{\partial x}\lVert_{\infty},\sqrt{\frac{4T}{\pi D_\mathtt{B}}}\lVert \frac{\partial f_{\mathtt{B}}}{\partial x}\lVert_{\infty}\},\nonumber\\
&\sigma=\max\{\rVert\frac{\partial\phi_\mathtt{A}}{\partial x}\lVert_{1},\rVert\frac{\partial\phi_\mathtt{B}}{\partial x}\lVert_{1},\rVert\frac{\partial\phi_\mathtt{C}}{\partial x}\lVert_{1}\}=\sqrt{\frac{4T}{\pi\min\{D_\mathtt{A},D_\mathtt{B},D_{\mathtt{C}}\}}}.
\end{align}
Then, for any $0<\lambda<\frac{1}{T(12M_0+10\gamma)}$, we have the following equations for any $i\geq 1$
\begin{align}
&\lambda^{i}\rVert[\mathtt{A}]_{i}\lVert_{\infty}\leq \frac{M_0}{2^{i}},~
\lambda^{i}\rVert\frac{\partial [\mathtt{A}]_{i}}{\partial t}\lVert_{\infty}\leq \frac{N_0}{3(2^{i})},~\lambda^{i}\rVert\frac{\partial [\mathtt{A}]_{i}}{\partial x}\lVert_{\infty}\leq \frac{\sigma M_0 i}{4T(2^i)},\\&\lambda^{i}\rVert\frac{\partial^2 [\mathtt{A}]_{i}}{\partial x^2}\lVert_{\infty}\leq(\frac{1}{2})^{i}
(\frac{\sigma(2M_0+\gamma) G_0}{4T(M_0+\gamma)}+\frac{\sigma^2(2M_{0}^{2}+
\gamma M_{0})}{32T^2(M_{0}+\gamma)}i(i-1)).
\end{align}
\end{lem}
Using this lemma and assuming $|\lambda|<\frac{1}{T(12M_0+10\gamma)}$ where $M_0$ is defined in Lemma \ref{LemmaBounds}, we obtain
$\sum_{i=N}^{\infty}
\lambda^{i}\rVert[\mathtt{A}]_{i}\lVert_{\infty}\leq M_0/2^{N-1}$.
Since $M_0/(2^{N-1})$ is a universal upper bound that does not depend on $(\lambda, x, t)$, the power series $\sum_{i=0}^{\infty}\lambda^{i}
[\mathtt{A}]_{i}(x,t)$ will uniformly converge. Proof of the convergence of the other power series given above is similar.
It remains to prove Lemma \ref{LemmaBounds}.
\begin{proof}[Proof of Lemma \ref{LemmaBounds}]
The proof is by induction on $i$.
Suppose $M_i,i\geq1$, is an upper bound on $\rVert[\mathtt{A}]_i\lVert_{\infty}$, $\lVert[\mathtt{B}]_i\rVert_{\infty}$, and $\lVert[\mathtt{C}]_i\rVert_{\infty}$ we have:
\begin{align}
\rVert[\mathtt{A}]_i\lVert_{\infty}&=\rVert \phi_\mathtt{A}**([\mathtt{A}]_{0:i-1}*^d[\mathtt{B}]_{0:i-1}-\gamma[\mathtt{C}]_{i-1})\lVert_{\infty}\leq\rVert\phi_{\mathtt{A}}\lVert_{1}(\sum_{j=0}^{i-1}\rVert [\mathtt{A}]_{j}\lVert_{\infty}
\rVert [\mathtt{B}]_{i-1-j}\lVert_{\infty}+\gamma\rVert [\mathtt{C}]_{i-1}\lVert_{\infty})\nonumber\\
&\leq
T (\sum_{j=0}^{i-1} M_jM_{i-1-j}+\gamma M_{i-1})\leq T(1+\frac{\gamma}{M_{0}})\sum_{j=0}^{i-1} M_jM_{i-1-j}
\end{align}
A similar bound for $\rVert[\mathtt{B}]_i\lVert_{\infty}$ and $\rVert[\mathtt{C}]_i\lVert_{\infty}$ can be written. Thus, using induction we can set $M_i=T(1+\frac{\gamma}{M_{0}}) \sum_{j=0}^{i-1} M_jM_{i-1-j}$. The solution of this recursive equation is given in the following form:
$ M_i=(T+\frac{\gamma T}{M_0})^iM_{0}^{i+1}\mathcal{C}_i
,~i\geq 0,
$ where $\mathcal{C}_i$ is the Catalan number and has an explicit formula:
$ \mathcal{C}_i=\frac{(2i)!}{(i+1)!(i)!}.
$ Using the Sterling formula, we have $
\mathcal{C}_i\leq 4^i.
$
Hence for $i\geq 0$ we obtain:
\begin{align}
\lambda^{i}\max(\rVert[\mathtt{A}]_i\lVert_{\infty}, \rVert[\mathtt{B}]_i\lVert_{\infty})&\leq M_0(\frac{4M_0+4\gamma}{12M_0+10\gamma})^i\leq \frac{M_0}{2^i}.
\end{align}
Suppose $N_i$ for $i\geq1$ is an upper bound on $\rVert\frac{\partial[\mathtt{A}]_i}{\partial t}\lVert_{\infty},\lVert
\frac{\partial[\mathtt{B}]_i}{\partial t}\rVert_{\infty}$ and $\lVert
\frac{\partial[\mathtt{C}]_i}{\partial t}\rVert_{\infty}$ .
We have:
\begin{align}
\rVert\frac{\partial[\mathtt{A}]_i}{\partial t}\lVert_{\infty}&=\rVert \phi_\mathtt{A}**\frac{\partial}{\partial t}([\mathtt{A}]_{0:i-1}*^d[\mathtt{B}]_{0:i-1}-\gamma [\mathtt{C}]_{i-1})\lVert_{\infty}\nonumber\\ &\leq \rVert\phi_{\mathtt{A}}\lVert_{1}(2\sum_{j=0}^{i-1} \rVert\frac{\partial[\mathtt{A}]_j}{\partial t}\lVert_{\infty}
\rVert [\mathtt{B}]_{i-1-j}\lVert_{\infty}+\gamma \rVert\frac{\partial[\mathtt{C}]_{i-1}}{\partial t}\lVert_{\infty})\nonumber\\&\leq (2T+\frac{\gamma T}{M_0})\sum_{j=0}^{i-1}N_{j}M_{i-1-j}\leq
(2TM_0+\gamma T) \sum_{j=0}^{i-1} N_j(4TM_0)^{i-1-j}.
\end{align}
A similar equation can be written for $\rVert\frac{\partial[\mathtt{B}]_i}{\partial t}\lVert_{\infty}$ and $\rVert\frac{\partial[\mathtt{C}]_i}{\partial t}\lVert_{\infty}$.
Thus, setting
$N_i=(2TM_0+\gamma T) \sum_{j=0}^{i-1} N_j(4TM_0)^{i-1-j}$ yields a valid upper bound on
$\rVert\frac{\partial[\mathtt{A}]_i}{\partial t}\lVert_{\infty},\lVert
\frac{\partial[\mathtt{B}]_i}{\partial t}\rVert_{\infty}$ and $\lVert
\frac{\partial[\mathtt{C}]_i}{\partial t}\rVert_{\infty}$
by induction.
The solution of this recursive equation is as follows:
\begin{equation}
N_i=N_0(\frac{2M_0+\gamma}{6M_0+5\gamma})(6TM_0+5T\gamma)^{i},~i\geq 1.
\end{equation}
For $0<\lambda<\frac{1}{T(12M_0+10\gamma)}$, we obtain that
\begin{equation}
\lambda^{i}\max(\rVert\frac{\partial [\mathtt{A}]_{i}}{\partial t}\lVert_{\infty},\rVert\frac{\partial [\mathtt{B}]_{i}}{\partial t}\lVert_{\infty}\big)\leq \frac{N_0}{3(2^{i})}.
\end{equation}
Next, we have
\begin{align}
\rVert\frac{\partial[\mathtt{A}]_i}{\partial x}\lVert_{\infty}&=\rVert \frac{\partial\phi_\mathtt{A}}{\partial x} **([\mathtt{A}]_{0:i-1}*^d[\mathtt{B}]_{0:i-1}-\gamma [\mathtt{C}]_{i-1})\lVert_{\infty}\nonumber\\&\leq \rVert\frac{\partial\phi_\mathtt{A}}{\partial x}\lVert_{1}(\sum_{j=0}^{i-1} \rVert[\mathtt{A}]_j\lVert_{\infty}
\rVert [\mathtt{B}]_{i-1-j}\lVert_{\infty}+\gamma \rVert [\mathtt{C}]_{i-1}\lVert_{\infty} )\leq \sigma(1+\frac{\gamma}{M_0})\sum_{j=0}^{i-1}M_{j}M_{i-1-j}\nonumber\\
&\leq
\frac{\sigma M_0}{4T} \sum_{j=0}^{i-1} (4T(M_0+\gamma))^{i}
\leq \frac{\sigma M_0}{4T}i
(4T(M_0+\gamma))^{i},
\end{align}
thus,
$
G_i=\frac{\sigma M_0}{4T}i
(4T(M_0+\gamma))^{i},
$
for $i\geq 1$, serves as an upper bound on
$\rVert\frac{\partial[\mathtt{A}]_i}{\partial x}\lVert_{\infty}$ (and by a similar argument $\lVert
\frac{\partial[\mathtt{B}]_i}{\partial x}\rVert_{\infty}$ and $\lVert
\frac{\partial[\mathtt{C}]_i}{\partial x}\rVert_{\infty}$ ).
Finally,
\begin{align}
\rVert\frac{\partial^2[\mathtt{A}]_i}{\partial x^2}\lVert_{\infty}&=\rVert
\frac{\partial \phi_\mathtt{A}}{\partial x}
**(\frac{\partial}{\partial x}([\mathtt{A}]_{0:i-1}*^d[\mathtt{B}]_{0:i-1}-\gamma [\mathtt{C}]_{i-1})\lVert_{\infty}
\nonumber\\&\leq \rVert\frac{\partial \phi_\mathtt{A}}{\partial x} \lVert_{1}(2\sum_{j=0}^{i-1} \rVert\frac{\partial[\mathtt{A}]_j}{\partial x}\lVert_{\infty}
\rVert [\mathtt{B}]_{i-1-j}\lVert_{\infty}+
\gamma \rVert\frac{\partial[\mathtt{C}]_{i-1}}{\partial x}\lVert_{\infty}
)\leq\sigma(2\sum_{j=0}^{i-1}
G_jM_{i-1-j}+\gamma G_{i-1})
\nonumber\\&\leq\sigma(2+\frac{\gamma}{M_0}) \sum_{j=0}^{i-1}
G_jM_{i-1-j}=\sigma(2+\frac{\gamma}{M_0})(G_0 M_{i-1}+\sum_{j=1}^{i-1} G_jM_{i-1-j})\nonumber\\&\leq(\frac{\sigma(2M_0+\gamma) G_0}{4T(M_0+\gamma)}+\frac{\sigma^2(2M_{0}^{2}+
\gamma M_{0})}{32T^2(M_{0}+\gamma)}i(i-1))\times
(4T(M_0+\gamma))^{i}.
\end{align}
Therefore,
$
H_i=(\frac{\sigma(2M_0+\gamma) G_0}{4T(M_0+\gamma)}+\frac{\sigma^2(2M_{0}^{2}+
\gamma M_{0})}{32T^2(M_{0}+\gamma)}i(i-1))\times
(4T(M_0+\gamma))^{i}, i\geq 1,
$
is an upper bound on $\rVert\frac{\partial^2[\mathtt{A}]_i}{\partial x^2}\lVert_{\infty}$ (an on $\lVert
\frac{\partial^2[\mathtt{B}]_i}{\partial x^2}\rVert_{\infty}$ and $\lVert
\frac{\partial^2[\mathtt{C}]_i}{\partial x^2}\rVert_{\infty}$ by a similar argument). The proof is complete.
\end{proof}
\iffalse
\section{Increasing Radius of Convergence }\label{AppA:1 }
For simplicity, assume $\gamma=0$.
We can extend our solution for given$(\lambda,T,f_{\mathtt A}(x,t),f_{\mathtt B}(x,t))$. If $\lambda\leq 1/(12TM_0)$ we proved in Appendix \ref{AppA0 } of paper, the solution is convergent($M_0=max\{\lVert \phi_{\mathtt A}**f_{\mathtt A} \rVert_{\infty},\lVert \phi_{\mathtt B}**f_{\mathtt B} \rVert_{\infty}\}$).
Let $\lambda_{0}=1/(12TM_0)$ and $[\mathtt A]^{\lambda_0}$ and $[\mathtt B]^{\lambda_0}$ be solution of reaction-diffusion equation for $\lambda=\lambda_0$. Now for $\lambda>\lambda_{0}$,consider the solution as $[\mathtt{A}]=\sum_{i=0}^{\infty}(\lambda-\lambda_0)^{i}[\mathtt{A}]_{i}$ and $[\mathtt{B}]=\sum_{i=0}^{\infty}(\lambda-\lambda_0)^{i}[\mathtt{B}]_{i}$. By substituting these form in reaction-diffusion equations, we have :
\begin{align}
&[\mathtt A]_0=[\mathtt A]^{\lambda_0},~~[\mathtt B]_0=[\mathtt B]^{\lambda_0},\\
&[\mathtt A]_i=-\phi_{\mathtt A}**([\mathtt A]_{0:i-1}*^{d}[\mathtt B]_{0:i-1}) ~,i\geq 1,\\
&[\mathtt B]_i=-\phi_{\mathtt B}**([\mathtt A]_{0:i-1}*^{d}[\mathtt B]_{0:i-1}) ~,i\geq 1.
\end{align}
The solution are convergent if $\lambda-\lambda_0\leq 1/(12TM')$, where $M'=max\{\lVert [\mathtt A]^{\lambda_0} \rVert_{\infty},\lVert [\mathtt B]^{\lambda_0} \rVert_{\infty}\}$. In the following lemma we show $M'\leq M_0$.
\begin{lem}
$M'\leq M$
\end{lem}
\begin{proof}
It is suffient to prove $[\mathtt A](x,t)\leq \phi_{\mathtt A}**f_\mathtt{A}$. Concentration of molecules are non negative functions, i.e.,$[\mathtt A](x,t),[\mathtt B](x,t)\geq 0$, we have,
\begin{align*}
& [\mathtt A](x,t)=\phi_{\mathtt A}**(f_{\mathtt A}-\lambda [\mathtt A][\mathtt B])=\phi_{\mathtt A}**f_{\mathtt A}-\lambda \phi_{\mathtt A}** [\mathtt A][\mathtt B]\leq \phi_{\mathtt A}**f_{\mathtt A},
\end{align*}
similarly $[\mathtt B](x,t)]\leq \phi_{\mathtt B}**f_{\mathtt B} $, hence $M'\leq M$.
\end{proof}
Therefore, in the worse case $\lambda-\lambda_0\leq 1/(12TM_0)=\lambda_0$, so we have $\lambda\leq 2\lambda_0$. For $\lambda=k\lambda_0+r(r<\lambda_0)$ ,we repeat $k+1$ times this procedure. We proved that for any inputs one can compute solution by repeating perturbation method.
\fi
\iffalse
\section{Low Rate Region for Especial reaction-diffusion Equation}\label{AppA:2 }
For $g(\lambda)=\sum_{i=0}^{\infty}a_i\lambda^{i}$,
and given $\alpha$, we define two $n-$th approximations as below:
\begin{align}
&Approximation~1:~g(\lambda)\approx \sum_{i=0}^{n}a_i\lambda^{i}~~if~~ \lambda
\lVert \frac{a_{n+1}}{a_n} \rVert_{\infty}\leq \alpha,\\
&Approximation~2:~g(\lambda)\approx \sum_{i=0}^{n}a_i\lambda^{i}~~if~~
\lVert \frac{\sum_{i=n+1}^{\infty}a_i\lambda^{i}}{\sum_{i=0}^{n}a_i\lambda^{i}} \rVert_{\infty}\leq \alpha.
\end{align}
For given $\alpha$ and observation time $T$, if $\lambda\leq\frac{3\alpha}{T^2}$, we have first kind of approximation of order $1$. For
For given $n$ and $\alpha$, we say $\lambda$ is low rate of order $n$ if $\lambda$ satisfies approximation $1$ or $2$.
The common values for $\alpha$ are $.1,.05$.
In this paper we approximate solution up tp first term.
Unfortunately, in chemical reaction-diffusion, derivation explicit formula for low rate region of solution is not tractable. In other words for given
arbitrary $f_{\mathtt{A}},f_{\mathtt{B}},D_{\mathtt{A}},D_\mathtt{B}$ and $\alpha$ the range of low rate solution can be obtained only by simulation.
Here we consider very special case.
For given $\alpha$ and observation time $T$, if $\lambda\leq\frac{3\alpha}{T^2}$, we have first kind of approximation of order $1$. For second kind of approximation of order $1$, consider
$I=\sum_{i=1}^{\infty}\lambda^{i}[\mathtt{A}]_i=\sum_{i=0}^{\infty}\lambda^{i}(-1)^{i}\alpha_{i}t^{2i+1}=I_1-I_2$, where $I_1=\sum_{i=1}^{\infty}\alpha_{2i}\lambda^{2i}t^{4i+1}$ and $I_2=\sum_{i=0}^{\infty}\alpha_{2i+1}\lambda^{2i+1}t^{4i+3}$. Using upper and lower bound on $\alpha_{i}$, for $\lambda\leq\frac{2}{T^2}$ we have:
\begin{align*}
&\dfrac{\lambda^2t^5}{9-\lambda^2t^4}\leq I_1\leq
\dfrac{\lambda^2t^5}{4-\lambda^2t^4}\\
&\dfrac{3\lambda t^3}{9-\lambda^2t^4}\leq I_2\leq
\dfrac{2\lambda t^3}{4-\lambda^2t^4},
\end{align*}
we obtain:
\begin{align*}
&\dfrac{\lambda^2t^4}{9-\lambda^2t^4}-\dfrac{2\lambda t^2}{4-\lambda^2t^4}\leq \frac{I}{[\mathtt{A}]_{0}}\leq
\dfrac{\lambda^2t^4}{4-\lambda^2t^4}-\dfrac{3\lambda t^2}{9-\lambda^2t^4}.
\end{align*}
Let $u=\lambda t^2$, we have:
\begin{align*}
&g_d(u)=\dfrac{u^2}{9-u^2}-\dfrac{2u}{4-u^2}\leq \frac{I}{[\mathtt{A}]_{0}}\leq
g_{up}(u)=\dfrac{u^2}{4-u^2}-\dfrac{3u}{9-u^2}.
\end{align*}
For given $\alpha$, we obtain two fourth order equations($g_{up}(u)=\pm \alpha, g_{d}(u)=- \alpha $). By solving those equations, we can obtain upper bound on $\lambda$. For example for $\alpha=0.1$ we have $\lambda\leq \frac{1}{5T^2}$, and for $\alpha=0.01$ we obtain $\lambda\leq \frac{1}{50T^2}$.
\fi
\section{Proof of Theorem \ref{thmN1} }\label{AppA:3}
The zero-order terms in the perturbation method satisfy
\begin{align}
&\frac{\partial [\mathtt{A}]_{0}}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]_{0}}{{\partial x}^{2}}+f_\mathtt{A}(x,t),~~
\frac{\partial [\mathtt{B}]_{0}}{\partial t}=D_\mathtt{B}~\frac{\partial^{2} [\mathtt{B}]_{0}}{{\partial x}^{2}}+f_\mathtt{B}(x,t),~~
\frac{\partial [\mathtt{C}]_{0}}{\partial t}=D_\mathtt{C}~\frac{\partial^{2} [\mathtt{C}]_{0}}{{\partial x}^{2}},\\
&[\mathtt{A}]_{0}(x,0)=g_{\mathtt{A}}(x), ~[\mathtt{B}]_{0}(x,0)=g_{\mathtt{B}}(x), ~[\mathtt{C}]_{0}(x,0)=g_{\mathtt{C}}(x).
\end{align}
Setting $t=0$ in above equations we have:
\begin{align}
&\frac{\partial [\mathtt{A}]_{0}}{\partial t}(x,0)=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]_{0}(x,0)}{{\partial x}^{2}}+f_\mathtt{A}(x,0)=D_\mathtt{A}
g_{\mathtt{A}}''(x)+f_\mathtt{A}(x,0).
\end{align}
Thus,
$[\mathtt{A}]_{0}(x,t )=[\mathtt{A}]_{0}(x,0)+\frac{\partial [\mathtt{A}]_{0}}{\partial t}(x,0)t +o(t )=g_{\mathtt{A}}(x)+t D_\mathtt{A}
g_{\mathtt{A}}''(x)+t f_\mathtt{A}(x,0)+o(t ).
$ For other terms in perturbation solution we have ($i\geq 1$):
\begin{align}
&\frac{\partial[\mathtt{A}]_{i}}{\partial t}=D_\mathtt{A}~\frac{\partial^{2}[\mathtt{A}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1},~\quad[\mathtt{A}]_{i}(x,0)=0,\\
&\frac{\partial[\mathtt{B}]_{i}}{\partial t}=D_\mathtt{B}~\frac{\partial^{2}[\mathtt{B}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}+\gamma[\mathtt{C}]_{i-1},~\quad[\mathtt{B}]_{i}(x,0)=0,\\
&\frac{\partial[\mathtt{C}]_{i}}{\partial t}=D_\mathtt{C}~\frac{\partial^{2}[\mathtt{B}]_{i}}{{\partial x}^{2}}-[\mathtt{A}]_{0:i-1}*^d
[\mathtt{B}]_{0:i-1}-\gamma[\mathtt{C}]_{i-1},~\quad[\mathtt{C}]_{i}(x,0)=0.
\end{align}
Setting $t=0$ in above equations we obtain:
\begin{align}
\frac{\partial[\mathtt{A}]_{1}}{\partial t}(x,0)=
-g_{\mathtt{A}}(x)g_{\mathtt{B}}(x)+\gamma g_{\mathtt{C}}(x),\qquad
\frac{\partial[\mathtt{A}]_{i}}{\partial t}(x,0)=0,\quad~~i\geq 2.
\end{align}
From the Taylor series expansion we have:
\begin{align}
&[\mathtt{A}]_{1}(x,t )=[\mathtt{A}]_{1}(x,0)+\frac{\partial [\mathtt{A}]_{1}}{\partial t}(x,0)t +o(t )=-t g_{\mathtt{A}}(x)g_{\mathtt{B}}(x)+t \gamma g_{\mathtt{C}}(x)+o(t ),\\
&[\mathtt{A}]_{i}(x,t )=[\mathtt{A}]_{i}(x,0)+
\frac{\partial [\mathtt{A}]_{i}}{\partial t}(x,0)t +
\frac{\partial^{2}[\mathtt{A}]_{i}}{{\partial t}^2}(x,0)t ^2+o(t ^2)=\mathcal{O}(t ^2),~i\geq 2.
\end{align}
Therefore the perturbation solution is as follows:
\begin{align}
&[\mathtt{A}]_{PER}(x,t )=[\mathtt{A}]_{0}(x,t )+\lambda[\mathtt{A}]_{1}(x,t )+\sum_{i=2}^{\infty}\lambda^i[\mathtt{A}]_{i}(x,t )\nonumber\\
&=g_{\mathtt{A}}(x)+t D_\mathtt{A}
g_{\mathtt{A}}''(x)+t f_\mathtt{A}(x,0)+\lambda(-t g_{\mathtt{A}}(x)g_{\mathtt{B}}(x)+t \gamma g_{\mathtt{C}}(x))+o(t )+\mathcal{O}(t^2).
\end{align}
Hence,
$
[\mathtt{A}]_{PER}(x,t )=[\mathtt{A}]_{FDM}(x,t )+\mathcal{O}(t^2).$
\section{Poisson Hypothesis Testing }\label{AppB}
In a Poisson hypothesis testing problem with two hypotheses, we take a sample from a Poisson random variable $X$ whose mean is either $\rho_0$ (under hypothesis $\textbf{H}_{0}$) or $\rho_1$ (under hypothesis $\textbf{H}_{1}$).
Assume that the two hypotheses $\textbf{H}_{0}$ and $\textbf{H}_{1}$ are equally likely and $\rho_0<\rho_1$. The MAP decision rule compares the observed $X$ with threshold $\mathbb{T}_{h}=({\rho_1-\rho_0})/({\log{\rho_1}-\log{\rho_0}})$ and declares $\textbf{H}_{0}$ if and only if $X<\mathbb{T}_{h}$. The error probability of the MAP decision rule, denoted by $P_e(\rho_0, \rho_1)$, equals
\begin{align}
\frac{1}{2}\sum_{n\in\mathbb{Z}:~n\geq \mathbb{T}_{h}}^{\infty}\frac{e^{-\rho_0}\rho_{0}^{n}}{n!}+
\frac{1}{2}\sum_{n\in\mathbb{Z}:~0\leq n<\mathbb{T}_{h}}\frac{e^{-\rho_1}\rho_{1}^{n}}{n!}
=\frac{1}{2}-TV(\mathsf{Poisson}(\rho_0),\mathsf{Poisson}(\rho_1))
\end{align}
where
$TV(\cdot,\cdot))$ is the total variation distance.
\begin{lem}\label{decrese:error:pro}
Fix some $\rho_0$. Then $P_e(\rho_0,\rho_1)$ is a decreasing continuous function of $\rho_1$ for $\rho_1\geq \rho_0$.
\end{lem}
\begin{proof}
The above statement is equivalent with
$TV(\mathsf{Poisson}(\rho_0),\mathsf{Poisson}(\rho_1))$ being an increasing continuous function of $\rho_1$ for $\rho_1\geq \rho_0$. When $\rho_1=\rho_0$, the total variation distance is zero and as we increase $\rho_1$, this distance increases. Since the distribution of $\mathsf{Poisson}(\rho_1)$ varies continuously, the changes in the total variation distance is also continuous in $\rho_1$.
\end{proof}
\section{Support Lemma}\label{AppC}
Let $\mathbb{P}$ be the space of all unnormalized probability distributions on the interval $[0,T]$ (\emph{i.e.}, nonnegative functions with finite nonzero integral). For a distribution $p(t)\in\mathbb{P}$ and a continuous function $f(t)$, we define $
\mathbb{E}_{p}(f)=\int_{0}^{T}
p(t)f(t) dt.
$
\begin{lem} \label{SupportLemma} Take arbitrary continuous functions $f(t)$ and $f_i(t)$ for $i=1,2,\cdots, n$, and an arbitrary unnormalized distribution $p\in\mathbb{P}$. Then, there is another \emph{discrete} unnormalized distribution $q\in\mathbb{P}$
taking values
in a set of size $n$, \emph{i.e.,}
$q(t)=\sum_{i=1}^{n}{a}_{i}\delta(t-t_i)$
for some $a_i\geq 0$ and $t_i\in[0,T]$
such that
\begin{align}\mathbb{E}_{q}(f)\leq \mathbb{E}_{p}(f)\label{eqn:C12}\end{align}
and
\begin{align}\mathbb{E}_{q}(f_i)= \mathbb{E}_{p}(f_i), \quad \text{for}~~i=1,\cdots, n.\label{eqn:Cn}\end{align}
\end{lem}
This lemma shows that by replacing $p$ with $q$, we preserve the $n$ linear constraints \eqref{eqn:Cn} and impose one linear inequality constraint \eqref{eqn:C12}. The support of $q$ (the number of delta functions) is at most the number of constraints which is $n$.
Support lemmas of this type are commonly used in information theory, and follow from Fenchel-Bunt's extension of the Caratheodory theorem. For completeness, we give an intuitive sketch of the proof. For simplicity assume that $p(t)$ is discrete but with support on an arbitrarily large set, \emph{i.e.,}
$p(t)=\sum_{i=1}^{N}g_{i}\delta(t-\tilde{t}_i)$
for some large $N$, and $g_i\geq 0$, $\tilde{t}_i\in[0,T]$. Consider functions of the form
$\tilde{q}(t)=\sum_{i=1}^{N}x_{i}\delta(t-\tilde{t}_i)
$ for some $x_1,\cdots, x_N\geq 0$. Consider the set of $(x_1, \cdots, x_N)$ for which we have $\mathbb{E}_{\tilde q}(f_i)= \mathbb{E}_{p}(f_i)$, for $i=1,\cdots, n$. This imposes $n$ linear constraints on $(x_1, \cdots, x_n)$. These $n$ linear constraints along with the inequality constraints $x_1, \cdots, x_N\geq 0$ define a polytope in the $N$-dimensional region. This polytope is nonempty since $(x_1, \cdots, x_N)=(g_1, \cdots, g_N)$ belongs to this polytope. To enforce the inequality \eqref{eqn:C12}, let us minimize $\mathbb{E}_{\tilde q}(f)$, which is a linear function in $(x_1, \cdots, x_N)$ over this polytope. The minimum of a linear function occurs at a vertex of the polytope. The key observation is that each vertex of the polytope has at most $n$ nonzero entries, \emph{i.e.,} if $(x^*_1, \cdots, x^*_N)$ is a vertex of the polytope, at most $n$ entries of $(x^*_1, \cdots, x^*_N)$ are nonzero. This would imply the desired identification for $q(x)$. To see this, observe that since the polytope is in $N$ dimensions, every vertex of the polytope lies at the intersection of $N$ hyperplanes. The hyperplanes defining the polytope are the $n$ linear constraints along with $x_1,\cdots, x_N\geq 0$. Any vertex has to satisfy $N$ of these equations with equality. Thus, the vertex needs to pick at least $N-n$ inequalities of the form $x_i\geq 0$ and satisfy them with equality. In other words, for any vertex, at least $N-n$ entries must be zero, meaning that the number of nonzero entries is at most $n$.
\iffalse
\section*{Acknowledgment}
The authors would like to thank...\cite{6189388}
\fi
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\section{Further Examples}\label{ex2,3}
\subsection{Example 2: Reaction for Channel Amplification}
The main example in Section \ref{sec::generalperturbation} used chemical reaction as a means to produce molecules that are detected by the receiver. However, reaction may be used for other purposes as well. For instance, it may be used to enhance the channel between the transmitter and the receiver. This concept is considered in the example below.
Consider the following example with one transmitter and one receiver. The receiver can only measure the density of molecules of type $\mathtt{A}$ at its location. The transmitter is also able to release molecules of types $\mathtt{A}$ and $\mathtt{B}$ into the medium. Assume that there is an enzyme $\mathtt{C}$ in the medium (outside of our control) which reacts with molecules of type $\mathtt{A}$. If the level of enzyme $\mathtt{C}$ is high, molecules of type $\mathtt{A}$ are mostly dissolved before reaching the receiver. To overcome this, the transmitter may release molecules of a different type $\mathtt{B}$, which would also react with the enzyme $\mathtt{C}$ and thereby reduce the concentration of $\mathtt{C}$ in the medium. This ``cleaning" of the medium from molecules of type $\mathtt{C}$ would enhance the channel from the transmitter to the receiver. More specifically, assume that the medium is governed by the following chemical reactions:
\begin{align}
\ce{\mathtt{A} + \mathtt{C}&->[\lambda_1]\mathtt{P}_1},\label{eqnNew3}\\
\ce{\mathtt{B} + \beta \mathtt{C} &->[\lambda_2] \mathtt{P}_2},\label{eqnNew4}
\end{align}
where $\mathtt{P}_1$ and $\mathtt{P}_2$ are some products which do not include molecules of type $\mathtt{A},\mathtt{B}$ or $\mathtt{C}$. Here $\beta$ is a natural number. As an example, $\mathtt{A}$ and $\mathtt{B}$ can be two different acids, and $\mathtt{C}$ is a base substance (which reacts with acids $\mathtt{A}$ and $\mathtt{B}$). If $\mathtt{B}$ is a stronger acid than $\mathtt{A}$, the coefficient $\beta$ can be large, and release of $\mathtt{B}$ can be effective in canceling $\mathtt{C}$ from the medium.
In \eqref{eqnNew3} and \eqref{eqnNew4}, let $\gamma=\lambda_2/\lambda_1$ be the ratio of the reaction rates.
For simplicity of notation, set $\lambda_1=\lambda,\lambda_2=\gamma\lambda$. Assume that the reaction occurs in a two-dimensional medium. For our modeling purposes, suppose that
the transmitter is located at the origin and the receiver is located at $(d,0)$. There is also an independent source that releases molecule of type $\mathtt{C}$ in the medium with a known concentration $f_{\mathtt{c}}(x,y,t)$ at time $t$. The source is assumed to be located at location $r_0$. See Fig \ref{fig:ex2} for a depiction of this setting. Following equations describe the system dynamic.
\begin{align}
&\frac{\partial[\mathtt{A}]}{\partial t}=D_\mathtt{A}~\bigtriangledown^2 [\mathtt{A}]-\lambda~[\mathtt{A}]
[\mathtt{C}]+f_{\mathtt{A}}(x,y,t),\label{cartesian2dim:1}\\
&\frac{\partial[\mathtt{C}]}{\partial t}=D_\mathtt{C}~\bigtriangledown^2 [\mathtt{C}]-\lambda~([\mathtt{A}][\mathtt{C}]+\gamma[\mathtt{B}][\mathtt{C}]^{\beta})+f_{\mathtt{C}}(x,y,t),\label{cartesian2dim:2}\\
& \frac{\partial[\mathtt{B}]}{\partial t}=D_\mathtt{B}~\bigtriangledown^2 [\mathtt{B}]-\gamma\lambda~[\mathtt{B}]
[\mathtt{C}]^{\beta}+f_{\mathtt{B}}(x,y,t),\label{cartesian2dim:3}
\end{align}
where $f_{\mathtt{A}}(x,y,t),f_{\mathtt{B}}(x,y,t)$ are input signals of the transmitter.
The initial conditions for concentration of molecule of type $\mathtt{A}$ and type $\mathtt{B}$ are set to zero.
For molecule of type $\mathtt{C}$, the initial condition is $[\mathtt{C}](x,y,0)=\mathcal{I}_{\text{int}}(x,y)$. We assume that $\mathcal{I}_{\text{int}}(x,y)$ is completely known. We assume that the diffusion occurs in the entire $\mathbb{R}^2$ and do not assume any boundaries for the medium.
For the case of no reaction, $\lambda=0$, the system of equations has a closed-form solution.
Consider a solution of the following form:
\begin{align}\label{teylorex:2}
&[\mathtt{A}](x,y,t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{A}]_{i}(x,y ,t),~
[\mathtt{C}](x,y,t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{C}]_{i}(x,y ,t),~
[\mathtt{B}](x,y,t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{B}]_{i}(x,y ,t).
\end{align}
By substituting \eqref{teylorex:2} in
\eqref{cartesian2dim:1}, \eqref{cartesian2dim:2}, \eqref{cartesian2dim:3}, and
matching the coefficients of $\lambda^0,\lambda^1,\lambda^2$ on both side of equations, we obtain the following equations:
For particle $\mathtt{A}$ we have:
\begin{align}
&\frac{\partial[\mathtt{A}]_0}{\partial t}=D_\mathtt{A}~\nabla^2 [\mathtt{A}]_0+f_{\mathtt{A}}(x,y,t),~
\frac{\partial[\mathtt{A}]_1}{\partial t}=D_\mathtt{A}~\nabla^2 [\mathtt{A}]_1-[\mathtt{A}]_0 [\mathtt{C}]_0,\label{eq:ex2:A:01}\\
&\frac{\partial[\mathtt{A}]_2}{\partial t}=D_\mathtt{A}~\nabla^2 [\mathtt{A}]_2-([\mathtt{A}]_0 [\mathtt{C}]_1+[\mathtt{A}]_1 [\mathtt{C}]_0).\label{eq:ex2:A:2}
\end{align}
\newpage
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=.5]{slide2.png}
\caption{MC system for Example $2$.}
\label{fig:ex2}
\vspace{-3em}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[trim={1cm 0cm 0cm 0cm}, scale=0.5]{slide3.png}
\caption{MC system for Example $3$.}
\label{fig:ex3}
\end{figure}
For molecule of type $\mathtt{C}$ we have:
\begin{align}\label{eq:ex2:C}
& \frac{\partial[\mathtt{C}_0]}{\partial t}=D_\mathtt{C}~\nabla^2 [\mathtt{C}]_0+f_{\mathtt{C}}(x,y,t),~~
\frac{\partial[\mathtt{C}]_1}{\partial t}=D_\mathtt{C}~\nabla^2 [\mathtt{C}]_1-(\gamma[\mathtt{C}]_{0}^{\beta} [\mathtt{B}]_0+[\mathtt{C}]_{0}[\mathtt{A}]_0).
\end{align}
To impose the initial condition for concentration of molecule of type $\mathtt{C}$,
We set $[\mathtt{C}]_{0}(x,y,0)=\mathcal{I}_{\text{int}}(x,y)$ and for $i\geq 1$ we set $[\mathtt{C}]_{i}(x,y,0)=0$.
For particle $\mathtt{B}$ we have:
\begin{align}\label{eq:ex2:B}
&\frac{\partial[\mathtt{B}]_0}{\partial t}=D_\mathtt{B}~\nabla^2 [\mathtt{B}]_0+f_{\mathtt{B}}(x,y,t).
\end{align}
Let $\phi_{\mathtt{A}}(x,y,t)$ be the impulse response of heat equation with $D_\mathtt{A}$,
\begin{align}
\frac{\partial \phi_\mathtt{A}}{\partial t}=D_\mathtt{A}~\nabla^2(\phi_\mathtt{A})+\delta(x)\delta(y)\delta(t).
\end{align}
we have:
\begin{align}\label{greentwo}
\phi_\mathtt{A}(x,y,t)=\frac{1}{4\pi D_\mathtt{A} t}\exp(-\frac{x^2+y^2}{4 D_\mathtt{A} t}),~~t\geq 0.
\end{align}
Similarly, $\phi_{\mathtt{B}}(x,y,t)$ and $\phi_{\mathtt{C}}(x,y,t)$ are the impulse responses of the heat equations with diffusion coefficients $D_\mathtt{B}$ and $D_\mathtt{C}$ respectively. The solutions of \eqref{eq:ex2:A:01},\eqref{eq:ex2:A:2},\eqref{eq:ex2:C}, and \eqref{eq:ex2:B} are:
\begin{align}
&[\mathtt{A}]_0=\phi_\mathtt{A}**f_{\mathtt{A}},~[\mathtt{C}]_0=
\phi_\mathtt{C}**(f_{\mathtt{C}}+D_{\mathtt{C}}\nabla^{2}\mathcal{I}_{\text{int}})+\mathcal{I}_{\text{int}},
~[\mathtt{B}]_0=\phi_\mathtt{B}**f_{\mathtt{B}},\label{ex2:final:sol:1}\\&[\mathtt{A}]_1=- \phi_\mathtt{A}**([\mathtt{A}]_0 [\mathtt{C}]_0),
~[\mathtt{C}]_1=-\phi_\mathtt{C}**(\gamma[\mathtt{B}]_0 [\mathtt{C}]_0^{\beta}+[\mathtt{C}]_0 [\mathtt{A}]_0),\label{ex2:final:sol:2}\\
&[\mathtt{A}]_2=-\phi_{\mathtt{A}}**([\mathtt{A}]_0 [\mathtt{C}]_1+ [\mathtt{A}]_1 [\mathtt{C}]_0).\label{ex2:final:sol:3}
\end{align}
The density of molecule $\mathtt{A}$ is equal to:
\begin{align}\label{ex2:Asol}
[\mathtt{A}](x,y,t)= [\mathtt{A}]_0+\lambda [\mathtt{A}]_1+\lambda^{2} [\mathtt{A}]_2 +\mathcal{O}(\lambda^3)
\end{align}
For low reaction rates, we can approximate \eqref{ex2:Asol} as follows:
\begin{align}\label{ex2:Asol2}
[\mathtt{A}](x,y,t)\approx [\mathtt{A}]_0+\lambda [\mathtt{A}]_1+\lambda^{2} [\mathtt{A}]_2(x,y,t)
\end{align}
Using the above model, we design a modulation scheme in Section \ref{mod:ex2}.
\subsection{Modulation Design For Example 2}\label{mod:ex2}
Consider a communication scenario consisting of a transmitter and a receiver. The transmitter releases molecules of type $\mathtt{A}$ and $\mathtt{B}$ into the medium to encodes a message $M_{\mathtt{A}}\in\{0,1\}$. The concentration of the released molecules of type $\mathtt{A}$ is described by the input signal $f_{\mathtt{A}}^{i}(x,y,t) =a_i(t)\delta(x)\delta(y)$,$0\le t\le T$ for $i=0,1$. In other words, the density of released molecules of type $\mathtt{A}$ at time $t$ is $a_{M_{\mathtt{A}}}(t)$ if message $M_{\mathtt{A}}$ is transmitted for $t\in[0,T]$. Similarly, $f_{\mathtt{B}}(x,y,t)=b_{M_{\mathtt{A}}}(t)\delta(x)\delta(y)$ is the released concentration of molecules of type $\mathtt{B}$ where $b_0(t)$ and $b_1(t)$ are two nonnegative waveforms for $t\in[0,T]$.
The total amount of released molecules of type $\mathtt A ,\mathtt B $ during the transmission period $T$ is assumed to be at most $s_{\mathtt A},s_{\mathtt B}$ respectively, \emph{i.e.,}
\begin{align}\int_{0}^T a_i(t)dt\leq s_{\mathtt A},~\int_{0}^T b_i(t)dt\leq s_{\mathtt B} \quad i=0,1.\label{eqn:power2a}
\end{align}
Finally, we assume that molecules of type $\mathtt C$ are being continuously generated throughout the medium according to some function $f_{\mathtt{C}}(x,y,t)$ for $x,y\in\mathbb{R}$, $t\in[0,T]$. The function $f_{\mathtt{C}}(x,y,t)$ and the initial density $[\mathtt{C}](x,y,0)$ at time $t=0$ are assumed to be completely known to the receiver and the transmitter.
Receiver samples the number of molecules of type $\mathtt{A}$ at $(x=0,y=d,t=T)$. Similar to Example 1, receiver gets a number from a Poisson distribution with parameter $[\mathtt{A}](0,d,T)$.
As before, the probability of error is defined as
\begin{equation}
Pr(e)=Pr\{ M_{\mathtt{A}}\neq \hat{M}_\mathtt{A}\},
\end{equation}
where $\hat{M}_\mathtt{A}\in\{0,1\}$ is the receiver's decoded message bit. Similar to Example 1, we wish to minimize the error probability by choosing the best possible waveforms $a_i(t)$ and $b_j(t)$.
The following theorem states that in the low reaction rate regime (for the approximate reaction-diffusion equations when we consider first terms in the Taylor series expansion), one possible optimal waveforms $a_i(t)$ and $b_i(t)$ that achieve minimum probability of error are as follows:
\begin{theorem}\label{th::2}
For any given noise source $f_{\mathtt{C}}(x,y,t)$, and any arbitrary initial condition $\mathcal{I}_{int}(x,y)$, one choice for optimal order one waveforms $a_i(t),b_i(t)$ are as follows:
\begin{align}
&a_0(t)=0,~a_1(t)= s_\mathtt{A}\delta(t-t^{[a_1]}),~b_0(t)=0,~b_1(t)=s_{\mathtt{B}}\delta(t-t^{[b_1]}).
\end{align}
for some $t^{[a_1]},t^{[b_1]}\in [0,T]$ which depend on
$f_{\mathtt{C}}(x,y,t)$, and $\mathcal{I}_{int}(x,y)$.
\end{theorem}
\begin{proof}\label{AppF}
Substituting $f_\mathtt{A}^{i}(x,t),f_\mathtt{B}^{i}(x,t),i=0,1$ in \eqref{recu-eq1}-\eqref{recu-eq3}, one obtains an explicit formula for the (first-order approximation of the) concentration of molecules $\mathtt{A}$ in terms of the input signals. In particular, the concentration $[\mathtt{A}]$ at receiver's location at the sampling time $T$ is as follows:
\begin{equation*}\label{rho::ex2}
\begin{aligned}
[\mathtt{A}](d,0,T)=&[\mathtt{A}]_{0}(d,0,T)+\lambda
[\mathtt{A}]_{1}(d,0,T)+\lambda^2[\mathtt{A}]_{2}(d,0,T).\\
\end{aligned}
\end{equation*}
Observe from \eqref{ex2:final:sol:1}, \eqref{ex2:final:sol:2}, and \eqref{ex2:final:sol:3} that $[A]_{0}(x,y,t)$ and $[A]_{1}(x,y,t)$ are linear functions of $a_{M_{\mathtt{A}}}(t)$ and do not depend on $b_{M_{\mathtt{A}}}(t)$. However, $[A]_{2}(x,y,t)$ depends on both $a_{M_{\mathtt{A}}}(t)$ and $b_{M_{\mathtt{A}}}(t)$. It is a nonlinear function of $a_{M_{\mathtt{A}}}(t)$. However, if we fix $a_{M_{\mathtt{A}}}(t)$, $[A]_{2}(x,y,t)$ is a linear and positive function in $b_{M_{\mathtt{A}}}(t)$.
Let
\begin{equation}\label{densitiees::ex2:final}
\rho=\left\{ {\begin{array}{*{20}{llll}}
\rho_{0}=[\mathtt{A}](d,0,T)& & \text{if}&M_{\mathtt{A}}=0\\
\rho_{1}=[\mathtt{A}](d,0,T)& & \text{if}&M_{\mathtt{A}}=1.\\
\end{array}}\right.
\end{equation}
Receiver is assumed to be transparent and due to the particle counting noise, it observes a sample from
$\mathsf{Poisson}(V\rho_{
M_{\mathtt{A}}})
$
for some constant $V$.
Since the effect of molecule $\mathtt{B}$ appears in $\lambda^{2}$ coefficient in the expansion $[\mathtt{A}]=\sum_{=i=0}^{\infty}\lambda^{i}[A]_{i}(x,y,t)$ (i.e, $[\mathtt{A}]_{2}(d,0,T)$), we find optimal waveform of order two for $b_{M_{\mathtt{A}}}(t)$. For $a_{M_{\mathtt{A}}}(t)$ we find optimal waveform of order one.
According to Lemma \ref{decrese:error:pro} in Appendix \ref{AppB},
the error probability is a decreasing function of
$|\rho_1-\rho_0|$, hence by minimizing $\rho_0$ and maximizing $\rho_1$ the optimal solution can be obtained. Since $\rho_0\geq 0$, we set $a_0(t),b_0(t)=0$ to minimize $\rho_0$ and we get $\rho_0=0$. Fix some $a_1(t)$, and let $b_{1}^{\text{opt}}(t)$ be a maximizer of $\rho_1$. Since the expression for
$\rho_1(a_1(t),b_{1}(t))$, up to its second-order term, is linear in $b_1(t)$ for a fixed value of $a_1(t)$, according to Lemma \ref{SupportLemma} there is waveform of the form $b_{1}^{*}(t)=\hat{b}_{1}\delta(t-t^{[b_1]})$ with a total power \eqref{eqn:power2a} less than or equal to the power of $b_{1}^{\text{opt}}(t)$
such that (up to the second-order terms), we have $\rho_1(a_1(t),b_{1}^{\text{opt}}(t))=\rho_1(a_1(t),b_{1}^{*}(t))$.
In other words, replacing $b_{1}^{\text{opt}}(t)$ with $b_{1}^{*}(t)$ would not change the value of $\rho_1$, or the error probability.
As
$b_{1}^{*}(t)=\hat{b}_{1}\delta(t-t^{[b_1]})$ has a total power constraint $s_{\mathtt{B}}$, we obtain $b_{1}\leq s_{\mathtt{B}}$.
Since
$\rho_1(a_1(t),b_{1}^{*}(t))$ is an increasing function in $\hat{b}_{1}$, we deduce that $\hat{b}_1=s_{\mathtt{B}}$ maximizes $\rho_1$.
Next, in order to obtain the first-order optimal waveform for
$a_1(t)$, we recall that $\rho_1$ up to the first-order is a linear function of only $a_1(t)$. According to Lemma \ref{SupportLemma} for any arbitrary $a_{1}(t)$ there is some
${a}_{1}^{*}(t)=\hat{a}_{1}\delta(t-t^{[a_1]})$ with a total power less than or equal to the power of $a_1(t)$
such that (up to order one terms) $\rho_{1}(a_1(t))=\rho_{1}(a_{1}^{*}(t))$. Then, as above the power constraint implies that $\hat{a}_{1}\leq s_{\mathtt{A}}$. Since (up to order one terms), $\rho_{1}(a_{1}^{*}(t))$ is an increasing function with respect to $\hat{a}_{1}$
we deduce $\hat{a}_{1}=s_{\mathtt{A}}$. That completes the proof.
\end{proof}
In order to determine parameters in the statement of Theorem \ref{th::2} we need optimize over constants $t^{[a_1]},t^{[b_1]}$.
As an example, consider the values of system parameters as follows in Table \ref{tabex2}.
\begin{table}
\centering
\caption{Parameters}
\label{tabex2}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(D_{\mathtt{A}},D_{\mathtt{B}},D_{\mathtt{C}})[m^{2}s^{-1}]$} &\multirow{2}{*} {$10^{-10}\times(10,1.1,1)$}\\
&\\
\hline
\multirow{2}{*} {$\lambda[molecules^{-1}.m^{3}.s^{-1}]$} & \multirow{2}{*}{$ 10^{-23}$}\\
&\\
\hline
\multirow{2}{*}{$V[m^{3}]$} &\multirow{2}{*} {$2.5\times10^{-7}$}\\
&\\
\hline
\multirow{2}{*}{$(T[s],\beta,\gamma)$ }& \multirow{2}{*}{$(10,2,1)$} \\
&\\
\hline
\multirow{2}{*}{$(d_{\mathtt{B}} [m],d_{\mathtt{c}} [m])$} & \multirow{2}{*}{$5\times 10^{-5}\times((1,1),(10,0))$ } \\
&\\
\hline
\multirow{2}{*}{$(\mathcal{I}_{int}(x,y),f_{mathtt{c}}(x,y,t))$} & \multirow{2}{*}{$(0,4\times 10^{4}\delta(x-5\times 10^{5})\delta(y-5\times 10^{5})\delta(t))$} \\
&\\
\hline
\multirow{2}{*}{$s_{\mathtt{A}}[molecules.m^{-3}],s_{\mathtt{B}}[molecules.m^{-3}])$} & \multirow{2}{*}{$10^8\times(5,24)$} \\
&\\
\hline
\end{tabular}
\end{table}
Simulation results yield the optimal values for
$t^{[a_1]},t^{[b_1]}$ as $t^{[a_1]}=5.62,t^{[b_1]}=0$.
Also, when we vary $s_{\mathtt A}\in [500,5\times 10^{11}]$ while we fix the other parameters, we observe that $t^{[b_1]}$ remains zero. In other words, it is best to release molecules of type $\mathtt B$ as soon as interfering molecules of type $\mathtt C$ are released. Moreover, the value of
$t^{[a_1]}$ is almost constant as we vary $s_{\mathtt A}$.
\subsection{Example 3: Reaction for Two-way Communication }
Finally, our third example considers two transceivers (who are able to both transmit and receive signals) in a three-dimensional setting.
Consider two molecular transceivers $\mathsf T_1$ and $\mathsf T_2$. The transceiver $\mathsf T_1$ is able to release molecules of type $\mathtt{A}$ and receive molecules of type $\mathtt{B}$. On the other hand, the transceiver $\mathsf T_2$ is able to release molecules of type $\mathtt{B}$ and receive molecules of type $\mathtt{A}$. If there is no reaction between $\mathtt{A}$ and $\mathtt{B}$, we have two distinct directional channels between the two transceivers (e.g., one channel is formed by $\mathsf T_1$ releasing molecules of type $\mathtt{A}$ and $\mathsf T_2$ receiving them). However, if $\mathtt{A}$ reacts with $\mathtt{B}$ in the medium, the two channels become entangled. This would impact the transmission strategy of the two transceivers if they wish to establish a two-way communication channel, and send and receive messages at the same time. The chemical reaction would weaken both signals in this case. Assume that
\begin{equation}
\ce{\mathtt{A} + \mathtt{B} ->[\lambda] \mathtt{P}}.
\end{equation}
The following equations describe the dynamic of the system:
\begin{align}
&\frac{\partial[\mathtt{A}]}{\partial t}=D_\mathtt{A}\nabla^2[\mathtt{A}]-\lambda~[\mathtt{A}][\mathtt{B}]+f_\mathtt{A}(\vec{x},t),~\frac{\partial[\mathtt{B}]}{\partial t}=D_\mathtt{B}\nabla^2[\mathtt{B}]-\lambda~[\mathtt{A}][\mathtt{B}]+f_\mathtt{B}(\vec{x},t),\label{ex3:eq}
\end{align}
where $f_\mathtt{A}(\vec{x},t),f_\mathtt{B}(\vec{x},t)$ are input signals of the two transceivers. The initial conditions are set to zero, and we allow diffusion in the entire $\mathbb{R}^3$ with no boundaries. As before, for
$\lambda=0$ the system of equations is linear and analytically solvable.
Consider a solution of the following form:
\begin{align}\label{teylorex:2N}
&[\mathtt{A}](\vec{x},t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{A}]_{i}(\vec{x} ,t),~[\mathtt{B}](\vec{x},t)=\sum_{i=0}^{\infty}\lambda^{i}[\mathtt{B}]_{i}(\vec{x} ,t).
\end{align}
By matching the coefficients of $\lambda^0,\lambda^1$ on the both side of equations, we can find a solution for in low reaction rate regime.
By matching the constant terms, we obtain
\begin{align}\label{ex3:firstterm}
&\frac{\partial[\mathtt{A}]_0}{\partial t}=D_\mathtt{A}
\nabla^{2}[\mathtt{A}]_0 +f_\mathtt{A}(\vec{x},t),~\frac{\partial[\mathtt{B}]_0}{\partial t}=D_\mathtt{B}\nabla^{2}[\mathtt{B}]_0 +f_\mathtt{B}(\vec{x},t).
\end{align}
By matching the coefficients of $\lambda$, we get
\begin{align}\label{ex3:secondterm}
&\frac{\partial[\mathtt{A}]_1}{\partial t}=D_\mathtt{A}
\nabla^{2}[\mathtt{A}]_1 -[\mathtt{A}]_{0}[\mathtt{B}]_{0},~\frac{\partial[\mathtt{B}]_1}{\partial t}=D_\mathtt{B}\nabla^{2}[\mathtt{B}]_1 -[\mathtt{A}]_{0}[\mathtt{B}]_{0}.
\end{align}
The impulse response for heat equation is as follows:
\begin{equation}\label{greenthree}
\phi_\mathtt{A}(\vec{x},t)=\frac{1}{(4\pi D_\mathtt{A}t)^{\frac{3}{2}} }\exp(-\frac{\lVert\vec{x}\rVert_{2}^{2}}{4 D_\mathtt{A} t}), ~~\vec{x}\in \mathbb{R}^{3}~,t\geq 0.
\end{equation}
The solution of \eqref{ex3:firstterm},
\eqref{ex3:secondterm} are given in the following form.
\begin{align}\label{ex:3:final:density}
&[\mathtt{A}]_0=\phi**f_\mathtt{A},~
[\mathtt{B}]_0=\phi**f_\mathtt{B},~
[\mathtt{A}]_1=- \phi_{\mathtt{A}}**([\mathtt{A}]_0 [\mathtt{B}]_0),~
[\mathtt{B}]_1=- \phi_{\mathtt{B}}**([\mathtt{A}]_0 [\mathtt{B}]_0).
\end{align}
This results in the following solution for low reaction rates:
\begin{align}\label{ex3::finalsol}
&[\mathtt{A}](\vec{x},t)= [\mathtt{A}]_0(\vec{x},t)+\lambda [\mathtt{A}]_1(\vec{x}, ,t)+\mathcal{O}(\lambda^2),~[\mathtt{B}](\vec{x},t)= [\mathtt{B}]_0(\vec{x},t)+\lambda [\mathtt{B}]_1(\vec{x}, ,t)+\mathcal{O}(\lambda^2).
\end{align}
Using these equations, we design a modulation scheme in Section \ref{mod:ex3}.
\subsection{Design of Modulation For Example 3}\label{mod:ex3}
Suppose transceivers $ \mathsf T_{\mathtt {A}}, \mathsf T_{\mathtt {B}}$ encode messages $M_{\mathtt{A}},M_{\mathtt{B}}\in\{0,1\}$ with input signals $f_{\mathtt{A}}^{i}(\vec{x},t) =a_i(t)\delta(\vec{x}),f_{\mathtt{B}}^{i}(\vec{x},t) =b_i(t)\delta(\vec{x}-\vec{d_{\mathtt{B}}})$, for $0\le t\le T$ and $i=0,1$,
where $\vec{d_\mathtt{B}}=(d,0,0)
$ is the location of $\mathsf T_{\mathtt {B}}$. In other words, transceiver $ \mathsf T_{\mathtt {A}}$ releases a concentration of $a_{M_{\mathtt{A}}}(t)$ at origin, and
transceiver $ \mathsf T_{\mathtt {B}}$ releases a concentration of $b_{M_{\mathtt{B}}}(t)$ at its location $\vec{d_\mathtt{B}}$. As before, we restrict the total amount of released molecules of types $\mathtt A$ and $\mathtt B$ during the transmission period $T$ as follows:
\begin{align}\int_{0}^T a_i(t)dt\leq s_{\mathtt A},~\int_{0}^T b_j(t)dt\leq s_{\mathtt B} \quad i=0,1.\label{eqn:power3a}
\end{align}
Transceiver $\mathsf{T_{\mathtt {A}}}$ samples the medium for molecules of type $\mathtt {A}$ at the end of the time slot at time $t=T$, and uses its observation to decode the message $\hat{M}_{\mathtt{B}}$. Similarly, transceiver
$\mathsf{T_{\mathtt {B}}}$ decodes the message $\hat{M}_{\mathtt{A}}$ using its sample at time $T$.
Four error probabilities could be considered in our problem: since transceiver $ \mathsf T_{\mathtt {B}}$ knows transmitted message $M_{\mathtt{B}}$, for this transceiver we can consider error probabilities
\begin{align*}&J_1:= Pr(M_{\mathtt{A}}\neq \hat M_{\mathtt{A}}|M_{\mathtt{B}}=0),
~J_2:= Pr(M_{\mathtt{A}}\neq \hat M_{\mathtt{A}}|M_{\mathtt{B}}=1)\end{align*}
Similarly, for transceiver $ \mathsf T_{\mathtt {A}}$, we can consider
\begin{align*}&J_3:= Pr(M_{\mathtt{B}}\neq \hat M_{\mathtt{B}}|M_{\mathtt{A}}=0),~
J_4:= Pr(M_{\mathtt{B}}\neq \hat M_{\mathtt{B}}|M_{\mathtt{A}}=1)\end{align*}
We would like to make $J_1, J_2, J_3, J_4$ as small as possible
by choosing optimal waveforms $a_i(t)$ and $b_j(t)$. Since there is a tradeoff between minimizing $J_1, J_2, J_3$ and $J_4$, we can choose a cost function
$\mathcal{H}:[0,1]^{4}\rightarrow \mathbb{R}$
and minimize
$\mathcal{H}(J_{1},J_{2},J_{3},J_{4})$. While we could choose any arbitrary cost function, for simplicity of exposition, we adopt $\mathcal{H}(J_1,J_2,J_3,J_4)=\sum_{i=1}^{4}\omega_{i}\log(J_{i})
$ for some constants $\omega_{i}, i=1, \cdots,4$. This particular choice for $\mathcal{H}$ leads to very simple optimal waveforms $a_i(t)$ and $b_j(t)$ in the low reaction regime as shown in the following theorem (other choices for $\mathcal{H}$ result in more delta terms in the optimal waveform).
\begin{theorem}\label{th::3}
One choice for optimal order one waveforms $a_i(t), b_i(t)$ is as follows:
\begin{align}
&a_0(t)=0,~a_1(t)=\hat{a}_{1}\delta(t-t^{[a_1]}),~b_0(t)=0,~b_1(t)=\hat{b}_{1}\delta(t-t^{[b_1]}).
\end{align}
for some $\hat{a}_{1},\hat{b}_{1},t^{[a_1]},t^{[b_1]}$.
\end{theorem}
\begin{proof}
By substituting $f_{\mathtt{A}}^{0}(\vec{x},t),f_{\mathtt{A}}^{1}(\vec{x},t),f_{\mathtt{B}}^{0}(\vec{x},t)$ and $f_{\mathtt{B}}^{1}
(\vec{x},t)$ in \eqref{ex:3:final:density},
\eqref{ex3::finalsol}, the density of molecule $\mathtt{A}$ at $\vec{d}_{\mathtt{B}}$ and
density of molecule $\mathtt{B}$ at $\vec{d}_{\mathtt{A}}=\vec{0}$ have an explicit formula with respect to input signals.
The concentration of molecule of type $\mathtt{A}$
at location of $\mathsf{T_{\mathtt {B}}}$ at the sampling time $T$ is $ [\mathtt{A}](\vec{d}_
{\mathtt{B}},T)=\zeta_1-\zeta_2,$ where:
\begin{align}
\zeta_1=\int_{0}^{T}
\phi_{\mathtt{A}}(\vec{d}_{\mathtt{B}},T-t^{'})a_{M_\mathtt{A}}(t^{'})dt^{'},
\end{align}
and
\begin{align}
\zeta_2=
\lambda
\int_{0}^{T}
\int_{\mathbb{R}^3}\int_{0}^{t'}
\int_{0}^{t'}
&\phi_{\mathtt{A}}(\vec{d}_{\mathtt{B}}-\vec{x}',T-t')\phi_{\mathtt{A}}(\vec{x}',t'-t_1)
\phi_{\mathtt{B}}(\vec{x}'-\vec{d}_{\mathtt{B}},t'-t_2)
\nonumber\\&a_{M_{\mathtt{A}}}(t_1)
b_{M_{\mathtt{B}}}(t_2)
dt_{1}dt_{2}d\vec{x}'dt'.
\end{align}
A similar expression holds for the concentration of molecule $[\mathtt B]$ at the location of $\mathsf{T_{\mathtt {A}}}$ at sampling time $T$ (by simply swapping $\mathtt{A}\rightleftharpoons\mathtt{B}$ in the above formula).
Observe that $[\mathtt{A}](\vec{d}_{\mathtt{B}},T)$ and $[\mathtt{B}](\vec{0},T)$ are bilinear functions with respect to input signals.
Let
$\rho_{m_{\mathtt{B}}=0}^{\mathtt{A}}$ be the concentration of molecules of type $\mathtt{A}$ at transceiver $\mathtt{B}$ if $m_{\mathtt{B}}=0$. Similarly, we define $\rho_{m_{\mathtt{B}}=1}^{\mathtt{A}}$. Also,
$\rho_{m_{\mathtt{A}}=1}^{\mathtt{B}}$ denotes the concentration of molecules of type $\mathtt{B}$ at transceiver $\mathtt{A}$ if $m_{\mathtt{A}}=0$, etc.
Transceivers are assumed to be transparent. Due to the counting noise transceiver $\mathtt{A}$ observes a sample from $\mathsf{Poisson}(V\rho^{\mathtt{A}}_{m_{\mathtt{B}}})$
for some constant $V$. A similar statement holds for transceiver $\mathtt{B}$.
According to Lemma \ref{decrese:error:pro} error probabilities $J_1,J_2,J_3,J_4$ are decreasing functions of
$|\rho_{01}^{\mathtt{B}}-\rho_{00}^{\mathtt{B}}|,
|\rho_{11}^{\mathtt{B}}-\rho_{10}^{\mathtt{B}}|,
|\rho_{01}^{\mathtt{A}}-\rho_{00}^{\mathtt{A}}|,
|\rho_{11}^{\mathtt{A}}-\rho_{10}^{\mathtt{A}}|$ respectively. By minimizing non negative numbers $\rho_{00}^{\mathtt{B}}, \rho_{10}^{\mathtt{B}}$,
$\rho_{00}^{\mathtt{A}},\rho_{10}^{\mathtt{A}}$ the probabilities reduce. We set $a_0(t),b_0(t)=0$ to achieve $(\rho_{00}^{\mathtt{B}}, \rho_{10}^{\mathtt{B}},
\rho_{00}^{\mathtt{A}},\rho_{10}^{\mathtt{A}})=(0,0,0,0)$. To obtain $a_1(t),b_1(t)$, we consider cost function $\mathcal{H}(J_1,J_2,J_3,J_4)=\sum_{i=1}^{4}\omega_{i}\log(J_{i})$ for some constants $\omega_{i}, i=1, \cdots,4$.
The expression for $\mathcal{H}(J_1,J_2,J_3,J_4)$ becomes a bilinear function of $a_1(t),b_1(t)$ since
\begin{align}
\mathcal{H}(J_1,J_2,J_3,J_4)=&-\omega_{1}\zeta_{1}(\vec{d}_{\mathtt{B}},a_{1})-\omega_{2}\zeta_{2}(\vec{d}_{\mathtt{B}},a_{1},b_{1})-\omega_{3}\zeta_{1}(\vec{d}_{\mathtt{A}},b_{1})-\omega_{4}\zeta_{2}(\vec{d}_{\mathtt{A}},a_{1},b_{1})\nonumber\\&-\sum_{i=1}^{4}\omega_{i}\log(2).
\end{align}
Fix some $b_{1}(t)$ and let $a_{1}^{\text{opt}}(t)$ be maximizer of $\mathcal{H}$, according to Lemma \ref{SupportLemma} there is a waveform of the form
${a}_{1}^{*}(t)=\hat{a}_{1}\delta(t-t^{[a_1]})$
with a total power \eqref{eqn:power3a} less than or equal to $a_{1}^{\text{opt}}(t)$ such that it dos
not affect the value of $\mathcal{H}$, \emph{i.e.}, $\mathcal{H}(a_{1}^{\text{opt}}(t),b_{1}(t))=\mathcal{H}(a_{1}^{*}(t),b_{1}(t))$. As ${a}_{1}^{*}(t)=\hat{a}_1\delta(t-t^{[a_1]})$ has a total power constraint $s_{\mathtt{A}}$, we obtain $\hat{a}_{1}\leq s_{\mathtt{A}}$. Hence ${a}_{1}^{\text{opt}}(t)=\hat{a}_1\delta(t-t^{[a_1]})$. By similar argument we can deduce that the one choice for optimal waveform $b_{1}(t)$ is
${b}_{1}^{*}(t)=\hat{b}_1\delta(t-t^{[b_1]})$ for some constant $\hat{b}_1\leq s_{\mathtt{B}},t^{[b_1]}\leq T$ . That completes the proof.
\end{proof}
In order to determine parameters in the statement of Theorem \ref{th::3} we need optimize over constants $\hat{a}_{1},\hat{b}_{1},t^{[a_1]},t^{[b_1]}$.
\begin{table}
\centering
\caption{Parameters}
\label{tabex3}
\begin{tabular}{ |c|c|c| }
\hline
\multirow{2}{*}{$(D_{\mathtt{A}},D_{\mathtt{B}})[m^{2}s^{-1}]$} &\multirow{2}{*} {$10^{-10}\times(10,1.1)$}\\
&\\
\hline
\multirow{2}{*} {$\lambda[molecules^{-1}.m^{3}.s^{-1}]$} & \multirow{2}{*}{$ 10^{-30}$}\\
&\\
\hline
\multirow{2}{*}{$V[m^{3}]$} &\multirow{2}{*} {$2.5\times10^{-14}$}\\
&\\
\hline
\multirow{2}{*}{$T[s]$ }& \multirow{2}{*}{$10$} \\
&\\
\hline
\multirow{2}{*}{$d_{\mathtt{B}} [m]$} & \multirow{2}{*}{$10^{-4}\times(1,1,1)$ } \\
&\\
\hline
\multirow{2}{*}{$(s_{\mathtt{A}}[molecules.m^{-3}],s_{\mathtt{B}}[molecules.m^{-3}])$} & \multirow{2}{*}{$10^8\times(5,24)$} \\
&\\
\hline
\end{tabular}
\end{table}
Simulation results yield the optimal values for $\hat{a}_{1},\hat{b}_{1},t^{[a_1]},t^{[b_1]}$ as $\hat{a}_{1}=2\times 10^{8} ,\hat{b}_{1}=2.4\times 10^{9},t^{[a_1]}=0,$ and $t^{[b_1]}=0$.
\bibliographystyle{IEEEtran}
| 7fa3fa74c9b897606d4df01a887b2dfd0d7e51aa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\setcounter{section}{1}\setcounter{equation}{0}
We consider the nonlinear wave equation with an inverse-square potential in $\Rm^d$,
i.e.,
\begin{equation}
\label{wave-La-p}
\left\{\begin{aligned}& u_{tt} +\Big(- \Delta + \frac{a}{|x|^2} \Big) u = - |u|^{p-1} u, \qquad (x,t)\in \Rm^d \times \Rm,\\
& (u,u_t)|_{t=0} = (u_0,u_1) \in \dot{H}^1 \times L^2,\end{aligned}\right.
\end{equation}
where $d\geq 3$, $p \in [1+\frac{4}{d-1},1+\frac{4}{d-2})$ and $a >-\frac{(d-2)^2}{4} $.
The restriction $a>-\frac{(d-2)^2}{4}$ ensures that the operator $\mathcal{L}_a
= -\Delta +\frac{a}{|x|^2}$ is positive
definite via the sharp Hardy inequalities.
In this article, we consider $\mathcal{L}_a$ as the Friedrichs extension
of this operator defined initially on $C_c^\infty(\Rm^d\setminus \{0\})$ via the quadratic form
\begin{equation*}
Q(f)= \int_{\Rm^d} \Big(|\nabla f(x)|^2 +\frac{a}{|x|^2} |f(x)|^2\Big) dx, \text{ for } f\in C_c^\infty(\Rm^d\setminus \{0\}) .
\end{equation*}
It is easy to find that the solution $u$ to the equation \eqref{wave-La-p}
satisfies an energy conservation law:
\begin{align*}
E(u,u_t)&\doteq\int_{\Rm^d} \Big( \frac{1}{2} |\nabla u(x,t) |^2+ \frac{a}{2} \frac{|u(x)|^2}{|x|^2}+ \frac{1}{2} |\partial_t u(x,t)|^2 +\frac{1}{p+1} |u(x,t)|^{p+1}\Big)dx\nonumber\\
& =E(u_0,u_1).
\end{align*}
If $u$ is a solution to \eqref{wave-La-p}, then, for $\lambda>0$,
the associated scaling transform
$u_\lambda(x,t)=\lambda^{-\frac{2}{p-1}} u( \tfrac{x}{\lambda} ,\tfrac{t}{\lambda} )$
is also a solution and its norm
in space $\dot H^{s_p}(\Rm^d)$
is invariant, where $s_p= \tfrac{d}{2}-\tfrac{2}{p-1}$.
When $s_p=1$(i.e., $p=p_e\doteq1+\frac{4}{d-2}$), we call the equation \eqref{wave-La-p} is energy critical. And we call the equation \eqref{wave-La-p} is energy subcritical(or supercritical) for $s_p<1$(or $s_p>1$). In particular, when $p$ equals to the conformal exponent $p_{conf}\doteq1+\frac{4}{d-1}$, we have
$s_p=\frac12$.
\subsection{Background and Motivations}
First, we recall the well-posedness and scattering theory of the defocusing nonlinear wave equation
\begin{equation}
\label{wave-p}
\left\{\begin{array}{l} u_{tt} - \Delta u = - |u|^{p-1} u, \qquad (x,t)\in \Rm^d \times \Rm,\\
(u,u_t)|_{t=0} = (u_0,u_1) \in \dot{H}^1 \times L^2,\end{array}\right.
\end{equation}
where $d\geq 3$ and $p \in [1+\frac{4}{d-1},1+\frac{4}{d-2})$.
By some suitable Strichartz estimates and the fixed point
argument, one may obtain the
well-posedness for the Cauchy problem \eqref{wave-p},
see for example \cite{LS-1985}.
For the scattering theory of the solution $u$,
we mean that there exist free wave solutions $u_\pm$ such that
$u-u_\pm $ tend to zero in some suitable space as $t\to\pm\infty$.
For
the
energy critical case $p=p_e$, the
global well-posedness and scattering theory of energy solution to the equation \eqref{wave-p} has been extensively studied by Grillakis
\cite{Grillakis-1990, Grillakis-1992},
Shatah-Struwe \cite{SS-1993,SS-1994},
Bahouri-Shatah \cite{BS-1998}, and
Bahouri-G\'{e}rard \cite{BG-1999}.
In the case of $p\neq p_e$, there is no conservation law of same scaling of $\dot H^{s_p}\times \dot H^{s_p-1} $ that provides
an a priori bound on the critical norm of the solution.
Under the assumption that
the $\dot H^{s_p} \times \dot H^{s_p-1} $ norm of the solution is uniformly bounded, many authors
studied the scattering problem of solution for the problem \eqref{wave-p},
see for example
\cite{KM-2011,Shen-2014,DL-2015-1,DL-2015-2,DLMM-2018,Rodriguez-2017}. We also refer to
\cite{DR-2017, DY-2018}
for general norms.
For the defocusing subcritical case, one may obtain the global well-posedness of solutions with finite energy via the Strichartz estimates and energy conservation law.
Nevertheless, in this case the scattering theory in energy spaces remains open.
Ginibre-Velo \cite{GV-1987} and Hidano\cite{Hidano-2003} proved that if the initial datum are in the conformal energy spaces,
then the solutions scatter in both time directions. In \cite{Shen-2017}, Shen
established
the scattering results for solutions in larger weighted spaces.
Latter, Dodson
obtained
the scattering results in
the critical Sobolev spaces for
solutions with radial initial datum in the critical space
$\dot B^2_{1,1}\times\dot B^1_{1,1}$ for $(d,p)=(3,3)$
in \cite{Dodson-2018-APDE}
and later $\dot H^{s_p} \times \dot H^{s_p-1} $,
for $d=3$ and $3\leq p<5$ in \cite{Dodson-2018-1,Dodson-2018-2}.
Miao-Yang-Zhao \cite{MYZ-2019} considered the case $(d,p)=(5,2)$ for initial datum in the critical space $\dot B^3_{1,1}\times\dot B^2_{1,1}.$
Recently, Shen developed the energy flux methods and showed the energy scattering results for solutions
wit
initial datum in the
weighed energy spaces in \cite{Shen-2018-1,shenenergy,Shen-2019-1,shenhdradial}.
For the non-radial scattering results, we refer to
Shen\cite{shen3dnonradial,shenhd} and Yang\cite{Yang-2019}.
Now we consider the scattering theory of the solutions to the Cauchy problem of \eqref{wave-La-p}. The well-posedness can be obtained by the Strichartz estimates of
Burq, Planchon, Stalker, and Tahvildar-Zadeh \cite{BPST-2003}. We also refer to
Killip-Miao-Visan-Zheng-Zhang\cite{KMVZZ-2018}
for the development of
harmonic analysis tools of the operator $\mathcal{L}_a$ by utilizing the heat kernel bounds.
Later, by employing these tools and the concentration compactness arguments, Miao-Murphy-Zheng\cite{MMZ-2019} considered the scattering results in the energy critical case
and show the solutions approach the linear solutions of the operator $\vec S_a(t) $. Here $\vec S_a(t) $ denotes the evolution operator of the linear equation
$\partial_t^2 u+\big(-\Delta +\tfrac{a}{|x|^2}\big)u=0$. However, the scattering of the energy subcritical situation seems still open for us.
This article is devoted to
the study of scattering theory of solutions for \eqref{wave-La-p} in sub-critical cases with initial datum in weighted energy spaces.
\subsection{Main Results}
The first results we obtain is
the
following scattering result
in the region outside the light cone.
\begin{theorem}[Exterior scattering] \label{main 1}
Suppose that $3\leq d \leq 6$, $p\in[1+\frac{4}{d-1},1+\frac{4}{d-2})$,
and $a>-\frac{(d-2)^2}4+\big(\frac{(d-2)p-d}{2p}\big)^2$.
Let $u$ be a radial solution to \eqref{wave-La-p} with a finite energy. Then there exist finite-energy free waves $ v^{\pm}$(solutions to the linear wave equation $\partial_t^2 v - \Delta v=0$)
so that
\[
\lim_{t \rightarrow \pm\infty} \int_{ |x|>|t|-\eta} \left(|\nabla u(x,t) - \nabla v^{\pm}(x,t)|^2 + |u_t(x,t)- v^{\pm}_t(x,t)|^2 \right) dx = 0, \;\forall \eta \in \Rm.
\]
\end{theorem}
Next, we show the scattering theory in the whole spatial space for solutions with some additional energy decay.
\begin{theorem} \label{main 2}
Suppose that $3\leq d \leq 6$, $p\in[1+\frac{4}{d-1},1+\frac{4}{d-2})$,
and $a>-\frac{(d-2)^2}4+\big(\frac{(d-2)p-d}{2p}\big)^2$.
Let $u$ be a radial solution to \eqref{wave-La-p} with initial data $(u_0,u_1) \in \dot{H}^1 \times L^2$
so that the inequality
\begin{equation}\label{weight-energy}
E_{\kappa} (u_0,u_1) \doteq \int_{\Rm^d} (|x|^\kappa + 1)\left(|\nabla u_0(x) |^2 +|u_1(x)|^2 + |u_0(x)|^{p+1} \right) dx < +\infty
\end{equation}
holds for a constant $\kappa \geq \kappa_0 \doteq \frac{(d+2)-(d-2)p}{p+1}$. Then the solution $u$ scatters, that is, there exist finite-energy free waves $v^\pm$ so that
\[
\lim_{t \rightarrow \pm \infty} \int_{\Rm^d} \left(|\nabla u(x,t) - \nabla v^\pm(x,t)|^2 + |u_t(x,t)-v^\pm_t(x,t)|^2 \right) dx = 0.
\]
\end{theorem}
\begin{remark}\label{rem-1.3}
We remark that these theorems depend on a suitable local theory with $(u,u_t) \in C(\Rm; \dot{H}^1 \times L^2)$ and $|u|^{p-1} u \in L_{loc}^1 L^2(\Rm \times \Rm^d)$.
The condition $|u|^{p-1} u \in L_{loc}^1 L^2 (\Rm \times \Rm^d)$ will be very useful when we need to apply a smooth approximation argument.
This inspired us to use the Strichartz norm
$L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x $.
When $d=3$,
$\frac{2p}{(d-2)p-d}=\frac{2p}{p-3}>2, $
which holds for $p\in[3,5).$
On the hand
if $d\geq 4$, the conditions
$\frac{2p}{(d-2)p-d}\geq 2$ and $p\in[p_{conf},p_e) $
imply that we need the restriction $d\leq 6.$ For the higher dimension $d\geq7$, one may consider the exotic Strichartz estimates as in \cite{Foschi-2005, Visan-2007}.
On the other hand, the assumption of $a$
origins from the Strichartz estimates of \cite{BPST-2003}.
In fact,
$\big(\tfrac{(d-2)p-d}{2p}\big)^2$ is increasing of $p$ and
corresponds to
the exponents in
\cite{MMZ-2019} when $d=3,4$ and $p=p_e$.
That is why we need to make additional assumption on $a$ and $d$.
\end{remark}
Now,
we outline the proof of these theorems and make of some suitable reduction.
For the case of $a>0$,
the proof is similar to the case $a=0$ of \cite{shenhdradial}. In fact, one can show the energy flux estimates on the light cones by using the energy flux formula
\begin{align}\nonumber
&E(t_2; B(0,t_2-\eta)) - E(t_1; B(t_1-\eta)) \\
=& \frac{1}{\sqrt{2}} \int_{|x|=t-\eta, t_1<t<t_2} \Big(\frac{1}{2}|u_r+u_t|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2} + \frac{1}{p+1}|u|^{p+1}\Big) dS.
\label{energy-flux-a>0}
\end{align}
By making use of
the characteristic line method, one can show Theorem \ref{main 1} holds since in this case
$a \frac{u}{|x|^2}$
is a defocusing term with a good spatial decay.
Next, for the solution with weighted energy, one can show the
scattering results by utilizing the energy distribution estimates
\begin{align}\nonumber
& \int_{R<|t|<T} \int_{|x|< R} \left(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\right) dxdt \\
\leq& \int_{-R}^R \int_{|x|> R} \left(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\right) dxdt
,
\label{coro-Morawetz-a>0}
\end{align}
which follow from the Morawetz estimates
\begin{align}\nonumber
&\frac1{2R} \int_{T_1}^{T_2} \int_{|x|< R}
\left(|\nabla u|^2 +a \frac{|u|^2}{|x|^2}+| u_t|^2 + \frac{(d-1)(p-1)-2}{p+1} |u|^{p+1} \right) dxdt \\
& + \frac{d-1}{4R^2}\int_{T_1}^{T_2} \int_{|x|=R} |u|^2dS(x)dt + \int_{T_1}^{T_2}\int_{|x|\geq R} \left( \frac{(d-1)(p-1)}{2(p+1)} \frac{|u|^{p+1}}{|x|}+ (a+\lambda_d) \frac{|u|^2}{|x|^3}\right) dxdt\nonumber \\
\leq& 2E ,\label{Morawetz-a>0}
\end{align}
for some constant $\lambda_d\geq0.$ For more details, we refer to \cite{shenhdradial}.
We only focus on the case $a<0$ from now on. From the formula
\eqref{energy-flux-a>0}, we prove the energy flux estimates on light cones by using the Hardy inequality and the radial decay properties in Lemma \ref{pointwise estimate 2} of the radial solutions.
Thanks to these estimates and the radiation
field
for free waves, we are able to prove Theorem \ref{main 1} by utilizing the characteristics lines method.
Next,
for the proof of Theorem \ref{main 2}, since the 、
estimate \eqref{Morawetz-a>0} do not seem to
hold in this case, we establish a
modified
version of
Morawetz estimates, which is based on the Hardy inequalities on a local region.
Employing the modified Morawetz estimates and careful treatments
of the potential term $a\frac{u}{|x|^2}$, we prove the
scattering results for solutions with an additional energy decay.
\begin{remark}
As noted in \cite{shenhdradial}, there exists radial smooth initial datum $(u_0,u_1)$ such that the assumptions of Theorem \ref{main 2}, but $(u_0,u_1)$ do not belong to the scaling invariant spaces $\dot H^{s_p}\times \dot H^{s_p}(\Rm^d) $.
In fact, suppose
\begin{equation}
u_0(x)=(1+|x|)^{-\frac{2(p+d+1)}{(p+1)^2}-\varepsilon} \text{ ~~and ~~} \left|\nabla u_0(x)\right| \approx (1+|x|)^{-\frac{2(p+d+1)}{(p+1)^2}-1-\varepsilon}, |x|\gg1.
\end{equation}
By a direct computation, we have $(u_0,0)$ satisfies the assumption of
Theorem \ref{main 2} with $\kappa=\kappa_0$. But $u_0 \not\in L^\frac{d(p-1)}{2} (\Rm^d)$ when $\varepsilon$ is sufficiently small. Thus $u_0 \not\in \dot H^{s_p}(\Rm^d)$ by the Sobolev embedding.
\end{remark}
\begin{remark}
For the case $-\lambda_d<a$, one may improve the
result of Theorem \ref{main 2} slightly by showing the inward/outward energy flux and
the weighted Morawetz estimates as in \cite{shen3dnonradial}. Here $\lambda_d = (d-1)(d-3)/4$.
But the minimal decay rate $\kappa_0$ of
energy in \eqref{weight-energy} can not be further improved by the inward/outward energy theory.
While for the case $a<-\lambda_d$, the inward/outward energy flux formula may be not true
since the extra term
$\iint_{\Rm\times\Rm^3} \frac{|u|^2}{|x|^3}dxdt$ may be infinite.
\end{remark}
\begin{remark}
We consider the scattering theory of the linear wave equation
\[ \label{wave-Linear}
\left\{\begin{array}{l} u_{tt} +\left(- \Delta + \frac{a}{|x|^2} \right) u = 0, \qquad (x,t)\in \Rm^d \times \Rm,\\
(u,u_t)|_{t=0} = (u_0,u_1),\end{array}\right. \tag{$LW_a$}
\]
where $d\geq3$ and $a>-\frac{(d-2)^2}{4}$.
By the Strichartz estimates of $\vec S_0(t)$ and the generalized Morawetz estimates, one may obtain the scattering result of solution of \eqref{wave-Linear} in $\dot H^\frac12 \times \dot H^{-\frac12}(\Rm^d)$.
In fact, by the Duhamel formula, for $t_1<t_2 $, we have
\begin{equation}\label{Duhamel-1}
\vec S_0(-t_2)(u(t_2),u_t(t_2))(x) - \vec S_0 (-t_1)(u(t_1),u_t(t_1))(x) =
\int_{t_1}^{t_2} \vec S_0(-s) \big[0, \tfrac{a}{|\cdot|^2} u(\cdot,s) \big](x)ds.
\end{equation}
Then, by square function estimates\footnote{ One can also use the Strichartz estimates on $L^2_t L^{\tfrac{2d}{d-2},2 }_x $ here but with radial assumption if $d=3$. For the cases $d\geq 4,$ we refer to \cite{KT-1998}. If $d=3$,
this follows
the Strichartz estimates for radial data
(\cite[Theorem 1.3]{Sterbenz2005})and real interpolation estimates in \cite[Section 6]{KT-1998}.} in Appendix
and generalized H\"older inequality, we have
\begin{align}
& \left\| \vec S_0(-t_1)(u(t_1),u_t(t_1)) - \vec S_0(-t_2)(u(t_2),u_t(t_2)) \right\|_{\dot H^\frac12 \times \dot H^{-\frac12} (\Rm^d)}\nonumber\\
\lesssim~ & \left\|\tfrac{a}{|x|^2}u(x,t) \right\|_{L^{\tfrac{2d}{d+2},2 }_x L^2_t ([t_1,t_2] \times \Rm^d ) }\nonumber\\
\lesssim ~& \left\| |x|^{- 1}\right\|_{L^{d ,\infty}_x(\Rm^d)} \left\| \tfrac{a}{|x| } u(x,t) \right\|_{L^2_{t,x} (\Rm^d\times[t_1,t_2])} \to 0,
\label{Duhamel-3}
\end{align}
as $t_2>t_1\to \infty,$
where $L^{\alpha,\beta}_x$ are Lorentz spaces and we have used the generalized Morawetz estimate as in \cite[Theorem 2]{BPST-2003}, i.e.
\begin{equation}
\| \tfrac{1}{|x| } u(x,t) \|_{L^2_{t,x} (\Rm\times\Rm^d)} \lesssim \|(u_0,u_1)\|_{\dot H^\frac12 \times \dot H^{-\frac12} (\Rm^d)}.
\end{equation}
However,
generally speaking,
for scattering results in the energy space $\dot H^1\times L^2(\Rm^d)$,
one may need the associated square function estimates on $L^{\frac{2d}{d-1},2}_xL^2_t(\Rm\times\Rm^d)$ or
the Strichartz estimates on $L^2_tL^{\frac{2d}{d-1},2}_x(\Rm\times\Rm^d)$.
But, the both tools are not true even for radial case. In fact, One easy to see that square function estimates do not hold via Plancherel and the fact $| \widehat{ d\sigma}_{S^{d-1} }(\xi) | = O((1+|\xi|)^{\frac{1-d}{2}})$); On the other hand, by making use of Proposition \ref{decay-out-cone} and H\"older inequalities, we have
\begin{equation}
\| |e^{it|\nabla|} f \|_{L^q_x{ (||x|-|t||\lesssim 1) }} \gtrsim
\| |e^{it|\nabla|} f \|_{L^2_x{ (||x|-|t||\lesssim 1) }} |t|^{-(n-1)(\frac12-\frac1q)}
\gtrsim |t|^{-(n-1)(\frac12-\frac1q)}.
\end{equation}
This inequality
shows
us that $L^p_tL^q_x(\Rm\times\Rm^d)$ is acceptable if
$\frac{1}{p}+\frac{n-1}{q}<\frac{n-1}{2}.$
Fortunately, we can obtain scattering theory of radial energy solutions to the linear equation \eqref{wave-Linear} in the Appendix by the argument of \cite{shenenergy} and a series of estimates. This fact shows us that the radial solutions of the energy critical nonlinear wave equations scatter to free waves, see \cite{MMZ-2019}.
\end{remark}
\subsection{The structure of this paper}
This article is organized as follows: In Section \ref{basic}, we give
technical lemmas for latter use and well-posedness theory of Cauchy problem \eqref{wave-La-p}.
In Section \ref{energyflux}, we prove the energy flux formula of the solution to the problem \eqref{wave-La-p}.
We establish a class of local Morawetz estimates associated with the problem\eqref{wave-La-p} in Section \ref{Morawetz-sec}.
In Section \ref{mtheod-characteristic}, we prove Theorem \ref{main 1} by making use of the characteristic lines method.
In Section \ref{further} we establish a series of local energy estimates,
and complete the proof of Theorem \ref{main 2} by making use of Morawetz estimates established in Section \ref{Morawetz-sec}. In the appendix, we prove the asymptotic behaviour of linear solution, and establish the square function estimate in Lorentz space by Stein-Tomas restriction theorem.
We conclude this section by introducing some notations. We denote by $A\lesssim B$ the fact that the inequality $A\le C B$ holds with an implicit constant C depending on $p$ and $a$. Sometimes we also allow these constants to depend on the energy $E$ and weighted energy $E_\kappa$, especially in the last two sections.
The derivative operator $\nabla$ refers to the spatial variable only. $\partial_r=\frac{x}{|x|} \cdot \nabla$ denotes the derivative in the radial direction and $\slashed{\nabla}=\nabla - \frac{x}{|x|} \partial_r$ denotes the angular derivative.
Let $u(x, t)$ be a spatially radial function. By convention $u(r, t)$ represents
the value of $u(x, t)$ when $|x|=r$.
\section{Technical Lemmas and well-posedness theory}\label{basic}
\setcounter{section}{2}\setcounter{equation}{0}
First, we recall the sharp Hardy inequality, for $d\geq 3,$
\begin{equation}
\frac{(d-2)^2}{4} \int_{\Rm^d} \frac{|u(x)|^2}{|x|^2} \leq
\int_{\Rm^d} |\nabla u(x)|^2 dx.\label{basic-eq-1}
\end{equation}
For $a \in(-\frac{(d-2)^2}{4},\infty )$, we have the rough identity as follows
\begin{equation} E \simeq \int_{\Rm^d} \Big(|\nabla u(x,t)|^2 + \frac{|u(x,t)|^2}{|x|^2}+| u_t(x,t)|^2 + |u(x,t)|^{p+1}\Big) dx. \label{basic-eq-2}
\end{equation}
Thus for any region $\Sigma \subset \Rm^d$ we have
\begin{equation}
\int_{\Sigma} \Big(\frac{1}{2}|\nabla u(x,t)|^2 +\frac{a}{2} \frac{|u(x,t)|^2}{|x|^2}+\frac{1}{2}| u_t(x,t)|^2 + \frac{1}{p+1}|u(x,t)|^{p+1}\Big) dx \lesssim E.
\label{basic-eq-3}
\end{equation}
\begin{lemma}[Pointwise Estimate, see \cite{shenenergy} \cite{shenhdradial}] \label{pointwise estimate 2}
All radial $\dot{H}^1(\Rm^d)$ functions $u$ satisfy
$$ |u(r)| \lesssim_d r^{-(d-2)/2} \|u\|_{\dot{H}^1}, \qquad r>0.$$
If $u$ also satisfies $u \in L^{p+1} (\Rm^d)$, then its decay is stronger as $r \rightarrow +\infty$, i.e.
\begin{equation}
|u(r)| \lesssim_d r^{-\frac{2(d-1)}{p+3}} \|u\|_{\dot{H}^1}^{\frac{2}{p+3}} \|u\|_{L^{p+1}}^{\frac{p+1}{p+3}}, \qquad r>0.
\label{basic-eq-4}\end{equation}
\end{lemma}
Now, we recall some properties of the operator $\mathcal{L}_a=-\Delta +\frac{a }{|x|^2}.$
For $r\in(1,\infty)$, we write $\dot H^{1,r}_a$ and $H^{1,r}_a$ for the homogeneous and inhomogeneous space defined in terms of $\mathcal{L}_a$; these spaces have norms
$$\|f\|_{\dot H^{1,r}_a}= \|\sqrt{\mathcal{L}_a } f \|_{L^r},\quad \;
\|f\|_{ H^{1,r}_a}= \|\sqrt{1+\mathcal{L}_a } f \|_{L^r}. $$
We write $\dot H^{1,2}_a= \dot H^1_a$ and denote $\sigma= \frac{d-2}{2}-\sqrt{\left(\frac{d-2}{2} \right)^2+a } $.
\begin{lemma}[Equivalence of Sobolev spaces, \cite{KMVZZ-2018}]
Let $d \geq 3$, $a > -\frac{(d-2)^2}{4}$, and $s\in(0,2)$.
\begin{itemize}
\item \; If $p\in (1,\infty)$ satifies $\frac{s+\sigma}{d}<\frac1p<\min\left\{ 1,\frac{d-\sigma}{d}
\right\}$, then
\[ \||\nabla|^s f\|_{L^p(\mathbb{R}^d)} \lesssim \|\mathcal{L}^\frac{s}{2}_a f\|_{L^p(\mathbb{R}^d)}. \]
\item \; If $p\in (1,\infty)$ satifies $\max\left\{ \frac{s}{d}, \frac{\sigma}{d} \right\}<\frac1p<\min\left\{ 1,\frac{d-\sigma}{d}
\right\}$, then
\[ \|\mathcal{L}^\frac{s}{2}_a f\|_{L^p(\mathbb{R}^d)} \lesssim \||\nabla|^s f\|_{L^p(\mathbb{R}^d)}. \]
\end{itemize}
\end{lemma}
\subsection{Local theory}\label{local}
We consider the inhomogeneous wave equation
\begin{equation}\label{inh-wave-La}
\left\{ \begin{aligned}
\partial_t^2 u +\mathcal{L}_a u&=F(x,t),\quad (x,t)\in \mathbb{R}^d\times\mathbb{R} \\
(u,u_t)(x,0)&=(u_0,u_1)\in \dot H^1\times L^2(\mathbb{R}^d).
\end{aligned}\right.
\end{equation}
First, we record the Strichartz estimates.
\begin{proposition}[Strichartz estimates \cite{BPST-2003} \cite{MMZ-2019}]\label{Strichartz}
Let $ 2\leq q \leq \infty, 2\leq r < \infty$ such that the admissible condition
$\frac{1}{q} +\frac{d-1}{2r} \leq \frac{d-1}{4}$
holds,
where if $d=3$, we additionally require $q>2$.
Define $\gamma$ via the scaling condition
$ \frac1q+\frac{d}r =\frac{d}2- \gamma.$
Then, we have
\begin{equation}\label{Strichartz-equ-1}
\bigg\|\cos (t\sqrt{\mathcal{L}_a}) f + \frac{\sin (t\sqrt{\mathcal{L}_a)}}{\sqrt{\mathcal{L}_a}} g\bigg\|_{L^q_tL^r_x(\mathbb{R} \times \Rm^d)} \leq C \big(\|f\|_{\dot{H}^\gamma} + \|g\|_{\dot{H}^{\gamma-1}}\big)
\end{equation}
provided that
\[ -\min\left\{ 1,\sqrt{a +\tfrac94}-\tfrac12,\sqrt{a +\tfrac14}+1 \right\} <\gamma <\min\left\{
2, \sqrt{a +\tfrac94}+\tfrac12,\sqrt{a+\tfrac14 }+1 -\tfrac1q \right\} \] when $d=3$;
\begin{align*}
& -\min\left\{ \tfrac{d}{2} - \tfrac{d+3}{2(d-1)},\sqrt{a +\tfrac{d^2}4}-\tfrac{d+3}{2(d-1)},\sqrt{a +\tfrac{(d-2)^2}4}+1 \right\}\\
& \qquad <\gamma <\min\left\{
\frac{d+1}{2}, \sqrt{a +\tfrac{d^2}4}+\tfrac12,\sqrt{a+\tfrac{(d-2)^2}4 }+1 -\tfrac1q \right\}
\end{align*} when $d\geq4$.
\end{proposition}
\begin{remark}
We will also need
an inhomogeneous version of \eqref{Strichartz-equ-1},
\begin{equation}\label{inh-Strichartz}
\Big\|\int_0^t \frac{\sin ((t-s)\sqrt{\mathcal{L}_a)}}{\sqrt{\mathcal{L}_a}} F(s) ds \Big\|_{L_t^\infty([0,T]; \dot{H}^1\times L^2) \cap L_t^q L_x^r ([0,T]\times \Rm^d)} \leq C \|F\|_{L^1_t L^2_x([0,T]\times \Rm^d) },
\end{equation}
where parameters $(q,r,\gamma) = (q,r,1)$ satisfy conditions in Proposition \ref{Strichartz}.
In fact, by Strichartz estimates we have
$$ \Big\|\chi_{t\geq s}(t) \frac{\sin ((t-s)\sqrt{\mathcal{L}_a)}}{\sqrt{\mathcal{L}_a}} F(s)\Big\|_{L_t^\infty([0,T]; \dot{H}^1\times L^2) \cap L_t^q L_x^r ([0,T]\times \Rm^d)} \leq C \|F(s) \|_{L^2_x(\Rm^d)}.
$$
This inequality implies that
$$\Big\|\int_0^T \chi_{t\geq s}(t) \frac{\sin ((t-s)\sqrt{\mathcal{L}_a)}}{\sqrt{\mathcal{L}_a}} F(s) ds \Big\|_{L_t^\infty([0,T]; \dot{H}^1\times L^2) \cap L_t^q L_x^r ([0,T]\times \Rm^d)} \leq C \|F\|_{L^1_t L^2_x([0,T]\times \Rm^d) },$$
which is exactly inequality \eqref{inh-Strichartz}.
\end{remark}
For the well-posedness theory for
the problem \eqref{wave-La-p}, we need to use the Strichartz norm such as
$\|\cdot\|_{L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x} $.
Based on the discussion in Remark \ref{rem-1.3}, for each $p\in[p_{conf},p_e)$ with $3\le d\le6$, we have by \eqref{inh-Strichartz}
\begin{align}\nonumber
& \Big\| \int_{0}^t \frac{ \sin((t-s)\sqrt{\mathcal{L}_a} )}{\sqrt{\mathcal{L}_a}} \big( |u|^{p-1} u \big)ds \Big\|_{L_t^\infty(\dot H^1_x\times L^2_x) \cap L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x (I\times \mathbb{R}^d) } \\
\leq & C \| |u|^{p-1} u \|_{L^1_tL^2_x(I\times \mathbb{R}^d)} \nonumber\\
\leq &C T^{\frac{(d+2)-(d-2)p}{2}} \|u\|^p_{ L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \mathbb{R}^d)},\label{nonlinear-1}
\end{align}
and
\begin{align}\nonumber
& \Big\| \int_{0}^t \frac{ \sin((t-s)\sqrt{\mathcal{L}_a} )}{\sqrt{\mathcal{L}_a}} \Big( |u|^{p-1} u - |v|^{p-1} v \Big)ds \Big\|_{L_t^\infty(\dot H^1_x\times L^2_x) \cap L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x (I\times \mathbb{R}^d) } \\
\leq & ~~C T^{\frac{(d+2)-(d-2)p}{2}} (\|u\|^{p-1}_{ L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \mathbb{R}^d)}+\|v\|^{p-1}_{ L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \mathbb{R}^d)}) \|u-v\|_{ L_t^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \mathbb{R}^d)},\label{nonlinear-2}
\end{align}
where $I=[0,T]$ and
$$a>-\tfrac{(d-2)^2}4+\left(\tfrac{(d-2)p-d}{2p}\right)^2.$$
Then, by the Duhamel formula
\[
u= \cos(t \sqrt{\mathcal{L}_a} ) u_0 +\sin(t\sqrt{\mathcal{L}_a} )u_1 -
\int_{0}^t \frac{ \sin((t-s)\sqrt{\mathcal{L}_a} )}{\sqrt{\mathcal{L}_a}} \left( |u|^{p-1} u \right)ds,
\]
one can show the local well-posedness of \eqref{wave-La-p} via the concentration map argument.
Moreover, by energy conservation, one can obtain the global existence of the solution.
\begin{proposition}[Well-posedness theory]\label{well-p}
Assume that $3\leq d \leq 6$ and $a>-\frac{(d-2)^2}4+\left(\frac{(d-2)p-d}{2p}\right)^2$.
\begin{enumerate}
\item[\rm(i)] (Existence)
For any initial data $(u_0,u_1)\in \dot H^1\times L^2(\mathbb{R}^d)$,
there are a time interval $I\owns0$ and
a unique solution $u$ to \eqref{wave-La-p} on $I$, such that
$(u,u_t)\in C(I,\dot H^1\times L^2(\mathbb{R}^d)) $ and
$ u \in L_{t}^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \mathbb{R}^d). $
\item[\rm(ii)] (Long time perturbation)
Let $I=[0,T]$ be a time interval with $0<T<\infty$. For any positive $M<\infty$, there exists $\varepsilon_0=\varepsilon_0(T,M)$ such that if $0<\varepsilon<\varepsilon_0$, then the following holds:
If $\tilde u$ be a solution to $\partial_{tt} \tilde u+\mathcal{L}_a \tilde u =-|\tilde u|^{p-1}\tilde u$
and
\begin{align*}
& \|\tilde u(0),\partial_t \tilde u(0)\|_{\dot H^1\times L^2(\Rm^d)}<\infty ,\\
&|\tilde u\|_{L_{t}^{\frac{2p}{(d-2)p-d} } L^{2p}_x(I\times \Rm^d)} <M, \\
&\|(u_0,u_1)-(\tilde u(0),\partial_t \tilde u(0))\|_{\dot H^1\times L^2(\Rm^d)}<\varepsilon. \end{align*}
Then there exists a solution $u$ to \eqref{wave-La-p} on $I$ such that
\begin{align*}
& \|u-\tilde u\|_{L_{t}^{\frac{2p}{(d-2)p-d} } L^{2p}_x (I\times \Rm^d)} < C(T,M)\varepsilon, \\
& \sup_{t\in I}\|(u(t),\partial u(t))-(\tilde u(t),\partial_t \tilde u(t))\|_{\dot H^1\times L^2(\Rm^d)}< C(T,M)\varepsilon.\end{align*}
\item[\rm(iii)](energy conservation)
If $E(u_0,u_1)<\infty$, then we have
$E(u(t),u_t(t))\equiv E(u_0,u_1),$ for $t\in I_{max}$.
\end{enumerate}
By {\rm(i)} and {\rm (iii)}, we can obtain the global well-posedness of $u$.
\end{proposition}
\begin{proof}
The proof of {\rm (i)} and {\rm (ii)} follows from the estimates
\eqref{nonlinear-1}-\eqref{nonlinear-2} and standard fixed point argument, see for example
\cite{LS-1985, Shen-2014}.
Moreover,
by $(ii)$ and
Fatou's Lemma,
time reverse invariance of \eqref{wave-La-p}, it suffices to consider the
initial data in $(u_0,u_1)\in H^1\times L^2(\Rm^d)$.
From $u_0\in L^2(\Rm^d)$ and $(u,\partial_t u)\in C(I,\dot H\times L^2(\Rm^d)),$ we have
$(u,\partial_t u)\in C(I, H^1\times L^2(\Rm^d)).$
Then,
we have the energy conservation law, by Theorem 4.1 in Strauss\cite{Strauss-1966}, if we take $V= H^1(\Rm^d),H=W=L^2(\Rm^d), A(t)=\mathcal{L}_a$ and $Z=L^2((0,T);L^2(\Rm^d))$.
\end{proof}
\section{Energy Flux}\label{energyflux}
\setcounter{section}{3}\setcounter{equation}{0}
In this section, we consider the energy flux of the solutions given by Proposition \ref{well-p}.
Given any region $\Sigma \subset \Rm^d$, we use the following notation to represent local energy for convenience.
\begin{equation}\label{energy-flux-1}
E(t;\Sigma) \doteq \int_{\Sigma} \Big(\frac{1}{2}|\nabla u(x,t)|^2 +\frac{a}{2} \frac{|u(x,t)|^2}{|x|^2}+\frac{1}{2}| u_t(x,t)|^2 + \frac{1}{p+1}|u(x,t)|^{p+1}\Big) dx.
\end{equation}
We have the following energy flux formula for $t_2>t_1>\eta$
\begin{align}\nonumber
&E(t_2; B(0,t_2-\eta)) - E(t_1; B(t_1-\eta)) \\
=& \frac{1}{\sqrt{2}} \int_{|x|=t-\eta, t_1<t<t_2} \Big(\frac{1}{2}|u_r+u_t|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}|\slashed{\nabla} u|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dS.
\label{energy-flux-2}\end{align}
Since the local energy $E(t;\Sigma)$ remains uniformly bounded, the energy flux above is always bounded by a constant multiple of energy.
Letting $t_1\rightarrow \eta$ and $t_2 \rightarrow +\infty$, we obtain the energy flux through a whole light cone:
\begin{equation}\label{energy-flux-3}
\int_{|x|=t-\eta} \left(\frac{1}{2}|u_r+u_t|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}|\slashed{\nabla} u|^2 + \frac{1}{p+1}|u|^{p+1}\right) dS \lesssim E.
\end{equation}
The limit process $t_1\rightarrow \eta$ and $t_2 \rightarrow +\infty$ is valid at least in the radial case although it involves the integral of negative functions $\frac{a}{2}\frac{|u|^2}{|x|^2}$. In fact,
\begin{itemize}
\item \; The limit $t_1\rightarrow \eta$ can be done for a.e. $\eta \in \Rm$ by making use of
$$ \int_{T_1}^{T_2} \int_{\Rm^d} \Big(\frac{1}{2}|u_r+u_t|^2 +\frac{|a|}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}|\slashed{\nabla} u|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dx dt \lesssim (T_2-T_1) E < +\infty.
$$
\item\; In the radial case, we have $u(x,t) \lesssim |x|^{-\frac{2(d-1)}{p+3}}$. Thus the integral $\int_{|x|=t-\eta>1} \frac{|u|^2}{|x|^2} dS$ is always finite.
This fact tell us that the limit $t_2\rightarrow \infty$ can be done.
\end{itemize}
Next we have a key observation
\begin{equation}\label{energy-flux-4}
\int_{|x|=t-\eta} |u_r+u_t|^2 dS \geq \Big(\frac{ d-2 }{2} \Big)^2 \int_{|x|=t-\eta} \frac{|u|^2}{|x|^2} dS.
\end{equation}
In fact, it suffices to apply the sharp Hardy inequality on $\tilde{u}(x) = u(x, |x|+\eta) \in \dot{H}^1 (\Rm^d)$.
Hence, it immediately follows that
\begin{proposition}[Energy flux] \label{energy flux}
Let $u$ be a finite-energy radial solution to \eqref{wave-La-p}. Then we have the energy flux formula
\begin{align}
&E(t_2; B(0,t_2-\eta)) - E(t_1; B(t_1-\eta))\nonumber \\
=& \frac{1}{\sqrt{2}} \int_{|x|=t-\eta, t_1<t<t_2} \Big(\frac{1}{2}|u_r+u_t|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+ \frac{1}{p+1}|u|^{p+1}\Big) dS.
\label{energy-flux-5}\end{align}
In addition, the full energy flux through light cone satisfies
\begin{equation}\label{energy-flux-6}
\int_{|x|=t-\eta} \Big(|u_r+u_t|^2 + \frac{|u|^2}{|x|^2}+ |u|^{p+1}\Big) dS \lesssim E.
\end{equation}
\end{proposition}
\section{Morawetz estimates}\label{Morawetz-sec}
\setcounter{section}{4}\setcounter{equation}{0}
In this section,
we shall prove a class of local Morawetz estimates for solutions in Proposition \ref{well-p}.
For this purpose, we give a Hardy inequality on a local domain.
\begin{lemma}\label{cutoff energy}
Let $u \in \dot{H}^1(\Rm^d)$. Then for any $R>0$, we have
\begin{equation}\label{morawetz-eq-1}
f(R) \doteq \int_{|x|<R} \Big(|\nabla u(x)|^2 + a\tfrac{|u(x)|^2}{|x|^2}\Big) dx +\sigma \int_{|x|=R} \tfrac{|u(x)|^2}{|x|} dS(x)\geq 0,
\end{equation}
where
$$\sigma =\sigma_a \doteq \tfrac{d-2}{2}-\sqrt{\tfrac{(d-2)^2}{4}+a}.$$
Moreover, we have
\begin{equation}\label{morawetz-eq-2}
\int_{|x|<R} \Big(|\nabla u(x)|^2 + \frac{|u(x)|^2}{|x|^2}\Big) dx \lesssim f(R) + \int_{|x|=R} \frac{|u(x)|^2}{|x|} dS(x).
\end{equation}
\end{lemma}
\begin{proof}
Using the fact $a=\sigma^2-(d-2)\sigma$ and
integration by parts, we have
\begin{equation}\label{morawetz-eq-3}
f (R)=\int_{|x|< R}\Big |\nabla u(x)+\sigma \tfrac{x}{|x|^2}u(x)\Big|^2 dx \geq 0.
\end{equation}
This inequality implies that
\begin{align*}
\int_{|x|<R} |\nabla u(x)|^2 dx \lesssim& \int_{|x|< R} \Big(\Big |\nabla u(x)+\sigma \tfrac{x}{|x|^2}u(x)\Big|^2 + \Big|\tfrac{x}{|x|^2}u(x)\Big|^2\Big)dx\\
\lesssim& f(R) + \int_{|x|\leq R} \tfrac{|u(x)|^2}{|x|^2} dx.
\end{align*}
And, by the definition of $\sigma$ and
the fact that $f(R)\geq 0$ for $a$ replaced by
$a-\eps$, $\eps>0$ is a small constant,
we have
\begin{align*}
\int_{|x|<R} \frac{| u(x)|^2}{ |x|^2} dx \lesssim f(R) + \int_{|x|= R} \frac{|u(x)|^2}{|x|} dS(x).
\end{align*}
\end{proof}
\begin{corollary} \label{lower bound inner}
All $\dot{H}_{rad}^1(\Rm^d)$ functions $u$ satisfy
\begin{equation}\label{morawetz-eq-4}
\int_{|x|<R} \Big(|\nabla u(x)|^2 + a\frac{|u(x)|^2}{|x|^2}\Big) dx \geq -c_d \sigma R^{d-2}|u(R)|^2 = -\frac{\sigma}{ R}\int_{|x|=R} |u(x)|^2 dS(x),
\end{equation}
where $c_d $ is the surface area of the unit ball in $\Rm^d.$
\end{corollary}
\begin{remark}
The constants $-c_d\sigma$ in the first inequality of\label{morawetz-eq-4} and $-\sigma$ in \label{morawetz-eq-1} are optimal.
For the proofs of the optimality, one can take test functions
$u(x)=|x|^{-\sigma}$.
\end{remark}
\begin{proposition}[Morawetz estimates]\label{morawetz-prop-0}
Let $u$ be a finite-energy radial solution to \eqref{wave-La-p}. Then for any time $T_1<T_2$ we have
\begin{align}\nonumber
&\frac1{2R} \int_{T_1}^{T_2} \int_{|x|< R}
\Big(|\nabla u|^2 +a \frac{|u|^2}{|x|^2}+| u_t|^2 + \frac{(d-1)(p-1)-2}{p+1} |u|^{p+1} \Big) dxdt \\
& + \frac{d-1}{4R^2}\int_{T_1}^{T_2} \int_{|x|=R} |u|^2dS(x)dt + \int_{T_1}^{T_2}\int_{|x|\geq R} \Big(\frac{(d-1)(p-1)}{2(p+1)} \frac{|u|^{p+1}}{|x|}+ (a+\lambda_d) \frac{|u|^2}{|x|^3}\Big) dxdt \nonumber\\
& + \frac{ 1 }{ 2R^2} (\mu_d+a-2\sigma) \int_{|x|\leq R} \big(|u(x,T_1)|^2 +|u(x,T_2)|^2 \big) dx\nonumber \\
\leq& 2E + \frac12 \big(-\lambda_d+|a|\big) \int_{|x|>R} \frac{|u(x,T_1)|^2+|u(x,T_2)|^2}{|x|^2} dx,
\label{Morawetz}
\end{align}
where $\mu_d=\frac{d^2-1}{4}$, $\lambda_d=\frac{(d-1)(d-3)}{4}$.
\end{proposition}
\begin{proof}[Proof]
We may assume $u$ has sufficient regularity, otherwise we can use the standard techniques of smooth approximation.
Let
\begin{align}
M(t)= \int_{|x|\leq R} u_t \Big( x\cdot \nabla u + \frac{d-1}{2} u\Big) dx + R\int_{|x|\geq R} u_t \Big(\frac{x}{|x|} \cdot\nabla u + \frac{d-1}{2} \frac{u}{|x|}\Big) dx.
\end{align}
A straightforward calculation shows that the sum of double integrals in the left hand side of \eqref{Morawetz},
i.e.
\begin{equation}
\hbox{sum of first two lines of \eqref{Morawetz}} = -\frac{1}{R} \int_{T_1}^{T_2} M'(t) dt = \frac{1}{R}(M(T_1)-M(T_2)). \label{expression of left}
\end{equation}
The major difference between the cases with $a>0$ and $a<0$ is to find the upper bound of $|M(t)|$. By Cauchy-Schwarz and integration by parts, we have
\begin{align}
|M(t)| & \leq
R^{\frac12}\|u_t\|_2\Big[ \frac{1}{R}\int_{|x|\leq R} \Big|x\cdot\nabla u+ \frac{d-1}{2} u\Big|^2dx +R \int_{|x|\geq R} \Big|\frac{x}{|x|} \cdot\nabla u + \frac{d-1}{2} \frac{u}{|x|} \Big|^2dx \Big]^\frac12 \nonumber\\
& \leq
R^{\frac12}\|u_t\|_2\Big[ \frac{1}{R}\int_{|x|\leq R} \Big( |x\cdot\nabla u |^2 -\mu_d|u|^2\Big)dx +R \int_{|x|\geq R} \Big(\Big|\frac{x}{|x|} \cdot\nabla u \Big|^2 -\lambda_d \frac{|u|^2}{|x|^2}\Big)dx \Big]^\frac12 \nonumber\\
& \leq \frac{R}{2} \|u_t\|_2^2+ \frac{1}{2R}\Big[\int_{|x|\leq R} \Big( |x\cdot\nabla u |^2 -\mu_d|u|^2\Big)dx + R^2 \int_{|x|\geq R} \Big( \Big|\frac{x}{|x|}\cdot \nabla u \Big|^2 -\lambda_d \frac{|u|^2}{|x|^2} \Big)dx\Big]\nonumber\\
& \doteq \frac{R}{2} \|u_t\|_2^2+\frac{1}{2R}I.\label{upper Mt}
\end{align
We may find an upper bound of $I$
\begin{align*}
I\leq & \int_{|x| \leq R} \Big( |x|^2 \cdot |\nabla u |^2 + a|u|^2\Big)dx -(\mu_d+a) \int_{|x|\leq R} |u|^2 dx\\
& \qquad + R^2 \int_{|x|\geq R} \Big(|\nabla u|^2 + a\frac{|u|^2}{|x|^2}\Big) dx -(\lambda_d+a) R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx\\
= & \int_{\Rm^d} \min\{|x|^2, R^2\}\Big(|\nabla u|^2 + a\frac{|u|^2}{|x|^2}\Big) dx -(\mu_d+a) \int_{|x|\leq R} |u|^2 dx -(\lambda_d+a) R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx\\
= & \int_0^R 2r\Big[ \int_{|x|>r} \Big(|\nabla u|^2 + a\frac{|u|^2}{|x|^2}\Big) dx \Big] dr -(\mu_d+a) \int_{|x|\leq R} |u|^2 dx-(\lambda_d+a)R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx.
\end{align*}
One easily obtain by Corollary \ref{lower bound inner}
\begin{align*}
\int_{|x|>r} \Big(|\nabla u|^2 + a\frac{|u|^2}{|x|^2}\Big) dx & = \mathcal{E} - \int_{|x|<r} \Big(|\nabla u|^2 + a\frac{|u|^2}{|x|^2}\Big) dx\\
& \leq \mathcal{E} + \frac{\sigma }{r}\int_{|x|=r} |u(x)|^2 dS(x).
\end{align*}
where $\mathcal{E} = \int_{\Rm^d} \big(|\nabla u(x)|^2 + a\frac{|u|^2}{|x|^2}\big)dx$. We apply this upper bound to $I$ and obtain
\begin{align*}
I & \leq R^2 \mathcal{E} + 2 \sigma\int_0^R \Big[\int_{|x|=r} |u(x)|^2 dS(x)\Big] dr -(\mu_d+a) \int_{|x|\leq R} |u|^2 dx-(\lambda_d+a)R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx\\
& = R^2 \mathcal{E} -(\mu_d+a-2\sigma) \int_{|x|\leq R} |u|^2 dx -(\lambda_d+a) R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx\\
& = R^2 \int_{\Rm^d} \left(|\nabla u(x)|^2 + a\frac{|u|^2}{|x|^2}\right)dx -(\mu_d+a-2\sigma)\int_{|x|\leq R} |u|^2 dx -\left(\lambda_d+a\right) R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx.
\end{align*}
Inserting this upper bound in \eqref{upper Mt}, we have
\begin{align*}
|M(t)| & \leq \frac{R}{2} \int_{\Rm^d}\Big( |u_t|^2+|\nabla u(x)|^2 + a\frac{|u|^2}{|x|^2}\Big)dx\\
& \quad -\frac{1}{2R}\Big[ (\mu_d+a-2\sigma)\int_{|x|\leq R} |u|^2 dx { +} \Big(\lambda_d+a\Big) R^2 \int_{|x|>R} \frac{|u|^2}{|x|^2} dx\Big]\\
& \leq RE - \frac{ 1 }{2R} (\mu_d+a-2\sigma) \int_{|x|\leq R} |u|^2 dx - \Big(\lambda_d+a\Big) \frac{ R}{2} \int_{|x|>R} \frac{|u|^2}{|x|^2} dx.
\end{align*}
Combining this upper bound with \eqref{expression of left},
we finish the proof of Proposition \ref{morawetz-prop-0}.
\end{proof}
\begin{corollary} \label{retard fast energy}
Let $u$ be a finite-energy radial solution to \eqref{wave-La-p}. Then for any $0<R<T$ we have
\begin{align}
& \int_R^T \int_{|x|< R} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dx dt \nonumber\\
&\quad \leq \int_{-R}^R \int_{|x|> R} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dxdt\nonumber\\
& \qquad +R(|a|-\lambda_d)\Big[\int_{-R}^T \int_{|x|>R} \frac{|u|^2}{|x|^3} dxdt
+ \frac{1}{2} \int_{|x|>R} \frac{|u(x,-R)|^2+ |u(x,T)|^2}{|x|^2} dx \Big]. \label{morawetz-eq-6}
\end{align}
\end{corollary}
\begin{proof}
We apply Morawetz estimate \eqref{Morawetz} with $T_1 = -R$ and $T_2 = T$. Our assumption $p\geq 1+\frac{4}{d-1}$ means that $\frac{(d-1)(p-1)-2}{2(p+1)} \geq \frac{1}{p+1}$.
And by the assumption of $a$ and $d$ in Proposition \ref{well-p}, we have
$\mu_d+a-2\sigma\geq\frac34. $
Thus we may ignore all other positive terms in the left hand side of Morawetz estimate \eqref{Morawetz} and obtain
\begin{align*}
\frac{1}{R}\int_{ -R}^T \int_{|x|< R} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dxdt \leq 2E +(|a|-\lambda_d) I,
\end{align*}
where $I$ represents the sum of three integrals enclosed in middle brackets above. We may multiply $R$ in both sides and utilize the energy conservation law
to get
\begin{align*}
& \int_{ -R}^T \int_{|x|< R} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dxdt \\
& \leq \int_{-R}^R \int_{\Rm^d} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dx dt +(|a|-\lambda_d)RI.
\end{align*}
Finally,
subtracting $\int_{-R}^R \int_{|x|<R} \Big(\frac{1}{2}|\nabla u|^2 +\frac{a}{2} \frac{|u|^2}{|x|^2}+\frac{1}{2}| u_t|^2 + \frac{1}{p+1}|u|^{p+1}\Big) dx dt$ from both sides,
we obtain \eqref{morawetz-eq-4}. This completes the proof of Corollary \ref{retard fast energy}.
\end{proof}
\section{Method of Characteristic Lines}\label{mtheod-characteristic}
\setcounter{section}{5}\setcounter{equation}{0}
In this section, we prove Theorem \ref{main 1} by method of characteristic lines.
First, we import radiation
field theorem
of the free wave equation.
\begin{proposition}[Radiation field, \cite{DKM-2019,Friedlander-1962,Friedlander-1980,shenhdradial}]\label{radiation}
For any radial free wave solution $u$ with finite energy, there exists a
unique function $g_+\in L^2(\Rm)$ such that
\begin{equation}\label{characteristic-eq-1}
\lim_{t\to \infty} \int_0^\infty \Big( \Big| r^{\frac{d-1}{2}} u_r(r,t)+g_+(t-r) \Big|^2 + \Big| r^{\frac{d-1}{2}} u_t(r,t)-g_+(t-r) \Big|^2 \Big)dr=0.
\end{equation}
On the other hand, for $g_+\in L^2(\Rm)$, there exists a
unique radial free wave solution $u$ with finite energy
such that the equality \eqref{characteristic-eq-1}
holds.
\end{proposition}
Next, we recall the decay property of free waves.
\begin{proposition}[\cite{shenhdradial}]\label{decay-out-cone}
Assume $d\geq3$. Let $u$ be a solution to the free wave equation
$\Big(\partial_t^2-\Delta\Big) v=0$ with initial data $\dot H^1\times L^2(\Rm^d).$ Then we have
\begin{equation}
\lim_{\eta\to\infty} \sup_{t>\eta}
\Big\{ \int_{|x|<t-\eta} \Big(|\nabla v(x,t)|^2+|v_t(x,t)|^2\Big)dx \Big\} =0,
\label{free limit 2}
\end{equation}
\vspace{-0.2cm}
\begin{equation}
\lim_{R \rightarrow \infty} \sup_{t>0} \Big\{\int_{|x|>t+R} \Big(|\nabla v(x,t)|^2 + |v_t(x,t)|^2\Big) dx \Big\} = 0, \label{free limit 3}
\end{equation}
\end{proposition}
Assume that $u$ is a finite-energy, radial solution to \eqref{wave-La-p}. Let $w = r^{\frac{d-1}{2}}u(r,t)$,
then, we have
\begin{equation}\label{characteristic-eq-2}
(\partial_t^2 -\partial_r^2)w=- r^{\frac{d-1}{2}} |u|^{p-1} u-(\lambda_d+a) r^{\frac{d-5}{2} } u.
\end{equation}
As in \cite{shenenergy,shenhdradial}, by the characteristic method and radial Sobolev, we have
\begin{proposition} \label{variation of w}
For $\tau+1 < t_1 < t_2 < s-1$, we have
\begin{align*}
|(w_t+w_r)(s-t_2,t_2)-(w_t+w_r)(s-t_1,t_1)| &\lesssim (s-t_2)^{-\beta(d,p)},\\
|(w_t-w_r)( t_2-\tau,t_2)-(w_t+w_r)(t_1-\tau,t_1)| &\lesssim (t_1-\tau)^{-\beta(d,p)},
\end{align*}
where $\beta(d,p)= \frac{(d-1)(p-1)-2}{2(p+1)}\in (0,\frac12)$ by the assumption of $p$.
\end{proposition}
In fact, the additional term $-(\lambda_d+a)r^{\frac{d-5}{2}}u(r,t)$ can be dealt with by $|u(r)| \lesssim r^{-(d-2)/2}$.
As a consequence of Proposition \ref{variation of w}, we have the following proposition
\begin{proposition}
There exist $g_\pm\in L^2(\Rm)$ satisfying that
\begin{align*}
&\lim_{t \rightarrow +\infty} (w_t-w_r)(t-\eta,t) = 2g_+(\eta),& &\|g_+\|_{L^2(\Rm)} \lesssim E/c_d;&\\
&\lim_{t \rightarrow -\infty} (w_t+w_r)(s-t,t) = 2g_-(s),& &\|g_-\|_{L^2(\Rm)} \lesssim E/c_d.&
\end{align*}
\end{proposition}
\noindent
\textbf{Proof of Theorem \ref{main 1}}\; As a consequence of Proposition \ref{variation of w}, one can verify that
\begin{equation}\label{characteristic-eq-3}
\limsup_{t\rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}}^{t+R} (|(w_t+w_r)(r,t)-2g_-(r+t)|^2 + |(w_t-w_r)(r,t)-2g_+(t-r)|^2) dr \lesssim c.
\end{equation}
where $1-\kappa_0=2\beta(d,p) $, $c, R$ are positive constants.
A simple calculation shows
$$\int_{t-c\cdot t^{1-\kappa_0}}^{t+R} |g_-(r+t)|^2 dr \rightarrow 0,\quad\;\text{\rm as}\quad\; t\to\infty.$$
Therefore, we have
$$ \limsup_{t\rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}}^{t+R} (|(w_t+w_r)(r,t)|^2 + |(w_t-w_r)(r,t)-2g_+(t-r)|^2) dr \lesssim c. $$
This can be rewritten as
$$ \limsup_{t\rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}}^{t+R} (|w_t(r,t)-g_+(t-r)|^2 + |w_r(r,t)+g_+(t-r)|^2) dr \lesssim c. $$
Note that $w_r = r^{\frac{d-1}{2}} u_r + {\frac{d-1}{2}} r^{\frac{d-3}{2}}u$ and
$$ \lim_{t\rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}}^{t+R} r^{d-3 }|u(r,t)|^2 dr \lesssim \lim_{t\rightarrow +\infty}\int_{t-c\cdot t^{1-\kappa_0}}^{t+R} r^{d-3 } r^{-\frac{4(d-1)}{p+3}}dr = 0,
$$
by Lemma \ref{pointwise estimate 2}.
Thus, we have
\begin{align}
& \limsup_{t\rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}}^{t+R} (|r^{\frac{d-1}{2}} u_t(r,t)-g_+(t-r)|^2 + |r^{\frac{d-1}{2}} u_r(r,t)+g_+(t-r)|^2) dr \lesssim c .\label{u middle 2}
\end{align}
Similarly we have for $\eta, R>0$
\begin{equation} \label{u middle 1}
\limsup_{t\rightarrow +\infty} \int_{t-\eta}^{t+R} (|r^{\frac{d-1}{2}} u_t(r,t)-g_+(t-r)|^2 + |r^{ \frac{d-1}{2}} u_r(r,t)+g_+(t-r)|^2) dr = 0.
\end{equation}
Next, by Proposition \ref{radiation}, there exists free wave $v$ such that
\begin{align}
& \lim_{t\rightarrow +\infty} \int_{0}^{\infty} (|r^{ \frac{d-1}{2}} v_t(r,t)-g_+(t-r)|^2 + |r^{ \frac{d-1}{2}} v_r(r,t)+g_+(t-r)|^2) dr = 0. \label{free limit 1}
\end{align}
Combining \eqref{free limit 1} with \eqref{u middle 1} or \eqref{u middle 2}, we have
\begin{align}
& \lim_{t \rightarrow +\infty} \int_{t-\eta<|x|<t+R} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx = 0, \label{difference middle 1}\\
& \limsup_{t \rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}<|x|<t+R} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \lesssim c. \label{difference middle 2}
\end{align}
Finally,
to complete the proof of Theorem \ref{main 1}, it suffices to show
\begin{equation}
\lim_{R \rightarrow +\infty} \sup_{t>0} \Big\{\int_{|x|>t+R} (|\nabla u(x,t)-\nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2) dx \Big\} = 0. \label{difference outside}
\end{equation}
In fact, this is a direct consequence of
the finite propagation speed.
One can also see \eqref{free limit 3}, or employ Lemma \ref{decay of tail} with $\kappa=0$ if the decay of energy is unknown.
\section{Further estimates with energy decay and the proof of Theorem \ref{main 2} }\label{further}
\setcounter{section}{6}\setcounter{equation}{0}
In this section, we consider the scattering theory of
\eqref{wave-La-p} under the condition of
additional energy decay \eqref{weight-energy}.
\begin{lemma} \label{decay of tail}
Let $\kappa \in [0,1)$. Assume that initial data $(u_0,u_1)\in \dot{H}^1 \times L^2$ satisfy
\[
E_\kappa (u_0,u_1) = \int_{\Rm^d} (|x|^\kappa + 1)\left(|\nabla u_0(x)|^2 + |u_1(x)|^2 + |u_0(x)|^{p+1}\right) dx < +\infty.
\]
Then the corresponding solution $u$ to \eqref{wave-La-p} satisfies for each $r>0$
\begin{align}\label{basicfurther-eq-0}
\int_{|x|>r+|t|} \Big(|\nabla u(x,t)|^2 + \frac{|u(x,t)|^2}{|x|^2} + |u_t(x,t)|^2 + |u(x,t)|^{p+1}\Big)dx \lesssim r^{-\kappa} E_{\kappa,r}(u_0,u_1),
\end{align}
where $E_{\kappa,r}(u_0,u_1)$ is defined by
\[
E_{\kappa,r}(u_0,u_1) \doteq \int_{|x|>r/2} (|x|^\kappa + 1)\left(|\nabla u_0(x)|^2 + |u_1(x)|^2 + |u_0(x)|^{p+1}\right) dx \rightarrow 0, \; \hbox{as}\; r \rightarrow +\infty.
\]
If the initial data is also radial, we have
\begin{align}
|u(x,t)|\lesssim |x|^{-\frac{2(d-1)}{p+3}}(|x|-|t|)^{-\frac{2\kappa}{p+3}} E_{\kappa,|x|-|t|}(u_0,u_1)^{\frac{2}{p+3}},\;\;|x|>|t|.\label{basicfurther-eq-1}
\end{align}
\end{lemma}
\begin{proof}
Let us fix a radial, smooth cut-off function $\phi: \Rm^d \rightarrow [0,1]$ satisfying
\[
\phi(x) = \left\{
\begin{array}{ll} 0, & |x|\leq 1/2; \\
1, & |x|\geq 1,\end{array}\right.
\]
and define $(u_{0,r}, u_{1,r}) = \phi(x/r)(u_0,u_1)$ for some $r>0$ to be determined later. A simple calculation shows
\begin{align*}
E(u_{0,r}, u_{1,r}) & \lesssim \int_{\Rm^d} (|\nabla u_{0,r}|^2 + |u_{1,r}|^2+|u_{0,r}|^{p+1}) dx \\
& \lesssim \int_{|x|>r/2} (|\nabla u_0|^2 + |u_1|^2+|u_0|^{p+1}) dx + \int_{r/2<|x|<r} \frac{|u_0|^2}{|x|^2} dx\\
& \lesssim \int_{|x|>r/2} (|\nabla u_0|^2 + |u_1|^2+|u_0|^{p+1}) dx\\
& \lesssim r^{-\kappa} E_{\kappa,r}(u_0,u_1).
\end{align*}
Let $u^{(r)}$ be the solution to \eqref{wave-La-p} with initial data $(u_{0,r}, u_{1,r})$. By energy conservation law we immediately have
\begin{align}
\int_{\Rm^d} \Big(|\nabla u^{(r)}|^2 + \frac{|u^{(r)}|^2}{|x|^2} + |u_t^{(r)}|^2+|u^{(r)}|^{p+1}\Big) dx & \lesssim E(u_{0,r}, u_{1,r}) \lesssim r^{-\kappa} E_{\kappa,r}(u_0,u_1). \label{energy estimate tail}
\end{align}
This proves \eqref{basicfurther-eq-0} since we always have $u(x,t) \equiv u^{(r)}(x,t)$ for $|x|>r+|t|$ by finite speed of propagation. If the data is also radial, then we may apply Lemma \ref{pointwise estimate 2} to conclude
\begin{align*}
|u(x,t)| & = |u^{(|x|-|t|)}(x,t)| \lesssim |x|^{-\frac{2(d-1)}{p+3}} \|u^{(|x|-|t|)}(\cdot,t)\|_{\dot{H}^1}^{\frac{2}{p+3}} \|u^{(|x|-|t|)}(\cdot,t)\|_{L^{p+1}}^{\frac{p+1}{p+3}}\\
& \lesssim |x|^{-\frac{2(d-1)}{p+3}} \left((|x|-|t|)^{-\kappa} E_{\kappa,|x|-|t|}(u_0,u_1)\right)^{\frac{2}{p+3}},
\end{align*}
where we have used \eqref{energy estimate tail} with $r = |x|-|t|$.
\end{proof}
\begin{corollary} \label{integral estimate kappa}
Let $(u_0,u_1)$ be as in Theorem \ref{main 2}. The corresponding solution $u$ satisfies
\begin{align*}
&\int_{|x|>R} \frac{|u(x,t)|^2}{|x|^2} dx \lesssim R^{-\frac{p+5}{p+3}\kappa_0} E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}},\;\;|t|\leq R;\\
&\int_{|x|>R} \frac{|u(x,t)|^2}{|x|^2} dx \lesssim \Big[|t|^{-\frac{p+5}{p+3}\kappa_0} +(|t|-R) R^{-\frac{4(d-1
)}{p+3} +d-3}\Big]E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}},\;\; |t|>R;\\
&\int_{-T}^T \int_{|x|>R} \frac{|u(x,t)|^2}{|x|^3} dx dt \lesssim \Big[R^{-\frac{p+5}{p+3}\kappa_0} +(T-R)^2 R^{-\frac{4(d-1)}{p+3}+d-4}\Big]E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}},\;\;R<T<2R.
\end{align*}
\end{corollary}
\begin{proof}
By the pointwise estimate given in Lemma \ref{decay of tail}, if $|t|\leq R$ we have
\begin{align*}
\int_{|x|>R} \frac{|u(x,t)|^2}{|x|^2} dx & \lesssim \int_{|x|>R} |x|^{-\frac{4(d-1)}{p+3}-2}(|x|-|t|)^{-\frac{4\kappa_0}{p+3}} E_{\kappa_0,|x|-|t|}(u_0,u_1)^{\frac{4}{p+3}} dx\\
& \lesssim R^{-\frac{p+5}{p+3}\kappa_0} E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}}.
\end{align*}
Similarly if $|t|>R$, we have
\begin{align*}
\int_{|x|>|t|} \frac{|u(x,t)|^2}{|x|^2} dx & \lesssim \int_{|x|>|t|} |x|^{-\frac{4(d-1)}{p+3}-2}(|x|-|t|)^{-\frac{4\kappa_0}{p+3}} E_{\kappa_0,|x|-|t|}(u_0,u_1)^{\frac{4}{p+3}} dx\\
& \lesssim |t|^{-\frac{p+5}{p+3}\kappa_0} E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}}.
\end{align*}
In addition, we apply Lemma \ref{pointwise estimate 2} to obtain
\begin{align*}
\int_{R<|x|<t} \frac{|u(x,T)|^2}{|x|^2} dx & \lesssim \int_{R<|x|<|t|} |x|^{-\frac{4(d-1)}{p+3}-2} E(u_0,u_1)^{\frac{4}{p+3}} dx\\
& \lesssim (|t|-R) R^{-\frac{4(d-1)}{p+3}+d-3} E_{\kappa_0} (u_0,u_1)^{\frac{4}{p+3}}.
\end{align*}
These prove the first two inequalities. The last inequality is a direct consequence of the first two because we have
\[
\int_{-T}^T \int_{|x|>R} \frac{|u(x,t)|^2}{|x|^3} dx dt \leq \frac{1}{R} \int_{-T}^T \Big[\int_{|x|>R} \frac{|u(x,t)|^2}{|x|^2} dx\Big] dt.
\]
\end{proof}
\begin{proposition} \label{energy estimate inside}
Let $(u_0,u_1)$ be as in Theorem \ref{main 2}. The corresponding solution $u$ satisfies
\begin{equation}
\lim_{t\rightarrow +\infty} \int_{|x|<t-c\cdot t^{1-\kappa_0}} \Big(|\nabla u|^2 + \frac{|u|^2}{|x|^2} + |u_t|^2 +|u|^{p+1}\Big) dx = 0. \label{basicfurther-eq-3}
\end{equation}
\end{proposition}
\begin{proof}
It suffices to show
\begin{equation}
\limsup_{t\rightarrow +\infty} \int_{|x|<t-c\cdot t^{1-\kappa_0}} \Big(\frac{1}{2}|\nabla u|^2 + \frac{a}{2}\frac{|u|^2}{|x|^2} + \frac{1}{2}|u_t|^2 +\frac{1}{p+1}|u|^{p+1}\Big) dx \leq 0,
\label{basicfurther-eq-4}
\end{equation}
because of Lemma \ref{cutoff energy} and the inequality $ |u(x,t)| \lesssim |x|^{-\frac{2(d-1)}{p+3}}$. By energy flux formula given in Proposition \ref{energy flux} we have
$$
E(t'; B(t'-c\cdot t^{1-\kappa_0})) \geq E(t; B(0, t-c\cdot t^{1-\kappa_0})) + \frac{a}{2\sqrt{2}}\int_{\Sigma(t,t')} \frac{|u|^2}{|x|^2} dS,
$$
where
$$\Sigma(t,t')= \{(\bar{x},\bar{t}): |\bar{x}| = |\bar{t}|-c\cdot t^{1-\kappa_0}, t<\bar{t}<t'\}$$
is a part of the light cone.
We may also enlarge the radius and obtain for $t' \in (t,t+c\cdot t^{1-\kappa_0})$
\begin{align*}
E(t'; B(0,t)) & \geq E(t'; B(t'-c\cdot t^{1-\kappa_0}) + \frac{a}{2} \int_{t'-c\cdot t^{1-\kappa_0}<|x|<t} \frac{|u(x,t')|^2}{|x|^2} dx\\
& \geq E(t; B(0, t-c\cdot t^{1-\kappa_0})) + \frac{a}{2\sqrt{2}}\int_{\Sigma(t,t')} \frac{|u|^2}{|x|^2} dS + \frac{a}{2} \int_{t'-c\cdot t^{1-\kappa_0}<|x|<t} \frac{|u(x,t')|^2}{|x|^2} dx.
\end{align*}
By the pointwise estimate of radial solutions $|u(x,t)| \lesssim |x|^{-\frac{2(d-1)}{p+3}}$, we have
\[
\int_{\Sigma(t,t')} \frac{|u|^2}{|x|^2} dS + \int_{t'-c\cdot t^{1-\kappa_0}<|x|<t} \frac{|u(x,t')|^2}{|x|^2} dx \lesssim c \cdot t^{-\frac{4(d-1)}{p+3}+d-2-\kappa_0}, \;\; t' \in (t, t+c\cdot t^{1-\kappa_0}).
\]
Thus we have
\begin{align*}
E(t; B(t-c\cdot t^{1-\kappa_0}) \leq \frac{1}{c\cdot t^{1-\kappa_0}} \int_t^{t+c\cdot t^{1-\kappa_0}} E(t'; B(0,t))dt' + C t^{-\frac{4(d-1)}{p+3}+{ d-2}-\kappa_0}.
\end{align*}
Here $C$ is independent of $t$ and $\frac{4(d-1)}{p+3}-{ (d-2)}> 0$ since $p<1+\frac{4}{ d-2}$. As a result, we only need to prove
\begin{equation}
\limsup_{t\rightarrow +\infty} \frac{1}{c\cdot t^{1-\kappa_0}} \int_t^{t+c\cdot t^{1-\kappa_0}} E(t'; B(0,t))dt' \leq 0. \label{to prove inner 2}
\end{equation}
Next we recall Corollary \ref{retard fast energy} (with $R=t$, $T=t+c\cdot t^{1-\kappa_0}$) and write
\begin{align}
\int_t^{t+c\cdot t^{1-\kappa_0}} E(t'; B(0,t))dt' \leq &\int_{-t}^t E(t'; \{x\in \Rm^d: |x|>t\}) dt'\nonumber\\
& +t(|a|-\lambda_d)\int_{-t}^{t+c\cdot t^{1-\kappa_0}} \int_{|x|>t} \frac{|u|^2}{|x|^3} dxdt'\nonumber\\
&+ \frac{1}{2}t(|a|-\lambda_d) \int_{|x|>t} \frac{|u(x,-t)|^2+ |u(x,t+c\cdot t^{1-\kappa_0})|^2}{|x|^2} dx\nonumber\\
=& I_1 +I_2 +I_3. \label{basicfurther-eq-6}
\end{align}
According to Corollary \ref{integral estimate kappa}, we have
\begin{align} { |I_2+I_3|} \lesssim t^{1-\frac{p+5}{p+3}\kappa_0 }+ c t^{1-\kappa_0-\frac{4(d-1)}{p+3} +{ d-2}} +c^2 t^{2(1-\kappa_0)-\frac{4(d-1)}{p+3} +d-3 }. \label{basicfurther-eq-7}
\end{align}
By the assumption of $p$, we have $d-2-\frac{4(d-1)}{p+3}<0$.
Thus
\begin{align}\lim_{t \to \infty} \frac{1}{c\cdot t^{1-\kappa_0}}(I_2+I_3) = 0.\label{basicfurther-eq-7}
\end{align}
In order to deal with $I_1$,
note that $E(t'; \{x\in \Rm^d: |x|>t\})$ is always bounded, we obtain by Lemma \ref{decay of tail}
\begin{align}
I_1 & \lesssim \int_{-t+t^{(1-\kappa_0)/2}}^{t-t^{(1-\kappa_0)/2}} E(t'; \{x\in \Rm^d: |x|>t\}) dt' + t^{(1-\kappa_0)/2} \nonumber\\
& \lesssim \int_{-t+t^{(1-\kappa_0)/2}}^{t-t^{(1-\kappa_0)/2}} (t-|t'|)^{-\kappa_0} E_{\kappa_0, t-|t'|}(u_0,u_1) dt' + t^{(1-\kappa_0)/2}\nonumber\\
& \lesssim t^{1-\kappa_0} E_{\kappa_0, t^{(1-\kappa_0)/2}}(u_0,u_1) + t^{(1-\kappa_0)/2}.
\label{basicfurther-eq-8}
\end{align}
Finally we may plug the upper bounds of $I_1, I_2+I_3$ to \eqref{basicfurther-eq-6}, and recall the fact
$$E_{\kappa_0, r}(u_0,u_1) \rightarrow 0, \quad\text{\rm as}\quad\; r\rightarrow +\infty, $$
we obtain \eqref{to prove inner 2}. This complete the proof of Proposition \ref{energy estimate inside}.
\end{proof}
Now, {\bf we are ready to prove Theorem \ref{main 2}}.
Let $u$ be a radial solution to \eqref{wave-La-p} with initial data $(u_0,u_1)$ satisfying $ E_{\kappa_0}(u_0,u_1) \lesssim E_{\kappa} (u_0,u_1)< +\infty$. Then, for $v$ given as in Section \ref{mtheod-characteristic}, we have
\begin{align*}
\limsup_{t \rightarrow +\infty} \int_{\Rm^d} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \leq L_1 + L_2 + L_3,
\end{align*}
with
\begin{align*}
L_1 = & \limsup_{t \rightarrow +\infty} \int_{|x|<t-c\cdot t^{1-\kappa_0}} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx; \\
L_2 = & \limsup_{t \rightarrow +\infty} \int_{t-c\cdot t^{1-\kappa_0}<|x|<t+R} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx;\\
L_3 = & \limsup_{t \rightarrow +\infty} \int_{|x|>t+R} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx,
\end{align*}
where $c,R>0$ are constants to be determined.
We have already known $L_3 = 0$ and $L_2 \lesssim c$ by the conclusion of Theorem \ref{main 1} and \eqref{difference middle 2}. We also have $L_1 = 0$ by a combination of Proposition \ref{energy estimate inside} and \eqref{free limit 2}. In summary we have
\[
\limsup_{t \rightarrow +\infty} \int_{\Rm^d} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \lesssim c
\]
for all positive
constants
$c$. We then make $c\rightarrow 0$ and finish the proof.
\section{Appendix}
\setcounter{section}{7}\setcounter{equation}{0}
\subsection{ The asymptotic behaviour of linear solution}
Let $u$ be a finite-energy radial solution to the linear wave equation with an inverse-square potential
\[
\left\{\begin{array}{l} u_{tt} +\Big(- \Delta + \frac{a}{|x|^2} \Big) u = 0, \qquad (x,t)\in \Rm^d \times \Rm,\\
(u,u_t)|_{t=0} = (u_0,u_1) \in \dot{H}^1 \times L^2. \end{array}\right.
\]
Here $d \geq 3$ and $a > -\tfrac{(d-2)^2}{4}$. We prove that there exists a free wave $v$ to the linear wave equation $\partial_t^2 v - \Delta v = 0$ so that
\begin{equation} \label{appendix-eq-01}
\lim_{t \rightarrow +\infty} \|(u(\cdot,t), u_t(\cdot,t)) - (v(\cdot,t), v_t(\cdot,t))\|_{\dot{H}^1 \times L^2 } = 0.
\end{equation}
The proof of \eqref{appendix-eq-01} divides into two steps.
We first prove
\begin{equation} \label{outer part linear}
\limsup_{t \rightarrow +\infty} \int_{|x|>(1-\delta)t} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \lesssim \delta.
\end{equation}
Following the same argument as in Lemma \ref{decay of tail}, one can obtain
\begin{align}
&\int_{|x|>|t|+r} \Big(|\nabla u(x,t)|^2 + \frac{|u(x,t)|^2}{|x|^2} + |u_t(x,t)|^2\Big) dx \lesssim E_{0,r}, \label{tail estimate 1 linear}\\
&|u(x,t)| \lesssim E_{0,r}^{1/2} |x|^{-\frac{d-2}{2}}, \quad |x|\geq r+|t|\label{tail estimate 1 linear-0}
\end{align}
such that
\begin{equation}\label{appendix-eq-1}
E_{0,r} = \int_{|x|>r/2} \Big(|\nabla u_0(x)|^2 + |u_1(x)|^2\Big) dx \rightarrow 0, \quad \hbox{as}\; r\rightarrow \infty.
\end{equation}
We also have by a direct computation
\begin{align*}
|u(r_1,t) - u(r_2,t)| &= \int_{r_1}^{r_2} |u_r (r,t)| dr \lesssim \Big(\int_{r_1}^{r_2} r^{d-1} |u_r(r,t)|^2 dr\Big)^{1/2} \Big(\int_{r_1}^{r_2} r^{-(d-1)} dr\Big)^{1/2} \\
& \lesssim (r_2-r_1)^{1/2} r_1^{-\frac{d-1}{2}}.
\end{align*}
Thus if $(1-\delta)|t|\leq r\leq |t|+r_0\leq (1+\delta)|t|$ with a small constant $\delta \ll 1$, then we have
\begin{equation} \label{improved pointwise H1}
|u(r,t)| \leq |u(|t|+r_0,t)| + |u(r,t)- u(|t|+r_0,t)| \lesssim (E_{0,r_0}^{1/2} + \delta^{1/2}) r^{-\frac{d-2}{2}}.
\end{equation}
Define $w = r^{\frac{d-1}{2}} u(r,t)$, then $w$ satisfies
\begin{equation}\label{appendix-eq-02}
(\partial_t^2 -\partial_r^2)w = -(\lambda_d+a) r^{\frac{d-5}{2}} u.
\end{equation}
Note that $|u(r,t)| \lesssim r^{-\frac{d-2}{2}}$, applying method of characteristic lines, we can obtain $g_+, g_- \in L^2 (\Rm)$ such that
\begin{align*}
&|(w_t+w_r)(r,t)-2g_-(r+t)| \lesssim r^{-1/2},& &r>1;&\\
&|(w_t-w_r)(r,t)-2g_+(t-r)| \lesssim r^{-1/2},& &r>1.&
\end{align*}
Thus we have ($\delta \ll 1$, $R>0$ are constants)
\[
\limsup_{t\rightarrow +\infty} \int_{(1-\delta)t}^{t+R} (|w_t(r,t)-g_+(t-r)|^2 + |w_r(r,t)+g_+(t-r)|^2) dr \lesssim \delta.
\]
By the identity $w_r = r^{\frac{d-1}{2}} u_r + \frac{d-1}{2} r^{\frac{d-3}{2}} u$ and the upper bound
\[
\limsup_{t\rightarrow +\infty} \int_{(1-\delta)t}^{t+R} r^{d-3} |u(r,t)|^2 dr \lesssim \limsup_{t\rightarrow +\infty} \int_{(1-\delta)t}^{t+R} \frac{1}{r} dr \lesssim \delta.
\]
We have
\[
\limsup_{t\rightarrow +\infty} \int_{(1-\delta)t}^{t+R} (|r^{\frac{d-1}{2}} u_t(r,t)-g_+(t-r)|^2 + |r^{\frac{d-1}{2}} u_r(r,t)+g_+(t-r)|^2) dr \lesssim \delta.
\]
By radiation fields, there exists a free wave $v$ so that
\[
\limsup_{t \rightarrow +\infty} \int_{(1-\delta)t<|x|<t+R} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \lesssim \delta.
\]
Combining this with \eqref{tail estimate 1 linear}, we obtain \eqref{appendix-eq-01}.
Now we are in position to show
\begin{equation} \label{inner part linear}
\limsup_{t \rightarrow +\infty} \int_{|x|<(1-\delta)t} \Big(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\Big) dx \lesssim \delta.
\end{equation}
By \eqref{free limit 2}, \eqref{improved pointwise H1},
Lemma \ref{cutoff energy}
and an argument similar to Proposition \ref{energy estimate inside}, it suffices to show
\begin{equation} \label{energy retard estimate linear}
\limsup_{t\rightarrow \infty} \frac{1}{\delta t} \int_{t}^{t+\delta t} E_0(t'; B(0,t)) dt' \lesssim \delta.
\end{equation}
In this argument we need to apply \eqref{improved pointwise H1} in order to deal with
the integrals of $\frac{|u|^2}{|x|^2}$, and more importantly, $\int_{|x|=t} \tfrac{|u(x,t)|^2}{|x|} dS(x)$.
$E_0$ represents the energy in a region:
\[
E_0(t; \Sigma) = \int_{\Sigma} \Big(\frac{1}{2}|\nabla u(x,t)|^2 + \frac{a}{2}\cdot \frac{|u(x,t)|^2}{|x|^2} + \frac{1}{2}|u_t(x,t)|^2\Big) dx.
\]
We may prove a Morawetz estimate for linear solutions. Following the same argument as in Corollary \ref{retard fast energy}, we have
\begin{align}
& \int_{t}^{(1+\delta)t} E_0(t'; B(0,t))dt' \leq \int_{-t}^t E_0(t'; \{x: |x|>t\}) dt' \nonumber\\
& \qquad {+t(|a|-\lambda_d)\Big[\int_{-t}^{t+\delta t} \int_{|x|>t} \frac{|u|^2}{|x|^3} dxdt'
+ \frac{1}{2} \int_{|x|>t} \frac{|u(x,-t)|^2+ |u(x,t+\delta t)|^2}{|x|^2} dx \Big]} = I_1 + I_2. \label{I1I2}
\end{align}
By \eqref{tail estimate 1 linear} we have
\begin{align*}
I_1 &\lesssim \delta^2 t + \int_{-t+\delta^2 t}^{t-\delta^2 t} \int_{|x|>t} \Big(|\nabla u(x,t')|^2 +\frac{|u(x,t')|^2}{|x|^2} + |u_t(x,t')|^2\Big) dx dt'\\
& \lesssim \delta^2 t + t E_{0,\delta^2 t}.
\end{align*}
Next we deal with $I_2$
\begin{align*}
I_2 & \lesssim \int_{-t}^{t+\delta t} \int_{|x|>t} \frac{|u|^2}{|x|^2} dxdt' + t \int_{|x|>t} \frac{|u(x,-t)|^2}{|x|^2} dx + t \int_{|x|>t} \frac{|u(x,t+\delta t)|^2}{|x|^2} dx\\
& \lesssim \int_{-t}^{t+\delta t} \int_{|x|>|t'|+\delta^2 t} \frac{|u|^2}{|x|^2} dxdt' + \int_{-t}^{-t+\delta^2 t} \int_{t<|x|<|t'|+\delta^2 t} \frac{|u|^2}{|x|^2} dxdt' +
\int_{
t-\delta^2 t
}^{t+\delta t} \int_{t<|x|<t'+\delta^2 t} \frac{|u|^2}{|x|^2} dxdt'\\
& \qquad + t \int_{|x|>t+\delta^2 t} \frac{|u(x,-t)|^2}{|x|^2} dx + t \int_{t<|x|<t+\delta^2 t} \frac{|u(x,-t)|^2}{|x|^2} dx\\
& \qquad + t \int_{|x|>t+\delta t + \delta^2 t} \frac{|u(x,t+\delta t)|^2}{|x|^2} dx + t \int_{t<|x|<t+\delta t + \delta^2 t} \frac{|u(x,t+\delta t)|^2}{|x|^2} dx \\
& = (I_{2,1} + I_{2,2} + I_{2,3}) + (I_{2,4} + I_{2,5}) + (I_{2,6} + I_{2,7})
\end{align*}
We apply \eqref{tail estimate 1 linear} and obtain $I_{2,1} + I_{2,4} + I_{2,6} \lesssim t E_{0,\delta^2 t}$. By $|u(x,t)| \lesssim |x|^{-\frac{d-2}{2}}$ we also have
\[
\int_{r_1<|x|<r_2} \frac{|u(x,t)|^2}{|x|^2} dx \lesssim \int_{r_1<|x|<r_2} |x|^{-d} dx \lesssim \ln(r_2/r_1).
\]
This gives $I_{2,2} + I_{2,3}+I_{2,5} \lesssim \delta^2 t$. Finally we may utilize \eqref{improved pointwise H1} and obtain
\[
I_{2,7} \lesssim t \int_{t<|x|<t+\delta t + \delta^2 t} (\delta + E_{0,\delta^2 t})|x|^{-d} dx \lesssim \delta^2 t + \delta t E_{0,\delta^2 t}.
\]
In summary we have $I_2 \lesssim \delta^2 t + t E_{0,\delta^2 t}$. Plugging the upper bound of $I_1, I_2$ in \eqref{I1I2}, we have
\[
\int_{t}^{(1+\delta)t} E_0(t'; B(0,t))dt' \lesssim \delta^2 t + t E_{0,\delta^2 t}.
\]
This proves \eqref{energy retard estimate linear}, thus \eqref{inner part linear}. Finally we combine \eqref{outer part linear} with \eqref{inner part linear} to conclude
\[
\limsup_{t \rightarrow +\infty} \int_{\Rm^d} \left(|\nabla u(x,t) - \nabla v(x,t)|^2 + |u_t(x,t)-v_t(x,t)|^2\right) dx \lesssim \delta..
\]
We make $\delta\rightarrow 0$ and finish the proof of \eqref{outer part linear}.
\subsection{Square function estimates}
In this subsection, employing the Stein-Tomas restriction theorem in Lorentz space, we show
\begin{equation}\label{square-ets}
\Big\|\int_{\Rm} e^{is|\nabla|} F(\cdot,s) ds \Big\|_{\dot H^\frac12(\Rm^d)}
\leq C \|F\|_{L^{\tfrac{2d}{d+2},2 }_x L^2_t(\Rm\times\Rm^d) },
\end{equation}
for radial function $F\in {L^{\tfrac{2d}{d+2},2 }_x L^2_t(\Rm\times\Rm^d) }$.
\begin{proof}
The proof follows the arguments of \cite{RV-1993}.
By duality, it suffices to show
\begin{equation}
\| |\nabla |^{-\gamma}e^{it|\nabla|} f \|_{L^{q,2}_x L^2_t(\Rm\times\Rm^d)} \leq C \|f\|_{L^2(\Rm^d)},
\end{equation}
where $\gamma=\frac{d}2-\frac{1}{2}-\frac{d}{q}$ and $q\geq \frac{2(d+1)}{d-1}$.
Utilizing the polar coordinates, we have
\[
|\nabla |^{-\gamma}e^{it|\nabla|} f(x) =\int_0^\infty e^{it r} r^{-\gamma} \int_{|\xi|=r} e^{ix \xi} \hat f(\xi) d\sigma^{n-1}_r(\xi) dr.
\]
By Plancherel theorem with respective to variable $t$, we have
\begin{align*}
\| |\nabla |^{-\gamma}e^{it|\nabla|} f(x) \|_{L^{q,2}_xL^2_t} = &
\Big\| \Big( \int_0^\infty \Big| \int_{|\xi|=r} e^{ix \xi} \hat f(\xi) d\sigma^{n-1}_r(\xi) \Big|^2 r^{-2\gamma} dr \Big)^\frac12\Big\|_{L^{q,2}_x}\\
\leq &
\Big( \int_0^\infty \Big\| \int_{|\xi|=r} e^{ix \xi} \hat f(\xi) d\sigma^{n-1}_r(\xi) \Big\|_{L^{q,2}_x}^2 r^{-2\gamma} dr \Big)^\frac12 \\
\leq & \Big( \int_0^\infty \int_{|\xi|=r} | \hat f(\xi) |^2 d\sigma^{n-1}_r(\xi) dr \Big)^\frac12
\\
= & \|f\|_{L^2_x(\Rm^d)}.
\end{align*}
where we used the Stein-Tomas restriction theorem of the Lorentz spaces in \cite{Bak-Segger-2011} and the following Minkowski inequality.
\begin{align}\label{add-1}
\|f(x,y)\|_{L^{p,r}_x L^r_y} \le \|f(x,y)\|_{ L^r_y L^{p,r}_x}, \;\; 1\leq r< p <\infty.
\end{align}
This completes the proof the square function estimate in lorentz space.
For the sake of completeness, we give a proof of the Minkowski inequalities in Lorentz space. In fact, for $1\leq r< p<\infty$, we have
\begin{align*}
\|f(x,y)\|_{L^{p,r}_x L^r_y} = &
\Big\| |f(x,y)|^r \Big\|^{\frac1r}_{L^{\frac{p}{r},1}_x L^1_y} \\
\approx & \Big( \sup_{\|g\|_{L^{(\frac{p}{r})',\infty}} \leq 1 } \int |g(x)|\int |f(x,y)|^r dy dx \Big)^\frac1r \\
= & \Big( \sup_{\|g\|_{L^{(\frac{p}{r})',\infty}} \leq 1 } \int \int |f(x,y)|^r |g(x)| dx dy \Big)^\frac1r \\
\leq & \| |f|^r\|_{L^1_y L^{\frac{p}{r},1}_x}^\frac1r = \|f(x,y)\|_{ L^r_y L^{p,r}_x}.
\label{add-2}\end{align*}
\end{proof}
\textbf{Acknowledgments:} C. Miao is supported in part by the National Natural Science Foundation of China under grant No.11831004, No.11926303, R. Shen is supported by the National Natural Science Foundation of China under grant No.11771325, T. Zhao is supported in part by the Chinese
Postdoc Foundation Grant 2019M650457.
| 3796aa224efe77bf55666d2854c933c7b6a46036 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Crowdsourced software engineering offers many opportunities for reducing time-to-market, producing alternative solutions, employing experts, learning through work, and democratizing participation in software engineering. There are several types of crowdsourced software engineering. One of the oldest and most common is open source software development. Another approach is
competition-based crowdsourcing, where platforms such as TopCoder have increasingly become very popular with over 1,500,000 users.
A more recent form of crowdsourced software engineering is microtask programming.
Microtask programming decontextualizes work into self-contained microtasks, reducing the context necessary to onboard onto a software project and thereby decreasing joining barriers. At the same time, it may reduce the time to market for completing software work through parallelism.
Several prior systems have explored approaches for microtasking programming work, using either manual or automatic approaches for decomposing programming tasks into microtasks. Manual approaches rely on a developer~\cite{codeon2017} or client to author each microtask~\cite{lasecki2015apparition}. Microtask programming environments can reduce onboarding barriers through preconfigured web-based environments, such as Codepilot, CrowdCode, and Collabode\cite{CodePilot:2017,Collabode:2011,latoza2018microtask}.
However, existing approaches for crowdsourced software development have significant limitations.
Open-source software development and competition-based approaches suffer from onboarding barriers, both technical and social. Although there are countless examples of successful open-source projects, onboarding challenges for newcomers can make it difficult to quickly onboard new developers and dissuade casual contributors.
Microtask programming approaches can reduce onboarding barriers, both by offering a preconfigured environment as well as by enabling developers to do programming work with less prior knowledge or awareness of the complete project. But existing approaches are still limited in their support for design and architecture activities necessary to scale to larger software projects. Moreover, decontextualizing programming work is hard, bringing with it many challenges in doing it effectively. For example, conflicts may occur when two crowd workers make conflicting assumptions, necessitating approaches to reduce or repair conflicts as they occur~\cite{latoza2018microtask}.
In my work, I have been exploring new ways to increase the \textit{scale} of microtask programming. I have developed a new behavior-driven development approach to microtask programming and conducted a series of studies to investigate the costs and benefits of microtask programming. In future work, I expect to continue to work to scale microtask programming to larger and more complex projects.
\begin{figure*}
\includegraphics[width=\textwidth,width=18cm, height= 8cm,keepaspectratio,clip]{figures/Implement_task2.pdf}
\centering
\caption{In our behavior-driven microtask programming approach, developers complete microtasks where they (1) identify a behavior, (2) write a test for the behavior, and (3) implement the behavior~\cite{aghayi2019crowdsourced}. }
~\label{fig:workflow}
\end{figure*}
\section{Crowdsourced Behavior-Driven Development}
To make microtask programming more efficient and reduce the potential for conflicts between contributors, I developed a new behavior-driven approach to microtasking programming. In our approach, each microtask asks developers to identify a behavior behavior from a high-level description of a function, implement a unit test for it, implement the behavior, and debug it. It enables developers to work on functions in isolation through high-level function descriptions and stubs.
In addition, I developed the first approach for building microservices through microtasks. Building microservices through microtasks is a good match because our approach requires a client to first specify the functionality the crowd will create through an API. This API can then take the form of a microservice description. A traditional project may ask a crowd to implement a new microservice by simply describing the desired behavior in a API and recruiting a crowd. We implemented our approach in a web-based IDE, \textit{Crowd Microservices}\footnote{https://youtu.be/mIn2EOqsDYw} (Fig.~\ref{fig:workflow}). It includes an editor for clients to describe the system requirements through endpoint descriptions as well as a web-based programming environment where crowd workers can identify, test, implement, and debug behaviors (Figure \ref{fig:workflow}). The system automatically creates, manages, assigns microtasks. After the crowd finishes, the system automatically deploys the microservice to a hosting site.\par
\textit{Study 1: Feasibility.} To evaluate the feasibility of this approach, we conducted a small study where 9 developers together worked to build a microservice. The results were promising. Participants submitted their first microtask less than half an hour after beginning, successfully submitted 350 microtasks, implemented 13 functions and 36 tests, completed microtasks in a median time under 5 minutes, correctly implemented 27 of 34 behaviors, and together implemented most of a functioning ToDo microservice~\cite{aghayi2019crowdsourced}.
\textit{Study 2: Comparing microtask programming to traditional development.} To directly compare traditional programming to microtasked programming, we conducted a controlled experiment. Twenty-eight developers worked either on traditional programming tasks, described through issues, or programming microtasks. We found that, compared to traditional software development, microtasking had important advantages in reducing onboarding time and time-to-market and, surprisingly, in increasing the quality of code and individual developer productivity.\par
\textit{Study 3: Using microtask programming in industry.} Our early studies were conducted entirely in artificial contexts, using artificial tasks and developers recruited specifically to work in the study. To examine the potential for using microtask programming in industry, we partnered with NTT, a large telecommunication company, to conduct a study of microtask programming within a real software project. We found that a microtask programming approach was successful in implementing and testing a project with 8000 lines of code in 14 functions. We also found that developers took time to understand the new concepts in microtask programming approach and be productive. We also found the value of having dedicated developers responsible for managing microtask projects. \par
\section{Future Work}
My studies have offered initial evidence that microtask programming can be effective in small crowds with a few developers. But much of the promise of microtasking comes from large crowds, and there exists a direct relationship between the number of independent tasks and the parallelism in microtask programming which may reduce time to market.
However, there are a number of significant challenges in scaling microtask programming to larger crowds. \par
To date, the largest crowd we have used is only 9 developers. To begin to examine microtask programming at scale, we are planning to conduct a virtual hackathon with around 100 developers.
Based on our findings, we will develop new techniques to scale microtask programming and more fully encompass software development work.
One direction we expect to pursue is with team organization. In our current approach, clients specify a microservice and crowd developers work in a flat organization to complete microtasks. This organization may not work well for all projects. For instance, some projects have security considerations, like private APIs that they do not want to expose to the public. Or projects may benefit from more experienced team leads who tackle more complex tasks or help less experienced developers when they get stuck. We will explore ways to make better use of more experienced developers in crowds through defining separate roles for crowd workers.
TopCoder is one successful example of this hybrid collaboration, and we will explore ways to adapt these ideas within microtask programming~\cite{Stol2014:TwoCompany}.\par
\section*{Acknowledgment}
This work was supported in part by the National Science Foundation under grants CCF-1414197 and CCF-1845508.
\bibliographystyle{IEEEtran}
| 898e1daa6a89f9210a06346066c76e8c0f36edb1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Visual content recognition and understanding have greatly made progress based on the advances of deep learning methods that construct the discriminative model by training large-scale labeled data. In fact, two reasons limit the current deep learning methods for efficiently learning new categories. One is that human annotation cost is high for large-scale data (for example, thousands of the diversity samples in the same category and hundreds of the various categories in one cognition domain), the other is that the rare samples of some categories are not enough for the discriminative model training. Therefore, it is still a challenge question that the discriminative model is learned from the rare samples of the categories. To solve this question, few-shot learning \cite{vinyals2016matching} \cite{snell2017prototypical} \cite{finn2017model} \cite{RaviL17} \cite{sung2018learning} \cite{qi2018low} \cite{SatorrasE18} \cite{kim2019edge} \cite{lee2019meta} \cite{sun2019meta} \cite{peng2019few} \cite{chen2019a}proposed from the inspiration of human visual system has been an attracted research to generalize the learning model to new classes with the rare samples of each novel category by feature learning \cite{ghiasi2018dropblock} \cite{wu2018unsupervised} \cite{donahue2019large} \cite{8955914} \cite{8948295} \cite{saikia2020optimized}or meta-learning \cite{chen2019a} \cite{ravichandran2019few} \cite{dhillon2019baseline}\cite{fei2020meta} \cite{zhang2020rethinking} \cite{luo2019learning}. Feature learning emphasises on feature generation and extraction model construction based on invariance transfer information, while meta-learning focuses on the relevance model between the samples for mining the common relationship of data samples by the episode training.
Meta-learning can transfer the available knowledge between the collection of the separated tasks, and propagate the latent structure information to enhance the model generalization and to avoid the model overfitting. Therefore, meta-learning is one of most promising directions for few-shot learning. However, meta-learning is constructed based on the large-scale separated tasks, and each task have the respective metric criterion that causes the gap of the transfer information between the samples of the separated tasks(the details in figure \ref{fig-2}). Although existing methods can relieve this gap to a certain extend by the same sample filling into the different tasks, it is still difficult to build the approximated metric criterion of the different tasks for efficiently information transfer and propagation. Therefore, we present HOSP-GNN that attempts to construct the approximated metric criterion by mining high-order structure and updates these metric values between samples by constraining data manifold structure for few-shot learning. Figure \ref{fig-1} illustrates the difference between HOSP-GNN and the most meta-learning for few-shot learning conceptually.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1\linewidth]{fig1.png}
\end{center}
\vspace{-0.2in}
\caption{The illustration of the difference between HOSP-GNN and the most meta-learning for few-shot learning.$S$ stands for support set; $Q$ is query set;the different color circles describe the labeled samples of the different classes in $S$; the gray color circles represent unlabeled samples in $Q$;the black solid lines between circles show the structure relationship of the labeled samples;the black dot lines between circles are the predicted structure relationship between labeled and unlabeled samples; the blue dot lines between circles across tasks indicate the latent high-order structure of samples.}
\label{fig-1}
\end{figure*}
Our contributions mainly have two points as follow.
\begin{itemize}
\item One is to find the high-order structure for bridging the gap between the metric criterion of the separated tasks. The importance of this point is to balance the consistence of the same samples in the different tasks, and to enhance the transferability of the similar structure in the learning model.
\item Another is to smooth the structure evolution for improving the propagation stability of the model transfer by manifold structure constraints. This point try to minimize the difference of the transformation projection between the similar samples, and to maximize the divergence of the transformation projection between the dissimilar samples for efficiently preserving the graph structure learning of data samples.
\end{itemize}
\section{Related Works}
In recent few-shot learning, there mainly are two kinds of methods according to the different learning focuses. One is feature learning based on the data representation of model extraction, and another is meta-learning based on the metric relationship of model description.
\subsection{Feature Learning}
Feature learning \cite{cubuk2019autoaugment} \cite{cubuk2019randaugment} \cite{lim2019fast} \cite{ghiasi2018dropblock} \cite{wu2018unsupervised} \cite{donahue2019large}for few-shot learning expects to inherit and generalize the well characteristics of the pre-train model based on large-scale samples training for recognizing new classes with few samples.
Because few samples often can not satisfy the necessary of the whole model training, the recent representative methods usually optimize the part parameters or structure of the pre-trained model by few samples for feature learning. For example,Bayesian optimization to Hyperband (BOHB) optimizes hyper-parameters by searching the smaller parameter space to maximize the validation performance for generic feature learning \cite{saikia2020optimized}; Geometric constraints fine-tune the parameters of one network layer with a few training samples for extracting the discriminative features for the new categories \cite{8948295}; Bidirectional projection learning (BPL) \cite{8955914} utilizes semantic embedding to synthesize the unseen classes features for obtaining enough samples features by competitive learning. These methods attempt to find the features invariance by partly fine-tuning the pre-trained model with the different constraints for recognizing the new classes with the few instances.
However, these methods can not explicitly formulate the metric rules for learning the discriminative model between new categories, moreover, these methods need retrain the model to adapt the distribution of new categories. It can lead to the degraded classification performance for few-shot learning and the more complicated optimization strategy in validation and test phases.
\subsection{Meta-Learning}
\label{ML}
Meta-learning \cite{vinyals2016matching} \cite{sung2018learning} \cite{chen2019a}\cite{gidaris2018dynamic} \cite{oreshkin2018tadam} \cite{ravichandran2019few} \cite{dhillon2019baseline}for few-shot learning tries to construct the relevances between samples in the base classes for generalizing the model to the new classes. These methods can learn the common structure relationship between samples by training on the collection of the separated tasks.In terms of the coupling between the model and the data, meta-learning can mainly be divided into two groups.
One group is the model optimization to quickly fit the distribution of new categories. Typical methods attempt to update the model parameters or optimizer for this purpose. For instance, meta-learner long short-term memory (LSTM) \cite{RaviL17} can update the model parameters to initialize the classifier network for the quick training convergence in the few samples of each classes; Model-agnostic meta-learning (MAML) \cite{finn2017model} can train the small gradient updating based on few learning data from a new task to obtain the well generalization performance; Latent embedding optimization(LEO)\cite{rusu2019metalearning}
can learn the latent generative representation of model parameters based on data dependence to decouple the gradient adaptation from the high-dimension parameters space.
Another group is metric learning to describe the structure relationship of the samples between support and query data for directly simulating the similarity metric of the new categories. The recent methods trend to enhance the metric structure by constraint information for the better model generalization in the new categories. For example,edge-labeling graph neural network (EGNN)\cite{kim2019edge} can update graph structure relationship to directly exploit the intra-cluster similarity and the inter-cluster dissimilarity by iterative computation; Meta-learning across meta-tasks (MLMT) \cite{fei2020meta} can explore their relationships between the random tasks by meta-domain adaptation or meta-knowledge distillation for boosting the performance of existing few-shot learning methods; Absolute-relative Learning (ArL)\cite{zhang2020rethinking} can both consider the class concepts and the similarity learning to complement their structure relationship for improving the recognition performance of the new categories; Continual meta-learning approach with Bayesian graph neural networks(CML-BGNN) \cite{luo2019learning} can implement the continual learning of a sequence of tasks to preserve the intra-task and inter-task correlations by message-passing and history transition.
In recent work, meta-learning based on metric learning shows the promising performance for recognizing the new categories with the few samples. These methods initially focus on the structure relation exploitation between support set and query set by modeling metric distances, and subsequent works further mine the relevance by mimicking the dependence between the separated tasks for enhancing the discrimination of the new categories. However, these methods depend on the projection loss between the seen and unseen classes \cite{fei2020meta} or Bayesian inference based on low-order structure (the metric of the pairwise data) \cite{luo2019learning} for considering the structure relationship between the intra or inter tasks. It is difficult to describe the latent high-order structure from the global observation. Therefore, the proposed HOSP-GNN expects to capture the high-order structure relationship based on samples metric for naturally correlating the relevance between the intra or inter tasks for improving the performance of few-shot learning.
\section{High-order structure preserving graph neural network}
Few-shot classification attempts to learn a classifier model for identifying the new classes with the rare samples. $C_{e}$ or $C_{n}$ respectively stands for a existing classes set with the large samples or a new classes set with the rare samples, and $C_{e}\bigcap C_{n}=\emptyset$,but they belong to the same cognise domain. The existing classes data set $D_{e}=\{(x_{i},y_{i})|y_{i} \in C_{e}, i=1,...,|D_{e} |\}$, where $x_{i}$ indicates the $i$-th image with the class label $y_{i}$, $|D_{e}|$ is the number of the elements in $D_{e}$. Similarly, the new classes data set $D_{n}=\{(x_{i},y_{i})|y_{i} \in C_{n}, i=1,...,|D_{n} |\}$, where $x_{i}$ indicates the $i$-th image with the class label $y_{i}$, $|D_{n}|$ is the number of the elements in $D_{n}$. If each new class includes $K$ labeled samples, the new classes data set is $K$-shot sample set. In other word, $|D_{n}|=K|C_{n}|$, where $|C_{n}|$ is the number of the elements in $C_{n}$. Few-shot learning is to learn the discriminative model from $D_{n}$ to predict the label of the image sample in the test set $D_{t}$ that comes from $C_{n}$ and $D_{n} \bigcap D_{t}=\emptyset$.
\subsection{Meta-learning for few-shot learning based on graph neural network}
In meta-learning, the classifier model can be constructed based on the collection of the separated tasks $\mathcal{T}=\{S,Q\}$ that contains a support set $S$ from the labeled samples in $D_{n}$ and a query set $Q$ from unlabeled samples in $D_{t}$. To build the learning model for few-shot learning, $S$ includes $K$ labeled samples and $N$ classes, so this situation is called $N$-way-$K$-shot few-shot classification that is to distinguish the unlabeled samples from $N$ classes in $Q$.
In practise, few-shot classification often faces the insufficient model learning based on the new classes data set $D_{n}$ with the rare labeled samples and $D_{t}$ with unlabeled samples. In this situation, the model difficultly identifies the new categories. Therefore, many methods usually draw support from the transfer information of $D_{e}$ with a large labels samples to enhance the model learning for recognizing the new classes. Episodic training \cite{sung2018learning} \cite{kim2019edge} is an efficient meta-learning for few-shot classification. This method can mimic $N$-way-$K$-shot few-shot classification in $D_{n}$ and $D_{t}$ by randomly sampling the differently separated tasks in $D_{e}$ as the various episodics of the model training. In each episode, $\mathcal{T}_{ep}=(S_{ep},Q_{ep})$ indicates the separated tasks with $N$-way-$K$-shot $T$ query samples, where the support set $S_{ep}=\{(x_{i},y_{i})|y_{i}\in C_{ep}, i=1,...,N\times K\}$, the query set $Q_{ep}=\{(x_{i},y_{i})|y_{i}\in C_{ep}, i=1,...,N\times T\}$, $S_{ep}\cap Q_{ep}=\emptyset$, and the class number $|C_{ep}|=N$. In the training phase, the class set $C_{ep}\in C_{e}$, while in test phase, the class set $C_{ep}\in C_{n}$. Many episodic tasks can be randomly sampled from $D_{e}$ to simulate $N$-way-$K$-shot learning for training the few-shot model, whereas the learned model can test the random tasks from $D_{n}$ for few-shot classification by $N$-way-$K$-shot fashion. If we construct a graph $G_{ep}=(\mathcal{V}_{ep},\mathcal{E}_{ep},\mathcal{T}_{ep})$ (here, $\mathcal{V}_{ep}$ is the vertex set of the image features in $\mathcal{T}_{ep}$, and $\mathcal{E}_{ep}$ is the edge set between the image features in $\mathcal{T}_{ep}$.) for describing the sample structure relationship in each episodic task, meta-learning for few-shot learning based on $L$ layers graph neural network can be reformulated by the cross-entropy loss $L_{ep}$ as following.
\begin{align}
\label{Eq1}
\begin{aligned}
L_{ep}&=-\sum_{l=1}^{L}\sum_{(x_{i},y_{i})\in Q_{ep}}y_{i}\log (h_{W}^{l}(f(x_{i},W_{f});S_{ep},G_{ep}))\\
&=-\sum_{l=1}^{L}\sum_{(x_{i},y_{i})\in Q_{ep}}y_{i}\log (\hat{y_{i}^{l}})
\end{aligned}
\end{align}
\begin{align}
\label{Eq1-1}
\begin{aligned}
\hat{y_{i}^{l}}=softmax(\sum_{j\neq i~~and~~c\in C_{ep}}e_{ij}^{l}\delta(y_{i}=c))
\end{aligned}
\end{align}
here, $\hat{y_{i}^{l}}$ is the estimation value of $y_{i}$ in $l$th layer;$e_{ij}^{l}$ is edge feature of the $l$th layer in graph $G_{ep}$; $\delta(y_{i}=c)$ is equal one when $y_{i}=c$ and zero otherwise;$f(\bullet)$ with the parameter set $W_{f}$ denotes the feature extracting function or network shown in Figure \ref{fig-4}(a); $h_{W}^{l}(f(x_{i});S_{ep},G_{ep})$ indicates few-shot learning model in the $l$th layer by training on $S_{ep}$ and $G_{ep}$, and $W$ is the parameter set of this model. This few-shot learning model can be exploited by the meta-training minimizing the loss function \ref{Eq1}, and then recognize the new categories with the rare samples.
\subsection{High-order structure description}
In few-shot learning based on graph neural network, the evolution and generation of the graph plays a very important role for identifying the different classes. In each episodic task of meta-learning, existing methods usually measure the structure relationship of the samples by pairwise way, and an independence metric space with the unique metric criteria is formed by the similarity matrix in graph. In many episodic tasks training, the various metric criteria lead to the divergence between the different samples structure relationship in Figure \ref{fig-2}. It is the main reason that the unsatisfactory classification of the new categories.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1\linewidth]{fig2.png}
\end{center}
\vspace{-0.2in}
\caption{The difference between the metric criteria of the episodic tasks. In each episodic training and testing, $S_{ep}$ stands for support set; $Q_{ep}$ is query set;the different color circles describe the labeled sample of the different classes in $S_{ep}$; the gray color circles represent unlabeled samples in $Q_{ep}$;the black solid lines between circles show the structure relationship of the labeled samples;the black dot lines between circles are the predicted structure relationship between labeled and unlabeled samples.}
\label{fig-2}
\end{figure*}
To reduce the difference between the metric criteria of the episodic tasks, we attempt to explore the high-order structure of the samples by building the latent connection. The traditional pairwise metric loses the uniform bench marking because of the normalization of the sample separation in independence tasks. However, the absolutely uniform bench marking is difficult to build the high-order structure relation between the samples of the different tasks. Therefore, \textbf{we define the relative metric graph of multi-samples in a task as high-order structure relation}, and the same samples by random falling into the independence task make this relative metric relationship widely propagate to the other samples for approximating to the uniform bench marking under the consideration with the interaction relationship between the episodic tasks.
More concretely, the relative metric graph $\hat{G}_{ep}=(\hat{\mathcal{V}}_{ep},\hat{\mathcal{E}}_{ep},\mathcal{T}_{ep})$, where $\mathcal{T}_{ep}=\{(x_{i},y_{i})|(x_{i},y_{i})\in S_{ep}~~or~~(x_{i},y_{i})\in Q_{ep}, y_{i}\in C_{ep},S_{ep}\bigcap Q_{ep}=\emptyset, i=1,...,N\times (K+T)\}$, the vertex set $\hat{\mathcal{V}}_{ep}=\{v_{i}|i=1,...,N\times (K+T)\}$, the edge set $\hat{\mathcal{E}}_{ep}=\{e_{ij}|i=1,...,N\times (K+T)~~and~~j=1,...,N\times (K+T)\}$. To describe the relative relationship between features, we can build $L$ layers graph neural network for learning edge feature $e_{ij}^{l}$ (graph structure relationship) and feature representation $v_{i}^{l}$ in each layer, where $l=0,...,L$. In the initial layer, each vertex feature $v_{i}^{0}$ can be computed by feature difference as following.
\begin{align}
\label{Eq2-1}
\begin{aligned}
&u_{i}^{0}=f(x_{i}),~~~~~~i=1,...,N\times (K+T),
\end{aligned}
\end{align}
\begin{align}
\label{Eq2}
\begin{aligned}
v_{i}^{0}=\left\{
\begin{aligned}
&u_{i}^{0}-u_{i+1}^{0},~~~~~~i=1,...,N\times (K+T)-1, \\
&u_{i}^{0}-u_{1}^{0},~~~~~~~~~~i=N\times (K+T),
\end{aligned}
\right.
\end{aligned}
\end{align}
here, $f(\bullet)$ is the feature extracting network shown in Figure \ref{fig-4}(a). The vertex can represented by two ways. One is that the initial vertex feature $u_{i}^{0}$ is described by the original feature. Another is that $v_{i}^{0}$ is a relative metric based on $u_{i}^{0}$ in $0$th layer. We expect to construct the higher order structure $e_{ij1}^{l}$ (the first dimension value of edge feature between vertex $i$ and $j$ in $l$ layer) based on this relative metric for representing edge feature under the condition with the pairwise similarity structure $e_{ij2}^{l}$ and dissimilarity structure $e_{ij3}^{l}$(these initial value of $0$ layer is defined by the labeled information of $S_{ep}$ in Equation \ref{Eq3}). Therefore, the initial edge feature can be represented by the different metric method as following.
\begin{align}
\label{Eq3}
\begin{aligned}
e_{ij}^{0}=\left\{
\begin{aligned}
& [e_{ij1}^{0}~||~ e_{ij2}^{0}=1~||~ e_{ij3}^{0}=0],~~~~y_{i}=y_{j}~~and~~(x_{i},y_{i})\in S_{ep}, \\
& [e_{ij1}^{0}~||~ e_{ij2}^{0}=0~||~ e_{ij3}^{0}=1],~~~~y_{i}\neq y_{j}~~and~~(x_{i},y_{i})\in S_{ep}, \\
& [e_{ij1}^{0}~||~ e_{ij2}^{0}=0.5~||~ e_{ij3}^{0}=0.5],~~~~ otherwise,
\end{aligned}
\right.
\end{aligned}
\end{align}
here, $||$ is concatenation symbol, $e_{ij1}^{0}$ can be calculated by the metric distance of the difference in Equation \ref{Eq4}, and $e_{ij1}^{l}$ can be updated by Equation \ref{Eq7}. It shows the further relevance between the relative metric, and indicates the high-order structure relation of the original features.
\begin{align}
\label{Eq4}
\begin{aligned}
e_{ij1}^{0}=1-\parallel v_{i}^{0}-v_{j}^{0} \parallel_{2}/\sum_{k}\parallel v_{i}^{0}-v_{k}^{0} \parallel_{2},~~~~(x_{i},y_{i})\in S_{ep}\bigcup Q_{ep},
\end{aligned}
\end{align}
Figure \ref{fig-3} shows the relationship between pairwise metric and high-order metric in $l$th layer, and the high-order metric involves any triple vertex features $u_{i}^{l}$,$u_{j}^{l}$ and $u_{k}^{l}$ in $\hat{G}_{ep}$ in each task. In these features, $u_{j}^{l}$ is a benchmark feature that is randomly sampled by the separated tasks. The common benchmark feature can reduce the metric difference between samples of the separated tasks.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1\linewidth]{fig3.png}
\end{center}
\vspace{-0.2in}
\caption{The relationship between pairwise metric $d_{pairwise\_metric}^{l}$(left figure) and high-order metric $d_{high-order\_metric}^{l}$(right figure) in $l$th layer. $f_{p}^{l}(\bullet)$ and $W_{p}^{l}$ respectively are pairwise metric network projection and parameter set in $l$th layer, while $f_{h}^{l}(\bullet)$ and $W_{h}^{l}$ respectively are high-order metric network projection and parameter set in $l$th layer.The black vector indicates the original vertex, the blue vector is the low-order metric vector ( the relative metric based on the original vertex), and the red vector stands for the high-order metric vector.}
\label{fig-3}
\end{figure*}
\subsection{High-order structure preserving}
HOSP-GNN can construct $L$ layers graph neural network for evolving the graph structure by updating the vertex and edge features. Moreover, we expect to preserve the high-order structure layer by layer for learning the discriminative structure between samples in the separated tasks. $l=1,...,L$ is defined as the layer number. In detail, $u_{i}^{l}$ can be updated by $u_{i}^{l-1}$,$v_{i}^{l-1}$ and $e_{ij}^{l-1}$ in Equation \ref{Eq5}, while $e_{ij}^{l}$ can be updated by$u_{i}^{l-1}$, $v_{i}^{l-1}$ and $e_{ij}^{l-1}$ in Equation \ref{Eq7},\ref{Eq8} and \ref{Eq9}.
\begin{align}
\label{Eq5}
\begin{aligned}
u_{i}^{l}=f_{v}^{l}([\sum_{j}\tilde{e}_{ij1}^{l-1}v_{j}^{l-1}~||~\sum_{j}\tilde{e}_{ij2}^{l-1}u_{j}^{l-1}~||~\sum_{j}\tilde{e}_{ij3}^{l-1}u_{j}^{l-1}], W_{v}^{l}),
\end{aligned}
\end{align}
\begin{align}
\label{Eq5-1}
\begin{aligned}
v_{i}^{l}=\left\{
\begin{aligned}
&u_{i}^{l}-u_{i+1}^{l},~~~~~~i=1,...,N\times (K+T)-1, \\
&u_{i}^{l}-u_{1}^{l},~~~~~~~~~~i=N\times (K+T),
\end{aligned}
\right.
\end{aligned}
\end{align}
here,$||$ is concatenation symbol, $\tilde{e}_{ijk}^{l-1}=e_{ijk}^{l-1}/\sum_{k}e_{ijk}^{l-1}$ ($k=1,2,3$), and $f_{v}^{l}(\bullet)$ is the vertex feature updating network shown in Figure \ref{fig-4}(b),and $W_{v}^{l}$ is the network parameters in $l$th layer. This updating process shows that the current vertex feature is the aggregative transformation of the previous layer vertex and edge feature in the different metrics, and can propagate the representation information under the consideration with edge feature (high-order structure information) layer by layer evolution. In \ref{Eq5}, high-order structure influences the vertex representation by transforming aggregation computation, but can not efficiently transfer layer by layer. Therefore, we expect to preserve high-order structure layer by layer by updating edge features. According to manifold learning \cite{he2004locality} and structure fusion\cite{Lin2014146}, structure information (the similarity relationship of samples) can be held from the original space to the projection space by minimizing the metric difference of these spaces. Similarly, high-order evolution based on graph neural network may obey the same rule for computing edge feature of each layer with the vertex feature updating. Therefore, we can construct the manifold loss by layer-by-layer computation for constraining the model optimization.
\begin{align}
\label{Eq6}
\begin{aligned}
L_{ml}=&\sum_{i,j,l}f_{h}^{l}(\|v_{i}^{l}-v_{j}^{l}\|_{2},W_{h}^{l})e_{ij1}^{l-1}+\\
&\sum_{i,j,l}f_{p}^{l}(\|u_{i}^{l}-u_{j}^{l}\|_{2},W_{p}^{l})e_{ij2}^{l-1}+\\
&\sum_{i,j,l}(1-f_{p}^{l}(\|u_{i}^{l}-u_{j}^{l}\|_{2},W_{h}^{l}))e_{ij3}^{l-1},
\end{aligned}
\end{align}
here,$L_{ml}$ is the loss of the manifold structure in the different layer and metric method (The first term is the manifold constrain for high-order structure, while the second and third terms are respectively the manifold constrain for similarity and dissimilarity); $f_{h}^{l}(\bullet)$ is the high-order metric network in Figure \ref{fig-4}(c) between vertex features, and $W_{h}^{l}$ is the parameter set of this network in $l$th layer; $f_{p}^{l}(\bullet)$ is the pairwise metric network in Figure \ref{fig-4}(c) between vertex features, and $W_{p}^{l}$ is it's parameter set in $l$th layer. \ref{Eq6} shows that the different manifold structures between layers can be preserved for minimizing $L_{ml}$. The edge updating based on high-order structure preserving is as following.
\begin{align}
\label{Eq7}
\begin{aligned}
\bar{e}_{ij1}^{l}=\frac{f_{h}^{l}(\|v_{i}^{l}-v_{j}^{l}\|_{2},W_{h}^{l})e_{ij1}^{l-1}}{\sum_{k}f_{h}^{l}(\|v_{i}^{l}-v_{k}^{l}\|_{2},W_{h}^{l})e_{ik1}^{l-1}/\sum_{k}e_{ik1}^{l-1}},
\end{aligned}
\end{align}
\begin{align}
\label{Eq8}
\begin{aligned}
\bar{e}_{ij2}^{l}=\frac{f_{p}^{l}(\|u_{i}^{l}-u_{j}^{l}\|_{2},W_{p}^{l})e_{ij2}^{l-1}}{\sum_{k}f_{p}^{l}(\|u_{i}^{l}-u_{k}^{l}\|_{2},W_{p}^{l})e_{ik2}^{l-1}/\sum_{k}e_{ik2}^{l-1}},
\end{aligned}
\end{align}
\begin{align}
\label{Eq9}
\begin{aligned}
\bar{e}_{ij3}^{l}=\frac{(1-f_{p}^{l}(\|u_{i}^{l}-u_{j}^{l}\|_{2},W_{p}^{l}))e_{ij3}^{l-1}}{\sum_{k}(1-f_{p}^{l}(\|u_{i}^{l}-u_{k}^{l}\|_{2},W_{p}^{l}))e_{ik3}^{l-1}/\sum_{k}e_{ik3}^{l-1}},
\end{aligned}
\end{align}
\begin{align}
\label{Eq10}
\begin{aligned}
e_{ij}^{l}=\bar{e}_{ij}^{l}/\|\bar{e}_{ij}^{l}\|_{1}.
\end{aligned}
\end{align}
Therefore, The total loss $L_{total}$ of the whole network includes $L_{ep}$ and $L_{ml}$.
\begin{align}
\label{Eq11}
\begin{aligned}
L_{total}=L_{ep}+\lambda L_{ml},
\end{aligned}
\end{align}
here, $\lambda$ is the tradeoff parameter for balancing the influence of the different loss. Figure \ref{fig-4} shows the network architecture of the proposed HOSP-GNN.
\begin{figure*}[hbp]
\begin{center}
\includegraphics[width=1\linewidth]{fig4.png}
\end{center}
\vspace{-0.2in}
\caption{The network architecture of the proposed HOSP-GNN.(a) is the total network structure, (b) and (c) respectively are vertex and edge updating network in (a). MLP is a multilayer perceptron; DU indicates the difference unit for the relative metric; Conv stands for a convolutional block that includes $96$ channels of $1\times 1$ convolution kernel, batch normalization unit, and LeakReLU unit;$S_{ep}$ stands for support set; $Q_{ep}$ is query set;the different color circles describe the labeled samples of the different classes in $S_{ep}$; the gray color circles represent unlabeled samples in $Q_{ep}$;the black solid lines between circles show the structure relationship of the labeled samples;the black dot lines between circles are the predicted structure relationship between labeled and unlabeled samples;$L_{ep}$ is the loss metric between the real labels and the predicted labels; $L_{ml}$ is the loss metric between high structures layer by layer;$v_{i}^{l}$ is the $i$th vertex feature in the $l$th layer of graph;$e_{ij}^{l}$ is the edge feature between the vertex $i$ and $j$ in the $l$th layer of graph;$f(\bullet)$ denotes the feature extracting network;$f_{v}^{l}(\bullet)$ indicates the vertex feature updating network in the $l$th layer; $f_{h}^{l}(\bullet)$ denotes the high-order metric network between vertex features in the $l$th layer; $f_{p}^{l}(\bullet)$ stands for the pairwise metric network between vertex features in the $l$th layer.}
\label{fig-4}
\end{figure*}
To indicate the inference details of HOSP-GNN, algorithm \ref{algHOSP-GNN} shows the pseudo code of the proposed HOSP-GNN for predicting the labels of the rare samples. This algorithm process contains four steps. The first step (line 1 and line 2) initializes the vertex feature and the edge feature. The second step (line 4 and line 5) updates the vertex features layer by layer. The third step (from line 6 to line 8) updates the edge features layer by layer. The forth step (line 9) predicts the labels of the query samples.
\begin{algorithm}[ht]
\caption{The inference of the HOSP-GNN for few-shot learning}
\begin{algorithmic}[1]
\label{algHOSP-GNN}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\renewcommand{\algorithmicreturn}{\textbf{Iteration:}}
\REQUIRE \textit{Graph}, $\hat{G}_{ep}=(\hat{\mathcal{V}}_{ep},\hat{\mathcal{E}}_{ep},\mathcal{T}_{ep})$, where $\mathcal{T}_{ep}=\{(x_{i},y_{i})|(x_{i},y_{i})\in S_{ep}~~or~~x_{i}\in Q_{ep}, y_{i}\in C_{ep},S_{ep}\bigcap Q_{ep}=\emptyset, i=1,...,N\times (K+T)\}$,$\hat{\mathcal{V}}_{ep}=\{v_{i}|i=1,...,N\times (K+T)\}$, $\hat{\mathcal{E}}_{ep}=\{e_{ij}|i=1,...,N\times (K+T)~~and~~j=1,...,N\times (K+T)\}$;\\ \textit{Model parameter}, $W=\{W_{f},W_{v}^{l},W_{h}^{l}|l=1,...,L\}$
\ENSURE The query samples of the predicted labels $\{\hat{y}_{i}^{l}|i=1,...,N\times T~~~~and~~~~l=1,...,L\}$
\STATE Computing the initial vertex feature $v_{i}^{0}$ by feature difference in Equation \ref{Eq2}
\STATE Computing the initial edge feature $e_{ij}^{0}$ as high-order structure in Equation \ref{Eq3}
\FOR {$1\leq l\leq L$}
\FOR {$1\leq i\leq N\times (K+T)$}
\STATE Updating vertex feature $v_{i}^{l}$ by Equation \ref{Eq5}
\FOR {$1\leq j\leq N\times (K+T)$}
\STATE Updating edge feature $e_{ij}^{l}$ by Equation \ref{Eq7},\ref{Eq8},\ref{Eq9} and \ref{Eq10}
\ENDFOR
\STATE Predicting the query sample labels $\hat{y}_{i}^{l}$ by Equation\ref{Eq1-1}
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{Experiment}
To evaluating the proposed HOSP-GNN, we carry out four experiments. The first experiment involves the baseline methods comparison. The second experiment conducts the state-of-the-art methods comparison. The third experiment implements semi-supervised fashion for few-shot learning. The forth experiment assesses the layer effect for graph model, and the loss influence for the manifold constraint.
\subsection{Datasets}
In experiments, we use three benchmark datasets that are miniImageNet\cite{vinyals2016matching}, tieredImageNet \cite{Mengye2018}, and FC100\cite{NIPS20187352}. In miniImageNet dataset from ILSVRC-12 \cite{ILSVRC15}, RGB images include $100$ different classes, and each class has $600$ samples. We adopt the splits configuration \cite{kim2019edge} that respectively is 64,16,and 20 classes for training, validation and testing. In tieredImageNet dataset from ILSVRC-12 \cite{ILSVRC15}, there are more than $700k$ images from $608$ classes. Moreover, $608$ classes is collected for $34$ higher-level semantic classes, each of which has $10$ to $20$ classes. We also use the splits configuration \cite{kim2019edge} that respectively is $351$,$97$, and $160$ for training, validation and testing. Each class has about $1281$ images. In FC100 dataset from CIFAR-100\cite{Krizhevsky}, there are $100$ classes images grouped into $20$ higher-level classes. Classes respectively are divided into $60$,$20$, and $20$ for training, validation and testing. Each classes have $600$ images of size $32\times 32$. Table \ref{tab1} shows the statistics information of these datasets.
\begin{table*}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Datasets statistics information in experiments. $\sharp$ denotes the number. }
\label{tab1}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{lp{1.0cm}p{1.5cm}p{1.5cm}p{1.5cm}p{1.5cm}p{0.5cm}}
\hline
\bfseries Datasets & \bfseries \tabincell{l}{$\sharp$ Classes } & \bfseries \tabincell{l}{$\sharp$ training\\ classes} & \bfseries \tabincell{l}{$\sharp$ validation \\classes} & \bfseries \tabincell{l}{$\sharp$ testing \\classes} &\bfseries \tabincell{l}{$\sharp$ images}\\
\hline \hline
miniImageNet & $100$ &$64$& $16$ & $20$ & $60000$\\
\hline
tieredImageNet & $608$ &$351$& $97$ & $160$ & $778848$\\
\hline
FC100 & $100$ &$60$& $20$ & $20$ & $60000$\\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Experimental Configuration}
Figure \ref{fig-4} describes the network architecture of the proposed HOSP-GNN in details. The feature extracting network is the same architecture in the recent works\cite{vinyals2016matching} \cite{snell2017prototypical} \cite{finn2017model} \cite{kim2019edge}, and specifically includes four convolutional blocks with $3\times 3$ kernel, one linear unit, one bach normalization and one leakReLU unit for few-shot models. Other parts of network is detailed in figure \ref{fig-4}. To conveniently compare with other methods(baseline methods and state-of-the-art methods), we set the layer number $L$ to $3$ in the proposed HOSP-GNN.
To train the proposed HOSP-GNN model, we use Adam optimizer with the learning rate $5\times 10^{-4}$ and weight decay $10^{-6}$. The mini-batch size of meta-learning task is set to $40$ or $20$ for 5-way-1-shot or 5-way-5-shot experiments. The loss coefficient $\lambda$ is set to $10^{-5}$. Experimental results in this paper can be obtained by 100K iterations training for miniImageNet and FC100, 200K iterations training for tieredImageNet.
We implement 5-way-1-shot or 5-way-5-shot experiments for evaluating the proposed method. Specifically, we averagely sample $15$ queries from each classes, and randomly generate $600$ episodes from the test set for calculating the averaged performance of the queries classes.
\subsection{Comparison with baseline approaches}
The main framework of the proposed HOSP-GNN is constructed based on edge-labeling graph neural network (EGNN)\cite{kim2019edge}. Their differences are the graph construction and the manifold constraint for model training in episodic tasks. EGNN method mainly considers the similarity and dissimilarity relationship between the pair-wise samples, but does not involve the manifold structure constraint of each layer for learning few-shot model.In contrast, HOSP-GNN tries to capture the high-order structure relationship between multi-samples ,fuses the similarity and dissimilarity relationship between the pair-wise samples, and constrains the model training by layer by layer manifold structure loss. Therefore, the base-line methods include EGNN, HOSP-GNN-H-S(the proposed HOSP-GNN only considers the high-order structure relationship and the similarity relationship),HOSP-GNN-H-D(the proposed HOSP-GNN only considers the high-order structure relationship and the dissimilarity relationship),HOSP-GNN-H (the proposed HOSP-GNN only considers the high-order structure relationship),HOSP-GNN-S (the proposed HOSP-GNN only considers the similarity relationship),and HOSP-GNN-D (the proposed HOSP-GNN only considers the dissimilarity relationship), in which H denotes the high-order structure relationship, S strands for the similarity relationship, and D represents the dissimilarity relationship.
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of the methods related the high-order structure (HOSP-GNN,HOSP-GNN-H-S,HOSP-GNN-H-D,and HOSP-GNN-H)with baseline methods (EGNN,HOSP-GNN-S,and HOSP-GNN-D) for 5-way-1-shot learning. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab2}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries 5-way-1-shot &\bfseries &\bfseries \\
\cline{2-4}
\bfseries &\bfseries miniImageNet &\bfseries tieredImageNet &\bfseries FC100 \\
\hline \hline
EGNN \cite{kim2019edge} & $52.46\pm0.45$ &$57.94\pm0.42$ & $35.00\pm0.39$ \\
\hline
HOSP-GNN-D & $52.44\pm0.43$ &$57.91\pm0.39$ & $35.55\pm0.40$ \\
\hline
HOSP-GNN-S & $52.86\pm0.41$ &$57.84\pm0.44$ & $35.48\pm0.42$ \\
\hline\hline
HOSP-GNN-H & $69.52\pm0.41$ &$91.71\pm 0.28$ & $76.24\pm0.41$ \\
\hline
HOSP-GNN-H-D & $78.82\pm0.45$ &$82.63\pm0.26$ & $82.27\pm0.44$ \\
\hline
HOSP-GNN-H-S & $88.15\pm0.35$ &$\textbf{95.39}\pm\textbf{0.20}$ & $\textbf{83.65}\pm\textbf{0.38}$ \\
\hline
HOSP-GNN & $\textbf{93.93}\pm\textbf{0.37}$ &$94.00\pm0.24$ & $76.79\pm0.46$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of Comparison of the methods related the high-order structure (HOSP-GNN,HOSP-GNN-H-S,HOSP-GNN-H-D,and HOSP-GNN-H) with baseline methods (EGNN,HOSP-GNN-S,and HOSP-GNN-D) for 5-way-5-shot learning. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab3}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries 5-way-5-shot &\bfseries &\bfseries \\
\cline{2-4}
\bfseries &\bfseries miniImageNet &\bfseries tieredImageNet &\bfseries FC100 \\
\hline \hline
EGNN \cite{kim2019edge} & $67.33\pm0.40$ &$68.93\pm0.40$ & $47.77\pm0.42$ \\
\hline
HOSP-GNN-D & $65.75\pm0.43$ &$68.30\pm 0.40$ & $47.00\pm0.41$ \\
\hline
HOSP-GNN-S & $66.10\pm0.42$ &$68.64\pm0.41$ & $47.69\pm0.41$ \\
\hline\hline
HOSP-GNN-H & $69.19\pm0.44$ &$90.06\pm 0.30$ & $70.82\pm0.46$ \\
\hline
HOSP-GNN-H-D & $68.39\pm0.42$ &$91.11\pm0.29$ & $48.48\pm0.43$ \\
\hline
HOSP-GNN-H-S & $68.85\pm0.42$ &$91.16\pm0.29$ & $48.25\pm0.43$ \\
\hline
HOSP-GNN & $\textbf{95.98}\pm\textbf{0.21}$ &$\textbf{98.44}\pm\textbf{0.12}$ & $\textbf{70.94}\pm\textbf{0.51}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
In Table \ref{tab2} and \ref{tab3}, the methods related the high-order structure relationship show the better performance in the base-line methods. However, the performance of HOSP-GNN based on the high-order structure combination is different because of the adaptability and coupling between the high-order structure and the pair-wise structure (similarity or dissimilarity). Figure \ref{fig-5} demonstrates the validation accuracy with iteration increasing for 5-way-1-shot or 5-way-5-shot in the different datasets. These processes also indicate the effectiveness of the high-order structure for training few-shot model. The details is analyzed in section \ref{analysis}.
\begin{figure}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-1.png}
}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-2.png}
}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-3.png}
}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-4.png}
}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-5.png}
}
\subfigure[]{
\centering
\includegraphics[width=2.4in]{fig5-6.png}
}
\caption{Validation accuracy with iteration increasing for 5-way-1-shot or 5-way-5-shot in the different datasets.(a),(c) and (e) for 5-way-1-shot in miniImageNet,tieredImageNet and FC100; (b),(d) and (f) for 5-way-5-shot in miniImageNet,tieredImageNet and FC100.}
\label{fig-5}
\end{figure}
\subsection{Comparison with state-of-the-arts}
In this section, we compare the proposed HOSP-GNN with the state-of-the-art methods, which include EGNN\cite{kim2019edge},MLMT \cite{fei2020meta},ArL\cite{zhang2020rethinking},
and CML-BGNN\cite{luo2019learning}, which are detailed in section \ref{ML}. These methods can capture the structure relationship of the samples in the episodic tasks based on meta-learning for few-shot learning. The difference of these method are based on the various processing ways to mine the structure relationship for few-shot models. Therefore,these methods denote the different classification performance in the benchmark datasets. Table \ref{tab4} and \ref{tab5} express that the performance of the proposed HOSP-GNN is greatly better than that of other methods. It shows that the dependence of the episodic tasks can be better described by high-order structure based on HOSP-GNN. The detailed analysis is demonstrated in section \ref{analysis}.
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of HOSP-GNN method with state-of-art methods (EGNN,MLMT,ArL,and CML-BGNN) for 5-way-1-shot learning. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab4}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries 5-way-1-shot &\bfseries &\bfseries \\
\cline{2-4}
\bfseries &\bfseries miniImageNet &\bfseries tieredImageNet &\bfseries FC100 \\
\hline \hline
EGNN \cite{kim2019edge} & $52.46\pm0.45$ &$57.94\pm0.42$ & $35.00\pm0.39$ \\
\hline
MLMT \cite{fei2020meta} & $72.41\pm0.49$ &$72.82\pm0.52$ & $null$ \\
\hline
ArL \cite{zhang2020rethinking} & $59.12\pm0.67$ &$null$ & $null$ \\
\hline
CML-BGNN \cite{luo2019learning} & $88.62\pm0.43$ &$88.87\pm 0.51$ & $67.67\pm1.02$ \\
\hline\hline
HOSP-GNN & $\textbf{93.93}\pm\textbf{0.37}$ &$\textbf{94.00}\pm\textbf{0.24}$ & $\textbf{76.79}\pm\textbf{0.46}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of HOSP-GNN method with state-of-art methods (EGNN,MLMT,ArL,and CML-BGNN) for 5-way-5-shot learning. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab5}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries 5-way-5-shot &\bfseries &\bfseries \\
\cline{2-4}
\bfseries &\bfseries miniImageNet &\bfseries tieredImageNet &\bfseries FC100 \\
\hline \hline
EGNN \cite{kim2019edge} & $67.33\pm0.40$ &$68.93\pm0.40$ & $47.77\pm0.42$ \\
\hline
MLMT \cite{fei2020meta} & $84.96\pm0.34$ &$85.97\pm0.35$ & $null$ \\
\hline
ArL \cite{zhang2020rethinking} & $73.56\pm0.45$ &$null$ & $null$ \\
\hline
CML-BGNN \cite{luo2019learning} & $92.69\pm 0.31$ &$92.77\pm 0.28$ & $63.93\pm 0.67$ \\
\hline\hline
HOSP-GNN & $\textbf{95.98}\pm\textbf{0.21}$ &$\textbf{98.44}\pm\textbf{0.12}$ & $\textbf{70.94}\pm\textbf{0.51}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Semi-supervised few-shot learning}
In support set, we label the part of samples on all classes for the robust test of the learning model, and this situation is called semi-supervised few-shot learning. Therefore, we set $20\%$, $40\%$, and $100\%$ labeled samples of the support set for 5-way-5-shot learning in miniImageNet dataset. In this section, we compare the proposed HOSP-GNN with three graph related methods, which are GNN\cite{SatorrasE18},EGNN\cite{kim2019edge}
and CML-BGNN\cite{luo2019learning}. The common of these methods is based on graph for describing the structure of the samples, while the difference of these methods is the various ways for mining the structure of the samples. For example, GNN focuses on generic message-passing mechanism for optimizing the samples structure;EGNN emphasizes on updating mechanism for evolving the edge feature;CML-BGNN cares about the continual information of the episode tasks for structure complement; The proposed HOSP-GNN expects to mine the high-order structure for connecting the separated tasks and preserves the layer-by-layer manifold structure of the samples for constraining the model learning.The detailed analysis is indicated in section \ref{analysis}.
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Semi-supervised few-shot learning for the graph related methods(GNN,EGNN,CML-BGNN and the proposed HOSP-GNN) in miniImageNet dataset. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab6}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries miniImageNet &\bfseries 5-way-5-shot &\bfseries \\
\cline{2-4}
\bfseries &\bfseries $20\%$-labeled &\bfseries $40\%$-labeled &\bfseries $100\%$-labeled \\
\hline \hline
GNN \cite{SatorrasE18} & $52.45\pm0.88$ &$58.76\pm0.86$ & $66.41\pm0.63$ \\
\hline
EGNN \cite{kim2019edge} & $63.62\pm0.00$ &$64.32\pm0.00$ & $75.25\pm0.49$ \\
\hline
CML-BGNN \cite{luo2019learning} & $\textbf{88.95}\pm\textbf{0.32}$ &$\textbf{89.70}\pm\textbf{0.32}$ &$92.69\pm0.31$ \\
\hline\hline
HOSP-GNN & $65.93\pm0.38$ &$67.06\pm0.40$ & $\textbf{95.98}\pm\textbf{0.21}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Ablation experiments for the layer number and the loss}
The proposed HOSP-GNN have two key points about structure evolution. One is the influence of the layer for model learning in graph. Another is the layer-by-layer manifold structure constraint for generating the better model with the preserved structure. Therefore, we respectively evaluate these points by ablating the part of the components from the whole model. The first experiment is about layers ablation, in which we train one layer model, two layer model and tree layer model for few-shot learning in Table \ref{tab7}. The second experiment is about the different loss, in which we set the various losses propagation for optimizing the model in Table \ref{tab8}.Table \ref{tab9} shows the parameter $\lambda$ influence to the proposed HOSP-GNN. The detailed analysis of these experimental results is shown in section \ref{analysis}.
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of the different layer model for the graph related methods(GNN,EGNN,CML-BGNN and the proposed HOSP-GNN) in miniImageNet dataset. Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab7}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.5cm}p{2.5cm}p{2.5cm}p{2.5cm}p{2.5cm}}
\hline
\bfseries Method &\bfseries miniImageNet &\bfseries 5-way-1-shot &\bfseries \\
\cline{2-4}
\bfseries &\bfseries one layer model &\bfseries two layer model &\bfseries three layer model \\
\hline \hline
GNN \cite{SatorrasE18} & $48.25\pm0.65$ &$49.17\pm0.35$ & $50.32\pm0.41$ \\
\hline
EGNN \cite{kim2019edge} & $55.13\pm0.44$ &$57.47\pm0.53$ & $58.65\pm0.55$ \\
\hline
CML-BGNN \cite{luo2019learning} & $\textbf{85.75}\pm\textbf{0.47}$ &$87.67\pm0.47$ &$88.62\pm0.43$ \\
\hline
HOSP-GNN & $75.13\pm0.44$ &$\textbf{87.77}\pm\textbf{0.37}$ & $\textbf{93.93}\pm\textbf{0.37}$ \\
\hline\hline
\bfseries Method &\bfseries miniImageNet &\bfseries 5-way-5-shot &\bfseries \\
\cline{2-4}
\bfseries &\bfseries one layer model &\bfseries two layer model &\bfseries three layer model \\
\hline \hline
GNN \cite{SatorrasE18} & $65.58\pm0.34$ &$67.21\pm0.49$ & $66.99\pm0.43$ \\
\hline
EGNN \cite{kim2019edge} & $67.76\pm0.42$ &$74.70\pm0.46$ & $75.25\pm0.49$ \\
\hline
CML-BGNN \cite{luo2019learning} & $\textbf{90.85}\pm\textbf{0.27}$ &$\textbf{91.63}\pm\textbf{0.26}$ &$92.69\pm0.31$ \\
\hline
HOSP-GNN & $67.86\pm0.41$ &$72.48\pm0.37$ & $\textbf{95.98}\pm\textbf{0.21}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{Comparison of the different loss model for the proposed HOSP-GNN (HOSP-GNN-loss1 for label loss in support set , and HOSP-GNN for the consideration of the label and manifold structure loss). Average accuracy (\%)of the query classes is reported in random episodic tasks.}
\label{tab8}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{2.0cm}p{2.0cm}p{2.0cm}p{2.0cm}}
\hline
\bfseries Method &\bfseries 5-way-5-shot &\bfseries &\bfseries \\
\cline{2-4}
\bfseries &\bfseries miniImageNet &\bfseries tieredImageNet &\bfseries FC100 \\
\hline \hline
HOSP-GNN-loss1 & $92.29\pm0.28$ &$98.41\pm0.12$ & $65.47\pm0.51$ \\
\hline\hline
HOSP-GNN & $\textbf{95.98}\pm\textbf{0.21}$ &$\textbf{98.44}\pm\textbf{0.12}$ & $\textbf{70.94}\pm\textbf{0.51}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\small
\renewcommand{\arraystretch}{1.0}
\caption{The tradeoff parameter $\lambda$ influence to few-show learning in miniImageNet.}
\label{tab9}
\begin{center}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{tabular}{l|p{1.0cm}p{1.0cm}p{1.0cm}p{1.0cm}p{1.0cm}p{1.0cm}p{1.0cm}}
\hline
\bfseries Method &\bfseries miniImageNet &\bfseries &\bfseries 5-way-5-shot &\bfseries $\lambda$ &\bfseries \\
\cline{2-7}
\bfseries &\bfseries $10^{-2}$ &\bfseries $10^{-3}$ &\bfseries $10^{-4}$ &\bfseries $10^{-5}$ &\bfseries $10^{-6}$ &\bfseries $10^{-7}$ \\
\hline \hline
HOSP-GNN & $94.65\pm0.22$ &$93.71\pm0.25$ & $93.43\pm0.25$ & $95.98\pm0.21$ &$95.31\pm0.21$ & $91.89\pm0.30$\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Experimental results analysis}
\label{analysis}
In above experiments, there are ten methods used for comparing with the proposed HOSP-GNN. In the baseline methods (HOSP-GNN,EGNN \cite{kim2019edge}, HOSP-GNN-H-S, HOSP-GNN-H-D, HOSP-GNN-H, HOSP-GNN-S and HOSP-GNN-D),we can capture the various structure information of the samples for constructing the similar learning model. In the state-of-the-art methods (EGNN\cite{kim2019edge}, MLMT \cite{fei2020meta}, ArL\cite{zhang2020rethinking},
,CML-BGNN\cite{luo2019learning} and HOSP-GNN), we demonstrate the model learning results based on the different networks framework for mining the relevance between the separated tasks. In the semi-supervised methods (GNN\cite{SatorrasE18},EGNN\cite{kim2019edge}
, CML-BGNN\cite{luo2019learning} and HOSP-GNN), we can find the labeled samples number to the performance influence for the robust testing of these methods. In ablation experiments, we build the different layers model(one layer model, two layer model and three layer model) and the various loss model (HOSP-GNN-loss1 and HOSP-GNN) for indicating their effects. The proposed HOSP-GNN can jointly consider the high-order structure and the layer-by-layer manifold structure constraints to effectively recognize the new categories. From these experiments, we have the following observations and analysis.
\begin{itemize}
\item The proposed HOSP-GNN and its Variants (HOSP-GNN-H-S, HOSP-GNN-H-D and HOSP-GNN-H) greatly outperform the base-line methods(HOSP-GNN-S, HOSP-GNN-D and EGNN) in table \ref{tab2},table \ref{tab3}, and figure \ref{fig-5}. The common characteristic of these methods (the proposed HOSP-GNN and its Variants) involves the high-order structure for learning model. Therefore,it shows that the high-order structure can better associate with the samples from the different tasks for improving the performance of few-shot learning.
\item The performance of the proposed HOSP-GNN and its Variants(HOSP-GNN-H-S, HOSP-GNN-H-D and HOSP-GNN-H) indicate the different results in the various dataset and experimental configuration in table \ref{tab2}, table \ref{tab3}, and figure \ref{fig-5}. In 5-way-1-shot learning, HOSP-GNN has the better performance than other methods in miniImageNet, while HOSP-GNN-H-S indicates the better results than others in tieredImageNet and FC100. In 5-way-5-shot learning, HOSP-GNN also shows the better performance than others in miniImageNet,tierdImageNet,and FC100. It shows that similarity, dissimilarity and high-order structure have the different influence to the model performance in the various datasets. For example, similarity,dissimilarity and high-order structure have the positive effect for recognizing the new categories in miniImageNet and tierdImageNet, while dissimilarity produces the negative effect for learning model in FC100. In any situation, high-order structure has an important and positive role for improving the model performance.
\item The proposed HOSP-GNN obviously is superior to other state-of-the-art methods in table \ref{tab4} and \ref{tab5}. These methods focus on the different aspects, which are the graph information mining based on EGNN\cite{kim2019edge}, the across task information exploitation based on MLMT \cite{fei2020meta}, the semantic-class relationship utilization based on ArL\cite{zhang2020rethinking}, the history information association based on CML-BGNN\cite{luo2019learning}, and the high-order structure exploration based on HOSP-GNN. The proposed HOSP-GNN can not only exploit the across task structure by the extension association of the high-order structure , but also use the latent manifold structure to constrain the model learning, so the proposed HOSP-GNN obtains the best performance in these methods.
\item The proposed HOSP-GNN demonstrates the better performance than the graph related methods(GNN\cite{SatorrasE18}, EGNN\cite{kim2019edge}, and CML-BGNN\cite{luo2019learning}) based on the more labeled samples in table \ref{tab6}. The enhanced structure of the more labeled samples can efficiently propagate the discriminative information to the new categories by the high-order information evolution based on the graph. The labeled sample number has few influence on model learning based on CML-BGNN\cite{luo2019learning}. In contrast, labeled sample number has an important impact on model learning based on the graph related methods(GNN\cite{SatorrasE18}, EGNN\cite{kim2019edge}, and HOSP-GNN).
\item In the different layer model experiments, the proposed HOSP-GNN indicates the various performance with layer number changing in table \ref{tab7}. In 5-way-1-shot learning, HOSP-GNN has the better performance than other methods, while in 5-way-5-shot learning, CML-BGNN\cite{luo2019learning} shows the more challenging results than other methods. It demonstrates that layer number has an important impact on the high-order structure evolution. We can obtain the significant improvement based on the more layer model of HOSP-GNN for 5-way-5-shot in miniImageNet, while the performances of other methods almost are not changing with the layer number increasing. Therefore, the proposed HOSP-GNN trends to the more layers to exploit the high-order structure for few-shot learning.
\item In table \ref{tab8}, the different losses (supervised label loss and manifold structure loss) are considered for constructing few-shot model. HOSP-GNN (the method model based on supervised label loss and manifold structure loss) can show the better performance than HOSP-GNN-loss1(the approach involves the model with the supervised label loss). It expresses that manifold constraint with the layer-by-layer evolution can enhance the performance of model because of the intrinsic distribution consistence on the samples of the different task.
\end{itemize}
\section{Conclusion}
To associate and mine the samples relationship in the different tasks, we have presented high-order structure preserving graph neural network(HOSP-GNN) for few-shot learning. HOSP-GNN can not only describe high-order structure relationship by the relative metric in multi-samples, but also reformulate the updating rules of graph structure by the alternate computation between vertexes and edges based on high-order structure. Moreover, HOSP-GNN can enhance the model learning performance by the layer-by-layer manifold structure constraint for few-shot classification. Finally, HOSP-GNN can jointly consider similarity, dissimilarity and high-order structure to exploit the metric consistence between the separated tasks for recognizing the new categories. For evaluating the proposed HOSP-GNN, we carry out the comparison experiments about the baseline methods,the state of the art methods, the semi-supervised fashion, and the layer or loss ablation on miniImageNet, tieredImageNet and FC100. In experiments, HOSP-GNN demonstrates the prominent results for few-shot learning.
\section{Acknowledgements}
The authors would like to thank the anonymous reviewers for their insightful comments that help improve the quality of this paper. Especially, this work was supported by NSFC (Program No.61771386,Program No.61671376 and Program No.61671374), Research and Development Program of Shaanxi (Program No.2020SF-359).
| 92db6b3d3fcb6a13777b8a83750341d7fa5b6189 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $G=(V,E)$ be a simple undirected graph, with vertices $V$ and edges $E\subset~V\times V$, and let $\mathbf{p} = \{ p_1, p_2, \dots p_n \} \in {\mathbb R}^{d n}$ be a set of vectors $p_u\in{\mathbb R}^d$ assigned to the vertices, where $n=|V|$ is the vertex cardinality.
This assignment is a (real) \textit{embedding} of $G$ in ${\mathbb R}^d$.
Every embedding induces a set of non-negative real edge lengths $\bm{\lambda}= (\lVert p_u-p_v \rVert)_{(u,v) \in E}$,
where $\lVert \cdot \rVert$ denotes the Euclidean distance.
The distances $\bm{\lambda}$ represent a labelling of the graph edges; hence such graphs are also known as distance graphs.
The embedding $\mathbf{p}$ of $G$ is called \textit{rigid} if it admits only a finite number of other embeddings up to rigid motions, for the edge labelling induced by $\mathbf{p}$, otherwise it is called \textit{flexible}.
This basic dichotomy can be associated to the graph for almost all embeddings with no reference to a specific embedding.
This is achieved using the following genericity notion.
A graph embedding is called \textit{generic} if a small perturbation of the embedded vertices does not alter whether the embedding is rigid or flexible \cite{GSS93}.
Thus, $G$ is \textit{generically rigid} in ${\mathbb R}^{d}$ if it is rigid for every edge labelling induced by a generic embedding.
Embeddings of simple graphs are also defined in $\mathbb{C}^d$ and correspond to all possible configurations of $n$ points $\bm{x}\in\mathbb{C}^{d n}$ that satisfy the assignments
$$
\lambda_{u,v}^2= \sum\limits_{k=1}^d (x_{u,k}-x_{v,k})^2,
$$
for all edges $(u,v)\in E$.
Note that this does not represent a norm in $\mathbb{C}^d$.
Additionally, a rigid graph is (generically) {\em minimally} rigid if any edge removal breaks the rigidity.
A fundamental theorem in graph rigidity due to Maxwell gives a necessary (but not sufficient) condition on the edge count of a graph and all its subgraphs for the graph to be rigid \cite{Maxwell}.
More precisely, if a graph $G=(V,E)$ is minimally rigid in ${\mathbb R}^d$, then $|E|=d \cdot |V| -\binom{d+1}{2}$ and $|E'| \leq d\cdot |V'|-\binom{d+1}{2}$ for every subgraph $G'(V',E') \subset G$.
Minimally rigid graphs in ${\mathbb R}^2$ are known as \textit{Laman graphs}, because of G.~Laman's theorem that gives a full characterisation of them \cite{Laman}.
Laman's theorem states that Maxwell's condition is also sufficient for minimally rigid graphs in ${\mathbb R}^2$.
This condition is not the only one that verifies minimal rigidity in ${\mathbb R}^2$.
For example, Henneberg constructions, pebble games and Recski's theorem have been used alternatively \cite{handbook1,pebble}.
In fact, Laman rediscovered the forgotten results of H.\ Pollaczek-Geiringer \cite{Geiringer1927,Geiringer1932} and in order to honour her we call minimally rigid graphs in ${\mathbb R}^3$ \textit{Geiringer graphs} following \cite{belt,GKT17}.
In dimension~3, a well-known theorem by Cauchy states that all strictly convex simplicial polyhedra are rigid \cite{Whit84}.
The edge-skeleta of these polyhedra constitute the subclass of planar (in the graph-theoretical sense) Geiringer graphs.
On the other hand, there is no full characterization for the whole class of Geiringer graphs, since Maxwell's condition is not always sufficient for minimal rigidity in dimension 3 or higher (see Figure~\ref{fig:doubleb} for a counter-example).
\begin{figure} \begin{center}
\includegraphics[width=0.4\textwidth]{double_banana.png}
\caption{The double-banana graph satisfies Maxwell's condition in ${\mathbb R}^3$, but is not rigid since its two rigid components revolve around the dashed axis.\label{fig:doubleb}}
\end{center}\end{figure}
Rigid graphs in ${\mathbb R}^2$ and ${\mathbb R}^3$ have attracted the principal interest of research because of their numerous applications.
Such applications can be found in robotics \cite{Rob2,Drone},
molecular biology \cite{Bio1,Bio2}, sensor network localization \cite{sensor} and architecture \cite{arch1,arch2}.
Beyond these applications, rigidity theory has also a mathematical interest on its own and
results on rigidity can be extended to arbitrary dimension and to other manifolds \cite{handbook1,tay,NOP2012}.
More specifically, the set of minimally rigid graphs on every unit $d$-dimensional sphere $S^d$, coincides with the set of minimally rigid graphs in ${\mathbb R}^d$ \cite{Whiteley_cone}.
The main contribution of this paper is to employ a variety of algorithms from algebraic geometry in order to investigate minimally rigid graphs, relying on the observation that they can be adequately modeled by well-constrained algebraic systems.
The total number of edge constraints according to Maxwell's theorem equals the total number of degrees of freedom (dof) of the rigid graph embedding.
They equal the total number of vertex coordinates, namely $d\cdot n$, after subtracting the $\binom{d+1}{2}$ dof of rigid motions (rotations and translations).
Besides, the inequalities in Maxwell's condition exclude any over-constrained subsystem.
We shall concentrate on complex embeddings.
Note that the number of complex embeddings of a minimally rigid graph $G$ is the same for all generic assignments of lengths $\bm{\lambda}$
(see~\cite{Jackson} for Laman graphs and~\cite{SomWam} on coefficient-parameter theory).
This number is denoted by $c_d(G)$ and is bounded by upper bounds on the number of (complex) solutions of the algebraic system.
\paragraph{Previous work:}
A major open question in rigidity theory is to obtain optimal asymptotic upper bounds on the number of embeddings (see for example \cite{Jackson}).
According to Maxwell's condition, there are $O(d n)$ edge constraints that can be represented by the quadratic Euclidean edge length equations $\displaystyle\lVert \mathbf{x}_u-\mathbf{x}_v \rVert^2~=~\lambda_{u,v}^2$.
Applying B\'ezout's bound, this formulation yields $O(2^{dn})$ as an upper bound on the number of embeddings.
In \cite{Steffens} the authors used mixed volume techniques on a set of equations that reformulates the quadratic edge length constraint equations to compute better BKK bounds (named after Bernstein, Khovanskii and Kushnirenko) on the number of embeddings for Laman graphs.
We will also use this formulation, however, their polyhedral methods did not improve B\'ezout's asymptotic bound in the general case.
Besides edge lengths equations, there are distance geometry methods to verify embeddings in ${\mathbb R}^d$, subject to determinantal equations and inequalities obtained from Cayley-Menger matrices \cite{Blu}.
The best known upper bound for the number of embeddings of a rigid graph in dimension $d$ uses these determinantal varieties, applying a theorem on their degree \cite{Borcea} (for the degree of determinantal varieties, see~\cite{HarrisTu}):
$$
\prod\limits_{k=0}^{n-d-2} \frac{\binom{n-1+k}{n-d-1-k}}{\binom{2k+1}{k}} .
$$
This bound does not improve upon B\'ezout's bound asymptotically.
Let us now juxtapose these to lower bounds.
Direct computations for Laman graphs have proved that there are graphs with $2.50798^n$ complex embeddings in $\mathbb{C}^2$ and Geiringer graphs with $3.06825^n$ complex embeddings \cite{GKT17}.
On the sphere it is proven that there are graphs with $2.5698^n$ complex embeddings \cite{count_sphere}.
Computations on the number of embeddings for all graphs with given $n$ have been completed only for relatively small $n$ due to the amount of computation required (up to $n=13$ in dimension~2 and $n=10$ in dimension~3).
In any case, the gap between asymptotic upper bounds and experimental results remains enormous.
Finding the exact number of complex solutions requires some demanding computations.
In the case of embeddings of Laman graphs both in $\mathbb{C}^2$ and $S^2$, there are combinatorial algorithms \cite{Joseph_lam,count_sphere} that speed up computations a lot, but still it is almost infeasible to compute $c_d(G)$ for graphs with more than 18 vertices in a desktop computer.
Since no similar algorithm exists for Geiringer graphs, solving algebraic systems for random instances is the only option:
Groebner bases~\cite{GKT17} and homotopy solvers like \texttt{phcpy} \cite{phcpy,belt} have been employed.
For the 11-vertex case, \texttt{phcpy} fails to give all solutions for many graphs, while Groebner bases may take more then 3 days for a single graph (if the algorithm terminates).
Another homotopy solver that has come recently to our attention is \texttt{MonodromySolver} \cite{monodromySolver}, implemented in \texttt{Macaulay2}.
We tested this solver for a variety of graphs and it seems more accurate and considerably faster than \texttt{phcpy}.
Besides computing the exact number of embeddings, complex bounds have been considered as general estimates of the number of solutions.
These bounds can also serve as input to homotopy continuation solvers.
Mixed volume computation has been used, by applying suitable variable transformations \cite{etv,belt}.
Even if mixed volume computations are generally faster than exact estimations, the computational limits of this method are also rather restrained.
Let us comment that there are real algebraic bounds \cite{BihSot,Khovanskii} which are sharper than the complex ones for certain polynomial structures.
These bounds are much higher than the (more general) complex bounds already mentioned in the case of minimally rigid graphs.
\paragraph{Our contribution.}
In the present paper we use m-B\'ezout bounds on the number of complex embeddings.
We present a new recursive combinatorial algorithm that adopts a graph-theoretic approach in order to speed up the computation of m-B\'ezout bounds in our case, based on a standard partition of variables.
We also use matrix permanent computation, which are known to compute m-B\'ezout bounds \cite{emvid}, comparing runtimes with our algorithm.
Applying the best known upper bounds for orientations of planar graphs~\cite{Felsner} and permanents of $(0,1)$-matrices (known as Br\`egman-Minc bound \cite{Bre73,Minc63}), we improve asymptotic upper bounds for planar minimally rigid graphs in dimension~3, and for all minimally rigid graphs in dimensions $d\geq 5$.
We compare the m-B\'ezout bounds with mixed volume bounds and the actual number of complex embeddings of all Laman and Geiringer graphs with $n\leq 9$ vertices, and some selected Laman graphs up to $n=18$ and Geiringer graphs up to $n=12$, and observe that m-B\'ezout is exact for the large majority of spherical embeddings in the case of planar Laman graphs, while it is exact for all planar Geiringer graphs.
We adjust the Bernstein's discriminant conditions on the exactness of mixed volume to the case of m-B\'ezout bounds using Newton polytopes whose mixed volume equals to the m-B\'ezout.
Our method exploits the system structure to ensure that some conditions are verified {\em \`a priori}, reducing the number of checks required by an exponential factor.
This number of conditions remains exponential, but based on our experiments
we conjecture that it suffices to check only a linear number of cases overall.
These results are highlighted by certain examples.
The rest of the paper is structured as follows.
The next section discusses some background and introduces the algebraic formulation that we exploit.
Section~\ref{sec:mBez} offers two approaches for efficiently computing m-B\'ezout bounds: first a combinatorial method and, second, a reduction to the permanent thus improving bounds for planar graphs in dimension $3$ and for all graphs in the case of $d\geq 5$.
Section~\ref{sec:exact} studies the exactness of m-B\'ezout bounds.
The paper concludes with open questions.
\section{Henneberg constructions and algebraic formulation}\label{sec:eq-sphere}
In this section we present some preliminaries about the construction of rigid graphs and the computation of their embeddings.
In general, Maxwell's condition is not suitable to find the set (or a superset) of all minimally rigid graphs with a given number of vertices.
On the other hand, a sequence of moves known as Henneberg steps can construct such sets.
Additionally, a characterization of minimally rigid graphs up to the last Henneberg move can be used to separate cases that have a trivial number from the the non-trivial ones.
Subsequently, we present an algebraic formulation that is based on a variant of the quadratic edge lengths equations.
This formulation has been used several times in the context of studies on rigid graphs that exploit sparse elimination techniques \cite{Steffens,etv,belt}.
\subsection{Henneberg steps}\label{sec:Hsteps}
\begin{figure}[htb!]
\centering
\hspace*{-6mm}
\begin{tabular}{cc|cccc}
H1 & H2 \hspace*{1mm}& \hspace*{1mm}H1 & H2 & H3x & H3v\\
\HIiid & \HIIiid \hspace*{1mm} & \hspace*{1mm} \HIiiid & \HIIiiid & \HIIIx & \HIIIv \\
\small{2-degree} & \small{edge split} \hspace*{1mm} & \hspace*{1mm} \small{3-degree}& \small{edge split} & \small{X-replacement} & \small{double}\\
\small{vertex addition}& \small{in 2d} \hspace*{1mm} & \hspace*{1mm} \small{vertex addition}& \small{in 3d} & & \small{V-replacement} \\
\end{tabular}
\caption{Henneberg steps in dimensions 2 \& 3.}
\end{figure}
Minimally rigid graphs in ${\mathbb R}^d$ can be constructed as a sequence of Henneberg moves starting from the complete graph on $d$ vertices $K_d$.
In the case of $d=2$ all minimally rigid graphs can be obtained by Henneberg 1 (H1) and Henneberg 2 (H2) operations, giving one more method to characterize Laman graphs.
On the other hand these two moves are not sufficient to construct all the minimally rigid graphs in $d=3$, so an extended Henneberg step is required.
These 3 moves give a superset of Geiringer graphs.
It is conjectured that H1, H2 and H3 completely characterize rigid graphs in ${\mathbb R}^3$ \cite{handbook1,tay}.
These moves generalize to arbitrary dimension.
The H1 move in dimension $d$ adds a new $d$-degree vertex, while H2 adds a $(d+1)$-degree vertex removing also an edge.
It has been proven that these moves always preserve rigidity in all dimensions.
On the other hand, H3 step in dimension~4 is not always rigid~\cite{H3contre}.
We used Henneberg steps to construct sets of Laman and Geiringer graphs up to isomorphism, using canonical labeling as in \cite{Joseph_lam,GKT17}.
Since Henneberg moves add a vertex with a fixed degree, we can separate the sets of graphs with the same number of vertices up to their minimal degree.
So if a graph in dimension $2$ has minimal degree $2$, then it can be constructed with an H1 move in the last step.
If the minimal degree is $3$ it certainly requires an H2 move in the last step of the Henneberg sequence.
Notice that the H1 move trivially doubles the number of embeddings, since the new vertex lies in the intersection of $d$ different $(d-1)$-spheres.
This means that we may examine only graphs that are constructed with the other Henneberg moves.
Let us comment that the Geiringer graphs whose construction requires an H3 move have minimal degree $5$ and no such graph exist for any graph with $n\leq 11$ vertices.
\begin{table}
\caption{Numbers of Laman and Geiringer graphs up to the last Henneberg move and graph planarity.}
\centering
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|}
\multicolumn{10}{c}{\textbf{Laman graphs}} \vspace{1.5mm}\\
\hline
$\bm{n}$ & $\bm{3}$ & $\bm{4}$ & $\bm{5}$ & $\bm{6}$ & $\bm{7}$ & $\bm{8}$ & $\bm{9}$ & $\bm{10}$ & $\bm{11}$ \\
\hline
H1 planar & $1$ & $1$ & $3$ & $11$ & $62$ & $491$ & $5,041$ & $60,040$ & $791,195$ \\
H1 non-planar & - & - & - & - & $4$ & $85$ & $1,917$ & $46,903$ & $1,201,401$ \\
\hline
H2 planar & - & - & - & $1$ & $3$ & $18$ & $122$ & $1,037$ & $9,884$ \\
H2 non-planar & - & - & - & $1$ & $1$ & $14$ & $142$ & $2,152$ & $36,793$ \\
\hline
\multicolumn{10}{c}{ }
\end{tabular}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|}
\multicolumn{9}{c}{\textbf{Geiringer graphs}} \vspace{1.5mm}\\
\hline
$\bm{n}$ & $\bm{4}$ & $\bm{5}$ & $\bm{6}$ & $\bm{7}$ & $\bm{8}$ & $\bm{9}$ & $\bm{10}$ & $\bm{11}$ \\
\hline
H1 planar & $1$ & $1$ & $1$ & $4$ & $12$ & $45$ & $221$ & $1,215$ \\
H1 non-planar & - & - & $2$ & $16$ & $ 299 $ & $9,718$ & $527,250$ & $41,907,790$ \\
\hline
H2 planar & - & - & $1$ & $1$ & $2$ & $5$ & $12$ & $34$ \\
H2 non-planar & - & - & - & $5$ & $61$ & $1,719$ & $ 85,401 $ & $6,267,144$ \\
\hline
\end{tabular}
\label{tab:Hsteps}
\end{table}
\subsection{Sphere equations}
In order to compute the number of embeddings of a rigid graph we used some standard algebraic formulation \cite{Steffens,etv}.
First we remove rigid motions by fixing $\binom{d+1}{2}$ coordinates yielding a $0-$dimensional system.
In the case of dimension $2$, we may fix both coordinates of one vertex and one coordinate of a second vertex.
If these vertices are adjacent to one edge, then the length constraint imposes only one solution for the remaining coordinate of the second vertex up to rotations.
In general, if the graph contains a complete subgraph with $d$ vertices $v_1,v_2,\dots v_d $, then we can choose the coordinates of this $K_d$ graph in a way that they satisfy the edge lengths of this subgraph.
So, in the case of Laman graphs, we need to fix an edge, while in 3 dimensions a triangle should be fixed.
Note that for the first set of graphs there is always a $K_2$ (edge).
As for the 3-dimensional case, Geiringer graphs with no triangles ($K_3$) are very rare (the first one is the 10-vertex $K_{6,4}$).
In that situation, the vertices of an edge are fixed, while a third vertex is located on the same fixed plane as the edge, leaving two degrees of freedom of this vertex free.
In that way, every embedding is counted twice, so the number of solutions is divided by 2.
We use two types of equations in our systems.
The first set of equations at hand are the \textit{edge equations}, which represent the edge length constraints between the adjacent vertices of an edge.
Although these equations suffice to find the embeddings of a graph, we cannot take advantage of their structure and compute efficient bounds.
To overcome this problem, we define the set of \textit{magnitude equations} that introduce new variables representing the distance of each vertex from the origin.
Following \cite{belt}, we call the combination of these two sets of equations \textit{sphere equations}.
\begin{definition}
\label{def:sphereEquations}
Let $G=(V,E)$ be a graph with $n$ vertices.
We denote by $\bm{\lambda}=(\lambda_{u,v})_{(u,v)\in E}\in {\mathbb R}_+^{|E|}$ the (given) edge lengths and by $\widetilde{X}_u=\{x_{u,1},x_{u,2},\dots x_{u,d}\}$ the variables assigned to the coordinates of each vertex.
The following system of equations gives the embeddings of $G$:
\begin{equation} \label{sphere_eq} \begin{split}
\lVert \widetilde{X}_u \rVert^2 &= s_u , \qquad\, \forall\, u \in V\,, \\
s_u +s_v -2\langle \widetilde{X}_u,\widetilde{X}_v \rangle &= \lambda_{u,v}^2, \qquad \forall\, (u,v) \in E\backslash edges(K_d)\,,
\end{split}
\end{equation}
where $\langle \widetilde{X}_u,\widetilde{X}_v \rangle$ is the Euclidean inner product.
We will denote the set of variables $X_u=\widetilde{X}_u \bigcup \{s_u\}$ in the euclidean case using $s_u$ as the $(d+1)$-th variable $x_{u,d+1}$.
If a vertex is fixed, its variables are substituted with constant values.
This formulation can be obviously used in the case of embeddings on the unit $d$-dimensional sphere $S^d$ using $|\widetilde{X}_u|=d+1$ coordinates and setting $s_u=1$.
\end{definition}
This algebraic system has $m=d\cdot n-d^2$ edge equations and $n-d$ magnitude equations if there is at least one subgraph $K_d$ of $G$.
Notice that the edges of the fixed $K_d$ serve to specify the fixed vertices and are not included in this set of equations, so $m < |E|$.
We remark that the m-B\'ezout bound (or the mixed volume bound) of a graph $G$ may vary up to the choice of the fixed $K_d$, so one needs to compute m-B\'ezout bounds up to all fixed $K_d$'s in order to find the minimal one.
On the other hand $c_d(G)$ is invariant under different choices of fixed $K_d$.
\section{Computing m-B\'ezout bounds}\label{sec:mBez}
In this section we propose two methods for computing m-B\'ezout bounds for minimally rigid graphs.
First, we give definitions of two classical complex bounds (m-B\'ezout, BKK) and we propose a natural partition in the set of variables for the m-B\'ezout.
Subsequently, we introduce an algorithm based on a relation between indegree constrained graph orientations and the m-B\'ezout bound of minimally rigid graphs.
Besides that, we also give an alternative way to compute this bound via the matrix permanent.
Let us mention that matrix permanents have been already used to bound the number of Eulerian orientations (which are graph orientations with equal indegree and outdegree for every vertex) in~\cite{Schrijver1983}, but to the best of our knowledge there are no published results on the connection between matrix permanents and indegree constrained graph orientations in the general case.
Consequently, we use an existing bound on the orientations of planar graphs to improve the asymptotic upper bounds for planar Geiringer graphs.
We also improve asymptotic upper bounds for $d\geq5$ applying a bound on the permanents of $(0,1)$-matrices.
Finally, we compare our combinatorial algorithm with existing methods to compute upper bounds or the actual number of embeddings in the cases of $\mathbb{C}^2$ and $\mathbb{C}^3$.
\subsection{The m-B\'ezout bound of the sphere equations}\label{sec:SphmBez}
Let us start with defining the bound in question.
It is based on a classical theorem on bounding the number of solutions of 0-dimensional well-constrained (or square) algebraic systems, see e.g.\ \cite{Shafarevich2013}:
\begin{theorem}
\label{thm:mBezout}
Let $X_1,X_2,\dots,X_k$ be a partition of the $m$ variables of a square algebraic system, such that $m_i=|X_i|$ and $m=m_1+m_2+ \dots + m_k$.
Let $\alpha_{i,j}$ be the degree of the i-th equation in the j-th set of variables.
Then, if the system has a finite number of solutions, it is bounded from above by the coefficient of the monomial $X_1^{m_1}\cdot X_2^{m_2} \cdots X_k^{m_k}$ in the polynomial
\begin{equation}
\prod_{i=1}^{m} (\alpha_{i,1}\cdot X_1+\alpha_{i,2} \cdot X_2+ \dots + \alpha_{i,k} \cdot X_k).
\label{eq:mbez}
\end{equation}
\end{theorem}
This so-called m-B{\'e}zout bound improves considerably the classical B{\'e}zout bound, if the variables can be grouped in an appropriate way, exploiting the structure of the polynomial system.\\
The structure of polynomials can be exploited further using sparse elimination techniques.
One of the basic tools in sparse elimination theory is the notion of Newton Polytopes of a polynomial system.
Given
$$
f= \sum \limits_{\alpha\in\ZZ^m} c_{\alpha} \mathbf{x}^{\alpha} \in K[x_1,x_2, \dots , x_m ],\quad \mathbf{x}=(x_1,\dots,x_m),\; c_\alpha\ne 0,
$$
the Newton polytope $NP(f)$ is defined as the convex hull of its monomial exponent vectors $\alpha$.
A very important theorem relates the structure of these polytopes to the number of solutions of algebraic systems in the corresponding toric variety.
The toric variety is a projective variety defined essentially by the Newton polytopes of the given system and contains the topological torus $(\mathbb{C}^*)^m$ as a dense subset.
The set-theoretic difference of a toric variety and $(\mathbb{C}^*)^m$ is toric infinity in correspondence with projective infinity.
In practical applications, as in this paper, one is interested in roots in the torus $(\mathbb{C}^*)^m$, hence all toric roots except from those lying at toric infinity, in other words only in affine toric roots.
\begin{theorem}[BKK bound \cite{Bernshtein1975,Kouchnirenko1976,Khovanskii1978}]\label{thm:BKK}
Let $ (f_i)_{1 \leq i \leq m } $ be a square
system of equations in
$\mathbb{C}[x_1,x_2, \dots , x_m ]$, and let $(NP(f_i))_{1\leq i\leq m}$ be their Newton Polytopes.
Then, if the number of system's solutions in $(\mathbb{C}^*)^m$ is finite, it is bounded above by the mixed volume of these Newton Polytopes.
\end{theorem}
Roughly, without paying much attention to the underlying variety, we have the following relations as in~\cite{SomWam}:
$$
\# \text{complex solutions} \leq \text{ mixed volume} \leq \text{ m-B{\'e}zout} \leq \text{ B{\'e}zout}. $$
On the other hand, the complexity of computing bounds goes in the opposite direction.
Here, we concentrate on the m-B\'ezout bound of the sphere equations of a graph $G=(V,E)$ up to a fixed complete subgraph $K_d$.
For the rest of the text, unless further specified, $K_d$ will denote a given complete subgraph and not all possible choices.
In general finding the optimal multihomogeneous partition is not in APX \footnote{APX is the class of all NP optimization problems, for which there exist polynomial-time approximation algorithms.
This approximation is bounded by a constant (for further details see \cite{APX}).}, unless P=NP \cite{Malajovich2007}.
Here we choose a natural partition such that each subset of variables contains these ones which correspond to the coordinates and the magnitude of a single vertex $X_u$.
In order to compute the m-B{\'e}zout bound, we will separate the magnitude equations from the edge equations.
In the first ones, there is only one set of variables with degree 2, while in every edge equation the degree of the $k$-th set of variables is always~1:
\begin{equation*} \label{bez_sphere}
\begin{split}
\prod_{u \in V'} 2\cdot X_u \prod_{(u,v) \in E'} (X_{u}+X_{v}) =
2^{n-d} \cdot \prod_{u \in V'} X_u \prod_{(u,v) \in E'} (X_{u}+X_{v}),
\end{split}
\end{equation*}
where $E'=E\backslash edges(K_d)$, $V'=V\backslash vertices(K_d)$ and $X_{u} \equiv 0 $ if $u \in K_d$.
This means that we only need to find the coefficient of the monomial $ \displaystyle\prod\limits_{u \in V'} X_u^d$ in the polynomial of the product:
\begin{equation} \label{eq:graph2mBezout}
\prod_{(u,v) \in E'} (X_{u}+X_{v})
\end{equation}
Let us denote this coefficient by $mB_{E}(G,K_d)$, which is related only to the combinatorial structure of edge equations.
The m-B{\'e}zout bound for the number of embeddings of a graph in $\mathbb{C}^d$ up to a fixed $K_d$ is $mB(G,K_d)=2^{n-d} \cdot mB_{E}(G,K_d)$.
Notice that this bound is the same for spherical embeddings in $S^d$.
\subsection{A combinatorial algorithm to compute m-B{\'e}zout bounds}\label{sec:comb}
This subsection focuses on m-B{\'e}zout bounds for minimally rigid graphs.
Our method is inspired by two different approaches that characterize Laman graphs.
First, Recski's theorem states that if a graph is Laman then any multigraph obtained by doubling an edge should be the union of two spanning trees \cite{handbook1}.
Additionally, pebble games give a relation between the existence of an orientation and the number of constraints in a graph and its subgraphs \cite{pebble}.
The following theorem gives a combinatorial method to compute the m-B{\'e}zout bound, proving that $mB_{E}(G,K_d)$ is exactly the number of certain indegree-constrained orientations.
\begin{theorem}
Let $H_{K_d}(V,E')$ be a graph obtained after removing the edges of $K_d$ from $G=(V,E)$.
We define $\mathcal{H}_{K_d}$ as the set of all orientation of $H_{K_d}$ such that each non-fixed vertex has indegree $d$ and each fixed vertex has indegree $0$.
Then, the coefficient of the monomial $ \displaystyle\prod\limits_{u \in V'} X_u^d$ of the product in expression~(\ref{eq:graph2mBezout}) is exactly the same as $|\mathcal{H}_{K_d}|$.
\end{theorem}
\begin{proof}
By expanding the product $\prod\limits_{(u,v) \in E'} (X_{u}+X_{v})$, the monomial $ \displaystyle\prod\limits_{u \in V'} X_u^d$ can be obtained when each $X_{u}$ from a given edge contributes exactly $d$ times in that product.
This means that every time we shall choose one of the two sets of variables that correspond to the adjacent vertices of the edge represented in the parenthesis.
This choice yields an orientation in the directed graph and vice versa.
Thus, the number of different orientations in all edges gives us how many times this monomial will appear in the expansion, completing the proof.
\end{proof}
This theorem gives another way to prove that an H1 move doubles the m-B\'ezout bound of minimally rigid graphs.
Hence, minimally rigid graphs constructed only by H1 moves have at most $2^{n-d}$ embeddings (actually this bound is tight, see Section~\ref{sec:Hsteps}).
\begin{corollary}\label{cor:orientH1}
An H1 move always doubles the m-B{\'e}zout bound up to the same fixed $K_d$.
\end{corollary}
\begin{proof}
Let $|\mathcal{H}_{K_d}|$ be the number of indegree-constrained orientations for a graph $G$ with $n$ vertices up to a given $K_d$.
This means that the m-B\'ezout bound is
$$
mB(G,K_d)=2^{n-d}\cdot |\mathcal{H}_{K_d}| .
$$
Now, let $G^*$ be a graph obtained by an H1-move on the graph $G^*$.
Since H1 adds a degree-$d$ vertex to $G$, this means that there is only one way to reach indegree $d$ for the new vertex of $G^*$.
So the indegree-constrained orientations of $G^*$ up to the same $K_d$ are exactly $|\mathcal{H}_{K_d}|$ and
$$
mB(G^*,K_d)=2^{n+1-d}\cdot |\mathcal{H}_{K_d}|=2\cdot mB(G,K_d).
$$
The proof concludes by induction: Starting from $K_d$, the m-B\'ezout bound of a minimally graph constructed only by H1 moves is $2^{n-d}$.
\end{proof}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=0.7\textwidth]{orientations.png}
\caption{The orientations of graphs $L56$ and $G48$.
Notice that there is only one way to direct the red edges up to the choice of $K_d$ (dashed blue).
\label{fig:orient}}
\end{center}
\end{figure}
\begin{algorithm}[htp!]\label{alg:orient}
\caption{Count graph orientations}
\DontPrintSemicolon
\KwFn{orient}\\
\KwInput{$n$ (\# of vertices), $E$ (graph edges $\backslash K_d$), $indeg$ (desired indegree list. \\ If vertex $u$ fixed then $indeg[u]=0$, otherwise $indeg[u]=d$), }
\KwOutput{\# of indegree-constrained orientations}
deg = vertex degrees of graph $G([n],E)$ \\
\tcc{Ending condition for the recursion}
\If{$|E|=0$ }
{
\Return(1)
}
\tcc{No valid orientations in this case}
\If{$\exists u, indeg\left[u\right]>deg\left[u\right]$ or $indeg\left[u\right]<0$ }{
\Return(0)
}
\tcc{Examine the conditions yielding unique orientations}
\For{$u\leq n$}
{
\If{$indeg[u]=0$ \tcp*{$u$ admits only new outdirerected edge orientations}}
{
\For{all edges $(u,v) \in E$}{
$indeg[v]=indeg[v]-1$\\
$E'$=$E\backslash \{(u,v)\}$\\
$newdeg=$vertex degree of graph $G'(V,E')$\\
\Return(orient($n$, $Enew$, indeg, $newdeg$))
}
}
\ElseIf{$indeg[u]=deg[u]$ \tcp*{$u$ admits only new indirerected edge orientations}}
{
\For{all edges $(u,v) \in E$}{
$indeg\left[u\right]=indeg\left[u\right]-1$\\
$E'$=$E\backslash \{(u,v)\}$\\
$newdeg=$vertex degrees of graph $G'(V,E')$\\
\Return(orient($n$, $E'$, $indeg, newdeg$))
}
}
}
\tcc{No more unique orientations exists: set both orientations for 1st edge}
$(u,v)=E\left[1\right]$ \\
$indeg1\left[u\right]=indeg\left[u\right]-1$\\
$indeg2\left[v\right]=indeg\left[v\right]-1$\\
$E'$=$E\backslash \{(u,v)\}$\\
$newdeg=$vertex degree of graph $G'(V,E')$\\
orient1=orient($n$, $E'$, $indeg1, newdeg$)\\
orient2=orient($n$, $E'$, $indeg2, newdeg$)\\
\Return(orient1+orient2)
\end{algorithm}
Let us demonstrate our method examining one Laman and one Geiringer graph.
\begin{example}
Here are two examples of this counting method in the case of $L56$ graph in dimension $2$ and $G48$ \footnote{The graphs in this example and the following ones are named after their class ($L$ for Laman and $G$ for Geiringer) and the number of their embeddings in the correspondent euclidean space as in \cite{belt}.} in dimension $3$, which are both 7-vertex graphs (see Figure~\ref{fig:orient}).
Graph $L56$ has $56$ complex embeddings in the plane and $64$ embeddings on the sphere, while $G48$ has $48$ embeddings in $\mathbb{C}^3$ (these numbers coincide also with the maximum number of real embeddings) \cite{belt,GKT17,count_sphere}.
The mixed volumes of the algebraic systems are $64$ and $48$ respectively.
The dashed lines indicate the fixed edges of $K_d$.
The edge direction for any edge that includes the fixed points is always oriented towards the non-fixed vertex.
This yields a \emph{unique orientation} up to the fixed $K_d$, which is coloured in red for both graphs.
The rest of the graph admits $|\mathcal{H}_{K_2}|=2$ orientations for $L56$, while the number of different orientations for $G48$ is $|\mathcal{H}_{K_3}|=3$.
So the m-B{\'e}zout bound is $2^{7-2} \cdot 2=64$ for $L56$ and $2^{7-3} \cdot 3=48$ for $G48$.
\end{example}
We have implemented a software tool in Python to count the number of orientations for an arbitrary graph given the desired indegrees.
The basic part of this code (see Algorithm~\ref{alg:orient}) is to decide recursively which choices of direction are allowed in every step \cite{code_mBezout}.
\subsection{Computing m-B{\'e}zout bounds using the permanent}
The permanent of an $m \times m$ matrix $A=(a_{i,j})$ is defined as follows:
\begin{equation}\label{eq:permanent}
\text{per }(A)= \sum\limits_{\sigma \in S_m} \prod_{i=1}^{m} a_{i,\sigma(i)},
\end{equation}
where $S_m$ denotes the group of all permutations of $m$ integers.
One of the most efficient ways to compute the permanent is by using Ryser's formula~\cite{ryser}:
\begin{equation}\label{eq:Ryser}
\text{per }(A)= \sum \limits_{M \subseteq \{1,2,\dots ,m\}} (-1)^{m-|M|} \prod_{i=1}^{m} \sum\limits_{j \in M} a_{i,j} .
\end{equation}
There is a very relevant relation between $\text{per}(A)$ and the m-B{\'e}zout bound, see \cite{emvid}:
\begin{theorem}\label{thm:perm}
Given a system of algebraic equations and a partition of the variables in $k$ subsets, as in Theorem~\ref{thm:mBezout}, we define the square matrix $A$ with $m=\sum\limits_{j=1}^{k} m_j$ rows, where each set of variables corresponds to a block of $m_j$ rows.
Let $\alpha_{i,j}$ be the degree of the $i$-th equation in the $j$-th set of variables.
The columns of $A$ correspond to the equations, where the subvector of the $i$-th column associated to the $j$-th set of variables has $m_j$ entries, all equal to $\alpha_{i,j}$.
Then, the m-B{\'e}zout bound of the given system is equal to
\begin{equation}\label{per}
\frac{1}{m_1! \, m_2 ! \cdots m_k!} \cdot \mbox{\rm per}(A) .
\end{equation}
\end{theorem}
We will refer to $A$ matrix as the \textit{m-B\'ezout matrix} of a polynomial system.
This implies that in the case of minimally rigid graphs, we obtain a square m-B\'ezout matrix $A$ with columns associated to the equations of non-fixed edges, and $n-d$ blocks of $d$ rows each, corresponding to the non-fixed vertices.
An entry $(r,c)$ is one if the vertex corresponding to $r$ is adjacent to the edge corresponding to the equation indexing $c$, otherwise it is zero.
This is an instance of a $(0,1)$-permanent.
Therefore Theorem~\ref{thm:perm} gives the coefficient
$$
mB_{E}(G,K_d)=\left(\frac{1}{d!}\right)^{n-d} \cdot \mbox{\rm per}(A),
$$
in bounding the system's roots, since all $m_i=d$, while $k=n-d$.
The effect of the magnitude equations implies that we should multiply $mB_{E}(G,K_d)$ by $2^{n-d}$ as in Subsection~\ref{sec:SphmBez}.
This yields the corollary below.
\begin{corollary}
The m-B{\'e}zout bound of the sphere equations for an $n$-vertex rigid graph in $d$ dimensions up to a given $K_d$ is exactly
\begin{equation} \label{mbe}
mB(G,K_d)= \displaystyle \left(\frac{2}{d!}\right)^{n-d} \cdot \mbox{\rm per}(A),
\end{equation}
assuming that matrix $A$ is the m-B\'ezout matrix up to $K_d$ defined as above.
\end{corollary}
The permanent formulation for the computation of the m-B\'ezout bound gives us another way to prove Corollary~\ref{cor:orientH1}.
\begin{corollary}
Let $G$ be a minimally rigid graph in $\mathbb{C}^d$ with $n$ vertices and $A$ be its $(m \times m)$ m-B\'ezout matrix up to a fixed $K_d$.
Then, for every graph $G^*$ obtained by an H1 operation on $G$, the permanent of its m-B\'ezout matrix $A^*$ up to the same $K_d$ is $\text{per }(A^*)=d! \cdot \text{per }(A)$.
\end{corollary}
\begin{proof}
Without loss of generality, we consider that the last $d$ rows of matrix $A^*$ represent the new vertex, while the last $d$ columns of this matrix represent the edges adjacent to this vertex, since matrix permanent is invariant under row or column permutations.
The rest of the matrix is the same as $A$.
This yields the following structure:
$$
A^*= \begin{pmatrix}
A & A' \\\
\mathbf{0} & \mathbf{1}
\end{pmatrix}
$$
where $\mathbf{0}$ is a $(d \times m)$ zero submatrix, $\mathbf{1}$ is a $(d \times d)$ submatrix with ones and $A'$ the $(m \times d)$ submatrix of the new edge columns without the new rows.
It is clear from the definition of the permanent (See Equation~\ref{eq:permanent}), that column permutations that do not include a zero entry are counted as 1 in this sum, while if they include a zero entry the product is zero.
The only column permutations that do not include a zero entry for the $d$ last rows are those that are related to the $d$ last edges, so there are $d!$ nonzero column permutations for this block of rows.
This means that the permutations for the other $m$ rows exclude the last three columns, so they are exactly $\text{per }(A)$ permutations in this case.
Thus, $\text{per }(A^*)=d! \cdot \text{per }(A)$.
\end{proof}
Since $mB(G^*,K_d)=\displaystyle\left(\frac{2}{d!}\right)^{n+1-d} \cdot \text{per }(A^*)$, it follows that
$$
mB(G^*,K_d)=2\cdot mB(G,K_d),
$$
as in Corollary~\ref{cor:orientH1}. \\
Let us give an example of this counting method for a minimally rigid graph.
\begin{figure}[htp!]
\centering \Lmaxeight
\caption{The $L136$ graph. The dashed edge is the fixed one. \label{fig:L136}}
\end{figure}
\begin{example}
We use the $L136$ graph to provide an example for this formulation (other examples can be found in \cite{code_mBezout}).
$L136$ is the 8-vertex Laman graph with the maximal embedding number $c_2(G)=136$ among all Laman graphs with the same number of vertices \cite{Joseph_lam}.
On $S^2$, it has $192$ complex embeddings, which is also maximum (but not unique), since there is another graph sharing the same $c_{S^2}(G) $.
The m-B\'ezout matrix $A_{L136}$ for this graph for the fixed edge $(1,2)$ is the following:
\begin{center} \footnotesize
\hspace*{-9mm} \begin{tabular}{c|cccccccccccc|}
& $(1,4)$ & $(1,8)$ & $(2,3)$ & $(2,5)$ & $(2,7)$ & $(3,4)$ & $(3,5)$ & $(4,6)$ & $(4,8)$ & $(5,6)$ & $(6,7)$ & $(7,8)$
\\ \hline
$x_3$ & 0 & 0 & 1 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 \\
$y_3$ & 0 & 0 & 1 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 \\
$x_4$ & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 0 & 0 & 0 \\
$y_4$ & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 0 & 0 & 0 \\
$x_5$ & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 \\
$y_5$ & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 \\
$x_6$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 0 \\
$y_6$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 1 & 0 \\
$x_7$ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\
$y_7$ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\
$x_8$ & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1\\
$y_8$ & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1\\
\hline
\end{tabular}
\end{center}
and its permanent is $\text{per }(A_{L136})=192$, which gives the m-B\'ezout bound since $d=2$.
\end{example}
\subsection{Improving asymptotic upper bounds}
We make use of both approaches to improve upon asymptotic upper bounds on the embedding number.
Despite various algebraic formulations and approaches on root counts, it is surprising that the best existing asymptotic upper bounds on the embedding number of minimally rigid graphs are in the order of $O\left(2^{dn} \right)$ and, thus, essentially equal the most straightforward bound on the most immediate system, namely B\'ezout's bound on the edge equation system \cite{Borcea,Steffens}.
On the other hand, the lower bounds keep on improving (please refer to the Introduction) but still an important gap remains.
This is indicative of the hardness of the problem.
First, we make use of the following proposition on the asymptotic bounds for the orientations of planar graphs in order to improve the asymptotic upper bound of planar Geiringer graphs, which are the only fully characterized class of minimally rigid graphs in 3d space, and hence of special interest.
\begin{proposition}[Felsner and Zickfeld~\cite{Felsner}]
The number of indegree constrained orientations of a planar graph is bounded from above by
\begin{equation}
2^{n-4} \cdot \prod_{u \in I} \left(2^{-deg(u)+1} \cdot \binom{deg(u)}{indeg(u)} \right)
\end{equation}
where $I$ is an independent set of the graph, $deg(u)$ and $indeg(u)$ are respectively the degree and the indegree of a vertex $u$.
Furthermore, in the case of $indeg(u)=3$ this bound asymptoticaly behaves as $3.5565^n$.
\end{proposition}
Given the relation between m-B\'{e}zout bounds and graph orientations (see Section~\ref{sec:comb}), this proposition leads to the following improvement upon the asymptotic upper bound for the number of embeddings of the subclass of planar Geiringer graphs.
\begin{theorem}
Planar Geiringer graphs have at most $O\left(7.1131^n\right)$ embeddings.
\end{theorem}
We also employ the permanent to obtain asymptotic improvement upon B\'ezout's asymptotic bound for $d\ge 5$ by using the following bound.
\begin{proposition}[Br\`egman~\cite{Bre73}, Minc~\cite{Minc63}]
For a $(0,1)$-permanent $A$ of dimension $m$, it holds:
\begin{equation}\label{per_bound}
\text{per }(A) \leq \prod_{j=1}^m \left(r_j ! \right)^{1/r_j} ,
\end{equation}
where $r_j$ is the sum of the entries in the $j$-th column (or the $j$-th row).
\end{proposition}
This leads to the following result.
\begin{theorem}
For $d\geq 5$ the B\'ezout bound is strictly larger than the m-B\'ezout bound given by Equation~(\ref{mbe}) for any fixed $K_d$.
Given a fixed $d$, the new asymptotic upper bound derived from the Br\`egman-Minc inequality is
$$
O\left( \left(2\cdot \displaystyle \frac{\sqrt{(2d) !}}{d !} \right)^n \right) .
$$
\end{theorem}
\begin{proof}
In this proof $Be_d(n)$ and $mB_d(n)$ denote the B\'ezout and the maximal m-B\'ezout bound of minimally rigid graphs in $\mathbb{C}^d$ with $n$ vertices respectively.
Since the number of edge equations for minimally rigid graphs with $n$ vertices is $n\cdot d- d^2$, the B\'ezout bound is
$$
Be_d(n) = 2^{n\cdot d- d^2}.
$$
The sum of columns for the permanent that computes the m-B\'ezout bound is $r_j = d$ for the edges that include one fixed vertex and one non-fixed vertex and $r_j = 2d$ for these that include two non-fixed vertices.
We denote these sets of edges $E_{f.}$ and $E_{n.f.}$ respectively.
Applying the Br\`egman-Minc bound and Equation~(\ref{mbe})
we get
\begin{equation} \label{mbper}
mB_d(n)\leq \left(\frac{2}{d!}\right)^{n-d} \cdot \prod_{i=1}^{E_{f.}} (d!)^{1/d} \prod^{E_{n.f.}} (2d!)^{1/2d}\leq \left( 2\cdot \displaystyle \frac{\sqrt{(2d) !}}{d !} \right)^{n-d} .
\end{equation}
Combining these bounds we get a sufficient condition for $Be_d(n) > mB_d(n)$:
\begin{align*}
2^{n\cdot d- d^2} > \left( 2\cdot \displaystyle \frac{\sqrt{(2d) !}}{d !} \right)^{n-d}
\Leftrightarrow\, 2^{2d-2} \cdot (d !)^2 > (2d) !
\end{align*}
Robbins' bound on Stirling's approximation \cite{Robbins} yields the following:
$$
\sqrt{2\pi} \cdot d^{d+1/2} \cdot e^{-d} \cdot e^{R_-}< d! < \sqrt{2\pi} \cdot d^{d+1/2} \cdot e^{-d} \cdot e^{R_+},
$$
where $R_+= \displaystyle \frac{1}{12d}$ and $R_-= \displaystyle \frac{1}{12d+1} $. We now derive the following inequalities:
\begin{align*}
2^{2d-2} \cdot (d !)^2 \quad & > \; 2^{2d-2} \cdot 2\pi \cdot d^{2d+1} \cdot e^{-2d} \cdot e^{2R_-} & \\
& > \sqrt{2\pi} \cdot 2^{2d+1/2} \cdot d^{2d+1/2} \cdot e^{-2d} \cdot e^{R_+/2} & > \; 2d !
\end{align*}
that lead to a sufficient condition for $Be_d(n) >mB_d(n)$ to hold:
\begin{equation} \label{cond}
\sqrt{d} > \frac{4}{\sqrt{\pi}} \cdot e^{R_+/2-2R_-} ,
\end{equation}
which is true for every integer $d\geq 5$.
Additionally, inequality (\ref{mbper}) leads directly to the asymptotic bound
$$
mB_d(n) \in O\left( \left(2\cdot \displaystyle \frac{\sqrt{(2d)!}}{d !} \right)^n \right)
$$
for any given $d$.
\end{proof}
\begin{table}[htp]
\caption{Comparing B\'ezout's and permanent's asymptotic bounds (last two rows) for increasing $d$ (top row),}
\begin{tabular}{|c|c|c||c|c|c|c|c|c||c|}
\hline
2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 &30\\
\hline\hline
$4^n$ & $8^n$ & $16^n$ & $32^n$ & $64^n$ & $128^n$ & $256^n$ & $512^n$ & $1024^n$ & $(1.07\cdot 10^9)^{n}$ \\
\hline
$4.9^n$ & $8.9^n$ & $16.7^n$ & $31.7^n$ & $60.8^n$ & $117.2^n$ & $226.9^n$ & $441^n$ & $860^n$ & $(6.88\cdot 10^8)^n$\\
\hline
\end{tabular}
\end{table}
\subsection{Runtimes}
The computation of the m-B{\'e}zout bounds using our combinatorial algorithm up to a fixed $K_d$ is much faster than the computation of mixed volume and complex embeddings.
In order to compute the mixed volume we used \texttt{phcpy} in \texttt{SageMath} \cite{phcpy} and we computed complex solutions of the sphere equations using \texttt{phcpy} and and \texttt{MonodromySolver} \cite{monodromySolver}.
Let us notice that, \texttt{MonodromySolver} seems to be faster than mixed volume software we used in the case of Geiringer graphs.
We also compared our runtimes with the combinatorial algorithm that counts the exact number of complex embeddings in $\mathbb{C}^2$ \cite{Joseph_lam}.
\begin{table*}[htp!] \hspace*{-6mm}
\caption{Runtimes of different algorithms on graphs with maximal $c_d(G)$ up to $n=11$ and up to $n=10$ for Laman and Geiringer graphs, resp.
We compute $c_2(G)$ by \cite{Joseph_lam} and $c_3(G)$ by \texttt{phcpy} \cite{sourceCode} (fails to find all solutions for $n> 11$).
Also runtimes for computing $c_2(G)$, $c_3(G)$ by \texttt{MonodromySolver}.
We compute MV by \texttt{phcpy}, m-B\'ezout by \texttt{Maple}'s permanent and our \texttt{Python} code \cite{code_mBezout}.
Computation of the m-Bézout and MV is up to a fixed $K_d$ (edges or triangles).\vspace*{2mm}}
\centering
\begin{tabular}{|c||c|c|c|c|c|}
\hline
& \multicolumn{5}{c|}{\bf Laman graphs} \\
\hline
$n$ & \parbox[t]{1cm}{combin.\\ $c_2(G)$} & \parbox[t]{1.45cm}{\texttt{Monodromy} \\ \texttt{Solver}} & \parbox[t]{8mm}{\texttt{phcpy}\\ MV} & \parbox[t]{7mm}{\texttt{Maple}'s \\ perm.} & \parbox[t]{14mm}{m-B{\'e}zout \\ \texttt{Python}}\\
\hline\hline
6 & 0.0096s & 0.2334s & 0.0024s & 0.0003s & 0.0009s \\
\hline
7 & 0.0153s & 0.566s & 0.006s & 0.00045s & 0.0012s \\
\hline
8 & 0.0276s & 1.373s & 0.0122s & 0.00065 & 0.002s \\
\hline
9 & 0.066s& 4.934s & 0.0217s & 0.0018s & 0.0032s \\
\hline
10 & 0.176s & 12.78s & 0.043s & 0.0053s & 0.0045s \\
\hline
11 & 0.558s & 46.523s & 0.17s & 0.0077s & 0.0074s \\
\hline
12 & 6.36s & 2m47s & 0.39s & 0.049s & 0.013s \\
\hline
18 & 17h 5m & - & 1h 34m & 24s & 0.115s \\
\hline
\end{tabular}
\vspace*{0.3 cm}
\begin{tabular}{|c||c|c|c|c|c|}
\hline
& \multicolumn{5}{c|}{ \bf Geiringer graphs} \\
\hline
$n$ & \parbox[t]{8mm}{\texttt{phcpy}\\ solver} &
\parbox[t]{1.45cm}{\texttt{Monodromy} \\ \texttt{Solver}} &
\parbox[t]{8mm}{\texttt{phcpy}\\ MV} & \parbox[t]{7mm}{\texttt{Maple}'s\\ perm.} & \parbox[t]{14mm}{m~-~B{\'e}zout \\ \texttt{Python}} \\
\hline\hline
6 & 0.652s & 0.141s & 0.00945s & 0.0002s & 0.0097s \\
\hline
7 & 3.01s & 0.584s & 0.041s & 0.001s & 0.00165s \\
\hline
8 & 20.1s & 2.297s & 0.425s & 0.0025s & 0.00266s \\
\hline
9 & 2m 33s & 14.97s & 3.42s & 0.0075s & 0.006s \\
\hline
10 & 16m 1s & 1m23s & 1m 12s & 0.08s & 0.0105s \\
\hline
11 & 2h 14m & 9m22s & 27m31s & 0.49s & 0.024s \\
\hline
12 & - & 1h22m & $>6$days & 0.96s & 0.06s\\
\hline
\end{tabular}
\end{table*}
We will try to give some indicative cases for which we compared the runtimes.
For example, computing the mixed volume of the spherical embeddings up to one fixed edge for the maximal 12-vertex Laman graph for a given fixed $K_2$ takes around 390ms, while our algorithm for the m-B\'{e}zout bound required 13ms.
If we wanted to compute mixed volumes up to all fixed $K_2$ we needed 8.6s, while the m-B{\'e}zout computation took 270ms.
The runtime for the combinatorial algorithm that computes the number of complex embeddings is 6.363s for the same graph.
For larger graphs i.e.\ 18-vertex graphs, the combinatorial algorithm may take $\sim 17$h to compute the number of complex embeddings in $\mathbb{C}^2$.
We tested a 18-vertex graph that did not require more than 0.12s to compute one m-B{\'e}zout bound and 4s to compute m-B{\'e}zout bounds up to all choices of fixed edges.
In dimension 3 our model was the Icosahedron graph, which has~12 vertices.
The computation of the mixed volume took more than 6~days in this case, while our algorithm needed 60ms to give exactly the same result (54,272).
\texttt{MonodromySolver} could track all $ 54,272$ solutions in $\sim 1.3$ hour, while Gr\"obner basis computations failed multiple times to terminate.
Computing the permanent required more time compared to our algorithm.
For the Icosahedron the fastest computation could be done using \texttt{Maple}'s implementation in \texttt{LinearAlgebra} package.
It took $\sim 0.96$s to compute the permanent up to a given fixed triangle with this one.
On the other hand the implementations in \texttt{Python} and \texttt{Sage} took much more time for the same graph ($\sim 8$m and $\sim10$m respectively).
This seems reasonable since the combinatorial algorithm has to check at most $2^{m}$ cases, while according to \cite{emvid} the complexity to compute the permanent using Ryser's formula is in the order of $m^2 \cdot 2^{m}$.
\section{On the exactness of m-B\'ezout bounds}\label{sec:exact}
In this section we examine the exactness of m-B\'ezout bounds.
First, we use already published results (in the cases of $\mathbb{C}^2$ and $\mathbb{C}^3$) \cite{GKT17,belt} and our own computations (in $S^2$) to compare m-B\'ezout bounds with the mixed volumes and the actual number of embeddings.
Then, we present a general method to decide if the m-B\'ezout bound of a minimally rigid graph is tight or not based on Bernstein's second theorem on mixed volumes \cite{Bernshtein1975} without directly computing the embeddings.
We consider that the latter may be a first step to establish the existence of a particular class of graphs with tight m-B\'ezout bounds.
\subsection{Experimental results}\label{sec:res}
We compared m-B{\'e}zout bounds with the number of embeddings and mixed volumes using existing results \cite{GKT17,etv,belt} for the embeddings in $\mathbb{C}^2$ and $\mathbb{C}^3$.
We also computed the complex solutions of the equations that count embeddings on $S^2$ for all Laman graphs up to 8 vertices and a selection of graphs with up to 12 vertices that have a large number of embeddings.
We remind that in general the m-B\'ezout bound is not unique up to all choices of a fixed $K_d$.
It is natural to consider the minimal m-B\'ezout bound as the optimal upper bound of the embeddings for a given graph.
Let us notice that we checked if the m-B\'ezout is minimized when the fixed $K_d$ has a maximal sum of vertex degrees or when the vertex with the maximum degree belongs to the fixed $K_d$.
There are counter-examples for both of these hypotheseis.
\paragraph{Mixed volume and m-B{\'e}zout bound.}
In all cases we checked in $\mathbb{C}^3$ and $S^2$, the m-B{\'e}zout bound up to a fixed $K_d$ is exactly the same as the mixed volume up to the same fixed points.
There are some cases in $\mathbb{C}^2$ for which the m-B{\'e}zout bound is bigger than the mixed volume for certain choices of $K_2$.
We shall notice that these cases do not correspond to the minimal m-B{\'e}zout bound for the given graph (thus the minimum m-B{\'e}zout and the minimum mixed volume are the same for these graphs).
\paragraph{Spatial embeddings and the m-B{\'e}zout bound.}
As shown in~\cite{belt}, there are many cases for which the bounds of the sphere equations are larger than the actual number of complex embeddings.
Nevertheless, we observed that for all planar graphs up to $n=11$ the number of complex embeddings is exactly the same as the mixed volume bound and therefore the m-B{\'e}zout bound, while in the non-planar case the bounds are generally not tight.
What is also interesting is that the m-B{\'e}zout bound is invariant for all choices of fixed triangles in the case of planar Geiringer graphs.
\begin{table*}[!htb] \begin{center}
\caption{Mixed volumes, complex embedding numbers, and m-B{\'e}zout bounds for embeddings of Laman graphs in $\mathbb{C}^2$ and $S^2$.
These graphs have the maximal number of embeddings in $\mathbb{C}^2$.
The 12-vertex maximal Laman graph is the first non-planar in this category. \vspace*{1.5mm}}
\begin{tabular}{|c||c|c||c|c||c|} \hline
$\bm{n}$ & $\bm{\text{\textbf{MV}}_{2d}}$ & $\bm{c_2(G)}$ & $\bm{\text{\textbf{MV}}_{S^2}}$ & $\bm{c_{S^2}(G)}$ & \textbf{mB{\'e}zout} \\
\hline\hline
6 & 32 & 24 & 32 & 32 & 32 \\
\hline
7 & 64 & 56 & 64 & 64 & 64 \\
\hline
8 & 192 & 136 & 192 & 192& 192 \\
\hline
9 & 512 & 344 & 512 & 512 & 512 \\
\hline
10 & 1536 & 880 & 1536 & 1536 & 1536 \\
\hline
11 & 4096 & 2288 & 4096 & 4096 & 4096 \\
\hline
12 & 15630 & 6180 & 15630 & \textcolor{red}{\em 8704} & 15630 \\
\hline
\end{tabular} \end{center} \end{table*}
\paragraph{Embeddings of Laman graphs and the m-B{\'e}zout bound.}
For Laman graphs, the m-B{\'e}zout bound diverges from the number of actual embeddings in $\mathbb{C}^2$ more than in the case of Geiringer graphs.
That happens both for planar and non-planar graphs.
On the other hand the number of spherical embeddings of planar Laman graphs coincides with the minimum m-B{\'e}zout bound for a vast majority of cases (all planar graphs up to 6 vertices, 64/65 7-vertex planar graphs and 496/509 8-vertex planar graphs).
Notice that the m-B{\'e}zout bounds for different choices of the fixed edge are, in general, different for planar Laman graphs.
\subsection{Using Bernstein's second theorem}
Our computations indicate that the m-B{\'e}zout bound is tight for almost all planar Laman graphs in $S^2$ and all planar Geiringer graphs.
Therefore, we decided to apply Bernstein's second theorem to establish a method that determines whether this bound is exact.
We believe that a generalization of this method may show whether the experimental results provably hold for certain classes of graphs.
In this subsection for reasons of simplicity, in all examples we will use the variables $x_i,y_i,s_i$ and $x_i,y_i,z_i$ for $\mathbb{C}^2$ and $S^2$ respectively (instead of $x_{i,1}, \cdots x_{i,d},s_i$, used earlier).
Bernstein has given discriminant conditions such that the BKK bound (Theorem~\ref{thm:BKK}) for a given algebraic system may be tight.
In order to state Bernstein's conditions we first need the following definition.
\begin{definition}[Initial form]
Let $P$ be a polytope in ${\mathbb R}^m$, $w$ be a vector in ${\mathbb R}^m$ and $f = \sum\limits_{\alpha \in \mathcal{A}} c_{\alpha} {\bf x}^{\alpha}$ be a polynomial in $\mathbb{C}[x_1,x_2, \dots , x_m ]$.
Let also $P^{w}= \{v \in P\, | \,\langle v,w \rangle \leq \langle u,w \rangle, \, \forall u \in P \} $ be the subset of $P$ that minimizes the inner product with $w$.
The initial form of $f$ with respect to $w$ is defined as
$$
f^w= \sum \limits_{\beta \in P^w \cap \mathcal{A}} c_{\beta} \mathbf{x}^{\beta}.
$$
\end{definition}
The initial form $f^w$ contains precisely the monomials whose exponent vector minimizes the inner product with $w$ and excluding the others.
Clearly $P^w$ is a face of $P$ and $w$ is an inner normal to face $P^w$.
Hence, the algebraic system comprised of initial forms for a face normal $w$ shall be called \textit{face system}.
The necessary and sufficient condition of BKK exactness is stated below. Recall that the Minkowski sum of sets is the set of all sums between elements of the first and elements of the second set. The Minkowski sum of convex polytopes is a convex polytope containing the vector sums of all points in the two summand polytopes.
\begin{theorem}[Bernstein's second theorem \cite{Bernshtein1975}]\label{thm:Ber2}
Let $ (f_i)_{1 \leq i \leq m }$ be a system of equations in $\mathbb{C}[x_1,x_2, \dots , x_m ]$ and let
$$
P=\sum_{i=1}^m NP(f_i)
$$
be the Minkowski sum of their Newton Polytopes.
The number of solutions of $ (f_i)_{1 \leq i \leq m } $ in $(\mathbb{C}^*)^m$
equals exactly its mixed volume (counted with multiplicities) {if and only if}, for all $w \in {\mathbb R}^m$, such that $w$ is a face normal of $P$, the system of equations $(f^w_i)_{1 \leq i \leq m }$ has no solutions in $(\mathbb{C}^*)^m$.
\end{theorem}
Let us note that although there is an infinite number of vectors that may appear as inner normals, Bernstein's condition can be verified choosing only one inner normal vector for every different face of $P$.
The results in Section~\ref{sec:res} motivated us to examine these conditions closely in order to determine when the m-B{\'e}zout bound is exact.
The first step is to use Newton polytopes whose mixed volume equals to the m-B{\'e}zout bound (see for example \cite{SomWam}) since they are simpler than the Newton polytopes of the sphere equations.
\begin{definition}
We set $e_i=(0,0,\dots, \underset{\text{i-th position}}{1}, \dots,0 )$ and let $T_k^u$ be the simplex defined as the convex hull of the set
$$
\{\mathbf{0}, e_{k\cdot (u-1)+1},e_{k\cdot (u-1)+2}, \dots, e_{k\cdot (u-1)+k} \} ,
$$
where $\mathbf{0}=(0,0,\dots,0)$ is the origin.
Let $X_u$ be the $u$-th set of variables under a partition of all variables, with set cardinality $d_u=|X_u|$.
Then $T_{d_u}^u$ is the simplex that corresponds to the variables of this set.
The \emph{m-B{\'e}zout Polytope} of a polynomial, with respect to a partition of the variables, is the Minkowski sum of the $T_{d_u}^u$ for all $u$, such that each simplex is scaled by the degree of the polynomial in $X_u$.
\end{definition}
For a multihomogeneous system, simplices $T^u_{d_u}$ belong to complementary subspaces.
Then, each m-B{\'e}zout Polytope is the Newton polytope of the respective equation.
For general systems, our procedure amounts to finding the smallest polytopes that contain the system's Newton polytopes and can be written as Minkowski sum of simplices lying in the complementary subspaces specified by the variable partition.
In the case of rigid graphs in $\mathbb{C}^d$, every set of variables has $d+1$ elements.
Thus, the m-B{\'e}zout Polytope of the magnitude equations for a vertex $u$ is $2 \cdot T_{d+1}^u$, while the m-B{\'e}zout Polytope of the equation for edge $(u,v)$ is $T_{d+1}^u+T_{d+1}^v$.
This implies that the Minkowski sum of the m-B{\'e}zout Polytopes for the sphere equations of a minimally rigid graph $G=(V,E)$ is exactly
$$
B_G= \sum\limits_{u \in V'} ( \deg(u) +2) \cdot T_{d+1}^u ,
$$
where $\deg(u)$ is the degree of vertex $u$ in the graph and $V'$ the set of non-fixed edges.
In general, it is hard to compute the Minkowski sum of polytopes in high dimension.
But in the case of the m-B{\'e}zout Polytopes the following theorem describes the facet normals of $B_G$.
\begin{theorem}
Let $G=(V,E)$ be a minimally rigid graph in $\mathbb{C}^d$ and $B_G$ be the Minkowski sum defined above.
The set of the inner normal vectors of the facets of $B_G$ are exactly
\begin{itemize}
\item all unit vectors $e_i$, and
\item the $n-d$ vectors of the form
$$
\delta_u= \sum\limits_{j=1}^{d+1} -e_{(d+1)\cdot(u-1)+j} =(0,0,\dots, -1,-1,\dots,-1,\dots, 0 ),
$$
where there are $d+1$ nonzero entries corresponding to the variables that belong to the $u$-th variable set.
\end{itemize}
\end{theorem}
\begin{proof}
Since each $T_{d+1}^u$ belongs to a complementary subspace, $B_G$ can be seen as the product of polytopes $\prod\limits_{u \in V'} ( \deg(u)+2) \cdot \Delta_{d+1}$, where $\Delta_{d+1}$ is the unit $(d+1)$-simplex \footnote{The idea of using the product of polytopes is derived by a proof for the mixed volumes corresponding to the weighted m-B\'ezout bound in \cite{Pinaki}}.
The inner normal vectors of the facets of $\Delta_{d+1}$ in ${\mathbb R}^{d+1}$ are the unit vectors $e_i$ and $\delta=\sum\limits_{j=1}^{d+1} -e_{j}$ in ${\mathbb R}^{d+1}$.
The theorem follows since the normal fan of a product of polytopes is the direct product of the normal fans of each polytope~\cite{Ziegler}.
\end{proof}
This theorem yields a method to find the H-representation of $B_G$, in other words the polytope is described as the intersection of linear halfspaces and the respective equations are given by the theorem.
In all cases where $MV=mBe$, the polytopes $B_G$ can be used instead of the Newton Polytopes of the equations.
The verification of Bernstein's second theorem requires a certificate for the existence of roots of face systems for every face of $B_G$, where faces range from vertices of dimension~0 to facets of codimension 1.
We propose a method that confirms or rejects Bernstein's condition checking a much smaller number of systems based on the form of facet normals.
For this, we shall distinguish three cases below.
The normal of a lower dimensional face can be expressed as the vector sum of facet normals, whose cardinality actually equals the face codimension.
This means that we need to verify normals distinguished in the following three cases:
\begin{enumerate}
\item vector sums of one or more ``coordinate" normals $e_i$'s,
\item vector sums of one or more ``non-coordinate" normals $\delta_u$'s,
\item ``mixed" vector sums containing both $e_i$'s and $\delta_u$'s.
\end{enumerate}
Notice that since there are $(d+2)\cdot (n-d)$ different normals, in order to check all resulting face systems, $2^{(d+2)\cdot (n-d)}$ computations are required.
We now examine each of these three cases separately, in order to exclude a very significant fraction of these computations.
\paragraph{First case (coordinate normals).}
Let $F=(f_i )_{1\leq i \leq m}$ be the system of the sphere equations, let the initial forms be $f_i^{e}$ for some normal $e$, and let $F^e$ be the resulting face system.
We will deal with the coordinate normals case starting with an example.
\begin{example}\label{ex:e_i}
We present the equations of face system $F^{e_1}$ in $\mathbb{C}^2$.
Normal $e_1$ corresponds to variable $x_1$.
This means that the inner products with the exponent vectors of the monomials in the magnitude equation $f_1 = x_1^2+y_1^2-s_1$ are $2,0,0$.
Thus, $f_1^{e_1} = y_1^2-s_1$, excluding the monomial $x_1^2$.
In the case of the edge equation $f_{(1,2)}= s_1+s_2-2 (x_1x_2+y_1y_2)+\lambda_{1,2}^2$, for a generic edge length $\lambda_{1,2}$, the inner products are $0,0,1,0,0$.
It follows that $f_{(1,2)}^{e_1}= s_1+s_2-2 y_1y_2+\lambda_{1,2}^2$.
If the degree of $x_1$ in an equation $f_i$ is zero, then $f_i^{e_1}=f_i$, since the inner product of all the exponent vectors with $e_i$ is zero.
\end{example}
This example shows that since all $x_1$ monomials are removed, $F^{e_1}$ is an over-constrained system that has the same number of equations as $F$, but a smaller number of variables.
The same holds obviously for every $F^{e_i}$, while for $e=\sum\limits_{i \in I} e_i$ (where $I$ is an index set) the initial forms in $F^{e}$ are obtained after removing all monomials that include one or more of the variables corresponding to the $e_i$'s of the sum.
In other words, the initial forms in system $F^{e}$ can be obtained by evaluating to zero all the variables indexed by the set $I$.
\begin{lemma}
Let $e$ be a sum of $e_i$ normals as described above.
Now, $F$ does not verify Bernstein's condition in the coordinate normals case (and has an inexact BKK bound) due to system $F^e$ having a toric root $r'$, only if $F$ has a root $r$ with zero coordinate for at least one of the variables in $I$, such that the projection of $r$ to the coordinates $j\not\in I$ equals $r'$.
\end{lemma}
We can now exclude the case of sums of coordinate
normals from our examination, since it shall not generically occur, because the next lemma shows that $r$ has no zero coordinate.
\begin{lemma}\label{lem:e_i}
The set of solutions of the sphere equations for a rigid graph generically lies in $(\mathbb{C}^*)^{d\cdot n}$.
\end{lemma}
\begin{proof}
We indicate by $\text{C}(G,\bm{\lambda},\{p_1,p_2,\dots p_d\})$ the set of complex embeddings for a rigid graph $G$ up to an edge labeling $\bm{\lambda}$ and $d$ fixed points $\{p_1,p_2,\dots p_d\}$.
This set of embeddings is finite by definition.
If there is a zero coordinate in the solution set, there exists a vector $p' \in \mathbb{C}^d$, such that no zero coordinates belong to the zero set $\text{C}^*(G,\bm{\lambda},\{p_1+p',p_2+p',\dots p_d+p'\})$.
Since we want to verify Bernstein's condition for a generic number of complex embeddings of $G$, we can always use the second set of embeddings.
\end{proof}
This lemma excludes a total of $2^{(d+1)\cdot (n-d)}$ cases when verifying Bernstein's second theorem for a given algebraic system.
\paragraph{Second case (non-coordinate normals).}
In the second case, the inner product of exponent vectors with $\delta_u$ is minimized for all variables $X_u$ of a vertex $u$ with maximum degree.
Let us give again an example to explain this statement.
\begin{example}\label{ex:delta}
It is an example in $\mathbb{C}^2$ for face system $F^{\delta_1}$.
The inner products for the magnitude equation $f_1 = x_1^2+y_1^2-s_1$ and the edge equation $f_{(1,2)}= s_1+s_2-2\cdot(x_1x_2+ y_1y_2)+\lambda_{1,2}^2$, $\lambda_{1,2}$ being a generic edge length, are $-2,-2,-1$ and $-1,0,-1,-1,0$ respectively.
So, $f_1^{\delta_1} = x_1^2+y_1^2$ and $f_{(1,2)}^{\delta_1} = s_1-2\cdot(x_1x_2+ y_1y_2)$.
If the degree of a polynomial $f_i$ in the set of variables $X_u$ is zero, then $f_i^{\delta_u}=f_i$.
\end{example}
The number of equations of $F^{\delta_u}$ equals the number of variables.
Following Bernstein's proof on the discriminant conditions, we introduce a new
system by applying a suitable variable transformation from the initial variable vector $\mathbf{x}$ to a new variable vector $\mathbf{t}$ with same indexing.
This transformation uses an $m\times m$ full rank matrix $B$ such that every monomial $\mathbf{x}^{\alpha}$ is mapped to $\mathbf{t}^{B\cdot \alpha}$ (see \cite{Bernshtein1975,Cox2} for more details).
Furthermore, $|\det B| = 1$ so that the transformation preserves the mixed volume of $F$ \cite{Cox2}.
In our case, we construct matrix $B$ with the following properties:
\begin{eqnarray*}
B\cdot \delta_u^T & = & e_{(d+1)\cdot (u-1)+1}, \\
B\cdot e_{(d+1)\cdot (u-1)+j}^T & = & e_{(d+1)\cdot (u-1)+j} ,\ \forall j \in \{ 2,\dots, d+1\} , \\
\text{det}(B) & = & \pm 1 .
\end{eqnarray*}
The intuition behind these choices is given in Lemma~\ref{lem:Lcase2} and its proof.
This yields the following map from variables $\mathbf{x}$ to new variables $\mathbf{t}$:
\begin{align}\label{eq:deltamap}
x_{u,1} \mapsto \frac{1}{t_{u,1}}, & \,\ x_{u,2} \mapsto \frac{t_{u,2}}{t_{u,1}}, \,\ \cdots ,\,\ s_{u} \mapsto \frac{t_{u,d+1}}{t_{u,1}} .
\end{align}
We will refer to the set of $x_{u,1}$'s as the $\delta$-\textit{variables} of $F$, since the image of their exponent vectors is the set of $\delta_u$'s, while the exponent vectors for the other variables remain same.
This transformation maps system $F(\mathbf{x})$ to a new system $\widehat{F}(\mathbf{t})$ of Laurent polynomials in the new variables.
In the case of $\mathbb{C}^2$, the sphere equations are mapped as follows:
\begin{align*}
\widehat{f}_u= \frac{1}{t_{u,1}^2}+\frac{t_{u,2}^2}{t_{u,1}^2}-\frac{t_{u,3}}{t_{u,1}}, \,\,\ & \text{(magnitude equations)} \\
\widehat{f}_{(u,v)}= \frac{t_{u,3}}{t_{u,1}}+\frac{t_{v,3}}{t_{v,1}}-2\cdot \displaystyle\left(\frac{1}{t_{u,1}t_{v,1}}+\frac{t_{u,2}t_{v,2}}{t_{u,1}t_{v,1}}\right)+ \lambda_{u,v}^2 \,\,\ & \text{(edge equations) .}
\end{align*}
The degree $\alpha(\widehat{f},t_{u,1})$ of a polynomial $\widehat{f}$ with respect to the variable $t_{u,1}$ will be either zero or negative.
Let us now multiply every polynomial in $\widehat{F}(\mathbf{t})$ with each one of the monomials $t_{u,1}^{-\alpha(\widehat{f},t_{u,1})}$.
These monomials are defined as the least common multiple of the denominators in the Laurent polynomials $\widehat{f}$, yielding the following system $\widetilde{F}(\mathbf{t})$:
\begin{equation*}\begin{split}
\widetilde{f}_u = 1+t_{u,2}^2-t_{u,1}t_{u,3}, \,\,\ \text{(magnitude equations)} \\
\widetilde{f}_{(u,v)}= t_{v,1}t_{u,3}+t_{u,1}t_{v,3}-2\cdot \displaystyle\left(1+t_{u,2}t_{v,2}\right)+\lambda_{u,v}^2\cdot t_{u,1}t_{v,1}. \,\,\ \text{(edge equations)}
\end{split}
\end{equation*}
This transformation yields the necessary conditions to verify if the face systems of the non-coordinate normals have solutions in $(\mathbb{C}^*)^m$.
We will refer to $t_{u,1}$'s as the set of $\delta$-variables of $\widetilde{F}(\mathbf{t})$, while the rest should be the $e$-variables.
Note that the transformation gives a well-constrained system, while zero evaluations of the $\delta$-variables shall result to an over-constrained system, that should have no solutions if the bound is exact.
\begin{lemma}\label{lem:Lcase2}
There exists a sum $\delta$ of different $\delta_u$ normals, such that face system $F^{\delta}$ has a solution in $(\mathbb{C}^*)^m$ if the algebraic system $\widetilde{F}(\mathbf{t})$, which is defined above, has a zero solution for $t_{u,1}$ for one or more vertices $u$.
\end{lemma}
\begin{proof}
Matrix $B$ is constructed to change the variables $x_{u,1}$ to variables $t_{u,1}$, for vertices $u$.
From the definitions of $\widehat{F}(\mathbf{t})$ and $\widetilde{F}(\mathbf{t})$, it follows that given a monomial $\mathbf{t}^{\bm{\beta}}$ in a polynomial $\widehat{f}$ of $\widehat{F}(\mathbf{t})$, the inner product $\langle \bm{\beta},B\cdot \delta_u^T \rangle$ is not minimized among other monomials in $\widehat{f}$ if and only if the degree of $t_{u,1}$ in the respective monomial of the transformed polynomial $\widetilde{f}$ is positive.
Thus, the existence of toric solutions for the face system $F^{\delta_u}$ is equivalent to existence of toric solutions for the zero evaluation $\widetilde{F}(\mathbf{t},t_{u,1}=0)$.
So, if $\widetilde{F}(\mathbf{t},t_{u,1}=0)$ has a solution in $\mathbb{C}^m$ such that $\widetilde{F}(\mathbf{t},t_{u,1}=t_{{v_1},1}=\cdots=t_{{v_k},1}=0)$ has a solution in $(\mathbb{C}^*)^m$, then $F^{\delta_u+\delta_{v_1}+\cdots +\delta_{v_k}}$ has a solution in $(\mathbb{C}^*)^m$.
\end{proof}
The computational gain in this case is that, without the lemma, one would have checked every different combination of the $\delta_u$'s, namely a total of $2^{n-d}$ checks.
Now, it suffices to check only one zero evaluation for each of them, hence only $n-d$ checks.
\paragraph{Third case (mixed normals).}
The third case, that includes the sums of vectors $\delta_u$ and $e_i$ can be also treated with the transformation $\widetilde{F}(\mathbf{t})$ introduced above.
Since the minimization of the inner product is invariant for $d$ of the $d+1$ variables per vertex, the non-existence of zero solutions in $\widetilde{F}(\mathbf{t})$ implies that no $F^w$ has solutions in $(\mathbb{C}^*)^m$ for all vectors $w$ that are sums of vectors $\delta_u$ with those vectors $e_i$ for which the equality $B \cdot e_i^T=e_i$ holds.
In order to proceed we need the following lemma.
This shows that using $d$ of the $d+1$ $e_i$'s of a vertex suffices to verify if a face system of a mixed normal has solutions in $\mathbb{C}^m$.
\begin{lemma}\label{lem:alle}
Let us define a sum of normals
$$
-\delta_u =\sum\limits_{j=1}^{d+1} e_{(d+1)\cdot(u-1)+j} \in {\mathbb R}^{d+1} .
$$
For every $w \in {\mathbb R}^m$, such that $w$ is a sum of $-\delta_u$ and other normals outside the set\\
$\{\delta_u,e_{(d+1)\cdot(u-1)+1}, \dots, e_{(d+1)\cdot (u-1)+d+1}\}$ (hence in a subspace complementary to that of $-\delta_u$), the face system $F^{w}$ cannot have a solution in $(\mathbb{C}^*)^m$.
\end{lemma}
Note that $-\delta_u$ is the sum of $d+1$ normals in complementary subspaces.
\begin{proof}
We will treat the case of dimension $d=2$ for simplicity of notation; the proof generalizes to arbitrary $d$.
Without loss of generality, we consider $u=1$.
Then $w \in {\mathbb R}^{3}\times {\mathbb R}^{m-3} $ with $w=(-\delta_1,v)$ and $v \in {\mathbb R}^{m-3}$.
The inner products of $-\delta_1$ with the exponent vectors of the magnitude equation in ${\mathbb R}^2$ for the first coordinate are $2,2,1$, so $f_1^{-\delta_1}=-s$.
It is obvious that no $w$ which is a sum of $-\delta_1$ and normals not belonging to the set $\{\delta_1,e_{1}, e_{2}, e_3 \}$ defines a face system with no solutions in $(\mathbb{C}^*)^m$.
Similarly, the inner products of $-\delta_1$ with the exponent vectors of the magnitude equation $f_1=x_1^2+y_1^2+z_1^2-1$ on $S^2$ are $2,2,2,0$, yielding $f_1^{-\delta_1}=-1$ which has no solutions in $\mathbb{C}^m$.
\end{proof}
Lemma \ref{lem:alle} reveals that in order to verify the conditions of Bernstein's theorem, we can use the transformations $\widetilde{F}(\mathbf{t})$ for all choices of $d$ variables from every set $X_u$, since there is no need to check the cases that include the sum of all $e_i$ normals of a single vertex.
This result, combined with Lemmas~\ref{lem:e_i} \& \ref{lem:Lcase2} leads to the following corollary.
\begin{corollary}\label{cor:allconditions}
There is a vector $w \in {\mathbb R}^m$ such that the face system of the sphere equations $F^w$ has a toric root if and only if there is a choice of $\delta$-variables such that the transformed algebraic system $ \widetilde{F}(\bm{t})$ has a zero solution in $\mathbb{C}^m$ for at least one $\delta$-variable.
\end{corollary}
Since the first $d$ coordinate variables $x_{u,j}$ are symmetric (while $s_u$ variables are not), we can exploit these symmetries excluding some choices.
So, without loss of generality, we may keep $x_{1,1}$ as a $\delta$-variable from variable set $X_1$, and check all possible choices for $\delta$-variables from all other variable sets $X_u$, such that $u\ne 1$ and $u$ is not among the fixed vertices.
\paragraph{Summary of three cases.}
In general, if one selects to take into consideration all possible sums of facet normals, then $2^{(d+2)\cdot (n-d)}$ cases should be checked.
We have shown that the category of face systems defined by a sum of coordinate normals cannot have toric solutions, discarding $2^{(d+1)\cdot (n-d)}$ cases.
In the two other cases, the investigation of toric solutions can be combined using the $\widetilde{F}(\mathbf{t})$ transformation.
If a face system has toric solutions, then in the non-coordinate normals case some $\delta$-variables may have zero solutions, while in the mixed normals case both $\delta$-variables and $e$-variables may have zero solutions.
A naive approach to verify these two cases would result to $2^{(d+1)\cdot (n-d)} \cdot (2^{ (n-d)}-1)$ checks, but using Corollary~\ref{cor:allconditions} one needs to verify the zero evaluations of $\delta$-variables for all choices of $\delta$-variables.
The latter, can be further reduced from $d^{n-d}$ to $d^{n-d-1}$, due to the fact that the coordinate variables are symmetric.
Summarizing, when checking Bernstein's condition, for any of the $d^{n-d-1}$ choices of $\delta$-variable transformation that construct $\widetilde{F}(\mathbf{t})$, it suffices that $n-d$ zero evaluations should be applied for each of the $\delta$-variables.
\begin{theorem}\label{thm:allcases}
Bernstein's condition can be verified in the case of the sphere equations after checking a total of at most $(n-d)\cdot d^{n-d-1}$ face systems.
\end{theorem}
This discussion yields an effective algorithmic procedure (see Algorithm~\ref{alg:exact}) to verify whether the m-B\'ezout bound is exact.
Function \texttt{ConstructDeltaPoly} takes as input the system of the sphere equations $F$ and a list of $\delta$-variables to construct the polynomials $\widetilde{F}(\mathbf{t})$.
The central role is played by function \texttt{IsmBezoutOfGraphExact}, which verifies if the polynomials $\widetilde{F}$ have zero solutions for the $\delta$-variables.
\begin{algorithm}[htp!]\label{alg:exact}
\caption{m-B\'ezout exactness for minimally rigid graphs}
\DontPrintSemicolon
\KwFn{ConstructDeltaPoly}\\
\KwInput{$F$ (sphere equations), $V'$ (non-fixed vertices), $L$ (variable indices mapped to $\delta$-variables from each partition set)}
\KwOutput{$\widetilde{F}$}
\tcc{Transformation to $\widehat{F}(\mathbf{t})$}
$changevars=\left(\bigcup\limits_{u \in V'} \{x_{u, L(u)} \mapsto \frac{1}{t_{u, L(u)}}\}\right)
\bigcup
\left(\bigcup\limits_{\substack{u \in V' \\ l \in \{1,\cdots, d+1\}\backslash \{L(u)\}}}\{x_{u,l} \mapsto \frac{t_{u,l}}{t_{u,L(u)}}\}\right)$\\
$\widehat{F}(\mathbf{t})=F(changevars)$\\
$\widetilde{F}(\mathbf{t})=\{\}$\\
\For{$\widehat{f} \in \widehat{F}(\mathbf{t})$}
{\tcc{$\alpha(\widehat{f},t_{u,L(u)})=$ (non-positive) degree of $\widehat{f}$ in variable $t_{u,L(u)}$}
$\widetilde{f}=\prod\limits_{u \in V'} t_{u,L(u)}^{-\alpha(\widehat{f},t_{u,L(u)})} \cdot \widehat{f}$\\
$\widetilde{F}(\mathbf{t})=\widetilde{F}(\mathbf{t}) \bigcup \{\widetilde{f}\}$
}
\Return($\widetilde{F}(\mathbf{t})$) \vspace*{3mm}\\
\DontPrintSemicolon
\KwFn{IsmBezoutOfGraphExact}\\
\KwInput{$F$ (sphere equations), $V'$ (non-fixed vertices), Conjecture (If \textbf{True}, only one choice of $\widetilde{F}$, else all choices of $\widetilde{F}$)}
\KwOutput{\textbf{True} (m-B\'ezout$=c_d(G)$) or \textbf{False} (m-B\'ezout$>c_d(G)$)}
\tcc{Verification of Bernstein's condition}
\If{Conjecture=\textbf{True}}{
\tcc{Transformation to $\widetilde{F}(\mathbf{t})$}
$L=[1,1,\dots ,1]$ \tcp*{$L$ length $=$ number of $u \in V'$}
$\widetilde{F}$=ConstructDeltaPoly(F,V',L)
\tcc{Check for zero solutions of $\delta$-variables (main computation)}
\For{$u \in V'$}
{ \tcc{Check if zero evaluation has solutions using ideal of transformed face system}
\If{$\widetilde{F}(\mathbf{t},\{t_{u,l_u}=0\})$ has a solution}
{ \Return(\textbf{False})
}
}
}
\ElseIf{Conjecture=\textbf{False}}
{
\For{all choices of 1 out of $d$ variables in every $X_u$, $u\neq 1$}
{
$L=[1,(l_u, u \in V'\backslash\{1\})]$ \tcp*{ always same choice for $X_1$}
$\widetilde{F}$=ConstructDeltaPoly(F,V',L)\\
\tcc{Check for zero solutions of $\delta$-variables (main computation).}
\For{$u \in V'$}
{ \If{$\widetilde{F}(\mathbf{t},\{t_{u,l_u}=0\})$ has a solution}
{ \Return(\textbf{False})
}
}
}
}
\Return(\textbf{True})
\end{algorithm}
Let us present two examples, further treated in the code found in~\cite{code_mBezout}.
\begin{example}\label{ex:desBer}
The first (and the simplest non-trivial) example we have treated is an application of our method to the equations that give the embeddings of Desargues graph (double prism) in $\mathbb{C}^2$ and $S^2$ (see Figure~\ref{fig:Des}).
It is known that its number of embeddings in ${\mathbb R}^2$ is 24 \cite{Borcea} and on $S^2$ it is 32 \cite{belt}, and they both coincide with the generic number of complex solutions of the associated algebraic system.
The m-B\'ezout bound for these systems is $32$, hence it is inexact in the $\mathbb{C}^2$ case.
This shall be explained by the fact that the sphere equations in $\mathbb{C}^2$ have face systems of non-coordinate normals with toric roots.
\begin{figure}[htp]\label{fig:Des}
\begin{center}
\Des
\caption{Desargues graph (double prism)}
\end{center}
\end{figure}
The system of the sphere equations (with vertices~1 and~2 fixed) is a $12 \times 12$ well-constrained system, but we can easily eliminate the linear equations obtaining an 8 by 8 well-constrained system.
Subsequently, we can also fix vertex~3 up to reflection about the edge $(1,2)$, obtaining finally a system of 6 polynomials in the variables $\{x_4,y_4,x_5,y_5,x_6,y_6 \}$.
If we apply the transformation of variables mentioned above, we can construct a system of polynomials in variables $\{t_{4,1},t_{4,2},t_{5,1},t_{5,2},t_{6,1},t_{6,2} \}$, such that evaluating $t_{4,1},t_{5,1}$ or $t_{6,1}$ to zero corresponds to the face systems of $\delta_4, \delta_5$ or $ \delta_6$ respectively.
This is one possible choice of $\delta$-variables to construct $\widetilde{F}(\mathbf{t})$.
Solving these 3 different systems for every $\delta_u$ with Gr\"obner basis in \texttt{Maple} we find the existence of solutions in $\mathbb{C}^2$, indicating that the number of complex solutions is strictly smaller than the m-B\'ezout bound.
In order to get nonzero solutions in $\mathbb{C}^2$, we need to evaluate to zero all $t_{4,1},t_{5,1},t_{6,1}$ variables, implying that the normal direction for which Bernstein's second theorem shows mixed volume to be inexact is $\left(-1,-1,-1,-1,-1,-1 \right)$.
This is a normal of a 3-dimensional face, where the face dimension is obtained as 6-3.
The normal equals the sum of 3 facet normals.
In the spherical case, no solutions exist, not only for the first choice of $\widetilde{F}(\mathbf{t})$, but also for all the other possible ones (see Algorithm~\ref{alg:exact}), suggesting that the bound is tight, so the number of spherical embeddings is $32$ and equals the m-B\'ezout bound.
\end{example}
\begin{example}\label{ex:jacBer}
The Jackson-Owen graph has the form of a cube with an additional edge adjacent to two antisymmetric vertices (see Figure~\ref{fig:Jac}).
This graph is the smallest known case that has fewer real than complex embeddings, in ${\mathbb R}^2$ and $\mathbb{C}^2$ respectively \cite{Jackson}.
The m-B\'ezout bound up to the fixed edge shown in the figure is $192$, while the number of embeddings in $\mathbb{C}^2$ is $90$.
This shall be explained by a face system of mixed normals that has toric roots.
\begin{figure}[htp]
\begin{center}
\Jackson
\caption{The Jackson-Owen graph \label{fig:Jac}}
\end{center}
\end{figure}
The system of the sphere equations is $18 \times 18$, reduced to $13 \times 13$ after linear elimination.
The set of variables is $\{x_3,y_3,x_4,y_4,x_5,y_5,x_6,y_6,x_7,y_7,x_8,y_8,s_8 \}$.
We apply the transformation of variables, resulting to a new system of polynomials in variables $\{t_{3,1},t_{3,2},t_{4,1},t_{4,2},t_{5,1},t_{5,2},t_{6,1},t_{6,2},t_{7,1},t_{7,2},t_{8,1},t_{8,2},t_{8,3} \}$.
As in the case of Desargues' graph, the evaluation of $t_{3,1},t_{4,1},t_{5,1},t_{6,1},t_{7,1},t_{8,1}$ to zero corresponds to the face systems of $\delta_3, \delta_4, \delta_5,\delta_6,\delta_7,\delta_8$ respectively.
We found a solution following zero evaluation of all $\delta$-variables of $\widetilde{F}(\mathbf{t})$ and $t_{8,3}$.
The normal direction is $\left(-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,0 \right)$ $\in{\mathbb R}^{13}$.
This is an example for the mixed normals case, the normal being a sum of one $e_i$ and all 6 $\delta_u$'s.
\end{example}
Another way to apply our method is the computation of a suitable resultant matrix for a given over-constrained system, which follows from evaluating some variable to zero.
It is obvious that if the system has any solutions, the rank of the resultant matrix with sufficiently generic entries is strictly smaller than its size, otherwise it has full rank, assuming we have a determinantal resultant matrix.
We have used \texttt{multires} module for \texttt{Maple} \cite{multires} to examine the existence of solutions, repeating the previous results.
However, these tools seem to be slower than other techniques which directly compute the number of embeddings.\\
In our experimental computations, we noticed that the existence of zero solutions of only one choice of $\widetilde{F}(\mathbf{t})$ (and not all $d^{n-d-1}$) suffice to verify Bernstein's conditions.
Therefore, we state the following conjecture:
\begin{conjecture}\label{con:delta}
The conditions of Bernstein's second theorem for the system of the sphere equations $F$ are satisfied if and only if the system $\widetilde{F}(\mathbf{t})$ has solutions for every zero evaluation of the $\delta$-variables.
\end{conjecture}
If Conjecture \ref{con:delta} holds, then one only needs to check $n-d$ face systems that correspond to the zero evaluations of $\widetilde{F}(\mathbf{t})$ for each one of the $\delta$-variables, instead of the $(n-d)\cdot d^{n-d-1}$ face systems indicated in Theorem~\ref{thm:allcases}.
Algorithm~\ref{alg:exact} includes the option to consider the Conjecture~\ref{con:delta} to be either True or False.
The first option takes into consideration only one choice of $\delta$-variables, while in the second one all different choices of $\delta$-variables are checked, as in Theorem~\ref{thm:allcases}.
\section{Conclusion and open questions}
We presented new methods to compute efficiently the m-B\'{e}zout bound of the complex embedings of minimally rigid graphs using graph orientations and matrix permanents.
Exploiting existing bounds on planar graph orientations and matrix permanents, we improved the asymptotic upper bounds of the embeddings for planar graphs in dimension 3 and for all graphs in high dimension.
We also compared our experimental results with existing ones indicating that some classes of graphs have tight m-B\'{e}zout bounds.
Finally, we applied Bernstein's second theorem in the case of the m-B\'{e}zout bound for rigid graphs.
Our findings in this topic can be generalized for every class of polynomial systems that have no zero solutions.
Several open questions arise from our results.
First of all, we would like to improve asymptotic upper bounds, especially for $d=2$ and $d=3$, exploiting the sparseness of minimally rigid graphs, closing the gap between upper bounds and experimental results.
We would like to examine aspects of both extremal graph theory (for graph orientations) and permanent bounds of matrices with specific structure for this reason.
Besides that, finding the minimal m-B\'ezout bound requires the computation of bounds up to all possible choices for a fixed $K_d$.
Thus, it would be convenient to find a method to select the $K_d$ that attains the minimum without computing its bound.
Regarding the exactness of the m-B\'ezout bound and the application of Bernstein's second theorem, the first priority would be a possible proof (or refutation) of Conjecture \ref{con:delta}.
Another issue is to optimize this method using the appropriate tools.
A first idea is to construct resultant matrices that exploit the multihomogeneous structure (see for example \cite{EmMantz,DE2003}).
The rank of the matrix could indicate which zero evaluations have solutions for our systems.
Last but not least, we would like to investigate a possible relation between planarity and tight upper bounds, especially in the case of Geiringer graphs.
\paragraph{Acknowledgements}
IZE and JS are partially supported, and EB is fully supported by project ARCADES which has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement No~675789.
IZE and EB are members of team AROMATH, joint between INRIA Sophia-Antipolis and NKUA.
We thank Georg Grasegger for providing us with runtimes of the combinatorial algorithm that calculates $c_d(G)$ and for explaining to EB the counting of embeddings for triangle-free Geiringer graphs.
We thank Jan Verschelde for help with applying our method to the Desargues graph, and Timothy Duff for giving us an example on the use of \texttt{MonodromySolver} to find the embeddings of the Icosahedron.
| b6789e769ec18ab11bf1a00bd9e97fb226d866a9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Plasmonic lattices form a large class of optical metasurfaces. In recent years they gained great attention due to a combination of a simple for qualitative analysis design and unique optical properties. Hybrid photonic structures with inclusions of such lattices confine light and enhance light-matter interaction due to plasmonic resonances, whereas photonic counterpart provides high-quality resonances. These prerequisites lead to an application of such structures for sensing \cite{Shen2013,rodriguez2011,lodewijks2012,sterl2020design}, light emission enhancement and routing \cite{kolkowski2019lattice,vaskin2019light,rodriguez2012, Guo2015,vecchi2009,ramezani2016modified}, lasing \cite{schokker2017systematic,zhou2013,wang2017,schokker2016lasing}, holography \cite{zheng2015metasurface,ye2016spin,wei2017broadband} and many other purposes.
Plasmonic lattices in homogeneous environments are well-known by specific non-typical lattice plasmon resonances (LPRs). LPRs are the modes associated with collective oscillations of individual nanoparticles coupled to each other by free photons propagating along the lattice. These resonances are unique since they potentially allow us to achieve very narrow lines in a plasmonic structure without the implementation of a photonic waveguide or any other high-$Q$ dielectric resonator. Such resonances have been thoroughly investigated both theoretically and experimentally \cite{guo2017,Augie2008,Rajeeva2018,kravets2008,kravets2018plasmonic,wang2017rich,rodriguez2011,rodriguez2012,Augie2008,Zhao2003,gerasimov2019engineering,zakomirnyi2017refractory,Guo2015,zhou2013}, powerful computational methods that can simplify their consideration have been developed \cite{chen2017general,babicheva2018metasurfaces,babicheva2019analytical}. However, in most studies, only simple lattices have been considered \cite{Zhao2003,rodriguez2012,Garcia2007, gerasimov2019engineering,zakomirnyi2017refractory,rodriguez2011,guo2017,Rajeeva2018,kravets2018plasmonic,wang2017rich,Guo2015,zhou2013}. Recently, lattices with several particles in a cell attracted attention and were considered in a few studies \cite{baur2018, fradkin2020nanoparticle,liu2018light,kolkowski2019lattice,Humphrey2014}. Nevertheless, there are still many promising designs of plasmonic lattices, particularly in homogeneous media, that are of great interest but have not been considered yet.
In this paper, we consider a stack of two plasmonic lattices in a homogeneous environment. We study symmetries of their modes, observe Fabry-Perot resonances, show and explain the coupling of its lattice plasmon resonances and demonstrate that in some cases one of the hybridized LPRs arises exactly on the Rayleigh anomaly~(RA) \cite{maystre2012}. Such mode is narrower than a typical lattice resonance and is observed in a wide range of distances between lattices. Our results show that this effect can be efficiently used to set the resonance energy precisely since it is determined only by the structure period, which is a quantity most stably reproduced in an experiment. Although we do not provide a comprehensive classification of two-lattice structures and their modes, the presented techniques can be further applied to explore any similar structures.
\section{Results and discussion}
Physical phenomena demonstrated in this article do not require a specific choice of materials and particles, so we consider simple square lattices of silver nanodiscs embedded in silica ($\varepsilon_{\mathrm{SiO}_2}=2.1$). Nanodiscs have radii of 30~nm, 20~nm height, and are described by Johnson\&Christy optical constants \cite{JohnsonChristy1972}. Period of the structure, $a$, and thickness, $H$, (see Fig.~\ref{fig:1}) are varied along the study.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{14.png}
\caption{Schematic of the stack of two plasmonic lattices. Each lattice is square and consists of silver nanodiscs. The whole structure is embedded into infinite layer of silica.}
\label{fig:1}
\end{figure}
Small plasmonic nanoparticles can be effectively described in the dipole approximation if their dimensions are much smaller than both the distance between them and wavelength of light.
Such an approach has been developed and applied for plasmonic lattices in many papers \cite{schokker2017systematic,chen2017general,reshef2019multiresonant,Augie2008,Rajeeva2018,kravets2018plasmonic,rodriguez2012,reshef2019multiresonant,berkhout2019perfect,Evlyukhin2010,Humphrey2014,Zhao2003,Garcia2007,kolkowski2019lattice,vaskin2019light} and proved itself to be much faster than widespread finite element method (FEM) or finite-difference time-domain (FDTD). Concurrently, its high precision allows us to trust the results and use them for analysis of physical effects as well as for the design of optical devices. In this article all the calculations are conducted via the method developed by our group in recent papers \cite{fradkin2019fourier, fradkin2018fourier,fradkin2020nanoparticle}, which allows to construct a scattering matrix of a plasmonic lattice in dipole approximation \cite{chaumet2003} and integrate it in Fourier modal method (FMM) \cite{tikhodeev2002} also known as Rigorous coupled-wave analysis (RCWA) \cite{moharam1995}. In such representation, the interaction of any number of vertically-spaced lattices is easily accounted for by the calculation of their combined scattering matrix \cite{ko88}.
At the same time, in order to provide an intuitive explanation of the demonstrated effects and interpretation of the results, here we apply the dipole model, which accounts for two spaced lattices as constituents of a single layer. Dipole moments of particles, $\mathbf{P}$, in a certain cell can be found by an application of polarizability tensor, $\hat{\alpha}$, to a background electric field acting on them:
\begin{equation}
\begin{pmatrix}
\mathbf{P}_1 \\ \mathbf{P}_2
\end{pmatrix} = \begin{pmatrix}
\hat{\alpha} &0\\ 0&\hat{\alpha}
\end{pmatrix}\begin{pmatrix}
\mathbf{E}_1^{\mathrm{bg}} \\ \mathbf{E}_2^{\mathrm{bg}}
\end{pmatrix},
\end{equation}
where indices $1,2$ correspond to either upper or lower lattice. Dipole moments of particles in different cells are trivially connected with each other by Bloch's theorem. Taking into account that background field is a sum of an external electric field, $\mathbf{E}^{0}$, of illuminating light and the field rescattered by neighboring particles of both lattices, we obtain the self-consistent system of equations:
\begin{equation}
\begin{pmatrix}
\mathbf{P}_1 \\ \mathbf{P}_2
\end{pmatrix}=
\begin{pmatrix}
\hat{\alpha} &0\\ 0&\hat{\alpha}
\end{pmatrix}\left[
\begin{pmatrix}
\mathbf{E}_1^0 \\ \mathbf{E}_2^0
\end{pmatrix}
+
\begin{pmatrix}\hat{C}_{11}&\hat{C}_{12}\\\hat{C}_{21}&\hat{C}_{22}\end{pmatrix}
\begin{pmatrix}
\mathbf{P}_1 \\ \mathbf{P}_2
\end{pmatrix}\right],
\end{equation}
where $\hat{C}$ blocks are lattice sums of dyadic Green's function \cite{chen2017general,kwadrin2014diffractive} also known as dynamic interaction constants \cite{Belov2005}. Diagonal blocks are associated with the self-action of lattices, and since the background environment is homogeneous they are equal in our case $\hat{C}_{11}=\hat{C}_{22}$. Off-diagonal blocks correspond to an interaction of upper and lower lattices. It should be especially noted that off-diagonal blocks are not equal to each other, $\hat{C}_{12}\neq\hat{C}_{21}$, and equality is achieved only under some specific conditions, which is discussed in supporting information.
Due to the absence of a relative in-plane shift of two lattices and subsequent high symmetry of the considered structure (see~Fig.~\ref{fig:1})
the off-diagonal elements of all the $\hat{C}$ blocks vanish for normally incident light ($\mathbf{k}_\parallel=0$) (see supporting information). This means that the $x$, $y$, and $z$ -polarized solutions get separated in terms of both dipole moments and incident light. Hence, the $z$ component of the dipole moment can not be excited by normally incident light. Concurrently, $x$ and $y$ -polarized solutions are equivalent to each other which allows us to consider only $x$ polarization without loss of generality:
\begin{multline}
\begin{pmatrix}
P_1^x \\ P_2^x
\end{pmatrix}=
\begin{pmatrix}
\alpha_{xx} &0\\ 0&\alpha_{xx}
\end{pmatrix}\times\\\left[
\begin{pmatrix}
E_1^{0x} \\ E_2^{0x}
\end{pmatrix}
+
\begin{pmatrix}C^{xx}_{11}&C^{xx}_{12}\\C^{xx}_{21}&C^{xx}_{11}\end{pmatrix}
\begin{pmatrix}
P^{x}_1 \\ P^{x}_2
\end{pmatrix}\right].
\end{multline}
The fact that $C_{12}^{xx}=C_{21}^{xx}$ (see supporting information), makes the equations symmetric. In turn, such a symmetric system supports two modes corresponding to in-phase and antiphase dipole moment oscillations, which we denote by A and B subscripts. Further we refer to A and B modes as even and odd correspondingly. Appropriate choice of the basis makes the system diagonal:
\begin{multline}
\begin{pmatrix}
P_\mathrm{A} \\ P_\mathrm{B}
\end{pmatrix}=
\frac{\alpha_{xx}}{\sqrt{2}}\begin{pmatrix}
E_1^{0x}+E_2^{0x} \\ E_1^{0x}-E_2^{0x}
\end{pmatrix}
+\\\alpha_{xx}
\begin{pmatrix}C^{xx}_{11}+C^{xx}_{12}&0\\0&C^{xx}_{11}-C^{xx}_{12}\end{pmatrix}
\begin{pmatrix}
P_\mathrm{A} \\ P_\mathrm{B}
\end{pmatrix},
\end{multline}
where
\begin{align}
\begin{pmatrix}
P_\mathrm{A} \\ P_\mathrm{B}
\end{pmatrix}
&=\frac{1}{\sqrt{2}}\begin{pmatrix}
1&1\\1&-1
\end{pmatrix}
\begin{pmatrix}
P^{x}_1 \\ P^{x}_2
\end{pmatrix}.
\end{align}
Solution of this system gives us amplitudes of the modes in the explicit form:
\begin{align}
P_\mathrm{A}=\frac{E^{0x}_1(1+e^{ikH})/\sqrt{2}}{\alpha_{xx}^{-1}-(C^{xx}_{11}+C_{12}^{xx})},\label{eqn:dipmom1}\\
P_\mathrm{B}=\frac{E^{0x}_1(1-e^{ikH})/\sqrt{2}}{\alpha_{xx}^{-1}-(C^{xx}_{11}-C_{12}^{xx})},\label{eqn:dipmom2}
\end{align}
where $k=k_0\sqrt{\varepsilon_{\mathrm{SiO}_2}}$ is the wavenumber in silica and the external incident light is assumed coming from above.
It is important to emphasize that these expressions are strict in the framework of the dipole approximation. However, we derived them not to conduct computations and obtain the precise solution, but for qualitative analysis.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig2.pdf}
\caption{Thickness-dependent spectra of (a) even and (b) odd mode amplitudes. White dotted lines correspond to the energy of RA, thick (or thin) green dashed lines indicate the condition of no-excitation (or maximum excitation) on panels (a,b). Magenta lines show the estimations of the position of Fabry-Perot modes. In panels (a) and (b) thick lines correspond to modes of the even and odd parities correspondingly. Panel (c) shows the extinction spectrum of the same structure for the normal light incidence.}
\label{fig:2}
\end{figure*}
As a first example, we consider the structure of $a=395$~nm period and compute its spectra as a function of thickness, $H$. Since ordinary lattice plasmons are observed in close vicinity of Rayleigh anomalies (RA) we consider energies near the first order RAs - $\left<\pm1,0\right>$ and $\left<0,\pm1\right>$. By means of our approach \cite{fradkin2018fourier,fradkin2019fourier,fradkin2020nanoparticle}, we calculate $\left|P_\mathrm{A}\right|$ and $\left|P_\mathrm{B}\right|$ as they are most convenient for analysis (Fig.~\ref{fig:2}~(a-b)). There are lots of modes and peculiarities in these figures that we explain stepwise below.
In the expressions for the modes amplitudes, (see~Eqns.~\ref{eqn:dipmom1},\ref{eqn:dipmom2}) numerators represent the effectiveness of incident light coupling to corresponding modes. They are periodic functions of $H$ that oscillate between their maxima and zero. From the explicit expressions (Eqns.~\ref{eqn:dipmom1},\ref{eqn:dipmom2}) we derive that amplitude of even mode A is zero for $kH=\pi+2\pi n$ and amplitude of odd mode B for $kH=2\pi n$, $n\in \mathbb{N}$. In other words, when these conditions are satisfied, normally incident light is not able to excite one or another mode because of the symmetry mismatch.
As we can see from Fig.~\ref{fig:2}~(a-b) thick green dashed lines, which correspond to these conditions indeed lay in a dark blue zone of no-excitation. Thin green lines are defined by the same formulas but correspond to the excitation maxima of complementary modes.
When the photon energy matches the level of RA, its wavelength in silica is equal to the period of the structure ($k=2\pi/a$) by definition, which means that the intersection of thick green lines with a white dotted line of RA occurs at $H=a/2+an$ or $H=an$ depending on the symmetry.
As the next step, we consider the hybridization of lattice plasmons. Taking into account the mode's parity and their proximity to RAs, one can easily obtain that the even lattice mode A (or odd lattice mode B) cannot be excited at $H\approx200,600,$ and $1000$~nm (or $H\approx0,400,$ and $800$ nm). However, these modes can be observed at thicknesses in between these values, which is confirmed by Fig.~\ref{fig:2}~(a-b).
In Ref.~\cite{baur2018} it was shown that in the case when both lattices lie in the same plane ($H=0$) and the relative in-plane shift between them tends to zero, $x$-components of diagonal and off-diagonal blocks $C^{xx}$ are equal $C^{xx}_{11}=C_{12}^{xx}$ in close vicinity of the first RA. However, since both of them diverge at RA and this divergence is due to harmonics propagating along the structure ($k_z\approx0$), this relation remains valid for $H>0$ in some limits. Therefore, lattice sums add up in denominator for the A mode amplitude (Eqn.~\ref{eqn:dipmom1}) and we observe the lattice plasmon in Fig.~\ref{fig:2}~(a) when the condition $\mathrm{Re}\alpha_{xx}^{-1} = 2 C^{xx}_{11}$ is satisfied. Also, in this case, the sums diverge on RA and $P_\mathrm{A}$ goes to zero exactly as in the case of a single lattice.
However, much more interesting situation is observed for the odd mode B where the lattice sums cancel each other out, which means that in order to obtain meaningful results we should accurately account for their finite difference and finite contributions of other harmonics.
Here we analyze analytical expressions to explain the most important effects.
It is convenient to represent lattice sums in the vicinity of RA as a sum of diverging term (originating from diffraction harmonics of the first order) and the remaining part: $C^{xx}_{11}=\frac{4\pi}{s}\frac{i k_0^2}{k_z} +\tilde{C}^{xx}_{11}$, $C^{xx}_{12}=\frac{4\pi}{s}\frac{i k_0^2}{k_z}e^{i k_z H} +\tilde{C}^{xx}_{12}(H)$, where $k_z=\sqrt{k^2-(2\pi/a)^2}$. What is important, $\tilde{C}^{xx}_{11}$ is a smooth function of energy near RA and does not depend on $H$ at all, whereas $\tilde{C}^{xx}_{12}$ is smooth as well, but depend on $H$. For relatively large thicknesses (more than 150-200 nm) this dependence is mostly determined by the zero order channel, which forces it to oscillate with a period of the wavelength in silica, which is close to the period of the structure, $a$, near RA.
In this way, denominator can be represented as follows:
\begin{multline}
\alpha^{-1}_{xx}-C_{11}^{xx}+C_{12}^{xx}=\\ \alpha^{-1}_{xx}+\frac{4\pi}{s}\frac{i k_0^2}{k_z}(e^{i k_z H}-1)-\tilde{C}^{xx}_{11} +\tilde{C}^{xx}_{12}(H)
\end{multline}
The main intrigue comes from the expression $(e^{i k_z H}-1)/k_z$ which is indeterminate on RA, which is easily resolved by an expansion in Taylor series:
\begin{multline} \alpha^{-1}_{xx}-C_{11}^{xx}+C_{12}^{xx}\approx\\
\alpha^{-1}_{xx}+\frac{4\pi}{s}\frac{i k_0^2}{k_z}(i k_z H - k_z^2 H^2/2) -\tilde{C}^{xx}_{11} +\tilde{C}^{xx}_{12}(H) = \\ \alpha^{-1}_{xx}- \frac{4\pi k_0^2}{s} H - \frac{4\pi k_0^2}{s} i k_z H^2/2 -\tilde{C}^{xx}_{11} +\tilde{C}^{xx}_{12}(H) \label{eqn:9}
\end{multline}
We see that divergent contributions compensate each other effectively for small thicknesses $H$. The impact of the remaining smooth term $-\tilde{C}^{xx}_{11} +\tilde{C}^{xx}_{12}(H) $ is not enough to change the behaviour of the
structure and, therefore, we observe an ordinary wide line of localized plasmon resonance for $H$ less than 200-250~nm (see Fig.~\ref{fig:2}~(b)). For slightly larger thicknesses, the linear term $\frac{4\pi k_0^2}{s} H$ grows in such a way that at some moment it cancels out $\alpha_{xx}^{-1}$ in the denominator of Eqn.~\ref{eqn:dipmom2} and from $H\approx200$ nm we observe bright yellow line of odd lattice plasmon right on the Rayleigh anomaly. The quadratic term in thickness $- \frac{4\pi k_0^2}{s} i k_z H^2/2$ does not make any contribution when we consider energy of RA ($k_z=0$), however, it indicates that there should be a discontinuity of the energy derivative in this point. Moreover, jump of the derivative grows fast with thickness $H$.
As it will be shown further (see~Fig.~\ref{fig:4}~(b-e)) this leads to strongly non-Lorentzian shapes of resonances. The common effect of the shape violation near RA was already considered for some other structures \cite{akimov2011optical}.
Interestingly, this resonance is much narrower than conventional lattice plasmons. With the further increase of $H$, the resonance condition on RA remains valid until approximately 350~nm, which is very close to the thick green line of no-excitation. When this resonance comes back after the intersection with this line it is shifted to the red zone as it should be in the regime of relatively weak coupling between two lattices. In the limit of a very large thickness, $H$, there is no difference in energy of even, odd, and ordinary LPRs of a single lattice.
The last type of resonance to be discussed is a Fabry-Perot-like one. Indeed, we observe several relatively narrow lines above the RA (see~Fig. \ref{fig:2}~(a,b)). In contrast with LPRs, which are mainly associated with photons having $k_z\approx0$ and no phase difference between two lattices ($\mathrm{Re}k_z H=0$), these modes couple them by some sort of quantized photons. In order to understand their nature, we again analyze the denominators from Eqns.~(\ref{eqn:dipmom1}-\ref{eqn:dipmom2}), but in slightly another way. It is obvious that symmetric and antisymmetric resonances arise when $\alpha^{-1}_{xx}-C_{11}^{xx}\approx\pm C^{xx}_{12}(H)$.
We again assume that interaction
between lattices is mostly determined by the first diffraction harmonics
$C^{xx}_{12}\propto\frac{i}{k_z}e^{ik_z H}$ and, therefore, $k_z H=-\pi/2 \pm \mathrm{arg} (\alpha^{-1}_{xx}-C_{11}^{xx})+2\pi n$.
An expression $\alpha^{-1}_{xx}-C_{11}^{xx}$ corresponds to a self-action of lattice. If we consider spectrum of the single lattice (see supporting information) then we will see that there is a peak of localized plasmon resonance in the energy range of 2.25-2.35~eV, which means that the phase of the considered expression is approximately $-\pi/2$ in this range. Since $k_x=2\pi/a$, even modes, which are associated with $k_z H\approx \pi +2\pi n$ have the following $H$-dependence $k=\sqrt{ \left({2\pi}/{a}\right)^2+\left({\pi+2 \pi n}/{H}\right)^2}$. For odd modes we have $k_z H\approx 2\pi n$ and $k=\sqrt{\left({2\pi}/{a}\right)^2+\left({2 \pi n}/{H}\right)^2}$ correspondingly. Although these estimations are rough - they are verified by the fact that corresponding thick magenta lines match with rigorously calulated resonances (see~Fig.~\ref{fig:2}~(a,b)).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig3.pdf}
\caption{ Thickness dependence of plasmonic modes energy of 395~nm-period lattice. Extinction spectrum is shown in slightly faded colors at the background.}
\label{fig:3}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig4.pdf}
\caption{Panel (a) shows the extinction of lattice stack on Rayleigh anomaly as a function of its period and thickness. The crosshatched region with white edge indicates the area in which the peak of the resonance is located exactly on the RA. Red lines show the naive analytical estimation of the resonance ridge position and width, central line corresponds to the real part of the given expression, whereas each of the sidelines is shifted on its imaginary part. Multicolored markers correspond to geometric parameters of lattices considered in panels (b-e). Panels (b-e) show the extinction spectra of the structures defined by markers. The shape of the marker determines the group of the fixed period, whereas colors in each group correspond to different thicknesses. Insets in each panel show shape of lines in close vicinity of RAs.}
\label{fig:4}
\end{figure*}
It is worth noting that the first branches of this kind are better to associate with waveguide modes rather than Fabry-Perot ones due to the small phase difference. Also, it is interesting that only the zero phase shift provides us with both even and odd LPR modes.
Amplitudes of even and odd modes are very convenient for analysis, however, in practice, we can observe only integral characteristics that provide us a compound of modes of all parities. For instance, the thickness-dependent extinction spectrum of normally-incident light (Fig.~\ref{fig:2}~(c)) shows that lattice plasmons of complementary symmetries, as well as Fabry-Perot modes, alternate each other with the thickness increase.
We explained the origin of the resonances in this structure, but for further analysis, it is convenient to find energies of resonances as a function of structure parameters. Conventionally it is done by the search of poles of some response function (it would have been scattering matrix in our case) in the plane of complex energy.
However, the most exciting odd (B-type) LPR, which attracts our attention has a non-Lorentzian shape when being located near RA and, therefore, can not be described by a single pole. In this work we do not go into detail of the correct representation of such resonances \cite{akimov2011optical,weiss2017analytical}, but find eigenenergies just as local maxima of $\left|P_\mathrm{A}(E)\right|$ and $\left|P_\mathrm{B}(E)\right|$ functions.
This approach allows us to easily plot the dependence of both LPR energies on the thickness for the same structure.
From the Fig.~\ref{fig:3} we see how localized surface plasmon resonance (LSPR) gradually transforms into odd lattice plasmon with the thickness increase. Moreover, we see that the peak of $|P_\mathrm{B}|$ stands on RA not only for $H$ in between approximately 200 and 350~nm, but also for smaller thicknesses. Indeed, there is also a small peak, which is not seen by eye, but exists. Blue and red lines of lattice plasmons of complementary parities indeed get closer with the rise of thickness and have the same energy for $H \approx 900$ nm. Also, we clearly see the periodic oscillations of each resonance energy, which are due to the already discussed periodic behavior of $\tilde{C}_{12}^{xx}(H)$.
The presented calculations help us to understand the $H$-dependence of all the modes, including the odd B-type LPR. However, it is also important to explore how this mode behaves with the period variation. In order to do that we scan for periods and thicknesses simultaneously and plot the extinction of corresponding structures on RA (see~Fig.~\ref{fig:4}~(a)). First of all, we see that there is a bright ridge, which corresponds to the resonant conditions of odd LPR. This ridge is cut by the thick green dashed lines $H=a,2a$ that indicate the impossibility to excite odd modes as in Fig.~\ref{fig:2}~(b). The crosshatched region with a white edge in the upper-left part of the graph highlights the area of parameters for which the peak of $\left|P_\mathrm{B}(E)\right|$ is right on the RA. Interestingly, the edge of this region acts as a crest of the ridge and separates the parameter space into two parts in such a way that half of the bright area corresponds to the peak of resonance on RA and the other half to the peak in its close vicinity.
The shape of the bright ridge line can be easily estimated analytically. Indeed, since we consider energy of RA ($E_{\mathrm{RA}}=\frac{h c }{a\sqrt{\varepsilon_{\mathrm{SiO}_2}}}$) then $k_z=0$ and, according to Eqn.~\ref{eqn:9} resonant condition, which connects $H$ and $a$, has the following form:
\begin{equation}
H=\frac{s}{4\pi k_0^2}\left(\alpha^{-1}_{xx}(E_\mathrm{RA})-\tilde{C}^{xx}_{11}(E_\mathrm{RA},a)+\tilde{C}^{xx}_{12}(E_\mathrm{RA},a,H)\right).
\end{equation}
We have already seen that smooth functions of energy and period $\tilde{C}$ being added to the inverse polarizability tensor does not have significant impact on its behaviour. Therefore, in order to simplify our estimation we just omit them, but note that $\tilde{C}^{xx}_{11}$ has nearly constant complex contribution to the value of $H$ as a function of $a$, whereas $\tilde{C}^{xx}_{12}$ has periodic dependence of $H$ which should make the dependence slightly oscillating.
Taking into account that $s=a^2$ and $k_0^2 = 4\pi^2/a^2/\varepsilon_{\mathrm{SiO}_2}$ in this case, we obtain
\begin{equation}
H(a)\approx \varepsilon_{\mathrm{SiO}_2}\frac{a^4}{16\pi^3}\alpha^{-1}_{xx}(\frac{hc}{\sqrt{\varepsilon_{\mathrm{SiO}_2}} a}).
\end{equation}
As we see from the Fig.~\ref{fig:4}~(a) red lines, which show both the estimation of the ridge position (real part) as well as it is width (imaginary part) indeed match the precise calculation. Moreover, we see that the discrepancy is well explained by the an unaccounted contribution of secondary terms $\tilde{C}$.
In order to provide a full picture, we compute extinction spectra for several structures, whose geometric parameters are indicated by markers in Fig~\ref{fig:4}~(a). The first group denoted by circles, which are located at the most bright part of the first "hill" of the ridge, is plotted in Fig~\ref{fig:4}~(b). The main peculiarity of this group is that they illustrate the case in which RA lays right onto the localized resonance and strongly deforms its shape. The green line shows the spectrum of the structure, whose parameters satisfy the relation $H=a/2$, which means that even modes can not be excited and indeed, we observe only single odd resonance. Blue and red lines that have slightly different thicknesses $H$ both demonstrate small peaks of even lattice modes in the red zone. As we can see from the inset all the lines have strongly non-Lorentzian shape and red line in contrast to the other ones have its peak slightly shifted to the red zone, which is in complete agreement with a relative position of the circles with respect to the white line of the ridge.
The next group, depicted by squares, has the period, $a$, of 395 nm, which corresponds to the structures considered in the Figs.~\ref{fig:2}~and~\ref{fig:3}. From Fig.~\ref{fig:4}~(c) we see that in this case even LPRs are much more prominent and besides the two lines of lattice resonances we observe Fabry-Perot modes in the blue zone as well. Also, the extinction of the red line is strongly suppressed at the RA because of the vicinity of the green line of no-excitation.
Diamonds correspond to such a period that odd modes excitation in the vicinity of RA is strongly suppressed. From the inset in Fig~\ref{fig:4}~(d) we see that there is no resonance at all for the green line ($H=a$ in this case). The blue line peak is so small that it is not seen and the peak of the red one is slightly redshifted as it should be. Interestingly, the blue line in the vicinity of the RA has a step-like shape, which might be potentially applied for example in optical filtering.
The last group of triangles corresponds to the second "hill" of the ridge. We see that in this case large thicknesses $H$ result in the observation of several Fabry-Perot modes and very narrow lines of lattice modes due to their strong red shift far from the line of dissipative localized plasmon.
Similar to the case considered in Fig~\ref{fig:4}~(b), the green line does not demonstrate the peak of the even lattice plasmon due to the specific relation of thickness and period, $H=1.5a$. From the inset, we see that there is almost no significant difference between the lines near the RA.
What is important, these calculations demonstrate that energy of very narrow resonances can be easily set to $E_{\mathrm{RA}}$ by our approach. We hope that this unique property will be useful in problems, which require setting of high-$Q$ resonances at predetermined positions.
In this paper, we have considered the structure of a rather simple design. However, it has lots of degrees of freedom to be tuned for the achievement of the desired properties. For example, the relative in-plane shift of two lattices can interchange energies of even and odd modes or mix them. The choice of the shape and size of particles as well as the type of the lattice potentially allows to consider resonances with non-trivial polarization of dipole moments, demonstrate chiral and spin-orbit effects. Yet another way to go further in the exploration of plasmonic lattices is an increase of a number of particles in a cell or number of layers. Although it is a trivial step in terms of calculation, it can open exciting opportunities for light manipulation.
\section*{Conclusion}
In this paper, we considered a stack of two plasmonic lattices and analyzed hybridized modes emerging in such structure. We showed that aside from localized plasmon it supports Fabry-Perot modes and lattice plasmons of different symmetries. Particularly, the demonstrated antisymmetric LPR is very untypical. It has a very high $Q$-factor for the given detuning from the line of LSPR.
Moreover, it is located strictly on the Rayleigh anomaly for a wide range of structure parameters, which makes it possible to set the desired energy precisely by an appropriate period and independently choose its thickness from the wide range of accessible values. This unique property may be used either as an additional tuning parameter for optimization of the desired properties or as a protection against experimental errors.
Also, it makes stacks of plasmonic lattices attractive for application in problems for which the precise positioning of resonant lines is crucial. In particular, one can employ them in light emission enhancement and routing, lasing from plasmonic structures, optical filtering.
\section*{Acknowledgements}
This work was supported by the Russian Foundation for Basic Research (Grant No. 18-29-20032).
\bibliographystyle{apsrev4-1}
| 44640e77b7232269e518c42b011abb9e46b9c190 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Acknowledgment}
We thank anonymous reviewers for their helpful comments to improve this research.
Yanhao Wang has been supported by the MLDB project of Academy of Finland (decision number: 322046).
The research of Raymond Chi-Wing Wong is supported by HKRGC GRF 16219816.
\bibliographystyle{IEEEtran}
\section{Introduction}
\label{sec:intro}
In many real-world applications such as multi-criteria decision
making~\cite{DBLP:journals/pvldb/NanongkaiSLLX10},
web search~\cite{DBLP:conf/edbt/StoyanovichYJ18},
recommendation~\cite{DBLP:journals/tkde/WangLT19,DBLP:conf/recsys/LiuMLY11},
and data description~\cite{DBLP:conf/kdd/WangLT19},
it is crucial to find a succinct representative subset from a large database
to meet the requirements of various users.
For example, when a user queries for a hotel on a website (e.g.,~booking.com and~expedia.com),
she/he will receive thousands of available options as results.
The website would like to display the best choices in the first few pages
from which almost all users could find what they are most interested in.
A common method is to rank all results using a utility function
that denotes a user's preference on different attributes
(e.g.,~\emph{price}, \emph{rating}, and \emph{distance to destination} for hotels)
and only present the top-$k$ tuples with the highest scores
according to this function to the user.
However, due to the wide diversity of user preferences,
it is infeasible to represent the preferences of all users by any single utility function.
Therefore, to select a set of highly representative tuples,
it is necessary to take all (possible) user preferences into account.
A well-established approach to finding such representatives from databases
is the \emph{skyline} operator~\cite{DBLP:conf/icde/BorzsonyiKS01}
based on the concept of \emph{domination}: a tuple $p$ dominates a tuple $q$ iff
$p$ is as good as $q$ on all attributes and strictly better than $q$ on at least one attribute.
For a given database, a skyline query returns its \emph{Pareto-optimal} subset which
consists of all tuples that are not dominated by any tuple.
It is guaranteed that any user can find her/his best choice from the skyline
because the top-ranked result according to any monotone function must not be dominated.
Unfortunately, although skyline queries are effective for representing low-dimensional databases,
their result sizes cannot be controlled
and increase rapidly as the dimensionality (i.e.,~number of attributes in a tuple) grows,
particularly so for databases with anti-correlated attributes.
Recently, the $k$-regret minimizing set ($k$-RMS)
problem~\cite{DBLP:journals/pvldb/NanongkaiSLLX10,DBLP:conf/icde/PengW14,DBLP:journals/pvldb/ChesterTVW14,DBLP:conf/wea/AgarwalKSS17,DBLP:conf/icdt/CaoLWWWWZ17,DBLP:conf/sigmod/AsudehN0D17,DBLP:conf/alenex/KumarS18,DBLP:conf/sigmod/XieW0LL18}
was proposed to alleviate the deficiency of skyline queries.
Specifically, given a database $P$ of tuples with $d$ numeric attributes,
the $k$-RMS problem aims to find a subset $Q \subseteq P$ such that, for any possible utility function,
the top-$1$ tuple in $Q$ can approximate the top-$k$ tuples in $P$ within a small error.
Here, the \emph{maximum $k$-regret ratio}~\cite{DBLP:journals/pvldb/ChesterTVW14} ($\mathtt{mrr}_k$)
is used to measure how well $Q$ can represent $P$.
For a utility function $f$, the $k$-regret ratio ($\mathtt{rr}_k$) of $Q$ over $P$
is defined to be $0$ if the top-$1$ tuple in $Q$ is among the top-$k$ tuples in $P$ w.r.t.~$f$,
or otherwise, to be one minus the ratio between the score of the top-$1$ tuple in $Q$
and the score of the $k$\textsuperscript{th}-ranked tuple in $P$ w.r.t.~$f$.
Then, the \emph{maximum $k$-regret ratio} ($\mathtt{mrr}_k$) is
defined by the maximum of $\mathtt{rr}_k$ over a class of (possibly infinite) utility functions.
Given a positive integer $r$,
a $k$-RMS on a database $P$ returns a subset $Q \subseteq P$ of size $r$ to minimize $\mathtt{mrr}_k$.
As an illustrative example, the website could run a $k$-RMS on all available hotels to
pick a set of $r$ candidates from which all users can find at least one close to her/his top-$k$ choices.
The $k$-RMS problem has been extensively studied recently.
Theoretically, it is
NP-hard~\cite{DBLP:journals/pvldb/ChesterTVW14,DBLP:conf/wea/AgarwalKSS17,DBLP:conf/icdt/CaoLWWWWZ17}
on any database with $d \geq 3$.
In general, we categorize existing $k$-RMS algorithms into three types.
The first type is dynamic programming
algorithms~\cite{DBLP:conf/sigmod/AsudehN0D17,DBLP:journals/pvldb/ChesterTVW14,DBLP:conf/icdt/CaoLWWWWZ17}
for $k$-RMS on two-dimensional data.
Although they can provide optimal solutions when $d = 2$,
they are not suitable for higher dimensions due to the NP-hardness of $k$-RMS.
The second type is the greedy
heuristic~\cite{DBLP:journals/pvldb/NanongkaiSLLX10,DBLP:conf/icde/PengW14,DBLP:journals/pvldb/ChesterTVW14},
which always adds a tuple that maximally reduces $\mathtt{mrr}_k$ at each iteration.
Although these algorithms can provide high-quality results empirically,
they have no theoretical guarantee and suffer from low efficiency on high-dimensional data.
The third type is to transform $k$-RMS into another problem
such as $\varepsilon$-kernel~\cite{DBLP:conf/sigmod/XieW0LL18,DBLP:conf/alenex/KumarS18,DBLP:conf/wea/AgarwalKSS17,DBLP:conf/icdt/CaoLWWWWZ17},
discretized matrix min-max~\cite{DBLP:conf/sigmod/AsudehN0D17},
hitting set~\cite{DBLP:conf/alenex/KumarS18,DBLP:conf/wea/AgarwalKSS17},
and $k$-\textsc{medoid} clustering~\cite{Shetiya2019},
and then to utilize existing solutions of the transformed problem for $k$-RMS computation.
Although these algorithms are more efficient than greedy heuristics while having theoretical bounds,
they are designed for the static setting and
cannot process database updates efficiently.
Typically, most of them precompute the skyline as an input to compute the result of
$k$-RMS. Once a tuple insertion or deletion triggers any change in the skyline,
they are unable to maintain the result without re-running from scratch.
Hence, existing $k$-RMS algorithms become very inefficient in highly dynamic environments
where tuples in the databases are frequently inserted and deleted.
However, dynamic databases are very common in real-world scenarios, especially for online services.
For example, in a hotel booking system,
the \emph{prices} and \emph{availabilities} of rooms are frequently changed over time.
As another example, in an IoT network,
a large number of sensors may often connect or disconnect with the server.
Moreover, sensors also update their statistics regularly.
Therefore, it is essential to address the problem of maintaining an up-to-date
result for $k$-RMS when the database is frequently updated.
In this paper, we propose the first fully-dynamic $k$-RMS algorithm that
can efficiently maintain the result of $k$-RMS w.r.t.~any tuple
insertion and deletion in the database with both theoretical guarantee and good empirical performance.
Our main contributions are summarized as follows.
\begin{itemize}
\item We formally define the notion of \emph{maximum $k$-regret ratio}
and the \emph{$k$-regret minimizing set} ($k$-RMS) problem in a fully-dynamic setting.
(Section~\ref{sec:definition})
\item We propose the first fully-dynamic algorithm called FD-RMS
to maintain the $k$-RMS result over tuple insertions and deletions in a database.
Our basic idea is to transform \emph{fully-dynamic $k$-RMS} into
a \emph{dynamic set cover} problem. Specifically, FD-RMS computes the (approximate)
top-$k$ tuples for a set of randomly sampled utility functions
and builds a set system based on the top-$k$ results.
Then, the $k$-RMS result can be retrieved from an approximate solution
for \emph{set cover} on the set system.
Furthermore, we devise a novel algorithm for dynamic set cover by introducing
the notion of \emph{stable solution},
which is used to efficiently update the $k$-RMS result
whenever an insertion or deletion triggers some changes
in top-$k$ results as well as the set system.
We also provide detailed theoretical analyses of FD-RMS.
(Section~\ref{sec:framework})
\item We conduct extensive experiments on several real-world and synthetic datasets
to evaluate the performance of FD-RMS.
The results show that FD-RMS achieves up to four orders of magnitude speedup over existing
$k$-RMS algorithms while providing results of nearly equal quality in a fully dynamic setting.
(Section~\ref{sec:exp})
\end{itemize}
\section{The FD-RMS Algorithm}
\label{sec:framework}
\begin{figure}
\centering
\includegraphics[width=.475\textwidth]{framework.pdf}
\caption{An illustration of FD-RMS}
\label{fig:framework}
\end{figure}
In this section, we present our FD-RMS algorithm for $k$-RMS in a fully dynamic setting.
The general framework of FD-RMS is illustrated in Fig.~\ref{fig:framework}.
The basic idea is to transform \emph{fully-dynamic $k$-RMS}
to a \emph{dynamic set cover} problem.
Let us consider how to compute the result of $\mathtt{RMS}(k,r)$ on database $P_t$.
First of all, we draw a set of $m$ random utility vectors $\{u_1,\ldots,u_m\}$
from $ \mathbb{U} $ and maintain the $\varepsilon$-approximate top-$k$ result
of each $u_i$ ($i \in [1,m]$) on $P_t$, i.e., $ \Phi_{k,\varepsilon}(u_i,P_t) $.
Note that $\varepsilon$ should be given as an input parameter of FD-RMS
and we will discuss how to specify its value at the end of Section~\ref{sec:framework}.
Then, we construct a set system $ \Sigma = (\mathcal{U},\mathcal{S}) $
based on the approximate top-$k$ results,
where the universe $\mathcal{U}=\{u_1,\ldots,u_m\}$
and the collection $\mathcal{S}$ consists of $n_t$ sets ($n_t=|P_t|$)
each of which corresponds to one tuple in $P_t$.
Specifically, for each tuple $p \in P_t$, we define $S(p)$ as a set of
utility vectors for which $p$ is an $\varepsilon$-approximate top-$k$ result on $P_t$.
Or formally, $S(p) = \{u \in \mathcal{U} : p \in \Phi_{k,\varepsilon}(u,P_t)\}$
and $\mathcal{S}=\{ S(p) : p \in P_t \}$.
After that, we compute a result $Q_t$ for $\mathtt{RMS}(k,r)$ on $P_t$
using an (approximate) solution for set cover on $\Sigma$.
Let $\mathcal{C} \subseteq \mathcal{S}$ be a set-cover solution of $\Sigma$,
i.e., $\bigcup_{S(p) \in \mathcal{C}} S(p)=\mathcal{U}$.
We use the set $Q_t$ of tuples corresponding to $\mathcal{C}$,
i.e., $Q_t=\{p \in P_t : S(p) \in \mathcal{C}\}$,
as the result of $\mathtt{RMS}(k,r)$ on $P_t$.
Given the above framework, there are still two challenges of updating the result of $k$-RMS
in a fully dynamic setting. Firstly, because the size of $Q_t$ is restricted to $r$,
it is necessary to always keep an appropriate value of $m$ over time
so that $|\mathcal{C}| \leq r$.
Secondly, the updates in approximate top-$k$ results triggered by tuple insertions
and deletions in the database
lead to the changes in the set collection $\mathcal{S}$.
Therefore, it is essential to maintain the set-cover
solution $\mathcal{C}$ over time for the changes in $\mathcal{S}$.
In fact, both challenges can be treated as a \emph{dynamic set cover} problem that keeps
a set-cover solution w.r.t.~changes in both $\mathcal{U}$ and $\mathcal{S}$.
Therefore, we will first introduce the background on \emph{dynamic set cover}
in Section~\ref{subsec:set:cover}.
After that, we will elaborate on how FD-RMS processes $k$-RMS in a fully dynamic setting
using the \emph{dynamic set cover} algorithm in Section~\ref{subsec:alg}.
\subsection{Background: Dynamic Set Cover}
\label{subsec:set:cover}
Given a set system $\Sigma=(\mathcal{U},\mathcal{S})$, the \emph{set cover} problem asks for
the smallest subset $\mathcal{C}^{*}$ of $\mathcal{S}$ whose union equals to the universe $\mathcal{U}$.
It is one of Karp's 21 NP-complete problems~\cite{DBLP:conf/coco/Karp72},
and cannot be approximated to $(1-o(1))\cdot\ln{m}$ ($m=|\mathcal{U}|$) unless P=NP~\cite{DBLP:journals/jacm/Feige98}.
A common method to find an approximate set-cover solution is the greedy algorithm.
Starting from $\mathcal{C} = \varnothing$, it always adds the set that contains the largest number of uncovered
elements in $\mathcal{U}$ to $\mathcal{C}$ at each iteration until $\bigcup_{S \in \mathcal{C}}S=\mathcal{U}$.
Theoretically, the solution $\mathcal{C}$ achieves an approximation ratio of $(1+\ln{m})$,
i.e., $|\mathcal{C}| \leq (1+\ln{m}) \cdot |\mathcal{C}^{*}|$.
But obviously, the greedy algorithm cannot dynamically update the set-cover solution
when the set system $\Sigma$ is changed.
Recently, there are some theoretical advances on covering and relevant problems
(e.g., vertex cover, maximum matching, set cover, and maximal independent set) in dynamic
settings~\cite{DBLP:journals/siamcomp/BhattacharyaHI18,DBLP:conf/stoc/GuptaK0P17,DBLP:conf/stoc/AbboudA0PS19,DBLP:conf/stacs/HjulerIPS19}.
Although these theoretical results have opened up new ways to design dynamic set cover algorithms,
they cannot be directly applied to the update procedure of FD-RMS because of two limitations.
First, existing dynamic algorithms for set cover~\cite{DBLP:conf/stoc/AbboudA0PS19,DBLP:conf/stoc/GuptaK0P17}
can only handle the update in the universe $\mathcal{U}$ but assume that the set collection $\mathcal{S}$ is not changed.
But in our scenario, the changes in top-$k$ results lead to the update of $\mathcal{S}$.
Second, due to the extremely large constants introduced in their analyses,
the solutions returned may be far away from the optima in practice.
Therefore, we devise a more practical approach to dynamic set cover
that supports any update in both $\mathcal{U}$ and $\mathcal{S}$.
Our basic idea is to introduce the notion of \emph{stability} to a set-cover solution.
Then, we prove that any stable solution is $O(\log{m})$-approximate ($m=|\mathcal{U}|$) for set cover.
Based on this result, we are able to design an algorithm to maintain a set-cover solution
w.r.t.~any change in $\Sigma$ by guaranteeing its \emph{stability}.
We first formalize the concept of \emph{stability} of a set-cover solution.
Let $\mathcal{C} \subseteq \mathcal{S}$ be a set-cover solution on $\Sigma=(\mathcal{U},\mathcal{S})$.
We define an assignment $\phi$ from each element $u \in \mathcal{U}$ to a \emph{unique} set $S \in \mathcal{C}$
that contains $u$ (or formally, $\phi: \mathcal{U} \rightarrow \mathcal{C}$).
For each set $S \in \mathcal{C}$, its cover set $\mathtt{cov}(S)$ is defined as the set of elements assigned to $S$,
i.e., $\mathtt{cov}(S) = \{u \in \mathcal{U} \,:\, \phi(u)=S\}$.
By definition, the cover sets of different sets in $\mathcal{C}$ are \emph{mutually disjoint} from each other.
Then, we can organize the sets in $\mathcal{C}$ into hierarchies according to the numbers of elements covered by them.
Specifically, we put a set $S \in \mathcal{C}$ in a higher level if it covers more elements and vice versa.
We associate each level $\mathcal{L}_j$ ($j \in \mathbb{N}$) with a range of cover
number\footnote{Here, the base $2$ may be replaced by any constant greater than 1.}
$[2^{j},2^{j+1})$.
Each set $S \in \mathcal{C}$ is assigned to a level $\mathcal{L}_j$ if $2^{j} \leq |\mathtt{cov}(S)| < 2^{j+1}$.
We use $A_j$ to denote the set of elements assigned to any set in $\mathcal{L}_j$,
i.e., $A_j = \{u \in \mathcal{U} \,:\, \phi(u) \in \mathcal{L}_j\}$.
Moreover, the notations $\mathcal{L}$ with subscripts, i.e.,
$\mathcal{L}_{>j}$ or $\mathcal{L}_{\geq j}$ and $\mathcal{L}_{<j}$ or $\mathcal{L}_{\leq j}$,
represent the sets in all the levels above (excl.~or incl.) and below $\mathcal{L}_j$ (excl.~or incl.), respectively.
The same subscripts are also used for $A$.
Based on the above notions, we formally define the \emph{stability} of a set-cover solution in Definition~\ref{def:stable}
and give its approximation ratio in Theorem~\ref{thm:approx:ratio}.
\begin{definition}[Stable Set-Cover Solution]\label{def:stable}
A solution $\mathcal{C}$ for \emph{set cover} on $\Sigma=(\mathcal{U},\mathcal{S})$
is stable if:
\begin{enumerate}
\item For each set $S \in \mathcal{L}_j$, $2^{j} \leq |\mathtt{cov}(S)| < 2^{j+1}$;
\item For each level $\mathcal{L}_j$, there is no $S \in \mathcal{S}$ s.t. $|S \cap A_j| \geq 2^{j+1}$.
\end{enumerate}
\end{definition}
\begin{theorem}\label{thm:approx:ratio}
If a set-cover solution $\mathcal{C}$ is stable,
then it satisfies that $|\mathcal{C}| \leq O(\log m) \cdot |\mathcal{C}^{*}|$.
\end{theorem}
\begin{proof}
Let $\mathtt{OPT}=|\mathcal{C}^{*}|$, $\rho^{*}=\frac{m}{\mathtt{OPT}}$,
and $j^{*}$ be the level index such that $2^{j^{*}} \leq \rho^{*} < 2^{j^{*}+1}$.
According to Condition (1) of Definition~\ref{def:stable},
we have $|\mathtt{cov}(S)| \geq 2^{j^{*}}$ for any $S \in \mathcal{L}_{\geq j^{*}}$.
Thus, it holds that
$|\mathcal{L}_{\geq j^{*}}| \leq \frac{|A_{\geq j^{*}}|}{2^{j^{*}}} \leq \frac{m}{2^{j^{*}}}
\leq \frac{\rho^{*}}{2^{j^{*}}} \cdot \mathtt{OPT} \leq 2 \cdot \mathtt{OPT}$.
For some level $\mathcal{L}_j$ with $j < j^{*}$,
according to Condition (2) of Definition~\ref{def:stable}, any $S \in \mathcal{S}$
covers at most $2^{j+1}$ elements in $A_j$.
Hence, $\mathcal{S}^{*}$ needs at least $\frac{|A_j|}{2^{j+1}}$ sets
to cover $A_j$, i.e., $\mathtt{OPT} \geq \frac{|A_j|}{2^{j+1}}$.
Since $|\mathtt{cov}(S)| \geq 2^{j}$ for each $S \in \mathcal{L}_j$, it holds that
$|\mathcal{L}_j| \leq \frac{|A_j|}{2^{j}} \leq 2 \cdot \mathtt{OPT}$.
As $1 \leq |\mathtt{cov}(S)| \leq m$, the range of level index is $[0,\log_{2}{m}]$.
Thus, the number of levels below $\mathcal{L}_{j^{*}}$ is at most $\log_{2}{m}$.
To sum up, we prove that
\begin{equation*}\textstyle
|\mathcal{C}| = |\mathcal{L}_{\geq j^{*}}| + |\mathcal{L}_{< j^{*}}|
\leq (2 + 2 \log_{2}{m}) \cdot \mathtt{OPT}
\end{equation*}
and conclude the proof.
\end{proof}
\begin{algorithm}[t]
\caption{\textsc{Dynamic Set Cover}}\label{alg:set:cover}
\Input{Set system $\Sigma$, set operation $\sigma$}
\Output{Stable set-cover solution $\mathcal{C}$}
\tcc{compute an initial solution $\mathcal{C}$ on $\Sigma$}
$ \mathcal{C} \gets$ \textsc{Greedy}$(\Sigma)$\;\label{ln:cover:init}
\tcc{update $\mathcal{C}$ for $\sigma=(u,S,\pm)$ or $(u,\mathcal{U},\pm)$}
\uIf{$\sigma=(u,S,-)$ and $u \in \mathtt{cov}(S)$\label{ln:sigma:s}}{
$ \mathtt{cov}(S) \gets \mathtt{cov}(S) \setminus \{u\} $\;
$ \mathtt{cov}(S^+) \gets \mathtt{cov}(S^+) \cup \{u\} $ for $S^+ \in \mathcal{S}$
s.t.~$u \in S^+$\;
\textsc{ReLevel}$(S)$ and \textsc{ReLevel}$(S^+)$\;\label{ln:move:1}
}
\uElseIf{$\sigma=(u,\mathcal{U},+)$}{
$ \mathtt{cov}(S^+) \gets \mathtt{cov}(S^+) \cup \{u\} $ for $S^+ \in \mathcal{S}$
s.t.~$u \in S^+$\;
\textsc{ReLevel}$(S^+)$\;\label{ln:move:2}
}
\ElseIf{$\sigma=(u,\mathcal{U},-)$}{
$ \mathtt{cov}(S^-) \gets \mathtt{cov}(S^-) \setminus \{u\} $ if $u \in \mathtt{cov}(S^-)$\;
\textsc{ReLevel}$(S^-)$\;\label{ln:move:3}
}
\textsc{Stabilize}$(\mathcal{C})$\;\label{ln:sigma:t}
\BlankLine
\Fn{\textnormal{\textsc{Greedy}$(\Sigma)$}\label{ln:greedy:s}}{
$I \gets \mathcal{U}$, $\mathcal{L}_j \gets \varnothing$ for every $j \geq 0$\;
\While{$I \neq \varnothing$}{
$S^* \gets \arg\max_{S \in \mathcal{S}} |I \cap S|$,
$\mathtt{cov}(S^*) \gets I \cap S^*$\;\label{ln:cover:max}
Add $S^*$ to $\mathcal{L}_j$ s.t.~$2^{j} \leq |\mathtt{cov}(S^*)| < 2^{j+1}$\;
$I \gets I \setminus \mathtt{cov}(S^*)$\;
}
\Return{$\mathcal{C} \gets \bigcup_{j \geq 0} \mathcal{L}_j$}\;\label{ln:greedy:t}
}
\Fn{\textnormal{\textsc{ReLevel}$(S)$}\label{ln:move:s}}{
\uIf{$\mathtt{cov}(S) = \varnothing$}{
$\mathcal{C} \gets \mathcal{C} \setminus \{S\}$\;
}
\Else{
Let $\mathcal{L}_{j}$ be the current level of $S$\;
\If{$|\mathtt{cov}(S)| < 2^{j}$ or $|\mathtt{cov}(S)| \geq 2^{j+1}$}{
Let $j'$ be the index s.t. $2^{j'} \leq |\mathtt{cov}(S)| < 2^{j'+1}$\;
Move $S$ from $\mathcal{L}_{j}$ to $\mathcal{L}_{j'}$\;
}
}
\label{ln:move:t}
}
\Fn{\textnormal{\textsc{Stabilize}$(\mathcal{C})$}\label{ln:stable:s}}{
\While{$\exists S \in \mathcal{S}$ and $\mathcal{L}_j$ s.t. $|S \cap A_{j}|
\geq 2^{j+1}$}{
$\mathtt{cov}(S) \gets \mathtt{cov}(S) \cup (S \cap A_{j})$, \textsc{ReLevel}($S$)\;
\While{$\exists S' \in \mathcal{C} : \mathtt{cov}(S) \cap \mathtt{cov}(S') \neq
\varnothing$}{
$\mathtt{cov}(S') \gets \mathtt{cov}(S') \setminus \mathtt{cov}(S)$,
\textsc{ReLevel}($S'$)\;
}
\label{ln:stable:t}
}
}
\end{algorithm}
We then describe our method for \emph{dynamic set cover} in Algorithm~\ref{alg:set:cover}.
First of all, we use \textsc{Greedy} to initialize a set-cover solution $\mathcal{C}$
on $\Sigma$ (Line~\ref{ln:cover:init}).
As shown in Lines~\ref{ln:greedy:s}--\ref{ln:greedy:t},
\textsc{Greedy} follows the classic greedy algorithm for set cover, and
the only difference is that all the sets in $\mathcal{C}$ are assigned to different levels
according to the sizes of their cover sets.
Then, the procedure of updating $\mathcal{C}$ for set operation $\sigma$ is shown
in Lines~\ref{ln:sigma:s}--\ref{ln:sigma:t}.
Our method supports four types of set operations to update $\Sigma$ as follows:
$\sigma=(u,S,\pm)$, i.e., to add/remove an element $u$ to/from a set $S \in \mathcal{S}$;
$\sigma=(u,\mathcal{U},\pm)$,
i.e., to add/remove an element $u$ to/from the universe $\mathcal{U}$.
We identify three cases in which the assignment of $u$ must be changed
for $\sigma$. When $\sigma=(u,S,-)$ and $\phi(u)=S$, it will reassign
$u$ to another set containing $u$; For $\sigma=(u,\mathcal{U},\pm)$,
it will add or delete the assignment of $u$ accordingly.
After that, for each set with some change in its cover set,
it calls \textsc{ReLevel} (e.g., Lines~\ref{ln:move:1},~\ref{ln:move:2}, and~\ref{ln:move:3})
to check whether the set should be moved to a new level based on the updated size
of its cover set.
The detailed procedure of \textsc{ReLevel} is given
in Lines~\ref{ln:move:s}--\ref{ln:move:t}.
Finally, \textsc{Stabilize} (Line~\ref{ln:sigma:t}) is always called
for every $\sigma$ to guarantee the stability of $\mathcal{C}$ since
$\mathcal{C}$ may become unstable due to the changes in $\Sigma$ and $\phi(u)$.
The procedure of stabilization is presented in Lines~\ref{ln:stable:s}--\ref{ln:stable:t}.
It finds all sets that violate
Condition~(2) of Definition~\ref{def:stable} and adjust $\mathcal{C}$
for these sets until no set should be adjusted anymore.
\noindent\textbf{Theoretical Analysis:}
Next, we will analyze Algorithm~\ref{alg:set:cover} theoretically.
We first show that a set-cover solution returned by \textsc{Greedy} is stable.
Then, we prove that \textsc{Stabilize} converges to a stable solution in finite steps.
\begin{lemma}\label{lm:stability}
The solution $\mathcal{C}$ returned by \textnormal{\textsc{Greedy}} is stable.
\end{lemma}
\begin{proof}
First of all, it is obvious that each set $S \in \mathcal{C}$ is assigned to the correct level
according to the size of its cover set and Condition (1) of Definition~\ref{def:stable} is satisfied.
Then, we sort the sets in $\mathcal{C}$ as $S^*_1,\ldots,S^*_{|\mathcal{C}|}$
by the order in which they are added.
Let $S^*_i$ be the set s.t.~$|\mathtt{cov}(S^*_i)| < 2^{j+1}$
and $|\mathtt{cov}(S^*_{i'})| \geq 2^{j+1}$ for any $i'<i$,
i.e., $S^*_i$ is the first set added to level $\mathcal{L}_j$.
We have $|I \cap S^*_i| = |\mathtt{cov}(S^*_i)| < 2^{j+1}$
where $I$ is the set of uncovered elements before $S^*_i$ is added to $\mathcal{C}$.
If there were a set $S \in \mathcal{S}$ such that $|S \cap A_j| > 2^{j+1}$,
we would acquire $|I \cap S| \geq |S \cap A_j| > 2^{j+1}$ and $|I \cap S|>|I \cap S^*_i|$,
which contradicts with Line~\ref{ln:cover:max} of Algorithm~\ref{alg:rms:init}.
Thus, $\mathcal{C}$ must satisfy Condition (2) of Definition~\ref{def:stable}.
To sum up, $\mathcal{C}$ is a stable solution.
\end{proof}
\begin{lemma}\label{lm:termination}
The procedure \textnormal{\textsc{Stabilize}} converges to a stable solution in $O(m\log{m})$ steps.
\end{lemma}
\begin{proof}
For an iteration of the \texttt{while} loop (i.e., Lines~\ref{ln:stable:s}--\ref{ln:stable:t})
that picks a set $S$ and a level $\mathcal{L}_j$,
the new level $\mathcal{L}_{j'}$ of $S$ always satisfies $j' > j$.
Accordingly, all the elements in $\mathtt{cov}(S)$ are moved
from $A_{\leq j}$ to $A_{j'}$. At the same time, no element in $A_{\geq j'}$
is moved to lower levels. Since each level contains at most $m$ elements ($|A_j| \leq m$),
\textsc{Stabilize} moves at most $m$ elements across $O(\log m)$ levels.
Therefore, it must terminate in $O(m \log m)$ steps.
Furthermore, after termination, the set-cover solution $\mathcal{C}$ must satisfy
both conditions in Definition~\ref{def:stable}.
Thus, we conclude the proof.
\end{proof}
The above two lemmas can guarantee that the set-cover solution provided by
Algorithm~\ref{alg:set:cover} is always stable after any change in the set system.
In the next subsection, we will present how to use it for fully-dynamic $k$-RMS.
\subsection{Algorithmic Description}\label{subsec:alg}
Next, we will present how FD-RMS maintains the $k$-RMS result
by always keeping a \emph{stable} set-cover solution on a dynamic set system
built from the approximate top-$k$ results over tuple insertions and deletions.
\begin{algorithm}[t]
\caption{\textsc{Initialization}}\label{alg:rms:init}
\Input{Query $\mathtt{RMS}(k,r)$, initial database $P_0$,
parameters $\varepsilon \in (0,1)$ and $M \in \mathbb{Z}^+$ ($M>r$)}
\Output{Result $Q_0$ of $\mathtt{RMS}(k,r)$ on $P_0$}
Draw $M$ vectors $\{u_i \in \mathbb{U} : i \in [1,M]\}$
where the first $d$ are the standard basis of $\mathbb{R}^d_+$
and the remaining are uniformly sampled from $\mathbb{U}$\;
Compute $\Phi_{k,\varepsilon}(u_i,P_0)$ of every $u_i$ where $i \in [1,M]$\;
\label{ln:rms:init:topk}
$L \gets r$, $H \gets M$, $m \gets (L+H)/2$\;\label{ln:rms:search:s}
\While{$\mathtt{true}$}{
\ForEach{$p \in P_0$\label{ln:rms:set:s}}{
$S(p) \gets \{u_i \,:\, i \in [1,m] \wedge p \in \Phi_{k,\varepsilon}(u_i,P_0) \}$\;
}
$\Sigma = (\mathcal{U},\mathcal{S})$ where $\mathcal{U}=\{u_i : i \in [1,m]\}$
and $\mathcal{S}=\{S(p) : p \in P_0\})$\;\label{ln:rms:set:t}
$\mathcal{C} \gets$ \textsc{Greedy}$(\Sigma)$\;
\uIf{$|\mathcal{C}| < r$}{
$L \gets m+1$, $m \gets (L+H)/2$\;
}
\uElseIf{$|\mathcal{C}| > r$}{
$H \gets m-1$, $m \gets (L+H)/2$\;
}
\ElseIf{$|\mathcal{C}| = r$ or $m = M$}{
\textbf{break}\;\label{ln:rms:search:t}
}
}
\Return{$Q_0 \gets \{ p \in P_0 \,:\, S(p) \in \mathcal{C} \}$}\;
\label{ln:rms:init:return}
\end{algorithm}
\noindent\textbf{Initialization:}
We first present how FD-RMS computes an initial result $Q_0$ for
$\mathtt{RMS}(k,r)$ on $P_0$ from scratch in Algorithm~\ref{alg:rms:init}.
There are two parameters in FD-RMS:
the approximation factor of top-$k$ results $\varepsilon$
and the upper bound of sample size $M$. The lower bound of sample size is set to $r$
because we can always find a set-cover solution of size equal to the size of the universe
(i.e., $m$ in FD-RMS).
First of all, it draws $M$ utility vectors $\{u_1,\ldots,u_M\}$, where the first $d$ vectors
are the standard basis of $\mathbb{R}^d_+$
and the remaining are uniformly sampled from $\mathbb{U}$,
and computes the $\varepsilon$-approximate top-$k$ result of each vector.
Subsequently, it finds an appropriate $m \in [r,M]$
so that the size of the set-cover solution on the set system $\Sigma$ built on
$\mathcal{U} = \{u_1,\ldots,u_m\}$ is exactly $r$.
The detailed procedure is as presented in Lines~\ref{ln:rms:search:s}--\ref{ln:rms:search:t}.
Specifically, it performs a binary search on range $[r,M]$ to determine the value of $m$.
For a given $m$, it first constructs a set system $\Sigma$ according to
Lines~\ref{ln:rms:set:s}--\ref{ln:rms:set:t}.
Next, it runs \textsc{Greedy} in Algorithm~\ref{alg:set:cover}
to compute a set-cover solution $\mathcal{C}$ on $\Sigma$.
After that, if $|\mathcal{C}| \neq r$ and $m < M$,
it will refresh the value of $m$ and rerun the above procedures;
Otherwise, $m$ is determined and the current set-cover solution $\mathcal{C}$
will be used to compute $Q_0$ for $\mathtt{RMS}(k,r)$.
Finally, it returns all the tuples whose corresponding sets are included in $\mathcal{C}$
as the result $Q_0$ for $\mathtt{RMS}(k,r)$ on $P_0$ (Line~\ref{ln:rms:init:return}).
\begin{algorithm}[t]
\caption{\textsc{Update}}\label{alg:rms:update}
\Input{Query $\mathtt{RMS}(k,r)$, database $P_{t-1}$, operation $\Delta_t$,
set-cover solution $\mathcal{C}$}
\Output{Result $Q_t$ for $\mathtt{RMS}(k,r)$ on $P_t$}
Update $P_{t-1}$ to $P_t$ w.r.t.~$\Delta_t$\;\label{ln:update:database}
\For{$i \gets 1,\ldots,M$}{
Update $\Phi_{k,\varepsilon}(u_i,P_{t-1})$ to $\Phi_{k,\varepsilon}(u_i,P_{t})$ w.r.t.~$\Delta_t$\;\label{ln:update:topk}
}
Maintain $\Sigma$ based on $\Phi_{k,\varepsilon}(u_i,P_{t})$\;
\label{ln:update:sets}
\uIf{$\Delta_t=\langle p,+ \rangle$\label{ln:insert:s}}{
\ForEach{$u \in S(p)$}{
\If{$u \in \mathtt{cov}(S(p'))$ and $u \notin S(p')$}{
Update $\mathcal{C}$ for $\sigma=(u,S(p'),-)$\;\label{ln:insert:t}
}
}
}
\ElseIf{$\Delta_t=\langle p,- \rangle$\label{ln:delete:s}}{
Delete $S(p)$ from $\mathcal{C}$ if $S(p) \in \mathcal{C}$\;
\ForEach{$u \in \mathtt{cov}(S(p))$}{
Update $\mathcal{C}$ for $\sigma=(u,S(p),-)$\;\label{ln:delete:t}
}
}
\If{$|\mathcal{C}| \neq r$}{
$m, \mathcal{C} \gets$ \textsc{UpdateM}$(\Sigma)$\;
}
\Return{$Q_t \gets \{ p \in P_t \,:\, S(p) \in \mathcal{C} \}$}\;
\end{algorithm}
\begin{figure*}[t]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{example-dataset.pdf}
\caption{Dataset}\label{fig:example:database}
\end{subfigure}
\hfill
\begin{subfigure}{0.225\textwidth}
\centering
\includegraphics[width=\textwidth]{example-P0.pdf}
\caption{Initial construction}\label{fig:example:fdrms:P0}
\end{subfigure}
\hfill
\begin{subfigure}{0.225\textwidth}
\centering
\includegraphics[width=\textwidth]{example-P1.pdf}
\caption{Add tuple $p_9$}\label{fig:example:fdrms:P1}
\end{subfigure}
\hfill
\begin{subfigure}{0.225\textwidth}
\centering
\includegraphics[width=\textwidth]{example-P2.pdf}
\caption{Delete tuple $p_1$}\label{fig:example:fdrms:P2}
\end{subfigure}
\caption{An example of using FD-RMS to process a $k$-RMS with $k=1$ and $r=3$}
\label{fig:example:fdrms}
\end{figure*}
\noindent\textbf{Update:}
The procedure of updating the result of $\mathtt{RMS}(k,r)$
w.r.t.~$\Delta_t$ is shown in Algorithm~\ref{alg:rms:update}.
First, it updates the database from $P_{t-1}$ to $P_t$ and
the approximate top-$k$ result from $\Phi_{k,\varepsilon}(u_i,P_{t-1})$
to $\Phi_{k,\varepsilon}(u_i,P_t)$ for each $u_i$ w.r.t.~$\Delta_t$
(Lines~\ref{ln:update:database}--\ref{ln:update:topk}).
Then, it also maintains the set system $\Sigma$
according to the changes in approximate top-$k$ results
(Line~\ref{ln:update:sets}).
Next, it updates the set-cover solution $\mathcal{C}$
for the changes in $\Sigma$ as follows.
\begin{itemize}
\item \textbf{Insertion:} The procedure of updating $\mathcal{C}$
w.r.t.~an insertion $\Delta_t=\langle p,+ \rangle$
is presented in Lines~\ref{ln:insert:s}--\ref{ln:insert:t}.
The changes in top-$k$ results lead to two updates in $\Sigma$:
(1) the insertion of $S(p)$ to $\mathcal{S}$
and (2) a series of deletions each of which
represents a tuple $p'$ is deleted from $\Phi_{k,\varepsilon}(u,P_t)$
due to the insertion of $p$.
For each deletion, it needs to check whether $u$ is previously
assigned to $S(p')$.
If so, it will update $\mathcal{C}$ by reassigning $u$ to a new set according to
Algorithm~\ref{alg:set:cover} because $u$ has been deleted from $S(p')$.
\item \textbf{Deletion:} The procedure of updating $\mathcal{C}$
w.r.t.~a deletion $ \Delta_t=\langle p,- \rangle $
is shown in Lines~\ref{ln:delete:s}--\ref{ln:delete:t}.
In contrast to an insertion, the deletion of $p$
leads to the removal of $S(p)$ from $\mathcal{S}$
and a series of insertions.
Thus, it must delete $S(p)$ from $\mathcal{C}$.
Next, it will reassign each $u \in \mathtt{cov}(S(p))$ to a new set
according to Algorithm~\ref{alg:set:cover}.
\end{itemize}
Then, it checks whether the size of $\mathcal{C}$ is still $r$.
If not, it will update the sample size $m$ and the universe $\mathcal{U}$ so that
the set-cover solution $\mathcal{C}$ consists of $r$ sets.
The procedure of updating $m$ and $\mathcal{U}$ as well as maintaining $\mathcal{C}$
on the updated $\mathcal{U}$ is shown in Algorithm~\ref{alg:rms:update:m}.
When $|\mathcal{C}|<r$, it will add new utility vectors from $u_{m+1}$, and so on,
to the universe and maintain $\mathcal{C}$ until $|\mathcal{C}|=r$ or $m=M$.
On the contrary, if $|\mathcal{C}| > r$,
it will drop existing utility vectors from $u_{m}$, and so on, from the universe
and maintain $\mathcal{C}$ until $|\mathcal{C}|=r$.
Finally, the updated $m$ and $\mathcal{C}$ are returned.
After all above procedures, it also returns $Q_t$
corresponding to the set-cover solution $\mathcal{C}$ on the updated $\Sigma$
as the result of $\mathtt{RMS}(k,r)$ on $P_t$.
\begin{algorithm}[t]
\caption{\textsc{UpdateM}$(\Sigma)$}\label{alg:rms:update:m}
\Output{Updated sample size $m$ and solution $\mathcal{C}$ on $\Sigma$}
\uIf{$|\mathcal{C}| < r$}{
\While{$m < M$ and $|\mathcal{C}| < r$}{
$m \gets m+1$, $\mathcal{U} \gets \mathcal{U} \cup \{u_m\}$\;
\ForEach{$p \in \Phi_{k,\varepsilon}(u_m,P_t)$}{
$S(p) \gets S(p) \cup \{u_m\}$\;
}
Update $\mathcal{C}$ for $\sigma=(u_m,\mathcal{U},+)$\;
}
}
\ElseIf{$|\mathcal{C}| > r$}{
\While{$|\mathcal{C}| > r$}{
$\mathcal{U} \gets \mathcal{U} \setminus \{u_m\}$\;
\ForEach{$p \in \Phi_{k,\varepsilon}(u_m,P_t)$}{
$S(p) \gets S(p) \setminus \{u_m\}$\;
}
Update $\mathcal{C}$ for $\sigma=(u_m,\mathcal{U},-)$\;
$m \gets m-1$\;
}
}
\Return{$m, \mathcal{C}$}\;
\end{algorithm}
\begin{example}
Fig.~\ref{fig:example:fdrms} illustrates an example of using FD-RMS to process a $k$-RMS
with $k=1$ and $r=3$. Here, we set $\varepsilon=0.002$ and $M=9$.
In Fig.~\ref{fig:example:fdrms:P0}, we show how to compute $Q_0$ for $\mathtt{RMS}(1,3)$
on $P_0=\{p_1,\ldots,p_8\}$. It first uses $m=(3+9)/2=6$ and runs
\textnormal{\textsc{Greedy}} to get
a set-cover solution $\mathcal{C}=\{S(p_1), S(p_2), S(p_4)\}$.
Since $|\mathcal{C}|=3$, it does not change $m$ anymore and returns
$Q_0=\{p_1,p_2,p_4\}$ for $\mathtt{RMS}(1,3)$ on $P_0$. Then,
the result of FD-RMS after the update procedures for~$\Delta_1 = \langle p_9,+ \rangle$ as
Algorithm~\ref{alg:rms:update} is shown in Fig.~\ref{fig:example:fdrms:P1}.
For $\mathtt{RMS}(1,3)$ on $P_1=\{p_1,\ldots,p_9\}$,
the result $Q_1$ is updated to $\{p_1,p_4,p_9\}$.
Finally, after the update procedures for $ \Delta_2 = \langle p_1,- \rangle $,
as shown in Fig.~\ref{fig:example:fdrms:P2}, $m$ is updated to $4$ and the result
$Q_2$ for $\mathtt{RMS}(1,3)$ on $P_2$ is $\{p_4,p_7,p_9\}$.
\end{example}
\noindent\textbf{Theoretical Bound:}
The theoretical bound of FD-RMS is analyzed as follows.
First of all, we need to verify the set-cover solution $\mathcal{C}$ maintained by
Algorithms~\ref{alg:rms:init}--\ref{alg:rms:update:m} is always stable.
According to Lemma~\ref{lm:stability}, it is guaranteed that the set-cover
solution $\mathcal{C}$ returned by Algorithm~\ref{alg:rms:init} is stable.
Then, we need to show it remains stable
after the update procedures of Algorithms~\ref{alg:rms:update} and~\ref{alg:rms:update:m}.
In fact, both algorithms use Algorithm~\ref{alg:set:cover} to maintain
the set-cover solution $\mathcal{C}$. Hence, the stability of $\mathcal{C}$
can be guaranteed by Lemma~\ref{lm:termination} since \textsc{Stabilize}
is always called after every update in Algorithm~\ref{alg:set:cover}.
Next, we indicate the relationship between the result of $k$-RMS
and the \emph{set-cover} solution
and provide the bound on the maximum-$k$ regret ratio of $Q_{t}$
returned by FD-RMS on $P_t$.
\begin{theorem}\label{thm:rms:ratio}
The result $Q_t$ returned by FD-RMS is
a $\big(k,O(\varepsilon^{*}_{k,r'}+\delta)\big)$-regret set of $P_t$ with high probability
where $r'=O(\frac{r}{\log{m}})$ and $\delta=O(m^{-\frac{1}{d}})$.
\end{theorem}
\begin{proof}
Given a parameter $\delta > 0$, a $\delta$-net~\cite{DBLP:conf/wea/AgarwalKSS17}
of $\mathbb{U}$ is a finite set $U \subset \mathbb{U}$ where there
exists a vector $\overline{u}$ with $\| u-\overline{u} \| \leq \delta$ for any $u \in \mathbb{U}$.
Since a random set of $O(\frac{1}{\delta^{d-1}}\log\frac{1}{\delta})$ vectors in $\mathbb{U}$
is a $\delta$-net with probability at least $\frac{1}{2}$~\cite{DBLP:conf/wea/AgarwalKSS17},
one can generate a $\delta$-net of size $O(\frac{1}{\delta^{d-1}}\log\frac{1}{\delta})$
with high probability by random sampling from $\mathbb{U}$ in $O(1)$ repeated trials.
Let $B$ be the standard basis of $\mathbb{R}^d_+$
and $U$ be a $\delta$-net of $\mathbb{U}$
where $B=\{u_1,\ldots,u_d\} \subset U$.
Since $p \in P_t$ is scaled to $p[i] \leq 1$ for $i \in [1,d]$,
we have $\| p \| \leq \sqrt{d}$.
According to the definition of $\delta$-net,
there exists a vector $\overline{u} \in U$
such that $\| \overline{u} - u \| \leq \delta$ for every $u \in \mathbb{U}$.
Hence, for any tuple $p \in P_t$,
\begin{equation}\label{eq:1}
| \langle \overline{u},p \rangle - \langle u,p \rangle |
= | \langle \overline{u} - u,p \rangle |
\leq \| \overline{u} - u \| \cdot \| p \| \leq \delta \cdot \sqrt{d}
\end{equation}
Moreover, as $Q_{t}$ corresponds to a set-cover solution $\mathcal{C}$
on $\Sigma$, there exists a tuple $q \in Q_{t}$ such that
$\langle \overline{u},q \rangle \geq (1-\varepsilon) \cdot \omega_{k}(\overline{u},P_t)$
for any $\overline{u} \in U$.
We first consider a basis vector $u_i \in U$ for some $i \in [1,d]$.
We have $\omega(u_i,Q_t) \geq (1-\varepsilon) \cdot \omega_{k}(u_i,P_t)$ and
thus $ \omega(u_i,Q_t) \geq (1-\varepsilon) \cdot c $ where
$c = \min_{i \in [1,d]} \omega_{k}(u_i,P_t)$.
Since $\| u \| = 1$, there must exist some $i$ with $u[i] \geq \frac{1}{\sqrt{d}}$
for any $u \in \mathbb{U}$.
Therefore, it holds that $\omega(u,Q_{t}) \geq \omega(u_i,Q_{t}) \cdot \frac{1}{\sqrt{d}}
\geq (1-\varepsilon)\cdot\frac{c}{\sqrt{d}}$
for any $u \in \mathbb{U}$.
Next, we discuss two cases for $u \in \mathbb{U}$ separately.
\begin{itemize}
\item \textbf{Case 1 ($\omega_k(u,P_t) \leq \frac{c}{\sqrt{d}}$):}
In this case, there always exists $q \in Q_{t}$ such that
$\langle u,q \rangle \geq (1-\varepsilon)\cdot\omega_k(u,P_t)$.
\item \textbf{Case 2 ($\omega_k(u,P_t) > \frac{c}{\sqrt{d}}$):}
Let $\overline{u} \in U$ be the utility vector such that
$\| \overline{u} - u \| \leq \delta$.
Let $\Phi_{k}(u,P_t)=\{p_1,\ldots,p_k\}$ be the top-$k$ results of $u$ on $P_t$.
According to Equation~\ref{eq:1}, we have
$\langle \overline{u},p_i \rangle \geq \langle u,p_i \rangle
- \delta\cdot\sqrt{d}$ for all $i \in [1,k]$ and thus
$\langle \overline{u},p_i \rangle \geq \omega_{k}(u,P_t)
- \delta\cdot\sqrt{d}$. Thus, there exists $k$ tuples in $P_t$
with scores at least $\omega_{k}(u,P_t) - \delta\cdot\sqrt{d}$ for $\overline{u}$.
We can acquire
$\omega_{k}(\overline{u},P_t) \geq \omega_{k}(u,P_t) - \delta\cdot\sqrt{d}$.
Therefore, there exists $q \in Q_{t}$ such that
\begin{align*}
\langle u,q \rangle & \geq \langle \overline{u},q \rangle - \delta\cdot\sqrt{d}
\geq (1-\varepsilon) \cdot \omega_k(\overline{u},P_t) - \delta\cdot\sqrt{d} \\
& \geq (1-\varepsilon) \cdot \big(\omega_{k}(u,P_t) - \delta\cdot\sqrt{d}\big)
- \delta\cdot\sqrt{d} \\
& \geq \big(1-\varepsilon-\frac{(1-\varepsilon)d\delta}{c}-\frac{d\delta}{c}\big)
\cdot \omega_{k}(u,P_t) \\
& \geq (1-\varepsilon-\frac{2d\delta}{c})\cdot\omega_{k}(u,P_t)
\end{align*}
\end{itemize}
Considering both cases, we have $\omega(u,Q_t) \geq
(1-\varepsilon-\frac{2d\delta}{c})\cdot\omega_{k}(u,P_t)$ for any $u \in \mathbb{U}$
and thus $\mathtt{mrr}_k(Q_{t})$ over $P_t$ is at most $\varepsilon+\frac{2d\delta}{c}$.
In all of our experiments, the value of $c$ is always between $0.5$ and $1$,
and thus we regard $c$ as a constant in this proof.
Therefore, $Q_t$ is a $\big(k,O(\varepsilon+\delta)\big)$-regret set of $P_t$
with high probability for any $c,d=O(1)$.
Moreover, since FD-RMS uses $m$ utility vectors including $B$ to compute $Q_t$
and $m=O(\frac{1}{\delta^{d-1}}\log\frac{1}{\delta})$,
we can acquire $\delta=O(m^{-\frac{1}{d}})$.
Finally, because any $(k,\varepsilon)$-regret set of $P_t$ corresponds to a set-cover
solution on $\Sigma$
(otherwise, the regret ratio is larger than $\varepsilon$ for some utility vector)
and the size of the optimal set-cover solution on $\Sigma$ is $O(\frac{r}{\log{m}})$
according to Theorem~\ref{thm:approx:ratio},
the maximum $k$-regret ratio of any size-$r'$ subset of $P_t$ is at least $\varepsilon$
where $r'=O(\frac{r}{\log{m}})$, i.e., $\varepsilon^{*}_{k,r'} \geq \varepsilon$.
Therefore, we conclude that $Q_t$ is a $\big(k,O(\varepsilon^{*}_{k,r'}+\delta)\big)$-regret set of $P_t$
with high probability.
\end{proof}
Finally, the upper bound of the maximum $k$-regret ratio of $Q_{t}$ returned by FD-RMS
on $P_t$ is analyzed in the following corollary derived from the result of Theorem~\ref{thm:rms:ratio}.
\begin{corollary}\label{col:rms:bound}
It satisfies that $\mathtt{mrr}_k(Q_{t}) = O(r^{-\frac{1}{d}})$ with high probability
if we assume $\varepsilon = O(m^{-\frac{1}{d}})$.
\end{corollary}
\begin{proof}
As indicated in the proof of Theorem~\ref{thm:rms:ratio},
$\mathcal{U}=\{u_1,u_2,$ $\ldots,u_{m}\}$ is a $\delta$-net of $\mathbb{U}$
where $\delta=O(m^{-\frac{1}{d}})$ with high probability.
Moreover, we have $\mathtt{mrr}_k(Q_{t}) = O(\varepsilon + \delta)$
and thus $\mathtt{mrr}_k(Q_{t}) = O(m^{-\frac{1}{d}})$
if $\varepsilon = O(m^{-\frac{1}{d}})$.
In addition, at any time, $\mathcal{U}$ must have at least $r$ utility vectors,
i.e., $m \geq r$.
Thus, we have $\mathtt{mrr}_k(Q_{t}) = O(r^{-\frac{1}{d}})$
since $m^{-\frac{1}{d}} \leq r^{-\frac{1}{d}}$ for any $d>1$
and conclude the proof.
\end{proof}
Since $\varepsilon$ is tunable in FD-RMS, by trying different values of $\varepsilon$,
we can always find an appropriate one such that $\varepsilon = O(m^{-\frac{1}{d}})$.
Hence, from Corollary~\ref{col:rms:bound}, we show that
the upper bound of FD-RMS is slightly higher than that of
\textsc{Cube}~\cite{DBLP:journals/pvldb/NanongkaiSLLX10} and
\textsc{Sphere}~\cite{DBLP:conf/sigmod/XieW0LL18} (i.e., $O(r^{-\frac{1}{d-1}}$))
under a mild assumption.
\noindent\textbf{Complexity Analysis:}
First, we use tree-based methods to maintain
the approximate top-$k$ results for FD-RMS (see Section~\ref{subsec:impl} for details).
Here, the time complexity of each top-$k$ query is $O(n_0)$ where $n_0=|P_0|$
because the size of $\varepsilon$-approximate top-$k$ tuples can be $O(n_0)$.
Hence, it takes $O(M \cdot n_0)$ time to compute the top-$k$ results.
Then, \textsc{Greedy} runs $O(r)$ iterations to get a set-cover solution.
At each iteration, it evaluates $O(n_0)$ sets to find $S^*$ in Line~\ref{ln:cover:max}
of Algorithm~\ref{alg:set:cover}.
Thus, the time complexity of \textsc{Greedy} is $O(r \cdot n_0)$.
FD-RMS calls \textsc{Greedy} $O(\log{M})$ times to determine the value of $m$.
Therefore, the time complexity of computing $Q_{0}$ on $P_0$
is $O\big((M + r\log{M}) \cdot n_0\big)$.
In Algorithm~\ref{alg:rms:update},
the time complexity of updating the top-$k$ results and set system $\Sigma$
is $O\big(\mathtt{u}(\Delta_t)\cdot n_t \big)$ where $\mathtt{u}(\Delta_t)$
is the number of utility vectors whose top-$k$ results are changed by $\Delta_t$.
Then, the maximum number of reassignments in cover sets is
$|S(p)|$ for $\Delta_t$, which is bounded by $O(\mathtt{u}(\Delta_t))$.
In addition, the time complexity of \textsc{Stabilize} is $O(m \log{m})$ according to Lemma~\ref{lm:termination}.
Moreover, the maximum difference between the old and new values of $m$ is bounded by $O(m)$.
Hence, the total time complexity of updating $Q_t$ w.r.t.~$\Delta_t$
is $O\big(\mathtt{u}(\Delta_t) \cdot n_t + m^2 \log{m}\big)$.
\subsection{Implementation Issues}\label{subsec:impl}
\noindent\textbf{Index Structures:}
As indicated in Line~\ref{ln:rms:init:topk} of Algorithm~\ref{alg:rms:init}
and Line~\ref{ln:update:topk} of Algorithm~\ref{alg:rms:update},
FD-RMS should compute the $\varepsilon$-approximate top-$k$ result of each $u_i$ ($i \in [1,M]$)
on $P_0$ and update it w.r.t.~$\Delta_t$.
Here, we elaborate on our implementation for top-$k$ maintenance.
In order to process a large number of (approximate) top-$k$ queries with frequent updates in the database,
we implement a \emph{dual-tree}~\cite{DBLP:conf/kdd/RamG12,DBLP:conf/sigmod/YuAY12,DBLP:conf/sdm/CurtinGR13}
that comprises a tuple index $\mathtt{TI}$ and a utility index $\mathtt{UI}$.
The goal of $\mathtt{TI}$ is to efficiently retrieve the $\varepsilon$-approximate
top-$k$ result $\Phi_{k,\varepsilon}(u,P_t)$ of any utility vector $u$ on the up-to-date $P_t$.
Hence, any space-partitioning index, e.g.,
\emph{k-d tree}~\cite{DBLP:journals/cacm/Bentley75}
and \emph{Quadtree}~\cite{DBLP:journals/acta/FinkelB74},
can serve as $\mathtt{TI}$ for top-$k$ query processing.
In practice, we use \emph{k-d tree} as $\mathtt{TI}$.
We adopt the scheme of~\cite{DBLP:conf/recsys/BachrachFGKKNP14}
to transform a top-$k$ query in $\mathbb{R}^d$
into a $k$NN query in $\mathbb{R}^{d+1}$.
Then, we implement the standard top-down methods to construct $\mathtt{TI}$ on $P_0$
and update it w.r.t.~$\Delta_t$.
The branch-and-bound algorithm is used for top-$k$ queries on $\mathtt{TI}$.
The goal of $\mathtt{UI}$ is to cluster the sampled utility vectors
so as to efficiently find each vector whose $\varepsilon$-approximate
top-$k$ result is updated by~$\Delta_t$.
Since the top-$k$ results of linear functions are merely determined by directions,
the basic idea of $\mathtt{UI}$ is to cluster the utilities with high \emph{cosine similarities} together.
Therefore, we adopt an angular-based binary space partitioning tree
called \emph{cone tree}~\cite{DBLP:conf/kdd/RamG12} as $\mathtt{UI}$.
We generally follow Algorithms 8--9 in~\cite{DBLP:conf/kdd/RamG12} to build $\mathtt{UI}$
for $\{u_1,\ldots,u_M\}$.
We implement a top-down approach based on Section 3.2 of~\cite{DBLP:conf/sigmod/YuAY12}
to update the top-$k$ results affected by~$\Delta_t$.
\noindent\textbf{Parameter Tuning:}
Now, we discuss how to specify the values of $\varepsilon$, i.e., the approximation factor of
top-$k$ queries, and $M$, i.e., the upper bound of $m$, in FD-RMS.
In general, the value of $\varepsilon$ has direct effect on $m$
as well as the efficiency and quality of results of FD-RMS.
In particular, if $\varepsilon$ is larger, the $\varepsilon$-approximate
top-$k$ result of each utility vector will include more tuples and the set system
built on top-$k$ results will be more dense. As a result, to guarantee
the result size to be exactly $r$, FD-RMS will use more utility vectors (i.e., a larger $m$)
for a larger $\varepsilon$.
Therefore, a smaller $\varepsilon$ leads to higher efficiency and lower solution quality
due to smaller $m$ and larger $\delta$, and vice versa.
In our implementation, we use a trial-and-error method to find
appropriate values of $\varepsilon$ and $M$:
For each query $\mathtt{RMS}(k,r)$ on a dataset, we test different values of
$\varepsilon$ chosen from $[0.0001, \ldots, 0.1024]$ and,
for each value of $\varepsilon$, $M$ is set to the smallest one
chosen from $[2^{10},\ldots,2^{20}]$ that always guarantees $m < M$.
If the result size is still smaller than $r$ when $m=M=2^{20}$,
we will not use larger $M$ anymore due to efficiency issue.
The values of $\varepsilon$ and $M$ that strike
the best balance between efficiency and quality of results will be used.
In Fig.~\ref{fig:eps}, we present how the value of $\varepsilon$
affects the performance of FD-RMS empirically.
\section{Related Work}
\label{sec:related:work}
There have been extensive studies on the $k$-regret minimizing set ($k$-RMS) problem
(see~\cite{DBLP:journals/vldb/XieWL20} for a survey).
Nanongkai et al.~\cite{DBLP:journals/pvldb/NanongkaiSLLX10}
first introduced the notions of \emph{maximum regret ratio} and \emph{$r$-regret query}
(i.e., maximum $1$-regret ratio and $1$-RMS in this paper).
They proposed the \textsc{Cube} algorithm to provide an upper-bound guarantee
for the maximum regret ratio of the optimal solution of $1$-RMS.
They also proposed the \textsc{Greedy} heuristic for $1$-RMS, which always picked a tuple that
maximally reduced the maximum regret ratio at each iteration.
Peng and Wong~\cite{DBLP:conf/icde/PengW14} proposed the \textsc{GeoGreedy} algorithm to
improve the efficiency of \textsc{Greedy} by utilizing the geometric properties of $1$-RMS.
Asudeh et al.~\cite{DBLP:conf/sigmod/AsudehN0D17} proposed two discretized matrix min-max (DMM)
based algorithms for $1$-RMS. Xie et al.~\cite{DBLP:conf/sigmod/XieW0LL18}
designed the \textsc{Sphere} algorithm
for $1$-RMS based on the notion of $\varepsilon$-kernel~\cite{DBLP:journals/jacm/AgarwalHV04}.
Shetiya et al.~\cite{Shetiya2019} proposed a unified algorithm called URM
based on $k$-\textsc{Medoid} clustering for $l^p$-norm RMS
problems, of which $1$-RMS was a special case when $p=\infty$.
The aforementioned algorithms cannot be used for $k$-RMS when $k > 1$.
Chester et al.~\cite{DBLP:journals/pvldb/ChesterTVW14} first extended
the notion of $1$-RMS to $k$-RMS.
They also proposed a randomized \textsc{Greedy}$^*$ algorithm
that extended the \textsc{Greedy} heuristic to support $k$-RMS when $k > 1$.
The min-size version of $k$-RMS that returned the minimum subset whose maximum $k$-regret ratio
was at most $\varepsilon$ for a given $\varepsilon \in (0,1)$ was studied
in~\cite{DBLP:conf/wea/AgarwalKSS17,DBLP:conf/alenex/KumarS18}.
They proposed two algorithms for min-size $k$-RMS based on the notion of
$\varepsilon$-kernel~\cite{DBLP:journals/jacm/AgarwalHV04} and hitting-set,
respectively. However, all above algorithms are designed for the static setting and
very inefficient to process database updates.
To the best of our knowledge, FD-RMS is the first $k$-RMS algorithm that is optimized for
the fully dynamic setting and efficiently maintains the result for dynamic updates.
Different variations of regret minimizing set problems were also studied recently.
The $1$-RMS problem with nonlinear utility functions were studied
in~\cite{DBLP:journals/pvldb/FaulknerBL15,DBLP:journals/tods/QiZSY18,DBLP:conf/aaai/SomaY17a}.
Specifically, they generalized the class of utility functions to
\emph{convex functions}~\cite{DBLP:journals/pvldb/FaulknerBL15},
\emph{multiplicative functions}~\cite{DBLP:journals/tods/QiZSY18},
and \emph{submodular functions}~\cite{DBLP:conf/aaai/SomaY17a}, respectively.
Asudeh et al.~\cite{DBLP:conf/sigmod/AsudehN00J19} proposed the
rank-regret representative (RRR) problem.
The difference between RRR and RMS is that
the regret in RRR is defined by ranking while the regret in RMS is defined by score.
Several studies~\cite{DBLP:conf/icde/ZeighamiW19,DBLP:conf/aaai/StorandtF19,Shetiya2019}
investigated the \emph{average regret minimization} (ARM) problem.
Instead of minimizing the maximum regret ratio,
ARM returns a subset of $r$ tuples such that the average regret of all possible users is minimized.
The problem of \emph{interactive regret minimization} that aimed to enhance the regret minimization
problem with user interactions was studied
in~\cite{DBLP:conf/sigmod/NanongkaiLSM12,DBLP:conf/sigmod/XieWL19}.
Xie et al.~\cite{Xie2020} proposed a variation of min-size RMS
called $\alpha$-happiness query.
Since these variations have different formulations from the original $k$-RMS problem,
the algorithms proposed for them cannot be directly applied to the $k$-RMS problem.
Moreover, these algorithms are still proposed for the static setting
without considering database updates.
\section{Preliminaries}
\label{sec:definition}
In this section, we formally define the problem we study in this paper. We first introduce
the notion of maximum $k$-regret ratio. Then, we formulate the
$k$-regret minimizing set ($k$-RMS) problem in a fully dynamic setting.
Finally, we present the challenges of solving fully-dynamic $k$-RMS.
\begin{figure}
\centering
\includegraphics[width=.32\textwidth]{data.pdf}
\caption{A two-dimensional database of 8 tuples.}
\label{fig:database}
\end{figure}
\subsection{Maximum K-Regret Ratio}
Let us consider a database $P$ where each tuple $p \in P$
has $d$ nonnegative numerical attributes $p[1],\ldots,p[d]$
and is represented as a point in the nonnegative orthant $\mathbb{R}_+^d$.
A user's preference is denoted by a utility function
$f:\mathbb{R}_+^d \rightarrow \mathbb{R}^+$
that assigns a positive score $f(p)$ to each tuple $p$.
Following~\cite{DBLP:journals/pvldb/NanongkaiSLLX10,DBLP:conf/icde/PengW14,DBLP:journals/pvldb/ChesterTVW14,DBLP:conf/sigmod/AsudehN0D17,DBLP:conf/sigmod/XieW0LL18},
we restrict the class of utility functions to \emph{linear functions}.
A function $f$ is linear if and only if there exists
a $d$-dimensional vector $u=(u[1],\ldots,u[d]) \in \mathbb{R}_+^d$
such that $f(p)=\langle u,p \rangle =\sum_{i=1}^{d} u[i] \cdot p[i]$
for any $p \in \mathbb{R}_+^d$.
W.l.o.g., we assume the range of values on each dimension is scaled to $[0,1]$ and
any utility vector is normalized to be a unit\footnote{The normalization does not affect our results
because the maximum $k$-regret ratio is \emph{scale-invariant}~\cite{DBLP:journals/pvldb/NanongkaiSLLX10}.},
i.e., $\| u \|=1$.
Intuitively, the class of linear functions corresponds to the nonnegative orthant of
$d$-dimensional unit sphere
$ \mathbb{U}=\{ u \in \mathbb{R}_+^d \,:\, \lVert u \rVert=1 \} $.
We use $\varphi_j(u,P)$ to denote the tuple $p \in P$ with the $j$\textsuperscript{th}-largest
score w.r.t.~vector $u$ and $\omega_j(u,P)$ to denote its score.
Note that multiple tuples may have the same score w.r.t.~$u$
and any consistent rule can be adopted to break ties.
For brevity, we drop the subscript $j$ from the above notations when $j=1$,
i.e., $\varphi(u,P)=\arg\max_{p \in P} \langle u,p \rangle$
and $\omega(u,P)=\max_{p \in P} \langle u,p \rangle$.
The top-$k$ tuples in $P$ w.r.t.~$u$ is represented as
$\Phi_{k}(u,P)=\{\varphi_{j}(u,P) : 1 \leq j \leq k\}$.
Given a real number $\varepsilon \in (0,1)$,
the $\varepsilon$-approximate top-$k$ tuples in $P$ w.r.t.~$u$ is denoted
as $\Phi_{k,\varepsilon}(u,P)=\{p \in P \,:\, \langle u,p \rangle \geq (1-\varepsilon)\cdot\omega_{k}(u,P)\}$,
i.e., the set of tuples whose scores are at least $(1-\varepsilon)\cdot\omega_{k}(u,P)$.
For a subset $Q \subseteq P$ and an integer $k \geq 1$,
we define the \emph{$k$-regret ratio} of $Q$ over $P$ for a utility vector $u$ by
$\mathtt{rr}_k(u,Q)=\max\big(0,1-\frac{\omega(u,Q)}{\omega_{k}(u,P)}\big)$,
i.e., the relative loss of replacing the $k$\textsuperscript{th}-ranked tuple in $P$ by
the top-ranked tuple in $Q$.
Since it is required to consider the preferences of all possible users, our goal is to find
a subset whose $k$-regret ratio is small for an arbitrary utility vector.
Therefore, we define the \emph{maximum $k$-regret ratio} of $Q$ over $P$
by $\mathtt{mrr}_k(Q)=\max_{u \in \mathbb{U}}\mathtt{rr}_k(u,Q)$.
Intuitively, $\mathtt{mrr}_k(Q)$ measures how well the top-ranked tuple of $Q$
approximates the $k$\textsuperscript{th}-ranked tuple of $P$ in the worst case.
For a real number $\varepsilon \in (0,1)$,
$Q$ is said to be a $(k,\varepsilon)$-regret set of $P$ iff $\mathtt{mrr}_k(Q) \leq \varepsilon$,
or equivalently, $\varphi(u,Q) \in \Phi_{k,\varepsilon}(u,P)$ for any $u \in \mathbb{U}$.
By definition, it holds that $\mathtt{mrr}_k(Q) \in [0,1]$.
\begin{example}
Fig.~\ref{fig:database} illustrates a database $P$ in $\mathbb{R}^2_+$
with $8$ tuples $\{p_1,\ldots,p_8\}$.
For utility vectors $u_1=(0.42,0.91)$ and $u_2=(0.91,0.42)$,
their top-$2$ results are $\Phi_2(u_1,P)=\{p_1,p_2\}$ and $\Phi_2(u_2,P)=\{p_2,p_4\}$, respectively.
Given a subset $Q_1=\{p_3,p_4\}$ of $P$, $\mathtt{rr}_2(u_1,Q_1)=1-\frac{0.749}{0.98} \approx 0.236$
as $\omega(u_1,Q_1)=\langle u_1,p_3 \rangle=0.749$ and $\omega_2(u_1,P)=\langle u_1,p_2 \rangle=0.98$.
Furthermore, $\mathtt{mrr}_2(Q_1) \approx 0.444$
because $\mathtt{rr}_2(u,Q_1)$ is the maximum when $u=(0.0,1.0)$
with $\mathtt{rr}_2(u,Q_1) = 1-\frac{5}{9} \approx 0.444$.
Finally, $Q_2=\{p_1,p_2,p_4\}$ is a $(2,0)$-regret set of $P$ since $\mathtt{mrr}_2(Q_2)=0$.
\end{example}
\subsection{K-Regret Minimizing Set}
Based on the notion of \emph{maximum $k$-regret ratio},
we can formally define the $k$-\emph{regret minimizing set}
($k$-RMS) problem in the following.
\begin{definition}[$k$-Regret Minimizing Set]\label{def:k-rms}
Given a database $P \subset \mathbb{R}_+^d$ and a size constraint $r \in \mathbb{Z}^+$
($r \geq d$), the $k$-regret minimizing set ($k$-RMS) problem
returns a subset $Q^{*} \subseteq P$ of at most $r$ tuples
with the smallest \emph{maximum $k$-regret ratio},
i.e., $Q^{*}=\arg\min_{Q \subseteq P \,:\, |Q| \leq r} \mathtt{mrr}_k(Q)$.
\end{definition}
For any given $k$ and $r$, we denote the $k$-RMS problem by $\mathtt{RMS}(k,r)$
and the \emph{maximum $k$-regret ratio} of the optimal result $Q^{*}$ for $\mathtt{RMS}(k,r)$
by $\varepsilon^{*}_{k,r}$.
In particular, the $r$-regret query studied
in~\cite{DBLP:conf/icde/PengW14,DBLP:conf/sigmod/AsudehN0D17,DBLP:conf/sigmod/XieW0LL18,
DBLP:journals/pvldb/NanongkaiSLLX10}
is a special case of our $k$-RMS problem when $k=1$, i.e., $1$-RMS.
\begin{example}
Let us continue with the example in Fig.~\ref{fig:database}.
For a query $\mathtt{RMS}(2,2)$ on $P$, we have $Q^*=\{p_1,p_4\}$
with $\varepsilon^*_{2,2}=\mathtt{mrr}_2(Q^*) \approx 0.05$
because $\{p_1,p_4\}$ has the smallest maximum $2$-regret ratio
among all size-$2$ subsets of $P$.
\end{example}
In this paper, we focus on the fully-dynamic $k$-RMS problem.
We consider an initial database $P_0$ and a (possibly countably infinite) sequence of operations
$\Delta = \langle \Delta_1,\Delta_2,\ldots \rangle$.
At each timestamp $t$ ($t \in \mathbb{Z}^+$), the database is updated from $P_{t-1}$
to $P_t$ by performing an operation $\Delta_t$ of one of the following two types:
\begin{itemize}
\item \textbf{Tuple insertion} $\Delta_t=\langle p,+ \rangle$:
add a new tuple $p$ to $P_{t-1}$, i.e., $P_t \gets P_{t-1}\cup\{p\}$;
\item \textbf{Tuple deletion} $\Delta_t=\langle p,- \rangle$:
delete an existing tuple $p$ from $P_{t-1}$, i.e., $P_t \gets P_{t-1}\setminus\{p\}$.
\end{itemize}
Note that the update of a tuple can be processed by a deletion followed by an insertion,
and thus is not discussed separately in this paper.
Given an initial database $P_0$, a sequence of operations $\Delta$,
and a query $\mathtt{RMS}(k,r)$,
we aim to keep track of the result $Q^{*}_t$ for $\mathtt{RMS}(k,r)$ on $P_t$ at any time $t$.
Fully-dynamic $k$-RMS faces two challenges.
First, the $k$-RMS problem is
NP-hard~\cite{DBLP:conf/icdt/CaoLWWWWZ17,DBLP:journals/pvldb/ChesterTVW14,DBLP:conf/wea/AgarwalKSS17}
for any $d \geq 3$. Thus, the optimal solution of $k$-RMS is intractable
for any database with three or more attributes unless P=NP in both static and dynamic settings.
Hence, we will focus on maintaining an approximate result of $k$-RMS in this paper.
Second, existing $k$-RMS algorithms can only work in the static setting.
They must recompute the result from scratch once an operation
triggers any update in the skyline
(Note that since the result of $k$-RMS is a subset of the
skyline~\cite{DBLP:journals/pvldb/NanongkaiSLLX10,DBLP:journals/pvldb/ChesterTVW14},
it remains unchanged for any operation on non-skyline tuples).
However, frequent recomputation leads to significant overhead
and causes low efficiency on highly dynamic databases.
Therefore, we will propose a novel method for fully-dynamic $k$-RMS
that can maintain a high-quality result for $\mathtt{RMS}(k,r)$
on a database w.r.t.~any tuple insertion and deletion efficiently.
\section{Conclusion}\label{sec:conclusion}
In this paper, we studied the problem of maintaining $k$-regret minimizing sets
($k$-RMS) on dynamic datasets with arbitrary insertions and deletions of tuples.
We proposed the first fully-dynamic $k$-RMS algorithm called FD-RMS.
FD-RMS was based on transforming fully-dynamic $k$-RMS to a dynamic set cover problem,
and it could dynamically maintain the result of $k$-RMS with a theoretical guarantee.
Extensive experiments on real-world and synthetic datasets confirmed the efficiency,
effectiveness, and scalability of FD-RMS compared with existing static approaches to $k$-RMS.
For future work, it would be interesting to investigate whether our techniques
can be extended to $k$-RMS and related problems on higher dimensions (i.e.,~$d>10$)
or with nonlinear utility functions
(e.g.,~\cite{DBLP:journals/pvldb/FaulknerBL15,DBLP:journals/tods/QiZSY18,DBLP:conf/aaai/SomaY17a})
in dynamic settings.
\section{Experiments}\label{sec:exp}
In this section, we evaluate the performance of FD-RMS on real-world and synthetic datasets.
We first introduce the experimental setup in Section~\ref{subsec:setup}.
Then, we present the experimental results in Section~\ref{subsec:results}.
\subsection{Experimental Setup}\label{subsec:setup}
\noindent\textbf{Algorithms:} The algorithms compared are listed as follows.
\begin{itemize}
\item \textsc{Greedy}: the greedy algorithm for $1$-RMS in~\cite{DBLP:journals/pvldb/NanongkaiSLLX10}.
\item \textsc{Greedy}$^*$: the randomized greedy algorithm for $k$-RMS when $k>1$
proposed in~\cite{DBLP:journals/pvldb/ChesterTVW14}.
\item \textsc{GeoGreedy}: a variation of \textsc{Greedy} for $1$-RMS in~\cite{DBLP:conf/icde/PengW14}.
\item DMM-RRMS: a discretized matrix min-max based algorithm for $1$-RMS
in~\cite{DBLP:conf/sigmod/AsudehN0D17}.
\item $\varepsilon$-\textsc{Kernel}: computing an $\varepsilon$-kernel coreset
as the $k$-RMS result~\cite{DBLP:conf/wea/AgarwalKSS17,DBLP:conf/icdt/CaoLWWWWZ17} directly.
\item HS: a hitting-set based algorithm for $k$-RMS in~\cite{DBLP:conf/wea/AgarwalKSS17}.
\item \textsc{Sphere}: an algorithm that combines $\varepsilon$-\textsc{Kernel}
with \textsc{Greedy} for $1$-RMS in~\cite{DBLP:conf/sigmod/XieW0LL18}.
\item URM: a $k$-\textsc{medoid} clustering based algorithm for $1$-RMS in~\cite{Shetiya2019}.
Following~\cite{Shetiya2019}, we use DMM-RRMS to compute an initial solution for URM.
\item FD-RMS: our fully-dynamic $k$-RMS algorithm proposed in Section~\ref{sec:framework}.
\end{itemize}
The algorithms that only work in two dimensions are not compared.
All the above algorithms except FD-RMS and URM\footnote{
The original URM in~\cite{Shetiya2019} is also a static algorithm.
We extend URM to support dynamic updates as described in
Algorithm~\ref{alg:urm} of Appendix~\ref{appendix:dynamic:urm}.}
cannot directly work in a fully dynamic setting.
In our experiments, they rerun from scratch to compute the up-to-date $k$-RMS result
once the skyline is updated by any insertion or deletion.
In addition, the algorithms that are not applicable when $k>1$
are not compared for the experiments with varying $k$.
Since $\varepsilon$-\textsc{Kernel} and HS are proposed for min-size
$k$-RMS that returns the smallest subset whose maximum $k$-regret ratio is at most $\varepsilon$,
we adapt them to our problem by performing a binary search on $\varepsilon$ in the range
$(0,1)$ to find the smallest value of $\varepsilon$
that guarantees the result size is at most $r$.
Our implementation of FD-RMS and URM is in Java 8 and published on
GitHub\footnote{https://github.com/yhwang1990/dynamic-rms}.
We used the C++ implementations of baseline algorithms published by authors
and followed the default parameter settings as described in the original papers.
All the experiments were conducted on a server running
Ubuntu 18.04.1 with a 2.3GHz processor and 256GB memory.
\noindent\textbf{Datasets:}
The datasets we use are listed as follows.
\begin{itemize}
\item \textbf{BB}\footnote{\url{www.basketball-reference.com}}
is a basketball dataset that contains $21,961$ tuples, each of which represents one player/season combination
with $5$ attributes such as \emph{points} and \emph{rebounds}.
\item \textbf{AQ}\footnote{\url{archive.ics.uci.edu/ml/datasets/Beijing+Multi-Site+Air-Quality+Data}}
includes hourly air-pollution and weather data from 12 monitoring sites in Beijing.
It has $382,168$ tuples and each tuple has $9$ attributes including the concentrations of $6$ air pollutants like
$\text{PM}_{\text{2.5}}$, as well as $3$ meteorological parameters
like \emph{temperature}.
\item \textbf{CT}\footnote{\url{archive.ics.uci.edu/ml/datasets/covertype}}
contains the cartographic data of forest covers in the Roosevelt National Forest of northern Colorado.
It has $581,012$ tuples and we choose $8$ numerical attributes,
e.g., \emph{elevation} and \emph{slope}, for evaluation.
\item \textbf{Movie}\footnote{\url{grouplens.org/datasets/movielens}}
is the tag genome dataset published by MovieLens.
We extract the relevance scores of $13,176$ movies and $12$ tags for evaluation.
Each tuple represents the relevance scores of $12$ tags to a movie.
\item \textbf{Indep} is generated as described in~\cite{DBLP:conf/icde/BorzsonyiKS01}.
It is a set of uniform points on the unit hypercube where different attributes are
independent of each other.
\item \textbf{AntiCor} is also generated as described in~\cite{DBLP:conf/icde/BorzsonyiKS01}.
It is a set of random points with anti-correlated attributes.
\end{itemize}
\begin{table}[t]
\centering
\footnotesize
\caption{Statistics of datasets}\label{tbl:stats}
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Dataset} & $n$ & $d$ & \emph{\#skylines} & \emph{updates} (\%) \\ \hline
\textbf{BB} & $21,961$ & $5$ & $200$ & $1.07$ \\ \hline
\textbf{AQ} & $382,168$ & $9$ & $21,065$ & $5.60$ \\ \hline
\textbf{CT} & $581,012$ & $8$ & $77,217$ & $13.4$ \\ \hline
\textbf{Movie} & $13,176$ & $12$ & $3,293$ & $26.5$ \\ \hline
\textbf{Indep} & $100$K--$1$M & $2$--$10$ & \multicolumn{2}{c|}{see Fig.~\ref{fig:stats}} \\ \hline
\textbf{AntiCor} & $100$K--$1$M & $2$--$10$ & \multicolumn{2}{c|}{see Fig.~\ref{fig:stats}} \\ \hline
\end{tabular}
\end{table}
\begin{figure}[t]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{legend-dataset.pdf}
\end{subfigure}
\\
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{stat-d.pdf}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{stat-n.pdf}
\end{subfigure}
\caption{Sizes and update rates of the skylines of synthetic datasets}
\label{fig:stats}
\end{figure}
The statistics of datasets are reported in Table~\ref{tbl:stats}.
Here, $n$ is the number of tuples; $d$ is the dimensionality; \emph{\#skylines}
is the number of tuples on the skyline; and \emph{updates} (\%) is the percentage
of tuple operations that trigger any update on the skyline.
Note that we generated several \textbf{Indep} and \textbf{AntiCor} datasets
by varying $n$ from $100$K to $1$M and $d$ from $2$ to $10$ for scalability tests.
By default, we used the ones with $n=100$K and $d=6$.
The sizes and update rates of the skylines of synthetic datasets
are shown in Fig.~\ref{fig:stats}.
\noindent\textbf{Workloads:} The workload of each experiment was generated as follows:
First, we randomly picked $50\%$ of tuples as the initial dataset $P_0$;
Second, we inserted the remaining $50\%$ of tuples one by one into the dataset
to test the performances for insertions;
Third, we randomly deleted $50\%$ of tuples one by one from the dataset to test the performances for deletions.
It is guaranteed that the orders of operations kept the same for all algorithms.
The $k$-RMS results were recorded $10$ times when $10\%, 20\%, \ldots, 100\%$ of the operations were performed.
\noindent\textbf{Performance Measures:}
The efficiency of each algorithm was measured by \emph{average update time}, i.e.,
the average wall-clock time of an algorithm to update the result of $k$-RMS for each operation.
For the static algorithms, we only took the time for $k$-RMS computation into account
and ignored the time for skyline maintenance for fair comparison.
The quality of results was measured by the \emph{maximum $k$-regret ratio} ($\mathtt{mrr}_k$)
for a given size constraint $r$, and, of course, the smaller $\mathtt{mrr}_k$ the better.
To compute $\mathtt{mrr}_k(Q)$ of a result $Q$, we generated a test
set of $500$K random utility vectors and used the maximum regret value found as our estimate.
Since the $k$-RMS results were recorded $10$ times for each query,
we reported the average of the maximum $k$-regret ratios of $10$ results for evaluation.
\subsection{Experimental Results}\label{subsec:results}
\begin{figure*}[t]
\centering
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{NBA-eps.pdf}
\caption{BB}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AQ-eps.pdf}
\caption{AQ}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{CT-eps.pdf}
\caption{CT}
\end{subfigure}
\\
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Movie-eps.pdf}
\caption{Movie}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Indep-eps.pdf}
\caption{Indep}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AC-eps.pdf}
\caption{AntiCor}
\end{subfigure}
\caption{Performance of FD-RMS with varying $\varepsilon$
($k=1$; $r=20$ for BB and $r=50$ for other datasets).
Note that the red line represents the update time and the blue bars denote
the maximum regret ratios.}
\label{fig:eps}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{.96\textwidth}
\centering
\includegraphics[width=\textwidth]{legend-k1.pdf}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-NBA-time.pdf}
\includegraphics[width=.48\textwidth]{k1-NBA-mrr.pdf}
\caption{BB}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-AQ-time.pdf}
\includegraphics[width=.48\textwidth]{k1-AQ-mrr.pdf}
\caption{AQ}
\end{subfigure}
\\
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-CT-time.pdf}
\includegraphics[width=.48\textwidth]{k1-CT-mrr.pdf}
\caption{CT}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-Movie-time.pdf}
\includegraphics[width=.48\textwidth]{k1-Movie-mrr.pdf}
\caption{Movie}
\end{subfigure}
\\
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-Indep-time.pdf}
\includegraphics[width=.48\textwidth]{k1-Indep-mrr.pdf}
\caption{Indep}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k1-AC-time.pdf}
\includegraphics[width=.48\textwidth]{k1-AC-mrr.pdf}
\caption{AntiCor}
\end{subfigure}
\caption{Update time and maximum regret ratios with varying the result size $r$ ($k=1$)}
\label{fig:r}
\end{figure*}
\noindent\textbf{Effect of parameter $\varepsilon$ on FD-RMS:}
In Fig.~\ref{fig:eps}, we present the effect of the parameter $\varepsilon$
on the performance of FD-RMS.
We report the update time and maximum regret ratios
of FD-RMS for $k=1$ and $r=50$ on each dataset (except $r=20$ on \textbf{BB})
with varying $\varepsilon$.
We use the method described in Section~\ref{subsec:impl} to set the value of $M$
for each value of $\varepsilon$.
First of all, the update time of FD-RMS increases significantly
with $\varepsilon$. This is because both the time to process an $\varepsilon$-approximate
top-$k$ query and the number of top-$k$ queries (i.e.,~$M$) grow with $\varepsilon$,
which requires a larger overhead to maintain both top-$k$ results and set-cover solutions.
Meanwhile, the quality of results first becomes better when
$\varepsilon$ is larger but then could degrade if $\varepsilon$ is too large.
The improvement in quality with increasing $\varepsilon$ is attributed to
larger $m$ and thus smaller $\delta$.
However, once $\varepsilon$ is greater than the maximum regret ratio $\varepsilon^*_{k,r}$
of the optimal result (whose upper bound can be inferred from practical results),
e.g.,~$\varepsilon=0.0512$ on \textbf{BB},
the result of FD-RMS will contain less than $r$ tuples and
its maximum regret ratio will be close to $\varepsilon$ no matter how large $m$ is.
To sum up, by setting $\varepsilon$ to the one that is slightly lower
than $\varepsilon^*_{k,r}$ among $[0.0001,\ldots,0.1024]$, FD-RMS performs better
in terms of both efficiency and solution quality, and the values of $\varepsilon$
in FD-RMS are decided in this way for the remaining experiments.
\noindent\textbf{Effect of result size $r$:}
In Fig.~\ref{fig:r}, we present the performance of different algorithms for $1$-RMS
(a.k.a.~$r$-regret query) with varying $r$.
In particular, $r$ is ranged from $10$ to $100$ on each dataset
(except \textbf{BB} where $r$ is ranged from $5$ to $25$).
In general, the update time of each algorithm grows while the maximum regret ratios
drop with increasing $r$. But, for FD-RMS, it could take less update time
when $r$ is larger in some cases. The efficiency of FD-RMS is positively correlated
with $m$ but negatively correlated with $\varepsilon$. On a specific dataset, FD-RMS
typically chooses a smaller $\varepsilon$ when $r$ is large, and vice versa.
When $\varepsilon$ is smaller, $m$ may decrease even though $r$ is larger.
Therefore, the update time of FD-RMS decreases with $r$ in some cases
because a smaller $\varepsilon$ is used.
Among all algorithms tested, \textsc{Greedy} is the slowest and fails to
provide results within one day on \textbf{AQ}, \textbf{CT}, and \textbf{AntiCor} when $r>80$.
\textsc{GeoGreedy} runs much faster than \textsc{Greedy}
while having equivalent quality on low-dimensional data.
However, it cannot scale up to high dimensions (i.e., $d>7$) because the cost of
finding \emph{happy points} grows significantly with $d$.
DMM-RRMS suffers from two drawbacks:
(1) it also cannot scale up to $d>7$ due to huge memory consumption;
(2) its solution quality is not competitive when $r \geq 50$
because of the sparsity of space discretization.
The solution quality of $\varepsilon$-\textsc{Kernel} is generally inferior to
any other algorithm because the size of an $\varepsilon$-kernel coreset is much larger
than the size of the minimum $(1,\varepsilon)$-regret set.
Although HS provides results of good quality in most cases,
it runs several orders of magnitude slower than FD-RMS.
\textsc{Sphere} demonstrates better performance than other static algorithms.
Nevertheless, its efficiency is still much lower than FD-RMS,
especially on datasets with large skyline sizes, e.g., \textbf{CT} and \textbf{AntiCor},
where FD-RMS runs up to three orders of magnitude faster.
URM shows good performance in both efficiency and solution quality
for small $r$ (e.g., $r \leq 20$) and skyline sizes (e.g., on \textbf{BB} and \textbf{Indep}).
However, it scales poorly to large $r$ and skyline sizes because
the convergence of k-\textsc{medoid} becomes very slow and
the number of linear programs for regret computation grows rapidly
when $r$ and and skyline sizes increase.
In addition, URM does not provide high-quality results in many cases
since k-\textsc{medoid} cannot escape from local optima.
To sum up, FD-RMS outperforms all other algorithms for fully-dynamic $1$-RMS
in terms of efficiency.
Meanwhile, the maximum regret ratios of the results of FD-RMS are very close
(the differences are less than $0.01$ in almost all cases) to the best of static algorithms.
\begin{figure*}[t]
\centering
\begin{subfigure}{.75\textwidth}
\centering
\includegraphics[width=\textwidth]{legend-k5.pdf}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-NBA-time.pdf}
\includegraphics[width=.48\textwidth]{k5-NBA-mrr.pdf}
\caption{BB}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-AQ-time.pdf}
\includegraphics[width=.48\textwidth]{k5-AQ-mrr.pdf}
\caption{AQ}
\end{subfigure}
\\
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-CT-time.pdf}
\includegraphics[width=.48\textwidth]{k5-CT-mrr.pdf}
\caption{CT}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-Movie-time.pdf}
\includegraphics[width=.48\textwidth]{k5-Movie-mrr.pdf}
\caption{Movie}
\end{subfigure}
\\
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-Indep-time.pdf}
\includegraphics[width=.48\textwidth]{k5-Indep-mrr.pdf}
\caption{Indep}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{k5-AC-time.pdf}
\includegraphics[width=.48\textwidth]{k5-AC-mrr.pdf}
\caption{AntiCor}
\end{subfigure}
\caption{Update time and maximum regret ratios with varying $k$
($r=10$ for BB and Indep; $r=50$ for other datasets)}
\label{fig:k}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{.96\textwidth}
\centering
\includegraphics[width=\textwidth]{legend-k1.pdf}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{d-Indep-time.pdf}
\includegraphics[width=.48\textwidth]{d-Indep-mrr.pdf}
\caption{Indep, varying $d$}\label{fig:nd:a}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{d-AC-time.pdf}
\includegraphics[width=.48\textwidth]{d-AC-mrr.pdf}
\caption{AntiCor, varying $d$}\label{fig:nd:b}
\end{subfigure}
\\
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{n-Indep-time.pdf}
\includegraphics[width=.48\textwidth]{n-Indep-mrr.pdf}
\caption{Indep, varying $n$}\label{fig:nd:c}
\end{subfigure}
\begin{subfigure}{.36\textwidth}
\centering
\includegraphics[width=.48\textwidth]{n-AC-time.pdf}
\includegraphics[width=.48\textwidth]{n-AC-mrr.pdf}
\caption{AntiCor, varying $n$}\label{fig:nd:d}
\end{subfigure}
\caption{Scalability with varying the dimensionality $d$ and dataset size $n$
($k=1, \; r=50$)}
\label{fig:nd}
\end{figure*}
\noindent\textbf{Effect of $k$:}
The results for $k$-RMS with varying $k$ from $1$ to $5$ are illustrated
in Fig.~\ref{fig:k}.
We only compare FD-RMS with \textsc{Greedy}$^*$,
$\varepsilon$-\textsc{Kernel}, and HS
because other algorithms are not applicable to the case when $k>1$.
We set $r=10$ for \textbf{BB} and \textbf{Indep}
and $r=50$ for the other datasets.
The results of \textsc{Greedy}$^*$ for $k>1$ are only available
on \textbf{BB} and \textbf{Indep}. For the other datasets,
\textsc{Greedy}$^*$ fails to return any result within one day when $k>1$.
We can see all algorithms run much slower when $k$ increases.
For FD-RMS, lower efficiencies are caused by higher cost of maintaining top-$k$ results.
HS and $\varepsilon$-\textsc{Kernel} must consider all tuples in the datasets
instead of only skylines to validate that the maximum $k$-regret ratio is at most
$\varepsilon$ when $k>1$.
For \textsc{Greedy}$^*$, the number of linear programs to compute $k$-regret ratios
increases drastically with $k$.
Meanwhile, the maximum $k$-regret ratios drop with $k$,
which is obvious according to its definition.
FD-RMS achieves speedups of up to four orders of magnitude
than the baselines on all datasets.
At the same time, the solution quality of FD-RMS is also better
on all datasets except \textbf{Movie} and \textbf{CT},
where the results of HS are of slightly
higher quality in some cases.
\noindent\textbf{Scalability:}
Finally, we evaluate the scalability of different algorithms w.r.t.~the dimensionality $d$
and dataset size $n$.
To test the impact of $d$, we fix $n=100$K, $k=1$, $r=50$, and vary $d$ from $2$ to $10$.
The performance with varying $d$ is shown in Fig.~\ref{fig:nd:a}--\ref{fig:nd:b}.
Both the update time and maximum regret ratios of all algorithms increase dramatically with $d$.
Although almost all algorithms show good performance when $d=2,3$,
most of them quickly become very inefficient in high dimensions.
Nevertheless, FD-RMS has a significantly better scalability w.r.t.~$d$:
It achieves speedups of at least $100$ times over any other algorithm
while providing results of equivalent quality when $d \geq 7$.
To test the impact of $n$, we fix $d=6$, $k=1$, $r=50$, and vary $n$ from $100$K to $1$M.
The update time with varying $n$ is shown in Fig.~\ref{fig:nd:c}--\ref{fig:nd:d}.
For static algorithms, we observe different trends in efficiency on two datasets:
The update time slightly drops on \textbf{Indep} but keeps steady on \textbf{AntiCor}.
The efficiencies are determined by two factors, i.e.,
\emph{the number of tuples on the skyline} and
\emph{the frequency of skyline updates}.
As shown in Fig.~\ref{fig:stats}, when $n$ is larger, the number of tuples on the skyline
increases but the frequency of skyline updates decreases.
On \textbf{Indep}, the benefits of lower update frequencies outweigh the cost of more
skyline tuples; on \textbf{AntiCor}, two factors cancel each other.
FD-RMS runs slower when $n$ increases due to
higher cost of maintaining top-$k$ results on \textbf{Indep}.
But, on \textbf{AntiCor}, the update time keeps steady with $n$
because of smaller values of $\varepsilon$
and $m$, which cancel the higher cost of maintaining top-$k$ results.
We can observe that the maximum regret ratios of most algorithms
are not not significantly affected by $n$.
The solution quality of FD-RMS is close to the best of static algorithms.
Generally, FD-RMS always outperforms all baselines for different values of $n$.
\section{Frequently Used Notations}
A list of frequently used notations in this paper is summarized in Table~\ref{tbl:notation}.
\section{Additional Experimental Evaluation}
\subsection{Extension of Greedy for Dynamic Updates}
In Section~\ref{sec:exp}, all algorithms (except FD-RMS and URM) we compare are static algorithms
that recompute the results each time when the skyline is updated.
In fact, it is possible to extend a static algorithm, e.g.,
\textsc{Greedy}~\cite{DBLP:journals/pvldb/NanongkaiSLLX10}, to
support dynamic updates as follows: Initially, it computes a $k$-regret minimizing set
of size $r'=r+\Delta_r$ (where $\Delta_r$ is an integer parameter) using \textsc{Greedy}.
Among this set, a set of $r$ tuples is reported as the result.
The result is recomputed whenever (1) $\Delta_r$ tuples from the $k$-regret minimizing set
of size $r'$ are deleted (in which case it recomputes a $k$-regret minimizing set of size $r'$);
(2) there are a total of $\Delta_s$ number of insertions into the skyline
where $\Delta_s$ is another parameter.
We refer to this extension of \textsc{Greedy} as the Dynamic Greedy (DG) algorithm.
In this subsection, we compare the Dynamic Greedy (DG) algorithm with \textsc{Greedy}
and FD-RMS. Specifically, we test four sets of parameters $\Delta_s,\Delta_r$, namely
DG-1 with $\Delta_s=10,\Delta_r=0.25r$, DG-2 with $\Delta_s=25,\Delta_r=0.5r$,
DG-3 with $\Delta_s=50,\Delta_r=0.75r$, and DG-4 with $\Delta_s=100,\Delta_r=r$.
The original \textsc{Greedy} is a special case of DG with $\Delta_s=1,\Delta_r=0$.
The experimental results are shown in Fig.~\ref{fig:dynamic:greedy}.
First of all, when $\Delta_s$ and $\Delta_r$ increase, DG runs faster because
the recomputation is executed at lower frequencies. But unfortunately,
DG still runs much slower than FD-RMS even when $\Delta_s=100,\Delta_r=r$
because of the huge gap in the efficiency of \textsc{Greedy} and FD-RMS.
Meanwhile, the maximum regret ratios of the results of DG increase
with $\Delta_s$ and $\Delta_r$, and become significantly higher than
those of FD-RMS when $\Delta_s$ and $\Delta_r$ are large. This is because
the results of DG could often be obsolete w.r.t.~up-to-date
datasets. In general, DG is significantly inferior to FD-RMS in terms of
both efficiency and quality of results for different parameters.
Nevertheless, DG might be used to accelerate \textsc{Greedy} in dynamic settings
if the requirement for quality is not so strict.
\subsection{Effect of Update Patterns}
In Section~\ref{sec:exp}, we consider the updates to datasets consist of all insertions
first followed by all deletions to test the performance of different algorithms
for tuple insertions and deletions separately. In this subsection,
we evaluate the performance of FD-RMS in two update
patterns: (1) First all insertions followed by all deletions as Section~\ref{sec:exp} (I+D);
(2) A random mixture of insertions and deletions (Mixed).
We note that the efficiency and solution quality of a static
algorithm are not influenced by update patterns
because the update rates of skylines are similar for tuple insertions
and deletions, and it reruns from scratch for each skyline update.
The update time of FD-RMS in two update patterns is shown
in Fig.~\ref{fig:time:mixed}. We can see that the update time is nearly indifferent
to update patterns. In addition, we also confirm that the results returned by FD-RMS
remain the same in almost all cases regardless of the orders of tuple operations.
In addition, we show the update time of FD-RMS with varying $r$ for tuple
insertions and deletions in Fig.~\ref{fig:id}. The update time for deletions
is slightly longer than that for insertions on all datasets. This
is mainly attributed to the difference in the update of approximate top-$k$ results.
For insertions, the dual-tree can incrementally update the current top-$k$
result using the utility index only; but for deletions, the dual-tree must retrieve
the new top-$k$ result from scratch on the tuple index once a top-$k$ tuple is deleted.
\subsection{Extension of URM for Dynamic Updates}\label{appendix:dynamic:urm}
\begin{algorithm}
\caption{\textsc{Dynamic URM}}
\label{alg:urm}
\Input{Initial database $P_0$, set of operations $\Delta$}
\Output{Result $Q_t$ for $1$-RMS on $P_t$ at time $t$}
\tcc{compute an initial result $Q_0$ on $P_0$}
run DMM-RRMS to get an initial result $S_0$\;
run $k$-\textsc{medoid} from $S_0$ until convergence as $Q_0$\;
\tcc{update $Q_{t-1}$ to $Q_t$ for $\Delta_t = \langle p,\pm \rangle$}
Update $P_{t-1}$ to $P_t$ w.r.t.~$\Delta_t$\;
\uIf{$\Delta_t= \langle p,+ \rangle$}{
\uIf{there exists $p' \in Q_{t-1}$ such that the max regret ratio of $p$
in the region $R(p')$ of $p'$ is lower than $p'$}
{$S' \gets Q_{t-1} \setminus \{p'\} \cup \{p\}$\;
run $k$-\textsc{medoid} from $S'$ until convergence as $Q_t$\;}
\Else{ $Q_t \gets Q_{t-1}$\; }
}
\Else{\tcc{for a deletion $\Delta_t= \langle p,- \rangle$}
\uIf{$p \in Q_{t-1}$}
{$S' \gets Q_{t-1} \setminus \{p\} \cup \{p'\}$ where $p' \notin Q_{t-1}$ is top-ranked
for at least one utility vector in the region $R(p)$ of $p$\;
run $k$-\textsc{medoid} from $S'$ until convergence as $Q_t$\;}
\Else{ $Q_t \gets Q_{t-1}$\; }
}
\end{algorithm}
Unlike other static algorithms, the URM algorithm~\cite{Shetiya2019}
adopts an iterative framework based on the $k$-\textsc{medoid} clustering
for RMS computation. It is intuitive to extend the iterative framework to support dynamic updates.
The high-level idea of updating the result of URM for tuple operations
without fully recomputation is as follows: If an operation does not
affect the convergence of the current result, then skip it directly;
Otherwise, first adjust the result for this operation and then
run the iterative framework starting from the new result
until it converges again. The detailed procedures of (dynamic) URM
are presented in Algorithm~\ref{alg:urm}.
\begin{figure*}[t]
\captionsetup[subfigure]{aboveskip=2pt}
\centering
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[width=\textwidth]{NBA-Dynamic-r10.pdf}
\caption{BB ($r=10$)}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[width=\textwidth]{Movie-Dynamic-r50.pdf}
\caption{Movie ($r=50$)}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\centering
\includegraphics[width=\textwidth]{Indep-Dynamic-r50.pdf}
\caption{Indep ($d=6,r=50$)}
\end{subfigure}
\caption{Performance of the Dynamic Greedy (DG) algorithm in comparison with \textsc{Greedy}
and FD-RMS ($k=1$). Parameter settings for DG are listed as follows:
DG-1 -- $\Delta_s=10,\Delta_r=0.25r$; DG-2 -- $\Delta_s=25,\Delta_r=0.5r$;
DG-3 -- $\Delta_s=50,\Delta_r=0.75r$; DG-4 -- $\Delta_s=100,\Delta_r=r$.}
\label{fig:dynamic:greedy}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{NBA-mixed.pdf}
\caption{BB}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AQ-mixed.pdf}
\caption{AQ}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{CT-mixed.pdf}
\caption{CT}
\end{subfigure}
\\
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Movie-mixed.pdf}
\caption{Movie}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Indep-mixed.pdf}
\caption{Indep ($d=6$)}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AC-mixed.pdf}
\caption{AntiCor ($d=6$)}
\end{subfigure}
\caption{Update time of FD-RMS in different update patterns ($k=1$)}
\label{fig:time:mixed}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{NBA-id.pdf}
\caption{BB}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AQ-id.pdf}
\caption{AQ}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{CT-id.pdf}
\caption{CT}
\end{subfigure}
\\
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Movie-id.pdf}
\caption{Movie}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{Indep-id.pdf}
\caption{Indep}
\end{subfigure}
\begin{subfigure}{.18\textwidth}
\centering
\includegraphics[width=\textwidth]{AC-id.pdf}
\caption{AntiCor}
\end{subfigure}
\caption{Update time of FD-RMS for tuple insertion and deletion ($k=1$).
The red line denotes the average update time for all tuple operations.}
\label{fig:id}
\end{figure*}
| 2d4f96ec7293abd497d9af8d9d2a9971acaccdf4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
Algebra is a spacious part of the science of mathematics which provides the opportunity to express mathematical ideas precisely. In algebra, the Binomial expansion and Pascal's triangle are considered important.
Pascal's triangle is an arrangement of the binomial coefficients and one of the most known integer models.
Though it was named after French scientist Blaise Pascal, it was studied in ancient India, Persia, China, Germany, and Italy \cite{edwards}.\\[2ex]
In reality, the definition of the triangle was made centuries ago. In 450 BC, an Indian mathematician named Pingala is said to have introduced the definition of this triangle in a Sanskrit poetry book. At the same time, the commentators of this book acquainted with the diagonal surface of the triangle, which is the sum of the Fibonacci numbers. Chinese mathematicians had the same idea and named the triangle as ``Yang Hui's triangle”. Later, Persian mathematician Al-Karaji and Persian astronomer-poet Omar Khayyam named the triangle as the “Khayyam triangle”.
It also has multi-dimensional shapes, the three-dimensional shape is referred to as Pascal's pyramid or Pascal's tetrahedron, while the other general-shaped ones are called Pascal's simplifications.\\[2ex]
Various studies have been conducted in many different disciplines about Pascal's triangle. For the construction of Pascal's triangle, Sgroi \cite{Sgroi} stated that each line starts with 1 and ends with 1, and this series can be expanded with simple cross-joints. Jansson \cite{jansson} developed three geometric forms related to Pascal's triangle and included examples on each form. Toschi \cite{toschi} used various permutations to generate new forms of Pascal's triangles and generalized them. Duncan and Litwiller \cite{duncan} addressed the reconstruction of Pascal's triangle with the individuals. Here they collected data on the opinions of individuals using qualitative methods, and determined the methods of constructing the Pascal's triangle in different ways with the attained findings.\\[2ex]
Researches worked on Pascal's fascinating characteristics. Using the principle of permutation, Putz \cite{putz} designed the Pascal Polytope and linked it to the Fibonacci concept. Houghton \cite{houghton} gave the concept of the relationship between successive differential operation of a function and Pascal's triangle. With an application, he attempted to incorporate the idea of a differentiable function into Pascal's triangle. The relationship between Pascal's triangle and the Tower of Hanoi has been elucidated by Andreas M Hinz \cite{hinz}. Finding diagonal sum \cite{hoggatt}, k-Fibonacci sequence \cite{falcon}, recurrence relations \cite{green}, finding exponential $(e)$ \cite{brothers} were a part of those to describe the work that generates from the Pascal's triangle. Some fascinating properties of Pascal’s triangle are available in \cite{Bondarenko, Korec}. In 1956, Freund \cite{Freund} elicited that the generalized Pascal's triangles of $s^\text{th}$ order can be constructed from the generalized binomial coefficients of order $s$. Bankier \cite{bankier} gave the Freud’s alternative proof. Kall{\'o}s generalized Pascal's triangle from algebraic point of view by different bases. He tried to generalize Pascal's triangle using the power of integers \cite{Kall_1}, powers of base numbers \cite{Kall_2} and their connections with prime number \cite{Farkas}. kuhlmann tried to generate Pascal's triangle using the T-triangle concept \cite{kuhlmann}.\\[2ex]
The concept of the power of $11$ was first introduced by Sir Isaac Newton. He noticed that first five rows of Pascal's triangle are formed by the power of $11$ and claimed (without proof) that subsequent rows can also be generated by the power of eleven as well \cite{newton1736}. Arnold \textit{et al.} \cite{Arnold} showed if one assigns a place value to each of the individual terms in a certain row of the triangle, the pattern can be seen again. Morton \cite{Morton} noted the Pascal's triangle property by the power of 11 for 10 base numerals system. Mueller \cite{Mueller} noted that one can get the $n^\text{th}$ power of $11$ from the $n^\text{th}$ row of the Pascal's triangle with positional addition.\\[2ex]
It is clearly concluded that above mentioned works did not express the full row of Pascal's triangle from the power of $11$, or some variant of it, as Sir Isaac Newton hinted . This paper has worked on the generalization of Pascal's triangle by extending the power of eleven idea. Here, we have extend the concept of power of $11$ to the power of $101, 1001, 10001, \ldots$ and proved a general formula to achieve any row of Pascal's triangle. Using our formula one can generate any row of Pascal's triangle, regardless of the number of rows one can imagine.
\section{Methods} \label{sec: Methods}
The very basic definition to get any element of a row of the Pascal's triangle is the summation of two adjacent elements of the previous row. Each number in Pascal's triangle is the sum of two numbers above that number. Usually, the lines of Pascal's triangle are numbered starting from $n = 0$ on the top and the numbers in each line are starting from $k = 0$ on the left. For $k=0$, their is only one value 1. As the next lines are created, The remaining right most and left most element for new row is 1.
\[
\begin{array}{ccccccccccccc} {} & {} & {} &{} & {} & {} & {1} & {} & {} & {} &{} & {} & {} \\ {} & {} & {} &{} & {} & {1} & {} & {1} & {} & {} &{} & {} & {} \\ {} & {} & {} &{} & {1} & {} & {2} & {} & {1} & {} &{} & {} & {} \\ {} & {} & {} &{1} & {} & {3} & {} & {3} & {} & {1} &{} & {} & {}\\ {} & {} & {1} &{} & {4} & {} & {6} & {} & {4} & {} &{1} & {} & {}\\ {} & {1} & {} &{5} & {} & {10} & {} & {10} & {} & {5} &{} & {1} & {}\\ {1} & {} & {6} &{} & {15} & {} & {20} & {} & {15} & {} &{6} & {} & {1} \\ {} & {} & {\ldots} &{} & {} & {} & {\ldots} & {} & {} & {} &{\ldots} & {} & {} \end{array}
\]
\newline
\vspace{0.5cm}
\hspace{5cm}\textbf{Figure 1:} Pascal's triangle\\[1ex]
The concept of the power of $11$ leads to us $11^{1}=11$, $1^{\text{st}}$ row of Pascal's triangle and so $11^{2}=121$, $11^{3}=1331$ and $11^{4}=14641$ reveal $2^{\text{nd}}$, $3^{\text{rd}}$ and $4^{\text{th}}$ row respectively. Before finding the general rule for subsequent rows, we first elaborate the previous concept of power 11. The reason behind getting Pascal's triangle by the power of 11 lies on the general rule of multiplication. What do we get from multiplication of a number by $11?$\\
\[
\begin{array}{@{}r@{}}
2^{nd}\text{ row of Pascal's triangle} \rightarrow 121\\
{}\times 11\\
\hline
1${\color{red}2}$1 \\
\text{left shift of all digits by 1 place} \rightarrow 12${\color{red}1}$0 \\
\hline
3^{rd}\text{ row of Pascal's triangle} \rightarrow 13$\myul[red]{3}$1\\
\end{array}
\]
\newline
\vspace{0.5cm}
\hspace{4cm} \textbf{Figure 2:} Results after multiplication by $11$\\[1ex]
Figure 2 shows that multiplication of a number by $11$ gives an output which is the sum of the two adjacent numbers of previous row of Pascal's triangle.
\\\\
Patently $11^{5}=161051$ and $11^6=1771561$, but the $5^{\text{th}}$ and $6^{\text{th}}$ row of Pascal's triangle are
\begin{center}
{1} { } { } {5} { } { } {10} { } { } {10} { } { } {5} { } { } {1} {} \\
{} { } { } {} { } { } {} { } { } {} { } { } {} { $\qquad~~\text{and}$} { } {} { } { } {} {} \\
{1} { } { } {6} { } { } {15} { } { } {20} { } { } {15} { } { } {6} { } { } {1} {}\\
\end{center}
respectively. The above scheme fails for $11^5$ or $11^6$. Why are we not getting the $5^{th}$ row or why does the power of $11$ fail here? The answer is the middle values of the $5^{\text{th}}$ row of Pascal's triangle are two digit numbers, whereas the power of $11$ represents Pascal's row as a representation of one decimal place. So for finding $5^{\text{th}}$ or any row onward, we need a formula that can represent the number generated from the power of 11 have two or more digits. Now, we will endeavor to formulate a specific rule that generates the required number of digits for the representation of a row of the Pascal's triangle.\vspace{0.2cm}
\newline
At first, we attempt to generate the number of two digits using the very basic rules of multiplication. Figure 3, displays the impact of multiplication by $101.$\\[2ex]
\[
\begin{array}{@{}r@{}}
101 \\
{}\times 101\\
\hline
${\color{red}1}$01\\
\text{zeros cause the left shift of all digits by 1 place} \rightarrow $~{\color{red}00}$00\\
\text{left shift of all digits by 2 places} \rightarrow 1${\color{red}01}$00\\
\hline
1$\myul[red]{02}$01\\
{}\times 101\\
\hline
${\color{blue}1}{\color{green}02}$01\\
${\color{blue}00}{\color{green}00}$00\\
\text{left shift of all digits by 2 places} \rightarrow 1${\color{blue}02}{\color{green}01}$00\\
\hline
1$\myul[blue]{03}\myul[green]{03}$01\\
\end{array}
\]
\vspace{0.5cm}
\hspace{4cm} \textbf{Figure 3:} Results after multiplication by $101$\\[1ex]
The underlined numbers are same as the summation of two adjacent numbers of the previous row, but multiplication by 101 displays the rows as a representation of two digits number.\\[1.05ex]
Now, $101^{5}=10510100501$, from which we can construct $5^{\text{th}}$ row of Pascal's triangle by omitting extra zeros and separating the digits.
\begin{center}
{1} {} {5} {} {10} {} {10} {} {5} {} {1} {}
\end{center}
Similarly from $101^{6}=1061520150601$ and $101^{7}=107213535210701$, we can easily construct $6^{\text{th}}$ and $7^{\text{th}}$ row respectively.\\[0.1ex]
\begin{center}
{1} { } { } {06} { } { } {15} { } { } {20} { } { } {15} { } { } {06} { } { } {01} {}\\
{ $\quad\text{and}$}\\
{1} { } { } {07} { } { } {21} { } { } {35} { } { }{35} { } { } {21} { } { } {07} { } { } {01} {}
\end{center}
$101^{5}$, $101^{6}$ and $101^{7}$ all are representing $5^{th}$, $6^{th}$ and $7^{th}$ row of Pascal's triangle respectively as a representation of two digit numbers due to the insertion of one zero between $1$ and $1~$ in 11 such that $101$. $11^5=161051$, $11^6 = 1771561$ and $11^7=19487171$ could also represent the respective rows according to the Newton's claim but $101^n$ makes it precise.\\
Can a conclusion be drawn for generating any row of Pascal's triangle with the help of extended concept of the power of 11 such as $101^n$?
The $9^{\text{th}}$ row of Pascal's triangle is
$$1~~ 9~~ 36 ~84~~ 126~~ 126 ~~84 ~~36~ 9~~ 1$$
Plainly, $101^9 =1093685272684360901$ does not give the $9^{\text{th}}$ row because of the central element of this row contains three digits.
So the representation of three decimal places for each entry of Pascal's triangle requires a new formula to be generated.
The previous context directed that multiplication of a number by $11$ and $101$ makes the left shift of all digits by one and two places respectively. Therefore three digits representation requires multiplication by $1001$.\\[1.25ex]
Figure 4, proofs the left shift of all digits by $3$ times when a number is multiplied by $1001.$\\
\[
\begin{array}{@{}r@{}}
1001 \\
{}\times 1001\\
\hline
${\color{red}1}$001\\
${\color{red}00}$000\\
${\color{red}000}$000\\
\text{left shift of all digits by 3 time} \rightarrow 1${\color{red}001}$000\\
\hline
1$\myul[red]{002}$001\\
{}\times 1001\\
\hline
${\color{blue}1}{\color{green}002}$001\\
${\color{blue}00}{\color{green}000}$000\\
${\color{blue}000}{\color{green}000}$000\\
\text{left shift of all digits by 3 time} \rightarrow 1${\color{blue}002}{\color{green}001}$000\\
\hline
1$\myul[blue]{003}\myul[green]{003}$001
\end{array}
\]
\vspace{0.5cm}
\hspace{3.5cm} \textbf{Figure 4:} Results after multiplication by $1001$\\[1ex]
By continuing the multiplication by 1001 in Figure 4, we get $$1001^9 =1009036084126126084036009001$$
from which one may form the $9^{th}$ row
of Pascal's triangle by partitioning the digits of the number from the right three digits in each partition.
\begin{center}
$1~~009~~036~~084~~126~~126~~084~~036~~009~~001$
\end{center}
Similarly, $(1001)^{10}=1010045120210252210120045010001$,
$$1010045120210252210120045010001\longmapsto 1~~010~~045~~120~~210~~252~~210~~120~~045~~010~~001$$
the $10^{\text{th}}$ row of the Pascal's triangle.
\section{Results and discussion}
From the above study, it may be concluded that the representation of three decimal places requires the left shift of all digits by three places, and the left shift of all digits requires two zeros between $1$ and $1~$ in 11, that is $1001$. Why do we require three decimal places representation for $9^{th}$ and $10^{th}$ rows of Pascal's triangle? Because the central elements of $9^{\text{th}}$ and $10^{\text{th}}$ rows are of three digits. Similarly, we need two digits representation for $5^{th}$ to $8^{th}$ rows since the central element of these rows are numbers of two digits. And, the first four rows satisfy $11^n$ since the central element of the first four rows contains one digit only. So for any row, the number of decimal places representation should be equal to the number of digits in the central value of that row.\\[1.5ex]
The above discussion compels to generate a formula to find the central value of any row of the Pascal's triangle.
For an odd number, say $n=9,$ we get $n+1=10$ elements in $9^{th}$ row and so the central value should be $\left(\frac{10}{2}\right)^{th}=5^{th}$ observation of that row, which is $\binom{9}{5-1}=\binom{9}{4}=126$.
For an even number, say $n=10$ we get $n+1=11$ elements and the central value should be $\lceil\frac{11}{2}\rceil^{th}=6^{th}$ observation, which is $\binom{10}{6-1}=\binom{10}{5}=252$.\\[1.25ex]
By taking the \textit{floor} value of $\frac{n}{2}$, a formula for central value of $n^{th}$ row is $\binom{n}{\lfloor\frac{n}{2}\rfloor}$. But we never need a central element rather it is necessary to know how many digits the central element has. Applying the property of Logarithmic function, one can identify how many digits (or decimal places) of the central element has without knowing it.
Therefore the number of digits in the central value is given by
$$\lceil\log_{10} \binom{n}{\lfloor\frac{n}{2}\rfloor}\rceil$$
Since $\lceil\log_{10} (X)\rceil$ represents the number of digits of $X$ when $X \ne 10^{n}$, for $n \in \mathbb{N}$. For a central value of $d$ decimal places we require $d-1$ zeros between $1~\text{and}~1~$ in 11, such that $\left(1~\left(d-1\right)~zeros~1\right)^{n}$. So, the required number of zeros between $1~\text{and}~1~$ in 11 can be obtained by taking the \textit{floor} value of $\log_{10} \binom{n}{\lfloor\frac{n}{2}\rfloor}$.\\
If $\Theta$ represents the number of zeros between $1$ and $1~$ in 11. Then $$\Theta=\lfloor\log_{10} \binom{n}{\lfloor\frac{n}{2}\rfloor}\rfloor$$. We now verify it for an odd number $n=9$ and an even number $n=10$.\\ [1ex]
If, $n=9$ then $\Theta=2$, and if $n=10$ then $\Theta=2$.\\[1ex]
For both of the numbers we need $2$ zeros between $1$ and $1~$ in 11. So, to get $9^{th}$ and $10^{th}$ rows we have to calculate $1001^{9}$ and $1001^{10}$ respectively. Both of these cases have been discussed above.\\[1ex]
It's time to generate the formula to find any row of Pascal's triangle. The general formula for generating $n^{th}$ row of Pascal's triangle is $(1\Theta1)^n$.
For a random number such as $n=15$ we get $\Theta=3$.\\[1ex]
So, we have to insert 3 zeros and the $15^{th}$ row can be constructed from the following
$$10001^{15}=
1001501050455136530035005643564355005300313650455010500150001$$
Partitioning the digits from the right four digits at a time as shown below
$$1~0015~0105~0455~1365~3003~5005~6435~6435~5005~3003~1365~0455~0105~0015~0001$$
Notice the partitioning yields the $15^{\text{th}}$ row of the Pascal's triangle
$$1~15~105~455~1365~3003~5005~6435~6435~5005~3003~1365~455~105~15~1$$
Similarly, we may verify for $n=16$, $\Theta=4$ and\\[2ex]
$(100001)^{16} =
\seqsplit{%
100016001200056001820043680800811440128701144008008043680182000560001200001600001}
$\\[2ex]
This $16^\text{th}$ row can also be verified from the existing Pascal's triangle. The above formula can be used for a large $n$. We now exemplify $51^{st}$ row of Pascal's triangle. Hence $n=51$ gives $\Theta=14$\\[2ex]
We have to put $14$ zeros between $1~\text{and}~1~$ in 11, that is
$(1000000000000001)^{51}$.\\[2ex]
$(1000000000000001)^{51} =
1
{\color{red}000000000000051}
{\color{blue}000000000001275}
{\color{red}000000000020825}\\
{\color{blue}000000000249900}
{\color{red}000000002349060}
{\color{blue}000000018009460}
{\color{red}000000115775100}
{\color{blue}0000006367630}\\
{\color{blue}50}
{\color{red}000003042312350}
{\color{blue}000012777711870}
{\color{red}000047626016970}
{\color{blue}000158753389900}
{\color{red}00047626016}\\
{\color{red}9700}
{\color{blue}001292706174900}
{\color{red}003188675231420}
{\color{blue}007174519270695}
{\color{red}014771069086725}
{\color{blue}027900908}\\
{\color{blue}274925}
{\color{red}048459472266975}
{\color{blue}077535155627160}
{\color{red}114456658306760}
{\color{blue}156077261327400}
{\color{red}1967930}\\
{\color{red}68630200}
{\color{blue}229591913401900}
{\color{red}247959266474052}
{\color{blue}247959266474052}
{\color{red}229591913401900}
{\color{blue}19679}\\
{\color{blue}3068630200}
{\color{red}156077261327400}
{\color{blue}114456658306760}
{\color{red}077535155627160}
{\color{blue}048459472266975}
{\color{red}027}\\
{\color{red}900908274925}
{\color{blue}014771069086725}
{\color{red}007174519270695}
{\color{blue}003188675231420}
{\color{red}001292706174900}
{\color{blue}0}\\
{\color{blue}00476260169700}
{\color{red}00158753389900}
{\color{blue}000047626016970}
{\color{red}000012777711870}
{\color{blue}000003042312350}\\
{\color{red}000000636763050}
{\color{blue}000000115775100}
{\color{red}000000018009460}
{\color{blue}000000002349060}
{\color{red}0000000002499}\\
{\color{red}00}
{\color{blue}000000000020825}
{\color{red}000000000001275}
{\color{blue}000000000000051}
{\color{red}000000000000001}
$
\newline\newline
The desired $51^{\text{st}}$ row can be obtained by partitioning each $15$ digits from the right. For readers convenience, we marked each segment with different colors and showing that the above formula generates the $51^{\text{st}}$ row of the Pascal's triangle. Now we give a proof of the fact observed above.
We have the following inequalities
\begin{flalign}
\Theta=\lfloor\log_{10} \binom{n}{\lfloor\frac{n}{2}\rfloor}\rfloor
\end{flalign}
\begin{flalign*}
\Rightarrow 10^{\Theta+1} > \binom{n}{\lfloor\frac{n}{2}\rfloor}
\end{flalign*}
\begin{flalign}
&\Rightarrow 10^{\Theta+1}-1\ge \binom{n}{\lfloor\frac{n}{2}\rfloor}\\
&\binom{n}{\lfloor\frac{n}{2}\rfloor}\ge\binom{n}{r}\\
&10^{\Theta}\le\binom{n}{\lfloor\frac{n}{2}\rfloor}
\end{flalign}
And notice that $\left(1\Theta 1\right)^{n}=\left(10^{\Theta+1}+1\right)^{n}$.
\begin{lem}
\label{pascal1}
If $n, r\in\mathbb{N}$ and $0\le r\le n$, then
\begin{eqnarray*}
\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1<10^{r(\Theta+1)}.
\end{eqnarray*}
\end{lem}
\begin{proof}
By inequality 3, we have
\begin{eqnarray*}
&\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1\\
& <\binom{n}{\lfloor\frac{n}{2}\rfloor}\left(10^{(r-1)(\Theta+1)}+10^{(r-2)(\Theta+1)}+\cdots+1\right)\\
&=\binom{n}{\lfloor\frac{n}{2}\rfloor}\left(\frac{10^{r(\Theta+1)}-1}{10^{(\Theta+1)}-1}\right)
<\binom{n}{\lfloor\frac{n}{2}\rfloor}\left(\frac{10^{r(\Theta+1)}}{10^{(\Theta+1)}-1}\right).
\end{eqnarray*}
Dividing the last quantity by $10^{r(\Theta+1)}$ and using the inequality 2, we have
\begin{eqnarray*}
&\binom{n}{\lfloor\frac{n}{2}\rfloor}\left(\frac{10^{r(\Theta+1)}}{10^{(\Theta+1)}-1}\right)\left(\frac{1}{10^{r(\Theta+1)}}\right)=\binom{n}{\lfloor\frac{n}{2}\rfloor}\left(\frac{1}{10^{(\theta+1)}-1}\right)\\
&\le\binom{n}{\lfloor\frac{n}{2}\rfloor}\frac{1}{\binom{n}{\lfloor\frac{n}{2}\rfloor}}=1.
\end{eqnarray*}
Therefore
\begin{eqnarray*}
\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1<10^{r(\Theta+1)}.
\end{eqnarray*}
\end{proof}
\begin{Prop}
\label{pascal2}
If $n,r\in\mathbb{N}$ and $0\le r\le n$, then
\begin{flalign*}
&\left(10^{\Theta+1}+1\right)^{n}\mod 10^{r(\Theta+1)}\\
&=\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1
\end{flalign*}
\end{Prop}
\begin{proof}
Expanding $\left(10^{\Theta+1}+1\right)^{n}$ by binomial theorem, we have
\begin{flalign*}
&\left(10^{\Theta+1}+1\right)^{n}\mod 10^{r(\Theta+1)}\\
&=[10^{n(\Theta+1)}+\binom{n}{1}10^{(n-1)(\Theta+1)}+\cdots
+\binom{n}{n-r}10^{(r)(\Theta+1)}+\\
&\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}
+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1]
\mod 10^{r(\Theta+1)}\\
&=\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1,
\end{flalign*}
since by the Lemma \ref{pascal1}, we have
\begin{eqnarray*}
\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1< 10^{r(\Theta+1)}.
\end{eqnarray*}
\end{proof}
\begin{Cor}
\label{pascal3}
The integer $\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1$ has at best $r(\Theta+1)$ significant digits.
\end{Cor}
\begin{proof}
Since $\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1$ is the remainder when $\left(10^{(\Theta+1)}+1\right)^{n}$ is moded out by $10^{r(\Theta+1)}$.
\end{proof}
\begin{Cor}
\label{pascal4}
The left most $\left(\Theta+1\right)$ digits of the integer
\begin{eqnarray*}
\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1
\end{eqnarray*}
is
\begin{eqnarray*}
\binom{n}{n-(r-1)}=\binom{n}{r-1}.
\end{eqnarray*}
\end{Cor}
\begin{proof}
By Corollary \ref{pascal3}, the integer
\begin{eqnarray*}
\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}+\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\cdots+1
\end{eqnarray*}
has at most $r(\Theta+1)$ significant digits, and similarly
\begin{eqnarray*}
\binom{n}{n-(r-2)}10^{(r-2)(\Theta+1)}+\binom{n}{n-(r-3)}10^{(r-3)(\Theta+1)}+\cdots+1
\end{eqnarray*}
has at most $(r-1)(\Theta+1)$ significant digits. But
$\binom{n}{n-(r-1)}10^{(r-1)(\Theta+1)}$ has $(r-1)(\Theta+1)$ zeros to the right, the left most $(r)(\Theta+1)-(r-1)(\Theta+1)=\Theta+1$ digits gives
\begin{eqnarray*}
\binom{n}{n-(r-1)}=\binom{n}{r-1}.
\end{eqnarray*}
\end{proof}
\begin{thm}
The $r$-th block of $(\Theta+1)$ digits from the right of the integer $(1\Theta1)^{n}$ is the binomial coefficient $\binom{n}{r-1}$, where $\Theta=\lfloor\log_{10} \binom{n}{\lfloor\frac{n}{2}\rfloor}\rfloor$.
\end{thm}
\begin{proof}
The proof follows from the Corollary \ref{pascal3} and the Corollary \ref{pascal4}. Hence partitioning the digits of the integer $(1\Theta1)^{n}$ generates all the binomial coefficients or the $(n+1)^{\text{th}}$ row of the Pascal's triangle.
\end{proof}
\section{Conclusion}
Sir Isaac Newton hinted that binomial coefficients in the
$(n+1)^{\text{th}}$ row of the Pascal's triangle may be achieved from partitioning the digits in the $n^{\text{th}}$ power of some variation of $11$ \cite{newton1736}. It has been shown earlier the weighted sum of the values in the $(n+1)^{\text{th}}$ row of the Pascal's triangle is $(11)^{n}$ \cite{Arnold}. We have shown that $\left( \Theta +1\right)$ digit partition of $(1 \Theta1)^{n}$ from the right gives the values of the $(n+1)^{\text{th}}$ row of the Pascal's triangle.\\[5ex]
\vspace{0.25in}
\bibliographystyle{apa}
| 6184732529abd2c6c96049fad6b94243cf626402 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{QM2 altered setting}
\label{Sec:QM2A}
The altered setting of the {QM2} \ model we study is motivated from \cite[Appendix C]{shah2019sequentialME}, and we describe it below.
We have $k$ bins, one corresponding to each element of the support set, and a representative element in each bin. Any algorithm proceeds in rounds. In the $t^{th}$ round, the algorithm chooses a subset $\mathcal{C}(t)$ of these bins, and we compare the $t^{th}$ sample with the representative element from each bin in $\mathcal{C}(t)$. Note that these comparisons happen in parallel and the number of queries in round $t$ is the cardinality of the set $\mathcal{C}(t)$. In each round, the oracle response could be $+1$ for some bin $j$ $\in$ $\mathcal{C}(t)$ and $-1$ for all other bins in $\mathcal{C}(t)$ with probability $p_j$, or the response could be $-1$ for all bins in $\mathcal{C}(t)$ with probability $(1 - \sum_{j \in \mathcal{C}(t)} p_j)$. Based on the oracle responses obtained so far, the algorithm decides whether to stop or to proceed to the next round. When the algorithm decides to stop, it outputs an estimate $\widehat{S}$ of $S_{\mathcal{P}}^{\gamma}$, the set of support elements with a probability above $\gamma$.
Note that this setting is different from our original setting in {QM2}. Firstly, the number of bins is fixed with one bin corresponding to each element of the support and furthermore, there is a priori one representative element present in each bin. Secondly, in the modified setting, we choose the set $\mathcal{C}(t)$ at the start of each round and all the $|\mathcal{C}(t)|$ replies from the oracle come in parallel. However, in {QM2}, we we perform queries sequentially in each round and terminate the round as soon as we get a $+1$ response from any one of the bins in $\mathcal{C}(t)$. In spite of these differences, we believe that the query complexity for both these models will be quite similar and as we see below, the alternate setting can be placed in a framework that is fairly well studied and can potentially provide provide pointers towards solving the original problem.
We look at this new problem as a structured Multi-armed Bandit (MAB) problem \cite{MABbook} where there are $k$ arms, and each arm has a Bernoulli reward distribution with mean $p_i$. From the constraints of our original setup, the means must sum up to $1$ i.e. $\sum_{i} p_i = 1$. In each round we can pull a subset $\mathcal{C}(t)$ of arms and the output is a vector with $-1$ for all arms in $\mathcal{C}(t)$ with probability (1- $\sum_{i \in \mathcal{C}(t)} p_i$) or the output vector has $+ 1$ for some arm $j \in \mathcal{C}(t)$ and $-1$ for all other arms with probability $p_j$. The number of pulls in round $t$ is cardinality of the set $\mathcal{C}(t)$. Based on the responses from the arms, the algorithm decides whether to continue to the next round or stop and output an estimate for the set of arms with mean rewards above $\gamma$. The aim of the algorithm is to correctly identify this set of arms with probability at least (1- $\delta$). The total number of pulls across all rounds is defined as the query complexity of a $\delta$-true $\gamma$-threshold estimator in this setting.
Ideally, we would like to get a tight lower bound on the query complexity for the aforementioned structured MAB problem. The key challenge in doing so is the simplex constraint on the class of mean rewards imposed by $\sum_i p_i =1$. Although we are unable to provide a lower bound for this constraint, we are able to provide a lower bound under a slightly relaxed constraint given by
\begin{equation}{\label{QM2_pi_cond}}
p_1 + p_2 +...+p_k + 2 \gamma < 1.
\end{equation}
%
\begin{theorem}{\label{QM2_alter_lb}}
For a MAB setting described above where the mean rewards of the individual arms satisfy the condition in equation \ref{QM2_pi_cond}, any $\delta$-true $\gamma$-threshold algorithm has the following lower bound in expectation on the total number of pulls $N$:
$$
\mathbb{E}_{\mathcal{P}}[N] \geq \sum_{i=1}^{k} \frac{\log(\frac{1}{2.4\delta})}{2\cdot d(p_i||\gamma)}.
$$
\end{theorem}
Note that the expression in the lower bound above is very similar to the upper bound on the query complexity under the {QM2} \ model derived in Theorem~\ref{QM2ub}. Proving such a lower bound for the structured MAB under the true simplex constraint $\sum_{i} p_i = 1$ is part of our future work. There has been some recent work on similar problems which might provide us some pointers on how to pursue this problem. In particular, say we restrict attention to the class of schemes which compare to a single bin in each round, i.e., $|\mathcal{C}(t)| = 1$ for all $t$. The \textit{thresholding bandit problem} as described above, without the simplex constraint, was studied in \cite{locatelli2016optimal}, and the optimal query complexity expression turns out to be very similar to the one in Theorem~\ref{QM2_alter_lb}. On the other hand, \cite{simchowitz2017simulator} studies the related problem of identifying the arm with the largest mean reward and derives a tight lower bound on the query complexity under the simplex constraint.
\section{THRESHOLD-ESTIMATOR UNDER {QM1}}
\section{{QM1} \ THRESHOLD ESTIMATOR}
\label{Sec:QM1}
Under {QM1}, for every query with index $i$, the oracle returns the value $X_i$ of the $i^{th}$ sample. We now present an algorithm for this query model and analyze its query complexity, see Algorithm~\ref{QM1_alg}.
\subsection{Algorithm}
We first create $k$ bins numbered $1,2,...,k$. In each time step $t$, we query the oracle with the index $t$. The oracle reveals the value $X_t$ and the index $t$ is placed in the bin whose index matches $X_t$.
We define $Z_i^t$ for $i\in \{1,2,..,k\}$ used in Algorithm \ref{QM1_alg} as follows:
\begin{equation*}
Z_i^t=
\begin{cases}
1 & \text{if}\ X_t = i, \\
0 & \text{otherwise}.
\end{cases}
\end{equation*}
We can argue that for each given $i$, $\{Z_i^t\}$ is a collection of i.i.d. Bernoulli random variables with $\mathbb{E}[Z_i^t]=p_i$. Let $\tilde{p}_i^t $ denote the empirical probability estimate for support element $i$ at time $t$, and is given by
\begin{equation}\label{eq1}
\tilde{p}_i^t = \sum_{j=1}^{t} Z_i^j/t.
\end{equation}
Note that $\tilde{p}_i^t$ denotes the fraction of samples in Bin $i$, till time $t$.
At each round $t$ and for every $i$, we choose the confidence interval of Bin $i$ such that $p_i$ lies within the interval with "sufficiently" high probability. For a given time $t$, let $u_i(t)$ and $l_i(t)$ denote the upper confidence bound (UCB) and lower confidence bound (LCB) of Bin $i$ respectively and the confidence interval of Bin $i$ is given by $[l_i(t),u_i(t)]$.
{\begin{algorithm}{\label{QM1_alg}}
{ {
\SetAlgoLined
Create bins numbered $1,2...k$.\\
Initialize $\tilde{p}_i^{0}=0$, $l_i(0)=0$, $u_i(0)=1$, $\forall i\in\{1,2,...,k\}$.\\
$t = 0$ \\
\While{$\exists$ Bin $j$ s.t $l_j(t)$ $<$ $\gamma$ $<$ $u_j(t)$}{
{
Obtain $X_{t+1}$.
}
$t = t +1$.\\
Place the index $t$ in the bin numbered $X_t$.\\
\text{Update the empirical estimate} $\tilde{p}_i^{t}$ according to \eqref{eq1}; upper and lower confidence bounds $l_i(t)$ according to \eqref{eqnlb} and $u_i(t)$ \eqref{eqnub} respectively for all bins.\\
}
Output $\{i|l_i(t)>\gamma\}$.
\caption{Estimator for {QM1}}
}}
\end{algorithm}}
Motivated from the bounds in \cite{pmlr-v30-Kaufmann13}, for some given sequence of parameters $\{\beta^t\}$, we define $l_i(t)$ and $u_i(t)$ as follows:
\begin{align}
l_i(t) &= {\min}\{ q \in [0,\tilde{p}_i^t]: t\times d(\tilde{p}_i^t||q) \leq \beta^t \},\label{eqnlb} \\
u_i(t) &= {\max}\{ q \in [\tilde{p}_i^t,1]: t\times d(\tilde{p}_i^t||q) \leq \beta^t \}.\label{eqnub}
\end{align}
Here $d(p||q)$ denotes the Kullback-Leibler divergence between two Bernoulli distributions with means $p$ and $q$ respectively
The algorithm keeps querying the oracle and updating the confidence intervals for each bin, until every bin has either its LCB above $\gamma$ or its UCB below $\gamma$ at which point the Algorithm~\ref{QM1_alg} terminates. The estimate $\widehat{S}$ is chosen to consist of all bin indices which have their LCBs above $\gamma$. We choose the sequence $\{\beta^t\}$ in such a way that the probability of error is bounded by $\delta$.
The following lemma claims that Algorithm \ref{QM1_alg} returns the desired set $S_{\mathcal{P}}^{\gamma}$ for any underlying distribution $\mathcal{P}$ with probability at least $1- \delta$.
\begin{lemma}{\label{Correctness_QM1}}
Given the choice of $\beta^t = \log(2kt^2 / \delta)$ for each $t\ge 1$, Algorithm \ref{QM1_alg} is a $\delta$-true $\gamma$-threshold estimator under {QM1}.
\end{lemma}
\subsection{Query Complexity Analysis}
Given two values $x$ and $y$ satisfying $0 \leq x,y \leq 1$, we define $d^{*} (x,y) = d(z||x)$ where $z$ satisfies the condition $d(z||x) = d(z||y)$. The quantity $d^{*} (x,y)$ represents the \textit{Chernoff information} between two Bernoulli distributions with means $x$ and $y$ respectively, and is a relevant quantity in hypothesis testing problems \cite[Section 11.8]{10.5555/1146355}. This definition plays a key role in the following theorem which provides an upper bound on the query complexity of our proposed estimator in Algorithm \ref{QM1_alg}.
\begin{theorem}{\label{QM1ub}}
Let $\mathcal{A}$ denote the estimator in Algorithm \ref{QM1_alg} with $\beta^t = \log(2kt^2 / \delta)$ for each $t\ge 1$ and let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM1}. Then, we have
$$Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})\leq \max_{j\in\{m,m+1\}} \Biggl\{\frac{2e\log \Bigl(\sqrt{\frac{2k}{\delta}}\frac{2}{d^{*}(p_j,\gamma)}\Bigr)}{(e-1)d^{*}(p_j,\gamma)} \Biggr\},$$
with probability at least $1-\delta$.
\end{theorem}
This is essentially argued as follows. For each $i$, we identify a time $T_i$ such that with high probability, Bin $i$ would have been classified as being either above $\gamma$ or below $\gamma$ by time $T_i$. We then show that $T_i$ is bounded by the expression given in Theorem \ref{QM1ub}, $\forall i \in\{1,2,...,k\}$.
Next, we give a lower bound of the number of queries $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ for any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM1}.
\begin{theorem}\label{QM1lb}
For any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM1}, let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the query complexity. Then, we have
$$\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})]\geq \max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{d(p_j||\gamma)}\Bigg\}.$$
\end{theorem}
The proof of the above result is based on change of measure arguments similar to those in \cite{kaufmann2016complexity}. The detailed proof is given in Section~\ref{Sec:Thm2_proof}. Theorems~\ref{QM1ub} and \ref{QM1lb} provide a fairly tight characterization of the optimal query complexity under the QM1 query model, with the key difference between the upper and lower bound being the terms $d^{*}(p_j,\gamma)$ and $d(p_j||\gamma)$ respectively. {From the definition of $d^*(x,y)$, $\min\bigg\{d\Big(\frac{p_j+\gamma}{2}||p_j\Big),d\Big(\frac{p_j+\gamma}{2}||\gamma\Big)\bigg\}\leq d^{*}(p_j,\gamma) \leq \max\bigg\{d\Big(\frac{p_j+\gamma}{2}||p_j\Big),d\Big(\frac{p_j+\gamma}{2}||\gamma\Big)\bigg\}$. We also have $ 2(p_j-\gamma)^2 \leq d(p_j||\gamma) \leq \frac{(p_j-\gamma)^2}{\gamma(1-\gamma)}$ (\cite{popescu2016bounds}), and combining the above inequalities, we get $\frac{d(p_j||\gamma)}{d^{*}(p_j,\gamma)}\leq \frac{2}{\gamma(1-\gamma)}$. Thus, our bounds in Theorems \ref{QM1ub} and \ref{QM1lb} are tight up to logarithmic factors.}
\begin{remark*}
Our algorithms use the knowledge of the number of support elements ($k$) in the confidence intervals' design, whereas our lower bounds are independent of $k$. Whether the dependence on $k$ is fundamental or can it be removed by designing smarter algorithms and using more sophisticated concentration inequalities is an important question we want to address in the future.
\end{remark*}
%
\section{{QM1-N} \ THRESHOLD ESTIMATOR }
\label{Sec:QM1N}
Under {QM1-N}, when queried with an index $i$, the oracle returns the true value of $X_i$ with probability $1-p_e$ and a uniformly at random chosen value in $\{1,2,\ldots,k\}$ with probability $p_e$.
The responses of the oracle in this noisy setting would stochastically be the same as the responses from another oracle in the {QM1} \ model with an underlying distribution $p_i' = (1-p_e)p_i + {p_e}/{k}$ over the same support set. Since the responses from the oracles in the two settings described above are stochastically identical, Theorems \ref{QM1ub} and \ref{QM1lb} also provide upper and lower bounds on the query complexity of $\delta$-true $\gamma$-threshold estimators for the {QM1-N} \ model, by considering the underlying distribution as $\mathcal{P}'=\{p_1',p_2',...,p_k'\}$, and the threshold as $\gamma'$ = $(1-p_e)\gamma + {p_e}/{k}$. Note that the estimator proposed in Algorithm \ref{QM1_alg} for the {QM1} \ model is a $\delta$-true $\gamma$-threshold estimator for the noisy setting as well, with the threshold set to $\gamma' = (1-p_e)\gamma + {p_e}/{k}$.
\section{{QM2} \ THRESHOLD-ESTIMATOR }
\label{Sec:QM2}
Recall that under the {QM2} \ model, the oracle only makes pairwise comparisons. When queried with two indices $i$ and $j$, the oracle response indicates whether the values $X_i$ and $X_j$ are equal or not. We now present $\delta$-true $\gamma$-threshold estimator for this query model and analyse its query complexity
\subsection{Algorithm}
%
\textit{High-level description:} Recall that under the {QM1} \ model, we begin by creating $k$ bins, one corresponding to each element of the support of the underlying distribution. In each time slot $t$, the oracle reveals the value $X_t$ of the sample $t$, and the index $t$ is placed in the corresponding bin. A key challenge in the {QM2} \ model is that the oracle does not reveal the value of the samples. Instead, it just provides pairwise information about whether the values of two samples corresponding to the pair of queried indices are equal or not. Here, we create and update bins as the algorithm proceeds, while ensuring that each such bin contains samples representing the same support element.
As with the previous models, the algorithm proceeds in rounds. Let $\mathcal{B}(t)$ denote the set of bins created by the end of round $t$. For each round $t >1 $, we choose a subset of bins $\mathcal{C}(t-1)$ from $\mathcal{B}(t-1)$. We go over the bins in $\mathcal{C}(t-1)$ one by one, and then query the oracle to compare the $t^{th}$ sample with a representative element from the current bin.
Round $t$ stops when we either get a positive response from the oracle indicating that a matching bin has been found, or when all the bins in $\mathcal{C}(t-1)$ have been exhausted. As before, we maintain confidence intervals for the true probability value corresponding to each bin and our algorithm as described in Algorithm~\ref{QM2_alg} terminates, when its termination criterion is met, which we discuss below in the detailed description.
\textit{Detailed description:}
We define $\mathbb{Z}_i^t$ as follows:
{\small
\begin{equation*}
\mathbb{Z}_i^t =
\begin{cases}
1 & \text{ if } \mathcal{O}(i_r,t)=1 \text{ for } i_r \in \text{Bin } i , \text{ Bin } i \in \mathcal{B}(t-1)\\
0 & \text{otherwise}.
\end{cases}
\end{equation*}}
Let $\hat{p}_i^t${\footnote{Note that $\hat{p}_i^t$ denotes the fraction of samples in Bin $i$ and $\tilde{p}_i^t$ in Section \ref{Sec:QM1} denotes the fraction of samples of the support element $i$.}} denotes the fraction of samples present in Bin $i$ by the end of round $t$. We define it formally as follows:
\begin{equation}{\label{emp_QM2}}
\hat{p}_i^t = {\sum_{j=1}^{t} \mathbb{Z}_i^j }/{t}.
\end{equation}
\begin{algorithm}[h]{\label{QM2_alg}}
{\small
\SetAlgoLined
$t$=1\\
Create a new bin and add sample $t$ to it.\\
Update the empirical estimate $\hat{p}_i^{t}$ according to \eqref{emp_QM2}, the upper bound $\hat{u}_i(t)$ and the lower bound $\hat{l}_i(t)$ according to \eqref{eqnubbin} and \eqref{eqnlbbin} for the created bin.
Form $\mathcal{C}$($t$) according to $\eqref{eqn:c1t}$.\\
\While{( $t < T'$)}{
{$t=t+1$\\}
flag = 0.\\
\ForAll{ $j$ $\in$ $\mathcal{C}(t-1)$ }{
Obtain $j_l \in \text{Bin } j$.\\
\If{$\mathcal{O}(t,j_l) == +1$}{
Add sample $t$ to corresponding bin.\\
flag = 1. BREAK.\\
}
}
\If{flag==0 } {
Create a new bin and add sample $t$ to it.
}
Update the empirical estimate $\hat{p}_i^{t}$ according to \eqref{emp_QM2}, the upper bound $\hat{u}_i(t)$ and the lower bound $\hat{l}_i(t)$ according to \eqref{eqnubbin} and \eqref{eqnlbbin} for all the bins in $\mathcal{C}(t-1)$ and the newly created bin if created in this round.\\
{For all other bins, we keep the empirical means and LCB-UCB bounds unchanged.\\}
%
{Form $\mathcal{C}$($t$) according to $\eqref{eqn:c1t}$.}\\
}
Initialise $\mathcal{S}=\{x|\hat{l}_x(t)>\gamma\}$.\\
\While{$\mathcal{C}(t)\neq \Phi$}{
{ $t = t +1$}.\\
\ForAll{$j$ $\in$ $\mathcal{C}(t-1)$ }{
Obtain $j_l \in \text{Bin } j$.\\
\If{$\mathcal{O}(t,j_l) == +1$}{
Add sample $t$ to corresponding bin.
BREAK.\\
}
}
Update the empirical estimate $\hat{p}_i^{t}$ according to \eqref{emp_QM2}, the upper bound $\hat{u}_i(t)$ and the lower bound $\hat{l}_i(t)$ according to \eqref{eqnubbin} and \eqref{eqnlbbin} for all the bins in $\mathcal{C}(t-1)$.\\
Update $\mathcal{S}$ by adding the bins with $\hat{l}_x(t)>\gamma$.
%
{Form $\mathcal{C}$($t$) according to $\eqref{eqn:c2t}$.}
}
Output $\mathcal{S}$.
\caption{Estimator for {QM2}}
}
\end{algorithm}
As before, let $\hat{l}_i(t)$ and $\hat{u}_i{(t)}$ denote the lower and upper confidence bounds for Bin $i$ respectively, and as before are defined as follows for some appropriate choice of $\{\beta^t\}$:
\begin{align}
\label{eqnlbbin}
\hat{l}_i(t) & = {\min}\{ q \in [0,\hat{p}_i^t]: t\times d(\hat{p}_i^t||q) \leq \beta^t \},\\
\label{eqnubbin}
\hat{u}_i(t) &= {\max}\{ q \in [\hat{p}_i^t,1]: t\times d(\hat{p}_i^t||q) \leq \beta^t \}.
\end{align}
We run the algorithm in 2 phases. The goal of the first phase is to ensure that one bin is created corresponding to each support element $\{1,2,...,m\}$ in $S_{\mathcal{P}}^{\gamma}$ with "high" probability. The first phase runs from round $1$ to round ${T'=\big(\log (2k/\delta)\big)/{\big(\log\frac{1}{1-\gamma}}\big)}$. For this phase, the subset of bins which have their UCB above $\gamma$ form the subset $\mathcal{C}(t-1)$ of the created bins which are compared against in round $t$, i.e.,
\begin{align}\label{eqn:c1t}
\mathcal{C}(t)=\{x| \gamma<\hat{u}_x(t) , x \in \mathcal{B}(t) \}.
\end{align}
{In each round $t$, we choose a representative index ${i_r}\in$ Bin $i$ , $\forall i\in\mathcal{C}(t-1)$ and then go over the bins in $\mathcal{C}(t-1)$ one by one, querying the oracle with index pairs of the form ($i_r,t$)}. If we get a positive response from a bin, we place the index $t$ in the corresponding bin. If the replies from all the bins in $\mathcal{C}(t-1)$ is $-1$, we create a new bin with index $t$. We update the empirical estimates $\{\hat{p}_i^t\}$ as well as the UCB and LCBs of the created bins appropriately.
The second phase runs from round ${T'=\big(\log (2k/\delta)\big)/{\big(\log\frac{1}{1-\gamma}}\big)}$ onwards. In this phase, we do not create any new bins since the first phase guarantees that a bin corresponding to each element in $S_{\mathcal{P}}^{\gamma}$ has already been created. The goal of this phase is to correctly identify the bins belonging to elements in $S_{\mathcal{P}}^{\gamma}$ from amongst the created bins. Bins for which the corresponding LCB is greater than the threshold $\gamma$ are classified as belonging to $S_{\mathcal{P}}^{\gamma}$; and vice versa for bins with UCB at most $\gamma$.
At any round $t$ in this phase, the bins still in contention are those for which the UCB is above $\gamma$ and the LCB is below $\gamma$, and these are the ones that are chosen to form $\mathcal{C}(t)$, i.e.,
\begin{equation}
\label{eqn:c2t}
\mathcal{C}(t)=\{x| \hat{l}_x(t)< \gamma<\hat{u}_x(t) , x \in \mathcal{C}(t-1) \}.
\end{equation}
Similar to the first phase, in round $t$ we go over the bins in $\mathcal{C}(t-1)$ one by one, and then query the oracle to compare the $t^{th}$ sample with a representative element from the current bin. However, unlike the first phase, note that in this phase if we get a negative response from all the bins in $\mathcal{C}(t-1)$, we simply drop the index $t$.
The algorithm terminates when $\mathcal{C}(t)$ is empty, i.e., all the bins either have their LCB greater than $\gamma$ or their UCB below $\gamma$.
The following lemma claims that Algorithm \ref{QM2_alg} returns the desired set $S_{\mathcal{P}}^{\gamma}$ for any underlying distribution $\mathcal{P}$ under the {QM2} \ model, with probability at least ($1- \delta$).
\begin{lemma}{\label{Correctness_QM2}}
Given the choice of $\beta^t = \log(4kt^2/\delta)$ for each $t\ge1$, Algorithm \ref{QM2_alg} is a $\delta$-true $\gamma$-threshold estimator under {QM2}.
\end{lemma}
%
%
%
%
%
\subsection{Query Complexity Analysis}
The following theorem gives an upper bound on the query complexity of our proposed estimator in Algorithm \ref{QM2_alg}.
\begin{theorem}\label{QM2ub}
Let $\mathcal{A}$ denote the estimator in Algorithm~\ref{QM2_alg} with $\beta^t = \log(4kt^2 / \delta)$ for each $t\ge 1$ and let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM2}. We define $q$ as $\min\left\{k,{T'=\big(\log (2k/\delta)\big)/{\big(\log\frac{1}{1-\gamma}}\big)}\right\}$. Then, we have
\begin{align*}
Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})\leq \sum_{i=1}^{m} \max \Biggl\{ \frac{2e\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}(p_i,\gamma)}\Bigr)}{(e-1).d^{*}(p_i,\gamma)} ,{\frac{\log (2k/\delta)}{\log\frac{1}{1-\gamma}} \Biggr\}}
\hspace{0 em} + \sum_{i=m+1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}(p_i,\gamma)}\Bigr)}{(e-1).d^{*}(p_i,\gamma)},
\end{align*}
with probability at least $1-2\delta$.
\end{theorem}
We essentially argue this as follows. We show that with high probability, each bin created in the course of the algorithm \ref{QM2_alg} corresponding to the support element $i \in \mathcal{S}^{\gamma}_{\mathcal{P}} = \{1,2,...,m\}$ would be out of $\mathcal{C}(t)$ by $\max (T_i,T')$ rounds where $T_i$ is defined as in the argument of Theorem \ref{QM1ub}. Similarly, we show that each bin corresponding to a support element in $\{m+1,...,k\}$ would be out of $\mathcal{C}(t)$ by $T_i$ rounds with high probability. Combining together these two facts gives us the above result, details are provided in the formal proof.
The following result provides a lower bound on the expected number of queries for any $\delta$-true $\gamma$-threshold estimator under the {QM2} \ model
\begin{theorem}\label{QM2lb}
For any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM2}, let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the query complexity. Then, we have
$$\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})]\geq \max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{2\times d(p_j||\gamma)}\Bigg\}.$$
\end{theorem}
The proof of the above result involves constructing a $\delta$-true $\gamma$-threshold estimator $\mathcal{A}'$ for the {QM1} \ model using a $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM2}. The lower bound on the query complexity as given in the Theorem \ref{QM2lb} is close to the upper bound in Theorem~\ref{QM2ub} when $\min\{d^{*}(p_m||\gamma),d^{*}(p_{m+1}||\gamma)\} \ll d^{*}(p_i||\gamma)$ $\forall$ $i \notin \{m, m+1\}$ and $T'$ is smaller than $\frac{2e\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}(p_1,\gamma)}\Bigr)}{(e-1).d^{*}(p_1,\gamma)}$.
In this case, the terms corresponding to $i=m$ and $i=m+1$ dominate in the upper bound on the query complexity of Algorithm \ref{QM2_alg} as given in Theorem \ref{QM2ub}.
We would like to have a lower bound on the query complexity of $\delta$-true $\gamma$-threshold estimators which matches the upper bound more generally and towards this goal, we consider a slightly altered setting which relates closely to the thresholding problem in the Multi Armed Bandit (MAB) setting \cite{MABbook}. \remove{The details are given in Section 1 of the supplementary material.}
\remove{
}
\section{INTRODUCTION}
Estimating the likely `heavy hitters' amongst the possible outcomes of an unknown probability distribution can be a useful primitive for several applications, ranging from clustering and natural language processing to network flow / cache management and online advertising. In this work, we formulate and study a Probably Approximately Correct (PAC) sequential estimation problem in which the learner has access to a stream of independent and identically distributed (i.i.d.) samples from an underlying unknown distribution over a given support set via an oracle which it can query. The goal of the learner is to identify the set of elements in the support of the distribution whose probability value is above a pre-defined threshold. We will henceforth refer to this problem as \textit{threshold-based support identification}. We consider two natural models for oracle queries: \textit{(a) direct query} where the learner provides a sample index and the oracle responds with the value of the corresponding sample; and \textit{(b) pairwise query} where the learner queries the oracle with a pair of indices and the oracle responds with a binary answer confirming whether the sample values are identical or not. Note that in the pairwise query model, the true values of the samples are not revealed. While the former model has been a staple in the online learning literature \cite{devroye2002distribution}, the latter has also received significant attention recently under a wide variety of settings \cite{mazumdar2017clustering, chien2018query, jamieson2011active, mazumdar2016clustering,NIPS2016_6499}. The broad goal of our work is to design query-efficient schemes for these oracle models which can reliably estimate the support elements with probability values above a given threshold.
A concrete application of the above setting (and a key motivation for the formulation) can be found in a clustering problem where we have an underlying collection of items which can be partitioned into a given number of clusters based on some inherent property, for example, products in a shopping platform based on category or a population of individuals based on their political preferences. However, unlike the usual setting where the goal is \textit{full clustering}, i.e., mapping each individual item to its corresponding cluster \cite{mazumdar2017clustering,mazumdar2017query}, we consider \textit{partial clustering} where the goal is relaxed to identifying cluster indices with size larger than a certain fraction of the population. Now, say a stream of samples is generated from the population by picking an item uniformly at random in each instant. Then, it is easy to see that the probability of picking an item from a certain cluster is proportional to the size of the cluster and thus, the problem of identifying clusters with size greater than a fraction of the population would correspond to solving the threshold-based support identification problem under the corresponding sampling distribution as mentioned above. Our results provide schemes for performing partial clustering which can have significantly smaller query complexity than the naive approach of performing full clustering and then selecting the appropriate clusters. For example, say we have $n$ items with a ground truth clustering with $k = \Theta(\sqrt{n})$ clusters, where the three largest clusters have size $n^{7/8}$ and the remaining clusters are of size $O(\sqrt{n})$. If we have to identify all the clusters with size more than $n^{3/4}$ with access to a pairwise oracle, Theorem~\ref{QM2ub} shows that this can be performed using $O(n)$ queries whereas the naive approach would require $O(nk) = O(n^{3/2})$ queries \cite{mazumdar2016clustering}.
We make the following contributions to understanding the query complexity for this class of sequential estimation problems. For both the query models, we first design sequential algorithms which can provably solve the threshold-based support identification problem with large probability and also provide upper bounds on their query complexity. The estimators are based on maintaining empirical estimates and confidence intervals for the probability values associated with each element of the support. We also provide information-theoretic lower bounds on the query complexity of any reliable estimator under these query models. The bounds presented are instance-specific, i.e., they depend on the underlying probability distribution and the value of the threshold. Finally, we also consider noisy versions of both the direct query as well as the pairwise query models, and propose robust estimators which can effectively counter the noise in the oracle responses.
\subsection{Related Work}
Estimation of properties of distributions using samples is a direction of work which has a long and rich history. While some of the older works were concerned with the statistical properties of the estimators such as consistency etc., see for example \cite{chernoff1964estimation, parzen62estimation} which study mode estimation, there has been a lot of work recently on characterizing the optimal query complexity for estimating various properties of probability distributions including entropy \cite{caferov2015optimal, acharya2016estimating}, support size and coverage \cite{hao2019data, wu2018sample}, and `Lipschitz' properties \cite{hao2019unified} amongst others. Another related line of work is the `heavy hitter' estimation problem in the context of streaming algorithms \cite{sivaraman2017heavy, bhattacharyya2018optimal, karp2003simple} where given an \textit{arbitrary} stream of samples from a large alphabet, the goal is to identify those symbols whose frequency is above a certain threshold. The metric of performance is usually computational in nature such as the memory size, the number of passes or run time complexity. In contrast to these works, we study the threshold-based support identification problem in the \textit{stochastic} setting and our interest is in deriving instance-specific bounds on optimal query complexity which illustrate the dependence on the underlying distribution. Also, we study the query complexity for reliable estimation under several query models including the pairwise query model which isn't as prevalent in the literature.
Online decision making and active learning are also key features of the popular framework of Multi-Armed Bandits (MABs) \cite{MABbook}. In particular, the problem of \textit{thresholding bandits} \cite{locatelli2016optimal} aims to find the set of arms whose mean reward is above a certain threshold, using the minimum number of arm pulls. Thinking of each element in the support set as an arm, the MAB problem is indeed quite related to the setting studied in this work, especially the one with pairwise queries and we explore this relationship in this paper. One key difference between the MAB problem and our setup is that in the former we can choose to pull any arm at each instance, whereas in our problem the arm to be pulled is determined exogenously by the samples generated by the underlying probability distribution.
The two lines of work closest to ours in spirit are \cite{mazumdar2017clustering,mazumdar2017query} and \cite{shah2020sequential, 8732224}. The goal in \cite{mazumdar2017clustering,mazumdar2017query} is to fully cluster a collection of items and they characterize the optimal query complexity\footnote{Unlike our work, these results are worst-case and not instance-specific} for this task. In this context, our work corresponds to the less stringent objective of identifying the `significantly large' clusters which might be a natural objective in several unsupervised machine learning applications where we wish to associate labels to a large fraction of the unlabelled dataset and has not been studied before, to the best of our knowledge. On the other hand, \cite{shah2020sequential, 8732224} study the related problems of identifying the top-$m$ clusters and mode of an underlying distribution and find bounds on the optimal query complexity for these tasks. While the broad structure of our algorithms and lower bound techniques are similar, the details differ greatly, for example in the choice of confidence intervals in achievability and alternate instances for the converses. One additional technical challenge that we face due to our different objective is that the number of clusters to be recovered is not fixed apriori (all or one or $m$), but depends on the particular problem instance and this requires careful treatment in the algorithms as well as in the analysis.
\section{Proofs}
\label{Sec:Proofs}
\input{proof_lemma1.tex}
\input{proof_theorem1.tex}
\input{proof_theorem2.tex}
\input{proof_lemma2.tex}
\input{proof_theorem3.tex}
\input{proof_theorem4.tex}
\input{proof_theorem5.tex}
\input{proof_lemma3.tex}
\input{proof_theorem6.tex}
\input{proof_lemma10.tex}
\input{proof_ineq.tex}
\bibliographystyle{IEEEtran}
\section{Thresholding under QM1 noisy setting }
\section{Threshold-estimator under {QM2-N}}
\label{Sec:QM2N}
Recall that in this model, we make pair-wise comparisons between two samples, and the oracle responses are incorrect with a probability of error $p_e$.
\subsection{Algorithm}
Recall that for the noisy {QM1-N} \ model, we had established an equivalence to the noiseless {QM1} \ model with a modified threshold $\gamma '$. We would like to establish a similar relation of the {QM2-N} \ model to the noiseless {QM2} \ model studied before. But unlike the {QM1} \ and {QM1-N} \ models, where the $k$ distinct bins are available apriori, the bins need to be formed using pairwise queries in the {QM2} \ and {QM2-N} \ models. This represents the key challenge under the {QM2-N} \ model and in spite of the erroneous pairwise queries, we need to find a reliable method to create bins such that with high probability, all indices in a bin correspond to the same element in the support and different bins correspond to different elements.
Broadly speaking, our scheme operates in two phases. The objective of the first phase is to extract a collection of bins, each with at least a certain number of indices in it, which includes one corresponding to each element in $\mathcal{S}^{\gamma}_{\mathcal{P}}$. For this phase, we borrow ideas from \cite{mazumdar2017clustering, 8732224} which study the problem of clustering using noisy pairwise queries. The second phase is similar in spirit to the second phase of the estimator for the {QM2} \ model as described in Algorithm~\ref{QM2_alg}. We create and maintain confidence intervals for each bin by comparing a new sample in each round with representative indices from a subset of the bins extracted in the first phase. We compare the confidence intervals thus created with a a modified threshold $\gamma '$ to reliably identify bins corresponding to elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$.
We now describe the two phases of the estimator in some more detail, see Algorithm~\ref{QM2n_alg} for the pseudocode. In the first phase, we consider a natural number $T_0$ (defined as in \eqref{T0_defn}) and form a complete graph $\mathcal{G}$ using the first $T_0$ samples, such that each sample corresponds to a vertex of the graph. We query the oracle for each pair of vertices $i$ and $j$ and assign the oracle response as the weight to the edge between vertices $i$ and $j$. For any subgraph $\mathcal{S}$ of the graph $\mathcal{G}$, let $wt(\mathcal{S})$ denote its weight given by the sum of the weights on all the edges in $\mathcal{S}$. The maximum weighted subgraph (MWS) denotes the subgraph $\mathcal{S}$ corresponding to the largest weight $wt(\mathcal{S})$. Starting with the graph $\mathcal{G}$, we repeatedly extract and remove the MWS, as long as the size of the extracted MWS is greater than $S_0 = \frac{\gamma T_0}{4}$. Corresponding to each such extracted MWS, we create a different bin. Let $\mathcal{C}'(T_0)$ denotes the set of bins formed. We will argue that with high probability, each extracted bin corresponds to a unique element in the support and that there is a bin corresponding to each element in $\mathcal{S}^{\gamma}_{\mathcal{P}}$.
The second phase is from round $T_0$ onwards. In this phase, we first choose an index from each bin as a representative element of the bin. In particular, say for each $j\in \mathcal{C}'(T_0)$ we choose index $j_l$ from Bin $j$ as its representative element. For each round $t > T_0$, we consider the $t^{th}$ sample and define $\mathcal{Z}_j^t$ as follows.
\begin{equation}\label{indicator_noisy}
\mathcal{Z}_j^t=
\begin{cases}
1, & \text{if}\ \mathcal{O}(j_l,t)= +1 \\%\text{ for } j_l<T_0 \in \text{ Bin } j\\
0, & \text{otherwise}.
\end{cases}
\end{equation}
We define $\tilde{\rho}_j^t$ as the fraction of samples with index larger than $T_0$ that belong to Bin $j$, which can be formally written as follows.
\begin{equation}\label{eq_noisy}
\tilde{\rho}_j^t = \frac{\sum_{r=T_0+1}^{t} \mathcal{Z}_j^r }{t-T_0}.
\end{equation}
$\mathcal{L}_j(t)$ and $\mathcal{U}_j{(t)}$ denote the lower and upper confidence bounds of Bin $j$ respectively, and are defined as
\begin{equation}
\mathcal{L}_j(t)= {\min}\{ q \in [0,\tilde{\rho}_j^t]: (t-T_0)\times d(\tilde{\rho}_j^t||q) \leq \beta^t \},\label{eqn_noisylb}
\end{equation}
\begin{equation}
\mathcal{U}_j(t) = {\max}\{ q \in [\tilde{\rho}_j^t,1]: (t-T_0)\times d(\tilde{\rho}_j^t||q) \leq \beta^t \}.\label{eqn_noisyub}
\end{equation}
In this phase, $\mathcal{C}'(t)$ is defined as
\begin{align}\label{eqn:c2tnoise}
\mathcal{C}'(t)=\{x| \mathcal{L}_x(t)< \gamma'<\mathcal{U}_x(t) , x \in \mathcal{C}'(t-1) \}.
\end{align}
where, $\gamma'=(1-2p_e)\gamma+p_e$.
In each round $t> T_0$, we go over the bins in $\mathcal{C}'(t-1)$ one by one, and query about the samples ${j_l}$ and $t$, $\forall j\in\mathcal{C}'(t-1)$.
We add Sample $t$ to all the bins for which the oracle provides a positive response. In this phase, we also initialize an empty set $\mathcal{S}$, and we update it in each round by adding any bin index $x$ such that $\mathcal{L}_x(t)>\gamma'$. These are the bin indices which the algorithm believes corresponds to elements in $\mathcal{S}_{\mathcal{P}}^{\gamma}$. We run this phase as long as $ \mathcal{C}'(t)$ is non-empty and return the set of bin indices $\mathcal{S}$ upon termination.
The choice of the modified threshold $\gamma '$ above follows from the following observation. Consider a bin $j$ in the second phase and say it corresponds to some support element $i$. Then the expected fraction of indices added to each bin $j$ in the second phase by some round $t > T_0$, denoted by $\tilde{\rho}_j^t$, is equal to $(1-p_e)\times p_i + p_e \times (1 - p_i) = (1-2p_e)p_i + p_e$.
\remove{
{\color{red}
Intuitively, the idea behind choosing $\gamma'$ as $(1-2p_e)\gamma+p_e$ can be expressed as follows. This is because $\mathbb{E}[\mathcal{Z}_j^t]$ = ${p_i}'$ = $(1-2p_e)\times p_i+p_e$ if Bin $j$ contains indices corresponding to support element $i$. Since lower confidence bounds and upper confidence bounds are created with respect to (w.r.t) $\tilde{\rho}_j^t$, it makes more sense to create the termination criterion w.r.t $\gamma'= (1-2p_e)\gamma + p_e$. }
}
\begin{algorithm}{\label{QM2n_alg}}
\footnotesize{
\SetAlgoLined
Define $T_0$ as per \eqref{T0_defn} and $S_0= \frac{\gamma.T_0}{4}$\\
$t$= 1 \\
{Create a graph $\mathcal{G}$ with just one node labeled $t$.}\\
\While{$t$ {<} $T_0$}{
{$t=t+1$\\}
Create a new node labelled $t$.\\
\remove{. COMMENT: Owing to red line above, won't this create a node for $t = 1$ again Reply: Changed.. }
$j = 1$ \\
\While{$j < t$}{
Create an edge of weight $\mathcal{O}(j,t)$ between nodes $j$ and $t$.\\
$j = j + 1$\\
}
\remove{$t = t+1$}
}
Extract Maximum Weighted Sub-graph (say $G'$) from this graph.
\While{$|G'|> S_0$}{
Put all the nodes corresponding to $G'$ in a new bin.
Remove all nodes from $\mathcal{G}$ which were part of $G'$ and edges which were incident to these nodes.\\
Extract a new Maximum Weighted Sub-graph $G'$ from $\mathcal{G}$.\\
}
Denote this extracted set of bins by $\mathcal{C}'(T_0)$
\ForAll{ $j$ $\in$ $\mathcal{C}'(T_0)$}{
Pick representative index $j_l$ $\in$ Bin $j$;
}
Initialise $\mathcal{S}$ to $\Phi$.\\
\While{$\mathcal{C}'(t)$ $\neq$ $\phi$}{
{$t = t+1$\\}
\remove{\color{red}Update $\mathcal{C}'(t)$ according to \eqref{eqn:c2tnoise}.}
\ForAll{ $j$ $\in$ $\mathcal{C}'(t-1)$}{
\If{$\mathcal{O}(j_l,t) == +1$}{
Put index $t$ in bin $i$.
}
}
Update the empirical estimate $\tilde{\rho}_j^t$ according to \eqref{eq_noisy} and the confidence bounds $\mathcal{L}_j(t)$ and $\mathcal{U}_j(t)$ according to \eqref{eqn_noisylb} and \eqref{eqn_noisyub} respectively $\forall$ bins $\in$ $\mathcal{C}'(t-1)$.
Update $\mathcal{C}'(t)$ according to \eqref{eqn:c2tnoise}.\\
Update $\mathcal{S}$ by adding the bins with $\mathcal{L}_x(t)>\gamma$.
\remove{$t = t + 1$}
}
Output $\mathcal{S}$.\\
\remove{\color{red}This pseudocode needs work. Firstly $b(t)$ should be replaced appropriately by $C'(T_0)$. Also 'forall the' can be replaced by `forall'. Also, I think the various while loops are off by one unit because the $t = t+1$ line comes at end. For eg, the while loop for second phase starts at $t = T_0$ again even though that should have been dealt with in first phase. Also first phase seems to have $T_0 - 1$ samples the way it is written now. Also, text said a list $\mathcal{S}$ is maintained which is updated with any bin with $LCB > \gamma '$ but that is not mentioned here at all. We need to be consistent with the text. This needs a careful read and adjustment.}
\caption{Estimator for {QM2-N}}
}
\end{algorithm}
The following lemma claims that Algorithm \ref{QM2n_alg} returns the desired set $S_{\mathcal{P}}^{\gamma}$ for any underlying distribution $\mathcal{P}$ with probability at least $1- \delta$.
\begin{lemma}\label{correctness_noisy}
Given the choice of $\beta^t = \log (\frac{4k(t-T_0)^2}{\delta})$ for each $t > T_0$, where $T_0$ is as defined in \eqref{T0_defn}, Algorithm \ref{QM2n_alg} is a $\delta$-true $\gamma$-threshold estimator under {QM2-N}.
\end{lemma}
\subsection{Query complexity Analysis}
The following theorem provides an upper bound on the query complexity of our proposed estimator in Algorithm \ref{QM2n_alg}.
\begin{theorem}\label{QM2_noisyub}
Let $\mathcal{A}$ denote the estimator in Algorithm~\ref{QM2n_alg} with $\beta^t = \log (\frac{4k(t-T_0)^2}{\delta})$ for each $t > T_0$ where $T_0$ is as defined in \eqref{T0_defn}. Let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM2-N} \ and define $q = \min\{T_0,k\}$, ${p_i}'= (1-2p_e)\times p_i+p_e$ and $\gamma' = (1-2p_e)\gamma+p_e$. Then we have $$Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A}) \leq \Bigl(\sum_{i=1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}({{p}_i}',\gamma')}\Bigr)}{(e-1).d^{*}({p_i}',\gamma')} + \frac{T_0(T_0-1)}{2} \Bigr)$$
with probability at least $(1- 2\delta)$.
\end{theorem}
\subsection{Proof of Lemma \ref{first_ineq}}{\label{first_ineq_proof}}
Let us restate and prove Lemma \ref{first_ineq}.
\begin{lemma*
For all $1 \geq p_t, \gamma \geq 0$, the following inequality holds true.
$$p_t \log \Bigl(\frac{p_t}{\gamma}\Bigr)+ (2.\gamma) \log\Bigl(\frac{2.\gamma}{p_t + \gamma}\Bigr) \leq 2 d(p_t||\gamma).$$
\end{lemma*}
\begin{proof}
Consider the function $f(\gamma) = - p_t \log (\frac{p_t}{\gamma-\epsilon})- (2.\gamma) \log(\frac{2.\gamma}{p_t + \gamma}) + 2 d(p_t||\gamma)$.
On differentiating the function with respect to $\gamma$, we have
$$f'(\gamma) = -\frac{p_t}{\gamma} + \frac{2(1-p_t)}{(1-\gamma)} - 2.\log\Bigl(\frac{2.\gamma}{p_t+\gamma}\Bigr) - \frac{2.p_t}{p_t+\gamma}.$$
Note that the $f'(\gamma) = f(\gamma) = 0$ for $\gamma = p
_t$.
Now on double differentiating $f(\gamma)$, we have
\begin{align*}
f''(\gamma) = & \frac{p_t}{\gamma^2} + \frac{2(1-p_t)}{(1-\gamma)^2} - \frac{2.{p_t}^2}{\gamma{(p_t+\gamma)}^2}
= \frac{2(1-p_t)}{(1-\gamma)^2} + \frac{p_t({p_t}^2+\gamma^2)}{\gamma^2.{(p_t+\gamma)}^2} \geq 0.
\end{align*}
Using these results we conclude that $f(\gamma) \geq 0 $ $\forall$
$1 \geq p_t,\gamma \geq 0$, proving our lemma.
\end{proof}
\subsection{Proof of Lemma \ref{Correctness_QM1}}
We use the following lemmas to prove Lemma \ref{Correctness_QM1}. The following lemma is motivated from Lemma 4 in \cite{pmlr-v30-Kaufmann13}.
\begin{lemma}\label{lemma_cb}
Let $l_i(t)$ and $u_i(t)$ be the lower and upper confidence bounds respectively of bin index $i$ and are defined in equations \eqref{eqnlb} and \eqref{eqnub} respectively. Then,
\begin{center}
$\mathbb{P}(l_i(t) > p_i) \leq \exp(-\beta^t)$,\\
$\mathbb{P}(u_i(t) < p_i) \leq \exp(-\beta^t)$.
\end{center}
\end{lemma}
\begin{proof}
If $l_i(t)=0$, then the corresponding bound is trivial. For $l_i(t) > 0$, we prove $\mathbb{P}(l_i(t) > p_i)\leq \exp(-\beta^t)$. From equation \eqref{eqnlb}, the event $l_i(t)>p_i$ implies that $t\times d(\tilde{p}_i^t||p_i) > \beta^t$. From the properties of continuity and monotonicity of KL-divergence, there exists $x$ such that $p_i < x < \tilde{p}_i^t$ and $t\times d(x||p_i)=\beta^t$. Thus, we have $\mathbb{P}(l_i(t) > p_i) \le $ $\mathbb{P} (\tilde{p}_i^t >x)\overset{(a)}{\leq}\exp(-t\times d(x||p_i))=\exp(-\beta^t)$, where $(a)$ follows from the Chernoff bound for Binomial random variables, see \cite[Section 1.3]{Concentration_Inequalities}.
By following similar arguments, we can also prove that $\mathbb{P} (u_i(t) < p_i)\leq \exp(-\beta^t)$.
\end{proof}
\remove{
{\color{red}Comment: Added this lemma \ref{termination}- Might be needed for arguing termination... Have a look..
REPLY-NK: Let's discuss this separately.
I didn't remove this lemma. You may remove it if you feel unimportant.
}
\begin{lemma}{\label{termination}}
Given the choice of $\beta^t = \log(\frac{2kt^2}{\delta})$, there exists positive integer $T_{m}$ such that $t .d(\tilde{p}_i^t||\gamma) > \beta^t$ $\forall$ $t>T_{m}$ with probability 1 for every support element $i$.
\end{lemma}
\begin{proof}
Let us try to compute the probability of the complementary event (denoted by $A^c$) i.e, the probability that there does not exist $T_m$ such that $t.d(\tilde{p}_i^t||\gamma) > \beta^t$ $\forall$ $t>T_{m}$.
This event could equivalently expressed as there existing an infinite sequence of $\{t_i\}_{i \in \mathbb{N}}$ such that $t.d(\tilde{p}_i^t||\gamma) \leq \beta^t$ $\forall$ $t \in \{t_i\}_{i \in \mathbb{N}}$.
Now, we know that $\frac{\beta^t}{t}$ is decreasing for sufficiently large $t$ and asymptotically goes to 0, thus the above event implies that $\lim_{n \to \infty} d(\tilde{p}_i^{t_n}||\gamma) \leq \lim_{n \to \infty} \frac{\beta^{t_n}}{t_n} = 0$, hence $\lim_{n \to \infty} \tilde{p}_i^{t_i} = \gamma$.
Let us attempt to compute the probability of the event $A^c$.
\begin{align*}
& \mathbb{P}[A^c]\\
\leq & \lim_{n \to \infty} \mathbb{P}[\bigcup_i {t_n}.d(\tilde{p}_i^{t_n}||\gamma) \leq \beta^{t_n}]\\
\overset{(a)}{\leq} & \sum_{i=1}^{k} \mathbb{P}[\lim_{n \to \infty} {t_n}.d(\tilde{p}_i^{t_n}||\gamma) \leq \beta^{t_n}]\\
\overset{(b)}{\leq} & \sum_{i=1}^{k} \mathbb{P}[\lim_{n \to \infty} \tilde{p}_i^{t_n} = \gamma]\\
\overset{(c)}{\leq} & \sum_{i=1}^{k} \lim_{n \to \infty} \exp({t_n}.d(p_i||\gamma))\\
\overset{(d)}{\leq} & 0
\end{align*}
Note the inequality $(a)$ follows from continuity of probability measure and union bound. $(b)$ follows from the statement proven in previous paragraph. $(c)$ follows from Chernoff inequality on Bernoulli random variable-section 1.3 in \cite{10.5555/1146355}. $(d)$ follows from the fact that $d(p_i||\gamma) > 0$ $\forall$ $i$ as $p_i \neq \gamma$.
\end{proof}
}
Now we restate and prove Lemma \ref{Correctness_QM1}.
\begin{lemma*
Given the choice of $\beta^t = \log(2kt^2 / \delta)$ for each $t\ge 1$, Algorithm \ref{QM1_alg} is a $\delta$-true $\gamma$-threshold estimator.
\end{lemma*}
\begin{proof}
\remove{
We first argue the termination of the algorithm. From Lemma \ref{termination}, we know there exists $T_m$ such that $t.d(\tilde{p}_i^t||\gamma) > \beta^t$ $\forall$ $t>T_m$ $\forall$ $i$. This would imply from \eqref{eqnlb} and \eqref{eqnub} that either both $u_i(t)$ and $l_i(t)$ is greater than $\gamma$ or both are less than $\gamma$, thus the algorithm terminates with probability 1.
}
%
\remove{\color{red}Let $\mathcal{E}_i^t$ denote the event that $p_i$ does not lie in [$l_i(t)$, $u_i(t)$] at time $t$.}
For each $i \leq m$, let $\mathcal{E}_i^t$ denote the event that $p_i > u_i(t)$ at time $t$. On the other hand, for $i > m$, let $\mathcal{E}_i^t$ denote the event that $p_i < l_i(t)$ at time $t$. From Lemma \ref{lemma_cb}, we have that $\mathbb{P} (\mathcal{E}_i^t) \leq \delta / 2kt^2$. \remove{= \delta / 2kt^2.}
Let $\mathcal{E}$ denote the event that there exists a pair ($i$, $t$) such that $p_i$ lies above $u_i(t)$ if $i\leq m$, $t \geq 1$ or $p_i$ lies below $l_i(t)$ if $i>m$, $t \geq 1$.\remove{, $i \in \{1,2,\dots,k\}$.} Then,
\begin{equation}
\mathbb{P}[\mathcal{E}] = \displaystyle \mathbb{P}[ \bigcup_{i,t} \mathcal{E}_{i}^{t} ] \leq \sum_{i,t} \mathbb{P} [\mathcal{E}_{i}^{t}] \leq \sum_{i,t} \frac{\delta}{2kt^2} \leq \delta.
\end{equation}
We now argue that if the event $\mathcal{E}^{c}$ holds true, the algorithm will correctly return the desired set of support elements $\mathcal{S}_{\gamma}^{\mathcal{P}}$. The termination condition of the algorithm specifies that for each bin, either the LCB lies above $\gamma$ or the UCB lies below $\gamma$. Since we return the set of all bins which have their LCB above $\gamma$ as the estimate $\widehat{S}$ for $\mathcal{S}_{\gamma}^{\mathcal{P}}$, and the event $\mathcal{E}^{c}$ ensures that for each bin index $i\leq m$, $p_i \leq u_i(t)$, and for each index $i>m$ $p_i \geq l_i(t)$, the correctness of our estimator is guaranteed. In particular, the LCB can be greater than $\gamma$ only for indices $i$ such that $p_i > \gamma$. Similarly, the UCB can be smaller than $\gamma$ only for indices $j$ such that $p_j < \gamma$.
Thus, $\mathbb{P}_{\mathcal{P}}[\widehat{S} = \mathcal{S}_{\gamma}^{\mathcal{P}}] \geq \mathbb{P}_{\mathcal{P}}[\mathcal{E}^c]$ $\geq$ $1-\delta$ and hence Algorithm \ref{QM1_alg} is a $\delta$-true $\gamma$-threshold estimator.
\remove{\color{red} Comment: Need a line somewhere to argue that the algorithm will terminate. Not very sure whether we can argue that. Have to give more thought...}
\end{proof}
\subsection{Proof of Lemma \ref{lemma_extract}}{\label{lemma_extract_proof}}
We use the following claims to prove Lemma \ref{lemma_extract}. These claims and their proofs are very similar to those in \cite{mazumdar2017clustering, 8732224}. Recall the following terms defined in main paper.\\
$c_1= (\log 2 + 1) $; $c_2 = \frac{3} {4} (1 - 2p_e)^2 $ ; ${k_1}'= \frac{e}{2\pi}$ ; ${k_2}'= \frac{1}{2} \exp (2(1-2p_e)^2)$. Additionally, let $c$ denote the largest root of the equation $e^{x(1-2p_e)^2} = x$.
We now define $K'$ as follows.
\begin{align}\label{K_prime_defn}
K'= \max \Bigl\{c,\frac{2c_1}{c_2} ,\frac{c_1}{2.c_2}
+ \sqrt{ (1/c_2)(e/(e-1)) \Bigl(\log \Bigl(\frac{2k{k_1}'}{c_2 \delta}\Bigr)+ \Bigl(\frac{c_1^2}{4c_2}\Bigr)\Bigr) }, \nonumber\\
\frac{e}{e-1}\frac{\log\Bigl(\frac{{k_2}'}{ ((1-2p_e)^2)}\sqrt{\frac{2k}{\delta}} \Bigr)}{(1-2p_e)^2}\Bigr\}.
\end{align}
\begin{claim}\label{comp_subcluster}
Consider a graph $\mathcal{G}(\hat{V},\hat{E})$ where $|\hat{V}| \geq K'$ defined in \eqref{K_prime_defn} and edge weights are i.i.d. random variables taking the value $-1$ with probability $p_e$ $< \frac{1}{2}$ and $1$ with probability
$(1 - p_e)$. Then, $wt(\mathcal{G}) > wt(\mathcal{G}')$ for any subgraph $\mathcal{G}' \subset \mathcal{G}$, i.e.,
the MWS extracted from $\mathcal{G}$ will include the entire node set $\hat{V}$ with probability at least $(1- \frac{\delta}{k})$.
\end{claim}
\begin{proof}
We will use the following inequalities in the proof below.
\begin{equation}\label{bound1}
\mbox{Stirling's inequality \cite{feller1}}: \sqrt{2\pi} n^{n+\frac{1}{2}}.e^{-n} < n! < e.n^{n+\frac{1}{2}}.e^{-n} \ \forall \ n \ge 1.
\end{equation}
\begin{equation}\label{bound2}
\mbox{\cite[Appendix]{Concentration_Inequalities}} : \ \mathbb{P}(X \leq \mathbb{E}[X]-t) \leq
\exp\Bigl(-\frac{t^2}{2n}\Bigr),
\end{equation}
\begin{equation}
\text{where } t > 0, X = \sum_{i=1}^{n}X_i \text{ such that } \{X_i\} \text{is a set of i.i.d. radom varibles. } \nonumber
\end{equation}
Let $S$ be a subset of $\hat{V}$ and we try to compute the probability that the $wt(S)$ is greater than $wt(\hat{V})$. Let us denote the weights of the edge between node $i$ and node $j$ as $w_{ij}$. Then
\begin{align*}
\mathbb{P} \Big(\sum_{i,j \in \hat{V};i \neq j} w_{ij} < \sum_{i,j \in S; i \neq j ;S \subseteq \hat{V}} w_{ij} \Big)
=& \mathbb{P} \Big(\sum_{(i,j) \in (\hat{V},\hat{V});(i,j) \notin (S,S);i \neq j} w_{ij} < 0 \Big)\\
\overset{(a)}{\leq} & \exp\Big( -2 {(1-2p_e)}^2 \Big[ { |\hat{V}| \choose 2} - { S \choose 2} \Big] \Big).
\end{align*}
Note that $(a)$ follows from \eqref{bound2}.\\
Applying the union bound gives us,
\begin{align*}
\hspace{-10.0 em}\mathbb{P}(\text{MWS} \neq \hat{V})
\leq &\sum_{|S|=1}^{|\hat{V}|-1} { |\hat{V}| \choose |S|} \mathbb{P}\Big(\sum_{(i,j) \in (\hat{V},\hat{V});(i,j) \notin (S,S);i \neq j} w_{ij} < 0 \Big)\\
\leq &\sum_{|S|=1}^{|\hat{V}|-1} { |\hat{V}| \choose |S|} \exp\Big( -2 {(1-2p_e)}^2 \Big[ { |\hat{V}| \choose 2} - { |S| \choose 2} \Big] \Big) \\
= & \sum_{|S|=1}^{\frac{|\hat{V}|}{2}} { |\hat{V}| \choose |S|} \exp\Big( -2 {(1-2p_e)}^2 \Big[ { |\hat{V}| \choose 2} - { |S| \choose 2} \Big] \Big) \\
& + \sum_{|S|= \frac{|\hat{V}|}{2}+1}^{|\hat{V}|-1} { |\hat{V}| \choose |S|} \exp\Big( -2 {(1-2p_e)}^2 \Big[ { |\hat{V}| \choose 2} - { |S| \choose 2} \Big] \Big) \\
\overset{(b)}{\leq} & \sum_{|S|=1}^{\frac{|\hat{V}|}{2}} { |\hat{V}| \choose \frac{|\hat{V}|}{2}} \exp\Bigl(-(1-2p_e)^2 \Bigl( \frac{3|\hat{V}|^2}{4} - \frac {|\hat{V}|}{2} \Bigr)\Bigr)\\
&+ \sum_{|S|=\frac{|\hat{V}|}{2}+1}^{|\hat{V}|} { |\hat{V}| \choose {|\hat{V}|-1}} \exp(-2(1-2p_e)^2 \Bigl( |\hat{V}| - 1 \Bigr) \Bigr)\\
\hspace{-0em} \overset{(c)}{\leq} & \frac{|\hat{V}|}{2} \frac{e}{\pi} \frac{2^{| \hat{V}|}}{\sqrt{|\hat{V}|}} \exp( \Bigl(-(1-2p_e)^2 \Bigl( \frac{3|\hat{V}|^2}{4} - \frac {|\hat{V}|}{2} \Bigr)\Bigr) \Bigr)\\ &+{k_2}' |\hat{V}|^2 \exp(-2(1-2p_e)^2(|\hat{V}|)) \\
\overset{(d)} {\leq} & {k_1}' \sqrt {|\hat{V}|} \exp(|\hat{V}| (\log 2 + 1) - (1 - 2p_e)^2 . \frac{3} {4} {|\hat{V}|^2}))\\ &+{k_2}' |\hat{V}|^2 \exp(-2(1-2p_e)^2(|\hat{V}|))\\
\overset{(e)}{\leq}& \frac{\delta}{2k} +
\frac{\delta}{2k}
\leq \frac{\delta}{k}.
\end{align*}
The inequality for the first term in $(b)$ follows since $\sum_{|S|=1}^{\frac{|\hat{V}|}{2}} { |\hat{V}| \choose |S|} \exp\Big( -2 {(1-2p_e)}^2 \Big[ { |\hat{V}| \choose 2} - { |S| \choose 2} \Big] \Big)$ takes maximum value at $|S| = \frac{|\hat{V}|}{2}$. The inequality at second term in $(b)$ can be shown by arguing that term $\sum_{|S|=\frac{|\hat{V}|}{2}+1}^{|\hat{V}|} { |\hat{V}| \choose {|\hat{V}|-1}} \exp(-2(1-2p_e)^2 \Bigl( |\hat{V}| - 1 \Bigr) \Bigr) $ takes maximum value at $S = |\hat{V}| - 1$ which we show below.
Consider the function
\begin{align*}
f(a)
= &{ |\hat{V}| \choose {|\hat{V}|-a}} \exp\Biggl(-2(1-2p_e)^2 \Biggl( {|\hat{V}| \choose 2} - {(|\hat{V}|-a) \choose 2} \Biggr) \Biggr)\\
= & { |\hat{V}| \choose {|\hat{V}|-a}} \exp((-(1-2p_e)^2) (2a|\hat{V}| - a^2+a)).
\end{align*}
We wish to show that the maximum of $f(a)$ occurs at $a = 1$.
After the requisite algebraic simplification,
\begin{align*}
\frac{f(a)}{f(a+1)}
= \frac{a+1}{|\hat{V}|-a} \frac{e^{2(1-2p_e)^2|\hat{V}|}}{e^{(1-2p_e)^2(2a+2)}}
\overset{(f)}{\geq} \frac{2}{|\hat{V}|-1} \frac{e^{2(1-2p_e)^2|\hat{V}|}} {e^{(1-2p_e)^2(2.(\frac{|\hat{V}|}{2}-1)+2)}}
\geq \frac{2}{|\hat{V}|} e^{(1-2p_e)^2|\hat{V}|}
\overset{(g)}{\geq} 1.
\end{align*}
Note that $(f)$ follows on minimising each fraction individually. $(g)$ follows from the fact that $c$ is the largest solution of $e^{(1-2p_e)^2x} = x$ which implies that $e^{(1-2p_e)^2|\hat{V}|} > |\hat{V}|$ for $|\hat{V}| > c$. Therefore, $f(a)$ is a decreasing function of $a$ and takes maximum value at $a=1$.
The inequality for first term in $(c)$ follows on applying \eqref{bound1} whereas for the second term follows on applying ${k_2}'$ = $(1/2)\exp(2(1-2p_e)^2)$. The inequality at $(d)$ follows by bounding $\exp((1-2p_e)^2\frac{|\hat{V}|}{2})$ with $\exp(|\hat{V}|)$.
Now let us prove the inequality on first term in $(e)$ by proving
\begin{equation*}
{k_1}' \sqrt {|\hat{V}|} \exp\bigg(|\hat{V}| (\log 2 + 1) - (1 - 2p_e)^2 . \frac{3} {4} {|\hat{V}|^2}\bigg) \leq \frac{\delta}{2k}.
\end{equation*}
Since $|\hat{V}| > 2.\frac{c_1}{c_2}$ and $\frac{c_1}{c_2}>1$, $\sqrt {|\hat{V}|} < \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2$.
Now consider the first term in inequality in $(d)$.
\begin{align*}
{k_1}' \sqrt {|\hat{V}|} \exp\bigg(|\hat{V}|& (\log 2 + 1)- (1 - 2p_e)^2 . \frac{3} {4} {|\hat{V}|^2}\bigg)
= {k_1}' \sqrt {|\hat{V}|} \exp(c_1|\hat{V}| - c_2 {|\hat{V}|^2})\\
\overset{(j)}{\leq} & {k_1}' \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \exp\Bigl(\frac{c_1^2}{4c_2}\Bigr) \exp\Bigl(-c_2 \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \Bigr)
\overset{(l)}{\leq} \frac{\delta}{2}
\end{align*}
$(j)$ follows from $\sqrt {|\hat{V}|} < \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2$.
Let us now prove the inequality in $(l)$.
We can say that
\begin{align*}
& \hspace{-0em} |\hat{V}| > \frac{c_1}{2.c_2} + \sqrt{\frac{e}{c_2(e-1)} \bigg(\log \Bigl(\frac{2k{k_1}'}{c_2 \delta}\Bigr)+ \Bigl(\frac{c_1^2}{4c_2}\Bigr)\bigg)} \\
& \hspace{-0em} \Rightarrow - c_2\Bigl(|\hat{V}| - \frac{c_1}{2.c_2}\Bigr)^2 < -\frac{e}{e-1} \bigg(\log \Bigl(\frac{2k{k_1}'}{c_2 \delta}\Bigr)+ \Bigl(\frac{c_1^2}{4c_2}\Bigr)\bigg)\\
& \hspace{-0em} \overset{(h)}{\Rightarrow} -c_2\Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 < W_{-1} \Bigl(\frac{c_2.\delta}{2k{k_1}'}\exp\Bigl(-\frac{c_1^2}{4c_2}\Bigr)\Bigr)\\
& \hspace{-0em} \overset{(i)}{\Rightarrow} -c_2 \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \exp \Bigl(-c_2 \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \Bigr) > -c_2 \frac{\delta}{2k{k_1}'}\exp\Bigl(-\frac{c_1^2}{4c_2}\Bigr)\\
& \hspace{-0em} \Rightarrow {k_1}' \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \exp\Bigl(\frac{c_1^2}{4c_2}\Bigr) \exp\Bigl(-c_2 \Bigl(|\hat{V}|-\frac{c_1}{2.c_2}\Bigr)^2 \Bigr) < \frac{\delta}{2k}.
\end{align*}
Recall that $W_{-1}$ is the lower root in lambert function as defined in the proof of Theorem 1 of our paper. Now using Theorem 3.1 of \cite{Lambert}, we say that $W_{-1}(-\frac{c_2.\delta}{2k{k_1}'}.\exp(-\frac{c_1^2}{4c_2}) )$ is lower bounded by $-(e/(e-1)) (\log (\frac{2k{k_1}'}{c_2 \delta})+ (\frac{c_1^2}{4c_2}))$. This would in turn imply the implication in $(h)$.
Note that the implication in $(i)$ follows from the the fact that $W_{-1}$ is the lower root of the lambert function.
Thus for $|\hat{V}| > \max\Bigl(\frac{c_1}{2.c_2} + \sqrt{ (1/c_2)(e/(e-1)) (\log (\frac{2k{k_1}'}{c_2 \delta})+ (\frac{c_1^2}{4c_2})) },\frac{2.c_1}{c_2},c\Bigr) $, the inequality in $(l)$ is proven. Therefore, the first term in $(e)$ is upper bounded by $\frac{\delta}{2k}$.
Now consider the second term on the inequality in $(e)$.
We can say the following:
\begin{align*}
& |\hat{V}| > \frac{e}{e-1}\frac{\log\Bigl(\frac{{k_2}'}{ (1-2p_e)^2}\sqrt{\frac{2k}{\delta}} \Bigr)}{(1-2p_e)^2}\\
\Rightarrow & -(1-2p_e)^2 |\hat{V}| < \frac{e}{e-1}\log\Bigl(\frac{{k_2}'}{ (1-2p_e)^2}\sqrt{\frac{2k}{\delta}} \Bigr)\\
\overset{(m)}{\Rightarrow} & -(1-2p_e)^2 |\hat{V}| < W_{-1}\Bigl(-\frac{((1-2p_e)^2)}{{k_2}'}\sqrt{\frac{\delta}{2k}} \Bigr)\\
\overset{(n)}{\Rightarrow} & (-(1-2p_e)^2|\hat{V}| \exp(-(1-2p_e)^2|\hat{V}|) > -\frac{((1-2p_e)^2)}{{k_2}'}\sqrt{\frac{\delta}{2k}} \\
\Rightarrow & |\hat{V}| \exp(-(1-2p_e)^2|\hat{V}|`) < \frac{1}{{k_2}'}\sqrt{\frac{\delta}{2k}}\\
\Rightarrow & {k_2}' |\hat{V}|^2 \exp(-2(1-2p_e)^2(|\hat{V}|)) < \frac{\delta}{2k}.
\end{align*}
Note that the implication in $(m)$ follows from theorem 3.1 of \cite{Lambert} which implies that the value of $W_{-1}\Bigl((\frac{((1-2p_e)^2)}{{k_2}'}\sqrt{\frac{\delta}{2k}} \Bigr) \Bigr)$ is lower bounded by $\frac{e}{e-1}\log\Bigl(\frac{{k_2}'}{ ((1-2p_e)^2)}\sqrt{\frac{2k}{\delta}} \Bigr)$. The implication in $(n)$ follows from the definition of $W_{-1}$ similar to the reasoning in $(i)$. Thus, the second inequality in $(e)$ is proven implying that the claim is also proven.
\end{proof}
Now we state and prove the next claim which would be used to prove Lemma \ref{lemma_extract}.
\begin{claim}\label{no_overlap}
Consider a graph $G'$ whose vertices are partitioned into multiple clusters. The weight of edges between any pair of nodes in the same cluster are random variables which take value $-1$ with probability $p_e$ and $1$ with probability $(1-p_e)$. Also weight of edges between two nodes which does not lie in the same cluster takes value $1$ with probability $p_e$ and $-1$ with probability $(1-p_e)$. We assume that all the weights of the edges are independent.\\
Let S be the MWS of $G'$. If $ |S|
\Bigl(1+\frac{33\log (\frac{2}{\delta})}{(1-2p_e)^2}\Bigr) $ then we can say with probability at least $(1-\delta)$ that it can not contain nodes from multiple sub-clusters.
\end{claim}
\begin{proof}
Let S be a sub-component contain nodes from at least two sub-clusters. Let the clusters be denoted by $V_i$ and we denote $C_i=S \cap V_i$ and $j* = \argmin_{i:C_i \neq \phi} |C_i|$.
Let the weight of the edge in the graph between the nodes $i$ and $j$ be denoted by $w_{ij}$. We claim that $$\sum_{i,j \in S,i<j} w_{ij} < \sum_{i.j \in S \textbackslash C_{j*} , i<j } w_{ij} $$ with probability at least $(1-\delta)$. The above condition is equivalent to
$$\sum_{i,j \in C_{j*},i<j} w_{ij} + \sum_{i \in S, j \in S \textbackslash C_{j*} } w_{ij} < 0$$
We use the following equations from appendix of \cite{Concentration_Inequalities} in the proof.
\begin{equation}{\label{ ineq_1}}
\mathbb{P}(X>(1+\epsilon)\mathbb{E}[X])< \exp\Bigl(-\frac{\epsilon^2}{3} \mathbb{E}[X]\Bigr) \text{ } 0<\epsilon<1.
\end{equation}
\begin{equation}{\label{ ineq_2}}
\mathbb{P}(X<(1-\epsilon)\mathbb{E}[X])< \exp\Bigl(-\frac{\epsilon^2}{3} \mathbb{E}[X]\Bigr) \text{ } 0<\epsilon<1.\end{equation}
where $ X = \sum_{i} X_i$ such that $\{X_i\}$ is a set of i.i.d. random variables. We divide the proof into two cases.
\textbf {Case 1:}
$|C_{j*}| > \sqrt{\frac{108}{1-2p_e} \log (\frac{2}{\delta})}$\\
Now
\begin{align*}
\mathbb{P}\Bigg( \sum_{i,j \in C_{j*},i<j} w_{ij}> \Big(1+\frac{1}{3}\Big) (1-2p_e) {|C_{j*}| \choose 2} \Bigg)
\overset{(a)}{\leq} & \exp\Bigg(-(1/3)^3 (1-2p_e) {|C_{j*}| \choose 2}\Bigg)\\
< & \exp\bigg(-(1/3)^3 (1-2p_e) \frac{|C_{j*}|^2}{4} \bigg)
\leq \delta / 2.
\end{align*}
$(a)$ follows from \eqref{ ineq_1} by putting $\epsilon = 1/3$.
\begin{align*}
\mathbb{P}\Bigl( \sum_{i \in C_{j*}, j \in S \textbackslash C_{j*} } w_{ij}> -(1-\frac{1}{3}) (1-2p_e) |C_{j*}| |S \textbackslash C_{j*}|\Bigr)
\overset {(b)} {\leq} & \exp(-(1/3)^2 (1-2p_e) |C_{j*}| |S \textbackslash C_{j*}|)\\
\overset {(c)} {\leq} & \exp(-(1/3)^3 (1-2p_e) \frac{|C_{j*}|^2}{2})
\leq \delta /2.
\end{align*}
The inequality in $(b)$ holds from \eqref{ ineq_1} and the inequality in $(c)$ holds due to $|S \textbackslash C_{j*}| > |C_{j*}|$ since $C_{j*}$ is the smallest cluster.
Now we apply union bound on the previous two proven events and can say with probability at least $(1-\delta)$.
\begin{align*}
\sum_{i,j \in C_{j*},i<j} w_{ij} + \sum_{i \in C_{j*}, j \in S \textbackslash C_{j*} } w_{ij}
\leq & \Bigl((4/3) (1-2p_e) {|C_{j*}| \choose 2} - (2/3) (1-2p_e) |C_{j*}| |S \textbackslash C_{j*}|\Bigr) \\
\leq &\Bigl((4/3) (1-2p_e) \frac{|C_{j*}|^2}{2} -(2/3) (1-2p_e) |C_{j*}| |C_{j*}|\Bigr)\leq 0.
\end{align*}
\textbf {Case 2:}
$|C_{j*}|< \sqrt{\frac{108}{1-2p_e} \log (\frac{2}{\delta})}$\\
\begin{align*}
\mathbb{P}\Bigl( \sum_{i \in C_{j*}, j \in S \textbackslash C_{j*} } w_{ij}> -(1-\frac{1}{2}) (1-2p_e)& |C_{j*}| |S \textbackslash C_{j*}|\Bigr)\\
\leq &\exp(-(1/2)^2 (1/3) (1-2p_e) |C_{j*}| |S \textbackslash C_{j*}|)\\
\overset {(d)} {\leq} & \exp(-(1/2)^2 (1/3) (1-2p_e) (|S|-1) )
\overset {(e)} {\leq} \delta /2.
\end{align*}
The inequality $(d)$ follows since $|C_{j*}| |S \textbackslash C_{j*}|$ $\geq$ $(|S|-1)$.
The inequality $(e)$ follows since
\begin{align*}
|S| \geq & \Bigl(1+\frac{33\log (\frac{2}{\delta})}{(1-2p_e)^2}\Bigr)
\geq \Bigl(1+\frac{12\log (\frac{2}{\delta})}{(1-2p_e)^2}\Bigr).
\end{align*}
Now take $|C_{j*}| = x$.
Hence, $$\sum_{i,j \in C_{j*},i<j} w_{ij} \leq \frac{x^2}{2}.$$
Thus, we say with probability at least (1- $\delta$) that
\begin{align*}
\sum_{i,j \in C_{j*},i<j} w_{ij} + \sum_{i \in C_{j*}, j \in S \textbackslash C_{j*} } w_{ij}
&\leq \frac{x^2}{2} - (1-2p_e) |C_{j*}| |S \textbackslash C_{j*}|/2`\\
&\leq x^2 - (1/2) (1-2p_e) x(|S|-x)\\
&\leq x (\frac{3x}{2} - (1/2) (1-2p_e) |S|)
\overset{(f)}{<} 0.
\end{align*}
The inequality at $(f)$ is true since
\begin{align*}
|S| > \Bigl(1+\frac{33\log (\frac{2}{\delta})}{(1-2p_e)^2}\Bigr)\geq \frac{3 \sqrt{\frac{108}{1-2p_e} \log (\frac{2}{\delta})}}{1-2p_e}.
\end{align*}
Thus we say with probability at least $(1-\delta)$ that the MWS contains points from a single cluster.
\end{proof}
\begin{claim}{\label{new_claim}}
Consider a support element $a$ with less than $S_0$ samples denoting it in the graph. Consider another support element $b$ with at least than $2S_0$ samples denoting it. The probability that the weight of the sub-graph which is a subset of the samples corresponding to the support element $a$ is smaller than the weight of the sub-graph containing all the samples corresponding to the support element $b$ with probability at least $(1-\frac{\delta}{16k})$.
\end{claim}
\begin{proof}
Denote the nodes corresponding to a support element $a$ as a cluster $c_a$ and denote its subset as $s_{c_a}$. Similarly, we denote the nodes corresponding to a support element $b$ as $c_b$. We denote the weights of edges between nodes $i$ and $j$ as $w_{ij}$.
We consider the probability that $\sum_{i,j \in s_{c_a}} w_{ij} > \sum_{i,j \in c_b} w_{ij}$.
\begin{align*}
\mathbb{P}[\sum_{i,j \in {s_{c_a}}} w_{ij} > \sum_{i,j \in c_b} w_{ij}]
\leq & \mathbb{P}[\sum_{i,j \in c_b} w_{ij} - \sum_{i,j \in s_{c_a}} w_{ij}<0]\\
\overset{(a)}{\leq} &\exp\Bigl(-\frac{1}{3} \mathbb{E}\Bigl[\sum_{i,j \in c_b} w_{ij} - \sum_{i,j \in s_{c_a}} w_{ij}\Bigr]\Bigr)\\
\overset{(b)}{\leq} & \exp(-\frac{1}{6} (1-2p_e) ((|c_b|-1)^2-|s_{c_a}|^2))\\
\overset{(c)}{\leq} & \exp(-\frac{1}{2} (1-2p_e) {{S_0}^2})
\overset{(d)}{\leq} \frac{\delta}{16k}
\end{align*}
Note that $(a)$ follows from \eqref{ ineq_2} and substituting $\epsilon=1$. $(b)$ follows since $\mathbb{E}\Bigl[\sum_{i,j \in c_b} w_{ij} - \sum_{i,j \in s_{c_a}} w_{ij}\Bigr] = (1-2p_e) (\frac{(|c_b|)(|c_b|-1)}{2}- \frac{(|s_{c_a}|)(|s_{c_a}|-1)}{2}) \leq \frac{1}{2}(1-2p_e) ((|c_b|-1)^2-|s_{c_a}|^2)$.
$(c)$ follows since minimum value of $|c_b|$ is $2S_0$ whereas maximum value of $|s_{c_a}|$ is $(S_0-1)$, thus $((|c_b|-1)^2-|s_{c_a}|^2)$ is lower bounded by $3{S_0}^2$. $(d)$ follows since $S_0 \geq \Bigl(1+\frac{33\log (\frac{16k}{\delta})}{(1-2p_e)^2}\Bigr)$.
\end{proof}
Let us restate and prove Lemma \ref{lemma_extract}.
\begin{lemma*
The set of extracted bins from the graph $\mathcal{G}$ at the end of the first phase, denoted by $\mathcal{C}'(T_0)$, satisfies the following properties with probability at least $\Bigl(1-\frac{5\delta}{16}\Bigr)$.
\begin{enumerate}
\item No bin contains samples corresponding to two different support elements.
\item All the support elements which have at least $2S_0$ samples corresponding to it in the first phase have an extracted bin representing it. \remove{\color{red} the bin contains all the samples denoting a support element.}
\end{enumerate}
\end{lemma*}
\begin{proof}
Let $\tilde{K}$ be defined by substituting $\frac{\delta}{8}$ for $\delta$ in the expression of $K'$ in \eqref{K_prime_defn}.
By substituting $\frac{\delta}{8k}$ for $\delta$ in Claim \ref{no_overlap}, we can say any extracted Maximum Weighted SubGraph (MWS) of size larger than $\Bigl(1+\frac{33\log (\frac{16k}{\delta})}{(1-2p_e)^2}\Bigr)$ can have indices corresponding to multiple support element with probability at most $\frac{\delta}{8k}$.
Consider each extracted MWS of size larger than $S_0$. Since, $S_0 > \Bigl(1+\frac{33\log (\frac{16k}{\delta})}{(1-2p_e)^2}\Bigr)$, we say that this MWS extracted has indices representing only a support element with probability at least $(1-\frac{\delta}{8k})$.
Consider all support elements (denoted by $\tilde{S}_{min}$) with less than $S_0$ indices denoting it. Consider all support elements (denoted by $\tilde{S}_{max}$) with more than $2S_0$ indices denoting it.
We argue that the events in Claim \ref{new_claim} and Claim \ref{comp_subcluster} (substituting $\frac{\delta}{8}$ for $\delta$) and Claim \ref{no_overlap} (substituting $\frac{\delta}{8k}$ for $\delta$) would imply for each MWS extracted, we can say that no MWS representing an element in $\tilde{S}_{min}$ would occur till MWS corresponding to all support elements in $\tilde{S}_{max}$ have been extracted.
Suppose not. Consider that MWS corresponding to some support element in $\tilde{S}_{min}$ occurs before all elements in $\tilde{S}_{max}$ have occurred in some MWS created in previous rounds. Since there are some other elements in $\tilde{S}_{max}$ not a part of MWS extracted, it would imply that a subset of all indices denoting some support element in $\tilde{S}_{min}$ has higher weight than the sub-graph corresponding to the support element in $\tilde{S}_{max}$ implying event in Claim $\ref{new_claim}$ is violated.
Now the event that no MWS representing an element in $\tilde{S}_{min}$ occurs till MWS corresponding to all support elements in $\tilde{S}_{max}$ have been extracted and events in Claim \ref{new_claim} and Claim \ref{comp_subcluster} (substituting $\frac{\delta}{8}$ for $\delta$) and Claim \ref{no_overlap} (substituting $\frac{\delta}{8k}$ for $\delta$) imply that extraction of MWS does not stop till all elements in $\tilde{S}_{max}$ have been extracted in some MWS. Suppose not. This would imply that we get an MWS of size less than $S_0$ before all elements in $\tilde{S}_{max}$ have been put in some MWS implying the event that some element in $\tilde{S}_{min}$ occurs MWS before all support elements in $\tilde{S}_{max}$ have been extracted as Claim $\ref{comp_subcluster}$ holds true for all support elements of size larger than $S_0$.
Now the probability of the events described in Claim \ref{new_claim} and Claim \ref{comp_subcluster} (substituting $\frac{\delta}{8}$ for $\delta$) and Claim \ref{no_overlap} (substituting $\frac{\delta}{8k}$ for $\delta$) can be union bounded over all support elements to argue that the probability is at least $(1-\frac{5\delta}{16})$.
Thus, we argue that both the events in theorem hold true with probability at least $(1-\frac{5\delta}{16})$.
\remove{
Consider a support element $i$ which has at least $S_0$ samples denoting it. Since $S_0 \geq \tilde{K}$, we can say that there would exist no MWS which contains only some (not all) samples of support element $i$ with probability at least $(1-\frac{\delta}{8k})$ by substituting $\frac{\delta}{8}$ for $\delta$ in Claim \ref{comp_subcluster}. Thus with probability at least $(1-\frac{\delta}{8k})$, any MWS which contains support element $i$ would have size at least $S_0$ implying that it would be extracted from the graph and have a bin corresponding to it.
However we have a total of $k$ clusters in the graph. Union bounding both the events described above for all the clusters present in the graph, we can say both the properties in Lemma \ref{all_occur} hold true with probability at least $(1- \frac{\delta}{4})$.}
\remove{\color{red}I didn’t get the reason for this.}
\end{proof}
\subsection{Proof of Lemma \ref{Correctness_QM2}}
We use the following lemmas to complete the proof. The first lemma shows that for each index in $\mathcal{S}^{\gamma}_{\mathcal{P}} = \{1,2,\ldots,m\}$, at least one bin corresponding to it is created in the first phase with high probability, while the second lemma bounds the probability of misclassification of any index $i$. Finally, we argue that two bins corresponding to the same index cannot be returned as part of the estimate $\widehat{S}$ of $\mathcal{S}^{\gamma}_{\mathcal{P}}$ and use the union bound to upper bound the total probability of error of the proposed algorithm.
\begin{lemma}{\label{lemma1:QM2}}
The probability of the event that an element from the index set $\mathcal{S}^{\gamma}_{\mathcal{P}} =\{1,2,...,m\}$ does not have any bin corresponding to it after the first phase of Algorithm \ref{QM2_alg} is bounded by $\delta/2$.
\end{lemma}
\begin{proof}
The first phase runs for the first $T' = \frac{\log (\delta/2k)}{\log(1-\gamma)}$ rounds. Consider an element $i$ $\in$ $\{1,2,3,...,m\}$. The probability that $X_t \neq i$ for any $t\ge1$ is equal to ($1-p_i$) $<$ $(1-\gamma)$. Since the samples $X_l$ are i.i.d. $\forall$ $l \in \{1,2,...,T'\}$, we can say that the probability that $X_l \neq i$ $\forall$ $l \in \{1,2,...,T'\}$ is bounded by $(1-\gamma)^{T'}$ $\leq$ $\frac{\delta}{2k}$.
Applying union bound over all the $m$ elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$, the probability that
some element in $\{1,2,...,m\}$ does not have any bin corresponding to it after $T'$ rounds is bounded by $\frac{m\delta}{2k} \leq \frac{\delta}{2}$.
\end{proof}
\remove{
{\color{red} Changed the lemma and its proof as your suggestion. I feel it has become a bit simpler. However instead of restricting $p_i$ in $[l_i,u_i]$, restricted $p_i$ below $u_i$ for $i\leq m$ and $p_i$ above $l_i$ for $i > m$. This would reduce a factor of 2 inside $\log$ in beta expression if we do it in {QM1} \ as well.}
}
\begin{lemma}{\label{lemma2:QM2}}
For a support element $i\leq m$, define $\mathcal{E}_i^t$ as the event that ${p}_i$ lies above the UCB $u_i(t)$ (defined in \eqref{eqnlb}). Similarly, for any support element $j > m$, define $\mathcal{E}_j^t$ as the event that ${p}_j$ lies below the LCB $l_i(t)$ (defined in \eqref{eqnub}). Then, a bin corresponding to support element $i$ can be misclassified by the algorithm \remove{at some round }only if the event ${\mathcal{E}}_i^t$ holds true for some $t$.
\end{lemma}
\begin{proof}
%
Before we begin, we need to distinguish between two related quantities. For a bin $b(i)$ corresponding to support element $i$, $\hat{p}_{b(i)}^t$ as defined in equation \eqref{emp_QM2} denotes the fraction of the samples till round $t$ which are placed in the bin; on the other hand, $\tilde{p}_i^t$ as defined in equation \eqref{eq1} denotes the total fraction of samples till round $t$ corresponding to support element $i$. These two can in general be different since during the course of the algorithm, multiple bins can get created corresponding to same support element.
Next, let $m_u(i)$ denote the $u^{th}$ bin created corresponding to support element $i$.
\textbf{Case 1: $i\leq m$ }:
%
%
We prove the contrapositive statement here. The event $({\mathcal{E}}_i^t)^c$ $\forall$ $t$ would imply that ${p}_i$ would be less than $u_i(t)$ $\forall$ $t$ which would imply that the first bin corresponding to support element $i$ could never be classified below $\gamma$. As the first bin corresponding to support element $i$ cannot be classified below $\gamma$, multiple bins cannot be created for support element $i$ as per Algorithm \ref{QM2_alg}. This is because multiple bins can be created for a support element only if all previous bins corresponding to the same support element have been classified below $\gamma$.
Now the empirical probability of the first bin for support element $i$ $(\hat{p}_{m_1(i)}^t)$ is same as the empirical probability of support element $i$ $(\tilde{p}_i^t)$ implying that $\hat{u}_{m_1(i)}(t) = u_{i}(t)$ . Thus, the event $({\mathcal{E}}_i^t)^c$ $\forall$ $t$ would imply that the first and only bin corresponding to support element $i$ is never misclassified.
\textbf{Case 2: $i>m$}: Suppose the $l^{th}$ bin corresponding to support element $i$, $m_l(i)$ is misclassified at round $t$, i.e., the LCB corresponding to the bin ${\hat{l}_{m_l(i)}}(t) > \gamma$. We can say $\hat{p}_{m_l(i)}^t \leq {\tilde{p}}_i^t$ where equality would hold true iff $l = 1$. Now $\hat{p}_{m_l(i)}^t \leq \tilde{p}_i^t$ would imply that ${\hat{l}_{m_l(i)}}(t)$ $\leq$ ${l_i}(t)$. Thus ${\hat{l}_{m_l(i)}}(t) > \gamma$ would imply ${l_i}(t) \geq \gamma > p_i$ implying that event $\mathcal{E}_i^t$ occurs.
%
\end{proof}
%
Now we restate and prove Lemma $\ref{Correctness_QM2}$.
%
\begin{lemma*
Given the choice of $\beta^t = \log(4kt^2/\delta)$ for each $t\ge1$, Algorithm \ref{QM2_alg} is a $\delta$-true $\gamma$-threshold estimator under {QM2}.
\end{lemma*}
\begin{proof}
We first argue that there can be no two bins returned as part of the estimate $\widehat{S}$ of $\mathcal{S}^{\gamma}_{\mathcal{P}}$ can correspond to the same support element. We argue this as follows. Suppose there exists such a pair $B_1$ and $B_2$, then both these bins must have been created in the first phase. This is possible only if one of the bins, say $B_1$ was not in $\mathcal{C}(t)$ in the round $t$ when the other bin $B_2$ was created. This would imply that $B_1$ was classified as being below $\gamma$ in the first phase, which would contradict the fact that both the bins were returned.
Next, let $\mathcal{E}_1$ denote the event that some element in $\mathcal{S}^{\gamma}_{\mathcal{P}} = \{1,2,...,m\}$ is not present in any of the bins. From Lemma \ref{lemma1:QM2}, we have $\mathbb{P}[\mathcal{E}_1] < \delta/2$.
Next, let $\mathcal{E}_2$ denote the event that there exists a misclassified bin corresponding to some support element $i$. From Lemma \ref{lemma2:QM2}, this would imply that $\mathcal{E}_i^t$ occurs for some $(i,t)$ pair. However for $i\leq m$, $\mathcal{E}_i^t$ implies $u_i(t) > p_i$ whose probability can be bounded by Lemma \ref{lemma_cb} to be at most $\frac{\delta}{4kt^2}$. Similarly, for $i>m$, $\mathcal{E}_i^t$ implies $l_i(t)<p_i$ whose probability can be bounded by Lemma \ref{lemma_cb} to be at most $\frac{\delta}{4kt^2}$. {Taking the union bound over all such pairs $(i,t)$, we obtain $\mathbb{P}[\mathcal{E}_2]$ $\leq$ $\frac{\delta}{2}$}.
\remove{ {\color{red} Comment: how? what is the probability for a given pair $(i,t)$, before you take sum ? Reply: Added in the description.
}
}
%
We can see that Algorithm \ref{QM2_alg} returns an incorrect set of bins only if at least one of the two events $\mathcal{E}_1$, $\mathcal{E}_2$ has occurred. Applying the union bound on events $\mathcal{E}_1$ and $\mathcal{E}_2$, we get that the probability of Algorithm \ref{QM2_alg} returning an incorrect set of bins is bounded by $\delta$.
\end{proof}
\subsection{Proof of Lemma \ref{correctness_noisy}}
Let us now define the variable $T_0$ which represented the number of rounds in the first phase of Algorithm \ref{QM2n_alg}. To define $S_0$, we use the following new variables $c_1= (\log 2 + 1) $; $c_2 = \frac{3} {4} (1 - 2p_e)^2 $ ; ${k_1}'= \frac{e}{2\pi}$ ; ${k_2}'= \frac{1}{2} \exp (2(1-2p_e)^2)$. Additionally, let $c$ denote the largest root of the equation $e^{x(1-2p_e)^2} = x$. Then, $T_0$ is defined\footnote{If the expression in equation \eqref{T0_defn} is not integer, choose the smallest integer greater than or equal to it.} as follows:
\begin{align}
T_0 = \frac{4}{\gamma} \max \Bigg\{ c, \frac{2c_1}{c_2} ,\frac{c_1}{2.c_2} +
&\sqrt{ (1/c_2)(e/(e-1)) (\log \Bigl(\frac{16k{k_1}'}{c_2 \delta}\Bigr)+ \Bigl(\frac{c_1^2}{4c_2}\Bigr)\Bigr) },\nonumber \\
&\hspace{0.75in}\frac{e}{e-1}\frac{\log\Bigl(\frac{{k_2}'}{ ((1-2p_e)^2)}\sqrt{\frac{16k}{\delta}} \Bigr)}{(1-2p_e)^2},
\Bigl(1+\frac{33\log (\frac{16k}{\delta})}{(1-2p_e)^2}\Bigr) \Bigg\} .{\label{T0_defn}}
\end{align}
The above definition for $T_0$ is a result of the various constraints that come up while proving the lemmas below. Also, recall that we define $S_0 = \gamma T_0 / 4$. We use the following lemmas to prove Lemma \ref{correctness_noisy}.
\begin{lemma}{\label{lemma_extract}}
The set of extracted bins from the graph $\mathcal{G}$ at the end of the first phase, denoted by $\mathcal{C}'(T_0)$, satisfies the following properties with probability at least $\Bigl(1-\frac{5\delta}{16}\Bigr)$.
\begin{enumerate}
\item No bin contains samples corresponding to two different support elements.
\item All the support elements which have at least $2S_0$ samples corresponding to it in the in the first phase have an extracted bin representing it. \remove{\color{red} the bin contains all the samples denoting a support element.}
\end{enumerate}
\end{lemma}
\remove{\color{red}This lemma follows from the ideas used in \cite{mazumdar2017clustering, 8732224}.
We move the proof the above lemma to supplementary material.}
The proof of this lemma follows from ideas presented in \cite{mazumdar2017clustering, 8732224}. We move the proof to Section~\ref{lemma_extract_proof}.
\begin{lemma}\label{all_occur}
With probability at least $(1 - \frac{\delta}{16})$, for each support element $i$ which belongs to $\mathcal{S}^{\gamma}_{\mathcal{P}}$, and thus has a probability value above $\gamma$, at least $2S_0$ corresponding samples with value $i$ would have been seen by the end of the first phase after $T_0$ rounds.
\end{lemma}
\begin{proof}
We use the following inequality in our proof which follows from appendix of \cite{Concentration_Inequalities}.
\begin{equation}{\label{ ineq_2}}
\mathbb{P}(X<(1-\epsilon)\mathbb{E}[X])< \exp\Bigl(-\frac{\epsilon^2}{3} \mathbb{E}[X]\Bigr) \text{ } 0<\epsilon<1.\end{equation}
where $ X = \sum_{i} X_i$ such that $\{X_i\}$ is a set of i.i.d. random variables
For any support element $i$ in $\mathcal{S}^{\gamma}_{\mathcal{P}}$, the number of samples, say $N^i_{T_0}$, seen by the end of $T_0$ rounds with value $i$ satisfies the following:
%
\begin{align*}
\mathbb{P}\Big(N^i_{T_0} \le 2S_0\Big) = \mathbb{P}\Big(N^i_{T_0} \le \frac{\gamma T_0}{2}\Big)\overset{(a)}{\leq} \mathbb{P}\Big(N^i_{T_0} \le \frac{\mathbb{E}[N^i_{T_0}]}{2}\Big) &\overset{(b)}{\leq} \exp(-\mathbb{E}[N^i_{T_0}] / 12)\\&\overset{(c)}{\leq} \exp(-\gamma T_0 / 12)\overset{(d)}{\leq} \delta/16k.
\end{align*}
%
where $(a)$ and $(c)$ follow from $\mathbb{E}[N^i_{T_0}] \ge \gamma T_0$ for any $i \in \mathcal{S}^{\gamma}_{\mathcal{P}}$; $(b)$ follows from \eqref{ ineq_2} by substituting $\epsilon = \frac{1}{2}$ \remove{\color{red}which version of the bound is this? refer to it as discussed before Comment - SS : updated}and $(d)$ follows since from equation \eqref{T0_defn}.
%
\begin{align*}
\frac{\gamma T_0}{12} \geq \ \frac{1}{3}\cdot \Bigl(1+\frac{33\log (\frac{16k}{\delta})}{(1-2p_e)^2}\Bigr) > \ {\log \Bigl(\frac{16k}{\delta}\Bigr)}.
\end{align*}
%
The lemma follows by taking the union bound over all the support elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$.
\end{proof}
Now we restate and prove Lemma \ref{correctness_noisy}.
\begin{lemma*
Given the choice of $\beta^t = \log (\frac{4k(t-T_0)^2}{\delta})$ for each $t > T_0$, where $T_0$ is as defined in \eqref{T0_defn}, Algorithm \ref{QM2n_alg} is a $\delta$-true $\gamma$-threshold estimator under {QM2-N}.
\end{lemma*}
\begin{proof}
Let $\mathcal{E}$ denote the event that Algorithm \ref{QM2n_alg} correctly identifies the support elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$. Also, let $\mathcal{E}_1$ and $\mathcal{E}_2$ be the events that the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} respectively are satisfied. Then, we have
$$\mathbb{P}[\mathcal{E}^c] \leq \mathbb{P}[\mathcal{E}^c|(\mathcal{E}_1 \cap \mathcal{E}_2)] + \mathbb{P}[(\mathcal{E}_1 \cap \mathcal{E}_2)^c].$$
From Lemmas~\ref{lemma_extract} and \ref{all_occur}, we have
$$ \mathbb{P}[(\mathcal{E}_1 \cap \mathcal{E}_2)^c] = \mathbb{P}[\mathcal{E}_1^c \cup \mathcal{E}_2^c] \le \frac{\delta}{2}.$$
What remains is to show that $\mathbb{P}[\mathcal{E}^c|(\mathcal{E}_1 \cap \mathcal{E}_2)] \le \delta / 2$. Henceforth, assume that the events $\mathcal{E}_1$ and $\mathcal{E}_2$ hold true. Note that this implies that when bins are extracted at the end of the first phase of Algorithm \ref{QM2n_alg} after $T_0$ rounds, there will be a unique bin corresponding to each element in $\mathcal{S}^{\gamma}_{\mathcal{P}}$. There might be additional bins corresponding to other support elements as well.
Next, we consider the second phase of the algorithm where the goal is to identify the bins corresponding to support elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$. As done in the second phase of Algorithm \ref{QM2_alg} for {QM2}, for each round $t > T_0$, we compare the $t^{th}$ sample with a fixed representative element chosen from each bin belonging to a subset $\mathcal{C}'(t)$. Again similar to Algorithm \ref{QM2_alg}, confidence intervals are maintained for each bin and eventually those bins for which the LCB becomes larger than a threshold are identified as the ones corresponding to support elements in $\mathcal{S}^{\gamma}_{\mathcal{P}}$.
The only differences between the second phases of Algorithms \ref{QM2_alg} and \ref{QM2n_alg} are in the values of the empirical estimates and the confidence intervals associated with each bin as well as the value of the threshold against which they are compared. Recall that for $t > T_0$ and a bin $j$ representing support element $i$, $\tilde{\rho}_j^t$ defined in equation \eqref{eq_noisy} represents the fraction of samples since the beginning of the second phase for which the oracle provided a positive response when queried with the representative index from bin $j$. Note that $\tilde{\rho}_j^t$ is a running average of a sequence of i.i.d Bernoulli random variables, each with expected value ${p_i}'$= $p_i\times(1-p_e) + p_e\times(1-p_i) = (1-2p_e)\times p_i + p_e$. Thus as before, Lemma \ref{lemma_cb} applies and can be used to devise the confidence bounds $\mathcal{L}_j(t)$ and $\mathcal{U}_j(t)$ for ${p_i}'$. Finally, the modified threshold is given by $\gamma '$ = $(1-2p_e)\gamma + p_e$ and the bins whose LCB becomes larger than $\gamma '$ will be classified as corresponding to elements from $\mathcal{S}^{\gamma}_{\mathcal{P}}$.
Given the similarities of the two schemes, the arguments for proving the correctness of the above scheme run exactly parallel to the ones made in Lemma \ref{Correctness_QM2} for Algorithm \ref{QM2_alg} and we skip them here for brevity.
\end{proof}
\remove{
\begin{lemma}\label{correctness_cond_noisy}
Given that the events $\mathcal{E}_1$ and $\mathcal{E}_2$ occurred, for the choice of ${\beta}^t = \log \Bigl(\frac{4k(t-T_0)^2}{\delta}\Bigr)$ $\forall \ t > T_0$, Algorithm when terminates returns the desired set of bins with probability at least $\Bigl(1- \frac{\delta}{2}\Bigl)$
\end{lemma}
\begin{proof}
Note that the event $\epsilon_1$ and $\epsilon_2$ imply that there is a unique distinct bin corresponding to each support element in $\{1,2,...,m\}$. We might have some bins corresponding to other support elements as well.
Recall that in this phase of the algorithm we compare each new element with a fixed representative element chosen from each bin.
Recall the indicator random variable $\mathcal{Z}_j^t$ as defined in \ref{indicator_noisy} for $t > T_0$. Therefore, $\mathbb{E}[\mathcal{Z}_j^t]$ = ${p_i}'$= $p_i\times(1-p_e) + p_e\times(1-p_i) = (1-2p_e)\times p_i + p_e$ where $i$ denotes the support element that indices less than $T_0$ in bin $j$ represent.
We claim that Lemma \ref{lemma_cb} applies in our scenario for each bin where the confidence bounds are replaced by $\mathcal{L}_j(t)$ and $\mathcal{U}_j(t)$ respectively, the fraction of indices larger than $T_0$ in Bin $j$ post Round $T_0$ is $\tilde{\rho}^t_j$
and probability of a new index $t>T_0$ being classified in Bin $j$ is ${p_i}'$ where $i$ denotes the support element that indices in Bin $j$ less than $T_0$ represent. This is true because $\tilde{\rho}_j^t = \frac{\sum_{r=T_0}^{t} \mathcal{Z}_j^r }{t-T_0}$ where
$\mathcal{Z}_j^r$ denotes a set of i.i.d random variables for each $j$ with expectation ${p_i}'$ which would be sufficient for proving Lemma \ref{lemma_cb} in our scenario.
Thus the setting is similar to our {QM1} and
by replacing $\gamma$ by $\gamma'$ and using similar arguments of Lemma \ref{Correctness_QM2}, we prove Lemma \ref{correctness_cond_noisy}.
\end{proof}
Recall the events $\epsilon_1$ and $\epsilon_2$ defined in Lemmas \ref{correctness_cond_noisy} respectively.
Let $\epsilon$ denote the event that the algorithm when terminates, returns a correct set of bins.
Using conditional probability, we can show that
$\mathbb{P}[\epsilon^c] \leq \mathbb{P}[\epsilon^c|(\epsilon_1 \cap \epsilon_2)] + \mathbb{P}[(\epsilon_1 \cap \epsilon_2)^c] \leq \mathbb{P}[\epsilon^c|(\epsilon_1 \cap \epsilon_2)] + \mathbb{P}[\epsilon_1^{c} \cup \epsilon_{2}^{c}] \leq \mathbb{P}[\epsilon^c|(\epsilon_1 \cap \epsilon_2)] + \mathbb{P}[\epsilon_1^{c}] + \mathbb{P}[\epsilon_{2}^{c}].$
We use Lemma \ref{correctness_cond_noisy} to bound $\mathbb{P}[\epsilon^c|(\epsilon_1 \cap \epsilon_2)]$ by $\frac{\delta}{2}$, Lemma \ref{lemma_extract} to bound $\mathbb{P}[\epsilon_1^{c}]$ by $\frac{\delta}{4}$ and Lemma \ref{all_occur} to bound $\mathbb{P}[\epsilon_2^{c}]$ by $\frac{\delta}{4}$. Thus, we bound the probability of the algorithm returning an incorrect set of bins by ${\delta}$.
\end{proof}
}
\subsection{Proof of Theorem \ref{QM1ub}}
We use the following lemma to prove Theorem \ref{QM1ub} which effectively characterises the number of rounds by which a bin numbered $i$ is classified, i.e. either its LCB goes above $\gamma$ or UCB goes below $\gamma$.
\begin{lemma}\label{lemma_exittime}
Let $T_i$ be the smallest positive integer such that $\forall$ $t > T_i$ the condition $t \times d^{*} (p_i,\gamma) >\beta^t$ is satisfied. For any $t > T_i$, the bin corresponding to support element $i$ is not classified as either above $\gamma$ or below $\gamma$ after $t$ rounds with probability at most $\exp(-\beta^{t})$.
\end{lemma}
\begin{proof}
Consider the event that a bin numbered $i$ for which $p_i < \gamma$ is not yet classified after $t > T_i$ rounds. If this happens, the upper confidence bound $u_i(t)$ is still above $\gamma$ and the lower confidence bound $l_i(t)$ is still below $\gamma$. From equation \eqref{eqnlb} and \eqref{eqnub}, the event $l_i (t) < \gamma < u_i (t)$
implies that $t \times d(\tilde{p}_i^{t}||\gamma) <\beta^{t}$.
\remove{
Hence,
\begin{align*}
\mathbb{P}(l_i (T_i)<\gamma < u_i (T_i)) \
\leq \ & \mathbb{P}(T_i \times d(\tilde{p}_i^{T_i}||\gamma)<\beta^{T_i})
\end{align*}
}
{This implies that there exists $y$ such that $y < \gamma$ and $y < \tilde{p}_i^{t}$ such that $t \times d(y||\gamma) =\beta^{t}$. On the other hand, from the definition of $T_i$ in the statement of the lemma and given $p_i < \gamma$ and $t>T_i$, we have $y>p_i$. Thus, we have $p_i < y < \min\{\gamma, \tilde{p}_i^{t}\}$.
Given the series of implications mentioned above, we have the following series of inequalities:
\begin{align}
\nonumber
\mathbb{P}(l_i (t)<\gamma < u_i (t) ) \ \leq & \ \mathbb{P} (t \times d(\tilde{p}_i^{t}||\gamma) <\beta^{t})
\ \leq \ \mathbb{P}(\tilde{p_i}^{t} >y )
\ \leq \ \exp(-t \times d(y||p_i))
\label{eqn:ineq1}
\end{align}
%
where, the last inequality follows from the Chernoff bound for Binomial random variables, see \cite[Section 1.3]{Concentration_Inequalities}.
%
Since $t \times d^{*} (p_i,\gamma) >\beta^{t}$ and {$t \times d(y||\gamma) =\beta^{t}$}, we have $t \times d(y||p_i) >\beta^{t}$. Then, \remove{from equation \eqref{eqn:ineq1}}we have
$$
\mathbb{P}(l_i(t) <\gamma < u_i (t) ) \
\leq \ \exp(-\beta^{t}),
$$
%
thus proving the statement of the lemma for all bin indices $i$ such that $p_i < \gamma$. Similar arguments can be used to prove the result for the bins $j$ with $p_j > \gamma$.
}
\end{proof}
\remove{\color{red} Comments: I have some doubts on the proof of Lemma 6 as written here. Had some suggestions which I highlighted in red. I believe $T_i$ should be bounded by the larger root of the equation not smaller root and the function is not decreasing $\forall$ $t$ only decreases for sufficietly large $t$.}
\begin{lemma}\label{lemma_timeupper}
Given $\beta^t = \log(2kt^2/ \delta)$ for $t \ge 1$, let $T_i$ be the smallest positive integer such that $\forall$ $t > T_i$ the condition
$t \times d^{*} (p_i,\gamma) >\beta^t$ is satisfied. Then, setting $a_i = d^{*}(p_i,\gamma)/2$ and $b = \log(2k /\delta) / 2$, we have
%
$$
T_i \le \frac{e(b-\log(a_i))}{(e-1)a_i} .
$$
\end{lemma}
\begin{proof}
%
%
%
%
%
%
Since $\frac{\beta^t}{t} = \frac{\log(2kt^2/ \delta)}{t}$ is decreasing in $t$ for sufficiently large $t$. \remove{\color{red}sufficiently large $t$ Comment NK- Checking manually for small $t$, it seems it is always decreasing..so are there multiple roots for this eqn
Reply - SS: Yeah, there are two roots for this equation in most of the cases. Upto a very small value of $t=t_0$ the function increases very fast-, then rapidly decays. However $t_0$ is less than 2 (worst case for $k=1,\delta=1$),
}Thus, it is clear that $T_i$ is upper bounded by the largest root $r^*$ of the equation $t \times d^{*} (p_i,\gamma) = \log(2kt^2 / \delta)$. Letting $a_i$ = $d^{*}({p}_i,\gamma)/2$, $b$ = $\log(\frac{2k}{\delta}) / 2$, this largest root is given by $r^* = \frac{-1}{a_i}\times W_{-1}(-a_i e^{-b})$ where $W_{-1}(y)$ provides the smallest root of $xe^x=y$ for $y < 0$ and denotes the Lambert function \cite{Lambert}.
From \cite[Theorem 3.1]{Lambert}, we have $W_{-1}(y)> -\frac{e}{e-1}\log(-y)$ and thus $T_i \le r^* \le \frac{e(b-\log(a_i))}{(e-1)a_i}$. This concludes the proof of the lemma.
\remove{
Consider the equation $t \times d^{*} (p_i,\gamma) = \log(2kt^2 / \delta)$ which can be written as $t\times a_i = b+\log t$ and can be solved as $\frac{-1}{a_i}\times W(-a_i e^{-b})$ where $a_i$ = $d^{*}({p}_i,\gamma)/2$, $b$ = (1/2)* $\log(\frac{2k}{\delta})$, and W denotes Lambert function. Note that the lambert function denotes roots for the equation $xe^x=y$ as a function of $y$. Note that the Lambert function takes 2 values for $y<0$ and let $W_{-1}(y)$ denote its smaller root of equation $xe^x=y$ (only defined for $y<0$).
Using Theorem 3.1 of \cite{Lambert}, $W_{-1}(y)>-\frac{e}{e-1}\log(-y)$ and $xe^x>y$ $\forall x<W_{-1}(y)$. By applying this to $\frac{-1}{a_i}\times W_{-1}(-a_i e^{-b})$, we can say it is upper bounded by $\frac{e(b-\log(a_i))}{(e-1)a_i}$.
Note that $T_i$ as defined in Lemma \ref{lemma_timeupper} is $W_{-1}(-a_i e^{-b})$ which has been upper bounded by $\frac{e(b-\log(a_i))}{(e-1)a_i}$.
{\color{red} The above proof is currently in a really convoluted form and has to have a simpler exposition. It should be a simple thing. Please take another pass, even though I have a sense of what the argument is, its been written in such a winded form that I can't follow it}
}
\end{proof}
Now, we will restate and prove Theorem \ref{QM1ub}.
\begin{theorem*
Let $\mathcal{A}$ denote the estimator in Algorithm~\ref{QM1_alg} with $\beta^t = \log(2kt^2 / \delta)$ for each $t\ge 1$ and let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM1}. Then, we have
$$Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})\leq \max_{j\in\{m,m+1\}} \Biggl\{\frac{2e\log \Bigl(\sqrt{\frac{2k}{\delta}}\frac{2}{d^{*}(p_j,\gamma)}\Bigr)}{(e-1)d^{*}(p_j,\gamma)} \Biggr\},$$
with probability at least $1-\delta$.
\end{theorem*}
\begin{proof}
\remove{
{\color{red}Comment: Before editing the proof here, I want to understand two things: a) Why is the probability in the below paragraph $\frac{\delta}{2k{T_i}^2}$ and not $\frac{\delta}{4k{T_i}^2}$; and (b) why do we target the non-classification probability by $T_i$ for bin $i$ to be $\frac{\delta}{2k{T_i}^2}$ and not $\frac{\delta}{2k}$. Shouldn't this be sufficient since just a union bound over the various bins is needed and this bound will also suffice.}
}
From Lemmas \ref{lemma_exittime} and \ref{lemma_timeupper}, we have that for each $i \in \{1,2,\ldots,k\}$, the probability that bin $i$ has its UCB above $\gamma$ and LCB below $\gamma$ beyond $\frac{e.(b-\log(a_i))}{(e-1).a_i}$ rounds is bounded by $\frac{\delta}{2k{T_i}^2}$. Here, $a_i$ = $d^{*}({p}_i,\gamma)/2$ and $b = \frac{1}{2} \cdot \log(\frac{2k}{\delta})$.
Taking the worst case number of rounds and applying the union bound over all the $k$ bins, we get that the probability that all bins have been classified by $\max_{i\in [1,k]} \frac{e.(b-\log(a_i))}{(e-1).a_i}$ rounds is greater than or equal to $1-\delta$. It can be verified that $\frac{e.(b-\log(x))}{(e-1)x}$ is decreasing in $x$. Since $a_m \leq a_j$ $\forall$ $j$ $<$ $m$ and $a_{m+1} \leq a_j$ $\forall$ $j$ $>$ $m+1$, the expression for query complexity can be simplified from $\max_{i\in [1,k]} \frac{e.(b-\log(a_i))}{(e-1).a_i}$ to $\max\Bigl\{\frac{e.(b-\log(a_m))}{a_m.(e-1)},\frac{e.(b-\log(a_{m+1}))}{a_{m+1}.(e-1)}\Bigr\}$.
\remove{
{\color{red}. \\
COMMENT: Doesn't the above statement need that the expression for upper bound on $T_i$ be decreasing in $a_i$? Is that obvious?
We can check that $\frac{e.(b-\log(a_i))}{a_i.(e-1)}$ is decreasing with $a_i$ by differentiating as long as it is greater than 0.
}
}
\end{proof}
\subsection{Proof of Theorem \ref{QM1lb}}{\label{Sec:Thm2_proof}}
The following lemma follows from the proof of \cite[Theorem3]{shah2019sequentialME} and provides a recipe for deriving lower bounds on the query complexity of $\delta$-true $\gamma$-threshold estimators. The proof of this lemma follows along similar lines as that for \cite[Theorem3]{shah2019sequentialME} which is based on standard change of measure arguments \cite{kaufmann2016complexity}, and is skipped here for brevity.
\begin{lemma}{\label{shah_paper}}
For any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$, let $\tau$ be the stopping time of the algorithm. Then, we have
\begin{equation*}
\mathbb{E}_\mathcal{P}[\tau]\geq \frac{\log{\frac{1}{2.4\delta}}}{\displaystyle \inf_{\mathcal{P}': S_{\mathcal{P}}^{\gamma} \neq S_{\mathcal{P}'}^{\gamma}}D(\mathcal{P}||\mathcal{P}')}
\end{equation*}
\end{lemma}
Now we restate and prove Theorem \ref{QM1lb}.
\begin{theorem*
For any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM1}, let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the query complexity. Then, we have
$$\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})]\geq \max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{d(p_j||\gamma)}\Bigg\}.$$
\end{theorem*}
\begin{proof}
We show that $\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})]\geq \frac{\log{1 / 2.4\delta}}{d(p_m||\gamma)}$, the other inequality follows similarly.
Lemma~\ref{shah_paper} above requires a choice of distribution $\mathcal{P}'$ such that $S_{\mathcal{P}}^{\gamma} \neq S_{\mathcal{P}'}^{\gamma}$. For some small $\epsilon > 0$, we choose $\mathcal{P}'$ as follows:
\begin{align*}
p'_m=\gamma-\epsilon,\quad \text{ and }
p'_i=\frac{1-\gamma+\epsilon}{1-p_m}p_i , \ \forall \ i\neq m.
\end{align*}
Then, we have
%
\begin{align*}
D(\mathcal{P}||\mathcal{P}') &= p_m \log\left(\frac{p_m}{\gamma - \epsilon}\right) + \sum_{i \neq m} p_i \log\left(\frac{p_i}{\frac{1-\gamma+\epsilon}{1-p_m}p_i}\right) \\
&= p_m \log\left(\frac{p_m}{\gamma - \epsilon}\right) + (1-p_m)\log\left(\frac{1 - p_m}{1 - \gamma + \epsilon}\right)\overset{(a)}{\approx} d(p_m||\gamma),
\end{align*}
%
where $(a)$ follows since $\epsilon$ can be made arbitrarily small. Then, from Lemma~\ref{shah_paper}, we have
$$
\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})] \geq \frac{\log{\frac{1}{2.4\delta}}}{d(p_m||\gamma)}.
$$
%
\remove{ We can similarly prove the other lower bound as well.}
\end{proof}
\subsection{Proof of Theorem \ref{QM2ub}}
Recall that we defined $a_i = d^{*}({p}_i,\gamma)/2$ and $b = \frac{1}{2} \cdot \log(\frac{4k}{\delta})$. We use the following lemmas to prove Theorem \ref{QM2ub}.
\remove{
\begin{lemma}{\label{union_bound_error_for_all_t}}
Let $u_i(t)$ be defined as in $\eqref{eqnub}$. Then given any support element $i$, the probability that $\exists \ t>\frac{e(b-\log(a_i))}{(e-1)a_i}$ such that $u_i(t) > \gamma$ is bounded by $\frac{\delta}{2}$.
\end{lemma}
\begin{proof}
Let $K_i = \frac{e(b-\log(a_i))}{(e-1)a_i}$. Then,
\begin{align*}
\mathbb{P} [\exists\ t>K_i \text{ s.t } u_i(t) > \gamma]
\overset{(a)}{\leq} & \sum_{t=K_i+1} \mathbb{P} [u_i(t) > \gamma]\\
\overset{(b)}{\leq} & \sum_{t=K_i+1} \exp(-\beta^{t})\\
\leq & \frac{\delta}{2k}
\end{align*}
%
where $(a)$ follows from the union bound and {\color{red}$(b)$ follows from Lemma \ref{lemma_exittime}. COMMENT: I don't see how?}
\end{proof}
}
\remove{
\begin{lemma}{\label{union_bound_error_for_all_t}}
Let $l_i(t)$ and $u_i(t)$ be as defined in equations \eqref{eqnlb} and \eqref{eqnub} respectively. Then given any support element $i>m$\remove{and $(\mathcal{E}_i^t)^c$ (defined in Lemma \ref{lemma2:QM2}) occurs $\forall$ $i,t$}, the probability that $\exists \ t>\frac{e(b-\log(a_i))}{(e-1)a_i}$ such that $u_i(t) > \gamma>l_i(t)$ is bounded by $\frac{\delta}{2k}$.
\end{lemma}
\begin{proof}
Note that $l_i(t)$ and $u_i(t)$ denote the lower and upper confidence bounds for $p_i$ and involve $\tilde{p}_i^t$ which as defined in equation \eqref{eq1} denotes the total fraction of samples till round $t$ corresponding to support element $i$. Let $K_i = \frac{e(b-\log(a_i))}{(e-1)a_i}$. Then,
\begin{align*}
\mathbb{P} [\exists\ t>K_i \text{ s.t } u_i(t) > \gamma>l_i(t)]
\overset{(a)}{\leq} & \sum_{t=K_i+1}^{\infty} \mathbb{P} [u_i(t) > \gamma>l_i(t)]\\
\overset{(b)}{\leq} & \sum_{t=K_i+1}^{\infty} \exp(-\beta^{t})\\
\leq & \frac{\delta}{2k}
\end{align*}
%
where $(a)$ follows from the union bound and $(b)$ follows from Lemmas \ref{lemma_exittime} and \ref{lemma_timeupper}.
\remove{COMMENT: I don't see how?
Reply: $(b)$ follows since $l_i(t)<\gamma$, thus not classified. Thus from Lemma \ref{lemma_exittime} the probability of the bin not being classified after $t$ rounds is bounded by $\exp(-\beta^t)$.}
\end{proof}
}
\remove{\color{red}Comment (SS) : Slightly changed the wordings in the lemma to make the proof more clear. I think it would be good if the wordings in proof of theorem 3 also reflects this change. Added a sentence to reflect this slight change. Highlighted those in red.}
\begin{lemma}{\label{queries with each box}}
\remove{Given that \remove{event $(\mathcal{E}_i^t)^c$ (defined in Lemma \ref{lemma2:QM2}) occurs $\forall$ $i,t$} the algorithm correctly returns all the bins,} Each of the following statements is true with probability at most $\frac{\delta}{k}$.
\begin{itemize}{\Roman{enumii}}
\item The total number of queries with all the bins representing support element $i\leq m$ is greater than $\max\left\{\frac{e(b-\log(a_i))}{(e-1)a_i},T'\right\}$ and the algorithm {correctly classifies all the bins\remove{\color{red} with no misclassification at any round}.}\remove{corresponding to support element $i$}
\item The total number of queries with all the bins representing support element $i>m$ is greater than $\frac{e(b-\log(a_i))}{(e-1)a_i}$ and the algorithm correctly classifies all the bins\remove{\color{red} with no misclassification at any round}.
\end{itemize}
\end{lemma}
\begin{proof}
\remove{\color{red}Comment: Edited the proofs as per our discussion. }
{Consider any bin which contains indices representing support element $i$.}\\
\remove{We know from Lemma \ref{lemma2:QM2} that $(\mathcal{E}_i^t)^c$ $\forall$ $i,t$ implies that the algorithm correctly returns the desired set of bins.}
\begin{enumerate}
\item $ i \leq m$: Let us denote the bin with samples representing support element $i$ as $m_i$. All bins corresponding to support element $i$ being correctly classified and $i \leq m$ would imply that Bin $m_i$ would be classified above $\gamma$. According to Algorithm \ref{QM2_alg}, no bin classified above $\gamma$ would have another bin corresponding to the same support element created again. Therefore, $m_i$ is the first and last bin created for support element $i$ implying $\hat{p}_{m_i}^t=\tilde{p}_i^{t}$.
\remove{\color{red} Then Lemma \ref{lemma_exittime} tells us that after $\frac{e(b-\log(a_i))}{(e-1)a_i}$ rounds, Bin $m_i$ would have LCB below $\gamma$ and UCB above $\gamma$ with probability at most $\delta/ 2k$. However we know that no bin with UCB above $\gamma$ comes out of $\mathcal{C}(t)$ before $T'$ rounds.
{Thus with probability at most $\delta/ 2k$, we say the bin $m_i$ would not exit the set $\mathcal{C}(t)$ after \\ $\max\left\{\frac{e(b-\log(a_i))}{(e-1)a_i},T'\right\}$ queries with bins representing support element $i$ and the algorithm correctly classifies $m_i$. \remove{the set of bins corresponding to support element $i$.}}}
At round ${T_i}' = \max\left\{\frac{e(b-\log(a_i))}{(e-1)a_i},T'\right\}$, Algorithm \ref{QM2_alg} would have certainly proceeded to the second phase. Hence, the event that bin $m_i$ does not drop out of the subset of bins $\mathcal{C}(t)$ against which new samples are compared after ${T_i}'$ rounds would imply that the condition in equation \eqref{eqn:c2t} is satisfied after ${T_i}'$ rounds. This requires ${l}_i({T_i}') <\gamma < {u}_i({T_i}')$, which from Lemma \ref{lemma_exittime} and Lemma \ref{lemma_timeupper} is true with probability at most $\frac{\delta}{2k}$.
\remove{During a run of Algorithm \ref{QM2_alg}, there might be multiple bins created for support element $i$. Let $m_l(i)$ denote the the $l^{th}$ bin created for support element $i$, if created. Now we can say Bin $m_l(i)$ would be in $\mathcal{C}(t)$ until it is classified below $\gamma$.
Let us denote $\epsilon_i$ as the event that $\exists$ $t>\frac{e(b-\log(a_i))}{(e-1)a_i}$ s.t $u_i(t)>\gamma>l_i(t)$, where $u_i(t)$ and $l_i(t)$ are defined in \eqref{eqnub} and \eqref{eqnlb} respectively.
We wish to compute the probability that there exists bin $m_l(i)$ of support element $i$ which is not classified after $t > \frac{e(b-\log(a_i))}{(e-1)a_i}$ rounds. We can show that $\hat{p}_{m_l(i)}^t \leq \tilde{p}_i^{t}$ where equality holds iff $l = 1$. This is because for $l>1$, at least one sample denoting support element $i$ has gone into Bin $m_1(i)$. Now $\hat{p}_{m_l(i)}^t \leq \tilde{p}_i^{t}$ implies $\hat{u}_{m_l(i)}^t \leq {u}_i^t$. Thus, the event $\hat{u}_{m_l(i)}^t>\gamma$ implies the event $u_i^t>\gamma$. Thus, the event $\hat{u}_{m_l(i)}^t>\gamma$ for some $t>\frac{e(b-\log(a_i))}{(e-1)a_i}$ implies the event $\epsilon_i$ or event $l_i(t) > \gamma$ for some $t$. This is true for all bins $m_l(i)$ created for support element $i$, hence the event $\exists$ bin $m_l(i)$ of support element $i$ which is not classified after $t>\frac{e(b-\log(a_i))}{(e-1)a_i}$ rounds implies the event $\epsilon_i$ or event $l_i(t)>\gamma$.
Also the total number of rounds is lower bounded by the total number of queries with bins representing support element $i$. Thus the event that there exists bin $m_l(i)$ of support element $i$ which is not classified after more than $\frac{e(b-\log(a_i))}{(e-1)a_i}$ queries with bins representing support element $i$ implies the event that there exists bin $m_l(i)$ of support element $i$ which is not classified after $t > \frac{e(b-\log(a_i))}{(e-1)a_i}$ rounds which implies the event $\epsilon_i$ or event $l_i(t)>\gamma$ for some $t$ as shown in the previous paragraph.
Thus the probability of the event $\exists$ bin $m_l(i)$ of support element $i$ which is not classified after more than $\frac{e(b-\log(a_i))}{(e-1)a_i}$ queries with bins representing support element $i$ is bounded by the sum of probability of $\epsilon_i$ which is bounded by Lemma \ref{union_bound_error_for_all_t} by $\frac{\delta}{2k}$ and probability of the event $\exists t \text{ s.t }l_i(t)>\gamma$ which can be bounded by Lemma \ref{lemma_cb} by union bound to be $\sum\limits_t \frac{\delta}{4kt^2}= \frac{\delta}{2k}$.\remove{since $(\mathcal{E}_i^t)^c$ holds true $\forall$ $i,t$.}
Thus the probability of the event $\exists$ bin $m_l(i)$ of support element $i$ which is not classified after more than $\frac{e(b-\log(a_i))}{(e-1)a_i}$ queries with bins representing support element $i$ is bounded by $\frac{\delta}{k}$.}
\item $i > m$: In the course of a run of Algorithm \ref{QM2_alg}, there might be multiple bins created for support element $i>m$ even if the algorithm correctly classifies all the bins. Let $m_l(i)$ denote the the $l^{th}$ bin created for support element $i$, if created. Recall that $\hat{p}_{m_l(i)}^t$ as defined in equation \eqref{emp_QM2} denotes the fraction of the samples till round $t$ which are placed in the bin $m_l(i)$; on the other hand, $\tilde{p}_i^t$ as defined in equation \eqref{eq1} denotes the total fraction of samples till round $t$ corresponding to support element $i$. It is easy to see that $\hat{p}_{m_l(i)}^t \leq \tilde{p}_i^{t}$ where equality holds iff $l = 1$. The bin $m_1(i)$ not being out of $\mathcal{C}(t)$ after $K_i = \frac{e(b-\log(a_i))}{(e-1)a_i}$ rounds\remove{the bin $m_1(i)$ not being misclassified} would imply the event $\hat{l}_{m_1(i)}(K_i)< \gamma < \hat{u}_{m_1(i)}(K_i)$ which is equivalent to the event $l_{i}(K_i)< \gamma < u_{i}(K_i)$ \remove{and whose probability can be bounded using Lemma \ref{lemma_exittime} and \ref{lemma_timeupper} by $\frac{\delta}{2k}$} or the event $\gamma <\hat{l}_{m_1(i)}(K_i) $ which is equivalent to the event $p_i<\gamma< l_i(K_i)$. \remove{whose probability can be bounded by Lemma \ref{lemma_cb} $\forall t$ by $\sum_i \frac{\delta}{4kt^2} \leq \frac{\delta}{2k}$} The probability of the event $p_i<\gamma< l_i(K_i)$ is bounded using Lemma \ref{lemma_cb} by $\frac{\delta}{2k}$. The probability of the event $l_i(K_i)<\gamma<u_i(K_i)$ is bounded by Lemma \ref{lemma_exittime} and Lemma \ref{lemma_timeupper} by $\frac{\delta}{2k}$.
Thus the probability of the event bin $m_1(i)$ not being out of $\mathcal{C}(t)$ after $K_i$ rounds is bounded by sum of probabilities of the event $p_i<\gamma< l_i(K_i)$ and the event $l_{i}(K_i)< \gamma < u_{i}(K_i)$ which can be bounded by $\frac{\delta}{2k}+\frac{\delta}{2k} \leq \frac{\delta}{k}$.
\remove{{\color{red}Comment: Why should the LCB be below $\gamma$ at $K_i$.\\
Reply-SS : If the LCB is not below $\gamma$, the bin $m_1(i)$ would be misclassified at round $K_i$ as both LCB and UCB are above $\gamma$ though it may not be out of $\mathcal{C}(t)$ if $K_i <T'$. To avoid confusion, I am adding the part no misclassification at any round. }}
Now consider a bin $m_l(i)$, $l>1$ created at some round $t_l$. Since the bin $m_l(i)$ was created at some round $t_l$ and all bins are correctly classified, bin $m_1(i)$ must have been classified correctly below $\gamma$ in some previous round $t_1 < t_l$. It is easy to see that $\hat{p}_{m_l(i)}^{t_l} < \hat{p}_{m_1(i)}^{t_1}$ which would imply that $\hat{u}_{m_l(i)}^{t_l} < \hat{u}_{m_1(i)}^{t_1} \overset{(a)}{<} \gamma$ and thus the bin $m_l(i)$ would immediately be out of $\mathcal{C}(t)$ after its creation at round $t_l$ and there would be no queries with bin $m_l(i)$ $\forall$ $l>1$. Note that $(a)$ follows since the bin $m_1(i)$ is correctly classified below $\gamma$.
\remove{\color{red}What does classification mean here? That might not happen till the end of phase 1. Also, is it clear that LCB and UCB are monotonic in time
Reply-SS: UCB going below $\gamma$ immediately implies being out of $\mathcal{C}(t)$ irrespective of whichever phase the algorithm is in. Also since we know that the bounds i.e $\frac{\beta^t}{t}$ is decreasing, we can say if $\hat{p}_{m_l(i)}^{t_l} < \hat{p}_{m_1(i)}^{t_1}$ then, $\hat{u}_{m_l(i)}^{t_l} < \hat{u}_{m_1(i)}^{t_1}$ }
Thus the probability of total number of queries with all the bins denoting support element $i>m$ is greater than $\frac{e(b-\log(a_i))}{(e-1)a_i}$ and the algorithm correctly classifies all the bins is upper bounded by $\frac{\delta}{k}$.
\end{enumerate}
\end{proof}
Let us restate and prove Theorem \ref{QM2ub}
\begin{theorem*
Let $\mathcal{A}$ denote the estimator in Algorithm~\ref{QM2_alg} with $\beta^t = \log(4kt^2 / \delta)$ for each $t\ge 1$ and let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM2}. We define $q$ as $\min\left\{k,\frac{\log (\delta/2k)}{\log(1-\gamma)}\right\}$. Then, we have
\begin{align*}
Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})\leq \sum_{i=1}^{m} \max \Biggl\{ \frac{2e\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}(p_i,\gamma)}\Bigr)}{(e-1).d^{*}(p_i,\gamma)} ,\frac{\log (\delta/2k)}{\log(1-\gamma)} \Biggr\}
\hspace{0 em} + \sum_{i=m+1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}(p_i,\gamma)}\Bigr)}{(e-1).d^{*}(p_i,\gamma)},
\end{align*}
with probability at least $1-2\delta$.
\end{theorem*}
\remove{In the proof, we essentially bound the error probability by first conditioning on the correctness of algorithm and then bound the net error probability of algorithm.}
\begin{proof}
The total number of bins created during the course of the algorithm must be bounded by $T' = \frac{\log (\delta/2k)}{\log(1-\gamma)}$ since new bins are created only in the first phase of the algorithm and at most one new bin can be created in each round. Thus, the number of bins corresponding to distinct support elements is upper bounded by $q = \min\left\{k,\frac{\log (\delta/2k)}{\log(1-\gamma)}\right\}$. Furthermore, if the algorithm returns a correct set of bins, at least one bin corresponding to each element in $\{1,2,...,m\}$ must have been created.
\remove{
However, the total number of bins corresponding to different support elements must be bounded by $T'$. Thus the total number of bins corresponding to distinct support elements created must be upper bounded by $q=\min(k,T')$.
}
\remove{Given that the algorithm returns a correct set of bins,} Using Lemma \ref{queries with each box} and taking the union bound over all support elements $i \in \{1,2,...,k\}$, the probability of the event that the algorithm correctly classifies all the bins \remove{with no misclassification at any round}\remove{correctly classifies all the bins}and there is\remove{some support element has a bin representing itself in $\mathcal{C}(t)$} an unclassified bin after $\sum_{i=1}^{m} \max\left\{\frac{e(b-\log(a_i))}{(e-1)a_i},T'\right\}$ + $\sum_{i=m+1}^{q} \frac{e(b-\log(a_i))}{(e-1)a_i}$ queries is at most ${\delta}$.
{Also, using Lemma $\ref{Correctness_QM2}$, the probability that the algorithm returns an incorrect set of bins is bounded by $\delta$.}
\remove{\color{red} Also using Lemma \ref{lemma2:QM2}, the probability that the algorithm misclassifies some bin at some round would be bounded by $\sum_{i,t}\mathbb{P}[\mathcal{E}_i^t] \leq \delta/2$ as done in the proof of Lemma \ref{Correctness_QM2}} Combining together these two observations, we have that the probability that the Algorithm $\ref{QM2_alg}$ does not terminate after
$\sum_{i=1}^{m} \max\left\{\frac{e(b-\log(a_i))}{(e-1)a_i},T'\right\}$ + $\sum_{i=m+1}^{q} \frac{e(b-\log(a_i))}{(e-1)a_i}$ queries is bounded by 2$\delta$. This completes the proof of the result.
\remove{
Thus, the probability of algorithm having some bin in $\mathcal{C}(t)$ after $\sum_{i=1}^{m} max(\frac{e(b-\log(a_i))}{(e-1)a_i},T')$ + $\sum_{i=m+1}^{q} \frac{e(b-\log(a_i))}{(e-1)a_i}$ queries is bounded by 2$\delta$ implying .
}
\remove{
{\color{red}Read the above to check if its ok after the edits.
Reply- SS: Added the sentence highlighted in red to incorporate the slight change in lemma 11.
Comment - SS: Reverted the proof back to original.
}}
\end{proof}
\subsection{Proof of Theorem \ref{QM2lb}}
Now, we restate and prove Theorem \ref{QM2lb}.
\begin{theorem*
For any $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ under {QM2}, let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the query complexity. Then, we have
$$\mathbb{E}[Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})]\geq \max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{2\times d(p_j||\gamma)}\Bigg\}.$$
\end{theorem*}
\begin{proof}
Consider any $\delta$-true $\gamma-$threshold estimator $\mathcal{A}_3$ under {QM2} \ and let us denote the total number of queries by $\tau$ when the underlying distribution is $\mathcal{P}$. Using the above estimator let us construct a $\delta$ -true $\gamma-$threshold estimator for {QM1}. We create such an estimator $\mathcal{A}'$ by simply querying all the indices involved in {QM2}. Since we know that $\mathcal{A}_3$ is a $\delta$-true $\gamma-$threshold for {QM2}, we can argue that $\mathcal{A}'$ would also be a $\delta$-true $\gamma-$threshold for {QM1} \ and thus the query complexity of $\mathcal{A}_3$ would be $2.\tau$.
Thus if the expected query complexity $\mathbb{E}[\tau]$ of $\mathcal{A}_3$ is less than $\max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{2\times d(p_j||\gamma)}\Bigg\}$ we can construct an estimator $\mathcal{A}'$ for noiseless query model 1 whose expected query complexity is less than $\max_{j\in\{m,m+1\}}\Bigg\{\frac{\log{\frac{1}{2.4\delta}}}{d(p_j||\gamma)}\Bigg\}$ which contradicts the Theorem \ref{QM1lb}.
\end{proof}
\subsection{Proof of Theorem \ref{QM2_alter_lb}}
The following lemma, which we prove in \ref{first_ineq_proof}, will be used to prove this result.
\begin{lemma}{\label{first_ineq}}
For all $1 \geq p_t, \gamma \geq 0$, the following inequality holds true.
$$p_t \log \Bigl(\frac{p_t}{\gamma}\Bigr)+ (2.\gamma) \log\Bigl(\frac{2.\gamma}{p_t + \gamma}\Bigr) \leq 2 d(p_t||\gamma).$$
\end{lemma}
Now we restate and prove Theorem \ref{QM2_alter_lb}.
\begin{theorem*
For a MAB setting described above where the mean rewards of the individual arms satisfy the condition in equation \ref{QM2_pi_cond}, any $\delta$-true $\gamma$-threshold algorithm has the following lower bound in expectation on the total number of pulls $N$:
$$
\mathbb{E}_{\mathcal{P}}[N] \geq \sum_{i=1}^{k} \frac{\log(\frac{1}{2.4\delta})}{2\cdot d(p_i||\gamma)}.
$$
\end{theorem*}
The proof of this result follows along similar lines as that of of \cite[Theorem 7]{shah2019sequentialME}.
\begin{proof}
Consider an estimator $\mathcal{A}$ which can correctly identify the arms with mean reward distribution above $\gamma$ with probability at least $(1 - \delta)$. We consider two such distributions with mean reward profiles as follows:
\begin{align*}
\mathcal{P} = (p_1,p_2,..,p_l,..,p_k),\quad \text{ and } \quad
\mathcal{P}' = ({p_1}',{p_2}',..,{p_l}',..,{p_k}').
\end{align*}
where as before, we assume for $\mathcal{P}$ that $p_1 \geq p_2 \geq \ldots \geq p_m > \gamma > p_{m+1} \geq \ldots p_k$. Also, let $l \leq m$ with ${p_l}'=\gamma - \epsilon$ for some small $\epsilon > 0$ and ${p_i}' = p_i$ $\forall$ $i \neq t$.
Note that the sets of arms with mean reward distribution above $\gamma$ for the distributions $\mathcal{P}$ and $\mathcal{P}'$, $\mathcal{S}^{\gamma}_{\mathcal{P}}$ and $\mathcal{S}^{\gamma}_{\mathcal{P}'}$ respectively, are different.
Recall that, we may decide to pull any subset $S$ of the arms in any round and there can be $2^k$ such subsets. For any subset $S$, with probability $p_j$, the output vector in any round can be $+1$ for some arm $j \in S$ and $-1$ for all other arms; and with probability (1 - $\sum_{i} p_i$) the output vector is $-1$ for all arms . Let ($Y_{S_a,s}$) be the output vector observed while pulling the subset $S_a$ for the $s^{th}$ time. Based on the observations till round $t$, we define the likelihood ratio $L_t$ as follows:
\begin{equation*}
L_t = \sum_{a=1}^{2^k} \sum_{s=1}^{N_{S_a}(t)} \log \Bigl(\frac{f_{S_a}(Y_{S_a,s})}{f'_{S_a}(Y_{S_a,s})}\Bigr).
\end{equation*}
Here $N_{S_a}(t)$ denotes the number of times the subset of arms $S_a$ was pulled till round $t$. With a slight misuse of notation, we let $N_i(t)$ denote the number of times arm $i$ was pulled till round $t$, which sums over all subsets containing $i$.
We say
\begin{equation*}
\mathbb{E}_{\mathcal{P}}\Bigl[ \log \Bigl(\frac{f_{S_a}(Y_{S_a,s})}{f'_{S_a}(Y_{S_a,s})}\Bigr) \Bigr]= D(p_{S_a},{p_{S_a}}').
\end{equation*}
Applying Wald's stopping lemma to $L_{\sigma}$ where $\sigma$ is the stopping time associated with estimator $\mathcal{A}$ we have,
\begin{align}
\mathbb{E}_{\mathcal{P}}[L_{\sigma}] = \sum_{a=1}^{2^k} \mathbb{E}_{\mathcal{P}}[N_{S_a}(\sigma)] D(p_{S_a},{p_{S_a}}')
&\overset{(a)}{\leq} \mathbb{E}_{\mathcal{P}}[N_{l}(\sigma)] \max_{S_a: l \in S_a} D(p_{S_a},{p_{S_a}}'). {\label{expectation_likelihood}}
\end{align}
where $(a)$ follows from the definition of $\mathcal{P}$ and $\mathcal{P}'$ as $D(p_{S_a},{p_{S_a}}')$ is zero for those sets which don't contain arm $l$. Next, for any set $S_a$ such that $l \in S_a$ and $\sum_{i \in S_a \backslash \{l\}} p_i = s$, we have
\begin{align*}
D(p_{S_a},{p'_{S_a}}) = p_l.\log\Bigl(\frac{p_l}{\gamma-\epsilon}\Bigr)
\hspace{0em} +(1-(p_l+s))\log\Bigl(\frac{1-(p_l+s)}{1-(\gamma-\epsilon+s)}\Bigr).
\end{align*}
We can show that the second term is increasing with $s$ and hence takes its maximum value when $s = \sum_i{p_i}-p_l$. Thus,
\begin{align*}
&\max_{S_a:t \in S_a} D(p_{S_a},{p_{S_a}}') = p_l \log \Bigl(\frac{p_l}{\gamma-\epsilon}\Bigr) +
\hspace{0 em} (1-\sum_i{p_i})\log\Bigl(\frac{1-\sum_i{p_i}}{1-(\gamma-\epsilon+\sum_i p_i - p_l)}\Bigr)\\ & \overset{(a)}{\leq} p_l \log \Bigl(\frac{p_l}{\gamma-\epsilon}\Bigr)+ (2.\gamma) \log\Bigl(\frac{2.\gamma}{(p_l+\gamma+\epsilon)}\Bigr) \overset{(b)}{\approx} p_l \log \Bigl(\frac{p_l}{\gamma}\Bigr)+ (2.\gamma) \log\Bigl(\frac{2.\gamma}{(p_l+\gamma)}\Bigr) \overset{(c)}{\leq} 2 d(p_l||\gamma),
\end{align*}
where $(a)$ follows from the fact that $(1-\sum{p_i}) > 2\gamma$ {and $x \log (\frac{x}{x-\alpha})$ is decreasing in $x$ \remove{(Comment: is $\alpha$ positive in $(a)$ Reply:The function is decreasing for both positive and negative $\alpha$. Positive $\alpha$ needed for $l\leq m$, negative $\alpha$ for $l > m$.)}}; $(b)$ follows by making $\epsilon$ arbitrarily small; and (c) follows from Lemma~\ref{first_ineq}.
Then, from \eqref{expectation_likelihood} we have
\begin{equation*}
\mathbb{E}_{\mathcal{P}}[L_{\sigma}] \leq 2\times \mathbb{E}_{\mathcal{P}}[N_l(\sigma)]\times d(p_l||\gamma).
\end{equation*}
On the other hand, it follows from \cite[Lemma 19]{DBLP:journals/jmlr/KaufmannCG16} that since the estimator $\mathcal{A}$ can correctly recover the set of arms $\mathcal{S}^{\gamma}_{\mathcal{P}}$ with probability at least $(1-\delta)$,
$ \mathbb{E}_{\mathcal{P}}[L_{\sigma}] \geq \log \Bigl(\frac{1}{2.4\delta}\Bigr).$
Combining the two inequalities above and recalling the assumption that $l \le m$, we have
$$ \mathbb{E}_{\mathcal{P}}[N_l(\sigma)] \geq \frac{\log(\frac{1}{2.4\delta})}{2d(p_l||\gamma)} \ \forall \ l \leq m.$$
\remove{
Now we prove the same for $t > m$.
We reconstruct the two distributions with reward distributions as follows:
\begin{align*}
\mathcal{P} & = (p_1,p_2,..,p_t,..,p_k)\\
\mathcal{P}' &= ({p_1}',{p_2}',..,{p_t}',..,{p_k}')
\end{align*}
where $t > m$ is an integer. ${p_t}'=\gamma - \epsilon$ for some $\epsilon > 0$ and ${p_i}' = p_i$ $\forall$ $i \neq t$.
Note that the sets of arms with mean reward distribution above $\gamma$ is different for the distributions $\mathcal{P}$ and $\mathcal{P}'$.
By using similar arguments as that of previous case, we can prove \eqref{expectation_likelihood} in this case as well.
We can still show the following inequality (using similar arguments as used in previous case):
\begin{align*}
& \max_{S_a:t \in S_a} (D(p_{S_a},{p_{S_a}}'))\\
& = p_t \log \Bigl(\frac{p_t}{\gamma+\epsilon}\Bigr) +
\hspace{0 em} (1-\sum{p_i})\log\Bigl(\frac{1-\sum{p_i}}{1-(\sum{p_i}+\gamma-\epsilon+s-p_t)}\Bigr)\\
& \overset{(c)}{\leq} p_t \log \Bigl(\frac{p_t}{\gamma + \epsilon}\Bigr)+ (2.\gamma) \log\Bigl(\frac{2.\gamma}{p_t+\gamma-\epsilon}\Bigr)\\
& \overset{(d)}{\leq} 2 \Bigl(p_t \log \Bigl(\frac{p_t}{\gamma}\Bigr) + (1-p_t) \log \Bigl(\frac{1-p_t}{1-\gamma}\Bigr)\Bigr)\\
& = 2 d(p_t||\gamma).
\end{align*}
Note that (c) follows from the fact that $(1-\sum{p_i}) > 2\gamma$ and $x \log (\frac{x}{x+\alpha})$ is decreasing in $x$.
(d) follows from the first inequality \eqref{first_ineq} by making $\epsilon$ arbitrarily small.
We use same arguments as used in previous case to show the following inequalities.
$$\mathbb{E}_{\mathcal{P}}[L_{\sigma}] \geq \log \Bigl(\frac{1}{2.4\delta}\Bigr). $$.
$$ \mathbb{E}_{\mathcal{P}}[L_{\sigma}] \leq 2\times \mathbb{E}_{\mathcal{P}}[N_t(\sigma)]\times d(p_t||\gamma) ; t > m.$$
}
Using similar arguments, we can also show that
$$ \mathbb{E}_{\mathcal{P}}[N_l(\sigma)] \geq \frac{\log(\frac{1}{2.4\delta})}{2d(p_l||\gamma)} \ \forall \ l > m.$$
Hence, we have the following lower bound on the query complexity of any $\delta$-true $\gamma$-threshold estimator under this setting:
$$\mathbb{E}_{\mathcal{P}}[N]\geq \sum_{l=1}^{k} \mathbb{E}_{\mathcal{P}}[N_l(\sigma)] > \sum_{l=1}^{k} \frac{\log(\frac{1}{2.4\delta})}{2d(p_l||\gamma)}.$$
\end{proof}
\subsection{Proof of Theorem \ref{QM2_noisyub}}
In this section, we define ${a_i}'$ as $d^{*}({p_i}',\gamma')/2$ and $b$ = $\frac{1}{2} \cdot \log(\frac{4k}{\delta})$. We start with the following lemma.
\begin{lemma}\label{queries with each box noisy}
Assume that the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} are satisfied. Then, the total number of queries with the bin representing support element $i$ in the second phase of Algorithm \ref{QM2n_alg} is upper bounded by $\frac{e(b-\log({a_i}'))}{(e-1){a_i}'}$ with probability at least $(1-\delta/4k)$.
\end{lemma}
\begin{proof}
Since the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} are satisfied, every bin at the end of the first phase represents a different support element. Let bin $j$ denote support element $i$. Recall from equation \eqref{eq_noisy} that $\tilde{\rho}_j^t$ denotes the fraction of samples seen in the second phase till round $t$ that receive a positive response when compared to the representative element in Bin $j$. We have $\mathbb{E}[\tilde{\rho}_j^t] = {p_i}' = p_i\times (1-p_e) + p_e\times(1-p_i) = (1-2p_e)\times p_i + p_e$ and $\mathcal{L}_j(t)$ and $\mathcal{U}_j{(t)}$ denote the lower and upper confidence bounds of bin $j$ respectively, as defined in equations \eqref{eqn_noisylb} and \eqref{eqn_noisyub} respectively. Accordingly in equation \eqref{eqn:c2tnoise}, the LCB and UCB of bin $j$ are compared with a modified threshold given by $\gamma' = (1-2p_e)\times \gamma + p_e$ to decide how long it will be retained in the subset $\mathcal{C}'(t)$ of bins that new samples are compared against.
The above setting of the second phase is similar to the {QM1} \ model where each new sample with index above $T_0$ would fall in the bin representing support element $i$ with probability ${p_i}'$. Thus similar results apply and in particular, Lemmas~\ref{lemma_exittime} and \ref{lemma_timeupper} can be used to show that the total number of queries with a bin representing support element $i$ is upper bounded by $\frac{e(b-\log({a_i}'))}{(e-1){a_i}'}$ with probability at least $(1-\frac{\delta}{4k})$.
\remove{
This proof follows along the same lines as that of Lemmas \ref{lemma_exittime} and \ref{lemma_timeupper}. This is because,
Recall that $\mathbb{E}[\mathcal{Z}_j^t]= {p_i}' = p_i\times (1-p_e) + p_e\times(1-p_i) = (1-2p_e)\times p_i + p_e$ for $t > T_0$ where indices less than $T_0$ in Bin $j$ denotes support element $i$.
Thus the classifier would be $\gamma' = (1-2p_e)\times \gamma + p_e$ for $t > T_0$ to classify the bins which have the probability of the support elements of indices below $T_0$ above $\gamma$.
Now the setting is similar as that in {QM1} \ model where each new sample with index above $T_0$ would fall in bin with support element $i$ with probability ${p_i}'$. Thus the same theorems and lemmas as that in \ref{lemma_exittime} and \ref{lemma_timeupper} can be used to show that the total number of queries with a bin which denoted support element $i$ at Round $T_0$ is upper bounded by $\frac{e(b-\log({a_i}'))}{(e-1){a_i}'}$ with probability at least $(1-\frac{\delta}{4k})$.
}
\end{proof}
Now we restate and prove Theorem \ref{QM2_noisyub}.
\begin{theorem*
Let $\mathcal{A}$ denote the estimator in Algorithm~\ref{QM2n_alg} with $\beta^t = \log (\frac{4k(t-T_0)^2}{\delta})$ for each $t > T_0$ where $T_0$ is as defined in \eqref{T0_defn}. Let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ be the corresponding query complexity for a given distribution $\mathcal{P}$ under {QM2-N} \ and define $q = \min\{T_0,k\}$, ${p_i}'= (1-2p_e)\times p_i+p_e$ and $\gamma' = (1-2p_e)\gamma+p_e$. Then we have $$Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A}) \leq \sum_{i=1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}({{p}_i}',\gamma')}\Bigr)}{(e-1).d^{*}({p_i}',\gamma')} + \frac{T_0(T_0-1)}{2} $$
with probability at least $(1- 2\delta)$.
\end{theorem*}
\begin{proof}
Let us first bound the total number of queries in the second phase, i.e., post round $T_0$. Since the bins are created only at the end of the first phase, the total number of bins must be upper bounded by $T_0$. Furthermore, if the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} are satisfied, there is at most one bin corresponding to each support element which implies the total number of bins must be upper bounded by $q= \min\{k,T_0\}$.
Assuming that the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} are satisfied, from Lemma \ref{queries with each box noisy}, we have that the total number of queries with the bin representing support element $i$ is upper bounded by $\frac{e(b-\log({a_i}'))}{(e-1){a_i}'}$ with probability at least $(1-\frac{\delta}{4k})$. Taking the union bound over all the $q$ bins we can say with probability at least $(1-\delta)$ that total number of queries in the second phase is bounded by $\sum_{i=1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}({{p}_i}',\gamma')}\Bigr)}{(e-1).d^{*}({p_i}',\gamma')}$.
Now since the probability that the properties in Lemmas~\ref{lemma_extract} and \ref{all_occur} are all satisfied is at least $(1 - \delta)$, the total number of queries in the second phase is upper bounded by $\sum_{i=1}^{q} \frac{2e.\log \Bigl(\sqrt{\frac{4k}{\delta}}\frac{2}{d^{*}({{p}_i}',\gamma')}\Bigr)}{(e-1).d^{*}({p_i}',\gamma')}$ with probability at least $(1-2\delta)$. Finally, noting that there are exactly $\frac{T_0(T_0-1)}{2}$ queries in the first phase of the algorithm, our proof is complete.
\end{proof}
\section{PROBLEM FORMULATION} \label{sec: Problem}
Consider an unknown discrete probability distribution $\mathcal{P}=\{p_1,p_2,...,p_k\}$ over a finite support set $\{1,2,...,k\}$. A random variable $X$ is said to be sampled from a distribution $\mathcal{P}$, if $\mathbb{P}_{\mathcal{P}}(X=i) = p_i$ for all $i\in\{1,2,...,k\}$. Our goal in this paper is to
identify all the indices whose probability of occurrence is above a given threshold $\gamma$. Formally, we want to identify the set of support elements $S_{\mathcal{P}}^{\gamma} = \{i \ | \ p_i \geq \gamma\}$. Towards this, we have access to an independent and identically distributed (i.i.d.) sequence of samples $X_1,X_2,...$, from $\mathcal{P}$ via an oracle.
We consider the following four types of oracle queries.
\begin{itemize}
\item \textit{Noiseless query model 1 ({QM1})}: Queried with an index $i$, the oracle response is given by $\mathcal{O}(i) = X_i$, i.e., the oracle returns the value $X_i$ of the $i^{th}$ sample.
\item \textit{Noisy query model 1 ({QM1-N})}: Queried with an index $i$, the oracle returns the value $X_i$ of the $i^{th}$ sample with probability $1-p_e$ and a random value in $\{1,2,...,k\}$ with probability $p_e$, for some $p_e \in (0, 1/2)$. More formally, the oracle response\footnote{Note that the oracle response remains the same if the same index $i$ is queried repeatedly.} to query $i$ is given by $\mathcal{O}(i)= (1- K_i)\times X_i+ K_i \times U_i$, where $K_i$ is a Bernoulli random variable with mean $p_e$ and $U_i$ is a uniform random variable over the set $\{1,2,\ldots,k\}$. We assume that $\{U_i\}$, $\{K_i\}$ are all independent random variables and are also independent of the sequence of samples $X_1, X_2, \ldots$.
\item \textit{Noiseless query model 2 ({QM2})}: In this model, the oracle makes pairwise comparisons. Given two indices $i$ and $j$, the oracle tells us whether the values $X_i$ and $X_j$ are equal or not. Formally, we denote the oracle response to a query pair $(i,j)$ as:
\begin{equation*}
\mathcal {O}(i,j)=
\begin{cases}
1 & \text{if}\ X_i = X_j, \\
-1 & \text{otherwise.}
\end{cases}
\end{equation*}
Note that the true values of $X_i$ or $X_j$ are not revealed in this model.
\item \textit{Noisy query model 2 ({QM2-N})}: The oracle model here is the same as the one in {QM2}, except that the true noiseless
{QM2} \ oracle response is flipped with probability $p_e$. In particular, given two indices $i$ and $j$, if the values $X_i$ and $X_j$ are equal, the oracle returns $+1$ with probability $1-p_e$ and $-1$ with probability $p_e$. Similarly, if the values of $X_i$ and $X_j$ are not equal, the oracle answers $+1$ with probability $p_e$ and $-1$ with probability $1-p_e$. Formally, we denote the oracle response to a query pair $(i,j)$ as:
$$\mathcal{O} (i, j) = (2\times \mathbbm{1}_{X_i = X_j}-1) \times (1 - 2Z_{i,j}),$$
where $\mathbbm{1}_E$ denotes the indicator random variable corresponding to event $E$ and $Z_{i,j}$ denotes a Bernoulli random variable with mean $p_e$. We assume that $\{Z_{i,j}\}$ are independent random variables and are also independent of the sequence of samples $X_1, X_2, \ldots$.
\end{itemize}
For each of the query models described above, we aim to design efficient sequential algorithms which proceed in rounds as follows:
\begin{itemize}
\item The algorithm chooses a pair of indices (an index) for the {QM2} \textbackslash {QM2-N} \ ({QM1} \textbackslash {QM1-N}) models and queries the oracle with those indices (index).
\item Based on all the responses received from the oracle thus far, the algorithm decides either to terminate or proceed to the next round.
\end{itemize}
When the algorithm decides to terminate, it returns a set of indices denoted by $\widehat{S}$ as an estimate of the set $S_{\mathcal{P}}^{\gamma}$ consisting of all indices $i$ in the support set for which the corresponding probability value $p_i$ is above $\gamma$.
We measure the cost of an estimator in terms of the number of queries made before it stops and its accuracy in terms of the probability with which it successfully estimates the index set $S_{\mathcal{P}}^{\gamma}$. For $0<\delta<1$ and given $0<\gamma<1$, an algorithm is defined to be a $\delta$-true $\gamma$-threshold estimator, if for every underlying distribution $\mathcal{P}$, it identifies the set $S_{\mathcal{P}}^{\gamma}$ with probability at least $1- \delta$, i.e., $\mathbb{P}_\mathcal{P}[\widehat{S}= S_{\mathcal{P}}^{\gamma}] \geq (1- \delta)$, where $\widehat{S}$ is the estimated support set. In this work, we aim to design efficient $\delta$-true $\gamma$-threshold estimators which require as few queries as possible. For a $\delta$-true $\gamma$-threshold estimator $\mathcal{A}$ and a distribution $\mathcal{P}$, let $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ denote the number of queries made by the estimator. Note that $Q_{\delta,\gamma}^{\mathcal{P}}(\mathcal{A})$ is itself a random variable and our results hold either with high probability or in expectation.
Without loss of generality (W.L.O.G.), we assume $p_1 \geq p_2 \geq ...\geq p_m > \gamma> p_{m+1} \geq ...\geq p_k $, i.e., a $\delta$-true $\gamma$-threshold estimator must return the set $\{1,2,...,m\}$ with probability at least $1-\delta$. The rest of the paper is organized as follows. We propose sequential estimation algorithms for the {QM1} \ and {QM1-N} \ models in Sections~\ref{Sec:QM1} and \ref{Sec:QM1N} respectively, where we provide both upper and lower bounds on the query complexity. Section~\ref{Sec:QM2} discusses the query complexity and our proposed algorithm for the {QM2} \ model. Finally, Section~\ref{Sec:QM2N} proposes an estimator for the {QM2-N} \ model and analyses its query complexity. Some numerical results are provided in Section~\ref{Sec:Sim} and we relegate all the proofs to Section~\ref{Sec:Proofs}.
\section{NUMERICAL RESULTS}
\label{Sec:Sim}
In this section, we simulate Algorithms \ref{QM1_alg} and \ref{QM2_alg} for the {QM1} \ and {QM2} \ models respectively, under two different probability distributions.
\begin{figure}[h]
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{thresholdqm13.jpg}
\caption{ Query complexity plot of Algorithm \ref{QM1_alg} under different confidence bounds}
\label{QM1_with_p3}
\end{minipage}%
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{figure6}
\caption{ Query complexity plot for Algorithm \ref{QM2_alg} and a naive algorithm. }
\label{QM2_with_p3}
\end{minipage}%
\end{figure
\begin{enumerate}[label=(\alph*)]
\item In the first setting, we choose the support size $k=30$ and set $p_1= 0.35$, $p_2=0.28$, vary $p_3$ from $0.13$ to $0.19$, and for all $i=3,4,...,k$, set $p_i= \frac{1-p_1-p_2-p_3}{k-3}$. We set the threshold to $\gamma = 0.1$ and the required error probability $\delta = 0.1$. For each datapoint, we simulate Algorithms \ref{QM1_alg} and \ref{QM2_alg} under {QM1} \ and {QM2} \ respectively $15$ times each and plot the average number of queries required against $1/{d^{*}(p_3,\gamma)}$ in Fig. \ref{QM1_with_p3} and Fig \ref{QM2_with_p3} respectively. {In Fig. \ref{QM1_with_p3}, we compare the query complexity for Algorithm \ref{QM1_alg} (using KL-divergence based bounds) with those using other popular confidence bounds namely Hoeffding and Empirical Bernstein (used in \cite{shah2019sequentialME})}. As predicted by our theoretical result, the query complexity of Algorithm \ref{QM1_alg} under {QM1} \ increases (almost) linearly with ${1}/{d^{*}(p_3,\gamma)}$ in Fig. \ref{QM1_with_p3}. In Fig \ref{QM2_with_p3}, we compare the query complexity for Algorithm \ref{QM2_alg} with that of a naive algorithm which in each round, queries the next sample with all the bins created so far. {We observe that Algorithm \ref{QM1_alg} the one with KL-divergence based bounds performs better than its counterparts with other popular confidence bounds.}
We can see that the query complexity of our proposed estimator is much lower since it discards bins as we go along, thus reducing the number of queries.
\begin{figure}[h]
\centering
\begin{minipage}{0.495\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{gammaqm13.jpg}
\caption{Query complexity plot of Algorithm \ref{QM1_alg} under different confidence bounds with $\gamma$}
\label{QM1_with_gamma}
\end{minipage}
\centering
\begin{minipage}{0.495\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{gammaqm2.jpg}
\caption{Query complexity plot of Algorithm \ref{QM2_alg} and a naive algorithm with $\gamma$}
\label{QM2_with_gamma}
\end{minipage}
\end{figure}
\item In the second setting, we choose a probability distribution $\{0.3, 0.25, 0.2, 0.15, 0.1\}$ and vary $\gamma$ from $0.02$ to $0.4$. As before, we simulate Algorithms \ref{QM1_alg} and \ref{QM2_alg} under {QM1} \ and {QM2} \ respectively 15 times each, and plot the average number of queries required against $\gamma$ in Figures~\ref{QM1_with_gamma} \ and \ref{QM2_with_gamma}. {In Fig. \ref{QM1_with_gamma}, we compare the query complexity for algorithm \ref{QM1_alg} (using KL-divergence based bounds) with those other popular confidence bounds namely Hoeffding and Empirical Bernstein (used in \cite{shah2019sequentialME}).} In Fig \ref{QM2_with_gamma}, we compare the query complexity for Algorithm \ref{QM2_alg} with that of a naive algorithm which in each round, queries the next sample with all the bins created so far. The query complexity has multiple peaks, each corresponding to the case where $\gamma$ approaches some $p_i$.
\begin{figure}[h]
\centering
\begin{minipage}{0.495\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{zipfqm13}
\caption{Query complexity plot of Algorithm \ref{QM1_alg} using different confidence bounds for different values of Zipf parameter}
\label{QM1_with_zipf}
\end{minipage}
\centering
\begin{minipage}{0.495\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=0.2\textheight]{zipfqm21}
\caption{Query complexity plot of Algorithm \ref{QM2_alg} and its naive variant for different values of Zipf paramter}
\label{QM2_with_zipf}
\end{minipage}
\end{figure}
\item {In this setting we simulate our algorithms for a fixed value of threshold $\gamma$ = 0.1 against Zipf distributions for various values of Zipf parameter $\beta$ from 0.5 to 4.5. As before we simulate each algorithm 15 times and plot the average query complexity of algorithms corresponding to {QM1} \ and {QM2}\ with $\max\Bigl\{\frac{1}{d^{*}(p_m,\gamma)},\frac{1}{d^{*}(p_{m+1},\gamma)}\Bigr\}$ and $\sum_i \frac{1}{d^{*}(p_i,\gamma)}$ respectively. In Fig. \ref{QM1_with_zipf}, we compare the query complexity for algorithm \ref{QM1_alg} (using KL-divergence based bounds) with those other popular confidence bounds namely Hoeffding and Empirical Bernstein (used in \cite{shah2019sequentialME}). In Fig \ref{QM2_with_zipf}, we compare the query complexity for Algorithm \ref{QM2_alg} with that of a naive algorithm which in each round, queries the next sample with all the bins created so far. Note that a very similar linear variation is observed similar to those in Fig. \ref{QM1_with_p3} and Fig. \ref{QM2_with_p3}.}
\end{enumerate}
{
\textbf{Comparison on a real-world dataset}: We conduct a clustering experiment on a real-world purchase dataset \cite{clusterdatasets}, where we wish to only identify the clusters with size larger than a given threshold. We benchmark our proposed Algorithm \ref{QM2_alg} for pairwise queries and a naive variant of it with no UCB-based bin elimination against the full clustering algorithm of \cite{mazumdar2017clustering}. We use the dataset in \cite{clusterdatasets} to create a set of nodes (denoting the products in this case) with a label attached to each node such that all nodes attached with a common label represents a set of items belonging to the same product category.
From the given dataset, we chose the top $k = 100$ clusters for our experiment so that the total number of items is $n= 9,07,101$. The size of the largest cluster is $53,551$ and we chose our threshold size as $17,688$ with $11$ clusters having size larger than it, with the sizes for the $11$-th and $12$-largest clusters being $19390$ and $16298$ respectively. For 99\% target confidence, Algorithm \ref{QM2_alg} terminated with $\sim$ $2.85 \times 10^6$ pairwise queries. In contrast, even for a target confidence of 80\%, the naive variant of Algorithm \ref{QM2_alg}'s where all bins are queried in every round required $\sim$ $17 \times 10^6$ queries ($6\times$ more). On the other hand, the algorithm that does the full clustering first is expected to take around $nk \sim 90 \times 10^6$ queries ($31\times$ more).
{ We also benchmark the performance of our schemes on the Movielens dataset\\ (https://grouplens.org/datasets/movielens/), with each movie associated with its most popular tag, and let each tag represents a cluster. We consider the top 100 clusters which contain 15,241 movies and choose our threshold as 409 with exactly 3 clusters above it. For the QM1 model and with 99 \% confidence, Algorithm 1 required 2,05,394 queries whereas its variants with Hoeffding and Empirical Bernstein - based confidence intervals needed 2,37,346 and 3,06,976 queries respectively. Under the QM2 model, Algorithm 2 terminated after 8,22,124 queries whereas its naive variant required 61,26,357 queries even under 80\% confidence. Also, the full clustering scheme of Mazumdar and Saha 2017 is expected to take around $nk$=15,24,100 queries}.
}
| 4f00d2fb3cfd85e5241aaf53a15481f795c40710 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Intracranial aneurysms (IAs) are abnormal bulges mainly located at vessel bifurcations and arising from weakened vessel wall.
Despite being highly prevalent ($\sim$3.2\% of population), most IAs do not rupture throughout patient's lifetime and are not associated to any symptoms. In case of rupture, however, the subsequent subarachnoid hemorrhage (SAH) has 50\% fatality rate. Imaging studies have shown that aneurysm size and shape are one of the key factors for prediction rupture risk and deciding on optimal treatment plan for unruptured IAs \cite{ishibashi_toshihiro_unruptured_2009}. For instance, in ELAPSS score~\cite{backes_elapss_2017} the IA size can amount to 55\% of the final score, whereas ELAPSS also takes into consideration earlier SAH, location of aneurysm, age, population and shape of aneurysm.
Clinical treatment of small and medium IAs is yet not well defined. Note that size groups of IAs are small (0-4.9 mm in diameter), medium (5 to 9.9 mm), large (10 to 25 mm) and giant ($>$25 mm)~\cite{ishibashi_toshihiro_unruptured_2009}.
Studies have shown that risk of complications during procedure is larger than risk of spontaneous rupture \cite{brinjikji_risk_2016,ishibashi_toshihiro_unruptured_2009,brown_unruptured_2014,wiebers_unruptured_2003,chien_unruptured_2019}. There are many cases, where risk of spontaneous rupture is less likely than the risk due to treatment and where follow-up imaging is considered to be the best course of treatment. Follow-up imaging may involve several modalities like DSA, CTA and MRA. Therefore, an automatic and reproducible modality-independent process of aneurysm size and shape measurement is essential to provide best achievable treatment for patient and to correctly measure change of aneurysm morphology between consecutive imaging sessions.
Aneurysm size and shape quantification is possible through its segmentation and isolation from parent vessels. Most neurosurgeons still use manual methods to isolate IAs \cite{cardenes2013performance}.
Due to high intra- and inter-rater variability of manual aneurysm segmentation, and the time-consuming and cumbersome 3D image manipulation, there is a need for fast, accurate and consistent computer-assisted segmentation methods.
In this paper we propose surface point classification (SPC) method to segment and isolate the IA. Once the point cloud is extracted from 3D image, for instance by simple interactive thresholding, the procedure to isolate IAs remains the same for all modalities. Our segmentation is based on the point cloud data and is thus modality-independent. This was successfully verified on a total of 100 IA cases, with 57 DSA, 28 CTAs and 5 MRA scans. The proposed method achieved consistent performance across different image modalities and also improved segmentation over the state-of-the-art.
\section{Related work}
An approach mimicking manual IA isolation is by positioning a cutting plane. In past clinical practice the cutting plane was chosen manually with the help of the 3D visualization software~\cite{ma2004three}. Jerman et al.~\cite{jerman2019automated} developed method that automatically positions the cutting plane (ACP). This method works well on most of the aneurysms and is considered as current state-of-the-art. However, if the aneurysm is small or if the aneurysm is blended with surrounding vessels the ACP may fail. Furthermore, if the shape of IA's neck does not lie in a plane, the method may again fail to isolate aneurysm correctly as shown in Figure~\ref{model_output}.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=4.5in]{fail_and_gold.pdf}
\end{center}
\caption{\small Two examples of aneurysm segmentation. Current state-of-the-art-method (ACP) fails at isolate aneurysm with plane (\textit{grey surface on the left}), while our SPC method (\textit{yellow-colored surface}) is very close to reference manual segmentation (\textit{red curve}). }
\label{model_output}
\end{figure}
Non-planar curve for separation of aneurysm from the surrounding vessels seem a better option. Ruben et al.~\cite{cardenes2011automatic} used automatic approach based on the computation of a minimal path along a scalar field obtained on the vessel surface. In order to assure correct topology of the aneurysms bulb, the neck computation was constrained with a region defined by surface Voronoi diagram. The method was evaluated on 26 real cases against manual aneurysm isolation using cutting plane. This method works well for standard cases, but can fail if the configuration of the arteries is complex. Marina et al.~\cite{piccinelli2012automatic} proposed a method also based on Voronoi diagram with the combination of so-called tube function. The method was evaluated on 30 cases. In 20 cases, the manual and computed separation curves were considered as being very similar and both acceptable. In five cases manual curve was better then the computed, but still acceptable. In five cases the method did not produce similar curves compared to manual ones. Hence, robustness remains an issue with non-planar separation curves.
Sylivia et al.~\cite{saalfeld2018semiautomatic} proposed a semi-automatic separation curve reconstruction algorithm for automatic extraction of morphological parameters. The first drawback of this approach is that it requires a pronounced aneurysm neck to work properly. The second drawback is that the user has to click on the aneurysm to initialize the method's parameters. Lawonn et al.~\cite{lawonn2019geometric} proposed a geometric optimization approach for the detection and segmentation of multiple aneurysms in two stages. First, a set of aneurysm candidate regions was identified by segmenting regions of the vessels. Second stage was a classifier, that identified the actual aneurysms among the candidates. While the method was capable of detecting aneurysm location automatically, the user still had to use the brush tool to get correct aneurysm neck curve. To address this problem they proposed a smoothing algorithm based on ostium curve extraction to improve upon previous results. While smoothing algorithm worked fine, additional manual user input was still necessary for some aneurysm cases.
All of the aforementioned methods mostly provide an acceptable result for large saccular IAs that have a well defined aneurysm neck (the cross section of the inlet is significantly smaller than cross-section of the aneurysm). For other aneurysm size and shapes most of these methods fails to deliver correct result. To our knowledge, there is still no fully automatic method that would work well for IAs of various sizes and shapes, regardless of the configuration of surrounding vessels.
\section{Materials and methods}
Current clinical detection and measurements of IAs is performed on angiographic images like DSA, CTA and MRA (section~\ref{cases}). Procedures used to extract 3D surface mesh from angiographic images are explained in section~\ref{preprocesing}, data augmentation and training process in~\ref{training} and the inference on clinical images in~\ref{inference}.
\begin{figure}[]
\begin{center}
\includegraphics[width=4.7in]{flow_neck.pdf}
\end{center}
\caption{\small Flowchart of the proposed aneurysm segmentation method.}
\label{PointNet}
\end{figure}
\subsection{Case Information}\label{cases}
This study was performed with the approval of the institutional review board. A total of 100 patient angiographic 3D scans, including 57 DSA images, 28 CTA images and 5 MRA images, of the cerebrovascular region were obtained for the purpose of this study.
All images were acquired at University Medical Centre Ljubljana
using standard imaging protocols used in clinical routine.
Per patient incidence of IAs in our dataset was from zero to three. We found a total of 100 IAs, from which 27 are considered small ($<5$ mm), 49 medium sized ($5<$ size $<10$ mm) and 24 are large ($>25$ mm). The median diameter of the aneurysms was 6.99 mm.
\subsection{Mesh extraction from 3D angiographic image}\label{preprocesing}
First step in extracting the IA mesh from 3D angiographic image is IA detection. The search for IA is mainly still performed by skilled surgeon without any special computer-assisted tools. Many algorithms for automatic aneurysm detection were previously proposed \cite{jerman2017aneurysm,zhou2019intracranial,jin2019fully,duan2019automatic}, but all have the difficulty of presenting with to many false positives or are not able to detect small aneurysms. For our purposes detection was carried out manually using visual image inspection.
Following detection, a region of interest (ROI) containing the aneurysm and surrounding vessels was determined. Segmentation of vascular structures was performed by using interactive thresholding, followed by the application of marching cubes and smooth non-shrinking algorithms \cite{larrabide2011three,cebral2001medical}. In this way, the corresponding 3D surface mesh of aneurysm and surrounding vessels was reconstructed. The following procedures ware performed using image analysis software (RealGUIDE Software version 5.0; 3DIEMME Srl). Example mesh containing an aneurysm and surrounding vessels is shown in Figure~\ref{model_output}.
\subsection{Training}\label{training}
On every input mesh the aneurysm was manually isolated from surrounding vessels by skilled neurosurgeon by drawing a closed surface curve (see Figure \ref{model_output} for examples of manual reference curve for IA isolation). Aneurysm points were labeled according to the neurosurgeon's segmentation. Surgeon was not restricted to the specific shape of aneurysm neck. The obtained manual IA segmentation isolation were used for training and validation of multi-layer neural network (MNN) model~\cite{qi2017pointnet}.
To train MNN model, with network architecture as presented on Figure~\ref{PointNet}, that is robust to point ordering, rotations and scaling of input data we replicated each mesh \textit{n} times (\textit{n}=number of replications). Rotations were randomly applied with angles from 0 to 360 degrees in all three axes. Scaling factor was randomly set between values 0.5 and 1. From each augmented mesh we randomly picked \textit{m} points (\textit{m}=number of points) and saved them as point cloud. Augmented point clouds of one aneurysm appeared only in train, test or validation set.
\subsection{Inference - using trained model for isolation}\label{inference}
To further suppress MNN model's sensitivity to rotation, we applied a so-called multiple partial voting algorithm. Similar to the augmentation process before training, we used the same augmentation procedures for the target mesh until each point of the target mesh was used at least 20 times, each time with different rotation. For every augmentated instance we used \textit{m} random points from target mesh. Each augmented point cloud contributed one vote to segmentation for each point. The sum of all votes was divided by number of votes for each point. The resulting heatmaps represented the output segmentation, which was compared to the other methods as shown on Figure~\ref{results}.
To determine which points on heatmap form the aneurysm and which the surrounding vessels we used simple thresholding at 0.5. We also verified our decision on such threshold by plotting the ROC (receiver operating characteristic) curve.
\section{Experiment}
The training of MNN segmentation model was executed only once. As train dataset we used 70\% of all available data, i.e. 70 aneurysm meshes. Each mesh was reproduced with different rotation about 100 times (\textit{n}=100) and 3000 points (\textit{m}=3000) were randomly chosen after each rotation. For test dataset we used 25\% of available data, i.e. 25 aneurysm meshes, which were also rotated \textit{n} times and \textit{m} points were chosen. We used 5\% of data or 5 aneurysm meshes for validation purposes. Those aneurysm meshes were used to create a valid heatmap. The number of epochs was 100 and the learning rate was 0.001 for fist 20 epochs, 0.0005 for epochs between 21-40, 0.00025 for epochs between 40-60, 0.000125 for epochs between 61-80 and 0.000063 for the rest of the epochs. Batch size was set to 16.
To fairly compare our results with the state-of-the-art method proposed by Jerman et al.~\cite{jerman2019automated} we also calculated SPC with plane (SPCP) as follows: using least-square fit a cutting plane was positioned onto border points between aneurysm and vessel.
The model was trained on the Linux based computer with 32 GB RAM, Intel Core 8700K 6 Core 4,7 GHz processor and 11GB NVIDIA Graphic Card.
\section{Results}
After training our model on 70 meshes, we tested the performance on all datasets. Point-wise predictions were rounded to the nearest integer (0 or 1) for all \textit{n} rotation of each mesh, the aggregated heatmap values of each target mesh were normalized to interval [0, 1]. For each mesh we computed true positive rate (TPR; sensitivity). Table~\ref{table-res} shows sensitivity evaluation of all data subsets. Median sensitivity on learning dataset was 0.985, 0.983 on test dataset and 0.994 on validation dataset. Overall median sensitivity was 0.985. If we fit plane to the SPC model result we achieve median sensitivity 0.938 on training dataset, 0.924 on test dataset and 0.947 on validation dataset. Overall median sensitivity was 0.938. Current state-of-the-art method achieved accuracy of 0.830 on training, 0.775 on testing and 0.846 on validation dataset, with median sensitivity of 0.830. State-of-the-art method failed to isolate aneurysm from the vessels in 8\% of all cases, while our proposed method succeeded in accurately isolating aneurysm from the vessel across all cases. The distribution of sensitivity for all three tested methods is presented in box plots on Figure~\ref{results}.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=4.5in]{box_plot.pdf}
\end{center}
\caption{\small Point-jitter and box plots of sensitivity values for all meshes for three tested methods.}
\label{box_plots}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=4.7in]{results.pdf}
\end{center}
\caption{\small Images \textbf{A-D} depict four examples of our isolation compared to gold standard and previous work. Red line represents gold standard aneurysm neck, grey plane represents the ouput of ACP algorithm, orange plane was fitted from boarder points of SPC. Yellow surface is our result of aneurysm isolation from vessels. Dark purple surface is predicted vessel surface.}
\label{results}
\end{figure}
Average time to segment and isolate one aneurysm from the vessels with the current state-of-the-art method was approximately 11 minutes. On the other hand, the isolation of using novel method executed in under 23 seconds per input image.
\begin{table}[]
\caption{Aneurysm segmentation evaluation across all 100 datasets and across the training, testing and validation cases.\label{table-res}}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c|c}
\cline{2-13}
& \multicolumn{3}{c|}{\textbf{ALL}} & \multicolumn{3}{c|}{\textbf{TRAIN}} & \multicolumn{3}{c|}{\textbf{TEST}} & \multicolumn{3}{c|}{\textbf{VAL}} \\ \hline
\multicolumn{1}{|l|}{\textbf{\textbf{Method}}} & \textbf{SPC} & \textbf{SPCP} & \textbf{ACP} & \textbf{SPC} & \textbf{SPCP} & \textbf{ACP} & \textbf{SCP} & \textbf{SPCP} & \textbf{ACP} & \textbf{SCP} & \textbf{SPCP} & \multicolumn{1}{l|}{\textbf{ACP}} \\ \hline \hline
\multicolumn{1}{|l|}{min TPR} & \textbf{0.759} & 0.507 & 0 & \textbf{0.759} & 0.507 & 0 & \textbf{0.918} & 0.730 & 0.074 & \textbf{0.982 }& 0.892 & \multicolumn{1}{l|}{0.709} \\ \hline
\multicolumn{1}{|l|}{max TPR} & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0.985 & \textbf{0.999} & 0.972 & \multicolumn{1}{l|}{0.972} \\ \hline
\multicolumn{1}{|l|}{median TPR} & \textbf{0.985 } & 0.938 & 0.830 & \textbf{0.985} & 0.938 & 0.830 & \textbf{0.983} & 0.924 & 0.775 & \textbf{0.994} & 0.947 & \multicolumn{1}{l|}{0.846} \\ \hline
\end{tabular}
\end{table}
\section{Discussion and Conclusion}
We successfully developed and validated a novel MNN based method for aneurysm segmentation and isolation and compared its performances to current state-of-the-art ACP method~\cite{jerman2019automated}. To our knowledge our method is the first that works on all aneurysm cases regardless of their own shape or shape of the surrounding vessels. At the same time, the ACP method fails to provide correct results in approximately 10\% of the cases and provide partially satisfactory segmentations on 30\% of our cases. The computational times, compared to approximately 11 minutes for the ACP, were less then 23 seconds for the proposed method.
The proposed method achieved 15\% better median sensitivity than the ACP. For comparison we use the least-square minimization to fit a plane to the original output of our method and still achieved 10\% better median sensitivity than ACP. We succeeded to isolate aneurysm from vessel for all cases with the sensitivity of at least 0.759 (median sensitivity for all cases was 0.985, while for validation dataset the sensitivity was 0.994). The 5 validation cases were not used in training and were randomly chosen from the set of 100 aneurysms.
Figure~\ref{results} shows 2 cases where all methods work and 2 cases where current state-of-the-art ACP method fails. Red curve is manual reference neck curve determined by skilled neurosurgeon, yellow/purple areas on meshes represents our isolation of aneurysm from vessel, black surface is a manually positioned cutting plane, while the orange surface represents the plane fit on border points between aneurysm and vessel of our method's output. In sections \textbf{A} and \textbf{B} of Figure~\ref{results} all approaches worked well, while in sections \textbf{C} the ACP method failed, while the plane acquired from output of our method still provided good results, in section \textbf{D} only our method provided meaningful results.
We have successfully surpassed the sensitivity of MNN to rotation and scaling of input data with random rotations and scaling of training data. The additional rotation of target mesh contributed to better and more robust prediction of our MNN model. The output in the form of heatmap allows us to see the probability of each point being part of aneurysm or vessel. Probability map also emphasizes the area between aneurysm or vessel (the values smaller than one, but bigger than zero).
We have also shown that point cloud based approach is not sensitive to different input image modalities and artifacts associated to them. To create our data we used all common angiography modalities, namely the CTA, DSA and MRA.
Accurate and fast prediction of aneurysm neck is essential for computing aneurysm features that are based on neck curve determination, such as volume, surface area, sphericity index, surface are to volume ratio, etc. Those measurements are useful in a variety of clinical and research applications related to intracranial aneurysms. An important treatment procedure called coiling involves coil embolization for treating unruptured aneurysms. The quantity of inserted coil depends on the measured aneurysm volume. Reliable volume calculations can prevent complications during and after the procedure, for instance, caused by inaccurate amount of inserted coil that can cause aneurysm lumen reopening and eventually result in rupture~\cite{slob2005relation,kai2005evaluation}.
Surface area and volume have both been demonstrated to be important indicators of aneurysm rupture risk \cite{austin1989controlled,valencia2008blood}. As shown on Figure \ref{results} the segmentation and isolation with cutting plane does not seem robust enough to isolate aneurysm from the vessels regardless of the position and the shape of aneurysm. Though prior work on isolation with non-planar separation curves showed some promising results \cite{saalfeld2018semiautomatic,lawonn2019geometric}, none of them achieved clinically acceptable results and robust enough performance. Small aneurysms with curved veins proved to be extremely challenging. Our method showed high accuracy and sensitivity, even with a relatively small dataset of 100 aneurysms. The method was able to predict, which points belong to the aneurysm automatically with sensitivity of 0.985. We feel that the robustness of our model will be reconfirmed even on larger volumes of training and test data.
The proposed deep learning approach to aneurysm isolation can identify and label unordered aneurysm points in 3D vascular mesh automatically and with great sensitivity. This approach enables robust, repeatable and fast isolation of aneurysms from the surrounding vessels, thus the method seems readily usable in research and clinical applications.
\bibliographystyle{splncs04}
| 989ef179120c6baac3fcedf01ec1118858b79aa8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Feedback from active galactic nuclei (AGN) is widely believed to be a crucial element of the mass assembly of galaxies: with feedback acting on gas and stars forming from gas, feedback regulates the stellar mass growth of galaxies \citep{bower06,croton06,dimatteo08,sijacki07,dubois16, cielo18}. By leading to the acceleration or heating of the gas in the interstellar medium (ISM), feedback is
associated with winds and outflows that can even eject gas outside of galaxies.
It can also delay the deposition of new gas from the intracluster medium (ICM) or intergalactic medium (IGM) onto galaxies.
Both effects help the quenching of star formation through the lack of fuel for accretion, simultaneously terminating the black hole activity \citep{alatalo15,lanz16}.
It is widely believed that AGN feedback operates in a bimodal manner. The quasar mode, also called radiative or wind mode, occurs through wide-angle winds driven by radiative pressure in luminous AGN with accretion rates close to Eddington. This mode can be met at any redshift $z$, and it is very common in high-$z$, young QSOs \citep{king03,harrison17,bieri17}. The radio (or kinetic) mode involves the launch of collimated relativistic jet of particles (i.e.\,radio jet), and occurs mainly in low-luminosity AGN (<0.01L$_{\rm Edd}$). Radio jets launched by supermassive black holes (SMBH) are seen to have an impact even on large scales, e.g., of tens of kpc, delaying cooling flows from the ICM/IGM and the formation of extremely massive galaxies \citep{fabian12, namara12}. This is characteristically seen at the centers of fairly low $z$ clusters and groups. Of course, they can also have a strong impact on (sub-)kpc scales, altering the properties of the ISM and driving molecular outflows \citep{morganti15,dasyra16,fotopoulou19, oosterloo19,aalto20,juan20}, and even in low-luminosity AGN \citep{combes13,santi14,ane19,ruffa22}.
The observational findings are supported by 3D hydrodynamical simulations of relativistic jets coupling in inhomogeneous and clumpy ISM \citep{wagner11,wagner12, muk18}, showing that propagating radio jets can gradually disperse gas clouds and create a cocoon of shocked material associated with multi-phase outflows. As this jet-driven bubble expands, it also leads to the dispersal of molecular clouds. The final feedback impact depends on the jet kinetic energy and the distribution of the clumpy medium defining the jet's path of minimum resistance.
Yet, a systematic and statistically meaningful observational measurement of the number of radio galaxies (RGs) with significant gas reservoirs, upon which the radio-mode feedback could be operating, is missing. So far, it has been shown that most of the local RGs have low gas content compared to spirals or infrared selected galaxies (e.g., Ruffa et al. 2022). High-$z$ CO surveys of RGs started in the 1990s \citep{evans96} and eventually led to the discovery of large reservoirs of molecular gas (with $M_{\rm H_2}\sim10^{10}-10^{11}M_\odot$) in some powerful RGs at $z\gtrsim$2 or gas-depleted, dead hosts in other \citep{scoville97,papa00,breuck03, breuck05,greve04, nesvadba08,emonts11,emonts14, castignani19}. The increased sensibility achieved with the advent of interferometric facilities such as the Atacama Large Millimeter Array (ALMA) and the Northern Extended Millimeter Array (NOEMA) nowadays offer a great opportunity to study the evolution of the molecular gas in low and high-$z$ RGs.
In this paper, we present a new analysis of observations of CO molecular transitions from rotational numbers J=1-0 to J=4-3 performed with ALMA for RGs up to $z<$2.5. To build a statistically significant sample, we analyzed the CO emission in a sub-sample of RGs from the ALMA Calibrator Source Catalog, which were selected to be representative of the NRAO/VLA Sky Survey (NVSS) catalogue \citep{condon98}, with respect to its $z$ and 1.4 GHz flux distribution, down to a limit of 0.4 Jy. Complementing these with literature data, we carried out a large and complete archival CO survey of molecular gas in RGs. A main aim of our work was to analyze the evolution of the gas content of RGs with $z$, which is highly needed for an accurate benchmarking of the cosmological simulations. Such analysis also lead us to construct for the first time the CO luminosity function of RGs at low and high $z$. Another goal was to determine the occurrence of molecular outflows in the RGs in our sample.
The paper is structured as follows: in Section~\ref{sec:sample} we describe the statistics used to reproduce a representative sample of NVSS radio galaxies. In Section~\ref{sec:obs} we describe the data reduction of the ALMA observations. Main results from the analysis of the literature and ALMA observations are presented in Section~\ref{sec:results}. The presentation and discussion of the mass and luminosity functions are described Section~\ref{sec:lco}, followed by conclusions. In this paper we adopt a flat $\Lambda$CDM cosmological
model with H$_{\rm 0}$=70\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$} Mpc$^{-1}$, $\Omega_{\rm M}$=0.3, and $\Omega_{\rm \Lambda}$=0.7.
\color{black}
\section{Radio galaxies with ALMA observations}
\label{sec:sample}
\color{black}
\subsection{ ALMA data mining }
\label{subsec:sample_arc}
The selection of galaxies for our survey started from the ALMA Radio Source Catalogue (ARC), which comprises a list of mm and sub-mm bright objects that can serve as (flux, phase, bandpass, etc) calibrators. This catalogue includes a compilation of catalogues from several facilities other than ALMA, including the Very Large Array Calibrator Manual, Submillimeter Array and Atacama Compact Array catalogues, the Parkes survey, and the Combined Radio All-Sky Targeted Eight-GHz Survey. As such, the ensemble of the ARC covers the full sky, but in a non-homogeneous manner.
The ARC catalogue was downloaded from the European Southern Observatory (ESO) archival interface \footnote{\url{https://almascience.eso.org/sc/}}. Each catalogue entry corresponds to observations of a single object in a single band. More specifically, each entry shows the latest flux measurement of each object in a given band, with the exception of band 3. For band 3, two entries are provided, presenting the latest flux for each sideband. As of 01/08/2020, the downloaded catalogue contained 8679 entries. The number of unique objects (deduced by removing the information on the bands) was 3362. For some of the objects, the spectral windows of the calibrations covered the frequencies of molecular line transitions. Therefore, additional science on, e.g., the detection of CO lines, was feasible with the calibration data. Some of the sources in this catalogue were also observed as science targets. The combination of calibration and science related data led to a wealth of information that enabled our archival survey to be designed.
For each ARC source we used the "astroquery" tool \citep{astroquery} to find the available ALMA observation searching within a radius of 3\arcsec\ from the registered sky coordinates. Additionally, we iterated over all names given for each ARC catalogue entry in order to cross-check detection rates. In total, we found recorded 25827 observations from 1562 unique sources.
To determine whether a CO transition was in the spectral window of an observation, we needed the redshift of the observed source. For this purpose, we first queried the NASA Extragalactic Database (NED), Simbad, and Vizier using "astroquery". We performed an extra quality assessment by only keeping spectroscopically determined $z$ values, and in fact, by manually
examining whether each spectroscopic $z$ was measured using two or more spectral lines. Once having the $z$ information, we only kept observations of sources with CO(1-0), CO(2-1), CO(3-2), or CO(4-3) coverage in a spectral window. A total of 675 sources had at least one CO line with any integration time. We kept only sources with a minimum integration time of $5 \,\hbox{\hbox{min}}$ to ensure that the data analysis (e.g., the derivation of error bars) is meaningful.
As a next step, we had to ensure that only galaxies with radio emission associated with accretion onto black holes were kept. As the goal of our survey is to investigate the role of radio-mode feedback on galaxy evolution, all objects with radio emission associated with star formation activity had to be eradicated. For this reason, we imposed the 1.4GHz-to-24$\mu$m flux criterion of \citet{bonzini13}, identifying radio emission in excess of what supernovae can produce in star-forming galaxies via the quantity $q_{24}=\log(S_{\rm 24\mu m}/S_{1.4GHz})$, which has to be $<$0.5\footnote{The ``classical" definition of radio loudness via the $R$ indicator, the ratio between the rest-frame radio-to-optical flux density with typical values of R$\sim$10 characterizing RGs \citep{kellermann89}, is often insufficient to identify radio-quiet (RQ) objects, because both star-forming and RQ galaxies can have low R values.}.
The 1.4\,GHz fluxes of our sources were obtained directly from radio surveys, when available. Otherwise, we interpolated their values from
the nearest available radio data, fitting the radio data with a power law in $\nu F_\nu$. In total, this requirement left us with 204 sources. From here onward, we will refer to all radio galaxies with ALMA CO observations as the CO-ARC catalogue.
To serve as calibrators, the CO-ARC sources are point sources at millimeter and radio wavelengths (such as quasars or blazars). To also include RGs with spatially-resolved, extended radio jets, we performed an extensive bibliographic search, which provided us with a pool of either single-object or dedicated samples with CO observations \citep[e.g.][]{lim00,evans05,ocana10,ruffa19b,russell19, dabhade20}. The $q_{24}$ criterion described above was then applied to the selected literature sources, leaving us with a total of 152 galaxies which were then considered suitable to complement our CO-ARC sample. Hereafter, we will refer to the joint CO-ARC and bibliographic super-sample of 356 galaxies as the {\it extended} CO-ARC.
\begin{figure}[h!]
\resizebox{\hsize}{!}{\includegraphics{figures/SampleSize_algorithm.png}}\caption{D-statistic, i.e., maximum distance between two cumulative distributions used for the sample selection versus the sample size . Solid lines show a theoretical comparison of the NVSS and a shuffled NVSS samples with 10 and 30\% flux variations. Small black dots show how the sample D-statistic changes as a function of the number of bootstrapped sources (see text). The large circles show the final numbers for the CO-ARC (black) and for individual $z$ ranges (color). The same analysis is shown for the CO-ARC only using the ALMA calibrators (without including the literature sources) is indicated as squares and equivalent color code for the redshift bin. We reach an optimal sample size of 120 (CO-ARC in black circle) for a D-statistics smaller than 10\% and converging with the values computed to the theoretical shuffled NVSS sample down to 10\% flux variation.}
\label{fig:convergence}
\end{figure}
\subsection{Selection of a radio-representative sample}
\label{subsec:sample_parent}
Our next task was to find a sub-sample of the {\it extended} CO-ARC sample that is representative of a radio galaxy survey. As parent radio catalogue, we used the NVSS 1.4\,GHz imaging survey, performed by the Very Large Array (VLA) \citep{condon98}. The NVSS covers a large fraction of the sky for declinations above $\delta>-$40$^\circ$, and it is fairly deep in sensitivity, recovering 1.4 GHz flux densities as low as $\sim$2\,mJy. This limit ensures completeness at the levels we are interested in examining, of a few hundred mJy, given the 1.4 GHz fluxes of the CO-ARC sources. In fact,
we also examined a few other catalogues as potential parent surveys of ours:
the Fourth Cambridge Survey (4C) at 178 MHz \citep{pilkington65}, the Sixth Cambridge Survey (6C) at 151MHz \citep{hales93}, the Parkes Radio Sources Catalogue (PKS) at 2.7GHz \citep{wright90} and at 4.8 GHz \citep{griffith94}, the MIT-Green Bank 5GHz Survey \citep{bennett86}, and the Australia Telescope 20GHz Survey Catalog (AT20G; \citealt{murphy10}). We found that NVSS is the optimal parent survey because of its similarity in the flux distribution and of its well characterized completeness\footnote{Despite using the NVSS as radio parent sample, a derivation of a luminosity function with another sample representative of the 6th Cambridge Survey of radio sources (6C survey) is also discussed in Section~\ref{sec:lco}, for comparison purposes.}.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/Sample_NVSS_dist_linear_bins_stone.png}}
\caption{1.4 GHz flux distribution of NVSS and CO-ARC. The orange histogram shows the distribution for all NVSS parent sample, while the blue histogram is for the NVSS with redshift information. The thin black lines indicate the re-shuffled samples of NVSS with flux variations up to 10\%. The CO-ARC in green is well represented by both parent samples within the 10\% errors.}
\label{fig:flux_distributions}
\end{figure}
We then kept the largest possible sub-sample of the extended CO-ARC that is fully representative of the NVSS 1.4 GHz flux distribution. We used bootstrapping to iteratively and randomly remove more and more sources from our initial sample until we obtain a final sample consistent with the NVSS. Often, the convergence criterion of any two samples is checked with a Kolmogorov-Smirnov (KS) test: KS provides a significance level (p-value) based on the maximum distance between the two cumulative distributions (D-statistic), disproving the hypothesis that the two distributions are inconsistent. However, the p-value is rather arbitrary. Instead, we used our own precisely-defined D-statistic criterion, so that bootstrapping can reliably provide the maximum possible CO-ARC sub-sample that is compatible with the NVSS. For this definition, we simulated new, mock sets of NVSS data, starting from the initial fluxes in the catalogue and adding to them Gaussian random noise with dispersion $\sigma_f$. Then, we measured how the D-statistic between the original NVSS fluxes and any N-sized subsample of simulated NVSS fluxes changes as a function of N. The curves are shown in Fig.~\ref{fig:convergence}, and they are calculated using the average D-statistic of 1000 implementations for each N. They indicate how close to the parent catalogue a child catalogue would be if its fluxes varied by 1$\sigma_f$ and 3$\sigma_f$, respectively, using a typical flux uncertainty of $\sigma_f$=10\%. As the sample size is getting smaller, the corresponding D-statistic grows due to dominance of small number statistics: at small enough sample sizes any two samples can resemble each other and thus the statistics deteriorate.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/zbins.png}}
\caption{Redshift completeness of CO-ARC. {\it Top}: Number of sources as a function of $z$ in our final CO-ARC sample and in the NVSS parent catalogue (divided by 25 for comparison purposes). The NVSS distribution was scaled to a factor of 25 for displaying purposes. {\it Bottom:} Completeness correction so that our final sample reproduces the number of sources in the flux-limited NVSS with $z$. }
\label{fig:zbin}
\end{figure}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/sky_positions.png}}
\caption{Distribution of the sample in the sky in the aitoff projection. The color code corresponds to the redshift and the sizes are according to the FWHP of the primary beam of the observations. The dotted curve and the star indicate the Galactic plane and center.}
\label{fig:sky}
\end{figure}
Having defined our theoretical convergence criterion, we proceeded with replacing the theorized child sample with the actual extended CO-ARC sample. Starting from the initial number of sources in the extended CO-ARC, we randomly drew consecutively smaller and smaller sub-samples and compared their D-statistic with the value established for the theorized NVSS sub-sample.
Our bootstrapping again indicates how the D-statistic changes as a function of N: it decreases with decreasing N as (mainly bright) sources are excluded from the sample, until it reaches a minimum and starts rising again due to small number statistics. The two samples become fully compatible (down to $\sim$ 1$\sigma_f$ or 10\% flux variation) for 120 sources. These numbers are presented for a flux cut of 0.4\,Jy. Looping over different flux thresholds and repeating the above procedure indicated that this choice led to the maximum possible sample size. The 1.4\,GHz flux distribution of our final sample and the flux-limited NVSS parent sample are shown in Figure~\ref{fig:flux_distributions}.
We also show the same analysis for the CO-ARC only using the ALMA calibrators (without including the literature sources, squares in Figure~\ref{fig:convergence}) and the CO-ARC with ALMA calibrators \textit{and} literature sources (circles) for the total sample and for the respective redshift bins, indicated by different color codes in Figure~\ref{fig:convergence}. The inclusion of literature sources helps to improve the statistics and to converge the CO-ARC sample to the theoretical D-statistic distribution of the shuffled NVSS.
It is noteworthy that our final sample is representative of the NVSS flux distribution also in each of the 3 individually examined $z$ ranges. We have chosen to use the ranges 0.005$<$z$<$0.3, 0.3$<$z$<$1,1$<$z$<$2.5, to study the evolution of the galaxies in comparable time steps of about 3.5-4.5\,Gyr. The CO-ARC sources of the $z$ slices are from 1$\sigma_f$ to 3$\sigma_f$ away from the NVSS parent catalogue at the same $z$. For the sake of completion, we note that this also holds separately for the sources from the ARC and the sources from the literature. The only exception comes from the few CO-ARC sources at intermediate $z$, where the construction of the luminosity function is unfeasible due to small number statistics.
Having built a final sample that is representative of the flux-limited NVSS, we then calculated its completeness with respect to the parent survey number of sources in the sky as a function of $z$. For this purpose, we split both samples in even smaller $z$ steps (of 0.77\,Gyr) and present the comparison in Figure~\ref{fig:zbin}. The top panel shows the $z$ distribution of the two samples (with the flux-limited NVSS being divided by a factor of 25 for comparison purposes). It shows that the final CO-ARC is over-represented at $z$<0.2 and under-represented at z$\sim$0.8, but otherwise follows the parent catalogue's $z$ distribution. The completeness correction, i.e., the $z$-dependent number with which we need to multiply the CO-ARC galaxies to make them as frequent (in $z$ and in the sky) as the flux-limited NVSS, is shown in the bottom panel of Figure~\ref{fig:zbin} and is taken into account during the construction of the CO luminosity function in Section~\ref{sec:lco}. There is one more correction to be taken into account in the luminosity function construction: the fact that not all NVSS sources have known $z$. On average, 1 on every 4 NVSS sources have $z$ information. This correction is, in principle, also $z$-dependent, but its behaviour with $z$ is unknown. We, thus, also multiply the number density of sources times 4, except for the local Universe (at $z$ $<$0.3), in which we assume the redshift acquisition to be complete.
\begin{table*}
\footnotesize
\centering
\begin{tabular}{lccclllll}
\hline
Galaxy & Line & Beam size & $\sigma_{\rm rms}$ & S$_{\rm CO}\Delta v$ & FWHM & $v_{\rm res}$ & log(MH$_2$) & log(L'$\rm_{CO(1-0)}$) \\
& & ("$\times$") & (mJy/beam) & (Jy\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$}) & (\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$}) & (\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$}) & (M$\rm_\odot$) & (K\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$} pc$^2$) \\
\hline
J1000-3139 & CO(2-1) & 0.97$\times$0.71 & 0.40 & 18.38$\pm$ 0.59 & 368.53 & 20.00 & 7.85$\pm$6.87 & 7.21$\pm$6.22 \\
J1109-3732 & CO(2-1) & 0.6$\times$0.5 & 0.46 & 6.93$\pm$ 0.35 & 452.27 & 20.00 & 7.56$\pm$6.61 & 6.92$\pm$5.97 \\
J1336-3357 & CO(2-1) & 0.63$\times$0.57 & 0.34 & 1.89$\pm$ 0.29 & 781.79 & 20.00 & 7.17$\pm$6.33 & 6.52$\pm$5.68 \\
J1723-6500 & CO(2-1) & 0.27$\times$0.19 & 0.24 & 93.80$\pm$ 2.15 & 533.90 & 20.00 & 8.98$\pm$7.98 & 8.34$\pm$7.34 \\
J1945-5520 & CO(1-0) & 1.8$\times$1.6 & 2.20 & $<$17.99 & 300.00 & 100.00 & $<$8.91 & $<$8.28 \\
J0057+3021 & CO(2-1) & 0.5$\times$0.32 & 0.18 & 13.52$\pm$ 0.20 & 582.32 & 20.00 & 8.27$\pm$7.29 & 7.63$\pm$6.64 \\
J1301-3226 & CO(2-1) & 0.68$\times$0.65 & 0.31 & $<$5.15 & 300.00 & 100.00 & $<$7.87 & $<$7.22 \\
J2131-3837 & CO(2-1) & 0.67$\times$0.61 & 0.51 & 1.08$\pm$ 0.18 & 540.65 & 20.00 & 7.27$\pm$6.56 & 6.62$\pm$5.92 \\
J1407-2701 & CO(2-1) & 0.69$\times$0.59 & 0.32 & 3.75$\pm$ 0.50 & 147.77 & 20.00 & 7.96$\pm$7.00 & 7.31$\pm$6.35 \\
J2009-4849 & CO(2-1) & 0.64$\times$0.58 & 0.40 & 6.17$\pm$ 1.03 & 448.41 & 22.00 & 9.20$\pm$8.26 & 8.56$\pm$7.61 \\
J1008+0029 & CO(1-0) & 2.21$\times$1.95 & 0.64 & $<$0.74 & 300.00 & 100.00 & $<$9.16 & $<$8.52 \\
J1221+2813 & CO(1-0) & 2.14$\times$1.39 & 0.07 & $<$0.11 & 300.00 & 180.00 & $<$8.40 & $<$7.76 \\
J0623-6436 & CO(1-0) & 0.79$\times$0.6 & 0.28 & 4.68$\pm$ 0.91 & 671.67 & 100.00 & 10.21$\pm$9.33 & 9.57$\pm$8.69 \\
J1217+3007 & CO(1-0) & 2.56$\times$1.51 & 0.23 & $<$0.21 & 300.00 & 100.00 & $<$8.88 & $<$8.24 \\
J1427+2348 & CO(1-0) & 1.78$\times$1.32 & 0.27 & $<$0.28 & 300.00 & 100.00 & $<$9.18 & $<$8.55 \\
J1332+0200 & CO(1-0) & 2.74$\times$2.24 & 0.75 & $<$0.39 & 300.00 & 100.00 & $<$9.59 & $<$8.95 \\
J1356-3421 & CO(1-0) & 1.1$\times$0.88 & 0.28 & $<$0.34 & 300.00 & 100.00 & $<$9.56 & $<$8.93 \\
J0943-0819 & CO(1-0) & 4.78$\times$2.19 & 0.69 & $<$0.36 & 300.00 & 100.00 & $<$9.60 & $<$8.96 \\
J1220+0203 & CO(1-0) & 0.34$\times$0.26 & 0.36 & $<$1.39 & 300.00 & 102.00 & $<$10.24 & $<$9.60 \\
J1547+2052 & CO(1-0) & 0.23$\times$0.21 & 0.58 & $<$2.83 & 300.00 & 100.00 & $<$10.63 & $<$9.99 \\
J2341+0018 & CO(1-0) & 0.7$\times$0.57 & 0.55 & 4.16$\pm$ 0.56 & 350.00 & 20.00 & 10.84$\pm$9.93 & 10.20$\pm$9.29 \\
J1305-1033 & CO(1-0) & 0.47$\times$0.42 & 0.14 & 0.67$\pm$ 0.17 & 173.67 & 100.00 & 10.06$\pm$9.25 & 9.42$\pm$8.61 \\
J0242-2132 & CO(1-0) & 0.74$\times$0.63 & 0.30 & $<$0.41 & 300.00 & 100.00 & $<$9.95 & $<$9.31 \\
J0006-0623 & CO(1-0) & 2.27$\times$1.3 & 44.60 & $<$23.17 & 300.00 & 100.00 & $<$11.79 & $<$11.15 \\
J1505+0326 & CO(3-2) & 0.81$\times$0.44 & 0.81 & 3.28$\pm$ 0.26 & 376.56 & 20.00 & 10.15$\pm$9.31 & 9.50$\pm$8.66 \\
J0748+2400 & CO(3-2) & 1.22$\times$0.99 & 0.61 & $<$0.44 & 300.00 & 100.00 & $<$9.28 & $<$8.63 \\
J0510+1800 & CO(3-2) & 0.65$\times$0.46 & 0.41 & $<$0.26 & 300.00 & 20.00 & $<$9.07 & $<$8.42 \\
J2349+0534 & CO(3-2) & 7.77$\times$5.09 & 3.13 & $<$1.63 & 300.00 & 100.00 & $<$9.87 & $<$9.22 \\
J2141-3729 & CO(3-2) & 1.4$\times$1.04 & 0.96 & 0.73$\pm$ 0.24 & 255.50 & 20.00 & 9.53$\pm$8.99 & 8.88$\pm$8.34 \\
J0914+0245 & CO(3-2) & 0.05$\times$0.04 & 0.16 & $<$3.34 & 300.00 & 100.00 & $<$10.20 & $<$9.55 \\
J1038+0512 & CO(2-1) & 2.26$\times$1.94 & 0.95 & $<$0.49 & 300.00 & 100.00 & $<$9.80 & $<$9.16 \\
J0940+2603 & CO(3-2) & 0.11$\times$0.09 & 0.68 & $<$3.58 & 300.00 & 60.00 & $<$10.37 & $<$9.71 \\
J1610-3958 & CO(3-2) & 1.02$\times$0.81 & 0.53 & $<$0.41 & 300.00 & 100.00 & $<$9.46 & $<$8.80 \\
J2239-5701 & CO(3-2) & 6.96$\times$5.44 & 5.18 & $<$2.69 & 300.00 & 100.00 & $<$10.36 & $<$9.71 \\
J1058-8003 & CO(3-2) & 1.66$\times$1.09 & 0.89 & $<$0.46 & 300.00 & 100.00 & $<$9.61 & $<$8.96 \\
J0106-4034 & CO(4-3) & 7.07$\times$3.59 & 7.73 & $<$4.02 & 300.00 & 100.00 & $<$10.36 & $<$9.66 \\
J0217-0820 & CO(2-1) & 1.91$\times$1.46 & 0.58 & $<$0.30 & 300.00 & 100.00 & $<$9.81 & $<$9.17 \\
J2320+0513 & CO(2-1) & 1.2$\times$1.19 & 0.48 & $<$0.25 & 300.00 & 100.00 & $<$9.76 & $<$9.12 \\
J2000-1748 & CO(2-1) & 0.4$\times$0.27 & 0.36 & $<$0.68 & 300.00 & 100.00 & $<$10.23 & $<$9.59 \\
J1010-0200 & CO(4-3) & 1.17$\times$0.74 & 0.63 & $<$0.37 & 300.00 & 100.00 & $<$9.70 & $<$9.00 \\
J0329-2357 & CO(4-3) & 0.43$\times$0.33 & 0.54 & $<$0.80 & 300.00 & 100.00 & $<$10.03 & $<$9.33 \\
J0946+1017 & CO(2-1) & 0.39$\times$0.34 & 1.98 & $<$2.92 & 300.00 & 100.00 & $<$11.24 & $<$10.59 \\
J0909+0121 & CO(2-1) & 0.71$\times$0.66 & 1.12 & $<$0.89 & 300.00 & 100.00 & $<$10.74 & $<$10.09 \\
J1351-2912 & CO(2-1) & 2.32$\times$1.87 & 0.38 & $<$0.20 & 300.00 & 100.00 & $<$10.10 & $<$9.45 \\
J1743-0350 & CO(4-3) & 0.72$\times$0.55 & 0.06 & 2.26$\pm$ 0.04 & 145.99 & 80.00 & 10.62$\pm$9.65 & 9.92$\pm$8.95 \\
J0125-0005 & CO(2-1) & 3.26$\times$2.52 & 0.42 & $<$0.21 & 300.00 & 90.00 & $<$10.15 & $<$9.51 \\
J0239-0234 & CO(4-3) & 0.4$\times$0.32 & 0.47 & $<$0.70 & 300.00 & 100.00 & $<$10.16 & $<$9.46 \\
J0837+2454 & CO(2-1) & 0.9$\times$0.79 & 0.66 & $<$0.42 & 300.00 & 100.00 & $<$10.49 & $<$9.84 \\
J0118-2141 & CO(4-3) & 0.6$\times$0.53 & 0.77 & $<$0.72 & 300.00 & 100.00 & $<$10.21 & $<$9.51 \\
J0112-6634 & CO(2-1) & 0.54$\times$0.47 & 0.61 & $<$0.64 & 300.00 & 100.00 & $<$10.72 & $<$10.08 \\
J1304-0346 & CO(2-1) & 0.71$\times$0.65 & 0.31 & $<$0.23 & 300.00 & 100.00 & $<$10.33 & $<$9.68 \\
J2134-0153 & CO(2-1) & 1.52$\times$1.25 & 0.47 & $<$0.24 & 300.00 & 100.00 & $<$10.37 & $<$9.72 \\
J1359+0159 & CO(2-1) & 0.7$\times$0.55 & 0.31 & $<$0.26 & 300.00 & 100.00 & $<$10.41 & $<$9.77 \\
J0529-7245 & CO(2-1) & 2.38$\times$2.16 & 0.89 & $<$0.21 & 300.00 & 20.00 & $<$10.33 & $<$9.69 \\
J1147-0724 & CO(4-3) & 0.55$\times$0.37 & 0.58 & $<$0.66 & 300.00 & 100.00 & $<$10.29 & $<$9.59 \\
J0343-2530 & CO(2-1) & 1.47$\times$1.19 & 0.38 & $<$0.19 & 300.00 & 100.00 & $<$10.35 & $<$9.71 \\
J1419+0628 & CO(3-2) & 0.39$\times$0.3 & 0.17 & 0.98$\pm$ 0.18 & 410.76 & 100.00 & 10.72$\pm$9.95 & 10.07$\pm$9.30 \\
J0954+1743 & CO(4-3) & 0.87$\times$0.78 & 0.46 & 0.58$\pm$ 0.09 & 193.30 & 20.00 & 10.31$\pm$9.61 & 9.61$\pm$8.91 \\
J2056-4714 & CO(2-1) & 2.23$\times$1.77 & 0.23 & $<$0.12 & 300.00 & 100.00 & $<$10.18 & $<$9.54 \\
J1520+2016 & CO(3-2) & 0.47$\times$0.32 & 0.13 & $<$0.17 & 300.00 & 100.00 & $<$10.03 & $<$9.38 \\
J1107-4449 & CO(2-1) & 0.97$\times$0.8 & 0.91 & 4.98$\pm$ 0.15 & 43.74 & 20.00 & 11.86$\pm$10.90 & 11.21$\pm$10.26 \\
J0219+0120 & CO(2-1) & 0.86$\times$0.65 & 0.66 & $<$0.45 & 300.00 & 100.00 & $<$10.82 & $<$10.18 \\
J1136-0330 & CO(2-1) & 2.49$\times$2.19 & 0.55 & $<$0.29 & 300.00 & 100.00 & $<$10.64 & $<$10.00 \\
J1146-2447 & CO(4-3) & 2.3$\times$1.47 & 0.49 & $<$0.25 & 300.00 & 100.00 & $<$10.17 & $<$9.47 \\
J0403+2600 & CO(4-3) & 0.06$\times$0.04 & 0.11 & $<$1.07 & 300.00 & 100.00 & $<$10.86 & $<$10.16 \\
J0106-2718 & CO(3-2) & 1.43$\times$1.09 & 0.59 & $<$0.31 & 300.00 & 100.00 & $<$10.64 & $<$9.99 \\
\hline
\end{tabular}
\caption{CO properties of the 66 CO-ARC sample sources. Note that, even for targets with available multiple-J CO observations, only the properties of lowest-J CO transitions are listed here and used for the luminosity function computations.}\label{tab:detections}
\end{table*}
In Figure \ref{fig:sky}, we show the distribution of the 120 sources of our survey in the sky plane projection. The symbol sizes are indicative of the full width at half power (FWHP) of the primary beam of the observations. The total survey area covered by our sample was estimated by summing the FWHP of all individuals observations, taking into account the dependency of the FWHP on the frequency of each observed CO line. This is feasible as the sources are distributed randomly in the sky with almost no overlap between them.
Our final sample, consisting of 120 sources and being representative of the NVSS down to 0.4 Jy, contains 66 sources from the CO-ARC plus 54 literature sources. Details on the CO-ARC sources are given in Tables~B.1 and 1
, including the spectral line of interest per galaxy, the ID of the ALMA observations that we found and reduced per line, as well as the redshift and its source of origin per galaxy. For the bibliographic sources, all pertinent information is summarized in Table~\ref{tab:biblio} in the Appendix.
\section{ALMA Observations}
\label{sec:obs}
\subsection{Data reduction}
The ALMA observations that were analyzed as part of the CO-ARC survey were carried out between Cycle 0 and Cycle 5. The sources were observed as flux, phase, bandpass, or polarization calibrators. The relevant properties of each observing run (e.g.\,scan intent, on-source integration time, CO transition, etc.\,) are reported in Table~\ref{tab:proj}.
The data of the sources observed as flux, phase, and polarization calibrators were reduced using the Common Astronomy Software Application \citep[{\sc CASA};][]{casa} pipeline, which automatically processes the data by performing standard calibration and flagging. Various pipeline versions were used, according to those adopted for the corresponding archival data reduction. When one of our targets was observed in more than one project, once calibrated separately according to the appropriate {\sc CASA} version, the visibilities were then combined using the {\sc CASA} task {\tt concat}. In a minority of cases the projects were observed in the first cycles (0, 1 and early 2) and were calibrated using casa version <=4.2.2, which computes the visibility weights in a different way and therefore they had to be corrected to the same definition using the {\sc casa} task {\tt statwt}\footnote{for more details, see \url{https://casaguides.nrao.edu/index.php?title=DataWeightsAndCombination}}.
Bright quasars were typically used as flux calibrators, resulting in a standard 10\% flux calibration uncertainty.
A different procedure was required for sample sources observed as bandpass calibrators, as any line emission in the source would be flattened by the default pipeline corrections. This is because the bandpass calibration corrects the observed visibilities for frequency-dependent amplitude and phase corruptions. Such corrections are calculated using as model the bandpass calibrator, which is typically a bright quasar (i.e.\,a point-like source in the sky). Due to the Fourier Transform properties, a point-like sky brightness distribution ideally has flat and constant amplitudes in the visibility plane. The bandpass calibration calculates the corrections needed to make the data as similar as possible to such ideal model. Therefore, any line emission in the spectrum of the bandpass calibrator would be treated as instrumental fluctuation and thus flattened. In such cases, the data were then manually calibrated using customised data reduction scripts, which were modified so that the bandpass calibrator was substituted by the phase (or flux) calibrator in the relevant calibration steps. A standard data reduction process was then carried out also in these cases.
\subsection{Imaging}
Where detected, the CO line emission was identified and isolated in the $uv$-plane using the {\sc casa} task {\tt uvcontsub}. This latter forms a continuum model from linear fits in frequency to line-free channels, and then subtracts this model from the visibilities.
Deconvolution and imaging were performed using the {\sc CASA} task {\tt tclean} with Briggs weighting and two robust parameters: 0.5 and 2. The former was chosen as it provides a good trade-off between angular resolution and signal-to-noise ratio (S/N); the latter the best sensitivity. For each type of weighting, we produced three data cubes with different channel widths: 20, 60 and 100~km~s$^{-1}$. For low spectral resolution observations in which the raw channel width was larger than 20 or 60~km~s$^{-1}$, we adopted the lowest possible channel binning. The continuum-subtracted dirty cubes were iteratively cleaned in regions with line emission (first identified through auto-masking) down to selected
thresholds, and then primary-beam corrected.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/noise.png}}
\caption{Distribution of achieved rms sensitivities per ALMA band. The black circles represent the rms noise levels (in log scale and mJy units) achieve for the emission line datacubes of the ALMA calibrators sample in function of the respective observed frequencies (in GHz) or ALMA bands highlighted in colors. The dashed line indicates the median value of 0.482\,mJy for the 66 sources drawn from the CO-ARC.}
\label{fig:noise}
\end{figure}
The achieved synthesised beams and root mean square (rms) noise levels of both detections and non-detections are listed in Table~\ref{tab:detections}. The median rms noise obtained for the emission line data cubes of ALMA calibrators is 0.48\,mJy and some scatter is seen along this value, however, overall the rms noises per ALMA band are fairly homogenous and below 1\,mJy for most of the targets, as we can see in Figure~\ref{fig:noise} the distribution of $\sigma_{\rm rms}$ per frequency (or per ALMA band). We note that even though some of the CO-ARC sources have been presented in the literature before,
we opted to re-image them, so that our archival survey has homogeneous imaging procedure.
\section{Results}
\label{sec:results}
\subsection{Moment maps and spectra of CO-ARC detections}\label{mommaps}
Overall, we report the detection of CO emission in 17 out of the 66 CO-ARC sources listed in Table~\ref{tab:detections}. A marginal detection for J2341+0018 (that most likely reflects a potential flux loss/bad calibration of its data). The remaining 49 analyzed sources are undetected in CO. Together with the 25 CO detections in the 54 literature sources listed in Table~\ref{tab:biblio}, we find that 35\% of radio galaxies in the extended CO-ARC sample contain detectable gas reservoirs. The fraction differs for local and high-$z$ sources, varying from 31/60 sources in the local Universe to 2/23, and 9/37 sources at intermediate and high-$z$, respectively.
Integrated intensity (moment 0), mean line-of-sight velocity (moment 1), and velocity dispersion (moment 2) maps of the CO-ARC detections were created from the continuum-subtracted, cleaned data cubes, keeping pixels down to 3$\sigma$ noise levels. The resulting maps are shown in Figure~\ref{fig:detsam} and in the Appendix in Figure~\ref{fig:detall}.
\begin{figure*}
\centering
\includegraphics[width=17cm]{figures/detections_sample.png}
\caption{Example of integrated intensity (moment 0; left panels), mean line-of-sight velocity (moment 1; middle-left panels), velocity dispersion (moment 2; middle-right) maps and spectra (right panels) of 2 detections in the CO-ARC sample. The moments maps for all the 17 detections are shown in the Appendix in Figure~\ref{fig:detall}. The synthesized beam is shown as a blue ellipse in the bottom-left corner of each moment 0 map. The bar to the top of each moment map shows the color scales (in Jy~beam$^{-1}$~km~s$^{-1}$ and km~s$^{-1}$ for moment 0 and moment 1/2 maps, respectively). East is to the left and north to the top. The observed CO transition is indicated in the top-right corner of each spectral profile. The aperture within which the illustrated spectra have been extracted are overlain in green in each moment 0 maps.}
\label{fig:detsam}
\end{figure*}
Based on the integrated intensity maps (left panels of Fig.\,\ref{fig:detsam} and Fig.\,\ref{fig:detall}), spatially-resolved extended gas distributions are observed in six targets (NGC\,315, IC\,4374, NGC\,3100, J0623-6436, NGC\,6328, and J2009-4349), most of which presenting disc- or ring-like structures. Marginally-resolved or unresolved gas structures are instead observed in all the other cases. The mean line-of-sight velocity maps (Fig.\,\ref{fig:detsam} and Fig.\,\ref{fig:detall}, middle-left panels) show clear evidence of gas rotation in eight cases, that are NGC\,315, IC\,4296, NGC\,3557, NGC\,3100, J0623-6436, NGC\,6328, J2009-4849, and NGC\,7075. Among these, clear kinematic distortions (i.e.\,s-shaped iso-velocity contours) are observed in IC\,4296, NGC\,3100, NGC\,6328. Such disturbances usually trace the presence of unrelaxed sub-structures in the gas distribution, possibly caused by either warps or non-circular motions, or a combination of both.
Accurate 3D kinematic modelling can help differentiating between the two. This has been carried out in details for IC\,4296 and NGC\,3100 by \citet{ruffa19a,ruffa22}. In the former case, such an analysis led to the conclusion that the presence of a position angle warp best reproduces the observed kinematic asymmetries. The additional presence of non-circular motions, however, cannot be fully excluded nor fully established due to the fact that the gas distribution is only marginally resolved in IC\,4296. The features observed in NGC\,3100 are instead found to be the result of a more complex kinematics, including the presence of both a position angle and inclination warp and non-circular motions. These latter have been recently partly identified into a mild jet-induced outflow with $\dot{M}$=0.12~M$_\odot$~yr$^{-1}$ \citep{ruffa22}.
Another case of full kinematic modeling was that of NGC\,6328, which has been carried out in the context of the CO-ARC project and can be found in separate papers (\citealt{michalis21}; Papachristou et al. 2022 in prep.). In short, our detailed analysis suggests the presence of a jet-induced of outflow as likely explanation for the kinematics observed in NGC\,6328, in addition to a highly warped gas distribution. A massive molecular outflow with $\dot{M}$=2300~M$_\odot$~yr$^{-1}$ has been also recently reported by \citet{vayner17} in the galaxy 3C\,298.
The line-of-sight gas velocity dispersion ($\sigma_{\rm gas}$; Figure~\ref{fig:detsam} and Figure~\ref{fig:detall}, middle-right panels) varies significantly among different detections and across the gas sky distribution observed in each source. In most of the detections with disc- or ring-like shapes (e.g.\,NGC\,315, J1505+0326, IC\,4296, NGC\,3557, NGC\,3100, etc.\,), the gas is observed to be dynamically cold towards outskirts of CO distribution, with $\sigma_{\rm gas}\sim10-20$~km~s$^{-1}$ in agreement with the typical ranges expected in unperturbed cold gas clouds \citep[e.g.][]{Voort18}. $\sigma_{\rm gas}$ then progressively increases towards the central gas regions, with the largest values ($\geq50$~km~s$^{-1}$) typically observed around the location of the line peak(s). In first instance, such large values would imply the presence of turbulence within the gas clouds. Observed line-of-sight velocity dispersions, however, can be significantly overestimated due to observational effects. Among these, the limited spatial resolution and/or low detection significance can introduce severe beam smearing (i.e.\,contamination from partially resolved velocity gradients within the host galaxy), which usually dominates in areas of large velocity gradients (such as the central regions of galaxies). Full kinematic modelling is again the only way to derive reliable estimates of the intrinsic gas velocity dispersion: this has been carried out in details for NGC\,315 by \citet{boizelle21} and for IC\,4296, NGC\,3557, NGC\,3100, and NGC\,7075 by \citet{ruffa19a,ruffa22}.
The spatially-integrated CO spectra of the 17 CO-ARC detections are shown in the right panels of Figure~\ref{fig:detall}. These have been extracted from the cleaned data cubes within circular or elliptical apertures enclosing all the observed CO emission (as illustrated in the left panels of Fig.\,~\ref{fig:detsam} and Fig.\,~\ref{fig:detall}). All of the eight above-mentioned sources showing clear signs of rotation in their moment 1 maps also exhibit the classic double-horned spectral shape expected from a rotating disc. In addition, in IC\,4296 a strong absorption feature is observed. This has been analyzed in detail by \citet{ruffa19b} and mainly interpreted as absorption against a bright core continuum. In all the other detections, the spectral profiles show Gaussian-like, centrally peaked shapes. The integrated CO flux of each target, $S_{\rm CO}\Delta v$, has been estimated from the obtained spectral profiles and, together with the line full width at half-maximum (FWHM), is listed in Table~\ref{tab:detections}.
For non-detections, upper limits were calculated in a customary Poissonian way multiplying 3$\sigma_{\rm rms}$ (i.e.\,three times the root mean square noise level determined from line-free channels of each data cube) by the square root of the number of channels contained in an assumed line of FWHM=300\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$}\ . We used the $\sigma_{\rm rms}$ value of the data cubes made with robustness parameter equal to 2 and 100\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$} channel width. This is valid if we consider that the whole CO emission is concentrated within the synthesized beam, which is likely the case for the most distant sources. For nearby targets, to take into account possibly extended emission, we assumed a typical galaxy size of 10\,kpc and corrected the previous noise by multiplying it by the square root of the number of synthesized beams inside the assumed galaxy size.
\subsection{Luminosity and Molecular mass}
A common way to express the CO luminosity is via $L^\prime_{\rm CO}$ in units of $\rm K~km~s^{-1}~pc^{2}$, which is proportional to the surface brightness temperature, T$_{\rm B}$. Its advantage comes from its definition: since it originates from the Rayleigh-Jeans flux multiplied by the beam area, the frequency squared in the Rayleigh-Jeans law gets cancelled out with $1/\nu^2$ of the beam area.
More specifically, the equation of \citet{sol05} provides $L^\prime_{\rm CO}$
from the integrated fluxes $\rm S_{CO}\Delta V$ as:
\begin{equation}
{L^\prime_{\rm CO}}(K km s^{-1} pc^2) = 3.25\times10^7 \frac{S_{\rm CO}\Delta V}{(1+z)^3} \left( \frac{D_{\rm L}}{\nu_{\rm obs}} \right)^2,
\end{equation}
where $\nu_{\rm obs}$ is the observed frequency of the CO line, in GHz, $D_{\rm L}$ is the luminosity distance of the galaxy, in Mpc, $z$ is the redshift and $\rm S_{\rm CO}\Delta V$ is the integrated flux in Jy\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$}. The mid-J CO transitions can be converted to $L^\prime_{\rm CO(1-0)}$ adopting a line ratio, called r$_{\rm J\rightarrow1}=\frac{L^\prime_{\rm CO}(J\rightarrow {J-1} )}{L^\prime_{\rm CO}(1\rightarrow0)}$. For thermalized, optically thick gas emission, r$_{\rm J\rightarrow1}$=1. We adopted the r$_{\rm J\rightarrow1}$ values reported by \citet{carilli13}, r$_{\rm 2\rightarrow1}$=0.99, r$_{\rm 3\rightarrow1}$=0.97 and r$_{\rm 4\rightarrow1}$=0.87, for typical gas excitation conditions of QSOs. The values derived for the CO-ARC sources are in the range 6.5$<$log$L^\prime_{\rm CO(1-0)}<$11.1, while the values reported for sources in the literature sample are 5.4$<$log$L^\prime_{\rm CO(1-0)}<$10.7.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/MH2_z_all.png}}
\caption{Molecular masses derived for the CO-ARC sample. Detection are shown with filled circles and upper limits are indicated with open circles and arrows. For comparison, we show the predicted evolution of the molecular gas content with redshift from semi-empirical models of normal star-forming galaxies with halo masses of $M_{\rm vir}=10^{13}$ and $M_{\rm vir}=10^{14}$ \citep{popping15}.}
\label{fig:mgas}
\end{figure}
Molecular masses can be calculated from ${L^\prime_{\rm CO(1-0)}}$ using the CO-to-H$_2$ conversion factor $M_{\rm H_2}=\alpha_{\rm CO} {L^\prime_{\rm CO(1-0)}}$.
We adopted the standard Galactic CO-to-H2, conversion factor of $\alpha_{\rm CO}=4.36M_\odot$ ($\rm K~km~s^{-1}~pc^{2}$)$^{-1}$ \citep{tacconi13,bolatto13} for consistency throughout the sample. Based on these assumptions, we derived molecular masses in the range of 1.5$\times10^{7}<M_{\rm{H_2}}<7.2\times10^{11} M_\odot$ for the 66 CO-ARC radio galaxies. For the 54 literature-based galaxies, including detections and upper limits, the reported $M_{\rm H_2}$ values vary from 1$\times10^{6}$ to 4$\times10^{11}M_\odot$, converting to the same $\alpha_{\rm CO}=4.36$ used in this work for consistency.
Figure~\ref{fig:mgas} illustrates the molecular masses as a function of $z$ for all the sources analyzed in this work. We compare our mass measurements from semi-empirical predictions of the typical H$_2$ gas content of galaxy halos from \citet{popping15}, with halo masses of $M_{\rm vir}=10^{13}$ and $M_{\rm vir}=10^{14}$. Interestingly, we find that even though at the very local Universe all radio galaxies have molecular gas contents significantly (1-2 orders of magnitude) below those predicted for typical galaxy halo values, the picture changes at high-$z$. At $z$ $>$1, several radio galaxies have molecular gas reservoirs that are very similar to (and even higher than) those of the typical galaxy halo contents. The detected molecular gas reservoirs of radio galaxies at high-$z$ ranges from 5$\times$10$^9$ to $\sim$10$^{12}$ $M_\odot$, while at low $z$ it ranges from 10$^7$ to 10$^{10}M_\odot$. Still, at all $z$, the majority of galaxies have small (undetected) reservoirs than that of typical galaxy halos.
\begin{figure*}
\centering
\begin{minipage}{.47\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/LumCO_lowz_nozcor.png}
\end{minipage}
\begin{minipage}{.47\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/LumCO_highz.png}
\end{minipage}
\caption{Luminosity functions for $0.005<z<$0.3 in the left panel and for 1$<z<$2.5 in the right panel. For the local mass and luminosity functions we compare our findings with those of different surveys: the XCOLD GASS \citep[][open diamonds]{saintonge17,fletcher21}, as the gray dots the FIR-selected CO survey of \citet{keres03} from the IRAS bright galaxy sample, the empirical CO luminosity function derived from the \textit{Herschel} IR luminosity \citep[][black line]{vallini16} and the light orange boxes are results from the ASPECS LP \citep{decarli19}. The {\sc shark} semi-analytic models by \citet{lagos20} for the CO luminosity function of sub-mm selected galaxies are shown in dotted lines. Predictions from the SPRITZ simulation \citep{spritz} are show in orange for all galaxies, and for the different AGN populations in yellow for SF-AGN, in light blue for SB-AGN, and for AGN1 and AGN2 are indicated in red and dark blue, respectively.}
\label{fig:lco}
\end{figure*}
\subsection{CO Luminosity Function}\label{sec:lco}
Previous studies reported luminosity functions for normal and starburst galaxies in the local Universe \citep{keres03, obreschkow09, saintonge17}, however this is the first time that the CO luminosity function of radio galaxies is presented. To construct the luminosity function, we adopted the 1/$V_{\rm max}$ method of \citep{schmidt68}, which provides the number density for each luminosity bin:
\begin{equation}
\Phi (Mpc^{-3} dex^{-1}) = \frac{1}{\Delta\log L} \sum_{i} \frac{\omega_i}{ V_{max,i} },
\end{equation}
where $\Delta\log L$ is the luminosity bin size (common in logarithmic scale), $V_{max}$ is the maximum comoving volume that any galaxy $i$ could be observable given its measured 1.4 GHz flux and the survey limit and $\omega_i$ is the weight shown in Figure~\ref{fig:zbin} corresponding to the completeness correction with respect to the NVSS.
We present the luminosity function for two different redshift ranges. For the range
0.005$<$z$<$0.3 (average $<z>$=0.07), comprising 60 radio galaxies, and for the range 1$<$z$<$2.5 (with average $<z>$=1.53), containing 37 radio galaxies.
A luminosity function at intermediate redshifts is not presented due to the small number of detections. For either redshift range, we present both the luminosity function measured using detections only, and an alternative version using the combination of detections and upper-limits for non-detections. The H$_2$ luminosity (and consequently mass) function is shown in the left panel of Figure~\ref{fig:lco} for $z<$0.3 and in the right panel of the same figure for 1$<z<$2.5.
For the local luminosity function, we have enough number statistics and bins to fit the datapoints in logarithmic scale with an analytical \citet{schechter76} function, as defined in \citet{riechers19}:
\begin{equation}
\log\Phi(L^\prime) = \log\Phi^* + \alpha \log\left( \frac{L^\prime}{L^*} \right) - \left( \frac{L^\prime}{\ln(10)L^*} \right) + \log(\ln(10)).
\end{equation}
In this equation, $\Phi(L^\prime)$ is the number of galaxies per comoving volume per luminosity bin, $\Phi^*$ is the scale density of galaxies per
unit volume; $L^*$ is the scale luminosity defining the knee of
the LF and $\alpha$ is the slope of the faint end of the luminosity function. The best Schechter function fit to our data is shown in Figure~\ref{fig:lco} as a dashed line, with 95\% confidence intervals shown as shaded areas. We found a negative slope of $\alpha$=-0.6, scale luminosity $L^*$ of 5.83$\times$10$^9$\,K\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$} pc$^2$, and turn over at $\Phi^*$=2.39$\times$10$^{-7}$\,Mpc$^{-3}$dex$^{-1}$.
For the local Universe, we compared our results with those from the eXtended CO Legacy Database for GASS (XCOLD GASS) survey \citep{saintonge17, fletcher21}, the CO survey of FIR-selected galaxies from the IRAS bright galaxy sample \citep{keres03}, and the empirical CO luminosity function derived from the \textit{Herschel} IR luminosity \citep{vallini16}. Figure~\ref{fig:lco} shows that the gas mass content locked up in local radio galaxies is $\sim$2-4 orders magnitude less than that in normal star-forming galaxies, with the gap closing at high masses. The same applies when comparing the CO-ARC results to those of the CO observations in the blind ALMA Spectroscopic Survey of the \textit{Hubble} Ultra Deep Field (ASPECS LP) \citet{decarli19} (orange shaded boxes in Fig.~\ref{fig:lco}) for the average $z$ of each bin. ASPECS presents CO luminosity functions for galaxies of any kind, from the local Universe to $z\lesssim$4 and covers an area of 4.6\,arcmin$^2$. Deviations in the (bright-end) shape are seen for the sub-millimeter selected galaxies from the {\sc shark} semi-analytical model predictions by \cite{lagos20}.
The comparison with the \textit{s}pectro-\textit{p}hotometric \textit{r}ealisations of \textit{i}nfrared-selected \textit{t}argets at all-\textit{z} (SPRITZ) simulation predictions \citep{spritz} turns out to be very instructive, as SPRITZ galaxies are separated in different populations including: unobscured and obscured AGN-dominated systems (hereafter AGN1 and AGN2, respectively) with little star formation, and composite systems of AGN and star-forming galaxies. One population of composite systems is similar to spirals hosting a low luminostity AGN (SF-AGN) and the other population resembles starburst galaxies hosting an obscured AGN (SB-AGN).
The CO luminosity function in the local Universe interestingly shows that the number of radio galaxies with H2 gas detection is only little lower than that of pure type 1 or 2 AGN. The comparison between our findings and the SPRITZ simulation at low-$z$ shows that one on every $\sim$four RGs (i.e. a comparable number of RGs) have similar gas reservoirs to pure type 1 or type 2 AGNs. Surprisingly, even the number of RGs with large gas reservoirs at the bright-end of the COLF ($>$10$^{10}$M$_{\odot}$) is not considerably lower, so RGs are not entirely red and dead as it is often believed to be the case. The agreement between the molecular gas content in RGs and the AGN-dominated populations with little star formation (AGN1 and AGN2) from SPRITZ point out that we are comparing AGN populations where the star formation does not dominated the SED, since our $q_{\rm 24}$ selection criteria described in Section~\ref{subsec:sample_arc} assures that we selected galaxies whose radio emission is associated to black hole accretion instead of star formation. A deficit of gas at high masses is only seen for the very bright comparison sample, made to be representative of the 6C radio survey, flux limited to 3\,Jy (green points of Figure~\ref{fig:lco}, the bright sample properties are listed in Table~\ref{tab:bright} in the Appendix). At low CO luminosities, the SPRITZ simulations do not have predictions for $L^\prime<10^{7.25}$\,K\,\hbox{\hbox{km}\,\hbox{s}$^{-1}$} pc$^2$, although the shape of our Schechter function fit appears to agree with an extrapolated curve of the SPRITZ AGN1 and AGN2 predictions at lower luminosities.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{figures/LumL14.png}
\label{fig:l14}
\caption{Radio luminosity function at 1.4\,GHz for the low (left), intermediate (middle) and high (right) redshift bins. We also report the luminosity function of \citet{condon02} in the local Universe for star-forming and AGN population as open squares and stars, respectively. The SPRITZ predictions for radio-loud AGN are shown in gray diamonds \citet{spritz}, as well as the 1.4GHz luminosity function from the radio AGN survey in the VLA-COSMOS field \citep[][open green squares]{smolcic17} for each redshift bin.}
\label{fig:giga}
\end{figure*}
At high-$z$, the gas mass function of detected RGs with CO emission is lower (by up to 2.5 orders of magnitude) than that of local RGs with a corresponding mass content. It is even lower compared to other AGNs from the SPRITZ survey, and all galaxies in the ASPECS LP \citet{decarli19}. This discrepancy largely originates from the fact that only a small number of RGs from the 1.4\,GHz luminosity function are detectable in radio wavelengths at high-$z$. One of possible reason is due to the fact that the continuum spectrum of RGs at $\sim$GHz frequencies decreases with frequency and therefore it suffers from a strong K-correction, contrary to the CO lines and continuum at mm wavelength, whose fluxes increase with frequency, leading to a negative K-correction and the possibility to be detected more easily at high z. Another explanation comes from our survey flux-limit of 0.4\,Jy, implying that only the most extremely luminous galaxies at the bright-end of the radio luminosity function can fit this criteria (see Figure~\ref{fig:giga} and text below). In other words, bright RGs are rarer to find, so our result needs to be roughly corrected with respect to the number density of undetectable RGs.
A rough attempt to correct this bias is made from the fraction of RGs in the inaccessible part of the 1.4\,GHz luminosity function, assuming a similar gas mass distribution throughout. For this reason, we created the 1.4 GHz luminosity function of each $z$ bin (Figure~\ref{fig:giga}) and identified what part of it our observations cover. At low-$z$, our survey covers a large part of the luminosity range of the SPRITZ simulations or of the NVSS and UGC galaxy samples \citet{condon98, condon02}. However, at high-$z$, the integral of the covered 1.4GHz luminosity function is much lower than that of the SPRITZ simulations, indicating that we only see the brightest of every $\sim$7000 galaxies. Therefore $\sim$7000 galaxies are missed, and based on our previous results, 35\% of them would, on average, contain detectable gas fractions. When accounting for the $\sim$2000 missing faint RGs in this way, we find that the gas mass content would be significantly higher than that in the local Universe, and comparable to that of other populations at high-z.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{figures/pho_mh2_z.png}}
\caption{Evolution of the cosmic molecular gas density, $\rho(M\rm{_{H_2}})$, with redshift. Our results derived directly from the detection of the COLF are shown as the purple points and corresponding uncertainties boxes. The extrapolation of the number density of missing sources in the intermediate and high $z$ bins due to our flux limit is shown in the filled triangles. The $\rho(M\rm{_{H_2}})$ for the representative sample of the 6C catalog is indicated in the green box, and for the very bright end of our radio luminosity function ($L_{1.4GHz}>3\times10^{25}$\,W/Hz) is shown in the purple star. The light orange boxes indicate the results from the ASPECS LP survey \citep{decarli19} and the dotted box from the COLDz \citep{riechers19}. The dark and light blue lines represent the predictions from SPRITZ from all AGN hosts, including SF-AGN and SB-AGN, and the AGN-dominated (only AGN1+AGN2) populations.}
\label{fig:rho}
\end{figure}
\section{Discussion: evolution of the M$_{\rm H_2}$ content of gas in radio galaxies with $z$}
By integrating $L'\Phi(L')$ over all luminosity function bins and converting it into mass, we obtain the molecular gas content per comoving Mpc$^3$ of the Universe, $\rho(M\rm{_{H_2}})$, at different epochs (Figure~\ref{fig:rho}). Indeed, this plot shows that the derived molecular gas content of bright enough galaxies to be detected at 1.4 GHz (circles) shows a decrease with $z$ due to the Malmquist bias. When correcting for the density of missing RGs as indicated above, i.e., by multiplying the derived high-$z$ $\rho(M\rm{_{H_2}})$ by $\sim$2000, then the corrected value is less than an order of magnitude below that of AGN1 plus AGN2, as in the local Universe. So, under the rough assumption that detectable and undetectable RGs have a similar gas mass distribution, the amount of gas locked-up in RGs and AGNs is similar at $z>1$ - as the local Universe. Likewise, in the intermediate $z$, we only see the brightest of every 5000 galaxies, from which we deduce the appropriate $\rho(M\rm{_{H_2}})$ correction. However, we draw no further conclusions from this bin, as it suffers from small number statistics (only 2 detections out of 23 sources there).
We further investigated our findings by only examining the most luminous objects at different epochs. Due to the $z$ evolution of luminosities \citep{pracy16}, we avoided directly comparing equal-L galaxies at different $z$. Instead we compared the gas mass content of the brightest of every few thousand RGs at any given $z$. For comparison, we cut the local 1.4GHz luminosity function at 9$\times$10$^{25}$
W/Hz, thus using only the brightest of every 7000
local galaxies. We then reconstructed the CO luminosity function and local Universe gas mass density using these galaxies. The outcome of this computation is the open star in Figure~\ref{fig:rho}. We find that the same amount of molecular gas is locked up in the brightest 1/5000-1/7000 radio galaxies at any given redshift. This result indicates that the net effect of intergalactic medium inflows on one hand and black hole feedback and star formation on the other hand, integrated over time, leads to the same amount of residual molecular gas in the brightest radio galaxies. It is, thus, plausible that radio-mode feedback indeed acts as a ``maintenance'' mode. To further check this, an independent measurement of the star formation rate in these galaxies is needed. This will be presented in a future paper from the fitting of the IR data.
\section{Summary}
\label{sec:summary}
To quantify the gas mass content of radio galaxies at different epochs, we examined a sample of 120 radio galaxies from the local Universe up to z$\sim$2.5. Our sample was meticulously constructed to be complete, representative of the 1.4\,GHz NVSS survey when flux limited to 0.4\,Jy. All galaxies in it have radio emission associated with black hole activity, as indicated by a radio-to-infrared ratio q$_{\rm 24\mu m}<$0.5. Of the sources, 66 were ALMA calibrators, originating from the ALMA Radio-Source Catalogue, and were often observed as such. Another 54 sources, often comprising more extended radio emission,
were compiled from the literature. For all galaxies, CO (1-0) up to (4-3) data were used for the gas mass content derivation. Our main findings can be summarized as follows.
\begin{itemize}
\item
We detected CO emission in 35\% of the sample galaxies, with molecular masses ranging from $\rm 1\times 10^7 <M_{H_2}< 7 \times 10^{10} M_{\odot}$ at low-$z$ and $\rm 2\times 10^{10} <M_{H_2}< 7 \times 10^{11} M_{\odot}$ at high-$z$. Two sources with outflows were seen among the 17 sources with newly reduced ALMA data.
\item
One quarter of the 1$<$$z$$<$2.5 radio galaxies contain at least as much molecular gas as a typical galaxy halo of 10$^{13}$-10$^{14}$ $M_{\odot}$. Therefore, a large number of radio galaxies at high-$z$ would reside at a "normal" or even starbursty for that epoch host. Still, at any $z$, most radio galaxies are more depleted and thus evolved than the typical galaxy of a simulated halo.
\item
The CO luminosity function in the local Universe interestingly shows that the number of radio galaxies with \,\hbox{$\rm{H_ 2}$}\ gas detection is only little lower than that of pure type 1 or 2 AGN.
\item
At 1$<$$z$$<$2.5, the CO luminosity function is significantly lower than that predicted by simulations for other galaxy types, because only hard-to-find bright radio galaxies are detectable. A rough correction of the fraction of missing sources of the 1.4 GHz luminosity function implies that the CO luminosity function would be 2000 times higher. This would close the gap between the various high-$z$ populations and it would imply that the gas mass locked up in high-$z$ galaxies is considerably higher than in the local Universe
\item
The gas mass content of the brightest 1/5000-1/7000 radio galaxies at any $z$ is similar. It is plausible that these molecular gas reservoirs are maintained at similar levels thanks to both the consumption of gas due to star formation and the delay of inflows from the intergalactic medium due to radio activity.
\end{itemize}
This work demonstrate a potential case of exploring the wealth of the ALMA archival observations, including calibrators, for scientific purposes. The CO luminosity function of radio galaxies derived here might be useful for benchmarking of cosmological simulations and constraining the gaseous content of radio galaxies at the local Universe up to the star formation/activity history peak at $z\lesssim$2.5, where feedback could be more crucial for galaxy evolution.
\begin{acknowledgements}
We thank the referee for the constructive review and comments. KMD, AA, MP, JAFO, and IR acknowledge financial support by the Hellenic Foundation for Research and Innovation (HFRI), under the first call for the creation of research groups by postdoctoral researchers that was launched by the General Secretariat For Research and Technology under project number 1882 (PI Dasyra, Title: "Do massive winds induced by black-hole jets alter galaxy evolution? Evidence from galaxies in the ALMA Radio-source Catalog (ARC)"). AA also acknowledges the projects ``Quantifying the impact of quasar feedback on galaxy evolution'', with reference EUR2020-112266, funded by MICINN-AEI/10.13039/501100011033 and the European Union NextGenerationEU/PRTR, and from the Consejer\' ia de Econom\' ia, Conocimiento y Empleo del Gobierno de Canarias and the European Regional Development Fund (ERDF) under grant ``Quasar feedback and molecular gas reservoirs'', with reference ProID2020010105, ACCISI/FEDER, UE.
JAFO acknowledges the financial support from the Spanish Ministry of Science and Innovation and the European Union -- NextGenerationEU through the Recovery and Resilience Facility project ICTS-MRR-2021-03-CEFCA. IR also acknowledges support from the UK Science and Technology Facilities Council through grant ST/S00033X/1. This paper makes use of the ALMA data with project IDs listed in Tables~\ref{tab:proj} and 1. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. We made use of the NASA/IPAC Extragalactic Database (NED), and of the VizieR catalogue access tool, CDS, Strasbourg, France. This research made use of Astropy, a community developed core Python package for Astronomy.
\end{acknowledgements}
\bibliographystyle{aa}
| c9b065f45982f4f1843497eca8a8c3275e8ff9cf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{intro}
Quantum computers are deemed promising technologies for solving industrial problems from various sectors like automotive, chemical, insurance, and technology. One of the main problem domains for industrial problems is optimization, as identified in the report prepared by Quantum Technology and Application Consortium (QUTAC) \cite{bayerstadler2021industry}. Recently, there have been attempts to solve optimization problems using near-term quantum computers, through variational quantum eigensolver (VQE) \cite{tilly2021variational}, quantum approximate optimization algorithm (QAOA) \cite{farhi2014qaoa}, and quantum annealing (QA) \cite{das2008annealing}.
Quantum annealing is a heuristic method for solving optimization problems. It operates in the framework of quantum adiabatic computing, which is a quantum computing model alternative to gate-based. Since many optimization problems are proven to be NP-hard, quantum annealing has gained significant interest as an up-and-coming tool to target them. Quantum annealers are commercially available by the D-Wave company \cite{johnson2011quantum}, and a vast amount of research has been devoted to identifying potential use cases \cite{applicationdwave}. D-Wave quantum annealers have been utilized to solve problems from different domains such as transportation \cite{domino2021quadratic, salehi2022unconstrained, domino2021quantum, yarkoni2021multi}, finance \cite{mugel2022dynamic, kurowski2020hybrid}, chemistry \cite{genin2019quantum, teplukhin2020electronic, mato2021quantum}, and computer science \cite{asproni2020accuracy,jiang2018quantum}.
Identified among the use-cases of quantum computing by BMW Group \cite{luckow2021quantum}, optimization of pre-production vehicle configurations is one of the challenges in the automotive industry. Every year, new features and car components are launched by the companies, and various tests should be carried out before the series production. The tests under consideration range from the validation of the model's functionality to the evaluation of the new components. Consequently, pre-production vehicles are built for testing purposes. As the construction of pre-production vehicles is costly and complex \cite{tiepelt2016finding}, it is desirable to reduce the number of required test vehicles. Hence, the test cars should be configured to cover as many tests as possible while meeting some dependency constraints among the different features.
Some of the attempts in solving the test-vehicle configuration optimization problem use the framework of satisfiability. The features of the vehicle are represented by Boolean variables, indicating whether the feature or the component exists or not. As each vehicle configuration should satisfy the feasibility rules concerning the different features of the car and the requirements imposed by the tests, such rules can be modeled through Boolean constraints. In \cite{tiepelt2016finding}, the authors present a Max-SAT framework that uses a greedy approach and tests it on small-scale real-world data. The problem is also studied through finding minimum set cover in \cite{walter2015optimal}, where the authors formulate the problem as a minimum set cover problem and use SAT solver to check the feasibility of the configurations.
In this paper, our main goal is to exploit quantum annealing to solve the test-vehicle configuration problem. Our problem definition is based on the use case ``Optimizing the Production of Test Vehicles'' given in the BMW Quantum Computing Challenge \cite{bmwchannelge} and takes into account various buildability constraints. We use an optimization approach, where all the variables are Boolean, and the conditions are given through Boolean constraints, and we aim to minimize the number of vehicles that will be used in testing. Among the various solvers provided by D-Wave, we use the newly introduced hybrid solver for constrained quadratic models \cite{cqm}. The mentioned hybrid solver requires the problem to be encoded as a constrained quadratic model (CQM). In CQM, which is also known as the quadratically constrained quadratic programming in the literature, the problem is identified through a quadratic objective function and quadratic constraints defined over binary and integer variables. Once the problem is formulated, the hybrid solver takes advantage of both the classical heuristic methods and the D-Wave quantum processors. As the hybrid solver is the proprietary of the D-Wave company, the exact way it operates is not revealed.
While modeling the problem as a CQM, our primary concern is to use as few qubits as possible. We propose an optimization model for which the total number of qubits grows in the order $O(n(f + o + q ))$, where $n$, $f$, $o$, $q$ are the numbers of vehicles, features, vehicle types, and tests respectively, and the number of required qubits is independent of the number of constraints. The analysis applies for both SAT and Max-SAT formulations, the former answering the question ``Are $n$ cars sufficient to cover all the tests?", and the latter aiming to `maximize the number of tests covered using $n$ test vehicles`. To benchmark the results obtained from the CQM solver, we develop an integer linear programming formulation. Since both problems are notably time-consuming for current quantum and classical solvers, in order to benchmark the efficiency of the classical and quantum solvers, we analyze the performance of a greedy optimization procedure based on the Max-SAT formulation. We test the performance of the algorithms on the dataset provided by the BMW Quantum Computing Challenge~\cite{bmwchannelge} using D-Wave's hybrid CQM Solver, and the classical solvers CBC (COIN-OR Branch-and-Cut) and Gurobi. The results indicate that the classical optimization algorithms outperform the performance of D-Wave's hybrid solver in minimizing the number of required cars and running time. Furthermore, we consider a variation of SAT problem that includes the scheduling of the tests, however, this model requires far more qubits; thus we claim that it is particularly inefficient for practical purposes. As per the authors' knowledge, this is the first attempt to benchmark the performance of the hybrid solver for CQM compared to classical solvers.
The rest of the paper is organized as follows. In Sec.~\ref{sec:preliminaries}, we present basic concepts related to SAT, Max-SAT, linear programming, quantum annealing, and constrained quadratic model as well as discuss the problem statement. In Sec.~\ref{sec:problem-formulation} we describe the problem formulation and the optimization approach we follow for solving the problem including a brief discussion on the required resources. In Sec.~\ref{sec:results}, we present and discuss the experimental results. We conclude with final comments in Sec.~\ref{sec:conclusion}.
\section{Background} \label{sec:preliminaries}
In this section, we briefly explain the necessary background information on satisfiability problems, linear programming, and quantum annealing concepts.
\subsection{Satisfiability problems}
\textit{Satisfiability problem} \cite{gu1996algorithms} (SAT) is the problem of determining whether there exists an assignment to the binary variables that make a given Boolean expression true. A \textit{Boolean expression} consists of \textit{Boolean variables} $x_1,x_2,\ldots, x_n$, that are combined together using logical OR ($\lor$) and AND ($\land$) operations and the negation operator ($\neg$). For instance, the Booelan expression $\phi = (\neg x_1 \land x_2) \lor x_3$ is satisfiable as $x_1=0, x_2=1, x_3=0$ is a satisfying assignment for $\phi$. SAT is the first problem to be proven to be NP-Complete~\cite{cook1971complexity, garey1979guide}. Hence, solving a large family of problems recognized as core to a number of areas in computer science and mathematics is as hard as solving the SAT problem.
A formula is said to be in \textit{conjunctive normal form} (CNF), if it is written as conjunction of clauses. A \textit{clause} is a disjunction of \textit{literals}, where a literal is either a variable (positive literal) or its negation (negative literal). Any Boolean formula can be expressed in CNF. For instance we can express $\phi$ in CNF as $(\neg x_1 \lor x_3) \land (x_2 \lor x_3)$. SAT problem can be equivalently defined as the question of whether there exists an assignment to the Boolean variables that make the formula
\begin{equation}
C_1\land C_2\land C_3\land\ldots\land C_m = \bigwedge_{i=1}^m C_i
\end{equation}
satisfiable, where $C_i$ is a clause and there are $m$ clauses in total.
\textit{Maximum satisfiability problem} (Max-SAT) is a generalization of the SAT problem and can be considered as the optimization variant of the decision version. The goal is to find an assignment that maximizes the number of satisfied clauses. Note that knowing the optimal number of satisfied clauses, we can also deduce the solution to the SAT problem; therefore, we can conclude that the Max-SAT problem is NP-Hard.
A further generalization is the \textit{weighted maximum satisfiability problem} (weighted Max-SAT), in which each clause is associated with a weight, and the aim is to maximize the weighted sum of the satisfied clauses.
\subsection{Linear programming}
\textit{Linear programming} (LP) is concerned with optimization of an objective function subject to equality and inequality constraints, such that the objective and the corresponding constraints are linear. Linear programs can be expressed in canonical form as
\begin{alignat*}{3}
&\text{minimize} \hspace{1em}&& c^\intercal x \\
&\text{subject to} \hspace{1em }&&Ax \leq b \\
&{} &&x \geq 0,
\end{alignat*}
where $c \in \mathbb{R}^n$, $A \in \mathbb{R}^{n \times n}$ and $ b \in \mathbb{R}^n$. The aim is to find the vector of variables $x$ that minimizes the objective function subject to the given constraints. When all variables $x_i$ are integers, then the problem takes the name \textit{integer linear programming} (ILP) which is NP-Hard in general. If the variables are further restricted to the set $\{0,1\}$, then the problem is called \textit{0-1 linear programming} (0-1 LP). Any ILP can be converted into 0-1 LP. For solving ILPs, there are heuristic methods like simulated annealing \cite{kirkpatrick1983optimization} and exact methods including cutting plane \cite{kelley1960} and branch-and-bound methods \cite{Lawler1966}.
There are various commercial and open-source solvers and toolkits for solving linear programs. PulP is a Python library \cite{mitchell2011pulp} that provides tools for modeling problems and an interface for accessing various solvers. In this paper, we use PulP to model our problems and two different solvers to get the results: CBC (Coin-or branch and cut)\cite{forrest2005cbc} solver which is the default one in PulP, and Gurobi solver \cite{gurobi}.
\subsection{Quantum annealing}
\textit{Adiabatic quantum computing} (AQC) is an analog computational model that relies on quantum adiabatic theorem which states that a quantum state that is initially in the ground state is likely to stay in the ground state given that the evolution takes place slow enough. Some assumptions of AQC are lifted in \textit{Quantum annealing} (QA), which is a meta-heuristic method for solving optimization problems using the quantum adiabatic theorem.
The problem Hamiltonian $ H_p $ is designed so that its ground state encodes the solution to the problem of interest, and an initial Hamiltonian $ H_0 $ is picked whose ground state is known and easy to prepare. The system is initialized with the ground-state of $ H_0 $, and an adiabatic evolution path is followed so that the system ends up in the ground-state of $ H_p $. This is achieved by evolving the system with the time-varying Hamiltonian $ H(t) $ expressed as
\begin{equation}
H(t) = \left(1- \frac{t}{\tau}\right)H_0 + \frac{t}{\tau}H_p.
\end{equation}
The quantum adiabatic theorem assures that the system always remains at the ground state of $H(t)$ for sufficiently large real evolution time $\tau$. When $t=\tau$, $H(t)$ equals $H_p$ and ideally the system is expected to be in the ground state of $ H_p $.
Commercially available quantum annealers are provided by the D-Wave company. D-Wave quantum processing units (QPUs) implement the initial Hamiltonian $H_0 = -\sum_i \sigma_i^x$, where $\sigma_i^x$ is the Pauli-X operator acting on $i$-th qubit and the problem Hamiltonian should be stated in the form of an Ising model $H_p = \sum_{i>j}J_{ij}\sigma^z_i \sigma^z_j + \sum_{i}h_{i}\sigma^z_i $, where $ J_{ij} $ denotes the interaction between sites $ i $ and $ j $, $ h_{i} $ is the external magnetic field applied on site $ i $ and $\sigma_i^z$ is Pauli-Z operator. Nevertheless, it is much more convenient to formulate problems over binary variables. Any problem that is expressed as a quadratic unconstrained binary optimization (QUBO) problem can be easily converted into an Ising model by replacing the binary variables $b_i$ with $(1-s_i)/2$. QUBO involves the minimization of a quadratic objective function defined over binary variables. While it is an unconstrained model, by using the \textit{penalty method}, one can incorporate the constraints to the objective function \cite{lucas2014ising}. We would like to point out that any ILP formulation can be formulated as a QUBO and we refer readers to \cite{salehi2022unconstrained} for a detailed explanation.
When running a problem on D-Wave QPUs, the variables need to be mapped to the QPU architecture as the underlying graph representing the interactions in the QPU is not fully connected; this process is known as the minor embedding \cite{choi2008minor}. Hence, the number of variables in the QUBO formulation should be much smaller than the actual number of physical qubits, making it unachievable to run real-world problems when considering the fact that D-Wave Advantage QPU has 5640 qubits.
D-Wave hybrid solvers can solve much larger problem using a classical-quantum hybrid workflow. With the announcement of the new hybrid solver for constrained quadratic models (CQMs), D-Wave's hybrid solver service (HSS) now consists of three different solvers. Binary quadratic model (BQM) solver accepts problems defined in the form of QUBO. One can define problems over discrete variables in an unconstrained form and use the discrete quadratic model (DQM) solver. In this paper, we will use the CQM solver, which is described in more detail in Sec. \ref{sec:cqm}. All solvers in HSS follow the same workflow. After taking the input, classical heuristic solvers that run parallel on the cloud are called. Those solvers have heuristic and quantum modules that send queries to D-Wave Advantage QPU. The responses taken from the QPU are used to guide the classical heuristic process and improve the quality of the solutions obtained so far. Finally, the heuristic solver returns a solution to the user.
\subsection{Constrained quadratic model} \label{sec:cqm}
\textit{Constrained quadratic model} (CQM) is the name given by D-Wave to the model involving a quadratic objective function and quadratic constraints that are defined over binary or integer variables. In the literature, this is also known as \textit{quadratically constrained quadratic programming}. We can define a CQM as
\begin{alignat*}{3}
&\text{minimize} \hspace{1em}&& x^\intercal P_0 x + q_0^\intercal x \\
&\text{subject to} \hspace{1em }&&x^\intercal P_i x + q_i^\intercal x \leq r_i \hspace{1em}i=1,\dots,m \\
&{} &&x \geq 0,
\end{alignat*}
where $P_i \in \mathbb{R}^{n \times n}$, $q_i \in \mathbb{R}^{n}$ and $r_i \in \mathbb{R}$ for $i=0,1,\dots,m$. The goal is to find the vector $x$ that minimizes the objective function, where $x$ consists of binary and integer variables.
D-Wave has recently introduced the hybrid solver for CQM. Unlike the previous hybrid solvers and quantum annealers of D-Wave, the CQM solver natively supports equality and inequality constraints. This is advantageous for problems involving constraints compared to the QUBO, which is the standard formulation that has been used for quantum annealing so far. First of all, there is no need for removing the constraints through the penalty method, which in turn increases the number of variables if the constraints are in the form of inequalities. Secondly, it removes the difficulty of setting penalty coefficients, which is challenging as the model becomes sophisticated. Thirdly, CQM allows the inclusion of quadratic constraints directly into the model, which is not possible for QUBO. It is also suggested by D-Wave that the hybrid CQM solver should be preferred over the other hybrid solvers in case the problem naturally involves constraints. An experimental evidence for performance comparison is available in \cite{cqm}.
\subsection{Problem definition}
In this section, we will describe the details of the vehicle optimization problem. The configuration of each vehicle is determined by the presence or absence of the availability of the features, and each vehicle has a specific type. Let us discuss the buildability constraints that a vehicle configuration should satisfy \cite{bmwchannelge}.
\begin{itemize}
\item \textbf{Single type requirement:} Each vehicle should have a single type.
\item \textbf{Features allowed per type:} For a given type, only some of the features are available.
\item \textbf{Group features:} Certain groups of features cannot be implemented together (i.e. at most one can be implemented). This constraint is valid for all types.
\item \textbf{Rules per type:} For each type, there are rules which govern the feature set that the vehicle of a specific type should have. Those are mainly implication rules.
\end{itemize}
Finally, we have the \textbf{test requirements} that define the properties of the cars needed for the testing phase. We will assume that each test requires a single car and discuss how this assumption can be removed later on.
Consider $n$ test vehicles, a list of $ f $ features and assume that there are $o$ different types available. Let us assume that there are $ q $ test requirements. For vehicle $i \in [n]$, we will represent the presence/absence of feature $j \in [f]$ through the binary variables $b_{i,j}$, where $b_{i,j}=1$ iff vehicle $i$ has feature $j$. Next we represent the type of the vehicle using the variable $t_{i,j}$, where $t_{i,j} = 1$ iff vehicle $i \in [n]$ is of type $j \in [o]$. And finally we define the binary variables $ p_{i,j}$, to represent whether vehicle is used in test, where $p_{i,j} = 1$ iff vehicle $i\in [n]$ is used for test $j$.
Suppose that there are $ c $ buildability constraints in total. Note that the same set of constraints applies to each vehicle. Each constraint $\phi_k$ can be expressed using a Boolean expression over the binary variables we have defined above. First of all, we would like all buildability constraints to be satisfied for any given number of cars. This can be expressed using the following logical expression:
\begin{equation} \label{eq:build}
\bigwedge_{i=1}^n \bigwedge_{k=1}^c \phi_k(b_{i,1},\dots, b_{i,f},t_{i,1},\dots, t_{i,o}).
\end{equation}
Similarly, we can express the test requirements using Boolean expressions. To start with, each test requires absence or presence of certain features. We need constraint $ \psi_l $ to ensure that the variable $ p_{i,l} $ representing the $ l $'th test requirement is set correctly.
\begin{equation} \label{eq:testc}
\bigwedge_{i=1}^n \bigwedge_{l=1}^q \psi_l(b_{i,1},\dots, b_{i,f},p_{i,l}).
\end{equation}
Now we can identify two different problems based on how we interpret test requirements. Given $ n $ vehicles, the first problem is to decide whether there exist configurations for the given cars so that the buildability constraints are satisfied, and for each test requirement, there is at least one car satisfying the requirement. This results in the following \textit{satisfiability problem}:
\begin{equation} \label{eq:sat}
\bigwedge_{i=1}^n \bigwedge_{k=1}^c \phi_k(b_{i,1},\dots, b_{i,f},t_{i,1},\dots, t_{i,o}) \land \bigwedge_{i=1}^n \bigwedge_{l=1}^q \psi_l(b_{i,1},\dots, b_{i,f},p_{i,l}) \land
\bigwedge_{l=1}^q \bigvee_{i=1}^n p_{i,l}.
\end{equation}
If one wants to find the smallest number of vehicles for which the Boolean formula given in Eq.~\eqref{eq:sat} is true, then the bisection method can be used by starting with a large $ n $ and applying binary search to find the optimal value.
It can be the case that the number of vehicles is fixed, and the aim is to find a configuration of vehicles that satisfy the buildability constraints and maximize the number of satisfied test requirements. So, unlike in the case of SAT, it is not required that all the test requirements are satisfied. Furthermore, each test requirement can be assigned some weight, in which case the aim is to maximize the weighted sum of the fulfilled tests. This yields a variant of \textit{weighted maximum satisfiability problem} that is expressed mathematically as follows:
\begin{equation} \label{eq:maxsat}
\text{maximize}~ \sum_{l=1}^q w_l \prod_{i=1}^n p_{i,l},
\end{equation}
where $ w_l $ is the weight associated with test requirement $ l $, subject to the constraints
\begin{equation}
\bigwedge_{i=1}^n \bigwedge_{k=1}^c \phi_k(b_{i,1},\dots, b_{i,f},t_{i,1},\dots, t_{i,o}) \land \bigwedge_{i=1}^n \bigwedge_{l=1}^q \psi_l(b_{i,1},\dots, b_{i,f},p_{i,l}).
\end{equation}
\section{ILP formulation}\label{sec:problem-formulation}
Having discussed the general overview for the vehicle testing problem, now we are ready to express it as a linear program defined over binary variables. We will be expressing Boolean expressions that correspond to constraints using equalities and inequalities.
\subsection{Buildability constraints}
Here we discuss the formulation of the buildability constraints through linear equalities and inequalities.
\subsubsection{Single type requirement}
The fact that each vehicle should have a single type, can be incorporated using the constraints
\begin{equation}
\sum_{j=1}^o t_{i,j}= 1, \qquad i\in[n]. \label{eq:constr1}
\end{equation}
\subsubsection{Features allowed per type}
Some of the features are not allowed for a specific type. We encode this constraint through the features which are not allowed for the given type. In other words, for each type $j \in [o]$, there exists a collection of features $F_{j}$ such that $t_{i,j}
=1\implies b_{i,k}=0$ for all $k\in F_{j}$ for all vehicles $i\in [n]$. This is expressed by the inequality constraints
\begin{equation}
t_{i,j} + b_{i,k} \leq 1, \qquad i\in [n], j\in[o],k \in F_{j}. \label{eq:constr2}
\end{equation}
\subsubsection{Group features}
Let $ F_G $ denote a collection of feature groups. Given a group of features $ G \in F_G $, we need to make sure that at most one feature from each group $ G $ is implemented. This is equivalent to inequalities
\begin{equation}
\sum_{k\in G} b_{i,k} \leq 1, \qquad i\in [n], \label{eq:constr3}
\end{equation}
for each $G \in F_G$.
\subsubsection{Rules per type}
Depending on the type of the vehicle, one may define rules in the form of implications about the presence or absence of the features in the vehicle as
\begin{center}
\texttt{T1: F2 $\land$ $ \neg $F4 $\land$ $ \neg$F5 $\implies$ F1 $\lor$ F3},
\end{center}
where the first term is the type of the vehicle, and on the left and right sides of the implication, we have conjunction or disjunction of literals. The rule is saying that if a vehicle is of type 1, has feature 2, and does not have features 4 and 5, then it should have at least one of the features 1 or 3.
We group the constraints into several classes depending on the form of the constraint and label it with a 4-character string, where the first two characters are representing the LHS of implication, and the remaining two are the RHS. The first character describes whether the variables on the LHS of the implication are negated or not. There are three possibilities: ``\texttt{0}" if none of the variables are negated, ``\texttt{1}" if all variables are negated,
``\texttt{m}" if some of the variables are negated. Second character on LHS describes the operator used: ``\texttt{\&}" for AND, ``\texttt{$|$}" for OR, \texttt{1} if only a single variable exists. The next two characters have the same meaning but describe the RHS of the constraint. For example, the implication above belongs to the class \texttt{m\&0$|$}. The cases we will consider are inspired by the BMW Quantum Computing Challenge dataset and include the combinations of \texttt{m\&} and \texttt{0|} on the LHS and \texttt{0|}, \texttt{0\&}, \texttt{1|}, \texttt{1\&} on the RHS. We give the logical expression and the corresponding arithmetic expression for the mentioned cases in Table~\ref{tbl:cases}.
\begin{table}[h]
\centering
\begin{tabular}{ccll}
\hline
\multicolumn{1}{c}{Type} & \multicolumn{1}{c}{Position} & \multicolumn{1}{l}{Logical expression} & \multicolumn{1}{l}{\hspace{5pt}Corresponding constraint} \\[7pt]
\hline
\texttt{m\&} &LHS & $ t_{i,j} \land \bigwedge_{r=1}^M b_{i,j_r} \land \bigwedge_{r=1}^\mu b_{i,l_r} $ & \hspace{5pt} $ 1-t_{i,j} + M - \sum_{r=1}^M b_{i,j_r} + \sum_{r=1}^\mu b_{i,l_r}=0$ \\[7pt]
\texttt{0\&} & RHS & $ \bigwedge_{r=1}^N b_{i,k_r}$ & \hspace{5pt} $N - \sum_{r=1}^N b_{i,k_r} = 0$ \\[7pt]
\texttt{1\&} & RHS & $\bigwedge_{r=1}^N \neg b_{i,k_r}$ & \hspace{5pt} $ \sum_{r=1}^N b_{i,k_r} = 0 $ \\[7pt]
\texttt{0|} & RHS & $ \bigvee_{r=1}^N b_{i,k_r}$ & \hspace{5pt} $ 1 \leq \sum_{r=1}^N b_{i,k_r}$ \\[7pt]
\texttt{1\&} & RHS & $ \bigvee_{r=1}^N \neg b_{i,k_r}$ & \hspace{5pt} $ \sum_{r=1}^N b_{i,k_r} \leq N-1 $ \\[7pt] \hline
\end{tabular}%
\caption{Some logical expressions and their corresponding constraints.}
\label{tbl:cases}
\end{table}
Now we will investigate the implications of each type, making use of the arithmetic expressions given in Table~\ref{tbl:cases}.
\paragraph{Case \texttt{m\&0\&}} This category encapsulates \texttt{m\&01}, \texttt{0\&0\&}, \texttt{1\&0\&}, \texttt{0\&01}, \texttt{1\&01}, \texttt{010\&}, \texttt{0101}, \texttt{110\&}, \texttt{1101}. The constraints take the form
\begin{equation}\label{eq:case1}
N - \sum_{r=1}^N b_{i,k_r} \leq N \big(1-t_{i,j} + M -\sum_{r=1}^M b_{i,j_r}+\sum_{r=1}^\mu b_{i,l_r}\big), \qquad i \in [n].
\end{equation}
RHS is 0 iff the assumption is satisfied, in which case $ \sum_{r=1}^N b_{i,k_r} $ should be equal to $ N $. If the assumption is not satisfied, then RHS is at least $ N $, hence the inequality is still correct.
\paragraph{Case \texttt{m\&1\&}} This category encapsulates \texttt{m\&11}, \texttt{0\&1\&}, \texttt{1\&1\&}, \texttt{0\&11}, \texttt{011\&}, \texttt{0111} and the reasoning is the same as above
\begin{equation}\label{eq:case2}
\sum_{r=1}^N b_{i,k_r} \leq N(1-t_{i,j} + M- \sum_{r=1}^M b_{i,j_r} + \sum_{r=1}^\mu b_{i,l_r}), \qquad i \in [n].
\end{equation}
\paragraph{Case \texttt{m\&0|}} This category encapsulates \texttt{0\&0|}, \texttt{1\&0|}, \texttt{010|}, \texttt{110|}. The constraints take the form
\begin{equation}\label{eq:case3}
1 - \sum_{r=1}^N b_{i,k_r}\leq (1-t_{i,j} + M- \sum_{r=1}^M b_{i,j_r} + \sum_{r=1}^\mu b_{i,l_r}), \qquad i \in [n].
\end{equation}
Note that RHS is 0 iff the assumption is satisfied In this case $ \sum_{r=1}^N b_{i,k_r} \geq 1 $ should be true. If the assumption is not satisfied, then the inequality is still correct.
\paragraph{Case \texttt{m\&1|}} This category encapsulates \texttt{0\&1|}, \texttt{1\&1|}, \texttt{011|}, \texttt{111|}. The constraints take the form
\begin{equation}\label{eq:case4}
\sum_{r=1}^N b_{i,k_r} - N + 1 \leq (1-t_{i,j} + M- \sum_{r=1}^M b_{i,j_r} + \sum_{r=1}^\mu b_{i,l_r}), \qquad i \in [n].
\end{equation}
\paragraph{Case \texttt{0|0|}, \texttt{0|1|}} We take the contrapositives and obtain \texttt{1\&1\&} and \texttt{0\&1\&} respectively, which are already discussed above.
\paragraph{Case \texttt{0|1\&}, \texttt{0|0\&}}
The constraints of the form \texttt{0|1\&} are expressed as $ \displaystyle
t_{i,j} \land \bigvee_{r=1}^M b_{i,j_r} \implies \bigwedge_{r=1}^N \neg b_{i,k_r}
$ and equivalently, we have conditions
\begin{align}\label{eq:case5}
\centering
( t_{i,j} \land b_{i,j_1}) &\implies \bigwedge_{r=1}^N \neg b_{i,k_r},\\
&\;\;\;\;\vdots\nonumber\\
( t_{i,j} \land b_{i,j_M}) &\implies \bigwedge_{r=1}^N \neg b_{i,k_r}.
\end{align}
We have $M$ rules of the form \texttt{0\&1\&}. Similarly, the constraints of the form \texttt{0|0\&} translates to rules of the form \texttt{0\&0\&}.
\subsection{Test requirements}\label{sec:requirements}
Test requirements define the properties of cars needed for the testing phase. Each test requires the absence or presence of certain features. We will assume that the test $ j $ is in the form
\begin{equation}
b_{i,{k_1}} \land \cdots \land b_{i,k_{T_j^1}} \land \neg b_{i,l_1} \land \cdots \land \neg b_{i,l_{T_{j}^2}} \land ( b_{i,m_1} \lor \cdots \lor b_{i,m_{T_{j}^3}}).
\end{equation}
The values $ T_j^1 ,T_j^2$ and $ T_j^3 $ may be equal to 0.
Recall that $p_{i,j}$ indicates whether vehicle $i$ is used in test $j$. We need constraints to ensure that the binary variables $ p_{i,j} $ are properly set. Note that test $j$ either imposes some features to exist in the vehicle, in which case we can express it using the inequality
\begin{equation}
-b_{i,k_r} + p_{i,j} \leq 0, \qquad i \in [n],~ r \in [T_{j}^1], \label{eq:tc1}
\end{equation}
or imposes that some features should not exist in the vehicle, which results in the inequality
\begin{equation}
b_{i,l_r} + p_{i,j} \leq 1, \qquad i \in [n],~ r \in [T_{j}^2], \label{eq:tc2}
\end{equation}
or imposes disjunction of some features which translates as the inequality
\begin{equation}
p_{i,j} -\sum_{r=1}^{T_{j}^3} b_{i,m_r} \leq 0 , \qquad i \in [n]. \label{eq:tc3}
\end{equation}
We will call the constraints defined in \cref{eq:tc1,eq:tc2,eq:tc3} as the test constraints. The test constraints are needed both in SAT and Max-SAT approaches.
We will lift the assumption that each test requires only a single car. Let's assume that test $ j $ requires $ k_j $ cars that satisfy the required properties. In case we want to solve the SAT problem, we include the following constraint in our formulation to ensure that the number of vehicles that satisfy test $j$ is $k_j$ for each $j=1,\dots,q$:
\begin{equation}
\sum_{i=1}^n p_{i,j} = k_j, \qquad j \in [q]. \label{eq:constr11}
\end{equation}
If we want to solve the Max-SAT problem, then we need to define an objective function to maximize. One possibility is to take into account the number of cars that satisfy the test and reflect this in the weight. This results in the following objective function:
\begin{equation}
\sum_{i=1}^n \sum_{j=1}^q w_j p_{i,j}.
\label{eq:partial-objective}
\end{equation}
It is clear that there should be some upper bound on the number of cars that satisfy a specific test. For example, for a given test $j$, if more than $k_j$ cars satisfy the test, the excess ones should not add to the objective. Hence, we need the following constraint in the case of the Max-SAT approach:
\begin{equation}
\sum_{i=1}^n p_{i,j} \leq k_j, \qquad j \in [q].
\label{eq:kj-bound}
\end{equation}
To conclude, to solve both the SAT problem and Max-SAT problem, we need the buildability and the test constraints defined in \cref{eq:constr1,eq:constr2,eq:constr3,eq:case1,eq:case2,eq:case3,eq:case4,eq:case5,eq:tc1,eq:tc2,eq:tc3}. For the SAT problem, we need additionally the constraint defined in Eq.~\eqref{eq:constr11}. For the Max-SAT problem, we need additionally the constraint defined in Eq.~\eqref{eq:kj-bound} and the objective function is defined as in Eq.~\eqref{eq:partial-objective}.
\subsection{Scheduling}
A related problem in the automotive industry is vehicle test scheduling. Besides the configuration of the vehicles, there are additional constraints about when and in which order the tests will be performed and whether a vehicle can be used in more than one test, making the problem more complicated. The problem has been considered using classical approaches like constraint programming and mixed-integer linear programming in \cite{mitchell2011pulp, forrest2005cbc, shi2017analytical}. From our perspective, the presented model can be extended to incorporate scheduling constraints, as we will discuss briefly.
We assume that each test takes a single day, and the tests should be completed in $D$ days. We extend the previously introduced $p_{i,j}$ variables into $p_{i,j,d}$ where $d\in[D]$ is the day at which test $j$ is performed with vehicle $i$. We replace $p_{i,j}$ in each previously introduced condition with $p_{i,j,d}$, and if needed a summation over $d$ should be added. For example constraint given in Eq.~\eqref{eq:constr11} will be replaced with
\begin{equation}
\sum_{d=1}^D\sum_{i=1}^n p_{i,j,d} = k_j, \label{eq:constr11-scheduling}
\end{equation}
for each test $j\in[q]$.
From now on, we assume that only a single car is needed for each test. If the $j$-th test requires $k_j$ cars, we create variables $p_{ij_1},p_{ij_2},\dots,p_{ij_{k_j}}$ for test $ j $. Hence, the constraint presented in Eq.~\eqref{eq:constr11-scheduling} takes the form
\begin{equation}
\sum_{d=1}^D\sum_{i=1}^n p_{i,j,d} = 1,
\end{equation}
where $ j $ belongs to the extended list of tests.
To ensure that test $ j $ is performed within the time frame $[ t_j^{\text start} , t_j^{\text end}]$, we need the constraint
\begin{equation}
t_j^{\text start} \leq d \cdot \sum_{d=1}^D \sum_{i=1}^n p_{i,j,d} \leq t_j^{\text end}.
\end{equation}
If for a given test set $\mathcal J=\{j_1,\dots,j_m\}$ we need to use different vehicles for testing, then the vehicle should be used at most once for one of the tests in the given test group. This constraint can be imposed by
\begin{equation}
\sum_{d=1}^D \sum_{j\in \mathcal J} p_{i,j,d} = 1
\end{equation}
for all $i\in[n]$ and each test set $\mathcal J$.
Let us now consider other conditions on scheduling.
Let $K$ be the number of cars that can be tested per one day. We need to ensure that for each day $d\in [D]$, the number of tests performed is at most $K$, which is equivalent to
\begin{equation}
\sum_{i=1}^n \sum_{j=1}^q p_{i,j,d} \leq K. \label{eq:scheduling-constr1}
\end{equation}
Similarly one has to ensure that each car $i\in[n]$ is
tested at most once in each day $d\in [D]$, which is equivalent to
\begin{equation}
\sum_{j=1}^q p_{i,j,d} \leq 1. \label{eq:scheduling-constr2}
\end{equation}
Let us now consider how one can assign groups to each test to impose an order condition among tests from different groups. Let $g_j$ be the group id of the test $j$. We assume, that $g_j$ is an integer in $\{1,\dots,\bar g\}$, s.t. for two tests $j,j'$, where $j$ has to be performed before $j'$ if $g_{j} > g_{j'}$ (\emph{order condition}). In addition, we assume that the tests from group $g_j=1$ are full crash tests, thus not only that they have to be the final test, but also each car can be used only once for such test (\emph{crash condition}).
The order condition can be implemented as follows: For each vehicle $i\in[n]$, for each pair of tests $j,j'$ such that $g_j > g_{j'}$, and for each $d,d'\in[D]$ such that $d<d'$, we add the constraint
\begin{equation}
p_{i,j',d} + p_{i,j,d'} \leq 1.
\end{equation}
Note that with the approach above, we can handle even more complicated test ordering, like the one defined by partial order of tests \cite{partialorder}.
To implement the crash condition, together with the order condition, it is enough to ensure that the car is used for only one crash test. Let $J_{1}=\{j\in[q]: g_j=1\}$ be the set of tests resulting in a crash. The constraint takes the form for each vehicle $ i \in [n] $
\begin{equation}
\sum_{j\in J_1} \sum_{d=1}^D p_{i,j,d} \le 1
\end{equation}
\subsection{Resource analysis}
Let us analyze the number of variables and constraints required by the formulation. To start with, there exist $ n\cdot f $ binary variables $ b_{i,j} $, $ n\cdot o $ binary variables $ t_{i,j} $, and $ n\cdot q $ binary variables $ p_{i,j} $. Overall, we need $O( n(f+o+q)) $ binary variables, which grows linearly in the number of vehicles.
There are $ c $ buildability constraints. The number of test requirements depends on the individual tests and can be expressed as $q + \sum_{j=1}^q T_j^1 + T_j^2 + [T_j^3>0]$, where $ [T_j^3>0]=1 $ if $T_j^3>0 $. The first term results either from the constraint Eq.~\eqref{eq:constr11} or Eq.~\eqref{eq:kj-bound}, depending on the problem in consideration. Assuming that $ T_j^1$ and $ T_j^2$ are negligible compared to $ q $, the total number of constraints can be expressed as $ O(c+q) $.
Let us now consider the number of qubits used for scheduling constraints. Previous considerations are still valid up to part where $p_{i,j}$s were computed, as they are now replaced with $p_{i,j,d}$. So in total we need $O(n(o+f+qD))$ variables. Note that the polynomial is no longer quadratic.
\section{Experimental results}\label{sec:results}
In this section, we will describe the algorithm, the implementation details and present our results.
\subsection{Algorithm}
Based on the formulations presented, one can follow different approaches to find the minimum number of required vehicles. The first is the global bisection method, which is also proposed in BMW use-case specification. The idea is to start with a large $ n $ value and then use binary search to find out the optimal $ n $. As the number of variables grows as the product of the number of vehicles and number of constraints, the limitation with this approach is the large number of variable requirements. For instance, in the case of the CQM solver, the number of variables is limited to 5000. In the case of Gurobi, there is no limit on the number of variables and constraints that can be used in principle, however, the time needed for solving the problem increases as the number of variables and constraints increase, which may result in an intractable problem in practice.
Another approach would be the vehicle-greedy algorithm. We choose an extra parameter $n_{\rm bunch}$ which denotes the number of cars that will be configured at each iteration. After each iteration, the tests are updated by removing the satisfied ones and by diminishing the number of required cars for a test if it is partially satisfied. Then the optimization process is repeated again with the new $n_{\rm bunch}$ cars. The procedure stops after all the tests are covered. Note that we actually maximize the number of covered tests using this approach, and thus solving the Max-SAT problem instance.
\subsection{Implementation}
We used the dataset provided by the BMW Quantum Computing Challenge, which is created based on the BMW Series 2 Gran Coupe. The features and constraints are based on the actual numbers resulting in a real-world problem. The specifications of the dataset are given in Table \ref{tbl:dataset}.
\begin{table}[h]
\centering
\begin{tabular}{l@{\quad\quad}ll}
\hline
Constraint & \multicolumn{2}{l}{Number} \\
\hline
Features allowed per type & \multicolumn{2}{l}{25} \\
Rules per type & \multicolumn{2}{l}{4032} \\
Group features & \multicolumn{2}{l}{41} \\
Test requirements & \multicolumn{2}{l}{643} \\
\hline
\end{tabular}
\caption{A summary of the number of constraints based on the real-world problem.}
\label{tbl:dataset}
\end{table}
When analyzing the buildability constraints, we noticed a significant redundancy in the ``rules per type'' constraints. We realized that some of the constraints apply to all types. Secondly,
some of the constraints apply to all types. In this case, we used a type-independent constraint and heavily reduced the number of constraints. For instance, assuming that the inequality $b_{i,1} \leq 2-t_{i,j} - b_{i, 2}$ exists for all $ i \in [n] $ and $ j \in [o] $, it can be replaced by $b_{i,1} \leq 1 - b_{i, 2}$ for all $ i \in [n] $. In case some constraint was missing for a number of types, but its inclusion for the remaining types was not disruptive (since the left hand side of the implication could not be satisfied for the particular type with any combination of features), we assumed that it applies for all types and used the discussed simplification. When some features were not available for the given type yet appeared as a positive literal in the constraint, they were removed. Finally, some ``rule per type" constraints possessed redundant information because variables were repeated both on the left and right sides. Those constraints were simplified as well, resulting in new constraint types, which are implemented in a form similar to the ones previously mentioned. Besides the buildability constraints, we also performed simplification for the test requirements, by merging the test lines occurring multiple times in the file. Further details can be found in our implementation processing which can be found in the \cite{akash_kundu_2022_6012261} in [path to solver].
We used the vehicle greedy algorithm, taking $ n_{\rm bunch}=1$, hence optimizing a single vehicle at a time. After the mentioned simplifications, the model has 911 binary variables and 6313 constraints. We followed the Max-SAT approach taking Eq.~\eqref{eq:partial-objective} as the objective function and setting all weights equal to 1. We used \texttt{PuLP toolkit} and \texttt{dwave-ocean-sdk} to implement the code for generating the linear program and the constrained quadratic model. The experiments were run on a computer with the following specifications: Intel(R) Core(TM) i9-10900KF CPU @ 3.70GHz; Ubuntu 20.04.3 LTS, 64 GB RAM.
\subsection{Results and discussion}
As mentioned earlier, we obtained the results by setting $n_{\rm bunch}=1$ i.e. one car is configured at each iteration. Hence the number of successful iterations (which outputs a valid car) corresponds to the number of needed vehicles. The algorithm stops if no tests are remaining to be satisfied. For the classical solvers, the experiments are repeated 60 times, and for CQM solver, it is repeated 3 times. CBC and Gurobi algorithms always return the same result in each experiment, i.e. they terminate at the same iteration and return the same number of cars. Meanwhile, for CQM, we took the best possible outcome.
In the Fig.~\ref{fig:remaining-tests-1-car}, we illustrate the number of remaining tests after each iteration using CQM, CBC, and Gurobi solvers setting $n_{\rm bunch}=1$. CBC solver returns the smallest number of vehicles which is 62, and Gurobi solver returns 64. We would like to note that in the experiment with CQM solver, one test that requires a single vehicle remains after the 65'th iteration. Thus, we can conclude that the CQM solver returns 66; however, the solver fails to find a configuration that satisfies the test in the 66'th iteration.
\begin{figure}[tbh!]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=\linewidth]{merged_averaged_remaining-cars_vs_iterations}
\caption{}
\label{fig:remaining-tests-1-car}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.975\linewidth]{averaged_runtime_vs_iteration}
\caption{}
\label{fig:optimizing-1-cars}
\end{subfigure}
\caption{Illustration of (a) variation in runtime and (b) number remaining tests with respect to iterations for CQM, CBC and Gurobi solvers.}
\label{fig:plots-exp-results}
\end{figure}
The problem size gets smaller over the iterations as some tests are removed. We analyze how the runtime changes as the problem size gets smaller in Fig.~\ref{fig:optimizing-1-cars} for $n_{\rm bunch}=1$. The runtimes are calculated by taking the average over the repeated experiments. The primary observation while comparing the performance of CQM, CBC, and Gurobi is that after the first few iterations, the runtime of CQM solver saturates near $5$ seconds. This is because the default runtime for the solver is $5$ seconds, and one can not go below it. Meanwhile, the fluctuation in runtime for CBC solver with the number of iterations is visible and varies in the range 1--25 seconds, and the fluctuations in runtime comparatively stabilize for the number of iterations $\ge 30$. Finally, for Gurobi the runtime always stays $\le 1$ second; hence it takes the least amount of time to satisfy all the tests. In Table \ref{tab:averaged-overall-runtime}, a summary of the overall runtime taken by the solvers is depicted. Overall, it takes $6.309$ seconds for Gurobi to find the solution, which is significantly small compared to the other solvers. Although CBC solver takes a longer time, it returns the best solution. CQM solver performs worse both in runtime and minimizing the number of required cars, it takes $371.975$ seconds, in which only $0.001$ of the time is spent on QPU.
We have also checked the performance of the solvers when $ n_{\rm bunch}=5$, which is the maximum possible number that can be taken without exceeding the 5000 variables limit of the CQM solver. Gurobi solver returned 63 cars, so we can say that the result is slightly improved. However, the overall experiment took a significantly longer time (83 seconds). For the CQM solver, no feasible solutions were obtained with the default time limit of 5 seconds. We observed that the returned samples either violated the ``single type constraint" and the vehicles had no type (in that case all constraints related to ``rules per type" are automatically satisfied) or several other constraints were violated. When the time limit was increased to $10$ seconds, then the CQM solver was able to return a feasible solution at each iteration. However, after the 13'th iteration (after $65$ vehicles were configured), there were still $35$ tests remaining to be satisfied, hence the optimization quality was worse. We would like to remark that the CBC solver failed to return any result within a reasonable amount of time.
\begin{table}[t!]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
\multicolumn{2}{l}{Solver} & \multicolumn{2}{l}{Overall runtime (in seconds)} \\
\hline
\multicolumn{2}{l}{Gurobi} & \multicolumn{2}{l}{6.309} \\
\multicolumn{2}{l}{CBC} & \multicolumn{2}{l}{136.800} \\
\multicolumn{2}{l}{CQM} & \multicolumn{2}{l}{371.975} \\
\hline
\end{tabular}
\caption{The averaged overall runtime by the CBC, Gurobi (classical) and CQM (quantum) solvers.}
\label{tab:averaged-overall-runtime}
\end{table}
Investigating the experimental evidence, we can conclude that the classical solvers outperform the CQM solver. Nevertheless, CQM has the potential to be a promising tool for large problems in near future.
\section{Conclusion and future work}\label{sec:conclusion}
In this paper, we proposed a greedy algorithm for solving the optimization problem of the production of test vehicles using the new hybrid Constrained Quadratic Model Solver (CQM Solver) by D-Wave. We provided a constrained quadratic model formulation for the problem that requires number of qubits linearly proportional to the number of vehicles, car types, features, and tests. We implemented the code for generating the constrained quadratic model and solved the problem instance provided by BMW Quantum Computing Challenge on D-Wave CQM Solver. We benchmarked the results by implementing the integer linear program formulation and running the same algorithm using classical solvers like CBC and Gurobi. The results showed that CQM Solver is comparable to classical solvers in optimizing the number of required vehicles, yet the classical solvers provide the solutions faster. Keeping in mind that the challenges faced and the ongoing efforts in the development of quantum computers, the results obtained from CQM solver for a real-world problem are promising.
The current CQM Solver is limited to 5000 variables, while the real-world problems which are not tractable for classical solvers often require more than that. It is clear that the problem size is an important factor as the currently available quantum solvers are limited in the number of qubits. There are several works that try to formulate models that are more efficient in the number of qubits used \cite{glos2020space, tabi2020quantum, campbell2021qaoa, mohammadbagherpoor2021exploring} and further research can be pursued in this direction.
A related and more general problem is product configuration and reconfiguration, where the problem's scope is not restricted to vehicles and one may consider any product such as computer parts. Satisfiability based approaches have been considered in \cite{singh2013generation,walter2014remax}. The presented model can be extended for such problems and evoke potential use-cases for D-Wave CQM Solver.
\section*{Acknowledgements}
We would like to thank Jarosław Miszczak and Krzysztof Domino for the discussion on the subject and their valuable comments on the report. This work has been partially supported by Polish National Science Center under the grant agreement 2019/33/B/ST6/02011. AG has been also supported by Polish National Science Center under the grant agreements 2020/37/N/ST6/02220. We would like to thank the organizers of \textit{BMW Group Quantum Computing Challenge} for providing us with the exemplary dataset used in this manuscript
\bibliographystyle{splncs04}
| 08e7907e3c17dba99b33d9e6baf6423a52f19e25 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
Subsets and Splits